paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
1806.00638 | 1 | 1806 | 2018-06-02T14:05:24 | On Minrank and Forbidden Subgraphs | [
"cs.DS",
"cs.IT",
"math.CO",
"cs.IT"
] | The minrank over a field $\mathbb{F}$ of a graph $G$ on the vertex set $\{1,2,\ldots,n\}$ is the minimum possible rank of a matrix $M \in \mathbb{F}^{n \times n}$ such that $M_{i,i} \neq 0$ for every $i$, and $M_{i,j}=0$ for every distinct non-adjacent vertices $i$ and $j$ in $G$. For an integer $n$, a graph $H$, and a field $\mathbb{F}$, let $g(n,H,\mathbb{F})$ denote the maximum possible minrank over $\mathbb{F}$ of an $n$-vertex graph whose complement contains no copy of $H$. In this paper we study this quantity for various graphs $H$ and fields $\mathbb{F}$. For finite fields, we prove by a probabilistic argument a general lower bound on $g(n,H,\mathbb{F})$, which yields a nearly tight bound of $\Omega(\sqrt{n}/\log n)$ for the triangle $H=K_3$. For the real field, we prove by an explicit construction that for every non-bipartite graph $H$, $g(n,H,\mathbb{R}) \geq n^\delta$ for some $\delta = \delta(H)>0$. As a by-product of this construction, we disprove a conjecture of Codenotti, Pudl\'ak, and Resta. The results are motivated by questions in information theory, circuit complexity, and geometry. | cs.DS | cs |
On Minrank and Forbidden Subgraphs
Ishay Haviv∗
Abstract
The minrank over a field F of a graph G on the vertex set {1, 2, . . . , n} is the minimum pos-
sible rank of a matrix M ∈ Fn×n such that Mi,i 6= 0 for every i, and Mi,j = 0 for every distinct
non-adjacent vertices i and j in G. For an integer n, a graph H, and a field F, let g(n, H, F)
denote the maximum possible minrank over F of an n-vertex graph whose complement con-
tains no copy of H. In this paper we study this quantity for various graphs H and fields F.
For finite fields, we prove by a probabilistic argument a general lower bound on g(n, H, F),
which yields a nearly tight bound of Ω(√n/ log n) for the triangle H = K3. For the real field,
we prove by an explicit construction that for every non-bipartite graph H, g(n, H, R) ≥ nδ for
some δ = δ(H) > 0. As a by-product of this construction, we disprove a conjecture of Code-
notti, Pudl´ak, and Resta. The results are motivated by questions in information theory, circuit
complexity, and geometry.
1 Introduction
An n × n matrix M over a field F is said to represent a digraph G = (V, E) with vertex set V =
6= 0 for every i, and Mi,j = 0 for every distinct i, j such that (i, j) /∈ E. The
{1, 2, . . . , n} if Mi,i
minrank of G over F, denoted minrkF(G), is the minimum possible rank of a matrix M ∈ Fn×n
representing G. The definition is naturally extended to (undirected) graphs by replacing every
edge with two oppositely directed edges. It is easy to see that for every graph G the minrank
parameter is sandwiched between the independence number and the clique cover number, that is,
α(G) ≤ minrkF(G) ≤ χ(G). For example, minrkF(Kn) = 1 and minrkF(Kn) = n for every field F.
The minrank parameter was introduced by Haemers in 1979 [16], and since then has attracted a
significant attention motivated by its various applications in information theory and in theoretical
computer science (see, e.g., [17, 7, 32, 26, 25, 19, 10]).
In this work we address the extremal behavior of the minrank parameter of n-vertex graphs
whose complements are free of a fixed forbidden subgraph. For two graphs G and H, we say that
G is H-free if G contains no subgraph, induced or not, isomorphic to H. For an integer n, a graph
H, and a field F, let g(n, H, F) denote the maximum of minrkF(G) taken over all n-vertex graphs
G whose complement G is H-free. Our purpose is to study the quantity g(n, H, F) where H and F
are fixed and n is growing.
∗School of Computer Science, The Academic College of Tel Aviv-Yaffo, Tel Aviv 61083, Israel.
1
1.1 Our Contribution
We provide bounds on g(n, H, F) for various graph families and fields. We start with a simple
upper bound for a forest H.
Proposition 1.1. For every integer n, a field F, and a nontrivial forest H on h vertices,
g(n, H, F) ≤ h − 1.
Equality holds whenever H is a tree and n ≥ h − 1.
We next provide a general lower bound on g(n, H, F) for a graph H and a finite field F. To
state it, we need the following notation. For a graph H with h ≥ 3 vertices and f ≥ 3 edges define
γ(H) = h−2
f−1 and γ0(H) = minH′ γ(H′), where the minimum is taken over all subgraphs H′ of H
with at least 3 edges.
Theorem 1.2. For every graph H with at least 3 edges there exists c = c(H) > 0 such that for every
integer n and a finite field F,
g(n, H, F) ≥ c ·
n1−γ0(H)
log(n · F)
.
Note that for every finite field F, the quantity g(n, H, F) grows with n if and only if H is not
a forest. Indeed, if H is a forest then g(n, H, F) is bounded by some constant by Proposition 1.1,
whereas otherwise H satisfies γ0(H) < 1 and thus, by Theorem 1.2, g(n, H, F) ≥ Ω(nδ) for some
δ = δ(H) > 0. Note further that for the case H = K3, which is motivated by a question in
information theory (see Section 1.2), Theorem 1.2 implies that
g(n, K3, F) ≥ Ω(cid:16)
√n
log n(cid:17)
(1)
for every fixed finite field F. This is tight up to a plog n multiplicative term (see Proposition 3.8).
Theorem 1.2 is proved by a probabilistic argument based on the Lov´asz Local Lemma [13]. The
proof involves an approach of Spencer [29] to lower bounds on off-diagonal Ramsey numbers and
a technique of Golovnev, Regev, and Weinstein [15] for estimating the minrank of random graphs.
As our final result, we show that for every non-bipartite graph H there are H-free graphs with
low minrank over the real field R.
Theorem 1.3. For every non-bipartite graph H there exists δ = δ(H) > 0 such that for every sufficiently
large integer n, there exists an n-vertex H-free graph G such that minrkR(G) ≤ n1−δ.
This theorem is proved by an explicit construction from the family of generalized Kneser graphs,
whose minrank was recently studied in [18]. It is known that every n-vertex graph G satisfies
for every field F (see, e.g., [23, Remark 2.2]). This combined with the graphs given in Theorem 1.3
implies the following (explicit) lower bound on g(n, H, R) for non-bipartite graphs H.
minrkF(G) · minrkF(G) ≥ n
(2)
Corollary 1.4. For every non-bipartite graph H there exists δ = δ(H) > 0 such that for every sufficiently
large integer n, g(n, H, R) ≥ nδ.
As another application of Theorem 1.3, we disprove a conjecture of Codenotti, Pudl´ak, and Resta [11]
motivated by Valiant's approach to circuit lower bounds [31] (see Section 1.2).
2
1.2 Applications
The study of the quantity g(n, H, F) is motivated by questions in information theory, circuit com-
plexity, and geometry. We gather here several applications of our results.
Shannon Capacity. For an integer k and a graph G on the vertex set V, let Gk denote the graph
on the vertex set Vk in which two distinct vertices (u1, . . . , uk) and (v1, . . . , vk) are adjacent if for
every 1 ≤ i ≤ k it holds that ui and vi are either equal or adjacent in G. The Shannon capacity of
a graph G, introduced by Shannon in 1956 [28], is defined as the limit c(G) = limk→∞ (α(Gk))1/k.
This graph parameter is motivated by information theory, as it measures the zero-error capacity
of a noisy communication channel represented by G. An upper bound on c(G), known as the
Lov´asz ϑ-function, was introduced in [22], where it was used to show that the Shannon capacity
of the cycle on 5 vertices satisfies c(C5) = √5, whereas its independence number is 2. Haemers
introduced the minrank parameter in [16, 17] and showed that it forms another upper bound
on c(G) and that for certain graphs it is tighter than the ϑ-function. In general, computing the
Shannon capacity of a graph seems to be a very difficult task, and its exact value is not known
even for small graphs such as the cycle on 7 vertices.
The question of determining the largest possible Shannon capacity of a graph with a given
independence number is widely open. In fact, it is not even known if the Shannon capacity of
a graph with independence number 2 can be arbitrarily large [3]. Interestingly, Erd os, McEliece,
and Taylor [14] have shown that this question is closely related to determining an appropriate
multicolored Ramsey number, whose study in [33] implies that there exists a graph G with α(G) =
2 and c(G) > 3.199. A related question, originally asked by Lov´asz, is that of determining the
maximum possible ϑ-function of an n-vertex graph with independence number 2. This maximum
is known to be Θ(n1/3), where the upper bound was proved by Kashin and Konyagin [20, 21],
and the lower bound was proved by Alon [2] via an explicit construction. Here we consider the
analogue question of determining the maximum possible minrank, over any fixed finite field F, of
an n-vertex graph with independence number 2. Since the latter is precisely g(n, K3, F), our bound
in (1) implies that the minrank parameter is weaker than the ϑ-function with respect to the general
upper bounds that they provide on the Shannon capacity of n-vertex graphs with independence
number 2.
In 1977, Valiant [31] proposed the matrix rigidity ap-
The Odd Alternating Cycle Conjecture.
proach for proving superlinear circuit lower bounds, a major challenge in the area of circuit com-
plexity. Roughly speaking, the rigidity of a matrix M ∈ Fn×n for a constant ε > 0 is the minimum
number of entries that one has to change in M in order to reduce its rank over F to at most ε · n.
Valiant showed in [31] that matrices with large rigidity can be used to obtain superlinear lower
bounds on the size of logarithmic depth arithmetic circuits computing linear transformations.
With this motivation, Codenotti, Pudl´ak, and Resta [11] raised in the late nineties the Odd Al-
ternating Cycle Conjecture stated below, and proved that it implies, if true, that certain explicit
circulant matrices have superlinear rigidity. By an alternating odd cycle we refer to a digraph
which forms a cycle when the orientation of the edges is ignored, and such that the orientation of
the edges alternates with one exception.
3
Conjecture 1.5 (The Odd Alternating Cycle Conjecture [11]). For every field F there exist ε > 0 and
an odd integer ℓ such that every n-vertex digraph G with minrkF(G) ≤ ε · n contains an alternating cycle
of length ℓ.
Codenotti et al. [11] proved that the statement of Conjecture 1.5 does not hold for ℓ = 3 over
any field F. Specifically, they provided an explicit construction of n-vertex digraphs G, free of
alternating triangles, with minrkF(G) ≤ O(n2/3) for every field F. For the undirected case, which
is of more interest to us, a construction of [11] implies that there are n-vertex triangle-free graphs
G such that minrkF(G) ≤ O(n3/4) for every field F (see [8, Section 4.2] for a related construction
over the binary field as well as for an application of such graphs from the area of index coding).
Note that this yields, by (2), that g(n, K3, F) ≥ Ω(n1/4). In contrast, for the real field and the cycle
on 4 vertices, it was shown in [11] that every n-vertex C4-free graph G satisfies minrkR(G) > n
6 .
Yet, the question whether every n-vertex digraph with sublinear minrank contains an alternating
cycle of odd length ℓ ≥ 5 was left open in [11] for every field. Our Theorem 1.3 implies that for
every odd ℓ there are (undirected) Cℓ-free graphs G with sublinear minrkR(G), and in particular
disproves Conjecture 1.5 for the real field R.
Nearly Orthogonal Systems of Vectors. A system of nonzero vectors in Rm is said to be nearly
orthogonal if any set of three vectors of the system contains an orthogonal pair. It was proved
by Rosenfeld [27] that every such system has size at most 2m. An equivalent way to state this, is
that every n-vertex graph represented by a real positive semidefinite matrix of rank smaller than
n
2 contains a triangle. Note that the positive semidefiniteness assumption is essential in this result,
as follows from the aforementioned construction of [11] of n-vertex triangle-free graphs G with
minrkR(G) ≤ O(n3/4).
A related question was posed by Pudl´ak in [24]. He proved there that for some ε > 0, every
n-vertex graph represented by a real positive semidefinite matrix of rank at most ε · n contains a
cycle of length 5. Pudl´ak asked whether the assumption that the matrix is positive semidefinite
can be omitted. Our Theorem 1.3 applied to H = C5 implies that there are C5-free graphs G with
sublinear minrkR(G), and thus answers this question in the negative.
1.3 Outline
The rest of the paper is organized as follows. In Section 2 we present the simple proof of Propo-
sition 1.1. In Section 3 we provide some background on sparse-base matrices from [15] and then
prove Theorem 1.2. In the final Section 4, we prove Theorem 1.3.
2 Forests
In this section we prove Proposition 1.1. We use an argument from one of the proofs in [5].
Proof of Proposition 1.1: Fix a nontrivial h-vertex forest H and a field F. It suffices to consider
the case where H is a tree, as otherwise H is a subgraph of some h-vertex tree H′, and since every
H-free graph is also H′-free, we have g(n, H, F) ≤ g(n, H′, F).
4
Our goal is to show that every n-vertex graph G whose complement G is H-free satisfies
minrkF(G) ≤ h − 1. Let G be such a graph. We claim that G is (h − 2)-degenerate, that is, ev-
ery subgraph of G contains a vertex of degree at most h − 2. Indeed, otherwise G has a subgraph
G′ all of whose degrees are at least h − 1, and one can find a copy of H in G′ as follows: First iden-
tify an arbitrary vertex of G′ with an arbitrary vertex of H, and then iteratively identify a vertex
of G′ with a leaf added to the being constructed copy of the tree H. The process succeeds since H
has h vertices and every vertex of G′ has degree at least h − 1. As is well known, the fact that G is
(h − 2)-degenerate implies that G is (h − 1)-colorable, so we get that minrkF(G) ≤ χ(G) ≤ h − 1,
as required.
We finally observe that the bound is tight whenever H is a tree and n ≥ h − 1. Indeed, let G
be the n-vertex complete ⌈ n
h−1⌉-partite graph, that has h − 1 vertices in each of its parts, except
possibly one of them. Its complement G is a disjoint union of cliques, each of size at most h − 1,
and is thus H-free. Since α(G) = χ(G) = h − 1, it follows that minrkF(G) = h − 1 for every field
F, completing the proof.
3 A General Lower Bound on g(n, H, F)
In this section we prove Theorem 1.2 and discuss its tightness for H = K3. We start with some
needed preparations.
3.1 Lov´asz Local Lemma
The Lov´asz Local Lemma [13] stated below is a powerful probabilistic tool in Combinatorics (see,
e.g., [6, Chapter 5]). We denote by [N] the set of integers from 1 to N.
Lemma 3.1. [Lov´asz Local Lemma [13]] Let A1, . . . , AN be events in an arbitrary probability space. A
digraph D = (V, E) on the vertex set V = [N] is called a dependency digraph for the events A1, . . . , AN if
for every i ∈ [N], the event Ai is mutually independent of the events Aj with j 6= i and (i, j) /∈ E. Suppose
that D = (V, E) is a dependency digraph for the above events and suppose that there are real numbers
x1, . . . , xN ∈ [0, 1) such that
Pr [Ai] ≤ xi · ∏
(i,j)∈E
(1 − xj)
for all i ∈ [N]. Then, with positive probability no event Ai holds.
3.2 Sparse-base Matrices
Here we review several notions and lemmas due to Golovnev, Regev, and Weinstein [15]. For a
matrix M over a field F, let s(M) denote its sparsity, that is, the number of its nonzero entries.
We say that a matrix M over F with rank k contains an ℓ-sparse column (row) basis if M contains
k linearly independent columns (rows) with a total of at most ℓ nonzero entries. We first state
a lemma that provides an upper bound on the number of matrices with sparse column and row
bases.
Lemma 3.2 ([15]). The number of rank k matrices in Fn×n that contain ℓ-sparse column and row bases is
at most (n · F)6ℓ.
5
The following lemma relates the sparsity of a matrix with nonzero entries on the main diagonal
to its rank.
Lemma 3.3 ([15]). For every rank k matrix M ∈ Fn×n with nonzero entries on the main diagonal,
s(M) ≥
n2
4k
.
We also need the following notion. An (n, k, s, ℓ)-matrix over a field F is a matrix in Fn×n of
rank k and sparsity s that contains ℓ-sparse column and row bases and has nonzero entries on the
4k . For integers
n,k be the collection that
n . This
main diagonal. Note that by Lemma 3.3, an (n, k, s, ℓ)-matrix exists only if s ≥ n2
n, k, s′ and a field F (which will always be clear from the context), let M(s′)
consists of all (n′, k′, s′, 2s′k′
collection is motivated by the following lemma.
n′ )-matrices over F for all n′ ∈ [n] and k′ ∈ [k] such that k′
n′ ≤ k
Lemma 3.4 ([15]). Every matrix in Fn×n with rank at most k and nonzero entries on the main diagonal
has a principal sub-matrix that lies in M(s′)
n,k for some s′.
Now, for integers n, k, s′, let P (s′)
n,k be the collection that consists of all pairs (M, R) such that, for
some n′ ∈ [n], M is an n′ × n′ matrix in M(s′)
n,k and R is an n′-subset of [n]. Observe that Lemma 3.4
implies that for every digraph G on the vertex set [n] with minrkF(G) ≤ k there exist s′ and a pair
(M, R) in P (s′)
n,k such that M represents the induced subgraph G[R] of G on R, with respect to the
natural order of the vertices in R (from smallest to largest).
The following lemma provides an upper bound on the size of P (s′)
n,k .
Lemma 3.5. For every integers n, k, s′, P (s′)
Proof: To bound the size of P (s′)
pairs (M, R) where M is an (n′, k′, s′, 2s′k′
are at most (n′ · F)12s′ k′/n′ such matrices M, each of which occurs in ( n
n,k , we consider for every n′ ∈ [n] and k′ ∈ [k] such that k′
n′) pairs of P (s′)
n the
n′ )-matrix and R is an n′-subset of [n]. By Lemma 3.2 there
n,k . It follows
n,k ≤ (n · F)24s′ k/n.
n′ ≤ k
that
n′,k′(cid:18) n
n′(cid:19) · (n′ · F)12s′ k′/n′ ≤ n2 · max
P (s′)
n,k ≤ ∑
n′,k′ (cid:0)n3n′ · (n′ · F)12s′ k′/n′(cid:1) ≤ max
≤ max
n′,k′ (cid:0)(n · F)12s′ k′/n′+12s′k′/n′(cid:1) ≤ (n · F)24s′ k/n,
≤ max
n′,k′ (cid:0)nn′ · (n′ · F)12s′ k′/n′(cid:1)
n′,k′ (cid:0)(n · F)3n′+12s′k′/n′(cid:1)
where in the fifth inequality we have used the relation s′ ≥ n′2
we have used k′
n′ ≤ k
n .
4k′ from Lemma 3.3, and in the sixth
3.3 Proof of Theorem 1.2
We prove the following theorem and then derive Theorem 1.2. Recall that for a graph H with
h ≥ 3 vertices and f ≥ 3 edges, we denote γ(H) = h−2
f−1. We also let exp(x) stand for ex.
6
Theorem 3.6. For every graph H with at least 3 edges there exists c = c(H) > 0 such that for every
integer n and a finite field F,
g(n, H, F) ≥ c ·
n1−γ(H)
log(n · F)
.
Proof: Fix a graph H with h ≥ 3 vertices and f ≥ 3 edges and denote γ = γ(H) = h−2
> 0. The
f−1
proof is via the probabilistic method. Let ~G ∼ ~G(n, p) be a random digraph on the vertex set [n]
where each directed edge is taken randomly and independently with probability p. Set q = 1 − p.
Let G be the (undirected) graph on [n] in which two distinct vertices i, j are adjacent if both the
directed edges (i, j) and (j, i) are included in ~G. Notice that every two distinct vertices are adjacent
in G with probability p2 independently of the adjacencies between other vertex pairs.
To prove the theorem, we will show that for a certain choice of p the random graph G satisfies
with positive probability that its complement G is H-free and that minrkF(G) > k, where
k = c1 ·
n1−γ
ln (n · F)
(3)
for a constant c1 > 0 that depends only on H. To do so, we define two families of events as follows.
First, for every set I ⊆ [n] of size I = h, let AI be the event that the induced subgraph of G
on I contains a copy of H. Observe that
Pr [AI] ≤ h! · (1 − p2) f = h! · (1 − (1 − q)2) f ≤ h! · (2q) f .
Second, consider the collection P = ∪s′∈[n2]P (s′)
n,k (see Section 3.2). Recall that every element of
P is a pair (M, R) such that, for some n′ ∈ [n], M is an n′ × n′ matrix over F and R is an n′-subset
of [n]. Denote Ns′ = P (s′)
n,k . By Lemma 3.5, combined with (3), we have
Ns′ ≤ (n · F)24s′ k/n = exp(24c1 · s′ · n−γ).
(4)
Let S = {s′ ∈ [n2] Ns′ ≥ 1}. By Lemma 3.3, for every s′ ∈ S and an n′ × n′ matrix of rank k′ in
M(s′)
n , we have that
n,k where n′ ∈ [n], k′ ∈ [k], and k′
n′ ≤ k
n′
n′
k′ ≥
4 ·
n′
4 ·
s′ ≥
n
k
= n′ ·
nγ · ln(n · F)
4c1
.
(5)
Now, for every pair (M, R) ∈ P, let BM,R be the event that the matrix M represents over F the
induced subgraph ~G[R] of ~G on R with respect to the natural order of the vertices in R. For M to
represent ~G[R] we require that for every distinct i, j such that Mi,j 6= 0, there is an edge in ~G from
the ith to the jth vertex of R. Hence, for M ∈ Fn′×n′ of sparsity s′ and an n′-subset R of [n],
Pr [BM,R] = ps′−n′ ≤ ps′/2 = (1 − q)s′/2 ≤ exp(−qs′/2),
where for the first inequality we have used the inequality s′ ≥ 2n′ which follows from (5) for every
sufficiently large n.
We claim that it suffices to prove that with positive probability none of the events AI and BM,R
holds. Indeed, this implies that there exists an n-vertex digraph ~G that does not satisfy any of
7
these events. Since the AI's are not satisfied it immediately follows that the complement G of the
(undirected) graph G associated with ~G is H-free. We further claim that minrkF(G) > k. To see
this, assume by contradiction that there exists a matrix M ∈ Fn×n of rank at most k that represents
G, and thus, in particular, represents ~G. By Lemma 3.4, such an M has a principal n′ × n′ sub-
matrix M′ ∈ M(s′)
n,k for some n′ and s′. Hence, for some n′-subset R of [n], the matrix M′ represents
~G[R] with respect to the natural order of the vertices in R, in contradiction to the fact that the event
BM′,R with (M′, R) ∈ P does not hold.
To prove that with positive probability none of the events AI and BM,R holds, we apply the
Lov´asz Local Lemma (Lemma 3.1). To this end, construct a (symmetric) dependency digraph
D = (V, E) whose vertices represent all the events AI and BM,R, and whose edges are defined as
follows.
• An AI-vertex and an AI′ -vertex are joined by edges (in both directions) if I ∩ I′ ≥ 2. Notice
that the events AI and AI′ are independent when I ∩ I′ < 2.
• An AI-vertex and a BM,R-vertex are joined by edges if there are distinct i, j ∈ I ∩ R for which
the entry of M that corresponds to the edge (i, j) is nonzero. Notice that the events AI and
BM,R are independent when such i and j do not exist.
• Every two distinct BM,R-vertices are joined by edges.
Clearly, each event is mutually independent of all other events besides those adjacent to it in D,
and thus D is a dependency digraph for our events. Observe that every AI-vertex is adjacent to
at most (h
2) · nh−2 AI′-vertices. Additionally, every BM,R-vertex, where M is an n′ × n′
matrix of sparsity s′, is adjacent to at most (s′ − n′) · ( n
h−2) < s′ · nh−2 AI-vertices. Finally, every
vertex of D is adjacent to at most Ns′ BM,R-vertices with M ∈ M(s′)
n,k (that is, s(M) = s′).
To apply Lemma 3.1 we assign a number in [0, 1) to each vertex of D. Define
2) · ( n
h−2) ≤ (h
q = c2 · n−γ,
x = c3 · n−γ· f ,
and xs′ = exp(−c4 · s′ · n−γ)
for every s′ ∈ S,
where c2, c3, c4 > 0 are constants, depending only on H, to be determined. We assign the number
x to every AI-vertex, and the number xs′ to every BM,R-vertex with s(M) = s′. We present now
the conditions of Lemma 3.1. For every AI-vertex, recalling that Pr [AI] ≤ h! · (2q) f , we require
h! · (2q) f ≤ x · (1 − x)(h
2)·nh−2
(1 − xs′)Ns′ .
· ∏
s′∈S
(6)
Similarly, for every BM,R-vertex with s(M) = s′, recalling that Pr [BM,R] ≤ exp(−qs′/2), we re-
quire
exp(−qs′/2) ≤ xs′ · (1 − x)s′·nh−2
(1 − xs′)Ns′ .
· ∏
s′∈S
(7)
To complete the proof, it suffices to show that the constants c1, c2, c3, c4 > 0 can be chosen in a
way that satisfies the inequalities (6) and (7). Consider the following three constraints:
1. c2 > 2 · (2c3 + c4),
2. c3 ≥ h! · (2c2) f · exp(3), and
8
3. c4 ≥ 32 · c1.
It is easy to see that it is possible to choose the constants under the above constraints. Indeed, by
f ≥ 3, for a sufficiently small choice of c2 > 0 one can take c3 with, say, an equality in Item 2 so
that some c4 > 0 satisfies Item 1. Then, c1 can be chosen as a positive constant satisfying Item 3.
We show now that such a choice satisfies (6) and (7) for every sufficiently large n. Note that we
use below several times the inequality 1 − α ≥ exp(−2α), which holds for any α ∈ [0, 1/2].
First, use (4) and the condition c4 ≥ 32 · c1 to obtain that
exp((24c1 − c4) · s′ · n−γ) ≤ ∑
xs′ · Ns′ ≤ ∑
∑
s′∈S
s′∈S
s′∈S
where the third inequality follows by s′ ≥ nγ·ln(n·F)
which we get from (5), and the fourth by
S ≤ n2. Considering the term ∏s′∈S (1 − xs′)Ns′ , which appears in both (6) and (7), we derive
that
exp(−8c1 · s′ · n−γ) ≤ ∑
s′∈S
exp(−2 ln n) ≤ 1,
4c1
(1 − xs′ )Ns′ ≥ ∏
s′∈S
For inequality (6), observe that
∏
s′∈S
exp(−2xs′ · Ns′ ) = exp(cid:16) − 2 · ∑
s′∈S
xs′ · Ns′(cid:17) ≥ exp(−2).
x · (1 − x)(h
2)·nh−2
· ∏
s′∈S
(1 − xs′)Ns′ ≥ x · exp(cid:16) − 2x ·(cid:18)h
2(cid:19) · nh−2(cid:17) · exp(−2)
= c3 · n−γ· f · exp(cid:16) − 2c3 · n−γ· f ·(cid:18)h
≥ h! · (2c2) f · n−γ· f · exp(cid:16)1 − 2c3 ·(cid:18)h
≥ h! · (2q) f ,
2(cid:19) · nh−2 − 2(cid:17)
2(cid:19) · n−γ(cid:17)
where for the second inequality we use c3 ≥ h! · (2c2) f · exp(3) and γ = h−2
use the assumption that n is sufficiently large. For inequality (7), observe that
f−1, and for the third we
xs′ · (1 − x)s′·nh−2
· ∏
s′∈S
(1 − xs′ )Ns′ ≥ xs′ · exp(−2x · s′ · nh−2) · exp(−2)
= exp(−c4 · s′ · n−γ) · exp(−2c3 · n−γ· f · s′ · nh−2) · exp(−2)
= exp(−(2c3 + c4) · s′ · n−γ − 2)
≥ exp(−(c2/2) · s′ · n−γ)
= exp(−qs′/2),
where for the second equality we again use the definition of γ, and for the second inequality we
use the condition c2 > 2 · (2c3 + c4), the fact that s′ · n−γ = ω(1) by (5), and the assumption that n
is sufficiently large. This completes the proof.
We can derive now Theorem 1.2. Recall that γ0(H) = minH′ γ(H′), where the minimum is
over all subgraphs H′ of H with at least 3 edges.
Proof of Theorem 1.2: For a graph H with h ≥ 3 vertices and f ≥ 3 edges, let H′ be a subgraph
of H with at least 3 edges such that γ0(H) = γ(H′). By Theorem 3.6 there exists c > 0 such that
g(n, H′, F) ≥ c ·
n1−γ0(H)
log(n · F)
9
for every integer n and a finite field F. Since every H′-free graph is also H-free, it follows that
g(n, H, F) ≥ g(n, H′, F) and we are done.
3.4 The Minrank of Graphs with Small Independence Number
For an integer t ≥ 3, g(n, Kt, F) is the maximum possible minrank over F of an n-vertex graph
with independence number smaller than t. For this case we derive the following corollary.
Corollary 3.7. For every t ≥ 3 there exists c = c(t) > 0 such that for every integer n and a finite field F,
g(n, Kt, F) ≥ c ·
n1− 2
t+1
log(n · F)
.
Proof: Apply Theorem 1.2 to the graph H = Kt, and notice that γ0(Kt) = γ(Kt) = t−2
( t
2)−1
= 2
t+1.
For H = K3, we observe that our lower bound on g(n, K3, F) is nearly tight.
Proposition 3.8. There exist constants c1, c2 > 0 such that for every integer n and a finite field F,
√n
log(n · F) ≤ g(n, K3, F) ≤ c2 ·r n
log n
.
c1 ·
Proof: For the lower bound apply Corollary 3.7 with t = 3. To prove the upper bound we need a
result of Ajtai et al. [1] which says that every triangle-free n-vertex graph has an independent set
of size Ω(pn · log n). By repeatedly omitting such independent sets it follows that the chromatic
number of such a graph is O(pn/ log n). Now, let G be an n-vertex graph whose complement G
is triangle-free. We get that minrkF(G) ≤ χ(G) ≤ O(pn/ log n), as required.
4 Non-bipartite Graphs
In this section we show that for every non-bipartite graph H there are H-free graphs with low
minrank over R, confirming Theorem 1.3. We start with the case where H is an odd cycle, and
since every non-bipartite graph contains an odd cycle the general result follows easily. The proof
is by an explicit construction from the following family of graphs.
Definition 4.1. For integers m ≤ s ≤ d, the graph K<(d, s, m) is defined as follows: the vertices are all
the s-subsets of [d], and two distinct sets A, B are adjacent if A ∩ B < m.
The minrank of such graphs over finite fields was recently studied in [18] using tools from [4].
The proof technique of [18] can be used for the real field as well, as shown below.
Proposition 4.2. For every integers m ≤ s ≤ d,
minrkR(K<(d, s, m)) ≤
s−m
∑
i=0 (cid:18)d
i(cid:19).
10
Proof: Let f : {0, 1}d × {0, 1}d → R be the function defined by
xiyi − j(cid:17)
s−1
j=m(cid:16)
∏
f (x, y) =
d
∑
i=1
for every x, y ∈ {0, 1}d. Expanding f as a linear combination of monomials, the relation z2 = z for
z ∈ {0, 1} implies that one can reduce to 1 the exponent of each variable occuring in a monomial.
It follows that f can be represented as a multilinear polynomial in the 2d variables of x and y. By
combining terms involving the same monomial in the variables of x, one can write f as
f (x, y) =
R
∑
i=1
gi(x)hi(y)
for an integer R and functions gi, hi : {0, 1}d → R, i ∈ [R], such that the gi's are distinct multilinear
monomials of total degree at most s − m in d variables. It follows that R ≤ ∑s−m
Now, let M1 and M2 be the 2d × R matrices whose rows are indexed by {0, 1}d and whose
columns are indexed by [R], defined by (M1)x,i = gi(x) and (M2)x,i = hi(x). Then, the matrix
M = M1 · MT
Finally, let V be the vertex set of K<(d, s, m), that is, the collection of all s-subsets of [d], and
identify every vertex A ∈ V with an indicator vector cA ∈ {0, 1}d in the natural way. We claim
that the matrix M restricted to V × V represents the graph K<(d, s, m). Indeed, for every A, B ∈ V
we have
2 has rank at most R and for every x, y ∈ {0, 1}d it holds that Mx,y = f (x, y).
i=0 (d
i ).
McA,cB = f (cA, cB) =
s−1
j=m(cid:16)A ∩ B − j(cid:17).
∏
Hence, for every A ∈ V we have A = s and thus McA,cA 6= 0, whereas for every distinct non-
adjacent A, B ∈ V we have m ≤ A ∩ B ≤ s − 1 and thus McA,cB = 0. Since the restriction of M to
V × V has rank at most R it follows that minrkR(K<(d, s, m)) ≤ R, and we are done.
We turn to identify graphs K<(d, s, m) with no short odd cycles. For this purpose, take an
even integer d, s = d
2 , and m = ε · d for a small constant ε > 0. Every path in these graphs is a
sequence of d
2 -subsets of [d] such that the intersection size of every two consecutive sets is small.
This implies, for a sufficiently small ε, that the sets in the even positions of the path are almost
disjoint from the first set, whereas the sets in the odd positions of the path share with it many
elements, hence such a graph contains no short odd cycle. This is shown formally in the following
lemma.
Lemma 4.3. Let ℓ ≥ 3 be an odd integer. For every even integer d and an integer m ≤ d
K<(d, d
2 , m) contains no odd cycle of length at most ℓ.
2ℓ , the graph
Proof: Fix an odd integer ℓ ≥ 3, an even integer d, and an integer m ≤ d
2ℓ . We prove that for every
odd integer ℓ′, such that 3 ≤ ℓ′ ≤ ℓ, the graph K<(d, d
2 , m) contains no cycle of length ℓ′. For such
an ℓ′, let A1, A2, . . . , Aℓ′ be a sequence of ℓ′ vertices in the graph, i.e., d
2 -subsets of [d]. Assuming
that for every i ≤ ℓ′ − 1 the vertices Ai and Ai+1 are adjacent in the graph, that is, Ai ∩ Ai+1 < m,
our goal is to show that A1 and Aℓ′ are not.
11
To this end, we argue that for every i, such that 0 ≤ i ≤ ℓ′−1
2 , we have
A1 ∩ A2i+1 ≥
d
2 − 2i · m.
(8)
We prove this claim by induction on i. The case i = 0 follows immediately from A1 = d
that (8) holds for i − 1, that is, A1 ∩ A2i−1 ≥ d
2 − (2i − 2) · m. Observe that this implies that
2 . Assume
A1 ∩ A2i = A1 ∩ A2i ∩ A2i−1 + A1 ∩ A2i ∩ A2i−1
≤ A2i−1 ∩ A2i + A1 ∩ A2i−1
≤ m + A1 − A1 ∩ A2i−1
≤ m +
2 −(cid:16) d
d
2 − (2i − 2) · m(cid:17) = (2i − 1) · m,
where in the second inequality we have used A2i−1 ∩ A2i < m. We proceed by proving (8) for i.
Observe that
A1 ∩ A2i+1 = A2i+1 − A1 ∩ A2i+1
= A2i+1 − A1 ∩ A2i+1 ∩ A2i − A1 ∩ A2i+1 ∩ A2i
≥
d
2 − m − A1 ∩ A2i,
where we have used A2i ∩ A2i+1 < m. Notice that
A1 ∩ A2i = d − A1 ∪ A2i = d − (A1 + A2i − A1 ∩ A2i) = A1 ∩ A2i.
It follows that
A1 ∩ A2i+1 ≥
completing the proof of (8).
Finally, applying (8) to i = ℓ′−1
d
2 − m − A1 ∩ A2i ≥
d
2 − m − (2i − 1) · m =
d
2 − 2i · m,
2 , using the assumption m ≤ d
d
d
2 − ℓ′ · m + m ≥
2 − ℓ · m + m ≥ m,
d
2 − (ℓ′ − 1) · m =
2ℓ , we get that
A1 ∩ Aℓ′ ≥
hence A1 and Aℓ′ are not adjacent in the graph K<(d, d
no cycle of length ℓ′, as desired.
2 , m). It thus follows that the graph contains
Equipped with Proposition 4.2 and Lemma 4.3, we obtain the following.
Theorem 4.4. For every odd integer ℓ ≥ 3 there exists δ = δ(ℓ) > 0 such that for every sufficiently large
integer n, there exists an n-vertex graph G with no odd cycle of length at most ℓ such that
minrkR(G) ≤ n1−δ.
Proof: Fix an odd integer ℓ ≥ 3. For an integer d divisible by 2ℓ, consider the graph G =
K<(d, d
2ℓ . By Lemma 4.3, G contains no odd cycle of length at most ℓ. As for
the minrank, Proposition 4.2 implies that
2 , m) where m = d
minrkR(G) ≤
d/2−m
∑
i=0 (cid:18)d
i(cid:19) ≤ 2H( 1
2− m
d )·d = 2H( 1
2− 1
2ℓ )·d,
12
where H stands for the binary entropy function. Since G has V = ( d
d/2) = 2(1−o(1))·d vertices,
2ℓ ) < 1 − δ we have minrkR(G) ≤ V1−δ for every sufficiently
for any δ > 0 such that H( 1
large integer d. The proof is completed by considering, for every sufficiently large integer n, some
n-vertex subgraph of the graph defined above, where d is the smallest integer divisible by 2ℓ such
that n ≤ ( d
2 − 1
d/2).
Now, Theorem 1.3 follows easily from Theorem 4.4.
Proof of Theorem 1.3: Let H be a non-bipartite graph. Then, for some odd integer ℓ ≥ 3, the
cycle Cℓ is a subgraph of H. By Theorem 4.4, there exists δ > 0 such that for every sufficiently
large integer n, there exists an n-vertex Cℓ-free graph G satisfying minrkR(G) ≤ n1−δ. Since every
Cℓ-free graph is also H-free, the result follows.
Remark 4.5. As mentioned in the introduction, Theorem 1.3 implies a lower bound on g(n, H, R) for every
non-bipartite graph H (see Corollary 1.4). We note that upper bounds on certain Ramsey numbers can be
used to derive upper bounds on g(n, H, F) for a general field F. For example, it was shown in [12] that for
every ℓ ≥ 3, every n-vertex Cℓ-free graph has an independent set of size Ω(n1−1/k) for k = ⌈ ℓ
2⌉ (see [9, 30]
for slight improvements). By repeatedly omitting such independent sets it follows that the chromatic number
of such a graph is O(n1/k). This implies that every n-vertex graph G whose complement is Cℓ-free satisfies
minrkF(G) ≤ χ(G) ≤ O(n1/k), hence g(n, Cℓ, F) ≤ O(n1/k).
Acknowledgements
We are grateful to Alexander Golovnev and Pavel Pudl´ak for useful discussions and to the anony-
mous referees for their valuable suggestions.
References
[1] M. Ajtai, J. Koml ´os, and E. Szemer´edi. A note on Ramsey numbers. J. Comb. Theory, Ser. A,
29(3):354–360, 1980.
[2] N. Alon. Explicit Ramsey graphs and orthonormal labelings. Electr. J. Comb., 1(R12), 1994.
[3] N. Alon. Graph powers. In B. Bollob´as, editor, Contemporary Combinatorics, Bolyai Society
Mathematical Studies, pages 11–28. Springer, 2002.
[4] N. Alon, L. Babai, and H. Suzuki. Multilinear polynomials and Frankl–Ray-Chaudhuri–
Wilson type intersection theorems. J. Comb. Theory, Ser. A, 58(2):165–180, 1991.
[5] N. Alon, M. Krivelevich, and B. Sudakov. Maxcut in H-free graphs. Combinatorics, Probability
and Computing, 14(5-6):629–647, 2005.
[6] N. Alon and J. H. Spencer. The Probabilistic Method. Wiley Publishing, 4th edition, 2016.
[7] Z. Bar-Yossef, Y. Birk, T. S. Jayram, and T. Kol. Index coding with side information. In FOCS,
pages 197–206, 2006.
13
[8] A. Blasiak, R. Kleinberg, and E. Lubetzky. Broadcasting with side information: Bounding and
approximating the broadcast rate. IEEE Trans. Information Theory, 59(9):5811–5823, 2013.
[9] Y. Caro, Y. Li, C. C. Rousseau, and Y. Zhang. Asymptotic bounds for some bipartite graph:
complete graph Ramsey numbers. Discrete Mathematics, 220(1-3):51–56, 2000.
[10] E. Chlamt´ac and I. Haviv. Linear index coding via semidefinite programming. Combinatorics,
Probability & Computing, 23(2):223–247, 2014. Preliminary version in SODA'12.
[11] B. Codenotti, P. Pudl´ak, and G. Resta. Some structural properties of low-rank matrices related
to computational complexity. Theor. Comput. Sci., 235(1):89–107, 2000. Preliminary version in
ECCC'97.
[12] P. Erd os, R. J. Faudree, C. C. Rousseau, and R. H. Schelp. On cycle-complete graph Ramsey
numbers. J. Graph Theory, 2(1):53–64, 1978.
[13] P. Erd os and L. Lov´asz. Problems and results on 3-chromatic hypergraphs and some related
questions. In A. Hajnal, R. Rado, and V. T. S ´os, editors, Infinite and Finite Sets, pages 609–627.
North-Holland, Amsterdam, 1975.
[14] P. Erd os, R. J. McEliece, and H. Taylor. Ramsey bounds for graph products. Pacific J. Math.,
37(1):45–46, 1971.
[15] A. Golovnev, O. Regev, and O. Weinstein. The minrank of random graphs. In Randomization
and Approximation Techniques in Computer Science (RANDOM), pages 46:1–46:13, 2017.
[16] W. Haemers. On some problems of Lov´asz concerning the Shannon capacity of a graph. IEEE
Trans. Inform. Theory, 25(2):231–232, 1979.
[17] W. Haemers. An upper bound for the Shannon capacity of a graph. In Algebraic methods in
graph theory, Vol. I, II (Szeged, 1978), volume 25 of Colloq. Math. Soc. J´anos Bolyai, pages 267–272.
North-Holland, Amsterdam, 1981.
[18] I. Haviv. On minrank and the Lov´asz theta function. In Approximation Algorithms for Combi-
natorial Optimization Problems (APPROX), 2018. To appear.
[19] I. Haviv and M. Langberg. H-wise independence.
In Innovations in Theoretical Computer
Science (ITCS'13), pages 541–552, 2013.
[20] B. S. Kashin and S. V. Konyagin. Systems of vectors in Hilbert space.
In Number theory,
mathematical analysis, and their applications, volume 157 of Trudy Mat. Inst. Steklov., pages 64–
67. 1981.
[21] S. V. Konyagin. Systems of vectors in Euclidean space and an extremal problem for polyno-
mials. Mat. Zametki, 29(1):63–74, 1981.
[22] L. Lov´asz. On the Shannon capacity of a graph. IEEE Trans. Inform. Theory, 25(1):1–7, 1979.
[23] R. Peeters. Orthogonal representations over finite fields and the chromatic number of graphs.
Combinatorica, 16(3):417–431, 1996.
14
[24] P. Pudl´ak. Cycles of nonzero elements in low rank matrices. Combinatorica, 22(2):321–334,
2002.
[25] P. Pudl´ak, V. R odl, and J. Sgall. Boolean circuits, tensor ranks, and communication complex-
ity. SIAM J. Comput., 26(3):605–633, 1997.
[26] S. Riis. Information flows, graphs and their guessing numbers. Electr. J. Comb., 14(1), 2007.
[27] M. Rosenfeld. Almost orthogonal lines in Ed. DIMACS Series in Discrete Math., 4:489–492,
1991.
[28] C. E. Shannon. The zero error capacity of a noisy channel. Institute of Radio Engineers, Trans-
actions on Information Theory, IT-2:8–19, 1956.
[29] J. Spencer. Asymptotic lower bounds for Ramsey functions. Discrete Mathematics, 20:69–76,
1977.
[30] B. Sudakov. A note on odd cycle-complete graph Ramsey numbers. Electr. J. Comb., 9(1), 2002.
[31] L. G. Valiant. Graph-theoretic arguments in low-level complexity. In Mathematical Foundations
of Computer Science (MFCS), 6th Symposium, pages 162–176, 1977.
[32] L. G. Valiant. Why is Boolean complexity theory difficult? In Poceedings of the London Mathe-
matical Society symposium on Boolean function complexity, volume 169, pages 84–94, 1992.
[33] X. Xu, X. Zheng, G. Exoo, and S. P. Radziszowski. Constructive lower bounds on classical
multicolor Ramsey numbers. Electr. J. Comb., 11(1), 2004.
15
|
1805.10708 | 1 | 1805 | 2018-05-27T23:01:25 | Distributed Treewidth Computation | [
"cs.DS"
] | Of all the restricted graph families out there, the family of low treewidth graphs has continuously proven to admit many algorithmic applications. For example, many NP-hard algorithms can be solved in polynomial time on graphs of constant treewidth. Other algorithmic techniques, such as Baker's technique, partition the graph into components of low treewidth. Therefore, computing the treewidth of a graph remains an important problem in algorithm design. For graphs of constant treewidth, linear-time algorithms are known in the classical setting, and well as $\text{polylog}(n)$-time parallel algorithms for computing an $O(1)$-approximation to treewidth. However, nothing is yet known in the distributed setting.
In this paper, we give near-optimal algorithms for computing the treewidth on a distributed network. We show that for graphs of constant treewidth, an $O(1)$-approximation to the treewidth can be computed in near-optimal $\tilde O(D)$ time, where $D$ is the diameter of the network graph. In addition, we show that many NP-hard problems that are tractable on constant treewidth graphs can also be solved in $\tilde O(D)$ time on a distributed network of constant treewidth.
Our algorithms make use of the shortcuts framework of Ghaffari and Haeupler [SODA'16], which has proven to be a powerful tool in designing near-optimal distributed algorithms for restricted graph networks, such as planar graphs, low-treewidth graphs, and excluded minor graphs. | cs.DS | cs |
Distributed Treewidth Computation
Jason Li1
1 Carnegie Mellon University, [email protected]
Abstract
Of all the restricted graph families out there, the family of low treewidth graphs has continu-
ously proven to admit many algorithmic applications. For example, many NP-hard algorithms
can be solved in polynomial time on graphs of constant treewidth. Other algorithmic techniques,
such as Baker's technique, partition the graph into components of low treewidth. Therefore,
computing the treewidth of a graph remains an important problem in algorithm design. For
graphs of constant treewidth, linear-time algorithms are known in the classical setting, and well
as polylog(n)-time parallel algorithms for computing an O(1)-approximation to treewidth. How-
ever, nothing is yet known in the distributed setting.
In this paper, we give near-optimal algorithms for computing the treewidth on a distributed
network. We show that for graphs of constant treewidth, an O(1)-approximation to the treewidth
can be computed in near-optimal O(D) time, where D is the diameter of the network graph. In
addition, we show that many NP-hard problems that are tractable on constant treewidth graphs
can also be solved in O(D) time on a distributed network of constant treewidth.
Our algorithms make use of the shortcuts framework of Ghaffari and Haeupler [SODA'16],
which has proven to be a powerful tool in designing near-optimal distributed algorithms for
restricted graph networks, such as planar graphs, low-treewidth graphs, and excluded minor
graphs.
Introduction
1
In recent years, there has been an increasing interest in fast distributed algorithms on restricted
√
graph families. Part of this recent action stemmed from the widespread lower bound of Ω(
n+D)
for distributed algorithms on general graphs [5], which holds for many basic graph optimization
problems. By restricting the graph networks to exclude the pathological lower bound instances
in [5], researchers have found success in beating the lower bound on nontrivial families of graph
networks. For example, there are now distributed MST algorithms running in near-optimal
O(D · no(1)) time on planar graphs, bounded treewidth graphs, and graphs with small mixing
time [7, 10, 11, 8]. Adding onto this line of work, this paper investigates many algorithmic
problems on graphs networks of bounded treewidth and gives efficient distributed algorithms
running in near-optimal O(D) 1 rounds on these networks, where D is the diameter of the
network graph.
The concept of treewidth, which dates back to the study of graph minors of Robertson and
Seymour [18], has proven fruitful in the quest for efficient classical algorithms for computationally
intractable problems. For problems of bounded treewidth, many difficult, NP-hard problems can
be solved in polynomial time. Then, with the increasing popularity of the parallel PRAM model,
the classical bounded treewidth algorithms were adapted to run in parallel [15, 16]. However,
until this paper, nothing was yet known in the distributed setting.
1.1 Results
Our main result is a distributed O(1)-approximation algorithm to compute the treewidth of a
network graph in O(kO(k)
D) rounds of the CONGEST model, where k is the treewidth of the
1 We use O(·) notation to hide polylogarithmic factors in n.
© Jason Li;
licensed under Creative Commons License CC-BY
42nd Conference on Very Important Topics (CVIT 2016).
Editors: John Q. Open and Joan R. Access; Article No. 23; pp. 23:1–23:26
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
23:2
Distributed Treewidth Computation with Applications
network graph G. To state this result in an approximation setting, we say that the algorithm
distinguishes the instances where tw(G) > k and the instances tw(G) ≤ O(k).
(cid:73) Theorem 1. Given a network graph G and integer k, there is a distributed algorithm running
in O(kO(k)
D) rounds and either correctly concludes that tw(G) > k, or correctly concludes that
tw(G) ≤ 7k + 4. Every node in the network should know the conclusion of the algorithm.
Of course, to approximate the treewidth of a network graph G, we simply run the above
algorithm with increasing values of k = 1, 2, 3, . . . until the algorithm outputs "tw(G) ≤ 7k + 4";
the running time will be dominated by the last, successful k. Also, observe that the diameter
factor, D, is necessary in the running time.
Intuitively, this is because treewidth is a global
property of a graph;
in other words, one cannot say anything about the treewidth by only
looking at a local neighborhood around a vertex. For example, given any approximation factor
α, consider the network graph consisting of a path of length Ω(n) with a clique of size 2α attached
to one end. For the information of the 2α-clique to reach the node v on the other end of the
path, the number of rounds required is at least the length of the path, which is Ω(n). Before
the 2α-clique reaches node v, the only nodes that v can possibly learn in the network form a
path, which has treewidth 1. Therefore, for node v to distinguish between treewidth 1 (the path
without the clique) and treewidth 2α − 1 (the path with the clique attached), Ω(n) rounds are
needed. Hence, to obtain an α-approximation, Ω(n) = Ω(D) rounds are necessary.
Our distributed algorithm follows the outline of the parallel algorithm of Lagergren [15],
which approximates the treewidth in kO(k)polylog(n) parallel time. The algorithm of [15] makes
repeated calls to an algorithm that finds vertex disjoint paths between two given vertices. Our
main technical contribution is a distributed algorithm solving this vertex disjoint paths problem
on a graph network of treewidth k in O(kO(1)
D) time. This algorithm resembles the parallel
vertex disjoint paths problem [14], but new ideas are required to construct a distributed algorithm.
Our main insight is in viewing the algorithm of [14] in a graph contraction-based setting, and then
applying the recent technology of partwise aggregation in distributed computing [7, 9, 14]. Outside
of this subroutine, we adapt the framework of [15], which computes a treewidth decomposition
given this subroutine, to the distributed setting.
(cid:73) Lemma 2. Given a graph G = (V, E) of treewidth at most k and two vertices s, t ⊆ V , we
can either find k vertex-disjoint s-t paths, or output an s-t node cut of size less than k, in
O(kO(1)
D) rounds. In the former case, every node knows whether it is on a path, and if so,
its predecessor and successor on that path. In the latter case, every node knows the fact that k
vertex-disjoint paths do not exist, as well as whether it is in the node cut.
Perhaps more importantly, the algorithm of Theorem 1 also outputs a distributed version of a
treewidth decomposition. Using this decomposition, we can solve many computationally difficult
problems on bounded treewidth graph networks like in the classical setting.
(cid:73) Theorem 3. Let G be a graph network with treewidth k. The problems maximum independent
set, minimum vertex cover, chromatic number, and minimum dominating set can be solved in
O(kO(k)
D) rounds on network G.
We remark that we can extend Theorem 3 to solve many other optimization problems that
are tractable on bounded-treewidth graphs.
1.2 Related Work
The shortcuts framework was introduced by Ghaffari and Haeupler [7], who used it to solve
MST and (1 + )-approximate minimum cut on planar graphs in near-optimal O(D) time. This
framework was expanded on in [10, 11], generalizing these algorithms to run on bounded-genus
J. Li
23:3
and bounded-treewidth graphs in O(D) time. Lately, [13] studied the shortcuts framework on
minor-free graphs, leading to O(D
2)-round distributed algorithms for these problems on graphs
excluded a fixed minor.
Efficient algorithms that do not use the shortcuts framework also exist. Recently, Ghaffari
et al. [8] give a distributed algorithm for MST in time proportional to the mixing time of the
network graph. For well-mixing graphs, such as expanders and random graphs, the algorithm
runs in 2O(√log n log log n) time.
The graph-theoretic property treewidth was introduced in Robertson and Seymour's study
of graph minors [18]. Since then, it has seen many algorithmic applications in solving NP-
hard problems, such as maximum independent set and chromatic number, efficiently on bounded
treewidth graphs; for an extensive study, see [17].
The algorithmic problem of computing or approximating the treewidth of a graph has also
been studied extensively. Computing the treewidth exactly is NP-hard [2], but admits an
√
O(
log k) approximation in polynomial time [1, 6], where k is the treewidth of the input graph.
For small values of k, faster algorithms were known since the work of Robertson and Seymour [18].
Bodlaender [3] gave the first linear-time algorithm for fixed k, running in 2O(k3)
n time. In the par-
allel setting, computing an O(1)-approximation of the treewidth can be done in kO(k)polylog(n)
time using O(n) processors [15, 16].
Preliminaries
2
All of our algorithms work under the CONGEST model of distributed computing. There is a
network G = (V, E) of n nodes and diameter D. On each synchronous round, every node can
send an O(log n)-bit message to each of its neighbors in G, possibly a different message to each
neighbor. We assume that between synchronous rounds, every node can perform unbounded
local computation.
Throughout the paper, G will always represent the graph network, and D its diameter. Given
any graph H, let V (H) and E(H) denote the vertices and edges of H. For a vertex v ∈ V , denote
N(v) as the neighbors of v in G. For a vertex set S ⊆ V , denote N(S) as the neighbor of set
S in G, i.e., the vertices in V − S with a neighbor in S, and define N[S] := N(S) ∪ S. A path
in H is a sequence of vertices such that adjacent vertices are neighbors in H. A path is simple
if no vertex appears twice on the path. For vertices s, t ∈ V , an s-t path is a (not necessarily
simple) path connecting s and t, and a vertex set S is called a vertex cut if it intersects every
s-t path in G.
Lastly, for a positive integer r, let [r] denote the set {1, 2, . . . , r} of positive integers from 1
to r.
2.1 Partwise and Subgraph Aggregation
The shortcuts framework of Ghaffari and Haeupler [7] has proved fruitful in designing distributed
algorithms on restricted graph families. The inner workings of shortcuts is not necessary for the
scope of this paper. Rather, we abstract out the primary task that is accomplished through the
shortcuts framework, which we define as Partwise Aggregation (PA), following [9].
(cid:73) Definition 4 (Partwise Aggregation [9]). Let G = (V, E) be a network graph, and let P =
(P1, . . . , PP) be a collection of pairwise disjoint and connected subsets Pi ⊆ V , called parts.
For each part Pi, every node v ∈ Pi knows the set N(v) ∩ Pi, i.e., which of its neighbors belong
i Pi has an integer xv of O(log n) bits, and let ⊕
be an associative function operating on integers of length O(log n). Every node in Pi wants to
xv, i.e., the aggregate ⊕ of all of the values xv in Pi. We call such a task
partwise aggregation with operator ⊕.
to its part. Suppose that every node v ∈ S
learn the valueL
v∈Pi
CVIT 2016
23:4
Distributed Treewidth Computation with Applications
The power of the shortcuts framework is that it allows us to solve this partwise aggregation
(PA) problem quickly, especially if the network graph G has special structure. We provide an
intuitive description below, referring the reader to [7] for more details. In an ideal case, if for
every part Pi, the diameter of G[Pi] is O(D), then every part Pi can simply aggregate inside G[Pi]
in O(D) rounds. The trouble is when the diameter of G[Pi] much larger than D, the diameter
of G. The shortcuts framework resolves this issue by allowing these "long and skinny" parts to
borrow edges from the rest of the network G to aid in their partwise aggregation. For a part
Pi, the borrowed edges Hi should satisfy the property that the diameter of G[Pi] ∪ Hi is now
comparable to the diameter of G. On the other hand, it is not ideal for a single edge in G to be
borrowed by too many parts, since it would induce "congestion" along the edge. The shortcuts
framework computes an appropriate set Hi of borrowed edges for each part Pi while satisfying
two conditions: (i) the diameter of G[Pi] ∩ Hi is small for each Pi, and (ii) no edge in G appears
in too many Hi. Each part Pi then executes its aggregation on the graph G[Pi] ∩ Hi.
A recent line of work [10, 11, 13] has built on the initial shortcuts framework of [7], leading
to near-optimal PA algorithms for many special classes of graphs.
(cid:73) Theorem 5 ([7, 10, 11, 13]). For any associative operator ⊕, we can solve the partwise aggreg-
ation problem in O(QG) rounds, where QG is a parameter that depends on the graph G and its
diameter D, as follows:
√
1. For all graphs G, QG = O(
n + D).
√
2. If G has genus g, QG = O(
g + 1D).
3. If G has treewidth k, QG = O(kD).
4. If G excludes a fixed minor H, QG = O(f(H) · D
2), where f is a function depending only on
H.
We will define a PA round to be one iteration of PA, where every node participating in PA
initially knows its part ID, its value xv, and the common operator ⊕, and at the end, every node
learns the aggregate ⊕ of its part. Observe that the well-studied broadcast procedure can be
formulated as a PA problem: if a leader node v in a part Pi wants to broadcast its value x, then
we set xv ← x, xu ← −∞ (more precisely, some O(log n)-bit encoding of −∞) for all u ∈ Pi − v,
and ⊕ to be the max function.
The PA round assumes that every node knows the ID of its part. Often, we will not have this
luxury: each node does not know its part ID, but only which of its neighbors also belong to its
part, and in some cases, only a subset of this. Below, we formulate an aggregation task with this
weaker assumption, and show that it can be solved using O(log n) iterations of PA as defined in
Definition 4.
(cid:73) Definition 6 (Subgraph Aggregation). Let G = (V, E) be a network graph, let P = (P1, . . . , PP)
be a collection of parts, and for each Pi ∈ P, let Hi be a connected subgraph of G on the nodes
in Pi, not necessarily the induced graph G[Pi]. Suppose that for each subgraph Hi, every node
in V (Hi) knows its neighbors in the subgraph Hi and nothing else. Suppose that every node
i Pi has an integer xv of O(log n) bits, and let ⊕ be an associative function operating
xv, i.e., the
aggregate ⊕ of all of the values xv in Pi. We call such a task subgraph aggregation with
operator ⊕.
v ∈ S
on integers of length O(log n). Every node in Pi wants to learn the value L
v∈Pi
Likewise, we define SA round to be one iteration of subgraph aggregation (SA), where every
node v ∈ Pi knows its neighbors in Hi, its value xv, and the common operator ⊕, and at the
end, learns the aggregate ⊕ of its part. The following lemma shows that while SA has a weaker
assumption, it is no harder than partwise aggregation modulo an O(log n) factor. While this
result has been implied in the literature, e.g., in [12], this is the first time it has been stated
explicitly. We defer the proof of the following lemma to Appendix A.
J. Li
23:5
(cid:73) Lemma 7. One SA round can be solved in O(log n) PA rounds.
Combining Lemma 7 with Theorem 5 gives the following corollary for treewidth k graphs,
which is the result we will use in this paper. Since SA solves a stronger problem than PA or
broadcast, we will only use the term SA for the rest of the paper. In fact, the following corollary
will be the only result in this section that is used for the rest of the paper.
(cid:73) Corollary 8. On a graph network of treewidth k, one SA round can be solved in O(kD) rounds.
2.2 SA Helper Routines
Our main algorithms will use the following routines, all of which reduce to computing O(log n)
rounds of SA. The proof of the statement below is deferred to Appendix A.
(cid:73) Lemma 9 (Spanning Tree). Given a connected subgraph H ⊆ G of the network graph, we can
compute a spanning tree of G in O(log n) SA rounds. Every node knows its neighbors of the
spanning tree.
(cid:73) Lemma 10 (Rooted Tree Aggregation). Consider a tree T in G. Given a root vr ∈ V (T), we
can compute the tree T rooted at vr in O(log n) SA rounds, so that every node in V (T) − vr
knows its parent in the tree T rooted at vr. Moreover, if each node vi knows an integer xi,
and a common associative operator ⊕, then we can let each node vi learn the subtree aggregate
j∈T (vi) xj, where T(vi) is the subtree rooted at vi, i.e., all nodes in T whose path to the root
L
contains vi.
(cid:73) Lemma 11 (Path Aggregation). Consider a directed path P = {v1, . . . , v'} in G, where each
node vi knows its predecessor and successor on the path. In O(log n) SA rounds, each node vi
can learn the value of i, its index in the path. Moreover, if each node vi knows an integer xi and
j≤i xj and
a common associative operator ⊕, we can let each node vi learn the prefix aggregateL
suffix aggregateL
j≥i xj.
(cid:73) Lemma 12 (s-t Path). Given a connected subgraph H ⊆ G and two vertices s, t ∈ V , we can
compute a directed s-t path in G in O(log n) SA rounds. Every node knows whether it is on the
path, and if so, its predecessor and successor nodes on that path.
The Vertex-disjoint Paths Algorithm
3
This section is devoted to proving the following lemma. It is the most technical section of the
paper, as well as our main technical contribution.
(cid:73) Lemma 13. Given a graph G = (V, E) of treewidth at most k and two vertices s, t ⊆ V , we
can either find k vertex-disjoint s-t paths, or output an s-t node cut of size less than k, in
O(kO(1)
D) rounds. In the former case, every node knows whether it is on a path, and if so,
its predecessor and successor on that path. In the latter case, every node knows the fact that k
vertex-disjoint paths do not exist, as well as whether it is in the node cut.
When we are talking about graph algorithms in general, not necessarily in the distributed
setting, we will use the term vertices. When talking about actual nodes in a distributed network,
we will use the term nodes. This is to distinguish between graphs (in the graph-theoretic sense)
and the physical graph network. Often times, it is simpler to first explain an algorithm in a
classical setting, and then adapt it to run on a distributed network. For most of this section, we
will take this approach, explaining our distributed implementation in gray boxes.
CVIT 2016
23:6
Distributed Treewidth Computation with Applications
1
2
1
1
1
, . . . , u' and the node set v
1
, . . . , v' with a complete bipartite graph Kr,r.
Before introducing our distributed algorithm, we first make one important transformation
of the network graph that will be useful later on. Define G' to be the following graph:
for each node v ∈ G, we add r corresponding nodes v
, . . . , v' in G'. For each v ∈ G, we
connect the nodes v
, . . . , v' in a clique Kr, and for each edge (u, v) in G, we connect the
1
node set u
We would like to simulate the network G' using the network G. In particular, we show
how to simulate a T-round algorithm on the network G' in O('
T) rounds on the original
network G.
We let every node v ∈ V (G) in the original network simulate the nodes v
, . . . , v' in G'.
That is, node v performs the computations that nodes v
, . . . , v' perform in the algorithm
on G'. Observe that two nodes ui, vj ∈ V (G') can communicate in G' iff either u = v, or the
nodes u, v ∈ V (G) that simulate them can communicate in G. If u = v, then since the same
node u simulates ui and vj, no communication between nodes in G is needed. If u 6= v, then
the edge (u, v) ∈ E(G) in the original network is responsible for '
2 communicating edges in
G', namely the edges (ui, vj) for i, j ∈ [']. Therefore, for each round of the algorithm on G',
2 messages in G' through the edge (u, v) ∈ E(G).
we can take '
It is easy to see that the diameter of G' is also D. The claim below, whose proof is deferred
to Appendix B, bounds the treewidth of G'. By Corollary 8, one SA round on G' can be
solved in O(k'D) rounds on O(k'D), and the above argument shows that it can be simulated
in O(k'
(cid:73) Claim 14. If G has treewidth k, then G' has treewidth O(k').
2 rounds to pass the at most '
3
D) rounds on G.
For the rest of this section, we will always either run our distributed algorithm directly on
G, or run it first on G' for some ' and then simulate it on G. In the latter case, we will
simply state that the algorithm is run on G'; simulating it on G is implied.
The algorithm models off of the one of [14] for approximating the treewidth in a parallel
model. However, new ideas are required to adapt the algorithm in a distributed model.
In
particular, as we will see, we need to adopt a graph contraction-based approach to support the
use of SA.
The algorithm is iterative and represents the original Ford-Fulkerson algorithm for maximum
flow. It maintains a set of disjoint s-t paths, and, on each iteration, either increases the number
of disjoint paths by one through an augmenting path step, or certifies that it is not possible
to reach k disjoint paths.
On the first iteration, the algorithm simply needs to find a single s-t path. This can be done
by Lemma 12. On a general iteration, for the rest of this section, we assume that the algorithm
knows r vertex-disjoint paths for 1 ≤ r < k, and needs to find r + 1 vertex-disjoint paths.
In the distributed setting, we assume that each node in a path knows its neighbor(s) on
the path. We now make each node on a path learn its index on the path; this can be
accomplished with Lemma 11 (on network G). In addition, we would like to assign each of
the r paths a unique path ID in [r], known to all nodes in the path. This can be accomplished
on network G as follows: first, each node sets xv ← ID(v) if v belongs in a path, or −∞
otherwise. Then, we compute the maximum xv inside G using SA. The (unique) node u
with xv = maxv xv notifies this event to all nodes in its path P using SA. We assign path P
the ID r, known to all its nodes. Then, all nodes in path P drop out of the future path ID
computations (i.e., they set xv ← −∞ from now on). There are r − 1 paths left; we iterate
with r ← r − 1 until we are done.
J. Li
23:7
3.1 The Residual Graph
Recall the setting: the algorithm knows r vertex-disjoint paths and tries to find r + 1 vertex-
disjoint paths. Our algorithm maintains a reachability graph, a directed graph with a source
s and sink t, such that there is an augmenting path in G iff t is reachable from s in this directed
graph.
Its construction is directly modeled off of the residual graph from the Ford-Fulkerson
algorithm and is similar to the "graph decomposition into bridges" in [14]. In general, directed
reachability is a hard problem in distributed models and even parallel models, but we will exploit
the special structure of the residual graph in the k-vertex disjoint paths problem in order to
compute s → t reachability efficiently.
Construction of the reachability graph.
The steps in our construction are illustrated
in Figure 1. First, we construct a directed graph G0
res which represents the residual graph in
the Ford-Fulkerson algorithm following the standard reduction from the vertex-disjoint paths
for each vertex v ∈ V − {s, t}, create
problem to the (directed) edge-disjoint paths problem:
two vertices vin and vout with a directed edge (arc) (vin, vout), and for each (undirected) edge
(u, v) ∈ E, add the two arcs (uout, vin) and (vout, uin). The following fact follows from standard
analysis of the Ford-Fulkerson algorithm.
(cid:73) Fact 15. There is an augmenting path in G iff there is a directed s → t path in G0
res.
We now modify the Ford-Fulkerson residual graph G0
res as follows: for each vertex v ∈ V not
on one of the r vertex-disjoint paths, we contract the vertices {vin, vout} into a single vertex v.
The resulting graph is our residual graph Gres.
(cid:73) Corollary 16. There is an augmenting path in G iff there is a directed s → t path in Gres.
Proof. Observe that this contraction does not create new, simple s → t paths, since given any
simple directed s → t path P in Gres, for each vertex v ∈ P not on one of the r vertex-disjoint
paths, replace the occurrence of v with vin, vout in that order; the resulting path is a directed
s → t path P in G0
res. Clearly, since we only contract vertices, we do not destroy any s → t
paths. Therefore, there is a directed s → t path in Gres iff there is one in G0
res, and the statement
(cid:74)
follows from the equivalence in Fact 15.
In fact, since there is always a t → s path in Gres, we can translate this statement in terms
of strong connectivity.
(cid:73) Corollary 17. There is an augmenting path in G iff s and t are strongly connected in Gres.
The notion of strongly connected components forms the basis of our distributed algorithm.
Throughout the algorithm, such as in the next step, we will modify the reachability graph in ways
such that s and t are strongly connected after the contraction iff they were strongly connected
before.
1
2 ∈ V (G
, v
, t
Let Gres denote the graph Gres with its arcs replaced by undirected edges (with parallel
2 as follows: the
edges removed). Observe that we can "embed" Gres as a subgraph of G
nodes vin, vout ∈ V (Gres) map to v
1
2), and the nodes s, t map to s
1. Therefore,
2. From now on, every time we say the distributed
we can simulate the network Gres using G
algorithm runs on the network Gres, we mean that it runs on G
For each connected component B in G −S
G −S
j V (Pj), we also have B ⊆ V (Gres); we call B a
bridge in Gres, following the terminology of [14]. Suppose we number the bridges B1, B2, . . ..
We claim that Bi is strongly connected in the Gres. Indeed, for each edge (u, v) in the graph
res, so the contracted vertices
j V (Pj), the two arcs (uout, vin), (vout, uin) are present in G0
2 with this embedding.
CVIT 2016
23:8
Distributed Treewidth Computation with Applications
s
t
s
t
vin
vout
vB2
β1
vin
vout
s
β2
t
s
t
vB1
Figure 1 Top left: the graph G with source s and sink t. The black vertices mark the one existing
path found so far. Top right: The residual graph G0
res from the Ford-Fulkerson algorithm. Bottom
left: The reachability graph GR vertices only. The two ellipses are the bridges. Bottom right: The
same reachability graph GR. The red arcs mark a directed s → t path, which corresponds to a valid
augmenting path.
u, v in Gres are connected by both arcs (u, v) and (v, u). Since Bi is connected by such bi-directed
arcs, it is strongly connected in Gres.
We now proceed to contruct a reachability graph GR. First, for each bridge Bi, contract it
into a single vertex βi, since Bi is strongly connected, this does not change whether or not s and
t are strongly connected. For each edge (u, v) in one of the r current paths where u is to the
left of v, we remove the arc (vout, uin). This does not affect the SCCs, since uin is still reachable
from vout along the path vout → vin → uout → uin. For each of the r existing vertex-disjoint
paths P in G, the set of vertices {vin, vout} in the reachability graph now form a directed path
from t to s; see Figure 1. We number these directed paths P1, . . . , Pr. For each such directed
path Pi and two vertices u, v ∈ V (Pi), we say that u is to the left of v if v can reach u on the
directed path Pj. Equivalently, we say that v is to the right of u. We also form a linear ordering
of V (Pj) ∪ {−∞, +∞}, where u < v iff u is to the left of v, and v > −∞ and v < +∞ for all
v ∈ V (Pj). Also, for each j ∈ [r], arrange Pj in rightward order as v1, . . . , vPj (so that v1 is
adjacent to s), and for each i ∈ [Pj], define π(vi) := i, the index of vi on the path Pj.
j Pj) ∪ {βi : bridge Bi}, constitutes our
The resulting directed graph, whose vertices are (S
that we can communicate within each bridge Bi using SA. Initially, every node inS
reachability graph GR.
For the distributed setting, the motivation for viewing the bridges Bi as single vertices βi is
j V (Pj)
broadcasts to its neighbors the fact that it belongs to some Pj. Then, for each bridge Bi,
each node in VBi knows that its neighbors in Bi are precisely the neighbors from which it
did not receive a broadcast. This knowledge is exactly what is needed for a single SA round.
Construction of the bridge graph.
J. Li
23:9
We next construct the bridge graph similarly to [14].
For a bridge Bi and a path Pj, let lj
i be the leftmost ingoing neighbor of βi on Pj in GR,
(βi) ∩ Pj) according
or −∞ if such a neighbor does not exist. In other words, lj
to the linear ordering of Pj, or −∞ if the set N(Bi) ∩ Pj is empty. Similarly, let rj
i be the
rightmost outgoing neighbor of Bi on Pj, or ∞ if such a neighbor does not exist. In other words,
i − 1;
i ≥ lj
(βi) ∩ Pj). Observe that, by the construction of GR, we always have rj
+
i = max(N
rj
GR
see Figure 1.
i = min(N−
GR
By Lemma 11, we can assume that each node v ∈ Pj knows its index on the path. Then,
in 2k SA rounds, every node in every bridge Bi can learn the values lj
i . First, every
node on Pj broadcasts its index to all its neighbors in Gres. Then, for each bridge Bi, every
node v ∈ Bi sets xl
v as the minimum index received from an in-neighbor of Pj in Gres, and
v as the maximum index received from an out-neighbor of Pj in Gres. Then, in two
sets xr
SA rounds, every node learns the minimum xl
v within its bridge, which
are the values li
v and maximum xr
i and rj
j and ri
j.
We now construct the bridge graph GB.
1. The vertices of GB is the set {βi : Bi is a bridge}.
2. For two bridges Bi, Bx and j ∈ [r], add an arc (βi, βx) to a set Dj if:
a. For some path Py, ly
x ≤ ry
i . Intuitively, this means that we can reach Bx from Bi in GR
by traveling leftward from ry
i to ly
x along path Py.
b. We have rj
x > rj
i . Intuitively, this means that we make "progress" along path Pj, in that
we can now reach a vertex in Pj further to the right.
c. There is no βz such that βz satisfies the above two conditions, and either rj
x, or
x and ID(Bz) > ID(Bx). In other words, ties are broken by ID (assume that each
z > rj
z = rj
rj
bridge has a unique ID).
For each Dj, we add all arc in Dj to GB.
We say that a node βi is s-reachable if there is a node in Bi with s as an in-neighbor.
Likewise, we say that a node βi is t-reachable if there is a node in Bi with t as an out-neighbor.
x and rj
We can compute the arc set Dj in a distributed fashion, such that for each βi, every
node in Bi knows the arc (βi, βx) ∈ Dj, if any. We assume that every bridge Bx has
x for all j ∈ [k]. First, each node v ∈ Bx broadcasts the value
computed the values lj
x), ID(Bx)) to each in-neighbor in Gres on a path Py. Then, every path Py computes
(π(rj
a prefix maximum of the values (π(rj
x), ID(Bx)) sent on the previous step. By maximum,
we mean lexicographic, so ties are broken by highest ID(Bx); if nothing is sent over a given
prefix, then the maximum is −∞ with an arbitrary ID. By Lemma 11, this maximum can
be computed in parallel for each Pj in O(log n) SA rounds.
At this point, for each y ∈ [k] and node v ∈ Py, we have computed the maximum rj
x ≤ v. We now have every node v ∈ Py broadcast this maximum
over all nodes βx with ly
x), ID(Bx)) to its in-neighbors in Gres. Finally, every bridge Bi computes the maximum
(π(rj
(π(rj
x), ID(Bx)) received by one of its nodes through a SA round. If the maximum value rj
satisfies rj
i , then every node in Bi now knows the arc (βi, βx) in Dj.
x > rj
x
x
We remark that our bridge graph contruction is slightly different from the one in [14], in order
to make it more amenable to distributed computing. We have a statement similar to Theorem 4.1
from [14], stating an equivalence between the residual graph and the bridge graph. By Fact 15,
this equivalence also extends to augmenting paths in G. Because the proof resembles the one
in [14], we defer it to Appendix B.
CVIT 2016
23:10
Distributed Treewidth Computation with Applications
(cid:73) Lemma 18. There is an augmenting path between s and t in G iff there exists s-reachable βi
and t-reachable βx and a directed βi → βx path in GB.
Like in [14], our next goal is to determine whether there is an s → t path in GB. Of course,
since directed reachability is a difficult problem in general, we need to exploit the special structure
of GB. [14] proceeds by iteratively shortcutting the graph, while we proceed using contraction.
This deviation from [14] is the main technical contribution of the paper.
From now on, we abuse notation, sometimes referring to Dj as the directed graph whose arcs
are precisely Dj. Observe that for each j ∈ [r], every vertex has out-degree at most 1 in Dj.
Also, the directed graph Dj is acyclic, since an arc βi → βx implies that rj
i . It follows that
Dj is composed of rooted trees, where the arcs point from away from the leaves towards the root.
We now show that every rooted tree in Dj is in fact strongly connected in GR.
(cid:73) Claim 19. Suppose βi and βx belong in the same rooted tree in Dj. Then, βi and βx are
strongly connected in Gres.
Proof. It suffices to prove the statement for pairs βi, βx where arc (βi, βx) exists in Dj. By
condition 2(a) in the construction of GB, there ia a path Py with ly
i . Therefore, from βi,
we move to ry
x, then
i − 1), and finally back to βi, showing strong connectivity. (cid:74)
along Pj to lj
With the bridge graph GB computed, our two remaining steps are: (i) determine if there
exists s-reachable βi and t-reachable βx and a directed βi → βx path in GB, and (ii) return an
augmenting path or a node separator of size r, depending on the outcome of (i).
i , then (left) along Py to ly
i (since rj
x, and then to βx. From βx, we can move to rj
x ≤ ry
For the rest of this section, we will shift our main focus from the classical setting to the
distributed setting. That is, we will explain our algorithm from a distributed point of view,
rather than commenting on distributed implementations in gray boxes.
x > rj
x > rj
i ≥ lj
3.2 Solving the Bridge Graph
Our distributed algorithm differs from the one in [14] by using a contraction-based approach,
rather than a shortcutting-based one. A high-level outline of our algorithm is as follows. First,
we contract every rooted tree of D1, or equivalently, every connected component in D1, the
underlying undirected graph of D1. We now recompute the bridge graph with the corresponding
bridges of each connected component merged into a single super-bridge in Gres. We repeat
this process for the remaining j ∈ [r]: contract every connected component in Dj, recompute
the bridge graph, and repeat. The lemma below states the desired property of the contraction
algorithm. The proof is deferred to Appendix B due to its length.
(cid:73) Lemma 20. At the end of the contraction algorithm, there exists s-reachable βi and t-reachable
βx that contract to the same vertex iff s and t are strongly connected in Gres.
First, for each j ∈ [r], in the distributed computation of the bridge graph GB, we can augment
the computation of the arc (βi, βx) in Dj (if any) so that for each Bi, not only do the nodes know
βx, but also (i) ID(Bx), (ii) the minimum2 value y ∈ [r] for which ly
i , and (iii) the value
π(ly
x); this simply requires broadcasting the auxiliary information on each step, and breaking ties
by value of y. Observe that since βi and βx belong in the same SCC in Gres, so do all nodes
v ∈ Py with ly
We now describe our algorithm for contracting the connected components in D1. Our next
1
C in Gres spanning the
goal is to find, for each connected component C in D1, a subgraph H
x ≤ v ≤ ry
i .
x ≤ ry
2 It is not important that this is minimum and not maximum or even arbitrary. However, we assume
minimum because uniqueness will make our analysis easier to present.
J. Li
23:11
bridges Bi whose nodes βi are in that component. Every node in the subgraph should know its
neighbors in the subgraph. Moreover, this subgraph should be a tree, a property that will be
useful later for recovery.
For each arc (βi, βx) in C, consider the corresponding pair (ly
i into H
i ) as mentioned above. We
x, ry
want to connect together all nodes v ∈ Py with ly
x ≤ v ≤ ry
1
C. In other words, each node
v ∈ Py should know whether there exists an (ly
x < v ≤ ry
i ; if so, the node
i ) pair such that ly
x, ry
v connects to its leftward adjacent node on Pj. We accomplish this task as follows: first, every
bridge Bi with an arc (βi, βx) in D1 and corresponding pair (ly
x) to node
i , which is necessarily adjacent to Bi in Gres. Then, for each y ∈ [r], the nodes on Py compute
ry
x) values. It is clear that for each node v ∈ Py, there exists
a suffix minimum of the sent π(ly
(ly
i iff the suffix minimum at v is strictly less than π(v). Therefore, v joins
x, ry
its leftward neighbor iff its suffix minimum is less than π(v).
i ) sends the value π(ly
x < v ≤ ry
i ) with ly
x, ry
Observe that if node v has suffix minimum exactly π(v), then node v is at the left endpoint
x is exactly π(ly
x = ly
x0.
i ) pair. We claim that the converse is true.
of some (ly
(cid:73) Claim 21. Suppose the arc (βi, βx) exists in D1, and consider the minimum y ∈ [r] such that
x ≤ ry
x). Moreover, there does not exist an arc
i . Then, the suffix minimum at ly
ly
(βi0 , βx0) in D1 such that x 6= x0 and ly
Proof. Suppose the first statement is false. Then, there is some other arc (βi0 , βx0) in Dj such
1
that ly
x and ID(Bx0) > ID(Bx). Then,
1
i , βx0 violates condition 2(c) for (βi, βx) in the definition
since ly
1
1
of the bridge graph, contradiction. The other case, assuming that either r
x = r
x0
and ID(Bx) > ID(Bx0) is symmetric: we have βx violating condition 2(c) for (βi0 , βx0). For the
second statement, if such an arc (βi0 , βx0) exists, then without loss of generality, assume that
x and ID(Bx0) > ID(Bx) (we swap x and x0 otherwise). Then, since
1
either r
x0 ≤ ly
(cid:74)
ly
1
x0 > r
x, we can apply the proof of the first statement to get a contradiction.
i0. Assume that either r
i and r
x0 ≤ ly
x0 ≤ ly
x ≤ ry
x ≤ ry
1
x0, or r
x0 ≥ r
1
1
x, or r
1
x, or r
1
x0 > r
1
x0 = r
1
x0 = r
1
x > r
1
x > r
x, ry
This finishes the edges of H
1
C within each Pj. Of course, we can repeat the above in parallel
for each connected component C in D1. We now augment the above computation so that when
computing suffix minimum, the value ID(Bx) is recovered along with the minimum π(ly
x). Then,
if a node v ∈ Py has suffix minimum exactly π(v), then v connects to its neighbor(s) in Bx in
1
C, where C is the component containing βx. Then, for each arc (βi, βx) in D1 and minimum y
H
1
for which ly
C, where C is the component containing βi.
1
Finally, within each bridge Bi, add a spanning tree of Bi into the corresponding subgraph H
C
using Lemma 9. This concludes our construction of the subgraph for each connected component
in D1.
i , the bridge Bi connects to ry
x ≤ ry
i in H
We now prove some properties of the subgraphs H
1
C. The tree property below does not help
us in the contraction phase, but it will help in recovering paths in the recovery phase.
(cid:73) Claim 22. The subgraphs H
1. For every connected component C in D1, there is a subgraph spanning (precisely) the bridges
1
C satisfy the following properties:
in C.
2. Any two subgraphs are disjoint.
3. Every subgraph is a tree.
i → ly
x on Py and the path ry
Proof. Claim 21 implies the following statement: for any two arcs (βi, βx), (βi0 , βx0) in D1, either
x0 on Py are disjoint, or x = x0. If x = x0, then the arcs
the path ry
(βi, βx), (βi0 , βx0) belong to the same connected component in D1. Therefore, if (βi, βx), (βi0 , βx0)
i → ly
belong to different connected components, then the segments ry
x0 do not
1
intersect. It is easy to see by the algorithm that each constructed H
C is disjoint. This proves (1)
and (2).
i0 → ly
x and ry
i0 → ly
CVIT 2016
23:12
Distributed Treewidth Computation with Applications
1
r
1
1
r
2
1
r
3
1
r
4
1
r
5
B1/β1
D1
D1
B4/β4
D1
B2/β2
B5/β5
7
2
l
7
r
1
7
l
5
7
r
4
D1
B3/β3
9
4
l
1
r
1
1
r
2
9
r
2
9
r
3
1
r
3
1
r
4
B1/β1
B1/β1
B4/β4
B4/β4
B2/β2
B2/β2
1
r
5
B5/β5
B5/β5
7
2
l
7
r
1
7
l
5
7
r
4
B3/β3
B3/β3
9
4
l
9
r
2
9
r
3
P1
P7
P9
P1
P7
P9
Figure 2 Top: the bridges Bi with their relevant edges li
red edges mark the arcs in D1 connecting the βi. Bottom: the subgraph H1
component {β1, β2, β3, β4, β5} in D1.
j, rj
i
in green and blue. The dotted
C for the connected
For property (3), fix a subgraph H
1
C is acyclic. Suppose we take H
1
C by construction. Therefore, the subgraph H
C. For each βi ∈ C, the corresponding bridge Bi induces
1
1
C is acyclic iff the graph obtained by
a tree in H
1
contracting each bridge Bi in H
C and contract each Bi into a
vertex βi, so that the new graph, called H, is now a subgraph of GR. To prove that H is acyclic,
we start with D1 and transform it into H while preserving the acyclicity of D1. For each βx ∈ C
1
with positive in-degree in D1, let βx1 , . . . , βx' be the in-neighbors of βi. In H
C, the algorithm
1
C; in addition, for each βxi, the
adds the union of the (now undirected) paths ry
xi
algorithm adds an edge connecting Bxi to ry
. The set of edges added is a tree connecting the
vertices βx, βx1 , . . . , βx' of H; call this tree Tx. For each βx in D1, we delete the edges (βx, βxi)
in D1 and add T (as well as any extra vertices). Since we always delete a tree and add back a
1
C, proving
tree, the graph remains acyclic. At the end, we have added exactly the edges in H
(cid:74)
that H
1
C is connected by property (1), so it is a tree.
1
C is acyclic. Lastly, H
x to H
→ ly
xi
1
Since the subgraphs H
1
C are disjoint, they serve as the nodes after contracting every connected
component of D1 in the classical algorithm, and we can communicate within each contracted
component in one SA round. We would like to continue this algorithm for j = 2, . . . , r, always
maintaining subgraphs H j
C that are trees spanning the super-bridges in the connected components
of Dj, but we run into the following obstacle: when constructing the subgraph on the next
1
C inside the paths Pj.
iteration, we might reuse edges in E(H
Reusing these edges may destroy the tree property of H j
C.
We fix this problem as follows. Our goal is to construct a graph G
2r
, . . . , vr,
(the same way Gres embeds into G
connected together in a clique, with each node vj sharing the same neighbors as v. Let us call
+ embeds into (Gres)r and Gres embeds into G, which means
this network G
that G
C) ∩ E(S
2). Replace each node in v ∈S
j Pj), i.e., the edges in H
+ that "embeds" into G
j Pj with r copies v
+ embeds into G
We run the algorithm on G
+ instead, which we then simulate on G. For the first iteration
j = 1, every node v
i Pi. This way, in future iteration
j = 2, . . . , r, we always have a fresh set of nodes, namely the nodes vj, to use in iteration j. For
1 now takes the role of node v ∈ S
+. Observe that G
2r.
1
J. Li
23:13
takes the role of node v ∈S
iteration j = 2 . . . , r, we repeat the algorithm for j = 1 with three main differences: (i) before the
iteration, every super-bridge is now the nodes in some H j−1
C , (ii) we do not add a spanning tree
inside each bridge, since we already have one from iteration j − 1, and (iii) every node vj now
i Pi. Since our algorithm emulates the classical algorithm, we know
by Lemma 20 that after all r iterations, there is a directed s → t path iff there is an s-reachable
bridge and a t-reachable bridge in the same subgraph H r
C. Our next goal is to, depending on
this outcome, either find an augmenting path in Gres or find an s-t node cut of size at most r.
3.3 Finding an Augmenting Path
C. If there exists a subgraph H r
First of all, it is easy to test if there exists a subgraph H r
C with both an s-reachable and t-reachable
bridge: in two SA rounds, the s-reachable bridges and the t-reachable bridges broadcast in their
subgraphs H r
C with both s-reachable and t-reachable bridges,
then we show how find an augmenting path in Gres. Note that this task is not trivial, since while
there exists a (unique) path from the s-reachable and t-reachable bridges in H r
C, this path may
go rightward along a path Pj, which is not allowed.
1
C, there already exist an s-reachable bridge Bs
and a t-reachable bridge Bt. Let βx be the "root" of the component (tree) C in D1, i.e., the
1
C only goes left along
unique βx with no out-arc in D1. The unique path from Bs to Bx in H
the paths Pj, since every time we travel along an arc in D1 from Bs to Bx, we traverse leftward
along one path Pj from one bridge to another. Moreover, we can compute this path in O(log n)
SA rounds using Lemma 12. Next, we "trim" the path by removing all edges completely inside
Bs or Bx.
For illustration, suppose first that in some H
1
x > r
If Bx = Bt, then we skip the next step; otherwise, since there is a path from Bt to Bx in
t ≥ l
1
1
1
t . Therefore, we can extend
t . Combining this with r
D1, we have r
1
1
this path from Bx to travel left from r
t , and then enter Bt. Again, by using Lemma 12, we
x to l
can establish this path in O(log n) SA rounds, and then trim it by removing all edges completely
inside Bx or Bt. Finally, we connect the ends of the Bs-Bx and Bx-Bt paths inside Bx, which
can be done in O(log n) SA rounds.
t − 1 gives r
1
x ≥ l
1
Lastly, it remains to connect node s to the Bs-end of the path, and to connect the Bt-end of
the path to t. Again, these take O(log n) SA rounds.
C , H r−1
C , . . . , H
C
C
C
In general, we process the graphs in reverse order H r
1
C. For a graph H j
C case above, treating the connected components of H j−1
1
first repeat the algorithm for the H
the contracted bridges. Observe that all edges outside any component H j−1
along the paths Pj. We now erase all edges inside each traversed component H j−1
the two broken endpoints in H j−1
+ that only travels left along paths Pj.
P
The last issue is that unlike the H
C, we
as
C must travel leftward
and connect
on the next iteration. At the end, we have constructed a path
are r copies of each node inS
Then, for each node v ∈S
+ may not be simple. Indeed, since there
j Pj, a single arc in some Pj can be traversed left up to r times,
once in each copy. We can fix this issue with the following "shortcutting" step: first, number the
+ using Lemma 11, and suppose every other node gets value xv := 0.
nodes on P
j Pj, every node {v
, . . . , vr} updates xvi ← maxh xvh, which can be
, . . . , vr are connected by a clique. Then, every node v ∈ P
+
done in one round since the nodes v
computes a prefix maximum of the xu values in the path. For each v ∈ P
+, if this maximum is
+ can be pruned before xv is reached; see Figure 3. Therefore,
greater than xv, then the path P
node v ∈ P
+ back into G, giving
our desired simple augmenting path in Gres. Lastly, translating the augmenting path in Gres to
r + 1 vertex-disjoint paths in G can be done as in the Ford-Fulkerson algorithm.
+. Finally, we collapse the graph G
+ drops out of the path P
1
C case, this path P
1
1
+ from 1 to P
CVIT 2016
23:14
Distributed Treewidth Computation with Applications
7
6
15
5
14
4
13
3
12
15
14
13
Figure 3 Left: the path P +. Right: collapsing the path P +; the numbers are the updated xv
values.
C with both an s-reachable and t-reachable bridge, then there is no
3.4 Finding a Node Cut
If there is no subgraph H r
augmenting path, so the algorithm needs to find an s-t node cut S of size at most k.
C containing an s-reachable bridge. For
each bridge Bi to learn whether or not Bi ∈ Bs, we have every node inside a bridge that is
C; the bridges Bi ∈ Bs are precisely the ones that
adjacent to s broadcast to its subgraph H r
receive such a broadcast.
Let Bs denote all bridges Bi inside some subgraph H r
construction is presented in the proof of Theorem 4.1 in [14]. For path Pj, let w0
j over all bridges Bi ∈ B. Necessarily, w0
ri
corresponding to Pj; let wj be this node v. If this node does not exist, i.e., ri
bridges Bi ∈ B, then let wj be the first node on Pj (the one adjacent to s).
We now present a set of k nodes, one from each Pk, that form an s-t node cut; a similar
j be the rightmost
j = vin for some v on the vertex-disjoint path in Gres
j = −∞ for all
We now show how to compute the nodes wj for each j ∈ [r]. First, every bridge Bi ∈ Bs
i ; then, every node in v ∈ Pj sets xv to be its index on Pj if it is notified, and
notifies node rj
1 otherwise. The nodes in Pj then compute aggregate maximum of the values xv. Finally, the
node v ∈ Pj whose index is exactly its value xv becomes wj.
(cid:73) Lemma 23. The set {w1, . . . , wr} is a node cut of G.
Proof. First, we would like to extend Lemma 20 to the following statement: at the end of the
contraction algorithm, an s-reachable βi and a (not necessarily t-reachable) vertex βx contract
to the same vertex iff s and Bi are strongly connected in Gres. To do so, imagine changing the
graph Gres as follows: remove all arcs from any bridge to t and add a single arc from a vertex
in Bx to t. With this modification, the graphs GR and GB do not change, but now, only βx is
t-reachable; applying Lemma 20 proves the statement. From now on, we forget this modification,
i.e., we stick with the original Gres.
Observe that the distributed algorithm follows the contraction algorithm of Lemma 20; in
C contains the bridges Bi whose vertices βi get contracted to a single
particular, every subgraph H r
vertex in the contraction algorithm. Therefore, the bridges Bi ∈ B are precisely the βi that get
contracted to the same vertex as some s-reachable bridge. By the statement at the beginning of
this proof, these bridges Bi are precisely those strongly connected to s.
The rest of our proof resembles the last paragraph of the proof of Theorem 4.1 in [14].
Suppose for contradiction that there is a simple s-t path P in G − {w1, . . . , wr}. Consider the
j Pj. At some point, we must have two vertices
j Pj adjacent on this subsequence such that for the paths Pj, Pj0 containing v and v0
respectively, we have v < wj and v0 > wj0. The vertices v, v0 cannot be adjacent in P, so there
must be vertices inside a single bridge Bi in between the occurrences of v and v0 on P. This
i0 for some Bi0 ∈ B, vertex wj is reachable
bridge Bi satisfies lj
subsequence of vertices in P that are also inS
v, v0 ∈S
i > wj0. Since wj = rj
i < wj and rj0
J. Li
23:15
from s. Therefore, vertex lj
strongly connected in Gres. In particular, Bi ∈ B, so wj0 ≥ rj0
the assumption that rj0
i is also reachable from s, and so is Bi, which means that s and Bi are
i by definition of wj0, contradicting
(cid:74)
i > wj0.
This finishes the k-vertex disjoint paths algorithm and Lemma 2.
3.5 Running on Multiple Subgraphs
We have proved our main result of this section, Lemma 2, restated below for reference. Below,
we state some modifications of this result that are more directly useful in the next section.
(cid:73) Lemma 2. Given a graph G = (V, E) of treewidth at most k and two vertices s, t ⊆ V , we
can either find k vertex-disjoint s-t paths, or output an s-t node cut of size less than k, in
O(kO(1)
D) rounds. In the former case, every node knows whether it is on a path, and if so,
its predecessor and successor on that path. In the latter case, every node knows the fact that k
vertex-disjoint paths do not exist, as well as whether it is in the node cut.
First, we obtain a generalization where we want to find vertex-disjoint paths between two sets
of nodes, not just s, t, within a connected subgraph of G, not G itself. Moreover, this formulation
includes forbidden nodes, those which cannot appear in any vertex-disjoint path.
(cid:73) Corollary 24. Given a graph G = (V, E) of treewidth k, a set U ⊆ V such that G[U] is
connected, and three disjoint vertex sets A, B ⊆ V − U and X ⊆ U, we can either find k vertex-
disjoint A-B paths whose internal nodes belong in G[U] − X, or conclude that k vertex-disjoint
paths do not exist, in O(kO(1)
D) rounds. In the positive case, every vertex knows whether it is
on a path, and if so, its predecessor and successor on that path. In the negative case, every vertex
knows the fact that k vertex-disjoint paths do not exist.
Proof. The subgraph is not an issue, because any simulated network G[U]' in the algorithm is
a subgraph of G', so we can simulate network G[U]' on G' first, and then on G.
To address the A-B paths modification, imagine adding a node s whose neighbors are
precisely A, and a node t whose neighbors are precisely B. The virtual nodes s, t do not exist in
the network, but observe that through the algorithm of Lemma 2, the only times when nodes s
and t are active are (i) when we compute prefix/suffix aggregates on at most k paths, and (ii)
when we compute a path with s and/or t as an endpoint, in an attempt to find an augmenting
path. In case (i), for each path Py, the node v ∈ Py with π(v) = 2 (i.e., the node to the immediate
right of s) can take the role of s in prefix computations; likewise, the node with π(v) = Py − 1
can take the role of t in suffix computations.3 In case (ii), the nodes in A and B, which are in
the network, can take the role of nodes s and t. For example, if we have an s-reachable bridge
and we want a path from s to a specific node v in the bridge Bi, then we instead compute a path
from a node in N(s) ∩ Bi to v.
Lastly, the forbidden node set X is also not a problem: when computing the bridges Bi, these
(cid:74)
nodes purposefully do not join any bridge.
The next generalization really emphasizes the power of the shortcuts framework: suppose,
instead, that we want to solve k vertex-disjoint paths on a subgraph H ⊆ G. Actually, we want
to solve multiple instances of the problem on vertex-disjoint subgraphs H1, . . . , H'. Then, we
can solve them all simultaneously in O(kO(1)
D) rounds!
3 The corner case Py = 2 can be ignored, since we can greedily choose the single edge from s to t as a
path. Likewise, if any nodes in A and B are adjacent, we can greedily choose them as vertex-disjoint
paths.
CVIT 2016
23:16
Distributed Treewidth Computation with Applications
(cid:73) Corollary 25. Given multiple instances (Ui, Ai, Bi, Xi) in Corollary 24 such that the node
sets Ui are disjoint, we can, simultaneously for each (Ui, Ai, Bi, Xi), either find k vertex-disjoint
Ai-Bi paths whose internal nodes belong in G[Ui]− Xi, or conclude that k vertex-disjoint paths
do not exist, in O(kO(1)
Proof. For each instance (Ui, Ai, Bi, Xi), every step of the algorithm either runs on network
G[Ui], or on G[Ui]'i for some 'i ≤ k. Since the networks G[Ui]'i are disjoint subgraphs of Gk for
different Ui, we simulate every network G[Ui] or G[Ui]'i on Gk. Therefore, on each step, the SA
tasks of the different instances can be simultaneously solved on Gk in O(kO(1)
D) rounds, which
is then simulated on G in O(kO(1)
(cid:74)
D) total rounds.
D) rounds.
Algorithm Outside Disjoint Paths
4
In this section, we provide the rest of the algorithm for approximating treewidth, which uses the
k-vertex disjoint paths problem as a subroutine. It is a combination of the efficient sequential
and parallel algorithms in [16, 15]. It will make repeated calls to the algorithm of Corollary 25,
the corollary of Lemma 2 in the previous section.
The treewidth approximation algorithm, which is recursive, uses the concept of graph separ-
ators, defined below.
(cid:73) Definition 26 (Separation). Let G = (V, E) be a graph and let A, B, S ⊆ V . We say that S
separates G into A and B if A, B, S partition V and N(A) ⊆ S and N(B) ⊆ S. Note that
A and B are not necessarily connected, which means that there could be multiple choices for A
and B. In addition, for disjoint X, Y ⊆ S, we say that S is an X-Y separator if there exist
A, B ⊆ V with X ⊆ A, Y ⊆ B such that S separates G into A and B.
(cid:73) Definition 27 (Balanced Separation). Let G = (V, E) be a graph and let X ⊆ V . A set S ⊆ V
is an (X, α)-balanced separator of G if there exist A, B ⊆ V such that S separates G into A
and B and A ∩ X,B ∩ X ≤ αX. When X = V , we omit the X, using the term α-balanced
separator instead.
The following well-known fact states that if G has bounded treewidth, then it admits constant-
sized balanced separators for any X ⊆ V .
(cid:73) Lemma 28 (Lemma 7.20 of [4]). Let G = (V, E) be a graph of treewidth k. For any set X ⊆ V ,
there exists an (X, 2/3)-balanced separator of G of size k + 1.
The algorithm is recursive, always running on an instance (U, X) with U ⊆ V , X ⊆ N[U],
and X ≤ 7k + 4. It starts with (U, X) = (V,∅). The algorithm has two cases, depending on
whether the current recursion depth is odd or even (the initial instance (V,∅) has recursion depth
0). The even case finds a balanced separator S of size ≤ k + 1 in the graph G[U],4 separating
the current graph into components whose sizes are a constant fraction smaller, and for each
component U0 of vertices, recursively calls (U0, N[U0] ∩ (X ∪ S)). This guarantees the algorithm
O(log n) recursion depth, but comes at a cost: the size of X increases by ≤ k + 1 upon the next
recursion call (to the odd case). This increase is remedied in the odd case below, which on input
(U, X) finds a set S that separates G[U] into components, each with ≤ 2
3X vertices in X, and
for each component U0 of vertices, recursively calls (U0, N[U0] ∩ (X ∪ S)). If X ≥ 6k + 4, then
3(6k+4)+(k+1) = X− k− 1
(U0, N[U0]∩(X ∪ S)) ≤ 2
3 ,
3X+S = X− 1
3X+S ≤ X− 1
4 Of course, if no such balanced separator S is found, the algorithm can immediately exit and conclude
that tw(G) > k.
J. Li
23:17
so the size of X is reduced by k + 1. Thus, by alternating the recursion between odd and even
depth, we can maintain the invariant X ≤ 6k + 4 while reducing the size of U by a constant on
every two iterations.
The distributed implementation runs through the recursion tree in parallel. Namely, it
proceeds in T super-rounds, where T is the maximum recursion depth of the recursive
algorithm. On super-round t ∈ [T], the distributed algorithm processes all instances (U, X)
at recursion depth t. Here, we will use the crucial property that the sets U in this recursion
layer are connected and pairwise disjoint.
4.1 Odd Recursion Depth
In this case, our goal is to reduce the size of X sufficiently. We know by Lemma 28, there exists
an (X, 2/3)-balanced separator in G. Suppose S is this separator, which separates G into A and
B such that A ∩ X,B ∩ X ≤ (2/3)X. If we let Y := A ∩ X and Z := B ∩ X, then this
means that Y ,Z ≤ 2
3X and the set S contains X − (Y ∪ Z) and is an Y -Z separator. The
algorithm proceeds by trying all possible values of Y and Z and finding such a set S.
The algorithm for odd recursion depth proceeds as follows. For all partitions X0, Y, Z of X
with Y ,Z ≤ 2
3X, try to find an Y -Z separator in the graph G[U]−X0 of size ≤ (k+1)−X0,
which is an instance of (k+1−X0)-Vertex Disjoint Paths. If no such separator is found over
all Y, Z, terminate the algorithm and output the conclusion that tw(G) > k. Else, for partition
X0, Y, Z and Y -Z separator S0, let S := X0 ∪ S0, and for each component U0 of vertices in
G[U] − S, recursively call (U0, N[U0] ∩ (X ∪ S)).
If the graph G[U] − S is empty, then this
recursion branch terminates.
Like the above algorithm, the distributed algorithm iterates over all O(3k) partitions X0, Y, Z.
We can elect a leader in U to decide which partition X0, Y, Z to try next, and broadcast it to
the other nodes in U using SA. For (k + 1 − X0)-Vertex Disjoint Paths, one difficulty
is that we want our computation to depend on X (cid:40) U and yet run only on nodes in U, since
only the sets U are disjoint over instances, not X. This is the reason for the specifications
of Corollary 25. In fact, we simply solve (k + 1 − X0)-Vertex Disjoint Paths using
Corollary 25 with parameters (A, B, X) := (Y, Z, X0), taking time O(kO(1)
D).
4.2 Even Recursion Depth
In this case, our goal is to separate the current graph into components a constant factor smaller,
in order to bound the recursion depth by O(log n). We first introduce the concept of splitters
from [15].
(cid:73) Definition 29 (B-splitter). For a rooted, spanning tree T ⊆ G[U], a set of vertices R ⊆ U is a
B-splitter if R ≤ n/B and every connected component in T − R has less than B vertices.
For a vertex v ∈ U, define desc(v) and children(v) as the children and descendants of v,
respectively. Define sub_size(v) as the number of vertices in the subtree rooted at v. Let R be
all the vertices v ∈ U satisfying the following condition:
1 + X
u∈children(v)
(sub_size(u) mod B) > k.
(cid:73) Theorem 30 (Theorem 9.2 in [15]). The set R defined above is an R-splitter.
CVIT 2016
23:18
Distributed Treewidth Computation with Applications
The distributed algorithm first computes an arbitrary spanning tree T of G[U], e.g., by
computing an MST with arbitrary weights following [7], and roots it at an arbitrary vertex.
At this point, every node knows its parent and children in the rooted tree. To compute
the set R, each node v ∈ U first computes the size of its subtree in T; this can be done in
O(log n) SA rounds using tree aggregation techniques in [12]. Then, each node broadcasts
sub_size(v) to its parent node in a single round, so that each node can determine whether
it joins R.
The rest of this even recursion section is based on [16]. For each vertex r ∈ R, we define
wr := sub_size(r) − X
and for R0 ⊆ R, wR0 := P
P
r∈R0 wr. In other words, wr is the number of vertices v for which
r is the first vertex in R encountered on the path from v to the root. Also, observe that
r∈R wr = U. The following lemma states that if a separator X does not intersect R, then the
values wr approximately determine the size of a separated component.
(cid:73) Lemma 31 (Lemma 2 in [16]). Consider a set X ⊆ U with X ≤ k and X ∩ R = ∅ which
separates G into A and B. Then,
sub_size(r0),
r0∈desc(r)∩R
A − w(A ∩ R) ≤ kB,
and the same holds for B.
We compute an R-splitter with B := n/(12k), so that R ≤ 12k. By Lemma 28, there exists
an (X, 2/3)-balanced separator of G into A and B. If X ∩ R = ∅, then by Lemma 31,
w(A ∩ R) ≤ A + kB ≤ (2/3)U + (1/12)U = (9/12)U,
and the same holds for B. Therefore, letting Y := A ∩ R and Z := B ∩ R, we conclude that
there exists a partition Y, Z of R with w(Y ), w(Z) ≤ (9/12)U that admits a Y -Z separator in
G[U] − R of size ≤ k + 1. The algorithm tries all possible such Y, Z and tries to find a separator
for each. If a separator S is found, then by Lemma 31,
A ≤ w(A ∩ R) + kB ≤ (9/12)U + (1/12)U = (10/12)U,
so S is an (X, 10/12)-balanced separator, giving us the necessary constant factor decrease.
Otherwise, if no S is found, we must have X ∩ R 6= ∅. In this case, we brute-force over which
one of the R ≤ 12k vertices belongs in X. If we guess r ∈ R, then we would like to solve the
instance (U − r, X), except we look for a separator of size k instead of k + 1.
It is possible for U − r to be disconnected, in which case only the largest component still
needs to be separated, since the other components have size ≤ (1/2)U. Therefore, if U0 is the
largest component of G[U] − r, then we solve the instance (U0, N[U0] ∩ X) with k decreased by
1. This is a recursion that is completely contained inside the (U, X) instance; it has nothing to
do with the main recursion, so it does not distinguish between even and odd recursion levels.
Overall, an instance with value k results in ≤ 12k recursive instances of value k − 1. A
straightforward induction shows that this recursion tree has size ≤ kO(k).
The values wr can be computed in a distributed setting as follows: every node r ∈ R sets
xr := wr and every other node v ∈ V − R sets xv := 0. Then, every node computes
u∈desc(v) xu in O(log n) shortcut rounds using tree aggregation techniques in [12]. Since
every node r ∈ R already knows sub_size(r), it can locally compute wr = sub_size(r) −
P
P
u∈desc(r) xu.
The recursion within each (U, X) instance is done sequentially, which results in kO(k) se-
J. Li
23:19
quential calls to s-Vertex Disjoint Paths for s ≤ k + 1, taking total time O(kO(k)).
5
Applications
This section proves Theorem 3, restated below.
(cid:73) Theorem 3. Let G be a graph network with treewidth k. The problems maximum independent
set, minimum vertex cover, chromatic number, and minimum dominating set can be solved in
O(kO(k)
D) rounds on network G.
For conciseness, we only provide a distributed algorithm for maximum independent set; the
algorithms for the other problems are straightforward modifications.
We first introduce our notation for treewidth decompositions. A treewidth decomposition of
a graph G is a tree T whose vertices, called bags, are subsets of V (G). The tree T satisfies three
properties: (i) the union of vertices over all bags equals V (G); (ii) for each v ∈ V , the set of bags
containing v is connected in T ; (iii) for each edge (u, v) ∈ E(G), there is a bag containing both u
and v. The treewidth of a graph G is the minimum k such that there exists a tree decomposition
T of G whose bag sizes are at most k + 1.
5.1 The Classical Algorithm
Let us now sketch the traditional algorithm for maximum independent set on bounded treewidth
graphs. For an input graph G of treewidth k, the algorithm first computes a treewidth decom-
position of the graph with bag sizes bounded by O(k). Then, the algorithm applies dynamic
programming on this treewidth decomposition; we sketch this dynamic program below. This
presentation of the dynamic programming algorithm is not the most standard or the most effi-
cient, but it will translate more smoothly when we adapt it to the distributed setting.
Root the tree T at a root vertex r ∈ V (T ). For each vertex
The dynamic program.
v ∈ V (T ), let its bag be Bv. For each bag Bv and subset I ⊆ Bv, we will define dynamic
programming states Join(Bv, I). For each vertex v ∈ V (T ) − {r} and its parent p ∈ V (T ) in
the rooted tree T , for subsets Iv ⊆ Bv and Ip ⊆ Bp, we will define dynamic programming states
Extend(Bv, Iv, Bp, Ip). These are defined as follows:
1. For each vertex v ∈ V (T ) that is a leaf in the rooted tree T and each Iv ⊆ Bv,
Join(Bv, Iv) := Iv if Iv is an independent set in G[Bv],
− ∞ otherwise
(1)
Our goal is for the optimal size of the independent set to be the best value of Join(Br, Ir),
i.e., maxBr⊆Iv Join(Br, Ir). Clearly, if r is a leaf (i.e., the tree T is a single vertex), then this
is true. Otherwise, we will define Join for non-leaf vertices later.
2. For each vertex v ∈ V (T ) − {r} and its parent p, and each Iv ⊆ Bv, Ip ⊆ Bp,
Extend(Bv, Iv, Bp, Ip) := Join(Bv, Iv) + Ip − Iv if Iv ∩ (Bv ∩ Bp) = Ip ∩ (Bv ∩ Bp)
(2)
and Iv, Ip are independent sets in G[Bv], G[Bp],
− ∞ otherwise
In other words, we try to extend the state Iv in Bp to the state Ip in Bp, but this is only
valid if the sets Iv, Ip agree on the vertices shared by Bv and Bp, namely Bv ∩ Bp.
CVIT 2016
23:20
Distributed Treewidth Computation with Applications
3. For each non-leaf vertex p, let children(p) denote the children of p in T . For each Ip ⊆ Bp,
Join(Bp, Ip) := Ip + X
(Extend(Bv, Iv, Bp, Ip) − Ip)
max
Iv⊆Bv
v∈children(p)
if Ip is an independent set in G[Bp],
− ∞ otherwise
(3)
It is a routine exercise in algorithm design on bounded treewidth graphs to argue that this
algorithm is correct. The main observation is that in the dynamic program, once Join(Bp, Ip)
"forgets" the vertices in Bv − Bp for some v ∈ children(p), these vertices never appear again on
any Join(Bp0 , Ip0) for any p0 on the path from p to the root r, due to property (ii) of the treewidth
decomposition.
Recovering the solution. Thus, maxIr⊆Br Join(Br, Ir) =: OP T is the size of the maximum
independent set in G. To compute the actual maximum independent set, we follow the traditional
procedure of "reversing" the dynamic program, starting from the root r:
r , pick an arbitrary set Ir ⊆ Br that satisfies Join(Br, Ir) = OP T, i.e.,
1. For I∗
I∗
r := arg max
Ir⊆Br
Join(Br, Ir).
(4)
2. For each non-leaf vertex p and each v ∈ children(p), define
(cid:0)Extend(Bv, Iv, B∗
p(cid:1) .
I∗
v := arg max
Iv⊆Bv
3. At the end, the returned maximum independent set isS
p) − I∗
p , I∗
(5)
v∈V (T ) I∗
v.
5.2 Distributed Implementation
Suppose the input (and network) graph G has treewidth t. We first run the treewidth algorithm
of Theorem 1, computing a treewidth decomposition with maximum bag size O(k). However,
one immediate issue is that the nodes in each bag
The first attempt is to adapt the treewidth algorithm of Theorem 1 into computing an actual
treewidth decomposition. However, one caveat is that the nodes in each bagof the treewidth
decomposition may not be connected in the network G. Therefore, when performing the stand-
ard dynamic programming over a treewidth decomposition, the nodes in a bag cannot directly
communicate with each other.
We resolve this issue by exploiting the special structure of the treewidth algorithm of Section 4.
Recall that the algorithm is recursive: on each recursive instance (U, X), it either terminates
prematurely, concluding that tw(G) > k, or finds a node set S of size ≤ k + 1 and, for each
connected component U0 in G[U] − S, recursively calls (U0, N[U0] ∩ (X ∪ S)). In Section 4, we
observed several properties of the algorithm that will be helpful, listed below.
1. For each layer of the recursion depth, the sets U in the instances (U, X) are connected and
node-disjoint.
2. For each instance (U, X) in the algorithm, X ⊆ N[U].
3. If instance (U, X) computes separator S ⊆ U, then every recursive instance (U0, X0) called
by this instance has N[U0] ∩ S 6= ∅.
4. The tree in the treewidth decomposition has depth O(log n).
5. Every leaf instance (U, X) in the recursion tree has U ≤ k +1, since the computed separator
S satisfies S ≤ k + 1, and G[U] − S must be empty to end this recursion branch.
We now specify a tree decomposition T "implicitly" produced by the algorithm of Section 4.
For each instance (U, X) that produces separator S, create a bag BU,X in T with vertices X ∪ S.
J. Li
23:21
The children of this bag are the bags produced by all recursive instances (U0, X0) called by (U, X).
It is clear by the algorithm of Section 4 that T is a treewidth decomposition with maximum bag
size O(k). We would like to apply dynamic programming on this treewidth decomposition T in
a distributed fashion.
Modifications. We first augment the algorithm of Theorem 1, so that for each layer of the
recursion, for each (U, X) in that layer, all nodes v ∈ V know: (i) the depth of that layer, (ii) the
ID of U, set to be the smallest ID of a node in U, and (iii) the set X of size O(k). Moreover, if S
is the separator computed at instance (U, X), then for each instance (U0, X0) called recursively
at this instance, exactly one of the nodes in N[U0]∩ S knows the ID of U0 and its neighbors in U0;
we say that node v is in charge of U0. It is clear that all of this can be done in kO(1)polylog(n)
SA rounds.
The dynamic program. With this, we implement the dynamic program of Section 5.1
"bottom-up", from layer T := Θ(log n) to (the single instance in) layer 0 in that order. For
layer t ∈ [T], for each instance (U, X) in that layer, perform the following:
1. If no nodes in U have received anything so far, then instance (U, X) is a leaf in the recursion
tree.
In this case, the computed separator S in instance (U, X) is exactly U, so the bag
corresponding to (U, X) has node set X ∪ S = X ∪ U. The nodes in U first learn each other's
IDs as well as the IDs in X; this can be achieved because (i) U ≤ k + 1, (ii) U is connected,
(iii) X ⊆ N[U], and (iv) X = O(t). The bag BU,X for this instance is a leaf in T and has
node set X ∪ U, so for each IU,X ⊆ BU,X, we compute Join(BU,X , IU,X) according to (1).
This is 2O(k) values, one for each IU,X ⊆ BU,X; we then broadcast all of the Join(BU,X , IU,X)
pairs to N[U]. In total, this takes 2O(k) distributed rounds.
2. Otherwise, (U, X) must have called some recursive instances after computing its separator
S. Any such instance (U0, X0) satisfies N[U0] ∩ S 6= ∅, and instance (U0, X0) has already
broadcasted all Join(BU0,X0 , IU0,X0) pairs to the unique node in U in charge of U0. For each
node v ∈ U and recursive instance (U0, X0) such that v is in charge of U0, node v locally
computes, according to (2),
Extend(BU0,X0 , IU0,X0 , BU,X , IU,X) for each X ⊆ BU,X , X0 ⊆ BU0,X0 .
Note that a node can receive from multiple (U0, X0), but it can locally compute all such states
simultaneously. Then, for each IU,X ⊆ BU,X, each node v locally computes the sum
X max
IU0 ,X0⊆BU0 ,X0
(Extend(BU0,X0 , IU0,X0 , BU,X , IU,X) − IU,X) ,
where the sum is taken over all instances (U0, X0) such that v is in charge of U0, and then
the nodes in U compute an aggregate sum of all these values. Finally, each node can locally
compute Join(BU,X , IU,X) for each IU,X ⊆ BU,X using (3). In total, this step takes 2O(k) SA
rounds.
Recovering the solution. To recover the actual maximum independent set, we again follow
the process in Section 5.1. This time, we process the layers from layer 0 to layer T.
1. For the initial depth 0 with instance (U, X), recall that all nodes in U have learned Join(BU,X , IU,X)
for all IU,X ⊆ BU,X. All nodes in U compute I∗
ties lexicographically (so that all nodes in U can agree upon the same I∗
cast the set I∗
whether it belongs to I∗
(U0, X0), node v locally computes I∗
neighbors in U0. All of this can be done in O(k) distributed rounds.
U,X according to (4), where the arg max breaks
U,X), and then broad-
U,X to their neighbors. Since X ∪ S ⊆ N[U], each node in X ∪ S can learn
U,X. Then, for each node v ∈ U in charge of a recursive instance
U0,X0 to its
U0,X0 according to (5), and sends the set I∗
CVIT 2016
23:22
Distributed Treewidth Computation with Applications
2. For each instance (U0, X0) called by instance (U, X), the nodes in U0 adjacent to the node in
charge of (U0, X0) have received I∗
U0,X0 to all nodes in
U0, taking O(k) SA rounds. Then each node in U0 in charge of a recursive instance performs
the same as above.
U0,X0. These nodes broadcast the set I∗
At the end, every node in V knows whether it belongs to the maximum independent set
U,X I∗
U,X. This completes the distributed implementation, which runs in 2O(k) log n SA rounds.
This concludes Theorem 3.
S
1
2
References
Eyal Amir. Efficient approximation for triangulation of minimum treewidth. In Proceedings
of the Seventeenth conference on Uncertainty in artificial intelligence, pages 7–15. Morgan
Kaufmann Publishers Inc., 2001.
Stefan Arnborg, Derek G Corneil, and Andrzej Proskurowski. Complexity of finding em-
beddings in ak-tree. SIAM Journal on Algebraic Discrete Methods, 8(2):277–284, 1987.
3 Hans L Bodlaender. A linear-time algorithm for finding tree-decompositions of small
treewidth. SIAM Journal on computing, 25(6):1305–1317, 1996.
4 Marek Cygan, Fedor V Fomin, Łukasz Kowalik, Daniel Lokshtanov, Dániel Marx, Marcin
Pilipczuk, Michał Pilipczuk, and Saket Saurabh. Parameterized algorithms, volume 3.
Springer, 2015.
5 Atish Das Sarma, Stephan Holzer, Liah Kor, Amos Korman, Danupon Nanongkai, Gopal
Pandurangan, David Peleg, and Roger Wattenhofer. Distributed verification and hardness
of distributed approximation. In STOC, pages 363–372, 2011.
6 Uriel Feige, MohammadTaghi Hajiaghayi, and James R Lee.
Improved approximation
algorithms for minimum weight vertex separators. SIAM Journal on Computing, 38(2):629–
657, 2008.
7 Mohsen Ghaffari and Bernhard Haeupler. Distributed algorithms for planar networks ii:
Low-congestion shortcuts, mst, and min-cut. In Proceedings of the Twenty-Seventh Annual
ACM-SIAM Symposium on Discrete Algorithms, pages 202–219. Society for Industrial and
Applied Mathematics, 2016.
9
8 Mohsen Ghaffari, Fabian Kuhn, and Hsin-Hao Su. Distributed mst and routing in almost
mixing time. In Proceedings of the ACM Symposium on Principles of Distributed Comput-
ing, pages 131–140. ACM, 2017.
Bernhard Haeupler, D Ellis Hershkowitz, and David Wajc. Round-and message-optimal
distributed part-wise aggregation. arXiv preprint arXiv:1801.05127, 2018.
Bernhard Haeupler, Taisuke Izumi, and Goran Zuzic. Low-congestion shortcuts without
embedding. In Proceedings of the 2016 ACM Symposium on Principles of Distributed Com-
puting, pages 451–460. ACM, 2016.
Bernhard Haeupler, Taisuke Izumi, and Goran Zuzic. Near-optimal low-congestion short-
cuts on bounded parameter graphs. In International Symposium on Distributed Computing,
pages 158–172. Springer, 2016.
Bernhard Haeupler and Jason Li. Faster distributed shortest path approximations via
shortcuts. arXiv preprint arXiv:1802.03671, 2018.
Bernhard Haeupler, Jason Li, and Goran Zuzic. Minor excluded network families admit
fast distributed algorithms. arXiv preprint arXiv:1801.06237, 2018.
Samir Khuller and Baruch Schieber. Efficient parallel algorithms for testing k and finding
disjoint s-t paths in graphs. SIAM Journal on Computing, 20(2):352–375, 1991.
Jens Lagergren. Efficient parallel algorithms for graphs of bounded tree-width. Journal of
Algorithms, 20(1):20–44, 1996.
10
11
12
13
14
15
J. Li
16
17
Bruce A Reed. Finding approximate separators and computing tree width quickly.
In
Proceedings of the twenty-fourth annual ACM symposium on Theory of computing, pages
221–228. ACM, 1992.
Bruce A Reed. Algorithmic aspects of tree width. In Recent advances in algorithms and
combinatorics, pages 85–107. Springer, 2003.
18 Neil Robertson and Paul D. Seymour. Graph minors. ii. algorithmic aspects of tree-width.
Journal of algorithms, 7(3):309–322, 1986.
A
Omitted Proofs in Section 2
Proof (Lemma 7). Let H1, . . . , H' be the connected, disjoint subgraphs in an SA instance. The
reduction from SA to PA uses the Heads/Tails clustering technique from parallel graph contrac-
tion algorithms. We sketch the Heads/Tails clustering algorithm below, which, given a graph,
contracts each connected component into a single vertex. For O(log n) rounds, each vertex flips
Heads/Tails with probability 1/2 each and broadcasts the flip to its neighbors. If a vertex flips
Tails and has a neighbor with Heads, it notifies one of its Heads neighbors. Afterwards, every
Heads vertex and its Tails neighbors who notified it contract into a single vertex. Following the
standard analysis, w.h.p., every connected component contracts to a single vertex after O(log n)
iterations.
Now we describe the distributed SA algorithm. As input, each node in each subgraph Hi
knows its neighbors in that subgraph. Our goal is for the nodes in each Hi to agree on a common
ID, unique to each Hi. Then, a single round of PA solves the desired aggregation task.
At a high level, we want to run a distributed version of the Heads/Tails algorithm on the
graph H1 ∪ ··· ∪ H'. On each iteration, for each vertex v in the contracted graph, the nodes in
the original graph that contracted to v form a part for a PA round. In particular, every node in
a part should know its (unique) part ID. At the beginning, each node is its own part, and it can
set its part ID to be its node ID.
We now describe each of O(log n) steps of the clustering algorithm. In O(1) PA rounds, each
part collectively decides on a Heads/Tails flip for that part. Then, every node broadcasts its part
ID and its flip to all its neighbors in its own subgraph. Then, in O(1) PA rounds, any node that
received a Heads flip from a neighbor, and whose own part flipped Tails, notifies its part of the
part ID received by such a neighbor. Then, each part that flipped Tails collectively decides on
a common part ID (e.g., the minimum one) and sets its own part ID to be that one. The new
parts are the nodes with a common part ID; clearly, every part is still connected. Finally, after
O(log n) iterations, each taking O(1) PA rounds, all nodes within each subgraph Hi contracted
to a single vertex in the Heads/Tails algorithm, which means that they have agreed on a common
(cid:74)
part ID.
Proof (Lemma 9). We can assign arbitrary weights to the edges of H use the MST algorithm
of [7], which runs a modified Boruvka's algorithm with the current contracted components as the
(cid:74)
parts; for details, see [7, 12].
u whose aggregateL
aggregateL
Proof (Lemma 10). Computing subtree aggregation in the tree rooted at vr is covered in [12].
To determine the parent of each node except the root, we can set xi ← 1 for each node vi with
sum as the aggregate. For each node vi ∈ V (T) − vr, its parent is precisely the unique neighbor
(cid:74)
j∈T (u) xj is larger than the aggregateL
j∈T (vi) xj at vi.
j∈T (vi) xj is exactly the prefix aggregateL
Proof (Lemma 11). For prefix aggregation, let v' be the root of the tree P. Then, subtree
j≤i xj, so we can apply Lemma 10. For
each node vi to learn the index i, we set xi ← 1 for all i so that value i is exactly the prefix
23:23
CVIT 2016
23:24
Distributed Treewidth Computation with Applications
aggregateL
j≤i xj. Finally, for suffix aggregation, we compute subtree aggregates with v1 as the
(cid:74)
root instead.
Proof (Lemma 12). First, compute a spanning tree T ⊆ H using Lemma 9. Root the tree T at
t, and set xs ← 1 and xv ← 0 for all other nodes v ∈ V (H)− s. Now compute subtree aggregates
with sum as the operator. Observe that a node has nonzero subtree sum iff it is on the path from
6= 0 sends a message to its parent in the rooted tree, so
(cid:74)
s to t. Finally, each node v withP
that every node on the path learns its predecessor and successor.
j∈Tv
1
Omitted Proofs in Section 3
B
Proof (Claim 14). Since G has treewidth k, there exists a treewidth decomposition with max-
imum bag size k + 1. We now construct a treewidth decomposition5 T of G' with maximum bag
size (k+1)', which is sufficient to prove the claim. Starting from T , for each v ∈ G, we replace all
occurrences of v in bags in T with the vertices v
, . . . , v'. Clearly, the maximum bag size is now
at most (k + 1)'. We now claim that the new decomposition T 0 is a treewidth decomposition of
G'. Property (i) is clearly satisfied. For property (1), for each vertex vr ∈ V (G'), the set of bags
containing it is precisely the set of bags containing v in T , which is connected, proving property
(2). Finally, for property (3), for vertices ui, vj ∈ V (G') adjacent in G', either (i) u = v, in
which case any bag containing u in T contains both ui and vj, or (ii) u 6= v, in which case there
must be a bag in T containing both u and v, and this bag in T 0 contains both ui and vj.
(cid:74)
Proof (Lemma 18). By Corollary 17, the existence of an augmenting path is equivalent to s and
t being strongly connected in Gres. Since the transformation from Gres to GR preserves SCCs,
this is also equivalent to s and t being strongly connected in GR. Also, there is always a t → s
path in GR, so strong connectivity of s and t is equivalent to the existence of a directed s → t
path in GR. Therefore, we prove the following statement instead: there is a directed s → t path
in GR iff there exists s-reachable βi and t-reachable βx and a directed βi → βx path in GB. The
rest of the proof resembles the proof of Theorem 4.1 in [14].
For the if direction, suppose there is a directed βi → βx path in GB from s-reachable βi
to t-reachable βx; we will transform this path into an s → t path in GR. We replace each arc
(βj, βj0) with a directed path from βj to βj0 in GR as follows. By definition of GB, there is y ∈ [r]
satisfying ry
j ), the left
path along Py from ry
j0 , βj0). Finally, add the arcs (s, βi) and (βx, t)
to the path, completing the s → t path.
For the only if direction, suppose there is a directed s → t path P in GR; without loss of
generality, assume that P is simple. A simple path in GR consists of vertices βi with a leftward
path along some Pj in between every two consecutive βi. Let the βi vertices on P be βx1 , . . . , βx'
from left to right; by definition, βx1 is s-reachable and βx' is t-reachable. We now construct a
(not necessarily simple) path from βx1 to possibly a different t-reachable vertex in GB.
For i ∈ [' − 1], let yi be such that the path P travels along Pyi from βxi to βxi+1, and let y'
be an arbitrary integer in [r]. First, set x01 ← x1. Then, one by one, for i from 2 to ', we will
such that (i) either βx0
replace the vertex βxi with a vertex βx0
)
i−1 , βx0
≥ ryi
. Note that condition (ii) is satisfied by definition for
exists in GB, and (ii) we have ryi
x0
i = 1.
j0. We replace arc (βj, βj0) with the path composed of the arc (βj, ry
j0, and finally the arc (ly
Fix an i ∈ [2, ' − 1]; we assume that the invariant is satisfied at i. Since ryi
x0
x0, we can
to βxi+1 along path Pyi. Consider the arc (xi, x0) ∈ Dyi+1; we first assume that
travel from βx0
it exists. By condition 2(c) of the construction of the bridge graph with βxi+1 as βz, we have
≥ ryi
or the arc (βx0
i−1 = βx0
i
j ≥ ly
j to ly
i
xi
i
i
i
i
5 For our definition of treewidth and treewidth decomposition, we refer the reader to Section 5.
J. Li
23:25
i
'
'
i+1 ← βx0
Since βx' is t-reachable, we have ry
maintains the two properties for index i + 1.
x0 ≥ ryi+1
ryi+1
the two properties for index i + 1.
xi+1; otherwise, we contradict condition 2(c). Therefore, setting βx0
Now suppose arc (xi, x0) does not exist in Dyi+1. Then, we must have ryi+1
i+1 ← βx0 maintains
xi+1; otherwise,
βxi+1 satisfies conditions 2(a) and 2(b) of the bridge graph, so an arc in Dyi+1 must exist.
Therefore, setting βx0
= Py for all y ∈ [r], i.e., the rightmost out-neighbor
= Py', vertex βx0
≥ ry'
is also t-reachable. We
of vertex βx is t on each path Py. Since ry'
x0
'−1, obtaining our desired path in GB from
remove duplicates from the sequence βx0
2 , . . . , βx0
1 , βx0
(cid:74)
s-reachable βx1 to some t-reachable vertex.
Proof (Lemma 20). We define iteration i of the contraction algorithm as the iteration where
B be the bridge graph after iteration i ∈ [r],
the components of Dj are contracted. Also, let Gi
and G
j be the edge set Dj in Gi
B.
0
B := GB, and let Di
x0 ≥ ryi+1
x'
x'
j
For the if direction, we temporarily abuse notation, referring to vertices βi in GR and Gj
For the only if direction, we first show that if two vertices βi, βx contract to the same vertex,
then the bridges Bi, Bx are strongly connected in Gres. Before the contraction algorithm begins
(i.e., before iteration 1), this is clearly true. On iteration i, suppose the statement holds for
the bridge graph right before iteration i. Iteration i contracts the bridges inside each connected
component of Di−1
, which, by Claim 19, is strongly connected. Therefore, the statement holds
after iteration i as well. Applying induction on i proves the statement at the end of the algorithm.
Finally, taking βi and βx to be an s-reachable and t-reachable vertex finishes the only if direction.
B as
(super-) bridges as well as Bi. For bridge βi, consider all directed simple paths in GR that start
from βi and end at some t-reachable bridge. Define dist(βi) to be the minimum possible number
of bridges βx on such a path, minus 1. In particular, dist(βi) = ∞ iff there is no such path, and
dist(βi) = 0 iff βi is t-reachable. Also, for bridges βi, βx, we say that βi can directly reach βx in
GR if there is a directed βi → βx path in GR with no other bridges inside. We now prove, by
induction on d ≥ 0, that all bridges βi with dist(βi) = d contract to the a vertex containing an
t-reachable bridge. Since s and t are strongly connected in Gres, there is an s-reachable bridge
βi with finite dist(βi), proving the if direction.
The base case d = 0 is trivial; we now consider the case d = 1. Suppose a bridge βi satisfies
dist(βi) = 1. Then, exiting from bridge βi, we can move left along some path Py and then enter
x = Pj for all
an t-reachable bridge βx. Note that a t-reachable bridge necessarily satisfies rj
j ∈ [r], i.e., the rightmost out-neighbor of bridge βx is t on each path Pj. By definition of GB,
this means that for all j, there is an arc in Dj from βi to βx, or possibly a different t-reachable
bridge due to tie-breaking. In particular, this arc is in D1, so on iteration 1 of the algorithm, βi
and βx already contract to the same vertex.
Now suppose dist(βi) = d + 1 for d ≥ 1. First, if βi contracts to the same component as
some other βx with dist(βx) ≤ d, then by induction, βx contracts to a component containing a
t-reachable bridge. βi contracts to this same component, completing the inductive step.
Otherwise, consider the directed path P from βi to t in GR achieving dist(βi) = d + 1, and
suppose βi, βx0 , βy are the first three bridges on P. Suppose that, from βx0 to βy, the path P
x0 ≥ lj
y. Furthermore, either the
walks left along path Pj in GR; in particular, this means that rj
x ≥ rj
arc (βi, βx0) is in Dj, or some arc (βi, βx) exists in Dj such that rj
x0 and βi can directly
x ≥ rj
x0 ≥ lj
reach βx in GR; in the former case, we set x := x0. The inequalities rj
y imply that βx
can directly reach βy, and since dist(βy) = d − 1, we have dist(βx) ≤ d. Now consider iteration
j of the algorithm, which contracts all connected components in Dj in Gj−1
B . Before iteration j,
let βi0 and βx0 in Gj−1
B denote the super-bridges that βi and βx have contracted to; we assume
βi0 6= βx0, since otherwise we are in the first case. Observe that the corresponding super-bridge
Bi0 can still directly reach Bx0. In particular, if βx00 is the super-bridge for which (βi0 , βx00) is an
arc in Dj−1
x0 ≥ rj
y. In particular, some original bridge βx1 that contracts
x00 ≥ lj
y, which again implies that dist(βx1) ≤ d. Then,
to the super-bridge βx0 satisfies rj
x ≥ lj
x1 = rj
x00 ≥ rj
, then rj
j
CVIT 2016
23:26
Distributed Treewidth Computation with Applications
on iteration j, super-bridges βi0 , βx0 contract together, so bridges βi, βx1 contract to the same
super-bridge. The fact that dist(βx1) ≤ d completes the inductive step.
(cid:74)
|
1004.5600 | 1 | 1004 | 2010-04-30T19:42:14 | On the (Im)possibility of Preserving Utility and Privacy in Personalized Social Recommendations | [
"cs.DS"
] | With the recent surge of social networks like Facebook, new forms of recommendations have become possible -- personalized recommendations of ads, content, and even new social and product connections based on one's social interactions. In this paper, we study whether "social recommendations", or recommendations that utilize a user's social network, can be made without disclosing sensitive links between users. More precisely, we quantify the loss in utility when existing recommendation algorithms are modified to satisfy a strong notion of privacy called differential privacy. We propose lower bounds on the minimum loss in utility for any recommendation algorithm that is differentially private. We also propose two recommendation algorithms that satisfy differential privacy, analyze their performance in comparison to the lower bound, both analytically and experimentally, and show that good private social recommendations are feasible only for a few users in the social network or for a lenient setting of privacy parameters. | cs.DS | cs | On the (Im)possibility of Preserving Utility and Privacy in
Personalized Social Recommendations
Ashwin Machanavajjhala
[email protected]
Aleksandra Korolova
[email protected]
Atish Das Sarma
[email protected]
0
1
0
2
r
p
A
0
3
]
S
D
.
s
c
[
1
v
0
0
6
5
.
4
0
0
1
:
v
i
X
r
a
ABSTRACT
With the recent surge of social networks like Facebook, new
forms of recommendations have become possible -- personal-
ized recommendations of ads, content, and even new social
and product connections based on one's social interactions.
In this paper, we study whether "social recommendations",
or recommendations that utilize a user's social network, can
be made without disclosing sensitive links between users.
More precisely, we quantify the loss in utility when existing
recommendation algorithms are modified to satisfy a strong
notion of privacy called differential privacy. We propose
lower bounds on the minimum loss in utility for any recom-
mendation algorithm that is differentially private. We also
propose two recommendation algorithms that satisfy differ-
ential privacy, analyze their performance in comparison to
the lower bound, both analytically and experimentally, and
show that good private social recommendations are feasible
only for a few users in the social network or for a lenient
setting of privacy parameters.
1.
INTRODUCTION
Making recommendations or suggestions to users to in-
crease their degree of engagement is a common practice for
websites. For instance, Facebook recommends friends to
existing users, Amazon suggests products, and Netflix rec-
ommends movies, in each case with the goal of making as
relevant a recommendation to the user as possible. Recom-
mending the right content, product, or ad to an individual is
one of the most important tasks in today's web companies.
With the boom in social networking many companies are
striving to incorporate the likes and dislikes of an individ-
ual's social neighborhood. There has been much research
and industrial activity to solve two problems: (a) recom-
mending content, products, ads not only based on the indi-
vidual's prior history but also based on the history of those
the individual trusts [12, 2], and (b) recommending others
whom the individual might trust. Recommendations based
on social connections are especially effective for users who
have seen very few movies, bought only a couple of products,
or never clicked on ads; while traditional recommender sys-
tems default to generic recommendations, a social-network
aware system can provide useful recommendations based on
active friends. Companies like TrustedOpinion1 and SoMR2
generate content and ad recommendations by leveraging so-
cial networks. In fact, Facebook3, Yahoo!4 and Google5 are
opening their social networks to third party developers to
encourage social network-aware recommender systems.
In addition, a social network might want to use a different
underlying social network, such as one derived from e-mail
records or Instant Messenger connections, to suggest friends
(e.g. Facebook already uses contacts imported from an ad-
dress book as suggestions). Social connections could also be
used to recommend products or advertisements to users --
Netflix (or Opentable or Yelp) could recommend movies (or
restaurants) to a subscriber based on her friends' activities
and ratings.
In fact, rather than using the entire social
graph, the system could use only a subset of trusted edges
for that application (for instance, a user might only trust
the movie recommendations of a subset of her friends).
However, these improved recommendations based on so-
cial connections come at a cost -- a recommendation can
potentially lead to a privacy breach by revealing sensitive
information. For instance, while the social network links
might be public, both the user-product links and the user-
user-trust links must be kept secret. (Knowing that your
friend doesn't trust your judgement about books might be
a breach of privacy). Similarly, revealing an edge in an e-
mail graph, or revealing that a particular user purchased
a sensitive product, constitutes a potentially serious breach
of user privacy. Recommendations can indeed lead to such
privacy breaches even without the use of social connections
in the recommendation algorithm [5]. The privacy concerns
posed by recommender systems and use of the social network
graph have been at the forefront of industry discussion on
the topic. In 2007, Facebook attempted to incorporate the
product purchases made by one's friends into the stream of
news one receives while visiting the site through a product
called Beacon. Their launch showed that people interact
with many websites and products in a way that they would
not want their friends to know about, leading to several
privacy lawsuits, and an eventual complete removal of the
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
Copyright 200X ACM X-XXXXX-XX-X/XX/XX ...$5.00.
1http://www.trustedopinion.com/
2http://www.somrnetworks.com/
3http://developers.facebook.com/connect.php
4http://developer.yahoo.com/yos/intro/
5http://www.google.com/friendconnect/
feature by Facebook.
In this paper, we present the first theoretical study of
the privacy-utility trade-offs in social recommender systems.
While there are many different settings where social recom-
mendations are used (friend/product recommendations, or
trust propagation), and each leads to a slightly different for-
mulation of the privacy problem (the sensitive information
is different in each case), all these problems have the fol-
lowing common structure -- recommendations are made on
a graph where some subset of edges are sensitive. For clar-
ity of exposition, we ignore (by and large) scenario specific
constraints, and focus on the following general model. We
consider a graph where all the edges are sensitive, and an
algorithm that recommends a single node v in the graph
to some target node u. We assume that the algorithm is
based on a utility function that encodes the "goodness" of
recommending each node in the graph to this target node.
Suggestions for utility functions include number of common
neighbors, weighted paths and PageRank distributions [21].
We consider an attacker who wishes to deduce the existence
of a single edge (x, y) in the graph by passively observing the
recommendation (v, u). We measure the privacy of the algo-
rithm using differential privacy -- the ratio of the likelihoods
of the algorithm recommending (v, u) on the graphs with
the edge (x, y) and without the edge (x, y), respectively. In
this setting, we ask the question: to what extent can edge
recommendations be accurate while preserving differential
privacy?
Our Contributions and Overview.
In this paper we
present the following results on the accuracy of differentially
private social recommendations.
• We present a trade-off between the accuracy and pri-
vacy of any social recommendation algorithm that is
based on any general utility function. This trade-off
shows an inevitable lower bound on the privacy pa-
rameter that must be incurred by an algorithm that
wishes to guarantee any constant-factor approximation
of the maximum utility. (Section 4)
• We present lower bounds on accuracy and privacy for
algorithms based on specific utility functions previ-
ously suggested for recommending edges in a social
network -- number of common neighbors and weighted
paths [21]. Our trade-offs for these specific utility func-
tions present stronger lower bounds than the general
one that is applicable for all utility functions. (Section
5)
• We adapt two well known privacy preserving algo-
rithms from the differential privacy literature for the
problem of social recommendations. The first (which
we call Laplace),
is based on adding random noise
drawn from a Laplace distribution to the utility vector
[8] and then recommending the highest utility node.
The second (Exponential), is based on exponential smooth-
ing [15]. We analyze and compare the accuracy of the
two algorithms and comment on their relative merits.
(Section 6)
• We perform experiments on a real graph using the
number of common neighbors utility function. The
experiments compare the algorithms Laplace, Expo-
nential, and our lower bound. Our experiments sug-
gest three takeaways: (i) For most nodes, the lower
bounds suggest that there is a huge inevitable trade-
off between privacy and accuracy when making social
recommendations; (ii) The more natural Laplace algo-
rithm performs as well as Exponential; and (iii) For a
large fraction of nodes, both Laplace and Exponential
almost achieve the maximum accuracy level suggested
by our theoretical lower bound. (Section 7)
• We briefly consider the setting when an algorithm may
not know (or be able to compute efficiently) the entire
utility vector. We recognize that both Laplace and Ex-
ponential algorithms assume the knowledge of all the
utilities (for every node) when recommending to a tar-
get node. We propose and analyze a sampling based
linear smoothing algorithm that does not require all
utilities to be pre-computed. We conclude by men-
tioning several directions for future work. (Section 8)
We now discuss related work and then formalize the mod-
els and definitions in Section 3.
2. RELATED WORK
Several papers propose that the social connections avail-
able can be effectively utilized for enhancing online appli-
cations [12, 2]. Golbeck [10] uses the trust relationships
expressed through social connections for personalized movie
recommendations and shows that the accuracy of the ratings
outperform those produced by a collaborative filtering algo-
rithm not utilizing the social graph. Mislove et al. [16] at-
tempt an integration of web search with social networks and
explore the use of trust relationships, such as social links, to
thwart unwanted communication [17]. Approaches incorpo-
rating trust models into recommender systems are gaining
momentum both in academic research [25], [18], [23], and in
real products. Examples include, Chorus6, which provides
social app recommendations for the iPhone; Fruggo.com7,
a social e-commerce site; and WellNet's8 online social net-
working program for health care coordination9.
Calandrino et al.
[5] demonstrate that algorithms that
recommend products based on a friends' purchases have very
practical privacy concerns: "passive observations of Ama-
zon.com's recommendations are sufficient to make valid in-
ferences about individuals' purchase histories". McSherry et
al. [14] show how to adapt the leading algorithms used in the
Netflix prize for movie recommendations to make privacy-
preserving recommendations. Their work does not apply
to algorithms that rely on the underlying social graph be-
tween users, as the user-user connections have not been re-
leased as part of the Netflix competition. Aımeur et al. [1]
study the problem of personalized recommendations in gen-
eral. Dwork et al.
[9] pose the problem of constructing
differentially private analysis of social networks. Toubiana
et al.
[24] propose a framework for privacy preserving tar-
geted advertising -- while targeting based on user history
is considered, targeting based on social interactions is not
considered.
A related and independent work [4] considers the problem
of mining top-k frequent item-set. Although they consider
mechanisms analogous to the ones we propose, since we solve
6http://www.envionetworks.com/
7http://fruugo.com/
8http://www.wellnet.com/
9www.medicalnewstoday.com/articles/118628.php
two different problems, the focus of their analysis, notion of
utility, and conclusions substantially differ from ours.
3. MODEL
In this section, we describe the model for privacy-preserving
social recommendations. We first define a social recommen-
dation algorithm and then mention notions of monotonicity
and accuracy of an algorithm. We then define axioms fol-
lowed by typical utility functions that such algorithms are
based on. Finally, we define differential privacy.
3.1 Social Recommendation Algorithm
i
Let G = (V, E) be the graph that describes the social net-
work. Each recommendation is an edge (i, r), where node i
is recommended to the target node r. Given a graph G, and
a target node r, we denote the utility of recommending node
i to node r by uG,r
. Further, we assume that a recommen-
dation algorithm R is a probability vector on all nodes. Let
pG,r
denote the probability of recommending node i to node
i
r in graph G by a specified algorithm. When the graph G
and the source node r are clear from context, we drop G and
r from the notation -- ui denotes the utility of recommend-
ing i, and pi denotes the probability that R recommends i.
We further define umax = maxi ui.
i ui · pi) of each recommendation.
We consider algorithms that attempt to maximize the ex-
If we
assume (without loss of generality) that the utility of the
least useful recommendation is 0, the accuracy of such an
algorithm can be defined as:
pected utility ((cid:80)
.
umax
Definition 1
(cid:80) uipi
(Accuracy). An algorithm A is said to
be (1 − δ)-accurate if given any set of utilities ui (for all i)
denoted by (cid:126)u, A recommends node i with probability pi such
that (1 − δ) = min(cid:126)u
Therefore, an algorithm is said to be (1 − δ)-accurate if for
any utility vector, the algorithm's expected utility is at least
(1−δ) times the utility of the highest utility node in the given
utility vector. It is easy to check that for the case when the
utility of the least useful recommendation is umin, in all of
our subsequent discussions, the definition of accuracy we
use is equivalent to accuracy defined as the fraction of the
difference between umax and umin.
Scale Invariance of Sensitivity and Utility Func-
tions. We initiate a small discussion on what happens
when the utility values for all potential recommendable nodes
are scaled by a multiplicative factor, or changed by an addi-
tive constant. Intuitively, since the scale of utilities is chosen
arbitrarily, one would expect the algorithms and the anal-
ysis to be invariant to such numeric changes. However, be-
cause of the constraints imposed by the desire to be privacy-
preserving, where the privacy-preservation is with respect to
a presence or absence of a particular edge, the scale invari-
ance assumptions require a more careful articulation.
In
particular, the crucial point of interaction between the pri-
vacy requirement and the utility function is the concept of
sensitivity, denoted by ∆f , which is the maximum change
in a utility vector (cid:126)u that can occur due to an addition or
removal of one edge in the graph. Observe that if we scale a
utility function by a multiplicative constant, the sensitivity
of the utility function is scaled as well by the same constant.
Without loss of generality, and for ease of subsequent expo-
sition, we assume that ∆f = 1, an assumption that implies
that the magnitudes of the utilities are now meaningful, as
the higher utility magnitude corresponds to more edges that
need to be added or removed in the graph in order to achieve
it. Equivalently, we could have chosen to let the utilities be
scale invariant, but would then need to compute and reason
in terms of the sensitivity of the utility function.
Another property that is natural of a recommendation
algorithm is monotonicity:
Definition 2
(Monotonicity). An algorithm is said
to be monotonic if ∀i, j, ui ≥ uj implies that pi ≥ pj.
3.2 Axioms on Utility Functions
We now define two axioms that we believe should be sat-
isfied by any meaningful utility function in the context of
recommendations on a social network. These axioms are
later used in proving our theoretical results. Our axioms are
inspired by the work of [21] and the specific utility functions
they consider, which include: number of common neighbors,
sum of weighted paths, and PageRank based utility mea-
sures.
Axiom 1
(Exchangeability). Let G be a graph and
let h be an isomorphism on the nodes giving graph Gh, s.t.
for target node r, h(r) = r. Then ∀i : uG,r
i = uGh,r
h(i) .
This axiom captures the intuition that the utility of a
node i should not depend on the node's name. Rather, its
utility with respect to target node r only depends on the
structural properties of the graph, and so, nodes that are
isomorphic from the perspective of the target node r should
have the same utility.
V (G), such that S = β, and(cid:80)
(Concentration Axiom). There exists S ⊂
i∈V (G) ui.
i∈S ui ≥ Ω(1)(cid:80)
Axiom 2
This says that there are some β nodes that together have
at least a constant fraction of the total utility mass. This ax-
iom is likely to be satisfied for small enough β, since usually
there are some nodes that are very good for recommendation
and many that are not so good.
In the subsequent lower bound sections, we only consider
monotonic algorithms for utility functions that satisfy the
exchangeability axiom as well as the concentration axiom
for a reasonable choice of β.
A running example throughout the paper of a utility func-
tion that satisfies these axioms in practical settings and is
often deployed [21] is that of the number of common neigh-
bors utility function: given a target node r and a graph
G, the common neighbors utility metric assigns a utility
uG,r
i = c(i, r), where c(i, r) is the number of common neigh-
bors between i and r.
3.3 Differential privacy
Differential privacy [6] is a strong definition of privacy
that is based on the following principle: an algorithm pre-
serves the privacy of an entity if the algorithm's output is
not sensitive to the presence or absence of the entity's infor-
mation in the input data set. In our setting of graph-based
social recommendations, we wish to maintain the presence
(or absence) of an edge in the graph private. Hence, the
privacy definition can be formally stated as follows.
Definition 3. A recommendation algorithm R satisfies
-differential privacy if for any pair of graphs G and G(cid:48) that
differ in one edge (i.e., G = G(cid:48) + {e} or vice versa) and
every set of possible recommendations S,
P r[R(G) ∈ S] ≤ exp() × P r[R(G
(cid:48)
) ∈ S]
(1)
where the probabilities are over the random coins of R.
Differential privacy has been widely used in the privacy liter-
ature [3, 8, 13, 15, 7], since it is even resilient to adversaries
who know all but one edges in the graph, and guarantees
privacy for multiple runs of the algorithm. While weaker
notions of privacy have also been considered in the litera-
ture, in this paper we focus on the strong differential privacy
definition only. Since in social recommendations protecting
privacy is extremely important, it seems reasonable to first
explore and understand the strongest notions of privacy.
In this paper, we only consider the utility of a single social
recommendation. We note that in this setting, we can relax
the differential privacy definition such that Equation 1 only
holds for graphs G and G(cid:48) that differ in an edge e that is
not incident on r, the target of the recommendation. This
mirrors the natural setting where (a) one recommendation
is made to the attacker (r), (b) only the target node (the
attacker) sees the recommendation. By considering G and
G(cid:48) that differ in e = (i, r), the adversary can only learn
about his neighborhood (which he is aware of to start with)
and not learn whether two legitimate nodes in the graph
are connected. While we consider a single recommendation
throughout the paper, we use the relaxed variant of differ-
ential privacy only in Sections 5 and 7.
4. GENERAL LOWER BOUND
In this section we prove a lower bound on the privacy
parameter on any differentially private recommendation
algorithm that (a) achieves a constant accuracy and (b) is
based on any utility function that satisfies the exchangeabil-
ity and concentration.
Let us first sketch the proof technique for the lower bound
using the number of common neighbors utility metric, and
then state the lower bound for a general utility metric. An
interested reader can find the full proofs in the Appendix.
Recall that given a target node r and a graph G, the common
neighbors utility metric assigns a utility uG,r
i = c(i, r), where
c(i, r) is the number of common neighbors between i and r.
The nodes in any graph can be split into two groups -- V r
hi,
nodes which have a high utility for the target node r and
V r
lo, nodes that have a low utility. In the case of common
neighbors, all nodes i in the 2-hop neighborhood of r (who
have at least one common neighbor with r) can be part of
V r
hi and the rest in V r
lo. Since the recommendation algorithm
has to achieve a constant accuracy, it has to recommend one
of the high utility nodes with constant probability.
hi, but there are many nodes in V r
By the concentration axiom, there are only a few nodes
in V r
lo; in the case of com-
mon neighbors, node r may only have 10s or 100s of 2-hop
neighbors in a graph of millions of users. Hence, there exists
a node i in the high utility group and a node (cid:96) in the low
utility group such that Γ = pi/p(cid:96) is very large (Ω(n)). At
this point, we show that we can carefully modify the graph
G by adding and/or deleting a small number (t) of edges in
such a way that the node (cid:96) with the smallest probability of
being recommended in G becomes the node with the highest
utility in G(cid:48). By the exchangeability axiom, we can show
that there always exist some t edges that make this possi-
ble. For instance in the common neighbors case, we can do
this by adding edges between a node i and t of r's neigh-
bors, where t > maxi c(i, r). It now follows from differential
privacy that
≥ 1
t
log Γ
More generally, let V r
hi be the set of nodes 1, . . . , k each of
which have utility ui > (1− c)umax, and let V r
lo be the nodes
k + 1, . . . , n each of which have ui ≤ (1 − c)umax utility of
being recommended to target node r. Recall that umax is
the utility of the highest utility node. Let t be the number
of edge alterations required to turn a node with the smallest
probability of being recommended from the low utility group
into a node of maximum utility in the modified graph. The
following lemma states the main trade-off relationship be-
tween the accuracy parameter δ and the privacy parameter
of a recommendation algorithm.
Lemma 1. ≥ 1
t
δ ) + ln( n−k
(cid:0)ln( c−δ
k+1 )(cid:1)
This lemma gives us a lower bound on the privacy guar-
antee in terms of the utility parameter δ. Equivalently,
Corollary 1. 1 − δ ≤ 1 −
c(n−k)
n−k+(k+1)et
By using the concentration axiom with parameter β we
can prove the following.
Lemma 2. For (1 − δ) = Ω(1) and β = o(n/ log n),
≥ log n − o(log n)
t
(2)
This expression can be intuitively interpreted as follows:
in order to achieve good accuracy with a reasonable amount
of privacy (where is independent of n), either the number
of nodes with high utility needs to be very large (i.e. β needs
to be very large, Ω(n/ log n)), or the number of steps needed
to bring up any node's utility to the highest utility needs to
be large (i.e. t needs to be large, Ω(log n)).
We shall use this relationship from Lemma 2 in the sub-
sequent section to prove stronger lower bounds for specific
utility functions. Below we mention a generic lower bound
that applies to any utility function. Note that we only need
an upper bound on t. The tighter upper bound we are able
to prove on t, the better lower bound we get for .
Using the exchangeability axiom, we can show that t ≤
4∗ dmax in any graph. Consider the highest utility node and
the lowest utility node, say x and y respectively. These
nodes can be interchanged by deleting all of x's current
edges, adding edges from x to y's neighbors, and doing the
same for y. This requires at most 4 ∗ dmax changes. Hence,
Theorem 1. For a graph with maximum degree dmax =
α log n, a differentially private algorithm can guarantee con-
stant accuracy only if
(cid:18) 1
4
≥ 1
α
− o(1)
(cid:19)
(3)
In the next section, we present stronger lower bounds for
two well studied utility functions -- common neighbors and
weighted paths.
5. LOWER BOUNDS FOR SPECIFIC UTIL-
ITY FUNCTIONS
In this section, we start from Lemma 2 and prove stronger
lower bounds for specific utility functions by proving stronger
upper bounds on t. Proofs and more details can be found
in the Appendix.
5.1 Common neighbors lower bound
Consider a graph and a target node r. As we saw in the
previous section, we can make any node x have the highest
utility by adding edges to all of r's neighbors. If dr is r's
degree, it suffices to add t = dr + O(1) edges to make a node
the highest utility node. We state the following theorem for
a more general version of common neighbors utility function
below.
Theorem 2. Let U be a utility function that depends only
on and is monotonically increasing with c(x, y), the number
of common neighbors between x and y. A recommendation
algorithm based on U that guarantees any constant approx-
imation to utility for target node r has a lower bound on
privacy given by ≥ 1−o(1)
where dr = α log n.
α
As we will show in Section 7, this is a very strong lower
bound. Since a significant fraction of nodes in real-world
graphs have small dr (due to a power law degree distribu-
tion), we can expect no algorithm based on common neigh-
bors utility to be both accurate on most nodes and satisfy
differential with a reasonable .
5.2 Weighted Paths
A natural extension of the common neighbors utility func-
tion and one whose usefulness is supported by the literature
[21], is the weighted path utility function, defined as
score(s, y) =(cid:80)inf
l=2 γl−1paths(l)
(s,y),
(s,y) denotes the number of length l paths from
where paths(l)
s to y. Typically, one would consider using small values of
γ, such as γ = 0.005, so that the weighted paths score is a
"smoothed version" of the common neighbors score.
Again let r be the target node. To make node y the highest
utility node, we add edges such that y has cdr common
neighbors with r. Now, the goal is to choose c > 1 such
that this alone is sufficient to ensure that y has the highest
utility. This is done by showing that (a) no other node
has more than dr common neighbors with r, and (b) the
utility derived from paths of length ≥ 3 cannot offset the
additional common neighbors between y and r (for suitably
small γ). Finally, we show that this requires adding only
t < dr + 2 ∗ (c − 1) + O(1).
Theorem 3. For weighted paths based utility functions
with parameter γ, we have t ≤ (1 + o(1))dr when making
). Therefore, for
recommendations for node r, if γ = o(
an algorithm to guarantee constant approximation to utility,
the privacy must be ≥ 1
α (1 − o(1)) where dr = α log n.
dmax
1
6. PRIVACY-PRESERVING RECOMMENDA-
TION ALGORITHMS
There has been a wealth of literature on developing differ-
entially private algorithms [3, 8, 15]. In this section we will
adapt two well known tools, Laplace noise addition [8] and
exponential smoothing [15], to our problem. For the pur-
pose of this section, we will assume that given a graph and
a target node, our algorithm has access to (or can efficiently
compute) the utilities ui for all other nodes in the graph.
Given this vector of utilities, our goal is to compute a vec-
i ui · pi is maximized,
tor of probabilities pi such that (a)(cid:80)
and (b) differential privacy is satisfied.
Clearly, maximum accuracy is achieved by recommending
the node with utility umax. However, it is well known that
any algorithm that satisfies differential privacy must recom-
mend every node, even the ones that have zero utility, with a
non-zero probability [20]. Indeed, suppose for graph G and
target node r, an algorithm assigns 0 probability to some
node x with utility uG,r
and a positive probability to some
. Transform G into G(cid:48) as follows:
node y, with utility uG,r
connect x to all of y's neighbors in G and disconnect x from
all its neighbors in G. Do the same for y. This in turns cre-
ates an isomorphism h between G and G(cid:48), where h(r) = r.
Hence, by the exchangeability axiom, the algorithm will rec-
ommend y with 0 probability. Thus, there is a path from G
to G(cid:48) of length t such that py goes from a positive number
to 0. This leads to a breach of differential privacy.
x
y
The following two algorithms ensure differential privacy:
6.1 Exponential mechanism
The exponential mechanism creates a smooth probability
distribution from the utility vector and then samples from
that.
∆f ui /(cid:80)n
Definition 4. Exponential mechanism: Given nodes
with utilities (u1, . . . , ui, . . . , un), algorithm AE() recom-
mends node i with probability
∆f uk , where ≥ 0 is the privacy parameter,
e
and ∆f is the sensitivity of the utility function10.
k=1 e
Theorem 4. [15] AE() guarantees differential privacy.
6.2 Laplace mechanism
Unlike the exponential mechanism, the Laplace mecha-
nism mimics the optimal mechanism. It first adds random
noise drawn from a Laplace distribution, and like the opti-
mal mechanism, picks the node with the maximum noise-
infused utility.
Definition 5. Laplace mechanism: Given nodes with
utilities (u1, . . . , ui, . . . , un), algorithm AL() first computes
a modified utility vector (u(cid:48)
i = ui + r
where r is a random variable chosen from the Laplace distri-
bution with scale ( ∆f
)11 independently at random for each
i. Then, AL() recommends node z whose noisy utility is
maximal among all nodes, i.e. z = arg maxi u(cid:48)
i.
n) as follows: u(cid:48)
1, . . . , u(cid:48)
Theorem 5. AL() guarantees differential privacy.
Proof. The proof follows from the privacy proof of the
Laplace mechanism in the context of publishing privacy-
preserving histograms [8] by observing that one could treat
each node as a histogram bin and release the noisy count
for the value in that bin, u(cid:48)
i. Since AL() is effectively doing
post-processing by releasing only the name of the bin with
the highest noisy count, the algorithm remains private.
10∆f = maxr maxG,G(cid:48):G=G(cid:48)+e (cid:126)uG,r − (cid:126)uG(cid:48),r
11In this distribution, the pdf at y is
2∆f exp(−y/∆f )
An astute reader might remark at this point that the
Laplace mechanism as stated does not satisfy the mono-
tonicity property that we relied upon in our lower bound
proofs. Indeed, the Laplace mechanism satisfies the prop-
erty only in expectation; however, that is not an obstacle to
our analysis since in order to meaningfully compare the per-
formance of Laplace mechanism with other mechanisms and
with the theoretical bound on performance, we would need
to evaluate its expected, rather than one-time, performance.
6.3 Exponential vs Laplace Mechanisms
It is natural to ask whether there is an equivalence be-
tween the two approaches of transforming a non-private al-
gorithm to a privacy-preserving algorithm or how they would
compare, perhaps depending on the setting. We present pre-
liminary results on comparing the utilities when there are
only two possible recommendations (n = 2). The theorem
is stated below and the proof can be found in the Appendix.
Theorem 6. Let UE and UL denote the utilities achieved
by AE() and AL() on input vector (u1, u2), respectively.
Wlog, assume u1 ≥ u2. Then UE = u1
and
UL = u1(1 − 1
2 e−(u1−u2) − (u1−u2)
(u1−u2)
4e(u1−u2) )
4e(u1−u2 ) ) + u2( 1
2 e−(u1−u2) +
eu1
eu1 +eu2 +u2
eu2
eu1 +eu2
To our knowledge, in the course of the proof we give the
first explicit closed form expression for the probabilities of
each of the two nodes being recommended by Laplace mech-
anism (the work of [19] gives a formula that does not apply
to our setting).
Although the expressions for UE and UL are difficult to
compare by eye-balling, by plugging in various values of
u1 and u2 into the formulas, one infers that the Exponen-
tial mechanism slightly outperforms the Laplace mechanism,
when is very small and the difference between u1 and u2 is
large. We leave it for future work to simplify these as well
as extend the analysis to the n > 2 case.
Implementation efficiency. The Laplace mechanism is
more intuitive than the Exponential mechanism, and more
likely to receive executive buy-in in a real-world environ-
ment. Furthermore, it has the advantage that it can be
implemented more easily than the Exponential mechanism.
AL requires computing the noisy utilities and then selecting
the node with the highest noisy utility, which takes linear
time. AE requires first computing a set of smoothed utilities
and then sampling from the probability distribution induced
by them, which can be accomplished in linear time using the
alias-urn method suggested by [22], but likely slightly less
practically efficiently than AL.
7. UTILITY ACHIEVABLE IN PRACTICE
ON A REAL GRAPH
In this section we present experimental results on a real
graph and for the # of common neighbors utility function.
The experiments compare the algorithms Laplace, Exponen-
tial, and our lower bound. Our experiments suggest three
takeaways: (i) For most nodes, the lower bounds suggest
that there is a huge inevitable trade-off between privacy
and accuracy when making social recommendations; (ii) The
more natural Laplace mechanism performs as well as the Ex-
ponential mechanism; and (iii) For a large fraction of nodes,
the accuracy achieved by Laplace and Exponential mecha-
nisms does not substantially differ from the best possible
accuracy suggested by our theoretical lower bound.
7.1 Experimental Setup
For our experiments we use the Wikipedia vote network
[11] available from Stanford Network Analysis Package12.
Some users in Wikipedia are administrators, who have ac-
cess to additional technical features. Users are elected to
be administrators via a public vote of other users and ad-
ministrators. The Wikipedia vote network consists of all
users participating in the elections (either casting a vote or
being vote on), since inception of Wikipedia until January
2008. We turn the network of [11] into an undirected net-
work, where each node represents a user and an edge from
node i to node j represents that user i voted on user j or
user j voted on user i. This obtained network consists of
7,115 nodes and 100,762 edges. Although the Wikipedia
vote network is publicly available, and hence the edges in it
are not private, we believe that the graph itself exhibits the
structure and properties of some of the graphs in which one
would want to preserve privacy, such as the graph of social
connections and people's product purchases.
For each pair of nodes in the social network, except nodes
that share an edge, we compute the number of common
neighbors they have in the Wikipedia vote network. Then,
assuming we will make one recommendation for each node in
the graph, we compute the expected accuracy of recommen-
dation for that node. For the Exponential mechanism and
the theoretical bound, given the utilities of recommending
each node to a given node v, we can compute the expected
accuracy and the theoretical bound on accuracy exactly. For
the Laplace mechanism, we compute its expected accuracy
by running 1, 000 independent trials of the Laplace mecha-
nism, and averaging the utilities obtained in those trials, for
each node in the graph.13
7.2 Exponential vs Laplace in practice
We first observe in Figure 1 that for all nodes in the
Wikipedia vote network, the Laplace mechanism achieves
nearly identical accuracy as the Exponential mechanism.
This confirms our hypothesis of Section 6 that Exponential
and Laplace mechanisms are nearly equivalent in practical
settings, and implies that one can use the more intuitive and
easily implementable Laplace mechanism in practice.
7.3 Social Recommendations: Good or Pri-
vate?
We now proceed to evaluate the accuracy of the Exponen-
tial mechanism and compare it with the best accuracy one
can hope to achieve using a privacy-preserving recommen-
dation algorithm, as computed according to our theoretical
bound of Corollary 1.
For ease of visual presentation, we assume that we do not
care about node identities; we number the nodes in decreas-
ing order of the accuracy one can hope for when making
the recommendation for that node, as predicted by the the-
oretical bound. For each node, the graph in Figure 2 shows
the theoretical bound and the accuracy achieved by the Ex-
12http://snap.stanford.edu/data/wiki-Vote.html
13Out of the 7,115 nodes, there are 60 nodes that have no
common neighbors with anyone except nodes they are al-
ready connected to. We omit those nodes from our analysis.
Figure 1: Accuracy achieved by Exponential and
Laplace mechanisms on Wikipedia vote network using
# of common neighbors as a measure of utility. The
x-axis represents the node number, the y-axis - the ac-
curacy of recommendation for that node. The top graph
is for desired privacy guarantee of = 0.1, the bottom -
for = 0.5.
Figure 2: Accuracy achieved by Exponential mecha-
nisms and predicted by theoretical bound on Wikipedia
vote network using # of common neighbors as a measure
of utility. The x-axis represents the node number, the y-
axis - the expected accuracy of recommendation for that
node. The top graph is for desired privacy guarantee of
= 0.1, the bottom - for = 0.5.
ponential mechanism. Due to our chosen numbering of the
nodes, the theoretical bound is a smooth monotonically de-
creasing function of the node number, whereas the achieved
accuracy is not necessarily monotonically decreasing (and
thus, in places, does not appear as a line).
As can be seen in Figure 2 and Figure 3, for some nodes,
the Exponential mechanism performs quite well, achieving
nearly perfect accuracy. However, the number of such nodes
is fairly small - the Exponential mechanism achieves better
than 0.9 approximation for less than 1.5% of the nodes when
= 0.1 and less than 21% of the nodes when = 0.5, it
achieves better than 0.8 approximation for less than 2% of
the nodes when = 0.1 and less than 25.5% of the nodes
when = 0.5. This matches the intuition that by making
the privacy requirement more lenient, one can hope to make
better quality recommendations for more nodes; however,
this also pinpoints the fact that for most nodes, the Expo-
nential mechanism does not achieve good accuracy.
Although there is a possibility that one could develop bet-
ter privacy-preserving recommendation mechanisms than Ex-
ponential or Laplace, this experiment shows that for a large
number of target nodes, our theoretical bound limits the
best accuracy one can hope to achieve privately quite severely.
For example, for = 0.1, no privacy-preserving algorithm
can hope to achieve a better than 70% accuracy for more
than 9% of the nodes. This finding throws into serious ques-
tion the feasibility of developing social recommendation al-
gorithms that are both accurate and privacy-preserving for
many real-world settings.
Finally, in practice, it is the least connected nodes that
are likely to benefit most from receiving high quality rec-
ommendations. However, our experiments suggest that the
low degree nodes are also the most vulnerable to receiv-
ing low accuracy recommendations due to needs of privacy-
preservation: see Figure 4 for an illustration of how accuracy
depends on the degree of the node. This further suggests
that, in practice, one has to make a choice between preserv-
ing accuracy vs preserving privacy.
7.4 Are AE and AL good enough for utility func-
tion based on common neighbors?
As we have experimentally observed in Figure 2, the Ex-
ponential mechanism achieves good accuracy compared to
the best achievable accuracy predicted by our theoretical
bound. We can formalize this statement rigorously as fol-
lows (proved in the Appendix):
Lemma 3. Let AE denote the accuracy of the Exponential
mechanism, and AO denote the upper bound on the accuracy
that can be achieved by any privacy-preserving algorithm.
Then, for utility functions based on the number of common
neighbors between two nodes, AE
k+1 , where k is the
AO
number of nodes with non-zero utility.
≥ 1
Furthermore,
Lemma 4. For utility vector of the form
≥ k
u = (umax, . . . , umax, 0, . . . , 0), AE
AO
number of nodes with non-zero utility,
k+1 , where k is the
For real-world graphs, we expect the number of nodes with
non-zero utility k to be fairly small, and thus, the Expo-
Figure 3: Performance of the Exponential mechanisms
and predicted by theoretical bound on Wikipedia vote
network using # of common neighbors as a measure of
utility. The x-axis represents the accuracy, the y-axis -
the expected percentage of nodes for whom that accu-
racy of recommendation is achieved (or predicted by the
theoretical bound). The top graph is for desired privacy
guarantee of = 0.1, the bottom - for = 0.5.
nential mechanism to achieve a good approximation to the
best possible accuracy achievable by a privacy-preserving so-
cial recommendation algorithm. Furthermore, observe that
Corollary 1 merely gives an upper bound on accuracy achiev-
able in a privacy-preserving manner, but it might be the case
that tighter lower bounds can be obtained. Hence, in many
ways, the Exponential and Laplace mechanisms are repre-
sentative of the class of good privacy-preserving mechanisms
one can hope for.
8. EXTENSIONS AND FUTURE WORK
8.1 Vertex privacy and non-monotone algo-
rithms
We considered the setting of graph based social recom-
mendations where we wished to maintain private the infor-
mation about the presence or absence of an edge in the graph
but our reasoning and results can easily be generalized to a
setting where we would like to protect the entire identity of
a node. To achieve that, one would need to strengthen the
definition of the recommendation algorithm satisfying dif-
ferential privacy to consider graphs that differ in one node,
rather than one edge, and adjust the value of t, the number
of edge alterations to turn a node from the low utility group
into a node of maximum utility, respectively.
Furthermore, our results can be generalized to social rec-
ommendation algorithms that do not satisfy the monotonic-
ity property. For clarity of exposition, we omit the exact
statements and proofs of lemmas analogous to Lemmas 1
Figure 4: Accuracy achieved by Exponential mechanism
and predicted by Theoretical Bound as a function of node
degree, = 0.5
and 2 but remark that the statement formulations and our
qualitative conclusions will remain essentially unchanged,
with the exception of the meaning of variable t. Without
the monotonicity property, t would correspond to the num-
ber of edge alterations necessary to exchange the node with
the smallest probability of being recommended and the node
with the highest utility, rather than to the number of edge
alterations necessary to make the node with the smallest
probability of being recommended into the node with the
highest utility, leading to a higher value for t.
8.2 What if utility vectors are unknown?
Both the differentially private algorithms we considered
in Section 6 assume the knowledge of the entire utility vec-
tor. This assumption cannot be made in social networks for
various reasons. Firstly, computing as well as storing the
utility of n2 pairs is prohibitively expensive, when dealing
with graphs of several hundred million nodes. Secondly, even
if one could compute and store them, these graphs change
at staggering rates, therefore, utility vectors are also con-
stantly changing. We believe that this is a very important
and interesting problem. In this section, we explore a simple
algorithm that assumes no knowledge of the utility vector;
it only assumes that sampling from the utility vector can be
done efficiently.
8.2.1
Suppose we are given an algorithm A which is a γ ap-
proximation in terms of utility, and not provably private.
We show how to modify the algorithm A to guarantee dif-
ferential privacy, while still preserving, to some extent, the
utility approximation of A. The proof of the following the-
orem, and a note, are placed in the appendix.
Sampling and Linear Smoothing
Definition 6. Given algorithm A = (p1, . . . , pi, . . . , pn),
n +
algorithm AS(x) recommends node i with probability 1−x
xpi, where 0 ≤ x ≤ 1 is a parameter.
Theorem 7. AS(x) guarantees ln(1 + nx
1−x )-differential
privacy and a xγ approximation of utility.
Another idea worth exploring is perturbing the input graph
(by adding/deleting a fraction of possible edges) and then
sampling and recommending from it. What is the rela-
tionship between the extent of perturbation and the util-
ity/privacy guarantees?
8.3 Future Work
Several interesting questions remain unexplored in this
work. While we have considered some specific utility func-
tions in this paper, it would be nice to look more. Further,
our motivation was to look at the most stringent require-
ment in terms of privacy; however, a natural question is
to understand utility-privacy trade-offs for certain typical
graphs that arise in social networks.
This paper only considers lower bounds and algorithms
for making one single recommendation.
It would be very
interesting, and important, to explore how the effect on pri-
vacy compounds with multiple recommendations. Further,
some edges can be more sensitive than others. Perhaps the
solution should be methodological - enable opt-in/opt-out
settings to specify which nodes/edges are private. A closer
look at such dependences is required.
Also, most works on making recommendations deal with
static databases. Social networks clearly change over time
(and rather rapidly). This raises several issues, such as
not being able to assume the utility vector is known, sensi-
tivity changing, privacy impacts of dynamic databases etc.
Dealing with such temporal graphs and understanding there
trade-offs would be very interesting.
Finally, it would certainly be interesting to extend these
results for weaker notions of privacy than differential pri-
vacy. For instance, some privacy notions previously defined
include k-anonymity, (, δ)-differential privacy, and relaxing
the adversary's background knowledge to just the general
statistics of the graph.
9. ACKNOWLEDGMENTS
The authors are grateful to Arpita Ghosh and Tim Rough-
garden for thought-provoking discussions.
10. REFERENCES
[1] E. Aımeur, G. Brassard, J. M. Fernandez, and F. S.
Mani Onana. Alambic: a privacy-preserving
recommender system for electronic commerce. Int. J.
Inf. Secur., 7(5):307 -- 334, 2008.
[2] R. Andersen, C. Borgs, J. T. Chayes, U. Feige, A. D.
Flaxman, A. Kalai, V. S. Mirrokni, and
M. Tennenholtz. Trust-based recommendation
systems: an axiomatic approach. In WWW, pages
199 -- 208, 2008.
[3] B. Barak, K. Chaudhuri, C. Dwork, S. Kale,
F. McSherry, and K. Talwar. Privacy, accuracy and
consistency too: A holistic solution to contingency
table release. In PODS, 2007.
[4] R. Bhaskar, S. Laxman, A. Smith, and A. Thakurta.
Personal communication, 2010.
[5] L. Calandrino, A. Narayanan, E. Felten, and
V. Shmatikov. Don't review that book: Privacy risks
of collaborative filtering. Manuscript, 2009.
[6] C. Dwork. Differential privacy. In ICALP, 2006.
[7] C. Dwork. Differential privacy: A survey of results. In
TAMC, pages 1 -- 19, 2008.
[8] C. Dwork, F. McSherry, K. Nissim, and A. Smith.
Calibrating noise to sensitivity in private data
analysis. In TCC, pages 265 -- 284, 2006.
[9] C. Dwork and A. Smith. Differential privacy for
statistics: What we know and what we want to learn.
In NCHS/CDC Data Confidentiality Workshop, 2008.
[10] J. Golbeck. Generating predictive movie
recommendations from trust in social networks. In
iTrust, pages 93 -- 104, 2006.
[11] J. Leskovec, D. Huttenlocher, and J.Kleinberg.
Predicting positive and negative links in online social
networks. In WWW, 2010.
[12] H. Ma, I. King, and M. R. Lyu. Learning to
recommend with social trust ensemble. In SIGIR,
pages 203 -- 210, 2009.
[13] A. Machanavajjhala, D. Kifer, J. Abowd, J. Gehrke,
and L. Vihuber. Privacy: From theory to practice on
the map. In ICDE, 2008.
[14] F. McSherry and I. Mironov. Differentially private
recommender systems: building privacy into the net.
In KDD, pages 627 -- 636, 2009.
[15] F. McSherry and K. Talwar. Mechanism design via
differential privacy. In FOCS, pages 94 -- 103, 2007.
[16] A. Mislove, K. P. Gummadi, and P. Druschel.
Exploiting social networks for internet search. In
Proceedings of the 5th Workshop on Hot Topics in
Networks (HotNets'06), November 2006.
[17] A. Mislove, A. Post, K. P. Gummadi, and P. Druschel.
Ostra: Leverging trust to thwart unwanted
communication. In Proceedings of the 5th Symposium
on Networked Systems Design and Implementation
(NSDI'08), April 2008.
[18] M. Montaner, B. L´opez, and J. L. d. l. Rosa.
Opinion-based filtering through trust. In CIA '02:
Proceedings of the 6th International Workshop on
Cooperative Information Agents VI, pages 164 -- 178,
London, UK, 2002. Springer-Verlag.
[19] S. Nadarajah and S. Kotz. On the linear combination
of laplace random variables. Probab. Eng. Inf. Sci.,
19(4):463 -- 470, 2005.
[20] K. Nissim. Private data analysis via output
perturbation. In Privacy-Preserving Data Mining:
Models and Algorithms, pages 383 -- 414. Springer, 2008.
[21] D. L. Nowell and J. Kleinberg. The link prediction
problem for social networks. In CIKM, pages 556 -- 559,
2003.
[22] J. Peterson, Arthur V. and R. A. Kronmal. On
mixture methods for the computer generation of
random variables. The American Statistician,
36(3):184 -- 191, 1982.
[23] G. Swamynathan, C. Wilson, B. Boe, K. Almeroth,
and B. Y. Zhao. Do social networks improve
e-commerce?: a study on social marketplaces. In
WOSP '08: Proceedings of the first workshop on
Online social networks, pages 1 -- 6, New York, NY,
USA, 2008. ACM.
[24] V. Toubiana, A. Narayanan, D. Boneh,
H. Nissenbaum, and S. Barocas. Adnostic: Privacy
preserving targeted advertising. In NDSS, 2010.
[25] C.-N. Ziegler and G. Lausen. Analyzing correlation
between trust and user similarity in online
communities. In Proceedings of Second International
Conference on Trust Management, pages 251 -- 265.
Springer-Verlag, 2004.
Or (for small enough β)
APPENDIX
Proof of Lemma 1
Proof. We initiate the analysis with a simple claim.
Claim 1. In order to achieve (1 − δ) accuracy, at least
c−δ
c of the probability weight has to go to nodes in the high
utility group, so there exists a node x in the low utility group
of G1 that is recommended with probability of at most
c(n−k) ,
e.g. pG1
x ≤ δ
δ
c(n−k) .
p+umax + (1 − c)umaxp− ≥(cid:80)
Proof. Denote by p+ and p− the total probability that
goes to high/low utility nodes, respectively, and observe that
i uipi ≥ (1 − δ)umax and p+ +
p− ≤ 1, hence, p+ > c−δ
c , p− ≤ δ
c .
We now continue the proof of Lemma 1.
Let G2 be the graph that turns x, found according to the
Claim above, into a node of highest utility by addition of t
edges.
By differential privacy, we have p
p
In order to achieve (1 − δ) accuracy on G2, at least c−δ
of the probability weight has to go to nodes in the high
utility group, and hence by monotonicity P r[xG2] > c−δ
c(k+1) .
Combining the previous three inequalities, we obtain:
c
≤ et.
G2
x
G1
x
(c−δ)(n−k)
(k+1)δ =
c−δ
c(k+1)
c(n−k)
δ
< p
p
≤ et, hence
G2
x
G1
x
(cid:0)ln(
≥ 1
t
c − δ
δ
) + ln(
n − k
k + 1
)(cid:1)
This completes the proof.
Proof of Lemma 2
(cid:16)
1 − 1
(cid:17)
Proof. We first use the concentration axiom to prove
the following claim.
Claim 2. If c =
, then k = O(β log n) where
log n
(cid:17)
β is the parameter of the concentration axiom.
(cid:16)
log n . Let the total utility mass be U =(cid:80)
Proof. Now consider the case when c =
Therefore, k is the number of nodes that have utility at
least umax
i ui. Since
by concentration, the β highest utility nodes add up to a
total utility mass of Ω(1)∗ U , we have umax ≥ Ω( U
β ). There-
fore, k, the number of nodes with utility at least umax
log n is at
most U log n
umax
which is at most O(β log n).
1 − 1
log n
.
We now prove the Lemma using Lemma 1 and Claim 2.
Substituting these in the expression, if we need 1− c(n−k)
to be Ω(1), then require (k + 1)et to be Ω(n − k). (Notice
that if (k + 1)et = o(n − k), then
n−k+(k+1)et ≥ c − o(1),
which is 1 − o(1).).
Therefore, if we want an algorithm to obtain constant
approximation in utility, i.e. (1 − δ) = Ω(1), then we need
the following (assuming β to be small):
c(n−k)
n−k+(k+1)et
(O(β log n))et = Ω((n − O(β log n))
Simplifying
et = Ω(
n
β log n
)
≥ log n − log β − log log n
t
≥ log n − o(log n)
t
Proof of Theorem 2
Proof. Lower Bound for Common Neighbors We
formalize the intuition in terms of an upper bound on t in
the following claim.
Claim 3. For common neighbors based utility functions,
when recommendations for r are being made, we have t ≤
dr + 2, where dr is the degree of node r.
Proof. Observe that if the number of common neighbors
is the measure of the utility of recommendation, then one
can make any zero utility node, say x, for source node r into
a max utility node by adding dr edges to all of r's neighbors
and additionally adding two more edges (one each from r
and x) to some node with small utility. This is because the
highest utility node has at most dr common neighbors with r
(one of which could potentially be x). Further, adding these
edges cannot increase the number of common neighbors for
any other node beyond dr.
We now use this to get the theorem immediately by re-
placing t in the expression stated previously.
Proof of Theorem 3
Proof. Lower Bound for Sum of Weighted Paths
The number of paths of length l between two nodes is at
most dl−1
max. Let x be the highest utility node and let y be the
node we wish to make the highest utility node after adding
certain edges. If we are making recommendations for node
r, then the maximum number of common neighbors with r
is at most dr.
ux ≤ γdr
term as well.)
Currently denote the utility of x by ux. We know that
max. (In fact one can tighten the second
We rewire the graph as follows. Any (c−1)dr nodes (other
than y and the source node r) are picked; here c > 1 is to
be determined later. Both r and y are connected to these
(c − 1)dr nodes. Additionally, y is connected to all of r's dr
neighbors. Therefore, we now get the following.
(cid:80)inf
l=3 γl−1dl−1
uy ≥ γcdr
Now we wish to bound by above the utility of any other
node in the network in this rewired graph. Notice that every
other node still has at most dr paths of length 2 with the
source. Further, there are only two nodes in the graph that
have degree more than dmax + 1, and they have degree at
most (c + 1)dmax. Therefore, the number of paths of length l
for l ≥ 3 for any node is at most ((c+1)dmax)2·(dmax +1)l−3.
This can be further tightened to ((c + 1)dmax)2 · (dmax)l−3.
We therefore get the following for any x in the rewired graph,
Let p(cid:48)(cid:48)
i = 1−x
n + xpi. We have
(cid:48)(cid:48)
n
+ x,
≤ p
1 − x
n
i ≤ 1 − x
n(cid:88)
k uk = 1 and(cid:80) pkuk = γ.
1 − x
n
n(cid:88)
(
k=1
(cid:48)(cid:48)
k =
)uk+
k=1
The utility of AS is
since 0 ≤ pi ≤ 1.
n(cid:88)
where we use(cid:80)
U (AS) =
k=1
ukp
xpkuk =
1 − x
n
+xγ ≥ xγ,
For the privacy guarantee, note again that the upper and
lower bounds on p(cid:48)
i hold for any graph and utility function.
Therefore, the change in the probability of recommending i
for any two graphs G and G(cid:48) that differ in exactly one edge
is at most
pi(G)
pi(G(cid:48))
≤ x + 1−x
n
1−x
n
= 1 +
nx
1 − x
.
Therefore, AS is ln(1+ nx
plete the proof.
1−x )-differentially private. This com-
Further, note, to guarantee 2- differentially privacy for
1−x ) =
AS(x), we need to set the parameter x so that ln(1 + nx
2c ln n (rewriting = c ln n), namely
n2c − 1
x =
n2c − 1 + n
.
The algorithm AS guarantees a utility of at least xγ.
Proof of Lemma 3
Proof. Suppose the variations on the common neighbor
functions permitted are ui = di/z, where di is the number of
common neighbors node i has with the target node, and z is
a scaling constant. Pick c = 1, meaning that all nodes except
k have zero utility. Then UO ≤ umax(1 − δ) ≤ umax(1 −
n−k+(k+1)eumax ) = umax
n−k+(k+1)eumax .
(k+1)eumax
n−k
ezumax
Under our restricted privacy definition, the sensitivity of
the scaled number of common neighbors utility function is 1
z .
UE ≥ umax
eumax
Hence UE
k+1 and exponential algorithm gives a (k + 1)
UO
approximation of utility, which could be a fairly good ap-
proximation, if k is small compared to n, which is what we
expect, in real-world social network graphs.
n−k+kezumax = umax
≥ 1
n−k+keumax ≥ umax
eumax
n−k+(k+1)eumax .
ux ≤ γdr + (c + 1)2
∞(cid:88)
l=3
γl−1dl−1
max
Now consider the case where γ < 1
dmax
. We get
ux ≤ γdr +
(c + 1)2γ2d2
1 − γdmax
We now want uy ≥ ux. This reduces to
max
(c − 1) ≥ (c + 1)2γdmax
1 − γdmax
1
Now if γ = o(
) then it is sufficient to have (c − 1) =
Ω(γdmax) which can be achieved even with c = 1 + o(1).
Now notice that we only added dr + 2(c − 1)dr edges to the
graph. This completes the proof of the theorem.
dmax
1
Comment on relationship between common neigh-
bors and weighted paths: Since common neighbors is
an extreme case of weighted paths (as γ → 0), we are able
to obtain the same lower bound (up to o(1) terms) when
γ is made small (in particular, γ ≈ o(
). Can one ob-
tain (perhaps weaker) lower bounds when say γ = Θ(
)?
Notice that the proof only needs (c − 1) ≥ (c+1)2γdmax
. We
1−γdmax
then get a lower bound of ≥ 1
2c−1 ) where dr = α log n.
Setting γdmax = s, for some constant s, we can find the
smallest c that satisfies the expression (c − 1) ≥ (c+1)2s
.
1−s
Notice that this does give a nontrivial lower bound (i.e. a
lower bound tighter than the generic one presented in the
previous section), as long as s is a sufficiently small constant.
α ( 1−o(1)
dmax
dmax
1
Proof of Theorem 6
Proof. Utility of Laplace for n = 2:
Suppose we
have two elements, with utility t1 and t2, respectively, where
t1 ≥ t2 wlog.
(cid:90) ∞
−∞
Let φX (t) denote the characteristic function of the Laplace
distribution, it is known that φX (t) = 1
1+b2t2 . Moreover, it
is known that if X1 and X2 are independently distributed
random variables, then φX1+X2 (t) = φX1 (t)φX2 (t) =
Using the inversion formula, we can compute the pdf of
X = X1 + X2 as follows:
1
(1+b2t2)2 .
fX (x) = F
(cid:48)
X (x) =
1
2π
−itxφX (t)dt
e
4 e−x( 2
For x > 0, the pdf of X1 + X2 is fX (x) = 1
the cdf is FX (x) = 1 − 1
What is the probability that element 1 is recommended?
It's the P r[t1 + X1 > t2 + X2] = P r[X2 − X1 < t1 − t2] =
2 e−(t1−t2) − (t1−t2)
1 − 1
4 e−(t1−t2)( 2
4e(t1−t2 )
Hence, the Laplace mechanism recommends node 1 with
+ (t1 − t2)) = 1 − 1
4b (1 + x
+ x).
b and
b )e− x
probability
1 − 1
2
−(t1−t2) − (t1 − t2)
4e(t1−t2) ,
e
from which the desired statement about AL's utility fol-
lows.
Proof of Theorem 7
Proof. Sampling and Linear Smoothing
|
1610.00575 | 1 | 1610 | 2016-10-03T14:49:10 | Facility Leasing with Penalties | [
"cs.DS"
] | In this paper we study the facility leasing problem with penalties. We present a primal-dual algorithm which is a 3-approximation, based on the algorithm by Nagarajan and Williamson for the facility leasing problem and on the algorithm by Charikar et al. for the facility location problem with penalties. | cs.DS | cs |
Facility Leasing with Penalties
Murilo S. de Lima∗1, Mário C. San Felice† 2, and Orlando Lee‡ 1
1Institute of Computing, University of Campinas (Unicamp)
2Institute of Mathematics and Statistics, University of São Paulo (USP)
September 22, 2018
Abstract
In this paper we study the facility leasing problem with penalties. We present a primal-
dual algorithm which is a 3-approximation, based on the algorithm by Nagarajan and
Williamson for the facility leasing problem [NW13] and on the algorithm by Charikar et al.
for the facility location problem with penalties [CKMN01].
1 Introduction
In the facility location problem, one is given a set F of facilities, an opening cost for each
facility, a set D of clients and a metric distance function d between facilities and clients. The
objective is to choose a subset of the facilities to open and an assignment between clients and
facilities, so to minimize the cost of opening the facilities plus the sum of the distances be-
tween each client and the corresponding assigned facility. This is an NP-hard problem, and it
does not have a polynomial-time algorithm with approximation factor smaller than 1.463 un-
less P = NP [Svi02]. Currently the best approximation factor is 1.488, due to an algorithm by
Li [Li13].
In the facility location problem with penalties, we may not assign a client j to a facility if we
choose to pay a penalty πj. I.e., we must select a subset of the facilities to open, and a subset
of the clients we assign to open facilities; we are going to pay the penalties for the remaining
of the clients. The cost of a solution is, therefore, the cost of opening the selected facilities,
plus the distance between the client and its corresponding facility for each assigned client, plus
the penalty cost for each unassigned client. Clearly the facility location problem reduces to this
problem if we set πj = ∞ for every client j. Currently it is known a 1.5148-approximation
algorithm for this problem [LDXX15]; however, there is a simpler 3-approximation algorithm by
Charikar et al. [CKMN01]. Also, if the penalties obey a submodular function, then there is a
2-approximation algorithm [LDXX15].
In the facility leasing problem, client requests are distributed along the time, and instead
of opening facilities permanently, we may lease each facility for one of K different durations
δ1, . . . , δK . The cost for leasing a facility for δk units of time depends on the facility position,
as in the traditional facility location problem, but also on the leasing type k. Additionally, it
is reasonable to suppose that the leasing costs respect an economics of scale: the leasing cost
per unit of time decreases with the leasing duration, for a fixed facility location. A facility lease
may begin at any moment in the time. Then, we wish to select a set of facility leases that cover
the client requests and minimizes the leasing costs plus the distance between each client and
∗[email protected] Supported by FAPESP PhD Scholarship Process 2014/18781-1, and CNPq PhD Schol-
arship Process 142161/2014-4.
†[email protected] Supported by CAPES PNPD scholarship 1522390.
‡[email protected] Partially supported by Bolsa de Produtividade do CNPq Process 311373/2015-1, and
Edital Universal CNPq Process 477692/2012-5.
1
the facility lease that serves each of its requests. This problem was proposed by Nagarajan and
Williamson, who presented a simple 3-approximation primal-dual algorithm [NW13].
In this paper, we study the combination of the previous two problems, which we call the
facility leasing problem with penalties (PFLe). In this problem, facilities are leased in-
stead of permanently opened, as in the facility leasing problem, and some clients may be left
unassigned by paying for the penalty cost. We obtain a 3-approximation algorithm by combin-
ing the algorithm by Nagarajan and Williamson for the facility leasing problem [NW13] and the
algorithm by Charikar et al. for the facility location problem with penalties [CKMN01].
The leveraging scheme by Li et al. [LDXX15] implies that, for any covering problem with
an α-approximation algorithm, there is a (1 − e−1/α)−1-approximation algorithm for the corre-
sponding covering problem with submodular penalties. Combining this with the algorithm by
Nagarajan and Williamson (α = 3), one may obtain a 3.5277-approximation algorithm for the
facility leasing problem with submodular penalties. Note that our algorithm obtains a better
approximation ratio for the linear case.
2 Notation and Problem Definition
Let [K] := {1, . . . , K} be the set of lease types. We denote a facility lease by a triple
f = (pf , kf , tf ), where pf ∈ V is the point where f is located, kf ∈ [K] is the leasing type for f ,
and tf ∈ Z+ is the instant of time in which the lease for f begins. We write F := F × [K] × Z+
so to simplify our notation.
Similarly, we denote a client by a triple j = (pj , πj , tj), where pj ∈ V is the point where j
is located, πj ∈ R+ is the penalty for not assigning a facility lease to j, and tj is the instant in
which j arrives.
In order to simplify our notation, we write δf instead of δ(kf ), and γf instead of γ(pf , kf ),
for a facility lease f = (pf , kf , tf ) ∈ F. Also, for f = (pf , kf , tf ) ∈ F and j = (pj, πj , tj) ∈
V × R+ × Z+, we define the distance between j and f to be
d(j, f ) := (cid:26) d(pj , pf )
∞
if tj ∈ [tf , tf + δf ),
otherwise.
I.e., the distance between client j and facility lease f is infinity if the facility lease does not
cover tj.
Problem PFLe(V, d, F, K, γ, δ, D): The input consists of a set of points V , a distance
function d : V × V 7→ R+ between the points of V satisfying symmetry and triangle
inequality, a set F ⊆ V of potential facilities, an integer K > 0 that represents the
number of lease types, a cost γ(p, k) ∈ R+ for leasing facility p ∈ F with leasing type
k ∈ [K], a function δ : [K] 7→ N that maps each lease type to a length in days, and a
set D ⊆ V × R+ × Z+ of clients in the form j = (pj , πj , tj). The goal is to find a set
X ⊆ F := F × [K] × Z+ of facility leases in the form f = (pf , kf , tf ), and a function
a : D 7→ X ∪ {null} that maps each client j to an active facility leasing f ∈ X such that
tj ∈ [tf , tf + δf ) or to null, so to minimize
Xf ∈X
γf + Xj∈D:a(j)6=null
d(j, a(j)) + Xj∈D:a(j)=null
πj.
2
3 Primal-Dual Formulation
• Primal:
minimize Pf ∈F γf · yf +Pj∈DPf ∈F d(j, f ) · xjf +Pj∈D πj · zj
subject to
xjf ≤ yf
xf j + zj ≥ 1
∀f ∈ F, j ∈ D
∀j ∈ D
P f ∈F
tj ∈[tf ,tf +δf )
xf j , yf , zj ∈ {0, 1} ∀f ∈ F, j ∈ D
(Variable yf indicates whether facility f was leased, variable xjf indicates whether client j
was served by facility lease f , and variable zj indicates whether the algorithm decided to
pay the penalty associated with not serving j with a facility lease.)
• Dual relaxation:
maximize Pj∈D αj
subject to Pj∈D(αj − d(j, f ))+ ≤ γf ∀f ∈ F
αj ≤ πj ∀j ∈ D
αj ≥ 0
∀j ∈ D
(Economical interpretation: each client j is willing to pay αj to connect itself to some
facility lease. Part of this value covers the distance to the facility; the other part is a
contribution to pay for leasing the facility. However, the client is not willing to pay more
than its penalty.)
4 Algorithm
Our algorithm is based on the algorithm by Nagarajan and Williamson for the facility leasing
problem [NW13], and on the algorithm by Charikar et al. for the facility location problem with
penalties [CKMN01]. We say that a client j reaches a facility lease f if αj ≥ d(j, f ).
or
or
(a) αj = d(j, f ) for some j ∈ S and f ∈ X
set αj ← 0 for every j ∈ D
increase αj uniformly for every j ∈ S until
Algorithm Primal-DualPFLe(V, d, F, K, γ, δ, D)
01
02 X ← ∅, S ← D
03 while S 6= ∅ do
04
05
06
07
08
09
10
11
12
13
14
15
16
(c) αj = πj for some j ∈ S
build the graph GX with
X ← X ∪ {f ∈ F \ X : f satisfies (b)}
S ← S \ {j ∈ S : αj ≥ πj or j reaches some f ∈ X}
(b) γf = Pj∈D(αj − d(j, f ))+ for some f ∈ F \ X
V [GX ] ← X
E[GX ] ← {(f, f ′) : ∃j ∈ D : j reaches both f and f ′}
build a maximal independent set X ′ in GX greedily in decreasing order of δ
X ← {(pf , kf , tf − δk), f, (pf , kf , tf + δk) : f ∈ X ′}
3
17
18
19
20
21
22
for every j ∈ D do
if j reaches some f ∈ X then
a(j) ← arg minf ′∈ X{d(j, f ′)}
else
a(j) ← null
return ( X, a)
The algorithm maintains a dual variable αj for each client j ∈ D, a set X of temporarily
leased facilities, and a set S of the clients whose dual variable still is being increased, which is
initialized with the whole set of clients D. The increasing pauses when either: (a) a client reaches
an already temporarily leased facility, (b) the sum of the contributions towards a facility lease
pays for its cost or (c) the dual variable reaches the penalty cost for some client. We then add
to X the facilities that reach condition (b). Also, we remove from S the clients that reach some
temporarily leased facility or whose dual variable pays for the penalty cost, and then proceed
the increasing of the remaining dual variables until S becomes empty.
After that initial phase, we build an interference graph GX between the facility leases in X.
Graph GX has vertex set X and has an edge between facilities f and f ′ if there is some client
that reaches both f and f ′. Then, we order set X by decreasing order of lease duration and build
a maximal independent set X ′ in a greedily manner; i.e., we visit set X in that order and add a
facility f to X ′ if there is no other facility lease f ′ ∈ X ′ reached by some client that reaches f .
Thus X ′ satisfies the following properties:
1. Every client reaches at most one facility lease in X ′;
2. If facility leases f and f ′ in X are reached by the same client j, and if f ′ ∈ X ′, then
δf ≤ δf ′.
However, note that there may be some client j that reaches some f in X but is not covered
by any facility lease in X ′. But then remember that some f ′ ∈ X ′ shares a reaching client j′
with f , thus δf ≤ δf ′ and the intervals covered by facility leases f and f ′ overlap. Then, since
we buy X, which has three copies of f ′, beginning at instants tf ′ − δf ′, tf ′ and tf ′ + δf ′ , we have
that the interval formed by those three facilities, which is [tf ′ − δf ′ , tf ′ + 2δf ′), is a superset of
interval [tf , tf + δf ), and therefore one of them covers tj.
Finally, if some client j does not reach any facility lease in X, then its dual variable pays for
its penalty and we set a(j) to null.
Also, note that, although the number of potential facility leases is infinite, the algorithm
may be implemented in finite time, which is also polynomial in the input size: it is enough to
consider, for every facility point, a lease beginning at each instant in which we have a client
request.
5 Analysis
In this section we analyze the approximation factor of algorithm Primal-DualPFLe.
First note that, since the conditions (a), (b) and (c) correspond to constraints of the relax-
ation of the dual program, we have that α is a feasible dual solution. Therefore, by weak duality,
we have that
αj ≤ opt(V, d, F, K, γ, δ, D).
Xj∈D
We will show, then, that the cost of the primal solution ( X, a) returned by the algorithm is
at most 3 times the cost of the dual solution, and thus our algorithm is a 3-approximation to
problem PFLe.
4
For every client j ∈ D, we define numbers αC
j , αF
j , and αP
j
in the following manner:
1. If j reaches some f ∈ X ′, then let
αC
j := d(j, f ), αF
j := αj − d(f, j), αP
j := 0;
2. If j does not reach any facility lease in X ′ but reaches some f ∈ X, then we let
3. Finally, if j does not reach any facility lease in X, then we let
αC
j := αj , αF
j := 0, αP
j := 0;
αC
j := 0, αF
j := 0, αP
j := αj .
Note that, either case, we have that
αj = αC
j + αF
j + αP
j .
Now, first let us bound the facility leasing cost. Note that, by construction, we have that, for
every f ∈ X ′, every client that reaches f reaches only f in X ′. Also, by case (b) of the algorithm,
the leasing cost of f is totally paid by contributions from clients that reach f . Therefore, we
have that
Xf ∈X ′
γf = Xj∈D
αF
j .
Since X, which is the set of facility leases actually bought by the algorithm, consists of three
copies of each facility lease in X ′, we have that
Xf ∈ X
γf ≤ 3 · Xj∈D
αF
j .
Now we bound the penalty cost. We have that a client j has a(j) set to null if and only if it
j . Also, due to case (c) of the algorithm,
does not reach any facility lease in X, and then αj = αP
we have that αj = πj. Thus, it is straightforward to conclude that
Xj∈D:a(j)=null
πj = Xj∈D
αP
j .
Finally, we have to bound the client connection cost. Let DX ′ be the set of clients that reach
some facility in X ′. Note that those clients are connected to the closest facility lease in X. Since
every such client j reaches some f ∈ X ′, we have that
d(j, a(j)) ≤ d(j, f ) = αC
j .
Now let j be some client that reaches some f ∈ X but does not reach any facility lease in X ′.
There must be some f ′ ∈ X ′ and some j′ that reaches both f and f ′, by construction of X ′. But
then we must have that αj ≥ αj′, since when αj′ stopped increasing it reached both f and f ′,
and αj reached f when it stopped increasing. Then, since j′ reaches both f and f ′, we have that
αj′ ≥ d(j′, f ) and αj′ ≥ d(j′, f ′).
Since one of the three copies of f ′ in X must cover tj, by triangle inequality, we have that
d(j, a(j)) ≤ d(j, f ′) ≤ d(j, f ) + d(j′, f ) + d(j′, f ′) ≤ αj + αj′ + αj′ ≤ 3 · αj = 3 · αC
j .
Summing up the previous inequalities, we have that
Xf ∈ X
γf + Xj∈D:a(j)6=null
d(j, a(j)) + Xj∈D:a(j)=null
πj ≤ 3 · Xj∈D
and we conclude the following theorem.
αj ≤ 3 · opt(V, d, F, K, γ, δ, D),
Theorem 1: Algorithm Primal-DualPFLe is a 3-approximation.
5
References
[CKMN01] M. Charikar, S. Khuller, D. M. Mount, and G. Narasimhan. Algorithms for facility
location problems with outliers. In SODA'01: Proceedings of the 12th Annual ACM-
SIAM Symposium on Discrete Algorithms, pages 642 -- 651, 2001.
[LDXX15] Y. Li, D. Du, N. Xiu, and D. Xu. Improved approximation algorithms for the facility
location problems with linear/submodular penalties. Algorithmica, 73(2):460 -- 482,
2015.
[Li13]
[NW13]
[Svi02]
S. Li. A 1.488 approximation algorithm for the uncapacitated facility location prob-
lem. Information and Computation, 222:45 -- 58, 2013.
C. Nagarajan and D. P. Williamson. Offline and online facility leasing. Discrete
Optimization, 10(4):361 -- 370, 2013.
M. Sviridenko. An improved approximation algorithm for the metric uncapacitated
facility location problem. In W. J. Cook and A. S. Schulz, editors, Integer Program-
ming and Combinatorial Optimization, volume 2337 of Lecture Notes in Computer
Science, pages 240 -- 257. Springer Berlin Heidelberg, 2002.
6
|
1909.12441 | 1 | 1909 | 2019-09-27T00:02:57 | Total Least Squares Regression in Input Sparsity Time | [
"cs.DS",
"cs.LG",
"stat.ML"
] | In the total least squares problem, one is given an $m \times n$ matrix $A$, and an $m \times d$ matrix $B$, and one seeks to "correct" both $A$ and $B$, obtaining matrices $\hat{A}$ and $\hat{B}$, so that there exists an $X$ satisfying the equation $\hat{A}X = \hat{B}$. Typically the problem is overconstrained, meaning that $m \gg \max(n,d)$. The cost of the solution $\hat{A}, \hat{B}$ is given by $\|A-\hat{A}\|_F^2 + \|B - \hat{B}\|_F^2$. We give an algorithm for finding a solution $X$ to the linear system $\hat{A}X=\hat{B}$ for which the cost $\|A-\hat{A}\|_F^2 + \|B-\hat{B}\|_F^2$ is at most a multiplicative $(1+\epsilon)$ factor times the optimal cost, up to an additive error $\eta$ that may be an arbitrarily small function of $n$. Importantly, our running time is $\tilde{O}( \mathrm{nnz}(A) + \mathrm{nnz}(B) ) + \mathrm{poly}(n/\epsilon) \cdot d$, where for a matrix $C$, $\mathrm{nnz}(C)$ denotes its number of non-zero entries. Importantly, our running time does not directly depend on the large parameter $m$. As total least squares regression is known to be solvable via low rank approximation, a natural approach is to invoke fast algorithms for approximate low rank approximation, obtaining matrices $\hat{A}$ and $\hat{B}$ from this low rank approximation, and then solving for $X$ so that $\hat{A}X = \hat{B}$. However, existing algorithms do not apply since in total least squares the rank of the low rank approximation needs to be $n$, and so the running time of known methods would be at least $mn^2$. In contrast, we are able to achieve a much faster running time for finding $X$ by never explicitly forming the equation $\hat{A} X = \hat{B}$, but instead solving for an $X$ which is a solution to an implicit such equation. Finally, we generalize our algorithm to the total least squares problem with regularization. | cs.DS | cs | Total Least Squares Regression in Input Sparsity Time∗
Huaian Diao† Zhao Song‡ David P. Woodruff§ Xin Yang¶
Abstract
F +kB − bBk2
F . We
In the total least squares problem, one is given an m × n matrix A, and an m × d matrix
B, and one seeks to "correct" both A and B, obtaining matrices bA and bB, so that there exists
an X satisfying the equation bAX = bB. Typically the problem is overconstrained, meaning
that m ≫ max(n, d). The cost of the solution bA, bB is given by kA − bAk2
give an algorithm for finding a solution X to the linear system bAX = bB for which the cost
kA− bAk2
is eO(nnz(A) + nnz(B)) + poly(n/ǫ)· d, where for a matrix C, nnz(C) denotes its number of non-
F is at most a multiplicative (1 + ǫ) factor times the optimal cost, up to an
additive error η that may be an arbitrarily small function of n. Importantly, our running time
zero entries. Importantly, our running time does not directly depend on the large parameter
m. As total least squares regression is known to be solvable via low rank approximation, a
natural approach is to invoke fast algorithms for approximate low rank approximation, obtaining
F + kB − bBk2
matrices bA and bB from this low rank approximation, and then solving for X so that bAX = bB.
However, existing algorithms do not apply since in total least squares the rank of the low rank
approximation needs to be n, and so the running time of known methods would be at least
mn2. In contrast, we are able to achieve a much faster running time for finding X by never
explicitly forming the equation bAX = bB, but instead solving for an X which is a solution to an
implicit such equation. Finally, we generalize our algorithm to the total least squares problem
with regularization.
9
1
0
2
p
e
S
7
2
]
S
D
.
s
c
[
1
v
1
4
4
2
1
.
9
0
9
1
:
v
i
X
r
a
∗A preliminary version of this paper appeared in NeurIPS 2019.
†[email protected]. Northeast Normal University.
‡[email protected]. University of Washington. This work was partly done while Zhao Song was visiting the Simons
Institute for the Theory of Computing.
§[email protected]. Carnegie Mellon University. David Woodruff would like to thank support from the Office
of Naval Research (ONR) grant N00014-18-1-2562. This work was also partly done while David Woodruff was visiting
the Simons Institute for the Theory of Computing.
¶[email protected]. University of Washington.
1 Introduction
In the least squares regression problem, we are given an m × n matrix A and an m × 1 vector b,
and we seek to find an x ∈ Rn which minimizes kAx − bk2
2. A natural geometric interpretation
is that there is an unknown hyperplane in Rn+1, specified by the normal vector x, for which we
have m points on this hyperplane, the i-th of which is given by (Ai,hAi, xi), where Ai is the i-th
row of A. However, due to noisy observations, we do not see the points (Ai,hAi, xi), but rather
only see the point (Ai, bi), and we seek to find the hyperplane which best fits these points, where
we measure (squared) distance only on the (n + 1)-st coordinate. This naturally generalizes to
the setting in which B is an m × d matrix, and the true points have the form (Ai, AiX) for some
unknown n × d matrix X. This setting is called multiple-response regression, in which one seeks to
find X to minimize kAX − Bk2
F is its squared Frobenius norm, i.e.,
the sum of squares of each of its entries. This geometrically corresponds to the setting when the
points live in a lower n-dimensional flat of Rn+d, rather than in a hyperplane.
F , where for a matrix Y , kY k2
While extremely useful, in some settings the above regression model may not be entirely realistic.
For example, it is quite natural that the matrix A may also have been corrupted by measurement
noise. In this case, one should also be allowed to first change entries of A, obtaining a new m × n
matrix bA, then try to fit B to bA by solving a multiple-response regression problem. One should again
be penalized for how much one changes the entries of A, and this leads to a popular formulation
F + kbAX − Bk2
known as the total least squares optimization problem min bA,X kA − bAk2
F . Letting
C = [A, B], one can more compactly write this objective as min bC=[ bA, bB] kC − bCk2
F , where it is
required that the columns of bB are in the column span of bA. Total least squares can naturally
capture many scenarios that least squares cannot. For example, imagine a column of B is a large
multiple λ · a of a column a of A that has been corrupted and sent to 0. Then in least squares, one
needs to pay λ2kak2
2, but in total least squares one can "repair A" to contain the column a, and
just pay kak2
2. We refer the reader to [MVH07] for an overview of total least squares. There is also
a large amount of work on total least squares with regularization [RG04, LPT09, LV14].
Notice that bC has rank n, and therefore the optimal cost is at least kC − Cnk2
that the last d columns are in the column span of the first n columns, then the optimal solution bC
to total least squares problem is equal to Cn, and so the total least squares cost is the cost of the
best rank-n approximation to C. In this case, and only in this case, there is a closed-form solution.
However, in general, this need not be the case, and kC − Cnk2
F may be strictly smaller than the
total least squares cost. Fortunately, though, it cannot be much smaller, since one can take the first
n columns of Cn, and for each column that is not linearly independent of the remaining columns,
we can replace it with an arbitrarily small multiple of one of the last d columns of Cn which is not
in the span of the first n columns of Cn. Iterating this procedure, we find that there is a solution to
the total least squares problem which has cost which is arbitrarily close to kC − Cnk2
F . We describe
this procedure in more detail below.
F , where Cn is the
best rank-n approximation to C. If, in the optimal rank-n approximation Cn, one has the property
The above procedure of converting a best rank-n approximation to an arbitrarily close solution
to the total least squares problem can be done efficiently given Cn, so this shows one can get an
arbitrarily good approximation by computing a truncated singular value decompostion (SVD), which
is a standard way of solving for Cn in O(m(n+d)2) time. However, given the explosion of large-scale
datasets these days, this running time is often prohibitive, even for the simpler problem of multiple
response least squares regression. Motivated by this, an emerging body of literature has looked at
the sketch-and-solve paradigm, where one settles for randomized approximation algorithms which
run in much faster, often input sparsity time. Here by input-sparsity, we mean in time linear in the
1
number nnz(C) of non-zero entries of the input description C = [A, B]. By now, it is known, for
example, how to output a solution matrix X to multiple response least squares regression satisfying
kAX − Bk2
F , in nnz(A) + nnz(B) + poly(nd/ǫ) time. This algorithm
works for arbitrary input matrices A and B, and succeeds with high probability over the algorithm's
random coin tosses. For a survey of this and related results, we refer the reader to [Woo14].
F ≤ (1 + ǫ) minX ′ kAX′ − Bk2
Given the above characterization of total least squares as a low rank approximation problem, it is
natural to ask if one can directly apply sketch-and-solve techniques to solve it. Indeed, for low rank
approximation, it is known how to find a rank-k matrix bC for which kC − bCk2
F ≤ (1 + ǫ)kC − Ckk2
in time nnz(C) + m · k2/ǫ [ACW17], using the fastest known results. Here, recall, we assume
m ≥ n + d. From an approximation point of view, this is fine for the total least squares problem,
since this means after applying the procedure above to ensure the last d columns of bC are in the
span of the first n columns, and setting k = n in the low rank approximation problem, our cost will
be at most (1 + ǫ)kC − Cnk2
F + η, where η can be made an arbitrarily small function of n. Moreover,
the optimal total least squares cost is at least kC − Cnk2
F , so our cost is a (1 + ǫ)-relative error
approximation, up to an arbitarily small additive η.
F
Unfortunately, this approach is insufficient for total least squares, because in the total least
squares problem one sets k = n, and so the running time for approximate low rank approximation
becomes nnz(C)+m·n2/ǫ. Since n need not be that small, the m·n2 term is potentially prohibitively
large. Indeed, if d ≤ n, this may be much larger than the description of the input, which requires
at most mn parameters. Note that just outputting bC may take m · (n + d) parameters to describe.
However, as in the case of regression, one is often just interested in the matrix X or the hyperplane
x for ordinary least squares regression. Here the matrix X for total least squares can be described
using only nd parameters, and so one could hope for a much faster running time.
1.1 Our Contributions
Our main contribution is to develop a (1 + ǫ)-approximation to the total least squares regression
F ≤ (1 + ǫ)kC − Cnk2
not need to write such matrices down, as it is only interested in outputting the solution X to the
rank-n approximation to C, and η is an arbitrarily small function of n. Importantly, we achieve a
problem, returning a matrix X ∈ Rn×d for which there exist bA ∈ Rm×n and bB ∈ Rm×d for which
bAX = bB and kC − bCk2
F + η, where C = [A, B], bC = [bA, bB], Cn is the best
running time of eO(nnz(A) + nnz(B)) + poly(n/ǫ) · d.
Notice that this running time may be faster than the time it takes even to write down bA and
bB. Indeed, although one can write A and B down in nnz(A) + nnz(B) time, it could be that the
algorithm can only efficiently find an bA and a bB that are dense; nevertheless the algorithm does
equation bAX = bB. This is motivated by applications in which one wants generalization error. Given
X, and a future y ∈ Rn, one can compute yX to predict the remaining unknown d coordinates of
the extension of y to n + d dimensions.
Our algorithm is inspired by using dimensionality reduction techniques for low rank approx-
imation, such as fast oblivious "sketching" matrices, as well as leverage score sampling. The
rough idea is to quickly reduce the low rank approximation problem to a problem of the form
minrank - n Z∈Rd1×s1 k(D2CD1)Z(S1C) − D2CkF , where d1, s1 = O(n/ǫ), D2 and D1 are row and
column subset selection matrices, and S1 is a so-called CountSketch matrix, which is a fast oblivi-
ous projection matrix. We describe the matrices D1, D2, and S1 in more detail in the next section,
though the key takeaway message is that (D2CD1), (S1C), and (D2C) are each efficiently com-
putable small matrices with a number of non-zero entries no larger than that of C. Now the problem
is a small, rank-constrained regression problem for which there are closed form solutions for Z. We
then need additional technical work, of the form described above, in order to find an X ∈ Rn×d
2
sketching methods have not been applied to the total least squares problem before, and we consider
this application to be one of the main contributions of this paper.
given Z, and to ensure that X is the solution to an equation of the form bAX = bB. Surprisingly, fast
We carefully bound the running time at each step to achieve eO(nnz(A) + nnz(B) + poly(n/ǫ)d)
overall time, and prove its overall approximation ratio. Our main result is Theorem 3.10. We also
generalize the theorem to the important case of total least squares regression with regularization;
see Theorem 3.12 for a precise statement.
We empirically validate our algorithm on real and synthetic data sets. As expected, on a number
of datasets the total least squares error can be much smaller than the error of ordinary least squares
regression. We then implement our fast total least squares algorithm, and show it is roughly 20− 40
times faster than computing the exact solution to total least squares, while retaining 95% accuracy.
Notation. For a function f , we define eO(f ) to be f · logO(1)(f ). For vectors x, y ∈ Rn, let
hx, yi :=Pn
square matrix. Let kAkF denote the Frobenius norm of a matrix A, i.e., kAkF = (PiPj A2
i=1 xiyi denote the inner product of x and y. Let nnz(A) denote the number of nonzero
entries of A. Let det(A) denote the determinant of a square matrix A. Let A⊤ denote the transpose
of A. Let A† denote the Moore-Penrose pseudoinverse of A. Let A−1 denote the inverse of a full rank
i,j)1/2.
Sketching matrices play an important role in our algorithm. Their usefulness will be further
explained in Section 3. The reader can refer to Appendix B for detailed introduction.
2 Problem Formulation
We first give the precise definition of the exact (i.e., non-approximate) version of the total least
squares problem, and then define the approximate case.
Definition 2.1 (Exact total least squares). Given two matrices A ∈ Rm×n, B ∈ Rm×d, let C =
[A, B] ∈ Rm×(n+d). The goal is to solve the following minimization problem:
X∈Rn×d,∆A∈Rm×n,∆B∈Rm×d k[∆A, ∆B]kF
min
subject to (A + ∆A)X = (B + ∆B)
(1)
It is known that total least squares problem has a closed form solution. For a detailed discussion,
see Appendix E. It is natural to consider the approximate version of total least squares:
Definition 2.2 (Approximate total least squares problem). Given two matrices A ∈ Rm×n and
B ∈ Rm×d, let OPT = minrank −n C ′ kC′ − [A, B]kF , for parameters ǫ > 0, δ > 0. The goal is to
output X′ ∈ Rn×d so that there exists A′ ∈ Rm×n such that
k[A′, A′X′] − [A, B]kF ≤ (1 + ǫ) OPT +δ.
One could solve total least squares directly, but it is much slower than solving least squares (LS).
We will use fast randomized algorithms, the basis of which are sampling and sketching ideas [CW13,
NN13, MM13, Woo14, RSW16, PSW17, SWZ17, CCLY19, CLS19, LSZ19, SWY+19, SWZ19a,
SWZ19b, SWZ19c, DJS+19], to speed up solving total least squares in both theory and in practice.
The total least squares problem with regularization is also an important variant of this prob-
lem [LV10]. We consider the following version of the regularized total least squares problem.
Definition 2.3 (Approximate regularized total least squares problem). Given two matrices A ∈
Rm×n and B ∈ Rm×d and λ > 0, let OPT = minU∈Rm×n,V ∈Rn×(n+d) kU V − [A, B]k2
F +
λkV k2
F , for parameters ǫ > 0, δ > 0. The goal is to output X′ ∈ Rn×d so that there exist A′ ∈ Rm×n
F + λkUk2
3
Algorithm 1 Our Fast Total Least Squares Algorithm
1: procedure FastTotalLeastSquares(A, B, n, d, ǫ, δ)
2:
s1 ← O(n/ǫ), s2 ← O(n/ǫ), d1 ← eO(n/ǫ), d2 ← eO(n/ǫ)
Choose S1 ∈ Rs1×m to be a CountSketch matrix, then compute S1C
⊲ Definition B.1
⊲ Reduce n + d to O(n/ǫ)
if d > Ω(n/ǫ) then
Choose D⊤1 ∈ Rd1×(n+d) to be a leverage score sampling and rescaling matrix according
⊲ We do not need to use matrix D1
to the rows of (S1C)⊤, then compute CD1
⊲ Theorem 3.10
else
Choose D⊤1 ∈ R(n+d)×(n+d) to be the identity matrix
rows of CD1
Choose D2 ∈ Rd2×m to be a leverage score sampling and rescaling matrix according to the
Z2 ← minrank −n Z∈Rd1×s1 kD2CD1ZS1C − D2CkF
⊲ Theorem D.1
A, B, π ← Split(CD1, Z2, S1C, n, d, δ/ poly(m)), X ← minkAX − BkF
if Need CFTLS then
⊲ For experiments to evaluate the cost
⊲ Lemma 3.8
⊲ bC = CD1Z2S1C ; C = S2bC
⊲ bA = bC∗,[n], bB = bC∗,[n+d]\[n]; A = S2bA, B = S2bB
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
16:
17:
18:
19:
20:
21:
Evaluate(CD1, Z2, S1C, X, π, δ/ poly(m))
return X
13:
14: procedure Split(CD1, Z2, S1C, n, d, δ)
15:
Choose S2 ∈ Rs2×m to be a CountSketch matrix
C ← (S2 · CD1) · Z2 · S1C
A ← C∗,[n], B ← C∗,[n+d]\[n]
T ← ∅, π(i) = −1 for all i ∈ [n]
for i = 1 → n do
if A∗,i is linearly dependent of A∗,[n]\{i} then
π(i) ← j
return A, B, π
22:
23: procedure Evaluate(CD1, Z2, S1C, X, π, δ)
24:
bC ← CD1Z2S1C, bA ← bC∗,[n], bB ← C∗,[n+d]\[n]
for i = 1 → n do
if π(i) 6= −1 then
bA∗,i ← bA∗,i + δ · bB∗,π(i)
return k[bA, bAX] − CkF
25:
26:
27:
28:
j ← minj∈[d]\T{B∗,j is linearly independent of A}, A∗,i ← A∗,i +δ·B∗,j, T ← T ∪{j},
⊲ π : [n] → {−1} ∪ ([n + d]\[n])
⊲ Appendix F.9
, U′ ∈ Rm×n and V ′ ∈ Rn×(n+d) satisfying k[A′, A′X′]− U′V ′k2
(1 + ǫ) OPT +δ.
F ≤ δ and k[A′, A′X′]− [A, B]k2
F ≤
3 Fast Total Least Squares Algorithm
We present our algorithm in Algorithm 1 and give the analysis here. Readers can refer to Table 1
to check notations in Algorithm 1. To clearly give the intuition, we present a sequence of approxi-
mations, reducing the size of our problem step-by-step. We can focus on the case when d ≫ Ω(n/ǫ)
and the optimal solution bC to program (6) has the form [bA, bAbX] ∈ Rm×(n+d). For the other case
when d = O(n/ǫ), we do not need to use the sampling matrix D1. In the case when the solution
does not have the form [bA, bAbX], we need to include Split in the algorithm, since it will perturb
4
Table 1: Notations in Algorithm 1
Matrix Dim. Comment
S1
Not. Value Comment
O(n/ǫ) #rows in S1
s1
d1
d2
s2
eO(n/ǫ) #columns in D1 D1
eO(n/ǫ) #rows in D2
O(n/ǫ) #rows in S2
D2
S2
Z2
Rs1×m CountSketch matrix
Rn×d1
Leverage score sampling matrix
Rd2×m Leverage score sampling matrix
Rs2×m CountSketch matrix for fast regression
Rs1×d1 Low rank approximation solution matrix
Split, we can handle all cases.
some columns in bA with arbitrarily small noise to make sure bA has rank n. By applying procedure
Fix A ∈ Rm×n and B ∈ Rm×d. Let OPT = minrank −n C ′∈Rm×(n+d) kC′ − [A, B]kF . By using
techniques in low-rank approximation, we can find an approximation of a special form. More
precisely, let S1 ∈ Rs1×m be a CountSketch matrix with s1 = O(n/ǫ). Then we claim that it is
sufficient to look at solutions of the form U S1C.
Claim 3.1 (CountSketch matrix for low rank approximation problem). With probability 0.98,
min
rank −n U∈Rm×s1 kU S1C − Ck2
F ≤ (1 + ǫ)2 OPT2 .
We provide the proof in Appendix F.1. We shall mention that we cannot use leverage score
sampling here, because taking leverage score sampling on matrix C would take at least nnz(C)+(n+
d)2 time, while we are linear in d in the additive term in our running time eO(nnz(C)) + d· poly(n/ǫ).
Let U1 be the optimal solution of the program minU∈Rm×s1 kU S1C − Ck2
F , i.e.,
U1 = arg
min
rank −n U∈Rm×s1 kU S1C − Ck2
F .
(2)
If d is large compared to n, then program (2) is computationally expensive to solve. So we can
apply sketching techniques to reduce the size of the problem. Let D⊤1 ∈ Rd1×(n+d) denote a lever-
age score sampling and rescaling matrix according to the columns of S1C, with d1 = eO(n/ǫ)
nonzero entries on the diagonal of D1. Let U2 ∈ Rm×s1 denote the optimal solution to the problem
minrank −n U∈Rm×s1 kU S1CD1 − CD1k2
U2 = arg
F , i.e.,
rank −n U∈Rm×s1 kU S1CD1 − CD1k2
min
F .
(3)
Then the following claim comes from the constrained low-rank approximation result (Theorem D.1).
Claim 3.2 (Solving regression with leverage score sampling). Let U1 be defined in Eq. (2), and let
U2 be defined in Eq. (3). Then with probability 0.98,
kU2S1C − Ck2
F ≤ (1 + ǫ)2kU1S1C − Ck2
F .
We provide the proof in Appendix F.2. We now consider how to solve program (3). We observe
that
Claim 3.3. U2 ∈ colspan(CD1).
We can thus consider the following relaxation: given CD1, S1C and C, solve:
min
rank −n Z∈Rd1×s1 kCD1ZS1C − Ck2
F .
(4)
5
By setting CD1Z = U , we can check that program (4) is indeed a relaxation of program (3). Let
Z1 be the optimal solution to program (4). We show the following claim and delayed the proof in
F.3.
Claim 3.4 (Approximation ratio of relaxation). With probability 0.98,
kCD1Z1S1C − Ck2
F ≤ (1 + O(ǫ)) OPT2 .
However, program (4) still has a potentially large size, i.e., we need to work with an m × d1
matrix CD1. To handle this problem, we again apply sketching techniques. Let D2 ∈ Rd2×m be
a leverage score sampling and rescaling matrix according to the matrix CD1 ∈ Rm×d1, so that D2
has d2 = eO(n/ǫ) nonzeros on the diagonal. Now, we arrive at the small program that we are going
to directly solve:
min
rank −n Z∈Rd1×s1 kD2CD1ZS1C − D2Ck2
F .
(5)
We shall mention that here it is beneficial to apply leverage score sampling matrix because we
only need to compute leverage scores of a smaller matrix CD1, and computing D2C only involves
sampling a small fraction of the rows of C. On the other hand, if we were to use the CountSketch
matrix, then we would need to touch the whole matrix C when computing D2C. Overall, using
leverage score sampling at this step can reduce the constant factor of the nnz(C) term in the running
time, and may be useful in practice. Let rank-n Z2 ∈ Rd1×s1 be the optimal solution to this problem.
Claim 3.5 (Solving regression with a CountSketch matrix). With probability 0.98,
kCD1Z2S1C − Ck2
We provide the proof in Appendix F.4.
Our algorithm thus far is as follows: we compute matrices S1, D1, D2 accordingly, then solve
F ≤ (1 + ǫ)2kCD1Z1S1C − Ck2
F
program (5) to obtain Z2. At this point, we are able to obtain the low rank approximation bC =
CD1 · Z2 · S1C. We show the following claim and delayed the proof in Appendix F.5.
Claim 3.6 (Analysis of bC). With probability 0.94,
kbC − Ck2
F ≤ (1 + O(ǫ)) OPT2 .
Let bC = [bA, bB] where bA ∈ Rm×n and bB ∈ Rm×d. However, if our goal is to only output a
matrix X so that bAX = bB, then we can do this faster by not computing or storing the matrix bC.
Let S2 ∈ Rs2×m be a CountSketch matrix with s2 = O(n/ǫ). We solve a regression problem:
min
X∈Rn×d kS2bAX − S2bBk2
F .
optimal solution to the above problem.
Notice that S2bA and S2bB are computed directly from CD1, Z2, S1C and S2. Let X be the
Claim 3.7 (Approximation ratio guarantee). Assume bC = [bA, bAbX] for some bX ∈ Rn×d. Then
with probability at least 0.9,
k[bA, bAX] − [A, B]k2
F ≤ (1 + O(ǫ)) OPT2 .
6
We provide the proof in Appendix F.6.
small multiple of one of the last d columns that cannot be spanned to the first n columns until the
first n columns are full rank. Formally, we have
If the assumption bC = [bA, bAbX] in Claim 3.7 does not hold, then we need to apply procedure
Split. Because rank(bC) = n from our construction, if the first n columns of bC cannot span the last
d columns, then the first n columns of bC are not full rank. Hence we can keep adding a sufficiently
Lemma 3.8 (Analysis of procedure Split). Fix s1 = O(n/ǫ), s2 = O(n/ǫ), d1 = eO(n/ǫ). Given
CD1 ∈ Rm×d1, Z2 ∈ Rd1×s1, and S1C ∈ Rs1×(n+d) so that bC := CD1·Z2·S1C has rank n, procedure
SPLIT (Algorithm 1) returns A ∈ Rs2×n and B ∈ Rs2×d in time O(nnz(C) + d · poly(n/ǫ)) so that
there exists X ∈ Rn×d satisfying A · X = B. Moreover, letting bA be the matrix computed in lines
(24) to (27), then with probability 0.99,
We provide the proof in Appendix F.7. Now that we have A and B, and we can compute X by
k[bA, bAX] − CkF ≤ kbC − CkF + δ.
solving the regression problem minX∈Rn×d kAX − Bk2
F .
We next summarize the running time. Ommitted proofs are in Appendix F.8.
Lemma 3.9 (Running time analysis). Procedure FastTotalLeastSquares in Algorithm 1 runs
in time eO(nnz(A) + nnz(B) + d · poly(n/ǫ)).
To summarize, Theorem 3.10 shows the performance of our algorithm. Ommitted proofs are in
Appendix F.10.
Theorem 3.10 (Main Result). Given two matrices A ∈ Rm×n and B ∈ Rm×d, letting
OPT =
rank −n C ′∈Rm×(n+d) kC′ − [A, B]kF ,
min
we have that for any ǫ ∈ (0, 1), there is an algorithm (procedure FastTotalLeastSquares in
Algorithm 1) that runs in eO(nnz(A) + nnz(B)) + d· poly(n/ǫ) time and outputs a matrix X ∈ Rn×d
such that there is a matrix bA ∈ Rm×n satisfying that
holds with probability at least 9/10, where δ > 0 is arbitrarily small.
k[bA, bAX] − [A, B]kF ≤ (1 + ǫ) OPT +δ
Remark 3.11. The success probability 9/10 in Theorem 3.10 can be boosted to 1 − δ for any δ > 0
in a standard way. Namely, we run our FTLS algorithm O(log(1/δ)) times where in each run we
use independent randomness, and choose the solution found with the smallest cost. Note that for any
fixed output X, the cost k[ ¯A, ¯AX] − [A, B]kF can be efficiently approximated. To see this, let S be a
CountSketch matrix with O(ǫ−2) rows. Then kS[ ¯A, ¯AX] − S[A, B]kF = (1 ± ǫ)k[ ¯A, ¯AX] − [A, B]kF
with probability 9/10 (see, for example Lemma 40 of [CW13] ). We can compute kS[ ¯A, ¯AX] −
S[A, B]kF in time O(d · poly(n/ǫ)), and applying S can be done in nnz(A) + nnz(B) time. We
can then amplify the success probability by taking O(log(1/δ)) independent estimates and taking the
median of the estimates. This is a (1± ǫ)-approximation with probability at least 1− O(δ/ log(1/δ)).
We run our FTLS algorithm O(log(1/δ)) times, obtaining outputs X 1, . . . , X O(log(1/δ)) and for each
X i, apply the method above to estimate its cost. Since for each X i our estimate to the cost is within
1 ± ǫ with probability at least 1 − O(δ/(log(1/δ)), by a union bound the estimates for all X i are
within 1 ± ǫ with probability at least 1 − δ/2. Since also the solution with minimal cost is a 1 ± ǫ
approximation with probability at least 1 − δ/2, by a union bound we can achieve 1 − δ probability
with running time eO(log2(1/δ)) · (nnz(A) + nnz(B) + d · poly(n/ǫ))).
7
We further generalize our algorithm to handle regularization. Ommitted proofs can be found in
Appendix G.
Theorem 3.12 (Algorithm for regularized total least squares). Given two matrices A ∈ Rm×n and
B ∈ Rm×d and λ > 0, letting
OPT =
U∈Rm×n,V ∈Rn×(n+d) kU V − [A, B]k2
min
F + λkUk2
F + λkV k2
F ,
we have that for any ǫ ∈ (0, 1), there is an algorithm (procedure FastRegularizedTotalLeast-
Squares in Algorithm 3) that runs in
eO(nnz(A) + nnz(B) + d · poly(n/ǫ))
time and outputs a matrix X ∈ Rn×d such that there is a matrix bA ∈ Rm×n, bU ∈ Rm×n and
bV ∈ Rn×(n+d) satisfying k[bA, bAX] − bUbV k2
k[bA, bAX] − [A, B]k2
F ≤ δ and with probability 9/10,
F + λkbUk2
F ≤ (1 + ǫ) OPT +δ.
F + λkbV k2
4 Experiments
We conduct several experiments to verify the running time and optimality of our fast total least
squares algorithm 1. Let us first recall the multiple-response regression problem. Let A ∈ Rm×n
and B ∈ Rm×d. In this problem, we want to find X ∈ Rn×d so that AX ∼ B. The least squares
method (LS) solves the following optimization program:
cLS :=
min
X∈Rn×d,∆B∈Rm×d k∆Bk2
F ,
On the other hand, the total least squares method (TLS) solves the following optimization program:
subject to AX = B + ∆B.
cTLS :=
rank −n C ′∈Rm×(n+d) kC′ − [A B]kF .
min
The fast total least squares method (FTLS) returns X ∈ Rn×d, which provides an approximation
C′ = [bA bAX] to the TLS solution, and the cost is computed as cFTLS = kC′ − Ck2
Our numerical tests are carried out on an Intel Xeon E7-8850 v2 server with 2.30GHz and 4GB
F .
RAM under Matlab R2017b. 1
4.1 A Toy Example
We first run our FTLS algorithm on the following toy example, for which we have the analytical
solution exactly. Let A ∈ R3×2 be A11 = A22 = 1 and 0 everywhere else. Let B ∈ R3×1 be B3 = 3
and 0 everywhere else. We also consider the generalization of this example with larger dimension in
Appendix H. The cost of LS is 9, since AX can only have non-zero entries on the first 2 coordinates,
so the 3rd coordinate of AX − B must have absolute value 3. Hence the cost is at least 9. Moreover,
a cost 9 can be achieved by setting X = 0 and ∆B = −B. However, for the TLS algorithm, the cost
is only 1. Consider ∆A ∈ R3×2 where A11 = −1 and 0 everywhere else. Then C′ := [(A + ∆A), B]
has rank 2, and kC′ − CkF = 1.
1The code can be found at https://github.com/yangxinuw/total_least_squares_code.
8
2.5
2
1.5
1
0.5
0
0
TLS
LS
FTLS 0.9
FTLS 0.6
FTLS 0.3
FTLS 0.1
20
40
60
80
100
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
TLS
LS
FTLS 0.9
FTLS 0.6
FTLS 0.3
FTLS 0.1
0
20
40
60
80
100
2.5
2
1.5
1
0.5
0
0
TLS
LS
FTLS 0.9
FTLS 0.6
FTLS 0.3
FTLS 0.1
20
40
60
80
100
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
TLS
LS
FTLS 0.9
FTLS 0.6
FTLS 0.3
FTLS 0.1
20
40
60
80
100
Figure 1: Running time and accuracy of our FTLS algorithms. The left 2 figures are for the sparse matrix.
The right 2 pictures are for the Gaussian matrix. (Left) The y-axis is the running time of each algorithm
(counted in seconds); the x-axis is the size of the matrix. (Right) The y-axis is cost-TLS/cost-other, where
cost-other is the cost achieved by other algorithms. (Note we want to minimize the cost); the x-axis is the
size of the matrix.
We first run experiments on this small matrix. Since we know the solution of LS and TLS exactly
in this case, it is convenient for us to compare their results with that of the FTLS algorithm. When
we run the FTLS algorithm, we sample 2 rows in each of the sketching algorithms.
The experimental solution of LS is CLS = diag(0, 1, 3) which matches the theoretical solution.
The cost is 9. The experimental solution of TLS is CTLS = diag(1, 1, 0) which also matches the
theoretical result. The cost is 1.
FTLS is a randomized algorithm, so the output varies. We post several outputs:
CFTLS =
.06 −.01
−.01
.99
.76
.01
.25
.00
2.79
,
.14 −.26 −.22
−.26
.91 −.06
−.67 −.20
2.82
These solutions have cost of 1.55 and 1.47.
We run the FTLS multiple times to analyze the distribution of costs. Experimental result, which
can be found in Appendex I, shows that FTLS is a stable algorithm, and consistently performs better
than LS.
We also consider a second small toy example. Let A still be a 10 × 5 matrix and B be a 10 × 1
vector. Each entry A(i, j) is chosen i.i.d. from the normal distribution N (0, 1), and each entry B(i)
is chosen from N (0, 3). Because entries from A and B have different variance, we expect the results
of LS and TLS to be quite different. When we run the FTLS algorithm, we sample 6 rows.
We run FTLS 1000 times, and compute the distribution of costs. The results of this experiment,
which is in Appendex I, again demonstrates the stability of the algorithm.
4.2 Large Scale Problems
We have already seen that FTLS works pretty well on small matrices. We next show that the fast
total least squares method also provides a good estimate for large scale regression problems. The
setting for matrices is as follows: for k = 5, 10,··· , 100, we set A to be a 20k × 2k matrix where
A(i, i) = 1 for i = 1,··· , 2k and 0 everywhere else, and we set B to be a 20k × 1 vector where
B(2k + 1) = 3 and 0 elsewhere. As in the small case, the cost of TLS is 1, and the cost of LS is 9.
Recall that in the FTLS algorithm, we use Count-Sketch/leverage scores sampling/Gaussian
sketches to speed up the algorithm. In the experiments, we take sample density ρ = 0.1, 0.3, 0.6, 0.9
respectively to check our performance. The left 2 pictures in Figure 1 show the running time
together with the ratio TLS/FTLS for different sample densities.
9
Method Cost C-std Time T-std
Method Cost C-std Time T-std
0.05
TLS
0.0002
LS
0.0058
FTLS 0.9
0.0033
FTLS 0.6
0.0022
FTLS 0.3
0.0024
FTLS 0.1
Method Cost C-std Time T-std
0
0
0.0002
0.0003
0.0007
0.0016
1.12
0.0012
0.16
0.081
0.046
0.034
0.10
106
0.10
0.10
0.10
0.10
0.16
TLS
0.001
LS
0.025
FTLS 0.9
0.01
FTLS 0.6
0.005
FTLS 0.3
0.004
FTLS 0.1
Method Cost C-std Time T-std
1.36
0.0012
0.30
0.17
0.095
0.074
0
0
0.0032
0.0050
0.01
0.03
0.93
666
0.93
0.94
0.95
0.99
TLS
LS
FTLS 0.9
FTLS 0.6
FTLS 0.3
FTLS 0.1
1.85
2794
1.857
1.858
1.864
1.885
0
0
0.001
0.002
0.006
0.019
29.44
0.0022
3.12
1.62
0.77
0.60
1.44
0.001
0.081
0.054
0.027
0.017
TLS
LS
FTLS 0.9
FTLS 0.6
FTLS 0.3
FTLS 0.1
0.550
303
0.553
0.558
0.558
0.732
0
0
0.003
0.011
0.054
0.227
125.38
0.019
21.313
13.115
7.453
4.894
82.9
0.02
1.867
1.303
1.237
0.481
Table 2: Up Left: Airfoil Self-Noise. Up Right: Red wine. Down Left: White wine. Down Right:
Insurance Company Benchmark. C-std is the standard deviation for cost. T-std is the standard
deviation for running time.
We can see that the running time of FTLS is significantly smaller than that of TLS. This is
because the running time of TLS depends heavily on m, the size of matrix A. When we apply
sketching techniques, we significantly improve our running time. The fewer rows we sample, the
faster the algorithm runs. We can see that FTLS has pretty good performance; even with 10%
sample density, FTLS still performs better than LS. Moreover, the more we sample, the better
accuracy we achieve.
The above matrix is extremely sparse. We also consider another class of matrices. For k =
5, 10,··· , 100, we set A to be a 20k × 2k matrix where A(i, j) ∼ N (0, 1); we set B to be a 20k ×
1 vector where B(i) ∼ N (0, 3). As in previous experiments, we take sample densities of ρ =
0.1, 0.3, 0.6, 0.9, respectively, to check our performance. The results of this experiment are shown
in the right 2 pictures in Figure 1.
We see that compared to TLS, our FTLS sketching-based algorithm significantly reduces the
running time. FTLS is still slower than LS, though, because in the FTLS algorithm we still need
to solve a LS problem of the same size. However, as discussed, LS is inadequate in a number of
applications as it does not allow for changing the matrix A. The accuracy of our FTLS algorithms
is also shown.
We also conducted experiments on real datasets from the UCI Machine Learning Repository
[DKT17]. We choose datasets with regression task. Each dataset consists of input data and output
data. To turn it into a total least squares problem, we simply write down the input data as a matrix
A and the output data as a matrix B, then run the corresponding algorithm on (A, B). We have
four real datasets : Airfoil Self-Noise [UCIa] in Table 2(a), Wine Quality Red wine [UCIc, CCA+09]
in Table 2(b), Wine Quality White wine [UCIc, CCA+09] in Table 2(c), Insurance Company Bench-
mark (COIL 2000) Data Set [UCIb, PS] From the results„ we see that FTLS also performs well on
real data: when FTLS samples 10% of the rows, the result is within 5% of the optimal result of
TLS, while the running time is 20− 40 times faster. In this sense, FTLS achieves the advantages of
both TLS and LS: FTLS has almost the same accuracy as TLS, while FTLS is significantly faster.
10
References
[ACW17] Haim Avron, Kenneth L. Clarkson, and David P. Woodruff. Sharper bounds for regular-
ized data fitting. In Approximation, Randomization, and Combinatorial Optimization.
Algorithms and Techniques, APPROX/RANDOM 2017, August 16-18, 2017, Berkeley,
CA, USA, pages 27:1 -- 27:22, 2017.
[ALS+18] Alexandr Andoni, Chengyu Lin, Ying Sheng, Peilin Zhong, and Ruiqi Zhong. Sub-
In ICML. arXiv preprint
space embedding and linear regression with orlicz norm.
arXiv:1806.06430, 2018.
[BWZ16] Christos Boutsidis, David P Woodruff, and Peilin Zhong. Optimal principal compo-
nent analysis in distributed and streaming models. In Proceedings of the 48th Annual
ACM SIGACT Symposium on Theory of Computing (STOC), pages 236 -- 249. ACM,
https://arxiv.org/pdf/1504.06729, 2016.
[CCA+09] Paulo Cortez, António Cerdeira, Fernando Almeida, Telmo Matos, and José Reis. Mod-
eling wine preferences by data mining from physicochemical properties. Decision Support
Systems, 47(4):547 -- 553, 2009.
[CCF02] Moses Charikar, Kevin Chen, and Martin Farach-Colton. Finding frequent items in data
streams. In Automata, Languages and Programming, pages 693 -- 703. Springer, 2002.
[CCLY19] Michael B Cohen, Ben Cousins, Yin Tat Lee, and Xin Yang. A near-optimal algorithm
for approximating the john ellipsoid. In COLT, 2019.
[CLS19] Michael B Cohen, Yin Tat Lee, and Zhao Song. Solving linear programs in the current
In Proceedings of the 51st Annual ACM Symposium on
matrix multiplication time.
Theory of Computing (STOC). https://arxiv.org/pdf/1810.07896.pdf, 2019.
[CW87]
[CW13]
Don Coppersmith and Shmuel Winograd. Matrix multiplication via arithmetic progres-
sions. In Proceedings of the nineteenth annual ACM symposium on Theory of computing,
pages 1 -- 6. ACM, 1987.
Kenneth L. Clarkson and David P. Woodruff. Low rank approximation and regression in
input sparsity time. In Symposium on Theory of Computing Conference, STOC'13, Palo
Alto, CA, USA, June 1-4, 2013, pages 81 -- 90. https://arxiv.org/pdf/1207.6365,
2013.
[CWW19] Kenneth L. Clarkson, Ruosong Wang, and David P Woodruff. Dimensionality reduction
for tukey regression. In ICML. arXiv preprint arXiv:1904.05543, 2019.
[DJS+19] Huaian Diao, Rajesh Jayaram, Zhao Song, Wen Sun, and David P. Woodruff. Optimal
sketching for kronecker product regression and low rank approximation. In NeurIPS,
2019.
[DKT17] Dua Dheeru and Efi Karra Taniskidou. UCI machine learning repository, 2017.
[DMM06a] Petros Drineas, Michael W. Mahoney, and S. Muthukrishnan. Sampling algorithms for
l2 regression and applications. In Proceedings of the Seventeenth Annual ACM-SIAM
Symposium on Discrete Algorithm, SODA '06, pages 1127 -- 1136, Philadelphia, PA, USA,
2006. Society for Industrial and Applied Mathematics.
11
[DMM06b] Petros Drineas, Michael W. Mahoney, and S. Muthukrishnan. Subspace sampling and
relative-error matrix approximation: Column-row-based methods. In Algorithms - ESA
2006, 14th Annual European Symposium, Zurich, Switzerland, September 11-13, 2006,
Proceedings, pages 304 -- 314, 2006.
[DMMS11] Petros Drineas, Michael W Mahoney, S Muthukrishnan, and Tamás Sarlós. Faster least
squares approximation. Numerische mathematik, 117(2):219 -- 249, 2011.
[DSSW18] Huaian Diao, Zhao Song, Wen Sun, and David P. Woodruff. Sketching for kronecker
product regression and p-splines. AISTATS, 2018.
[FT07]
Shmuel Friedland and Anatoli Torokhti. Generalized rank-constrained matrix approxi-
mations. SIAM Journal on Matrix Analysis and Applications, 29(2):656 -- 659, 2007.
[LHW17] Xingguo Li, Jarvis Haupt, and David Woodruff. Near optimal sketching of low-rank
tensor regression. In Advances in Neural Information Processing Systems, pages 3466 --
3476, 2017.
[LPT09]
[LSZ19]
[LV10]
[LV14]
[MM13]
Shuai Lu, Sergei V Pereverzev, and Ulrich Tautenhahn. Regularized total least squares:
computational aspects and error bounds. SIAM Journal on Matrix Analysis and Appli-
cations, 31(3):918 -- 941, 2009.
Yin Tat Lee,
minimization
https://arxiv.org/pdf/1905.04447.pdf, 2019.
Zhao Song,
and Qiuyi Zhang.
in
the
current matrix multiplication
Solving
time.
empirical
risk
In COLT.
Jörg Lampe and Heinrich Voss. Solving regularized total least squares problems based
on eigenproblems. Taiwanese Journal of Mathematics, 14(3A):885 -- 909, 2010.
Jorg Lampe and Heinrich Voss. Large-scale dual regularized total least squares. Elec-
tronic Transactions on Numerical Analysis, 42:13 -- 40, 2014.
Xiangrui Meng and Michael W Mahoney. Low-distortion subspace embeddings in
In Proceedings of
input-sparsity time and applications to robust linear regression.
the forty-fifth annual ACM symposium on Theory of computing, pages 91 -- 100. ACM,
https://arxiv.org/pdf/1210.3135, 2013.
[MVH07]
Ivan Markovsky and Sabine Van Huffel. Overview of total least-squares methods. Signal
processing, 87(10):2283 -- 2302, 2007.
[NN13]
Jelani Nelson and Huy L Nguyên. Osnap: Faster numerical linear algebra algorithms via
sparser subspace embeddings. In 2013 IEEE 54th Annual Symposium on Foundations of
Computer Science (FOCS), pages 117 -- 126. IEEE, https://arxiv.org/pdf/1211.1002,
2013.
[PS]
P. van der Putten and M. van Someren. Coil challenge 2000: The insurance company
case. Technical report.
[PSW17]
Eric Price, Zhao Song, and David P. Woodruff. Fast regression with an ℓ∞ guarantee.
In ICALP, 2017.
[RG04]
Rosemary A Renaut and Hongbin Guo. Efficient algorithms for solution of regularized
total least squares. SIAM Journal on Matrix Analysis and Applications, 26(2):457 -- 476,
2004.
12
[RSW16]
Ilya Razenshteyn, Zhao Song, and David P Woodruff. Weighted low rank approxima-
tions with provable guarantees. In Proceedings of the 48th Annual Symposium on the
Theory of Computing (STOC), 2016.
[SWY+19] Zhao Song, Ruosong Wang, Lin F. Yang, Hongyang Zhang, and Peilin Zhong. Efficient
symmetric norm regression via linear sketching. In NeurIPS, 2019.
[SWZ17]
Zhao Song, David P Woodruff, and Peilin Zhong. Low rank approximation with en-
trywise ℓ1-norm error. In Proceedings of the 49th Annual Symposium on the Theory of
Computing (STOC). ACM, https://arxiv.org/pdf/1611.00898, 2017.
[SWZ19a] Zhao Song, David P Woodruff, and Peilin Zhong. Average case column subset selection
for entrywise ℓ1-norm loss. In NeurIPS, 2019.
[SWZ19b] Zhao Song, David P Woodruff, and Peilin Zhong. Relative error tensor low rank ap-
proximation. In SODA. https://arxiv.org/pdf/1704.08246, 2019.
[SWZ19c] Zhao Song, David P Woodruff, and Peilin Zhong. Towards a zero-one law for column
subset selection. In NeurIPS, 2019.
[TZ12]
[UCIa]
[UCIb]
Mikkel Thorup and Yin Zhang. Tabulation-based 5-independent hashing with applica-
tions to linear probing and second moment estimation. SIAM Journal on Computing,
41(2):293 -- 331, 2012.
UCI. Airfoil self-noise. In . https://archive.ics.uci.edu/ml/datasets/Airfoil+Self-Noise,
.
Insurance
UCI.
https://archive.ics.uci.edu/ml/datasets/Insurance+Company+Benchmark+%28COIL+2000%29,
.
benchmark
data
set.
company
(coil
2000)
In
.
[UCIc]
UCI. Wine quality. In . https://archive.ics.uci.edu/ml/datasets/Wine+Quality,
.
[VHV91]
Sabine Van Huffel and Joos Vandewalle. The total least squares problem: computational
aspects and analysis, volume 9. Siam, 1991.
[Wil12]
Virginia Vassilevska Williams. Multiplying matrices faster than coppersmith-winograd.
In Proceedings of the forty-fourth annual ACM symposium on Theory of computing
(STOC), pages 887 -- 898. ACM, 2012.
[Woo14]
David P. Woodruff. Sketching as a tool for numerical linear algebra. Foundations and
Trends in Theoretical Computer Science, 10(1-2):1 -- 157, 2014.
13
Appendix
A Notation
In addition to O(·) notation, for two functions f, g, we use the shorthand f . g (resp. &) to
indicate that f ≤ Cg (resp. ≥) for an absolute constant C. We use f h g to mean cf ≤ g ≤ Cf
for constants c, C.
B Oblivious and Non-oblivious sketching matrix
In this section we introduce techniques in sketching. In order to optimize performance, we introduce
multiple types of sketching matrices, which are used in Section 3. In Section B.1, we provide the
definition of CountSketch and Gaussian Transforms. In Section B.2, we introduce leverage scores
and sampling based on leverage scores.
B.1 CountSketch and Gaussian Transforms
CountSketch matrix comes from the data stream literature [CCF02, TZ12].
Definition B.1 (Sparse embedding matrix or CountSketch transform). A CountSketch transform
is defined to be Π = ΦD ∈ Rm×n. Here, D is an n × n random diagonal matrix with each diagonal
entry independently chosen to be +1 or −1 with equal probability, and Φ ∈ {0, 1}m×n is an m × n
binary matrix with Φh(i),i = 1 and all remaining entries 0, where h : [n] → [m] is a random map
such that for each i ∈ [n], h(i) = j with probability 1/m for each j ∈ [m]. For any matrix A ∈ Rn×d,
ΠA can be computed in O(nnz(A)) time.
To obtain the optimal number of rows, we need to apply Gaussian matrix, which is another
well-known oblivious sketching matrix.
Definition B.2 (Gaussian matrix or Gaussian transform). Let S = 1√m · G ∈ Rm×n where each
entry of G ∈ Rm×n is chosen independently from the standard Gaussian distribution. For any matrix
A ∈ Rn×d, SA can be computed in O(m · nnz(A)) time.
We can combine CountSketch and Gaussian transforms to achieve the following:
Definition B.3 (CountSketch + Gaussian transform). Let S′ = SΠ, where Π ∈ Rt×n is the
CountSketch transform (defined in Definition B.1) and S ∈ Rm×t is the Gaussian transform (defined
in Definition B.2). For any matrix A ∈ Rn×d, S′A can be computed in O(nnz(A) + dtmω−2) time,
where ω is the matrix multiplication exponent.
B.2 Leverage Scores
We do want to note that there are other ways of constructing sketching matrix though, such as
through sampling the rows of A via a certain distribution and reweighting them. This is called
leverage score sampling [DMM06b, DMM06a, DMMS11]. We first give the concrete definition of
leverage scores.
Definition B.4 (Leverage scores). Let U ∈ Rn×k have orthonormal columns with n ≥ k. We will
use the notation pi = u2
2 is referred to as the i-th leverage score of U .
i /k, where u2
i = ke⊤i Uk2
14
Next we explain the leverage score sampling. Given A ∈ Rn×d with rank k, let U ∈ Rn×k be
an orthonormal basis of the column span of A, and for each i let k · pi be the squared row norm
of the i-th row of U . Let pi denote the i-th leverage score of U . Let β > 0 be a constant and
q = (q1,··· , qn) denote a distribution such that, for each i ∈ [n], qi ≥ βpi. Let s be a parameter.
Construct an n×s sampling matrix B and an s×s rescaling matrix D as follows. Initially, B = 0n×s
and D = 0s×s. For the same column index j of B and of D, independently, and with replacement,
pick a row index i ∈ [n] with probability qi, and set Bi,j = 1 and Dj,j = 1/√qis. We denote this
procedure Leverage score sampling according to the matrix A.
Leverage score sampling is efficient in the sense that leverage score can be efficiently approxi-
mated.
Theorem B.5 (Running time of over-estimation of leverage score, Theorem 14 in [NN13]). For
any ǫ > 0, with probability at least 2/3, we can compute 1 ± ǫ approximation of all leverage scores
of matrix A ∈ Rn×d in time eO(nnz(A) + rωǫ−2ω) where r is the rank of A and ω ≈ 2.373 is the
exponent of matrix multiplication [CW87, Wil12].
In Section C we show how to apply matrix sketching to solve regression problems faster. In
Section D, we give a structural result on rank-constrained approximation problems.
C Multiple Regression
Linear regression is a fundamental problem in Machine Learning. There are a lot of attempts trying
to speed up the running time of different kind of linear regression problems via sketching matrices
[CW13, MM13, PSW17, LHW17, DSSW18, ALS+18, CWW19]. A natural generalization of linear
regression is multiple regression.
We first show how to use CountSketch to reduce to a multiple regression problem:
Theorem C.1 (Multiple regression, [Woo14]). Given A ∈ Rn×d and B ∈ Rn×m, let S ∈ Rs×n
denote a sampling and rescaling matrix according to A. Let X∗ denote arg minX kAX − Bk2
F and
X′ denote arg minX kSAX − SBk2
F . If S has s = O(d/ǫ) rows, then we have that
kAX′ − Bk2
F ≤ (1 + ǫ)kAX∗ − Bk2
F
holds with probability at least 0.999.
The following theorem says leverage score sampling solves multiple response regression:
Theorem C.2 (See, e.g., the combination of Corollary C.30 and Lemma C.31 in [SWZ19b]). Given
A ∈ Rn×d and B ∈ Rn×m, let D ∈ Rn×n denote a sampling and rescaling matrix according to A. Let
X∗ denote arg minX kAX − Bk2
F . If D has O(d log d + d/ǫ)
non-zeros in expectation, that is, this is the expected number of sampled rows, then we have that
F and X′ denote arg minX kDAX − SBk2
kAX′ − Bk2
F ≤ (1 + ǫ)kAX∗ − Bk2
F
holds with probability at least 0.999.
D Generalized Rank-Constrained Matrix Approximation
We state a tool which has been used in several recent works [BWZ16, SWZ17, SWZ19b].
15
Theorem D.1 (Generalized rank-constrained matrix approximation, Theorem 2 in [FT07]). Given
matrices A ∈ Rn×d, B ∈ Rn×p, and C ∈ Rq×d, let the singular value decomposition (SVD) of B be
B = UBΣBV ⊤B and the SVD of C be C = UCΣCV ⊤C . Then
B†(UBU⊤B AVCV ⊤C )kC† =
arg min
rank −k X∈Rp×qkA − BXCkF
where (UBU⊤B AVC V ⊤C )k ∈ Rn×d is of rank at most k and denotes the best rank-k approximation to
UBU⊤B AVCV ⊤C ∈ Rn×d in Frobenius norm.
Moreover, (UBU⊤B AVCV ⊤C )k can be computed by first computing the SVD decomposition of
UBU⊤B AVCV ⊤C in time O(nd2), then only keeping the largest k coordinates. Hence B†(UBU⊤B AVC V ⊤C )kC†
can be computed in O(nd2 + np2 + qd2) time.
E Closed Form for the Total Least Squares Problem
Markovsky and Huffel [MVH07] propose the following alternative formulation of total least squares
problem.
rank −n C ′∈Rm×(n+d) kC′ − CkF
min
(6)
When program (1) has a solution (X, ∆A, ∆B), we can see that (1) and (6) are in general equivalent
by setting C′ = [A + ∆A, B + ∆B]. However, there are cases when program (1) fails to have a
solution, while (6) always has a solution.
V21 V22(cid:21) where V11 ∈ Rn×n and V22 ∈ Rd×d.
As discussed, a solution to the total least squares problem can sometimes be written in closed
form. Letting C = [A, B], denote the singular value decomposition (SVD) of C by U ΣV ⊤, where
Σ = diag(σ1,··· , σn+d) ∈ Rm×(n+d) with σ1 ≥ σ2 ≥ ··· ≥ σn+d. Also we represent (n + d)× (n + d)
matrix V as (cid:20)V11 V12
Clearly bC = U diag(σ1,··· , σn, 0,··· , 0)V ⊤ is a minimizer of program (6). But whether a
solution to program (1) exists depends on the singularity of V22.
In the rest of this section we
introduce different cases of the solution to program (1), and discuss how our algorithm deals with
each case.
E.1 Unique Solution
We first consider the case when the Total Least Squares problem has a unique solution.
Theorem E.1 (Theorem 2.6 and Theorem 3.1 in [VHV91]). If σn > σn+1, and V22 is non-singular,
then the minimizer bC is given by U diag(σ1,··· , σn, 0,··· , 0)V ⊤, and the optimal solution bX is given
by −V12V −1
22 .
Our algorithm will first find a rank n matrix C′ = [A′, B′] so that kC′ − CkF is small, then
solve a regression problem to find X′ so that A′X′ = B′. In this sense, this is the most favorable
case to work with, because a unique optimal solution bC exists, so if C′ approximates bC well, then
the regression problem A′X′ = B′ is solvable.
E.2 Solution exists, but is not unique
If σn = σn+1, then it is still possible that the Total Least Squares problem has a unique solution,
although this time, the solution bX is not unique. Theorem E.2 is a generalization of Theorem E.1.
16
Algorithm 2 Least Squares and Total Least Squares Algorithms
1: procedure LeastSquares(A, B)
2:
X ← minX kAX − BkF
CLS ← [A, AX]
return CLS
3:
4:
5: procedure TotalLeastSquares(A, B)
6:
CTLS ← minrank −n C ′ kC − C′kF
return CTLS
7:
Theorem E.2 (Theorem 3.9 in [VHV91]). Let p ≤ n be a number so that σp > σp+1 = ··· = σn+1.
Let Vp be the submatrix that contains the last d rows and the last n − p + d columns of V . If Vp
is non-singular, then multiple minimizers bC = [bA, bB] exist, and there exists bX ∈ Rn×d so that
bAbX = bB.
We can also handle this case. As long as the Total Least Squares problem has a solution bX, we
are able to approximate it by first finding C′ = [A′, B′] and then solving a regression problem.
E.3 Solution does not exist
has no solution.
the cost of program (1). But there are cases where this cost is not approchable in program (1).
F , where bC is the optimal solution to program (6), always lower bounds
Notice that the cost kbC−Ck2
Theorem E.3 (Lemma 3.2 in [VHV91]). If V22 is singular, letting bC denote [bA, bB], then bAX = bB
Theorem E.3 shows that even if we can compute bC precisely, we cannot output X, because the
first n columns of bC cannot span the rest d columns. In order to generate a meaningful result, our
algorithm will perturb C′ by an arbitrarily small amount so that A′X′ = B′ has a solution. This
will introduce an arbitrarily small additive error in addition to our relative error guarantee.
F Omitted Proofs in Section 3
F.1 Proof of Claim 3.1
Proof. Let C∗ be the optimal solution of minrank −n C ′∈Rm×(n+d) kC′ − [A, B]kF . Since rank(C∗) =
n ≪ m, there exist U∗ ∈ Rm×s1 and V ∗ ∈ Rs1×(n+d) so that C∗ = U∗V ∗, and rank(U∗) =
rank(V ∗) = n. Therefore
Now consider the problem formed by multiplying by S1 on the left,
min
V ∈Rs1×(n+d) kU∗V − Ck2
F = OPT2 .
min
V ∈Rs1×(n+d) kS1U∗V − S1Ck2
F .
Letting V ′ be the minimizer to the above problem, we have
V ′ = (S1U∗)†S1C.
17
Thus, we have
min
rank −n U∈Rm×s1 kU S1C − Ck2
F
F
F ≤ kU∗(S1U∗)†S1C − Ck2
= kU∗V ′ − Ck2
≤ (1 + ǫ)kS1U∗V ′ − S1Ck2
≤ (1 + ǫ)kS1U∗V ∗ − S1Ck2
≤ (1 + ǫ)2kU∗V ∗ − Ck2
= (1 + ǫ)2 OPT2
F
F
F
where the first step uses the fact that U∗(S1U∗)†S1 ∈ Rm×s1 with rank n, the second step is
the definition of V ′, the third step follows from the definition of the Count-Sketch matrix S1 and
Theorem C.1, the fourth step uses the optimality of V ′, and the fifth step again uses Theorem
C.1.
F.2 Proof of Claim 3.2
Proof. We have
kU2S1C − Ck2
F ≤ (1 + ǫ)kU2S1CD1 − CD1k2
≤ (1 + ǫ)kU1S1CD1 − CD1k2
≤ (1 + ǫ)2kU1S1C − Ck2
F ,
F
F
where the first step uses the property of a leverage score sampling matrix D1, the second step follows
from the definition of U2 (i.e., U2 is the minimizer), and the last step follows from the property of
the leverage score sampling matrix D1 again.
F.3 Proof of Claim 3.4
Proof. From Claim 3.2 we have that U2 ∈ colspan(CD1). Hence we can choose Z so that CD1Z =
U2. Then by Claim 3.1 and Claim 3.2, we have
kCD1ZS1C − Ck2
F = kU2S1C − Ck2
F ≤ (1 + ǫ)4 OPT2 .
Since Z1 is the optimal solution, the objective value can only be smaller.
F.4 Proof of Claim 3.5
Proof. Recall that Z1 = arg minrank −n Z∈Rd1×s1 kCD1ZS1C − Ck2
F . Then we have
kCD1Z2S1C − Ck2
F ≤ (1 + ǫ)kD2CD1Z2S1C − D2Ck2
≤ (1 + ǫ)kD2CD1Z1S1C − D2Ck2
≤ (1 + ǫ)2kCD1Z1S1C − Ck2
F ,
F
F
where the first step uses the property of the leverage score sampling matrix D2, the second step
follows from the definition of Z2 (i.e., Z2 is a minimizer), and the last step follows from the property
of the leverage score sampling matrix D2.
18
F.5 Proof of Claim 3.6
Proof.
kbC − Ck2
F = kCD1 · Z2 · S1C − Ck2
≤ (1 + ǫ)2kCD1Z1S1C − Ck2
≤ (1 + O(ǫ)) OPT2
F
F
where the first step is the definition of bC, the second step is Claim 3.5, and the last step is Claim
3.4.
F.6 Proof of Claim 3.7
Proof. By the condition that bC = [bA bAbX], bB = bAbX, hence bX is the optimal solution to the program
minX∈Rn×d kbAX − bBk2
F . Hence by Theorem C.1, with probability at least 0.99,
Therefore
Then it follows from Claim 3.6.
kbAX − bBk2
k[bA, bAX] − [A, B]k2
F ≤ (1 + ǫ)kbAbX − bBk2
F = k[bA, bB] − Ck2
F = 0
F = kbC − Ck2
F .
F.7 Proof of Lemma 3.8
Proof. Proof of running time. Let us first check the running time. We can compute C = S2 · bC
by first computing S2 · CD1, then computing (S2CD1) · Z2, then finally computing S2CD1Z2S1C.
Notice that D1 is a leverage score sampling matrix, so nnz(CD1) ≤ nnz(C). So by Definition B.1,
we can compute S2 · CD1 in time O(nnz(C)). All the other matrices have smaller size, so we can do
matrix multiplication in time O(d· poly(n/ǫ)). Once we have C, the independence between columns
in A can be checked in time O(s2 · n). The FOR loop will be executed at most n times, and inside
each loop, line (21) will take at most d linear independence checks. So the running time of the FOR
loop is at most O(s2 · n) · n · d = O(d · poly(n/ǫ)). Therefore the running time is as desired.
Proof of Correctness. We next argue the correctness of procedure Split. Since rank(bC) = n,
with high probability rank(C) = rank(S2·bC) = n. Notice that B is never changed in this subroutine.
In order to show there exists an X so that AX = B, it is sufficient to show that at the end of
procedure Split, rank(A) = rank(C), because this means that the columns of A span each of the
columns of C, including B. Indeed, whenever rank(A∗,[i]) < i, line 25 will be executed. Then by
doing line 26, the rank of A will increase by 1, since by the choice of j, A∗,i + δ · B∗,j is independent
form A∗,[i−1]. Because rank(C) = n, at the end of the FOR loop we will have rank(A) = n.
Finally let us compute the cost. In line (10) we use δ/ poly(m), and thus
δ2
k[bA, bB] − bCk2
F ≤
poly(m) · kbBk2
F ≤ δ2.
(7)
We know that X is the optimal solution to the program minX∈Rn×d kS2bAX − S2bBk2
Theorem C.1, with probability 0.99,
F . Hence by
kbAX − bBk2
F ≤ (1 + ǫ) min
X∈Rn×d kS2bAX − S2bBk2
F = 0.
19
which implies bAX = bB. Hence we have
k[bA, bAX] − CkF ≤ k[bA, bAX] − bCkF + kbC − CkF
= k[bA, bB] − bCkF + kbC − CkF
≤ δ + kbC − CkF
where the first step follows by triangle inequality, and the last step follows by (7).
F.8 Proof of Lemma 3.9
Proof. We bound the time of each step:
1. Construct the s1 × m Count-Sketch matrix S1 and compute S1C with s1 = O(n/ǫ). This
step takes time nnz(C) + d · poly(n/ǫ).
2. Construct the (n + d) × d1 leverage sampling and rescaling matrix D1 with d1 = eO(n/ǫ)
nonzero diagonal entries and compute CD1. This step takes time eO(nnz(C) + d · poly(n/ǫ)).
3. Construct the d2 × m leverage sampling and rescaling matrix D2 with d2 = eO(n/ǫ) nonzero
diagonal entries. This step takes time eO(nnz(C) + d · poly(n/ǫ)) according to Theorem B.5.
4. Compute Z2 ∈ Rd1×s1 by solving the rank-constrained system:
rank −n Z∈Rd1×s1 kD2CD1ZS1C − D2Ck2
F .
min
Note that D2CD1 has size eO(n/ǫ) × eO(n/ǫ), S1C has size O(n/ǫ) × (n + d), and D2C has size
eO(n/ǫ) × (n + d), so according to Theorem D.1, we have an explicit closed form for Z2, and the
time taken is d · poly(n/ǫ).
5. Run procedure Split to get A ∈ Rs2×n and B ∈ Rs2×d with s2 = O(n/ǫ). By Lemma 3.8,
this step takes time O(nnz(C) + d · poly(n/ǫ)).
6. Compute X by solving the regression problem minX∈Rn×d kAX−Bk2
F in time O(d·poly(n/ǫ)).
This is because X = (A)†B, and A has size O(n/ǫ)×n, so we can compute (A)† in time O((n/ǫ)ω) =
poly(n/ǫ), and then compute X in time O((n/ǫ)2 · d) since B is an O(n/ǫ) × d matrix.
Notice that nnz(C) = nnz(A) + nnz(B), so we have the desired running time.
F.9 Procedure Evaluate
In this subsection we explain what procedure Evaluate does.
Ideally, we would like to apply
procedure Split on the matrix bC directly so that the linear system bAX = bB has a solution.
However, bC has m rows, which is computationally expensive to work with. So in the main algorithm
we actually apply procedure Split on the sketched matrix S2bC. When we need to compute the
cost, we shall redo the operations in procedure Split on bC to split bC correctly. This is precisely
what we are doing in lines (24) to (27).
F.10 Putting it all together
Proof. The running time follows from Lemma 3.9. For the approximation ratio, let bA, A be defined
as in Lemma 3.8. From Lemma 3.8, there exists X ∈ Rn×d satisfying AX = B. Since X is obtained
from solving the regression problem kAX − Bk2
F , we also have AX = B. Hence with probability
0.9,
where the first step uses Lemma 3.8 and the second step uses Claim 3.6. Rescaling ǫ gives the
k[bA, bAX] − CkF ≤ δ + kbC − CkF ≤ δ + (1 + O(ǫ)) OPT,
desired statement.
20
Algorithm 3 Our Fast Total Least Squares Algorithm with Regularization
3:
4:
5:
6:
7:
8:
⊲ Theorem G.7
rows of CS⊤2
1: procedure FastRegularizedTotalLeastSquares(A, B, n, d, λǫ, δ)
2:
s1 ← eO(n/ǫ), s2 ← eO(n/ǫ), , s3 ← eO(n/ǫ), d1 ← eO(n/ǫ)
Choose S1 ∈ Rs1×m to be a CountSketch matrix, then compute S1C
Choose S2 ∈ Rs2×(n+d) to be a CountSketch matrix, then compute CS⊤2
Choose D1 ∈ Rd1×m to be a leverage score sampling and rescaling matrix according to the
bZ1, bZ2 ← arg minZ1∈Rn×s1 ,Z2∈Rs2×n kD1CS⊤2 Z2Z1S1C − D1Ck2
F +
λkZ1S1Ck2
⊲ Theorem G.2
A, B, π ← Split(CS⊤2 , bZ1, bZ2, S1C, n, d, δ/ poly(m)), X ← minkAX − BkF
return X
Choose S3 ∈ Rs3×m to be a CountSketch matrix
C ← (S3 · CS⊤2 ) · bZ2 · bZ1 · S1C
A ← C∗,[n], B ← C∗,[n+d]\[n]
T ← ∅, π(i) = −1 for all i ∈ [n]
for i = 1 → n do
⊲ bC = CS⊤2 bZ2bZ1S1C ; C = S3bC
⊲ bA = bC∗,[n], bB = bC∗,[n+d]\[n]; A = S3bA, B = S3bB
9: procedure Split(CS⊤2 , bZ1, bZ2, S1C, n, d, δ)
F + λkD1CS⊤2 Z2k2
⊲ Lemma 3.8
F
if A∗,i is linearly dependent of A∗,[n]\{i} then
j ← minj∈[d]\T{B∗,j is linearly independent of A}, A∗,i ← A∗,i +δ·B∗,j, T ← T ∪{j},
π(i) ← j
return A, B, π
10:
11:
12:
13:
14:
15:
16:
17:
⊲ π : [n] → {−1} ∪ ([n + d]\[n])
G Extension to regularized total least squares problem
In this section we provide our algorithm for the regularized total least squares problem and prove
its correctness. Recall our regularized total least squares problem is defined as follows.
OPT :=
min
bA∈Rm×n,X∈Rn×d,U∈Rm×n,V ∈Rn×(n+d)kU V − [A, B]k2
subject to [bA, bAX] = U V
F + λkUk2
F + λkV k2
F
(8)
Definition G.1 (Statistical Dimension, e.g., see [ACW17]). For λ > 0 and rank k matrix A, the
statistical dimension of the ridge regression problem with regularizing weight λ is defined as
sdλ(A) := Xi∈[k]
1
1 + λ/σ2
i
where σi is the i-th singular value of A for i ∈ [k].
Notice that sdλ(A) is decreasing in λ, so we always have sdλ(A) ≤ sd0(A) = rank(A).
Lemma G.2 (Exact solution of low rank approximation with regularization, Lemma 27 of [ACW17]).
Given positive integers n1, n2, r, s, k and parameter λ ≥ 0. For C ∈ Rn1×r, D ∈ Rs×n2, B ∈ Rn1×n2,
the problem of finding
min
ZR∈Rr×k,ZS∈Rk×s kCZRZSD − Bk2
F + λkCZRk2
F + λkZS Dk2
F ,
21
and the minimizing of CZR ∈ Rn1×k and ZSD ∈ Rk×n2, can be solved in
O(n1r · rank(C) + n2s · rank(D) + rank(D) · n1(n2 + rC))
time.
Theorem G.3 (Sketching for solving ridge regression, Theorem 19 in [ACW17]). Fix m ≥ n. For
A ∈ Rm×n, B ∈ Rn×d and λ > 0, consider the rigid regression problem
min
X∈Rn×d kAX − Bk2
F + λkXk2
F .
Let S ∈ Rs×m be a CountSketch matrix with s = eO(sdλ(A)/ǫ) = eO(n/ǫ), then with probability 0.99,
min
X∈Rn×d kSAX − SBk2
F + λkXk2
F ≤ (1 + ǫ) min
X∈Rn×d kAX − Bk2
F + λkXk2
F
Moreover, SA, SB can be computed in time
O(nnz(A) + nnz(B)) + eO(cid:0)(n + d)(sdλ(A)/ǫ + sdλ(A)2)(cid:1) .
We claim that it is sufficient to look at solutions of the form CS⊤2 Z2Z1S1C.
Claim G.4 (CountSketch matrix for low rank approximation problem). Given matrix C ∈ Rm×(n+d).
Let OPT be defined as in (8). For any ǫ > 0, let S1 ∈ Rs1×m, S2 ∈ Rs2×m be the sketching matrices
defined in Algorithm 3, then with probability 0.98,
min
Z1∈Rn×s1 ,Z2∈Rs2×n kCS⊤2 Z2Z1S1C − Ck2
F + λkCS⊤2 Z2k2
F + λkZ1S1Ck2
F ≤ (1 + ǫ)2 OPT .
Proof. Let U∗ ∈ Rm×n and V ∗ ∈ Rn×(n+d) be the optimal solution to the program (8). Consider
the following optimization problem:
min
V ∈Rn×(n+d) kU∗V − Ck2
F + λkV k2
F
(9)
Clearly V ∗ ∈ Rn×(n+d) is the optimal solution to program (9), since for any solution V ∈ Rn×(n+d)
to program (9) with cost c, (U∗, V ) is a solution to program (8) with cost c + λkU∗k2
F .
Program (9) is a ridge regression problem. Hence we can take a CountSketch matrix S ∈ Rs1×m
with s1 = eO(n/ǫ) to obtain
(10)
min
V ∈Rn×(n+d) kS1U∗V − S1Ck2
F + λkV k2
F
Let V1 ∈ Rn×(n+d) be the minimizer of the above program, then we know
V1 =(cid:20)S1U∗√λIn(cid:21)†(cid:20)S1C
0 (cid:21) ,
which means V1 ∈ Rn×(n+d) lies in the row span of S1C ∈ Rs1×(n+d). Moreover, by Theorem G.3,
with probability at least 0.99 we have
kU∗V1 − Ck2
F + λkV1k2
F ≤ (1 + ǫ)kU∗V ∗ − Ck2
F + λkV ∗k2
F
(11)
22
Now consider the problem
min
U∈Rm×n kU V1 − Ck2
F + λkUk2
F
(12)
Let U0 ∈ Rm×n be the minimizer of program (12). Similarly, we can take a CountSketch matrix
S2 ∈ Rs2×(n+d) with s2 = eO(n/ǫ) to obtain
F + λkUk2
Let U1 ∈ Rm×n be the minimizer of program (13), then we know
U⊤1 =(cid:20)S2V ⊤1√λIn(cid:21)†(cid:20)S2C⊤
0 (cid:21) ,
U∈Rm×n kU V1S⊤2 − CS⊤2 k2
(13)
min
F
which means U1 ∈ Rm×n lies in the column span of CS⊤2 ∈ Rm×s2. Moreover, with probability
at least 0.99 we have
kU1V1 − Ck2
F + λkU1k2
F ≤ (1 + λ) · (kU0V1 − Ck2
≤ (1 + λ) · (kU∗V1 − Ck2
F + λkU0k2
F )
F + λkU∗k2
F )
(14)
where the first step we use Theorem G.3 and the second step follows that U0 is the minimizer.
F + λkU1k2
F + (kU1V1 − Ck2
F + (1 + ǫ)(kU∗V1 − Ck2
Now let us compute the cost.
kU1V1 − Ck2
= λkV1k2
≤ λkV1k2
≤ (1 + ǫ) ·(cid:0)λkU∗k2
≤ (1 + ǫ) ·(cid:0)λkU∗k2
≤ (1 + ǫ)2 · (kU∗V ∗ − Ck2
= (1 + ǫ)2 OPT
F
F + λkV1k2
F + λkU1k2
F )
F + λkU∗k2
F )
F )(cid:1)
F + λkV1k2
F + (kU∗V1 − Ck2
F )(cid:1)
F + (1 + ǫ)2 · (kU∗V ∗ − Ck2
F + λkV ∗k2
F + λkV ∗k2
F )
F + λkU∗k2
where the second step follows from (14), the fourth step follows from (11), and the last step follows
from the definition of U∗ ∈ Rm×n, V ∗ ∈ Rn×(n+d).
Finally, since V1 ∈ Rn×(n+d) lies in the row span of S1C ∈ Rs1×(n+d) and U1 ∈ Rm×n lies in the
column span of CS⊤2 ∈ Rm×s2, there exists Z∗1 ∈ Rn×s1 and Z∗2 ∈ Rs2×n so that V1 = Z∗1 S1C ∈
Rn×(n+d) and U1 = CS⊤2 Z∗2 ∈ Rm×n. Then the claim stated just follows from (Z∗1 , Z∗2 ) are also
feasible.
Now we just need to solve the optimization problem
min
Z1∈Rn×s1 ,Z2∈Rs2×n kCS⊤2 Z2Z1S1C − Ck2
F + λkCS⊤2 Z2k2
F + λkZ1S1Ck2
F
(15)
The size of this program is quite huge, i.e., we need to work with an m× d2 matrix CS⊤2 . To handle
this problem, we again apply sketching techniques. Let D1 ∈ Rd1×m be a leverage score sampling
and rescaling matrix according to the matrix CS2 ∈ Rm×s2, so that D1 has d1 = eO(n/ǫ) nonzeros
on the diagonal. Now, we arrive at the small program that we are going to directly solve:
F + λkZ1S1Ck2
Z1∈Rn×s1 ,Z2∈Rs2×n kD1CS⊤2 Z2Z1S1C − D1Ck2
F + λkD1CS⊤2 Z2k2
(16)
min
F
We have the following approximation guarantee.
23
Claim G.5. Let (Z∗1 , Z∗2 ) be the optimal solution to program (15). Let (bZ1, bZ2) be the optimal
solution to program (16). With probability 0.96,
kCS⊤2 bZ2bZ1S1C − Ck2
≤(1 + ǫ)2(kCS⊤2 Z∗2 Z∗1 S1C − Ck2
F + λkCS⊤2 bZ2k2
F + λkbZ1S1Ck2
F
F + λkCS⊤2 Z∗2k2
F + λkZ∗1 S1Ck2
F )
Proof. This is because
kCS⊤2 bZ2bZ1S1C − Ck2
F + λkCS⊤2 bZ2k2
≤(1 + ǫ)(cid:16)kD1CS⊤2 bZ2bZ1S1C − D1Ck2
≤(1 + ǫ)(cid:16)kD1CS⊤2 bZ2bZ1S1C − D1Ck2
≤(1 + ǫ)(cid:16)kD1CS⊤2 Z∗2 Z∗1 S1C − D1Ck2
≤(1 + ǫ)2(cid:16)kCS⊤2 Z∗2 Z∗1 S1C − Ck2
F
F + λkbZ1S1Ck2
F + λkD1CS⊤2 bZ2k2
F + λkD1CS⊤2 bZ2k2
F + λkD1CS⊤2 Z∗2k2
F
F(cid:17) + λkbZ1S1Ck2
F(cid:17)
F + λkbZ1S1Ck2
F(cid:17)
F + λkZ∗1 S1Ck2
F + λkCS⊤2 Z∗2k2
F(cid:17)
F + λkZ∗1 S1Ck2
where the first step uses property of the leverage score sampling matrix D1, the third step follows
score sampling matrix D1.
from (bZ1, bZ2) are minimizers of program (16), and the fourth step again uses property of the leverage
Let bU = CS⊤2 bZ2,bV = bZ1S1C and bC = bUbV . Combining Claim G.4 and Claim G.5 together, we
get with probability at least 0.91,
kbUbV − [A, B]k2
F + λkbUk2
F + λkbV k2
F ≤ (1 + ǫ)4 OPT
(17)
If the first n columns of bC can span the whole matrix bC, then we are in good shape. In this case
we have:
Claim G.6 (Perfect first n columns). Let S3 ∈ Rs3×m be the CountSketch matrix defined in Algo-
rithm 3. Write bC as [bA, bB] where bA ∈ Rm×n and bB ∈ Rm×d. If there exists bX ∈ Rn×d so that
bB = bAbX, let ¯X ∈ Rn×d be the minimizer of minX∈Rn×d kS3bAX − S3bBk2
F , then with probability 0.9,
Proof. We have with probability 0.99,
k[bA, bA ¯X] − [A, B]k2
kbA ¯X − bBk2
F + λkbUk2
F + λkbV k2
F ≤ (1 + ǫ)kbAbX − bBk2
F = 0
F ≤ (1 + ǫ)4 OPT
where the first step follows from Theorem C.1 and the second step follows from the assumption.
Recall that bC = bUbV , so
k[bA, bA ¯X] − [A, B]k2
=kbUbV − [A, B]k2
F + λkbUk2
F + λkbUk2
F + λkbV k2
F + λkbV k2
F
F ≤ (1 + ǫ)4 OPT
where the last step uses (17).
However, if bC does not have nice structure, then we need to apply our procedure Split, which
would introduce the additive error δ. Overall, by rescaling ǫ, our main result is summarized as
follows.
24
Theorem G.7 (Restatement of Theorem 3.12, algorithm for the regularized total least squares
problem). Given two matrices A ∈ Rm×n and B ∈ Rm×d and λ > 0, letting
OPT =
min
U∈Rm×n,V ∈Rn×(n+d) kU V − [A, B]k2
F + λkUk2
F + λkV k2
F ,
we have that for any ǫ ∈ (0, 1), there is an algorithm that runs in
eO(nnz(A) + nnz(B) + d · poly(n/ǫ))
time and outputs a matrix X ∈ Rn×d such that there is a matrix bA ∈ Rm×n, bU ∈ Rm×n and
bV ∈ Rn×(n+d) satisfying that k[bA, bAX] − bUbV k2
k[bA, bAX] − [A, B]kF + λkbUk2
F ≤ δ and
F + λkbV k2
F ≤ (1 + ǫ) OPT +δ
H Toy Example
We first run our FTLS algorithm on the following toy example, for which we have an analytical
solution exactly. Let A ∈ Rm×n be Aii = 1 for i = 1,··· , n and 0 everywhere else. Let B ∈ Rm×1
be Bn+1 = 3 and 0 everywhere else.
The cost of LS is 9, since AX can only have non-zero entries on the first n coordinates, so the
(n + 1)-th coordinate of AX − B must have absolute value 3. Hence the cost is at least 9. Moreover,
a cost 9 can be achieved by setting X = 0 and ∆B = −B.
However, for the TLS algorithm, the cost is only 1. Consider ∆A ∈ Rm×n where A11 = −1 and
0 everywhere else. Then C′ := [(A + ∆A), B] does have rank n, and kC′ − CkF = 1.
For a concrete example, we set m = 10, n = 5. That is,
C := [A, B] =
1 0 0 0 0 0
0 1 0 0 0 0
0 0 1 0 0 0
0 0 0 1 0 0
0 0 0 0 1 0
0 0 0 0 0 3
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
We first run experiments on this small matrix. Because we know the solution of LS and TLS exactly
in this case, it is convenient for us to compare their results with that of the FTLS algorithm. When
we run the FTLS algorithm, we sample 6 rows in all the sketching algorithms.
The experimental solution of LS is CLS which is the same as the theoretical solution. The cost
is 9. The experimental solution of TLS is CTLS which is also the same as the theoretical result. The
cost is 1.
25
CLS =
1 0 0 0 0 0
0 1 0 0 0 0
0 0 1 0 0 0
0 0 0 1 0 0
0 0 0 0 1 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
CTLS =
0 0 0 0 0 0
0 1 0 0 0 0
0 0 1 0 0 0
0 0 0 1 0 0
0 0 0 0 1 0
0 0 0 0 0 3
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
FTLS is a randomized algorithm, so the output varies. We post several outputs:
0
0
0
0
0
0
0
0
0 0.5 0.5 0
0
0
0
0
0
0
1
0
0.1 −0.3
0
0
0 −0.9
2.7
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
CFTLS =
This solution has a cost of 4.3.
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
2.7
0
0
0
0
0
0
0
0
0
0
0
0
0
0.09 0.09 0 0.27
0
0
0
0
0.82 0.82 0 2.45
0.5 −0.5
−0.5
0.5
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
bCFTLS =
0
0
0
0
This solution has a cost of 5.5455.
CFTLS =
This solution has a cost of 3.4.
0.5
0
0
0
0
0
0
0
0
0
0.5
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0 −0.9
0
0
0
0
0
0
0
0
26
I More Experiments
Figure 2 shows the experimental result described in Section 4.1. It collects 1000 runs of our FTLS
algorithm on 2 small toy examples. In both figures, the x-axis is the cost of the FTLS algorithm,
measured by kC′ − Ck2
F where C′ is the output of our FTLS algorithm; the y-axix is the frequency
of each cost that is grouped in suitable range.
Figure 2: Cost distribution of our fast least squares algorithm on toy examples. The x-axis is the
cost for FTLS. (Note that we want to minimize the cost); the y-axis is the frequency of each cost. (Left)
First toy example, TLS cost is 1, LS cost is 9. (Right) Second toy example, TLS cost is 1.30, LS cost is 40.4
27
|
1702.02321 | 2 | 1702 | 2017-04-18T01:39:24 | Position Heaps for Parameterized Strings | [
"cs.DS"
] | We propose a new indexing structure for parameterized strings, called parameterized position heap. Parameterized position heap is applicable for parameterized pattern matching problem, where the pattern matches a substring of the text if there exists a bijective mapping from the symbols of the pattern to the symbols of the substring. We propose an online construction algorithm of parameterized position heap of a text and show that our algorithm runs in linear time with respect to the text size. We also show that by using parameterized position heap, we can find all occurrences of a pattern in the text in linear time with respect to the product of the pattern size and the alphabet size. | cs.DS | cs |
Position Heaps for Parameterized Strings
Diptarama1, Takashi Katsura1, Yuhei Otomo1, Kazuyuki Narisawa1, and Ayumi
Shinohara1
1Graduate School of Information Sciences, Tohoku University, , 6-6-05 Aramaki
Aza Aoba, Aoba-ku, Sendai, Japan, {diptarama@shino., katsura@shino.,
otomo@shino., narisawa@, ayumi@}ecei.tohoku.ac.jp
November 16, 2018
Abstract
We propose a new indexing structure for parameterized strings, called parame-
terized position heap. Parameterized position heap is applicable for parameterized
pattern matching problem, where the pattern matches a substring of the text if there
exists a bijective mapping from the symbols of the pattern to the symbols of the
substring. We propose an online construction algorithm of parameterized position
heap of a text and show that our algorithm runs in linear time with respect to the
text size. We also show that by using parameterized position heap, we can find all
occurrences of a pattern in the text in linear time with respect to the product of
the pattern size and the alphabet size.
1
Introduction
String matching problem is to find occurrences of a pattern string in a text string.
Formally, given a text string t and a pattern string p over an alphabet Σ, output all
positions at which p occurs in t. Suffix tree and suffix array are most widely used data
structures and provide many applications for various string matchings (see e.g. [11, 6]).
Ehrenfeucht et al. [8] proposed an indexing structure for string matching, called
a position heap. Position heap uses less memory than suffix tree does, and provides
efficient search of patterns by preprocessing the text string, similarly to suffix tree
and suffix array. A position heap for a string t is a sequence hash tree [4] for the
ordered set of all suffixes of t. In [8], the suffixes are ordered in the ascending order of
length, and the proposed construction algorithm processes the text from right to left.
Later, Kucherov [13] considered the ordered set of suffixes in the descending order of
length and proposed a linear-time online construction algorithm based on the Ukkonen's
algorithm [16]. Nakashima et al. [14] proposed an algorithm to construct a position heap
for a set of strings, where the input is given as a trie of the set. Gagie et al. [10] proposed
a position heap with limited height and showed some relations between position heap
and suffix array.
The parameterized pattern matching that focuses on a structure of strings is intro-
duced by Baker [2]. Let Σ and Π be two disjoint sets of symbols. A string over Σ ∪ Π is
1
called a parameterized string (p-string for short). In the parameterized pattern matching
problem, given p-strings t and p, find positions of substrings of t that can be transformed
into p by applying one-to-one function that renames symbols in Π. The parameterized
pattern matching is motivated by applying to the software maintenance [1, 2, 3], the
plagiarism detection [9], the analysis of gene structure [15], and so on. Similar to the
basic string matching problem, some indexing structures that support the parameterized
pattern matching are proposed, such as parameterized suffix tree [2], structural suffix
tree [15], and parameterized suffix array [7, 12].
In this paper, we propose a new indexing structure called parameterized position
heap for the parameterized pattern matching. The parameterized position heap is a
sequence hash tree for the ordered set of prev-encoded [2] suffixes of a parameterized
string. We give an online construction algorithm of a parameterized position heap based
on Kucherov's algorithm [13] that runs in O(n log (Σ + Π)) time and an algorithm
that runs in O(m log (Σ + Π) + mΠ + occ) time to find the occurrences of a pattern
in the text, where n is the length of the text, m is the length of the pattern, Σ is the
number of constant symbols, Σ is the number of parameter symbols, and occ is the
number of occurrences of the pattern in the text.
2 Notation
Let Σ and Π be two disjoint sets of symbols. Σ is a set of constant symbols and Π
is a set of parameter symbols. An element of Σ∗ is called a string, and an element of
(Σ ∪ Π)∗ is called a parameterized string, or p-string for short. For a p-string w = xyz,
x, y, and z are called prefix, substring, and suffix of w, respectively. w denotes the
length of w, and w[i] denotes the i-th symbol of w for 1 ≤ i ≤ w. The substring of w
that begins at position i and ends at position j is denoted by w[i : j] for 1 ≤ i ≤ j ≤ w.
Moreover, let w[: i] = w[1 : i] and w[i :] = w[i : w] for 1 ≤ i ≤ w. The empty p-string
is denoted by ε, that is ε = 0. For convenience, let w[i : j] = ε if i > j. Let N denote
the set of all non-negative integers.
Given two p-strings w1 and w2, w1 and w2 are a parameterized match or p-match,
denoted by w1 ≈ w2, if there exists a bijection f from the symbols of w1 to the symbols
of w2, such that f is identity on the constant symbols [2]. We can determine whether
w1 ≈ w2 or not by using an encoding called prev-encoding defined as follows.
Definition 1 (Prev-encoding [2]). For a p-string w over Σ ∪ Π, the prev-encoding for
w, denoted by prev (w), is a string x of length w over Σ ∪ N defined by
w[i]
x[i] =
0
i − max{j w[j] = w[i] and 1 ≤ j < i}
if w[i] ∈ Σ,
if w[i] ∈ Π and w[i] (cid:54)= w[j] for 1 ≤ j < i,
otherwise.
For any p-strings w1 and w2, w1 ≈ w2 if and only if prev (w1) = prev (w2). For
example, given Σ = {a, b} and Π = {u, v, x, y}, s1 = uvuvauuvb and s2 = xyxyaxxyb
are p-matches where prev (w1) = prev (w2) = 0022a314b.
The parameterized pattern matching is a problem to find occurrences of a p-string
pattern in a p-string text defined as follows.
2
(a)
(b)
(c)
(a) A sequence hash tree for (aab, ab, bba, baa, aaba, baaba). (b) A posi-
Figure 1:
tion heap for a string abbaabaabaabab, (c) An augmented position heap for a string
abbaabaabaabab. Maximal-reach pointers for mrp(i) (cid:54)= i are illustrated by doublet
arrows.
Definition 2 (Parameterized pattern matching [2]). Given two p-strings, text t and
pattern p, find all positions i in t such that t[i : i + p − 1] ≈ p.
For example, let us consider a text t = uvaubuavbv and a pattern p = xayby over
Σ = {a, b} and Π = {u, v, x, y}. Because p ≈ t[2 : 6] and p ≈ t[6 : 10], we should output
2 and 6.
Throughout this paper, let t be a text of length n and p be a pattern of length m.
3 Position Heap
In this section, we briefly review the position heap for strings. First we introduce the
sequence hash tree that is a trie for hashing proposed by Coffman and Eve [4]. Each
edge of the trie is labeled by a symbol and each node can be identified with the string
obtained by concatenating all labels found on the path from root to the node.
Definition 3 (Sequence Hash Tree). Let W = (w1, . . . , wn) be an ordered set of strings
over Σ and Wi = (w1, . . . , wi) for 1 ≤ i ≤ n. A sequence hash tree SHT (W) = (Vn, En)
for W is a trie over Σ defined recursively as follows. Let SHT (Wi) = (Vi, Ei). Then,
(cid:40)
SHT (Wi) =
({ε},∅)
(Vi−1 ∪ {pi}, Ei−1 ∪ {(qi, c, pi)})
(if i = 0) ,
(if 1 ≤ i ≤ n) .
where pi is the shortest prefix of wi such that pi
c = wi[pi]. If no such pi exists, then Vi = Vi−1 and Ei = Ei−1.
(cid:54)∈ Vi−1, and qi = wi[1 : pi− 1],
Each node in a sequence hash tree stores one or several indices of strings in the input
set. An example of a sequence hash tree is shown in Fig. 1 (a).
The position heap proposed by Ehrenfeucht et al. [8] is a sequence hash tree for
the ordered set of all suffixes of a string. Two types of position heap are known. The
first one is proposed by Ehrenfeucht et al. [8], that constructed by the ordered set of
suffixes in ascending order of length and the second one is proposed by Kucherov [13],
which constructed in descending order. We adopt the Kucherov [13] type and his online
3
1.abb2.ab3.bba4.baa5.aaba6.baabaababaa 1 3 4 6 2 51234567891011121314abbaabaabaabab 1 8 4 7 105, 132, 14 11 3 6 12 9aaaaaabbbbbb1234567891011121314abbaabaabaabab 1 8 4 7 105, 132, 14 11 3 6 12 9aaaaaabbbbbb (a)
(b)
Figure 2: Let Σ = {a}, Π = {x, y} and t = xaxyxyxyyaxyx. (a) A parameterized
position heap PPH (t). Broken arrows denote suffix pointers.
(b) An augmented
parameterized position heap APPH (t). Parameterized maximal-reach pointers for
pmrp(i) (cid:54)= i are illustrated by doublet arrows.
construction algorithm for constructing position heaps for parameterized strings in
Section 4. Here we recall the definition of the position heap by Kucherov.
Definition 4 (Position Heap [13]). Given a string t ∈ Σn, let St = (t[1 :], t[2 :], . . . , t[n :])
be the ordered set of all suffixes of t except ε in descending order of length. The position
heap PH (t) for t is SHT (St).
Each node except the root in a position heap stores either one or two integers those
are beginning positions of corresponding suffixes. We call them regular node and double
node respectively. Assume that i and j are positions stored by a double node v in
PH (t) where i < j, i and j are called the primary position and the secondary position
respectively. Fig. 1 (b) shows an example of a position heap.
In order to find occurrences of the pattern in O(m + occ) time, Ehrenfeucht et al. [8]
and Kucherov [13] added additional pointer called maximal-reach pointer to the position
heap and called this extended data structure as augmented position heap. An example
of an augmented position heap is showed in Fig. 1 (c).
4 Parameterized Position Heap
In this section, we propose a new indexing structure called parameterized position heap.
It is based on the position heap proposed by Kucherov [13].
4.1 Definition and Property of Parameterized Position Heap
The parameterized position heap is a sequence hash tree [4] for the ordered set of
prev-encoded suffixes in the descending order of length.
Definition 5 (Parameterized Position Heap). Given a p-string t ∈ (Σ ∪ Π)n, let
St = (prev (t[1 :]), prev (t[2 :]), . . . , prev (t[n :])) be the ordered set of all prev-encoded
suffixes of the p-string t except ε in descending order of length. The parameterized
position heap PPH (t) for t is SHT (St).
4
1234567891011121314xaxyxyxyyaxyxy10a2022221a43222a0022221a432230022221a43224002221a4322500221a432260021a43227001a4322801a032290a032210a0022110022120021300140 3, 13 6 2 100a1, 14 8 7 94, 125, 110011122a ԋ 3, 13 6 2 100a1, 14 8 7 94, 125, 110011122a ԋ Fig. 2 (a) shows an example of a parameterized position heap. A parameterized
position heap PPH (t) for a p-string t of length n consists of the root and nodes that
corresponds to prev (t[1 :]), prev (t[2 :]), . . . , prev (t[n :]), so PPH (t) has at most n+1 nodes.
Each node in PPH (t) holds either one or two of beginning positions of corresponding
p-suffixes similar to the standard position heaps. We can specify each node in PPH (t)
by its primary position, its secondary position, or the string obtained by concatenating
labels found on the path from the root to the node.
Different from standard position heap, prev (t[i :]) = prev (t)[i :] does not necessarily
hold for some cases. For example, for t = xaxyxyxyyaxyxy, prev (t[3 :]) = 0022221a4322
while prev (t)[3 :] = 0222221a4322. Therefore, the construction and matching algorithms
for the standard position heaps cannot be directly applied for the parameterized position
heaps. However, we can similar properties to construct parameterized position heaps
efficiently.
Lemma 1. For i and j, where 1 ≤ i ≤ j ≤ n, if prev (t[i : j]) is represented in PPH (t),
then a prev-encoded string for any substring of t[i : j] is also represented in PPH (t).
Proof. First we will show that prev-encoding of any prefix of t[i : j] is represented in
PPH (t). From the definition of prev-encoding, prev (t[i : j])[1 : i − j] = prev (t[i : j − 1]).
In other words, prev (t[i : j − 1]) is a prefix of prev (t[i : j]). From the definition of
PPH (t), prefixes of prev (t[i : j]) are represented in PPH (t). Therefore, prev (t[i : j − 1])
is represented in PPH (t). Similarly, prev (t[i : j − 2]), ··· , prev (t[i : i]) are represented
in PPH (t).
Next, we will show that prev-encoding of any suffix of t[i : j] is represented in PPH (t).
From the above discussion, there are positions b0 < b1 < ··· < bj−i = i in t such that
prev (t[bk : bk + k]) = prev (t[i : i + k]). From the definition of parameterized position
heap, prev (t[b1 + 1 : b1 + 1]) is represented in PPH (t). Since prev (t[bk + 1 : bk + k]) is
a prefix of prev (t[bk+1 + 1 : bk+1 + k + 1]) for 0 < k < j − i, if prev (t[bk + 1 : bk + k]) is
represented in PPH (t) then prev (t[bk+1 + 1 : bk+1 + k + 1]) is also represented in PPH (t)
recursively. Therefore, prev (t[bj−i + 1 : bj−i + j − i]) = prev (t[i + 1 : j]) is represented
in PPH (t). Similarly, prev (t[i + 2 : j]), ··· , prev (t[j : j]) are represented in PPH (t).
Since any prefix and suffix of prev (t[i : j]) is represented in PPH (t), we can say that
any substring of prev (t[i : j]) is represented in PPH (t) by induction.
4.2 Online Construction Algorithm of Parameterized Position Heap
In this section, we propose an online algorithm that constructs parameterized position
heaps. Our algorithm is based on Kucherov's algorithm, although it cannot be applied
easily. The algorithm updates PH (t[1 : k]) to PH (t[1 : k + 1]) when t[k + 1] is read,
where 1 ≤ k ≤ n−1. Updating of the position heap begins from a special node, called the
active node. A position specified by the active node is called the active position. At first,
we show that there exists a position similar to the active position in the parameterized
position heap.
Lemma 2. If j is a secondary position of a double node in a parameterized position
heap, then j + 1 is also a secondary position.
5
Proof. Let i be the primary position and j be the secondary position of node v, where
i < j. This means there is a position h such that prev (t[i : h]) = prev (t[j :]). By Lemma 1,
there is a node that represents prev (t[i + 1 : h]). Since prev (t[j + 1 :]) = prev (t[i + 1 : h]),
then j + 1 will be the secondary positions of node prev (t[i + 1 : h]).
Lemma 2 means that there exists a position s which splits all positions in t[1 : n]
into two intervals, similar to the active position in [13]. Positions in [1 : s − 1] and [s : n]
are called primary and secondary positions, respectively. We also call the position s as
active position.
Assume we have constructed PPH (t[1 : k]) and we want to construct PPH (t[1 : k + 1])
from PPH (t[1 : k]). The primary positions 1, . . . , s − 1 in PPH (t[1 : k]) become primary
positions also in PPH (t[1 : k + 1]), because prev (t[i : k]) = prev (t[i : k + 1])[1 : k− 1 + 1]
holds for 1 ≤ i ≤ s − 1. Therefore, we do not need to update the primary positions.
On the other hand, the secondary positions s, . . . , k require some modifications. When
inserting a new symbol, two cases can occur. The first case is that prev (t[i : k + 1]) is
not represented in PPH (t[1 : k]). In this case, a new node prev (t[i : k + 1]) is created
as a child node of prev (t[i : k]) and position i becomes the primary position of the new
node. The second case is that prev (t[i : k + 1]) was already represented in PPH (t[1 : k]).
In this case, the secondary position i that is stored in prev (t[i : k]) currently should be
moved to the child node prev (t[i : k + 1]), and position i becomes the secondary position
of this node.
From Lemma 1, if the node prev (t[i : k]) has an edge to the node prev (t[i : k + 1]),
prev (t[i + 1 : k]) also has an edge to prev (t[i + 1 : k + 1]). Therefore, there exists r, with
1 ≤ s ≤ r ≤ k, that splits the interval [s : k] into two subintervals [s : r − 1] and
[r : k], such that the node prev (t[i : k]) does not have an edge to prev (t[i : k + 1]) for
s ≤ i ≤ r − 1, and does have such an edge for r ≤ i ≤ k.
The above analysis leads to the following lemma that specifies the modifications from
PPH (t[1 : k]) to PPH (t[1 : k + 1]).
Lemma 3. Given t ∈ (Σ ∪ Π)n, consider PPH (t[1 : k]) for k < n. Let s be the active
position, stored in the node prev (t[s : k]). Let r ≥ s be the smallest position such
that node prev (t[r : k]) has an outgoing edge labeled with prev (t[r : k + 1])[k − r + 2].
PPH (t[1 : k + 1]) can be obtained by modifying PPH (t[1 : k]) in the following way:
1. For each node prev (t[i : k]), s ≤ i < r, create a new child prev (t[i : k + 1]) linked
by an edge labeled prev (t[i : k + 1])[k − i + 2]. Delete the secondary position i
from the node prev (t[i : k]) and assign it as the primary position of the new node
prev (t[i : k + 1]),
2. For each node prev (t[i : k]), r ≤ i ≤ k, move the secondary position i from the
node prev (t[i : k]) to the node prev (t[i : k + 1]).
Moreover, r will be the active position in PPH (t[1 : k + 1]).
Proof. Consider the first case that i be a secondary position in PPH (t[1 : k]) and
s ≤ i < r. From the definition of r, there is no node prev (t[i : k + 1]) in PPH (t[i : k]).
Therefore, i will be a primary position of the node prev (t[i : k + 1]) in PPH (t[1 : k + 1]).
We can update the position heap from PPH (t[1 : k]) to PPH (t[1 : k + 1]) by delete i
6
(a)
(b)
An example of updating a parameterized position heap,
from (a)
Figure 3:
PPH (xaxyyxyx) to (b) PPH (xaxyyxyxx). The updated positions are colored red.
The secondary positions 6 and 7 in PPH (xaxyyxyx) are become primary positions
in PPH (xaxyyxyxx), while the secondary position 8 in PPH (xaxyyxyx) is become a
secondary position of another node in PPH (xaxyyxyxx). The active position is updated
from 6 to 8.
from secondary position of the node prev (t[i : k]) and create a new node prev (t[i : k + 1])
and assign i to its primary position for the case s ≤ i < r.
Next case, i be a secondary position in PPH (t[1 : k]) and r ≤ i ≤ k.
In this
case, there is a node prev (t[i : k + 1]) in PPH (t[i : k]) and the node prev (t[i : k + 1])
is also represented in PPH (t[i : k + 1]). Therefore, i will be a secondary position of
the node prev (t[i : k + 1]) in PPH (t[1 : k + 1]). We can update the position heap from
PPH (t[1 : k]) to PPH (t[1 : k + 1]) by delete i from secondary position of the node
prev (t[i : k]) and assign i as secondary position of the node prev (t[i : k + 1]) for the case
r ≤ i ≤ k.
Since position i for 1 ≤ i < r be a primary position in PPH (t[1 : k + 1]) and position
i for r ≤ i ≤ k + 1 be a secondary position in PPH (t[1 : k + 1]), r will be the active
position in PPH (t[1 : k + 1]).
Fig. 3 show an example of updating a parameterized position heap. The modifications
specified by Lemma 3 need to be applied to all secondary positions. In order to perform
these modifications efficiently, we use parameterized suffix pointers.
Definition 6 (Parameterized Suffix Pointer). For each node prev (t[i : j]) of PPH (t),
the parameterized suffix pointer of prev (t[i : j]) is defined by psp(prev (t[i : j])) =
prev (t[i + 1 : j]).
By Lemma 1, whenever the node prev (t[i : j]) exists, the node prev (t[i + 1 : j]) exists
too. This means that psp(prev (t[i : j])) always exists. During the construction of the
parameterized position heap, let ⊥ be the auxiliary node that works as the parent of
root and is connected to root with an edge labeled with any symbol c ∈ Σ∪ 0. We define
psp(root) = ⊥.
7
3,7 20a1,8 45,6012 ԋ 3 6 20a1,9 4,8 750112 ԋ1 Algorithm 1: Parameterized position heap online construction algorithm
Input: A p-string t ∈ (Σ ∪ Π)n
Output: A parameterized position heap PPH (t)
1 create root and ⊥ nodes;
2 psp(root) = ⊥;
3 child (⊥, c) = root for c ∈ Σ ∪ {0};
4 currentNode = root;
5 s = 1;
6 for i = 1 to n do
7
c = normalize(prev (t)[i], depth(currentNode));
lastCreateNode = undefined;
while child (currentNode, c) = null do
8
9
10
11
12
13
14
15
16
17
18
19
create newnode;
prim(newnode) = s;
child (currentNode, c) = newnode;
if lastCreateNode (cid:54)= undefined then psp(lastCreateNode) = newnode;
lastCreateNode = newnode;
currentNode = psp(currentNode);
c = normalize(prev (t)[i], depth(currentNode));
s = s + 1;
currentNode = child (currentNode, c);
if lastCreateNode (cid:54)= undefined then psp(lastCreateNode) = currentNode;
20 while s ≤ n do
21
22
23
sec(currentNode) = s;
currentNode = psp(currentNode);
s = s + 1;
When s is the active position in PPH (t[1 : k]), we call prev (t[s : k]) the active node.
If no node holds a secondary position, root becomes the active node and the active
position is set to k + 1. The nodes for the secondary positions s, s + 1, . . . , k can be
visited by traversing with the suffix pointers from the active node. Thus, the algorithm
only has to memorize the active position and the active node in order to visit any other
secondary positions.
Updating PPH (t[1 : k]) to PPH (t[1 : k + 1]) specified by Lemma 3 is processed as
the following procedures. The algorithm traverses with the suffix pointers from the active
node till the node that has the outgoing edge labeled with prev (t[i : k + 1])[k − i + 2]
is found, which is i = r. For each traversed node, a new node is created and linked by
an edge labeled with prev (t[i : k + 1])[k − i + 2] to each node. A suffix pointer to this
new node is set from the previously created node. When the node that has the outgoing
edge labeled with prev (t[i : k + 1])[k − i + 2] is traversed, the algorithm moves to the
node that is led to by this edge, and a suffix pointer to this node is set from the last
created node, then the algorithm assigns this node to be the active node.
A pseudocode of our proposed construction algorithm is given as Algorithm 1.
prim(v) and sec(v) denotes primary and secondary positions of v, respectively. From
8
the property of prev-encoding, prev (t[i + 1 : k + 1])[k − i + 1] = prev (t[i : k + 1])[k −
i + 2] if prev (t[i : k + 1])[k − i + 2] ∈ Σ or prev (t[i : k + 1])[k − i + 2] ≤ k − i and
prev (t[i + 1 : k])[k − i + 1] = 0 otherwise. Therefore, we use a function normalize(c, j)
that returns c if c ∈ Σ or c ≤ j and returns 0 otherwise.
The construction algorithm consists of n iterations. In the i-th iteration, the algorithm
read t[i] and make PPH (t[1 : i]). In the i-th iteration, the traversal of the suffix pointers
as explained above is done. Since the depth of the current node decreases by traversing
a suffix pointer, the number of the nodes that can be visited by traversal is O(n). For
each traversed node, all the operations such as creating a node, an edge and updating
position can be done in O(log (Σ + Π)). Therefore, the total time for the traversals is
O(n log (Σ + Π)).
From the above discussion, the following theorem is obtained.
Theorem 1. Given t ∈ (Σ∪ Π)n, Algorithm 1 constructs PPH (t) in O(n log (Σ + Π))
time and space.
4.3 Augmented Parameterized Position Heaps
We will describe augmented parameterized position heaps, the parameterized position
heaps with an additional data structure called the parameterized maximal-reach point-
ers similar to the maximal-reach pointers for the position heap [8]. The augmented
parameterized position heap gives an efficient algorithm for parameterized pattern
matching.
Definition 7 (Parameterized Maximal-Reach Pointer). For a position i on t, a param-
eterized maximal-reach pointer of pmrp(i) is a pointer from node i to the deepest node
whose path label is a prefix of prev (t[i :]).
Obviously, if i is a secondary position, then pmrp(i) is node i itself. We assume
that the parameterized maximal-reach pointer for a double node applies to the primary
position of this node. Fig. 2 (b) shows an example of an augmented parameterized
position heap. Given a prev-encoded p-string prev (w) represented in an augmented
parameterized position heap APPH (t) and a position 1 ≤ i ≤ n, we can determine
whether prev (w) is a prefix of prev (t[i :]) or not in O(1) time by checking whether
pmrp(i) is a descendant of prev (w) or not. It can be done in O(1) time by appropriately
preprocessing APPH (t) [5].
Parameterized maximal-reach pointers can be computed by using parameterized
suffix pointers, similar to [13]. Algorithm 2 shows an algorithm to compute parameterized
maximal-reach pointers. pmrp(i) is computed iteratively for i = 1, 2,··· , n. Assume
that we have computed pmrp(i) for some i. Let pmrp(i) = prev (t[i : l]). Obviously,
prev (t[i + 1 : l]) is a prefix of the string represented by pmrp(i + 1). Thus, in order to
compute pmrp(i + 1), we should extend the prefix prev (t[i + 1 : l]) = psp(prev (t[i : l]))
in PPH (t) until we found l(cid:48) such that node prev (t[i + 1 : l(cid:48)]) does not have outgoing
edge labeled with prev (t[i + 1 :])[l(cid:48) − i + 1] and set pmrp(i + 1) = prev (t[i + 1 : l(cid:48)]). In
this time, we need re-compute prev (t[i + 1 :]) by replacing prev (t[i + 1 :])[j] with 0 if we
found that prev (t[i + 1 :])[j] ≥ j. The total number of extending prev (t[i + 1 : l]) in the
algorithm is at most n because both i and l always increase in each iteration. In each
9
Algorithm 2: Augmented parameterized position heap construction algorithm
Input: A p-string t ∈ (Σ ∪ Π)n and PPH (t)
Output: An augmented parameterized position heap APPH (t)
1 let t[n + 1] = $ where $ is a symbol that does not appear in t elsewhere;
2 currentNode = root;
3 l = 1;
4 for i = 1 to n do
5
c = normalize(prev (t)[l], l − i);
while child (currentNode, c) (cid:54)= null do
6
7
8
9
10
11
currentNode = child (currentNode, c);
l = l + 1;
c = normalize(prev (t)[l], l − i);
pmrp(i) = currentNode;
currentNode = psp(currentNode);
iteration, operations such as traversing a child node can be done in O(log (Σ + Π)).
Therefore, we can get the following theorem.
Theorem 2. Parameterized maximal-reach pointers for PPH (t) can be computed in
O(n log (Σ + Π)) time.
4.4 Parameterized Pattern Matching with Augmented Parameterized
Position Heaps
Ehrenfeucht et al. [8] and Kucherov [13] split a pattern p into segments q1, q2,··· , qk,
then compute occurrences of q1q2 ··· qj iteratively for j = 1,··· , k. The correctness
depends on a simple fact that for strings x = t[i : i + x − 1] and y = t[i + x :
i + x + y − 1] implies xy = t[i : i + xy − 1]. However, when x, y, and t are p-strings,
prev (x) = prev (t[i : i + x − 1]) and prev (y) = prev (t[i + x : i + x + y − 1]) does
not necessarily implies prev (xy) = prev (t[i : i + xy − 1]). Therefore, we need to modify
the matching algorithm for parameterized strings.
Let x, y and w be p-strings such that w = xy, prev (x) = prev (w[: x]) and
prev (y) = prev (w[x + 1 :]). Let us consider the case that prev (xy) (cid:54)= prev (w). From
prev (x) = prev (w[: x]) and prev (y) = prev (w[x + 1 :]), x and y have the same struc-
ture of w[: x] and w[x + 1 :], respectively. However, the parameter symbols those
are prev-encoded into 0 in prev (y) and prev (w[x + 1 :]), might be encoded differ-
ently in prev (xy) and prev (w), respectively. Therefore, we need to check whether
prev (xy)[x + i] = prev (w)[x + i] if prev (y)[i] = 0. Given prev (xy) and the set of
positions of 0 in prev (y), Z = {i 1 ≤ i ≤ y such that prev (y)[i] = 0}. We need to
verify whether prev (xy)[x + i] = prev (w)[x + i] or not for i ∈ Z. Since the size of Z is
at most Π, this computation can be done in O(Π) time.
A pseudocode of proposed matching algorithm for the parameterized pattern matching
problem is shown in Algorithm 3. Des APPH (t)(u) denotes the set of all descendants of
node u in APPH (t) including node u itself. The occurrences of p in t have the following
properties on APPH (t).
10
Algorithm 3: Parameterized pattern matching algorithm with APPH
Input: t ∈ (Σ ∪ Π)n , p ∈ (Σ ∪ Π)m, and APPH (t)
Output: The list ans of position i such that prev (p) = prev (t[i : i + m − 1])
1 let w be the longest prefix of prev (p) represented in APPH (t) and u be the node
2 if w = m then
represents w;
3
4
5
6
v = root;
for i = 1 to m do
v = child (v, prev (p)[i]);
if pmrp(v) ∈ Des APPH (t)(u) then add prim(v) to ans;
add all primary and sedondary position of decendants of u to ans;
7
8 else
9
v = root;
i = 1, j = 1;
while i ≤ w do
v = child (v, prev (p)[i]);
i = i + 1;
if pmrp(v) = u then add prim(v) to ans;
while i (cid:54)= m do
j = i, v = root;
Z = empty list;
while i (cid:54)= m do
c = normalize(prev (p)[i], i − j);
if child (v, c) = null then break;
if c = 0 then add i to Z;
v = child (v, c);
i = i + 1;
if v = root then return empty list;
foreach i(cid:48) ∈ ans do
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
if i = m then
if pmrp(i(cid:48) + j − 1) /∈ Des APPH (t)(v) then remove i(cid:48) from ans;
if pmrp(i(cid:48) + j − 1) (cid:54)= v then remove i(cid:48) from ans;
if normalize(prev (t)[i(cid:48) + Z[k] − 1], Z[k] − 1) (cid:54)= prev (p)[Z[k]] then
else
for k = 1 to Z do
remove i(cid:48) from ans;
32
33 return ans;
Lemma 4. If prev (p) is represented in APPH (t) as a node u then p occurs at position
i iff pmrp(i) is u or its descendant.
Proof. Let u be a node represents prev (p). Assume p occurs at position i in t and
represented in APPH (t) as prev (t[i : k]). Since either prev (t[i : k]) is a prefix of prev (p)
or prev (p) is a prefix of prev (t[i : k]), then i is either an ancestor or descendant of u.
11
(a)
(b)
Figure 4: Examples of finding occurrence positions of a pattern using an augmented
parameterized position heap PPH (xaxyxyxyyaxyxy). (a) Finding xyxy (prev (xyxy) =
0022). (b) Finding axyx (prev (axyx) = a002).
For both cases pmrp(i) is a descendant of u, because p occurs at position i.
Next let i be a node such that pmrp(i) is a descendant of u and represents prev (t[i : k]).
In this case, prev (p) is a prefix of prev (t[i : k]). Therefore p occurs at i.
Lemma 5. Assume prev (p) is not represented in APPH (t). We can split p into
q1, q2,··· , qk such that qj is the longest prefix of prev (p[q1 ··· qj−1 + 1 :]) that is repre-
sented in APPH (t). If p occurs at position i in t, then pmrp(i +q1 ··· qj−1) is the node
prev (qj) for 1 ≤ j < k and pmrp(i + q1 ··· qk−1) is the node prev (qk) or its descendant.
Proof. Let p = q1q2 ··· qk occurs at position i in t. Since prev (q1) is a prefix of prev (p),
then pmrp(i) is the node that represents prev (q1) or its descendant. However, if pmrp(i)
is a descendant of node prev (q1), then we can extend q1 which contradicts with the
definition of q1. Therefore, pmrp(i) is the node represents prev (q1).
Similarly for 1 < j < k, prev (qj) is a prefix of prev (p[q1 ··· qj−1 + 1 :]) and occurs
at position i + q1 ··· qj−1 in t. Therefore, pmrp(i + q1 ··· qj−1) is the node represents
prev (qj). Last, since qk is a suffix of p, then pmrp(i + q1 ··· qj−1) can be the node
prev (qk) or its descendant.
Algorithm 3 utilizes Lemmas 4 and 5 to find occurrences of p in t by using APPH (t).
First, if prev (p) is represented in APPH (t) then the algorithm will output all position
i such that pmrp(i) is a node prev (p) or its descendant. Otherwise, it will split p into
q1q2 ··· qk and find their occurrences as described in Lemma 5. The algorithm also checks
whether prev (q1 ··· qj) occurs in t or not in each iteration as described the above.
Examples of parameterized pattern matching by using an augmented position heap
In Fig. 4 (a) we want
are given in Fig. 4. Let t = xaxyxyxyyaxyxy be the text.
to find the occurrence positions of a pattern p1 = xyxy in t.
In this case, since
prev (p1) = 0022 is represented in PPH (t), The algorithm outputs all positions i such
that pmrp(i) is the node 0022 or its descendants, those are 3, 4, 5, and 11. On the other
hand, Fig. 4 (b) shows how to find the occurrence positions of a pattern p2 = axyx
in t.
In this case, prev (p2) = a002 is not represented in PPH (t). Therefore, The
algorithm finds the longest prefix of prev (p2) that is represented in PPH (i), which
is prev (p2)[1 : 2] = a0. We can see that prmp(2) = pmrp(10) = a0, then we save
12
3, 13 6 2 100a1, 14 8 7 94, 125, 110011122a ԋ 3, 13 6 2 100a1, 14 8 7 94, 125, 110011122a ԋ 3, 13 6 2 100a1, 14 8 7 94, 125, 110011122a ԋ positions 2 and 10 as candidates to ans. Next, The algorithm finds the node that
represents the longest prefix of prev (p2[3 :]) = 00 which is prev (p2[3 :]) = 00 itself.
Since both of pmrp(2 + p2[1 : 2]) = pmrp(4) and pmrp(10 + p2[1 : 2]) = pmrp(12) is
descendants of the node 00, prev (t[2 : 5][3]) = prev (t[10 : 13][3]) = prev (p2)[[3]] = 0, and
prev (t[2 : 5][4]) = prev (t[10 : 13][4]) = prev (p2)[4] = 2, then the algorithm outputs 2 and
10.
The time complexity of the matching algorithm is as follow.
Theorem 3. Algorithm 3 runs in O(m log (Σ + Π) + mΠ + occ) time.
Proof. It is easily seen that we can compute line 4 to 7 in O(m log (Σ + Π) + occ) time.
Assume that p can be decomposed into q1, q2, ··· , qk such that q1 is the longest prefix of
p and qi is the longest prefix of prev (p[q1 ··· qj−1 + 1 :]) represented in APPH (t). The
loop for line 15 consists of k − 1 iterations. In the loop line 18 in j-th iteration, qj+1
is extended up to reach qj+1 length. This can be computed in O(qj+1 log (Σ + Π))
time. After k − 1 iterations, the total number of extending of qj+1 does not exceed m,
j=2qj < m. In the loop for line 25, the algorithm verifies elements of ans.
because Σk
In j-th iteration, the size of ans is at most qj. Thus, after k − 1 iterations, the total
number of elements verified in line 25 does not exceed m by the same reason for that of
line 18. In each verification in line 25, the number of checks for line 27 and 29 is at most
qj. Therefore, it can be computed from line 25 to 32 in O(mΠ) time.
5 Conclusion and Future Work
For the parameterized pattern matching problem, we proposed an indexing structure
called a parameterized position heap. Given a p-string t of length n over a constant size
alphabet, the parameterized position heap for t can be constructed in O(n log (Σ + Π))
time by our construction algorithm. We also proposed an algorithm for the parameterized
pattern matching problem. It can be computed in O(m log (Σ + Π) + mΠ + occ)
time using parameterized position heaps with parameterized maximal-reach pointers.
Gagie et al. [10] showed an interesting relationship between position heap and suffix
array of a string. We will examine this relation for parameterized position heap and
parameterized suffix array [7, 12] as a future work.
References
[1] B. S. Baker. A program for identifying duplicated code. Computing Science and
Statistics, 24:49 -- 57, 1992.
[2] B. S. Baker. A theory of parameterized pattern matching: algorithms and applica-
tions. In Proc. 25th annual ACM symposium on Theory of computing, pages 71 -- 80,
1993.
[3] B. S. Baker. Parameterized pattern matching: Algorithms and applications. Journal
of Computer and System Sciences, 52(1):28 -- 42, 1996.
13
[4] E. G. Coffman and J. Eve. File structures using hashing functions. Commun. ACM,
13(7):427 -- 432, 1970.
[5] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to Algorithms.
MIT press, 1990.
[6] M. Crochemore and W. Rytter. Jewels of Stringology: Text Algorithms. World
Scientific, 2003.
[7] S. Deguchi, F. Higashijima, H. Bannai, S. Inenaga, and M. Takeda. Parameterized
suffix arrays for binary strings. In Proceedings of the Prague Stringology Conference
2008, pages 84 -- 94, Czech Technical University in Prague, Czech Republic, 2008.
[8] A. Ehrenfeucht, R. M. McConnell, N. Osheim, and S.-W. Woo. Position heaps: A
simple and dynamic text indexing data structure. Journal of Discrete Algorithms,
9(1):100 -- 121, 2011.
[9] K. Fredriksson and M. Mozgovoy. Efficient parameterized string matching. Infor-
mation Processing Letters, 100(3):91 -- 96, 2006.
[10] T. Gagie, W.-K. Hon, and T.-H. Ku. New Algorithms for Position Heaps, pages
95 -- 106. Springer Berlin Heidelberg, Berlin, Heidelberg, 2013.
[11] D. Gusfield. Algorithms on Strings, Trees and Sequences: Computer Science and
Computational Biology. Cambridge University Press, 1997.
[12] T. I, S. Deguchi, H. Bannai, S. Inenaga, and M. Takeda. Lightweight Parameterized
Suffix Array Construction, pages 312 -- 323. Springer Berlin Heidelberg, Berlin,
Heidelberg, 2009.
[13] G. Kucherov. On-line construction of position heaps. Journal of Discrete Algorithms,
20:3 -- 11, 2013. StringMasters 2011 Special Issue.
[14] Y. Nakashima, T. I, S. Inenaga, H. Bannai, and M. Takeda. The Position Heap of
a Trie, pages 360 -- 371. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012.
[15] T. Shibuya. Generalization of a suffix tree for rna structural pattern matching.
Algorithmica, 39(1):1 -- 19, 2004.
[16] E. Ukkonen. On-line construction of suffix trees. Algorithmica, 14(3):249 -- 260, 1995.
14
|
1307.2724 | 1 | 1307 | 2013-07-10T09:13:35 | The technique of in-place associative sorting | [
"cs.DS"
] | In the first place, a novel, yet straightforward in-place integer value-sorting algorithm is presented. It sorts in linear time using constant amount of additional memory for storing counters and indices beside the input array. The technique is inspired from the principal idea behind one of the ordinal theories of "serial order in behavior" and explained by the analogy with the three main stages in the formation and retrieval of memory in cognitive neuroscience: (i) practicing, (ii) storage and (iii) retrieval. It is further improved in terms of time complexity as well as specialized for distinct integers, though still improper for rank-sorting.
Afterwards, another novel, yet straightforward technique is introduced which makes this efficient value-sorting technique proper for rank-sorting. Hence, given an array of n elements each have an integer key, the technique sorts the elements according to their integer keys in linear time using only constant amount of additional memory. The devised technique is very practical and efficient outperforming bucket sort, distribution counting sort and address calculation sort family of algorithms making it attractive in almost every case even when space is not a critical resource. | cs.DS | cs |
The technique of in-place associative sorting
A. Emre CETIN
July 3, 2018
Abstract
In the first place, a novel, yet straightforward in-place integer value-sorting al-
gorithm is presented. It sorts in linear time using constant amount of additional
memory for storing counters and indices beside the input array. The technique is
inspired from the principal idea behind one of the ordinal theories of "serial order
in behavior" and explained by the analogy with the three main stages in the forma-
tion and retrieval of memory in cognitive neuroscience: (i) practicing, (ii) storage
and (iii) retrieval.
It is further improved in terms of time complexity as well as
specialized for distinct integers, though still improper for rank-sorting.
Afterwards, another novel, yet straightforward technique is introduced which
makes this efficient value-sorting technique proper for rank-sorting. Hence, given
an array of n elements each have an integer key, the technique sorts the elements ac-
cording to their integer keys in linear time using only constant amount of additional
memory. The devised technique is very practical and efficient outperforming bucket
sort, distribution counting sort and address calculation sort family of algorithms
making it attractive in almost every case even when space is not a critical resource.
keywords: associative sort, in-situ permutation sort, stimulation sort, linear time sorting.
1
1
Introduction
The adjective "associative" derived from two facts where the first one will be realized with
the description of the technique. The second one is that, although it replaces all derivatives
of the content based sorting algorithms such as distribution counting sort [15,16], address
calculation sort [17 -- 22] and bucket sort [23, 24] on a RAM, it seems to be more efficient
on a"content addressable memory" (CAM) known as "associative memory" which in one
word time find a matching segment in tag portion of the word and reaches the remainder
of the word [29]. In the current version of associative sort developed on a RAM, the nodes
of the imaginary subspace (tagged words) and the integers of the array space (untagged
words) are processed sequentially which will be a matter of one word time for a CAM to
retrieve previous or next tagged or untagged word.
An integer value-sorting algorithm puts an array of integers into ascending or descend-
ing order by their values, whereas a rank-sorting algorithm puts an array of elements
(satellite information) into ascending or descending order by their numeric keys, each of
which is an integer. It is possible that a rank-sorting algorithm can be used in place of a
value-sorting algorithm, since if each element of the array to be sorted is itself an integer
and used as the key, then rank-sorting degenerates to value-sorting, but the converse is
not always true.
The technique described in this study is suitable for arrays where the elements are
laid out in contiguous locations of the memory. Zero-based indexing is considered while
accessing the elements, e.g., A[0] and A[n − 1] are the first and last elements of the array,
respectively, where n is the number of elements of the array.
Nervous system is considered to be closely related and described with the "serial order
in behavior" in cognitive neuroscience [6, 7] with three basic theories which cover almost
2
all abstract data types used in computer science. These are chaining theory, positional
theory and ordinal theory [8].
Chaining theory is the extension of reflex-chaining or response-chaining theory, where
each response becomes the stimulus for the next. From an information processing per-
spective, comparison based sorting algorithms that sort the arrays by making a series of
decisions relying on comparing keys can be classified under chaining theory. Each com-
parison becomes the stimulus for the next. Hence, keys themselves are associated with
each other. Some important examples are quick sort [9], shell sort [10], merge sort [11]
and heap sort [12].
Positional theory assumes order is stored by associating each element with its position
in the sequence. The order is retrieved by using each position to cue its associated
element. Conventional (Von Neumann) computers store and retrieve order using this
method, through routines accessing separate addresses in memory. Content-based sorting
algorithms where decisions rely on the contents of the keys can be classified under this
theory. Each key is associated with a position depending on its content. Some important
examples are distribution counting sort [15, 16], address calculation sort [17 -- 22], bucket
sort [23, 24] and radix sort [23 -- 26].
Ordinal theory assumes order is stored along a single dimension, where that order is
defined by relative rather than absolute values on that dimension. Order can be retrieved
by moving along the dimension in one or the other direction. This theory need not assume
either the item-item nor position-item associations of chaining and positional theories
respectively [8].
One of the ordinal theories of serial order in behavior is that of Shiffrin and Cook
[27] which suggests a model for short-term forgetting of item and order information of
3
the brain. It assumes associations between elements and a "node", but only the nodes
are associated with one another. By moving inwards from nodes representing the start
and end of the sequence, the associations between nodes allow the order of items to be
reconstructed [8].
The first technique presented in this study is in-place associative integer sorting [1 --
4].
Inspired from the ordinal model of Shiffrin and Cook, the technique assumes the
associations are between the integers in the array space and the nodes in an imaginary
linear subspace (ILS) that spans a predefined range of integers. The range of the integers
spanned by the ILS is upper bounded by the number of integers n but may be smaller
and can be located anywhere provided that its boundaries do not cross over that of the
array. This makes the technique in-place, i.e., beside the input array, only a constant
amount of memory locations are used for storing counters and indices. An association
between an integer of the array space and the ILS is created by a node using a monotone
bijective hash function that maps the integers in the predefined interval to the ILS. When
a particular distinct integer is mapped to the ILS, a node is created reserving all the bits
of the integer except for the most significant bit (MSB) which is used to tag the word as
a node of the ILS for interrogation purposes. The reserved bits become the record of the
node which then be used to count (practice) other occurrences of the particular integer
that created the node. When all the key of the predefined interval are practiced, the nodes
can be stored at the beginning of the array (short-term memory) retaining their relative
order together with the information (cue) required to construct the sorted permutation
of the practiced interval. Afterwards, the short-term memory is processed and the sorted
permutation of the practiced interval is retrieved over the array space in linear time using
only constant amount of additional memory.
4
Another ordinal theory is the original perturbation model of Estes [28]. Although
proposed to provide a reasonable qualitative fit of the forgetting dynamics of the short
term memory [8] in cognitive neuroscience, the principle idea behind the method is that the
order of the elements is inherent in the cyclic reactivation of the elements, i.e., reactivations
lead to reordering of the elements.
In-place associative integer sorting technique is an efficient in-place integer value-
sorting algorithm, though not suitable for rank-sorting. Therefore, in-place associative
permutation sort [5] technique is developed combining the principle idea behind the origi-
nal perturbation model with the technique of associative integer sorting making it suitable
for rank-sorting.
2 Definitions
The definition of rank-sorting is: given an array S of n elements, S[0], S[1], . . . , S[n − 1]
each have an integer key, the problem is to sort the elements of the array according to
their integer keys. To prevent repeating statements like "key of the element S[i]", S[i] is
used to refer the key.
The definition of value-sorting is: given an array S of n integers, S[0], S[1], . . . , S[n−1],
the problem is to sort the integers according to their values. In other words, the elements
of the array are integers and used as the keys.
To prevent confusion, the term "integer" is used while describing value-sorting tech-
niques, whereas "key" is used for rank-sorting.
The notations used throughout the study are:
(i) Universe of keys is assumed U = [0 . . . 2w − 1] where w is the fixed word length.
(ii) Maximum and minimum keys of an array are, max(S) = max(aa ∈ S) and
5
min(S) = min(aa ∈ S), respectively. Hence, range of the keys is, m = max(S) −
min(S) + 1.
(iii) The notation B ⊂ A is used to indicated that B is a proper subset of A.
(iv) For two arrays S1 and S2, max(S1) < min(S2) implies S1 < S2.
3
In-place Associative Integer Sorting
The most critical phase of associative integer sorting is derived from the cycle leader
permutation (in-situ permutation) approach. Hence, a separate section is devised for it.
3.1 Cycle Leader Permutation
Given n distinct integer keys S[0 . . . n−1] each in the range [δ, δ+m−1] where δ = min(S),
if m = n and S is the sorted permutation, then there is a bijective relation i = S[i] − δ
between each key and its position. From contradiction, i 6= S[i] − δ implies that the key
S[i] is not at its exact position.
Its exact position can be calculated by j = S[i] − δ.
Therefore, the simple monotone bijective hash function j = S[i] − δ that maps the keys
to j ∈ [0, n − 1] can sort the array in O(n) time using O(1) constant space. This is cycle
leader permutation where S is re-arranged by following the cycles of a permutation π.
First S[0] is sent to its final position π(0) (calculated by j = S[i] − δ). Then the element
that was in π(0) is sent to its final position π(π(0)). The process proceeds in this way
until the cycle is closed, that is until the key addressing the first position is found which
means that the association 0 = S[0] − δ is constructed between the key and its position.
Then the iterator is increased to continue with the key of S[1]. At the end, when all the
cycles of S[i] for i = 0, 1.., n − 1 are processed, all the keys are in their exact position and
the association i = S[i] − δ is constructed between the keys and their position resulting
6
in the sorted permutation of the array.
3.2
In-place Associative Distinct Key Sorting
If we look at the cycle leader permutation closer, we can interpret the technique en-
tirely different. We are indeed creating an imaginary linear subspace Im[0 . . . n − 1] over
S[0 . . . n − 1] where the relative basis of this imaginary linear subspace (ILS) coincides
with that of the array space in the physical memory. The ILS spans a predefined interval
of the range of keys and this interval is upper bounded by the number of keys n. Since the
range of the keys m is equal to n, it spans the entire range of the keys. The association
between the array space and the ILS is created by a node using the monotone bijective
hash function j = S[i] − δ that maps a particular key to the ILS. Since ILS is defined
over the array space, mapping a distinct key to the ILS is just an exchange operation
from where a node is created. When a node is created for a particular key, the redun-
dancy due to the association between the key and the position of the node releases the
word allocated for the key in the physical memory. Hence, we can clear the node and
set its tag bit, for instance its most significant bit (MSB) to discriminate it as a node
for interrogation purposes, and use the remaining w − 1 bits of the node for any other
purpose. When we want the key back to array space from the ILS, we can use the inverse
of hash function and get the key back by S[i] = i + δ to the array space through the
node. However, we don't use free bits of a node for other purposes during cycle leader
permutation because it is known that all the keys are distinct and only one key can be
mapped to a location creating a node. Therefore, instead of tagging the word as node
using its MSB, we use the key itself to tag the word "implicitly" as node, since when a
key is mapped to the imaginary subspace, it will always satisfy the monotone bijective
7
hash function i = S[i] − δ. Hence, the keys are "implicitly practiced" in this case.
This interpretation immediately motivates a rank-sorting algorithm for distinct keys.
Consider the problem of sorting n distinct keys S[0 . . . n−1] each in the range [δ, δ+m−1].
If m > n, it is not possible to construct a monotone bijective hash function that maps all
the keys of the array into j ∈ [0, n−1] without collisions and additional storage space [33].
However, a bijective hash function can be constructed as a partial function that assigns
all the keys in the range [δ, δ + n − 1] to exactly one element in j ∈ [0, n − 1]. Hence,
a cycle leader permutation only for the keys in the range [δ, δ + n − 1] with the hash
function j = S[i] − δ can sort these keys in the ILS Im[0 . . . n − 1] over S[0 . . . n − 1].
The hashed keys are said "implicitly practiced" and always satisfy the monotone bijective
hash function i = S[i] − δ.
After implicitly practicing the keys in the predefined interval, the next step is to
separate the array into practiced and unpracticed keys. This is simply a partitioning
problem where we store implicitly practiced keys at the beginning of the array. However,
instead of using a pivot for partitioning, we partition the keys that satisfy the hash
function i = S[i] − δ. Hence, we obtain a simple two step sorting algorithm for distinct
keys:
(i). find min(S) and max(S) and initialize δ = min(S), δ′ = max(S), nd = 0; n′
d = 0;
(ii). implicitly practice all the distinct keys of the interval [δ, δ+n−1]. We can count the
number of implicitly practiced keys in nd and find the minimum of the unpracticed
keys in δ′ ;
(iii). store all the implicitly practiced keys (that satisfy i = S[i] − δ) at the beginning of
the array;
8
(iv). If nd = n exit. Otherwise, set S = S[nd − 1 . . . n − 1], n = n − nd, δ = δ′,
δ′ = max(S), reset nd and goto step (ii).
The algorithm sorts an array of n elements S[0 . . . n−1] each have a distinct integer key
in the range [0, m − 1] using O(1) extra space in O(n + m) time for the worst, O(m) time
for the average (uniformly distributed keys) and O(n) time for the best case. Therefore,
the ratio m
n defines the efficiency (time-space trade-offs) letting very large arrays to be
sorted in-place.
The algorithm is surprisingly effective and efficient. Comparisons with Ω(n log n) quick
sort [9, 25] and merge sort [13] and heap sort [12, 14] which take O(n log n) time on all
inputs showed that associative sort for distinct keys is superior than all (up to 20 times)
provided that m
n ≤ c log n where c ≈ 4 for heap sort and merge sort. Quick sort gave
worser results (c ≈ 8) for distinct keys. These results are consistent with m calculated
theoretically making average case time complexity of the algorithm less than lower-bound
of comparison-based sorting algorithms, i.e., O(m) < Ω(n log n). Another very important
meaning of this inequality is that, since it does not require additional space other than
a constant amount, no matter how large is the array, the proposed algorithm will sort
faster than all provided that m = O(n log n).
Comparisons with value-sorting version of distribution counting sort (frequency count-
ing sort [23]) showed that associative sort for distinct keys is superior in every case. This
is expectable considering memory allocation overload of distribution counting sort since
time-complexities of both algorithms are the same. The performance of the algorithm is
even asymptotically better than 2 lines of code referred in textbooks for sorting n distinct
integers from [0 . . . n − 1] with indexing an auxiliary array B of the same size as the input
A with keys of A by B[A[i]] = A[i] for i = 0, 1, . . . , n − 1, and then reconstructing sorted
9
permutation of A back by A[i] = B[i] for i = 0, 1, . . . , n − 1, which is possibly due to
time-consuming memory allocation of the auxiliary array.
It is compared with radix sort [23 -- 26] and bucket sort, as well. The results showed
that it is superior than radix sort when m
n ≤ 8 and faster than bucket sort for n distinct
integer keys S[0...n − 1] each in the range [0, n − 1].
Finally, the dependency of the efficiency of the technique on the distribution of the keys
is only O(n) which means it replaces all the methods based on address calculation [17 --
22], that are known to be very efficient when the keys have known (usually uniform)
distribution and require additional space more or less proportional to n [26].
3.3
In-place Associative Integer Sorting
The technique introduced above that sorts distinct keys can be generalized to arrays with
repeating integers if we consider using released w − 1 bits of a node for other purposes.
The generalized version becomes a value-sorting algorithm whereas the former was a
rank-sorting algorithm. Henceforth, the term "integer" will be used instead of "key".
The main difficulties of all distributive sorting algorithms is that, when the integers
are distributed using a hash function according to their content, several integers may be
clustered around a loci, and several may be mapped to the same location. These problems
are solved by inherent three basic phases of in-place associative integer sorting [1] namely
(i) practicing, (ii) storage and (iii) retrieval.
We will consider the problem of sorting n integers S[0 . . . n − 1] each in the range
[δ, δ + m − 1]. If m > n, it is not possible to construct a monotone bijective hash function
that maps all the integers of the array into j ∈ [0, n − 1] without collisions and additional
storage space. However, a bijective hash function can be constructed as a partial function
10
that assigns all the integers in the range [δ, δ+n−1] to exactly one element in j ∈ [0, n−1].
3.3.1 Practicing
It is assumed that associations are between the integers in the array space and the nodes
in an imaginary linear subspace (ILS) that spans a predefined range of integers. The ILS
can be defined anywhere on the array space S[0 . . . n − 1] provided that its boundaries do
not cross over that of the array. The range of the integers spanned by the ILS is upper
bounded by the number of integers n revealing the asymptotic power of the technique with
increasing n with respect to the range of integers. However, an ILS may be smaller and
can be located anywhere over the array space making the technique in-place, i.e., beside
the input array, only a constant amount of memory locations are used for storing counters
and indices. An association between an integer and the ILS is created by a node using a
monotone bijective hash function that maps the integers in the predefined interval to the
ILS. Therefore, the monotone bijective hash function is a partial function that assigns to
each distinct integer in a predefined interval to exactly one node of the ILS. Since ILS is
defined over the array space, mapping a distinct integer to the imaginary subspace is just
an exchange operation from where a node is created. This is "practicing a distinct integer
of an interval". Once a node is created, the redundancy due to the association between the
integer and the position of the node (the position where the integer is mapped) releases
the word allocated to the integer in the physical memory except for most significant bit
(MSB) which tags the word as a node for interrogation purposes. Hence, the integer is
said to be sent to the ILS thorugh the node. Nevertheless, the tag bit discriminates the
word as a node and the position of the node lets the integer be retrieved back through
the node from the ILS using the inverse hash function. This is "integer retrieval" through
the node from ILS. All the bits of the node except the tag bit can be cleared and used
11
to store any information. Hence, they are the "record" of the node and the information
stored in the record is the "cue" by which cognitive neuro-scientists describe the way that
the brain recalls the successive items in an order during retrieval. For instance, it will be
foreknown from the tag bit that a node has already been created while another occurrence
of that particular integer is being practiced providing the opportunity to count other
occurrences using the record. The process of counting an other occurrence of a particular
integer is "practicing an idle integer of an interval". Repeating this process for all the
integers of an interval is "practicing an interval", i.e., rehearsing used by cognitive neuro-
scientists to describe the way that the brain manipulates the sequence before storing in
a short (or long) term memory. Practicing all the integers of an interval does not need
to alter the value of other occurrences. Only the first occurrence is altered while being
practiced from where a node is created. All other occurrences of that particular integer
remain in the array space but become meaningless. That is why they are "idle integers".
Furthermore, practicing does not need to alter the position of idle integers as well, unless
another distinct integer creates a node exactly at the position of an idle integer while being
practiced. In such a case, the idle integer is moved to the former position of the integer
that creates the new node in its place. This makes associative sort unstable, i.e., equal
integers may not retain their original relative order. However, an imaginary subspace can
create other subspaces and associations using the idle integers that were already practiced
by manipulating either their position or value or both. Hence, a part of linear algebra
and related fields of mathematics can be applied on subspaces to solve such problems.
From information processing perspective, practicing is a derivative of cycle leader
permutation (Section 3.1) where the nodes and the keys that are out of the practiced
interval are treated specially. Processing the array from left to right for i = 0, 1, . . . n − 1,
12
if S[i] is a node (tagged word) or a key that is out of the practiced interval (S[i] − δ ≥ n)
increase the iterator i and start over with S[i]. Otherwise, it is an integer that is to be
practiced (S[i] − δ < n). Therefore, the monotone bijective hash function j = S[i] − δ
maps the keys to j ∈ [0, n − 1]. If the integer S[j] at the target is a node, then increase
the record of the node by one, increase the iterator i by one and start over with S[i]. This
is "practicing an idle integer of an interval", because the idle integer S[i] is practiced by
its node S[j]. On the other hand, if S[j] is not a node, then S[i] is the first occurrence
that will create the node. Therefore, copy S[j] over S[i] and then clear S[j] and set its
MSB making it a node of ILS releasing its w − 1 bits free. This is "practicing a distinct
integer of an interval" where a node is created with an empty record of w − 1 bits. In
such a case, j < i implies that the integer that was at S[j] (now at S[i]) was processed
before. Therefore, increase the iterator i by one and start over with S[i]. On the other
hand, j ≥ i implies that the integer that was at S[j] (now at S[i]) has not been processed
yet. Hence, start over with S[i] without increasing the iterator i to continue with it.
3.3.2 Storage
Once all the integers in the predefined interval are practiced, the nodes are spread over
the ILS depending on the distribution of the integers with relative order. The next
step is to store the nodes in a systematic way closing the distance between them to
a direction retaining their relative order with respect to each other. This is the storage
phase of associative sort where the received, processed and combined information required
to construct the sorted permutation of the practiced interval is stored in the short-term
memory (beginning of the array). When the nodes are moved, it is not possible to retain
the associations between the nodes of the ILS and the integers of the array space because
the position of each node cues the recall of the corresponding integer and retrieve it
13
from the ILS using the inverse hash function. This motivates the idea to further use the
record of a node to store the node's former position, or maybe its relative position with
respect to the ILS or how much that node is moved relative to its absolute or relative
position or the other integers. Unfortunately, this requires a record of a particular node
is enough to store both the positional information of the node and the number of idle
integers practiced by that node. This is statistically impossible, but, as mentioned before,
further associations can be created using the idle integers that were already practiced by
manipulating either their position or value or both. Hence, if the record is enough, it
can store together the positional information and the number of practiced idle integers,
whereas an idle integer can be associated accompanying the node to supply additional
space to store the positional information if the record is not enough.
Let us assume for a moment that our universe of integers is U = [0 . . . 2w−1 − 1] where
w is the fixed word length and n ≤ 2w−1. We know that other occurrences of a particular
integer can be counted (practiced) using w − 1 bits (record) of a node. If we decide to
write the absolute position of a node into its record during storage, we need log n bits
of the record. Hence, it is logical to think that we can count up to 2w−1−log n − 1 idle
integers with the record of a node during practicing. Fortunately, this is not the case.
We can count using all w − 1 bits of the record during practicing, and while storing the
nodes at the beginning of the array (short-term memory), we can get an idle integer
immediately after the node that has practiced at least 2w−1−log n idle integers and write
the absolute position of that node over the accompanying idle integer. In such a case, the
record of the node (predecessor of the idle integer) only stores the number of practiced idle
integers. This definition immediately reminds one of the aforementioned main difficulties
of all distributive sorting algorithms. When the keys are distributed using a hash function
14
according to their content, several of them may be clustered around a loci. In such a case,
how an idle integer can be inserted immediately after a particular node if there are several
nodes immediately before and after that particular node during storing? The answer is
in the pigeonhole principle. Pigeonhole principle says that,
Corollary 3.1. Given n ≤ 2w−1 integers S[0 . . . n − 1], the maximum number of distinct
integers that may occur contemporary in S at least 2w−1−log n times is,
⌈
n
2w−1−log n ⌉
(3.1)
Hence, if the size of the array is say n = 2w−1, the maximum number of distinct integers
that may occur contemporary in S at least 1 time is n. But the node itself represents the
first occurrence which creates it. Therefore,
Corollary 3.2. The maximum number of nodes that each can practice at least 2w−1−log n
idle integers and hence need an idle integer immediately after itself during storage is equal
to,
and upper bounded by n/2.
This means that,
ǫ = ⌈
n/2
2w−1−log n ⌉
(3.2)
Corollary 3.3. If the integers that are in the predefined interval [δ, δ + n − ǫ − 1] are
practiced to Im[ǫ . . . n − 1] over S[ǫ . . . n − 1] where ǫ ∈ [0, n
2 ] is calculated by Eqn.3.2,
then there will be ǫ integers in S[0 . . . ǫ − 1] either idle or unpracticed (out of the practiced
interval) which will prevent collisions while inserting idle integers immediately after the
nodes that has practiced at least 2w−1−log n idle integers during storage.
15
3.3.3 Retrieval
Finally, the sorted permutation of the practiced interval is constructed in the array space,
using the information stored in the short-term memory. This is the retrieval phase of
associative sort. It is known that if the record is enough, it stores both the position of
the node and the number of practiced idle integers.
If not, an associated idle integer
accompanying the node stores the position of the node, whereas the record (predecessor
of the idle integer) stores the number of practiced idle integers. If the number of occur-
rences of a particular integer is ni, then there are ni − 1 idle integers in the array space.
But the nodes represent the first integers that are mapped into the imaginary subspace
through themselves. If all the idle integers are grouped on the right side of the short-term
memory, then one can process the information in the short-term memory from right to
left and distinguish whether there is an idle integer (untagged word) accompanying its
predecessor (the node on the left side of the idle integer). An idle integer implies that it
is accompanying the node on its left for additional storage. In such a case, the positional
information is read from the idle integer, whereas the number of practiced idle integers
is read from the record of the node.
If there is not an idle integer accompanying the
node, both the positional information and the number of practiced idle integers are read
from the record of the node. Afterwards, the positional information cues the recall of the
integer using the inverse hash function. This is "integer retrieval" from ILS. Hence, the
retrieved integer can be copied on the array space as many as it occurrs. At this point,
we have two options: sequential and recursive versions which will be described next.
16
3.3.4 Overall Algorithm
Consider n ≤ 2w−1 integers S[0...n − 1] each in the range [0, 2w−1 − 1]. All the integers
in the range [δ, δ + n − ǫ − 1] will be practiced to Im[ǫ, n − 1] over S[ǫ, n − 1] where
δ = min(S) and ǫ is calculated using Eqn. 3.2. Assume that there are nd nodes, nc idle
integers practiced by those nd nodes and n′
d = n − (nd + nc) unpracticed integers that are
out of the practiced interval. All these can be counted during practicing. Furthermore,
the minimum δ′ of the unpracticed integers can be found as well during practicing.
While nd nodes are being stored at the beginning of the array (short-term memory)
closing the distance between them in order of precedence, if the record of a node is enough,
i.e., the node has practiced less than 2w−1−log n idle integers, we write the absolute (former)
position of the node into its record together with the number of practiced idle integers.
Otherwise, we search to the right and get the first idle integer immediately after the node
and write the absolute position of that node over the idle integer. Let us assume that ǫ′
nodes are counted that have practiced at least 2w−1−log n idle integers and needed an idle
integer during storage. At the end, nd nodes and ǫ′ idle integers are stored in the short-
term memory S[0 . . . nd + ǫ′ − 1] with the necessary information required to construct the
sorted permutation of the practiced interval. On the other hand, nc − ǫ′ idle integers and
n′
d = n − (nd + nc) unpracticed integers are distributed disorderly in S[nd + ǫ′ . . . n − 1].
3.3.5 Sequential Version
Selecting the pivot equal to δ + n−ǫ−1, if nc −ǫ′ idle integers and n′
d unpracticed integers
that are distributed disorderly in S[nd+ǫ′ . . . n−1] are partitioned, then nc−ǫ′ idle integers
come immediately after the short-term memory resulting in S[0 . . . nd + nc − 1]. Hence, it
is immediate from here that the information in the short-term memory S[0 . . . nd + ǫ′ − 1]
17
can be processed from right to left backwards and the integers practiced by each node
can be copied over S[0 . . . nd + nc − 1] right to left backwards without collision with the
short-term memory.
3.3.6 Recursive Version
We can recursively practice and store saving nd, ǫ′ and δ in stack space. Although the
exact number of integers to be sorted in the next level of recursion is n′
d, the overall
number of integers in that recursion is n = n′
d + nc − ǫ′ where nc − ǫ′ of them are idle
integers of the previous recursion and meaningless. However, these idle integers increase
the interval of range of integers spanned by the ILS improving the overall time complexity
in each level of recursion. The recursion can continue until no any integer exists. In the
last recursion, retrieval phase can begin to construct the sorted permutation of nd + nc
integers from nd + ǫ′ records stored in the short term memory S[0 . . . nd + ǫ′ − 1] of that
recursion and expand over S[0 . . . n − 1] right to left backwards. Each level of recursion
should return the total number of integers copied on the array to the higher level to
let it know where it will start to expand its interval. It should be noticed that, in the
recursive version of the technique, there is no need to partition nc − ǫ′ idle integers from
n′
d unpracticed integers. Hence, one step is canceled improving the overall efficiency.
3.3.7 Relaxing the Restrictions
The technique of associative sorting is explained restricting the universe of integers to
U = [0 . . . 2w−1 − 1] where w is the fixed word length and n ≤ 2w−1.
When an integer is first practiced, a node is created releasing w bits of the integer
free. One bit is used to tag the word as a node. Hence, it is reasonable to doubt that the
tag bit limits the universe of integers because all the integers should be untagged and in
18
the range [0, 2w−1 − 1] before being practiced. Of course we always have the option to use
additional n bits to tag the nodes. However, we can,
(i) partition S into 2 disjoint sublists S1 < 2w−1 ≤ S2 in O(n) time with well known
in-place partitioning algorithms as well as in a stable manner with [30],
(ii) shift all the integers of S2 by −2w−1, sort S1 and S2 associatively and shift S2 by
2w−1.
There are other methods to overcome this problem. For instance,
(i) sort the sublist S[0 . . . (n/ log n) − 1] using the optimal in-place merge sort [31],
(ii) compress S[0 . . . (n/ log n) − 1] by Lemma 1 of [32] generating Ω(n) free bits,
(iii) sort S[(n/ log n) . . . n − 1] associatively using Ω(n) free bits as tag bits,
(iv) uncompress S[0 . . . (n/ log n)−1] and merge the two sorted sublists in-place in linear
time by [31].
If practicing a distinct integer lets us to use its w − 1 bits to practice other occurrences
of that particular integer, we have w − 1 free bits by which we can count up to 2w−1
occurrences including the node that represents the first integer that created the node.
Hence, it is reasonable to doubt again that there is another restriction on the size of
the arrays, i.e., n ≤ 2w−1 under the assumption that an integer may always occur more
than 2w−1 times for an array of n > 2w−1. But an array can be divided into two parts
in O(1) time and those parts can be merged in-place in linear time by [31] after sorted
associatively.
3.3.8 Complexity
From complexity point of view, associative sort shows similar characteristics with distri-
bution counting sort [15, 16] and bucket sort [23, 24]. It sorts n integers S[0 . . . n − 1] each
in the range [0, m − 1] using O(1) extra space in O(m + n) time for the worst, O(m) time
19
for the average (uniformly distributed integers) and O(n) time for the best case. The
ratio m
n defines the efficiency (time-space trade-offs) of the algorithm letting very larges
arrays to be sorted in-place.
The complexity of the algorithm depends on the number and the range of the integers.
It is known that the algorithm is capable of sorting in each iteration (or recursion) the
integers in the range [δ, δ + n − ǫ − 1] where ǫ is defined by Eqn. 3.2 and upper bounded by
n/2. If we restrict our problem to sorting n integers S[0 . . . n−1] each in the range [0, n−1],
the worst case is when n = 2w−1 and the complexity is the recursion T (n) = T ( n
2 ) + O(n)
yielding T (n) = O(n). On the other hand, if we look at Eqn. 3.2 closer, we see that ǫ = 0
when log n < w
2 . This means that, the best case is when n < 2w/2 which implies ǫ = 0
and the complexity is exactly T (n) = O(n) in this case.
For the general case, we should consider the problem of sorting n integers S[0 . . . n − 1]
each in the range [δ, δ + m − 1] where m > n. Practically it is always possible that n − 1
of the integers be in the range [δ, δ + n − ǫ − 1]. This is the best case and the algorithm
sorts the integers in O(n) time. On the contrary, fixing ǫ always to its maximum n/2, if
there is only 1 integer available in each practiced interval until the last, in any jth step,
the only integer s ∈ S that will be sorted satisfies s < jn−(j−1)
2
which implies that the
last alone integer satisfies s < jn−(j−1)
2
≤ m from where we can calculate j by j ≤ 2m−1
n−1 .
In such a case, the time complexity is O(n) + O(n − 1) + . . . + O(n − j) < (j + 1)O(n)
yielding O(2m + n).
The average case is more difficult to estimate. However, fixing ǫ always to its maximum
n/2 will let us to assert that the integers in the range [δ, δ + n
2 − 1] will be sorted in each
step. On the other hand, if the integers are uniformly distributed, defining β = m
n , we
can say that n
2β integers will be in the range [δ, δ + n
2 − 1] in each step. Therefore, the
20
algorithm will sort n
2β of the integers in O(n) time in each pass. This will continue until
all the integers are sorted and the overall time complexity is,
O(n)(cid:0)
(2β − 1)
2β
+
(2β − 1)2
(2β)2 + . . . +
(2β − 1)k−1
(2β)k−1
(cid:1)
(3.3)
which means the algorithm is upper bounded by 2βO(n) or 2O(m) in the average case.
3.3.9 Empirical Tests
Practical comparisons for 1 million 32 bit integers with quick sort showed that associative
sort is roughly 2 times faster for uniformly distributed integers when m = n. When
m
n = 10 performances are same. When m
n = 1
10 associative sort becomes more than 3
times faster than quick sort.
If the distribution is exponential, associative sort shows
better performance up to m
n ≈ 25 when compared with quick sort.
Practical comparisons for 1 million 32 bit integers showed that radix sort is 2 times
faster for uniformly distributed integers when m = n. However, associative sort is slightly
better than radix sort when m
n = 1
10 . Further decreasing the ratio to m
n = 1
100 , associative
sort becomes more than 2 times faster than radix sort.
Practical comparisons for 1 million 32 bit integers showed that associative sort is 2
times faster than bucket sort with each bucket of size one hence distribution counting sort
for m
n = 1. Associative sort is still slightly better but the performances get closer when
m
n decreases. On the other hand, value-sorting version of distribution counting sort is 2
times faster than associative sort for m
n = 1. Similarly, performances get closer when m
n
decreases.
Even omitting its space efficiency for a moment, associative sort asymptotically out-
performs all content based sorting algorithms when n is large relative to m.
21
3.4
Improved In-place Associative Integer Sorting
With a very simple revision, the associative sorting technique can be improved theoreti-
cally and practically. The only cost of the improved version is that a recursive implemen-
tation is not possible.
The tag bit discriminates the word as a node in the array space after practicing. During
storage where the nodes are stored at the beginning of the array in order of precedence,
the positional information (log n bits) of a node is stored either in its record or in an
idle integer accompanying the node. However, ignoring all the MSBs of the array, if only
the records (w − 1 bits) of the nodes are stored at the beginning of the array (short-
term memory), there will be unmoved nd nodes (tagged words) spread over the array
space with relative order and nd records in the short-term memory (S[0, . . . nd − 1]) with
the same order after storage. Hence, a one-to-one correspondence is obtained with the
stored records and the nodes of the array. This means that, selecting the pivot equal to
δ + n − 1, if nc idle integers and n′
d unpracticed integers that are distributed disorderly
in S[nd . . . n − 1] are partitioned, then nc idle integers come immediately after the short-
term memory resulting in S[0 . . . nd + nc − 1]. Hence, retrieval phase can search the
overall array from right to left backwards for the first node, retrieve the integer from the
imaginary subspace through that node using its position and inverse of hash function,
read its number of occurrences from its record S[nd − 1] and copy over S[0 . . . nd + nc − 1]
right to left backwards without collision with the short-term memory. Afterwards, the tag
bit of the processed node can be cleared and a new search to the left can be carried for the
next node which will correspond to the next record S[nd − 2] of the short-term memory.
This can continue until all the integers are retrieved resulting in the sorted permutation
of the practiced integers.
22
As the position of the node is not required during storage, there is no need to get an
idle integer immediately after a node that has practiced at least 2w−1−log n idle integers.
This means that ǫ is always zero. Hence, the improved version is capable of practicing
the integers in the interval [δ, δ + n − 1] in each iteration. Therefore, while the former
technique was capable of sorting integers that satisfy S[i] − δ + ǫ < n, the improved
version sorts the integers that satisfy S[i] − δ < n in each iteration. Hence, n integers
S[0, . . . n − 1] each in the range [δ, δ + n − 1] can be sorted exactly in O(n) time regardless
of the number of the integers.
3.5
Improved In-place Associative Distinct Integer Sorting
The improved associative integer sorting technique can be easily specialized for distinct
integers. Given n distinct integers S[0...n − 1] each in the range [δ, δ + m − 1], a monotone
bijective super hash function can be constructed as a partial function that assigns each
integer in the range [δ, δ + (w − 1)n − 1] to exactly one element in j ∈ [0, n − 1] and one
element in k ∈ [0, (w − 1) − 1],
j = (S[i] − δ) div (w − 1)
if S[i] − δ < (w − 1)n
k = (S[i] − δ) mod (w − 1)
if S[i] − δ < (w − 1)n
(3.4)
(3.5)
where mod is the remainder modulo. In this case, w −1 integers may collide and mapped
to the same node created at j ∈ [0, n − 1] (Eqn. 3.4) in the ILS. But we can use w − 1
free bits of a record to encode which of w − 1 distinct integers are mapped to the same
node by setting their corresponding bit determined by k (Eqn. 3.5). In other words, now
the ILS is two dimensional over the array space where the first dimension along the array
designates the node position and the second dimension along the bits of the node uniquely
determines the integers which are mapped to the ILS through that node.
23
Once all the distinct integers in the predefined interval are practiced, ignoring all the
MSBs of the array, if only the records (w − 1 bits) of the nodes are stored at the beginning
of the array (short-term memory) in order of precedence, there will be unmoved nd nodes
(tagged words) spread over the array space, and nd records in the short-term memory
(S[0 . . . nd − 1]) with the same order after storage. Hence, selecting the pivot equal to
δ + (w − 1)n − 1, if nc idle integers and n′
d unpracticed integers that are distributed
disorderly in S[nd . . . n − 1] are partitioned, then nc idle integers come immediately after
the short-term memory resulting in S[0 . . . nd + nc − 1]. Hence, the sorted permutation
of the practiced interval can be retrieved by searching the nodes (tagged words) of the
array backwards. When a tagged word (node) is found, the base of the integers mapped
to that node can be calculated using the position of the node and the inverse of Eqn.3.4.
Then, the position of the bits that are equal to 1 in the corresponding record S[nd − 1]
uniquely determine (with the inverse of Eqn.3.5) the integers mapped to the same node
from most significant to least and each can be copied over S[0 . . . nd + nc − 1] right to left
backwards without collision with the short-term memory.
As a result, an algorithm is obtained that sorts n distinct integers S[0 . . . n − 1] each in
the range [0, m−1] in exactly O(n) time if m < (w −1)n. Otherwise, it sort in O(n+ m
w−1)
time for the worst, O( m
w−1) time for the average (uniformly distributed keys) and O(n)
time for the best case using only O(1) extra space.
4
In-place Associative Permutation Sort
In-place associative integer sorting introduced in the previous section is a value-sorting
algorithm except the one for distinct integer keys (Section 3.2). All others are not suitable
for rank-sorting which is the main objective of all sorting algorithms. In this section, in-
24
place associative permutation technique will be introduced making the former technique
suitable for rank-sorting without degrading its performance.
Although proposed to provide a reasonable qualitative fit of the forgetting dynamics
of the short term memory [8] in cognitive neuroscience, the idea behind the original
perturbation model of Estes [28] is that the order of the elements is inherent in the cyclic
reactivation of the elements, i.e., reactivations lead to reordering of the elements. Hence,
when the idea behind the perturbation model is combined with the original technique
of associative integer sorting, in-place associative permutation sort is obtained where the
order of the practiced interval is inherent in the cyclic reactivation of all the keys of
the array. The technique consists of three phases namely (i) practicing, (ii) permutation
(reactivation) and (iii) restoring.
The term "reactivation" immediately reminds the cycle leader permutation because
when each key is sent to its final position it reactivates the key that was there. Associative
permutation sort is based on a very simple idea: if nd + nc keys of a predefined interval
become consecutive, distinct and in the range [0, nd + nc − 1], and if each new distinct
value of a modified key corresponds to its rank with respect to others pointing its exact
position in the sorted permutation, then the overall array can be reactivated with a special
form of cycle leader permutation that rearranges the practiced (and modified) keys at the
beginning of array in order.
4.1 Practicing
Only two new steps are added to practicing phase of associative integer sorting (Sec-
tion 3.3.1) before permutation (reactivation) phase. These are (i) accumulation and (ii)
re-practicing steps.
25
In accumulation step, all the records of the nodes are accumulated from left to right.
Hence, at the end, each record of a particular node keeps the exact position of the last
idle key practiced by that node.
As mentioned previously, an ILS can create other subspaces and associations using
the idle keys that were already practiced by manipulating either their position or value
or both. Hence, it is logical to use the nodes of ILS as discrete hash functions that define
the values of idle keys when they are re-practiced using the same monotone bijective hash
function. This is the re-practicing step. When an idle key is remapped to its node, it can
obtain its exact position (ticket) from the record of its node. The record of the node will
be decreased by one for each re-practiced idle key. This means that, when all the idle keys
of a particular node are re-practiced, the value of the node will point its exact position
in the sorted permutation as well, of course when its MSB is ignored. Hence, when all
the idle keys of all the nodes are re-practiced, all the idle keys and all the nodes (ignoring
their MSB) will become consecutive, distinct and in the range [0, nd + nc − 1] pointing to
their exact position in the sorted permutation.
4.2 Permutation (Reactivation)
After practicing, if the tag bits of the nodes are ignored, a simple cycle leader permutation
can trivially rearrange the practiced keys at the beginning of array in order. Unsurpris-
ingly each node precedes its own idle keys after this rearrangement. This simply puts the
elements (satellite information) of the keys in order. However, the modified keys should
be restored to their original values unless one intentionally wishes to normalize the keys
and put all of them in the range [0, n − 1] at the end. Hence, the cycle leader permu-
tation (reactivation) should take care of the nodes since the position of the nodes before
26
reactivation is used to recall the key and retrieve it from the ILS using the inverse hash
function. This immediately lets us to assert that if it would be possible to rearrange the
practiced interval at the beginning of the array with each node storing its former position
in its record, then it would be possible to restore all the modified keys to their original
values. A node has a record of w − 1 bits which stores the node's exact position before the
reactivation. Hence, while a node is being moved to its exact position, its former position
can be overwritten into its record as the cue which can be used to recall the key using
the inverse hash function. But, from information processing perspective, how one can
distinguish the nodes that are already moved, from the nodes that are not moved yet in
such a case? The idle and unpracticed keys (that are out of the practiced interval) are the
solution to this problem. If a node is not at its exact position, then it is evident that either
an idle or a key that is out of the practiced interval will address the position of that node.
Hence, a special outer cycle leader permutation that only reactivates the idle keys and
the keys that are out of the practiced interval will ensure that the corresponding one will
be moved to the actual position of the node giving a chance to start an inner cycle leader
permutation that reactivates only the nodes and ensures that the nodes will be moved to
their exact position storing their former position in their record as the cue. Once a node
is moved to its exact position, there can not be any other outer cycle leader which will
address that particular position. If another node is available where that particular node
is moved, then the inner cycle leader permutation can continue with that node. However,
if an idle or an unpracticed key is encountered, then the inner cycle leader permutation
terminates and the outer cycle leader permutation continues.
From information processing perspective, associative permutation (reactivation) phase
is a derivative of cycle leader permutation (Section 3.1) where the nodes, practiced idle
27
keys and unpracticed keys (that are out of the practiced interval) are treated specially.
Processing the array from left to right for i = 0, 1, . . . n−1, if S[i] is a node (tagged word),
do nothing, increase the iterator i and start over with S[i]. If S[i] is a key that is out of
the practiced interval (S[i] − δ ≥ n), exchange S[i] with S[k] starting with k = nd + nc,
and increase k every time a key out of the practiced interval is moved to its new position.
Besides, set j = k − 1 and start over with S[i] if it is either an idle or a key that is out
of practiced interval. On the other hand, if S[i] is an idle key (S[i] − δ < n), implicitly
practice it by exchanging S[i] with S[j] where j = S[i] and start over with S[i] if it is
either an idle or a key that is out of practiced interval. If the key that came to S[i] from
j is a node, in other words, if the idle key or the key that is out of the practiced interval
is moved to a position where a node is there, then start an inner cycle permutation that
reactivates only the nodes until a new idle key or a key that is out of the practiced interval
is encountered. To do this, clear MSB of the node S[i] and read where it point by p = S[i].
Copy S[p] over S[i] and write the former position j of the node by S[p] = j and set MSB
of S[p] making it a node. Now S[i] is the new key to be processed and j = p is where
it came from. If S[i] is a node again, continue the cycle leader permutation. However, if
S[i] is an idle key or a key that is out of the practiced interval, start over with S[i].
4.3 Restoring
With a final scan of the short-term memory (S[0 . . . nd + nc − 1]), one can obtain the exact
values of practiced keys from their preceding nodes (tagged words). Each node stores its
absolute position in its record (w − 1 bits) and cues the recall of the key using the inverse
hash function. The value of the key can be copied over all the succeeding keys until a
new node is found or the short-term memory ends.
28
4.4 Empirical Tests
Practical comparisons for arrays up to one million integer keys all in the range [0, n−1] on
a Pentium machine with radix sort and bucket sort indicate that associative permutation
sort is slower roughly 2 times than radix sort and slower roughly 3 times than bucket sort.
On the other hand, it is faster than quick sort for the same arrays roughly 1.5 times.
Although its time complexity is similar to that of in-place associative sort [1] and
practically slower, in-place associative permutation sort is proposed for integer key sorting
problem. Hence, it sorts n records S[0 . . . n − 1] each have an integer key in the range
[0, m − 1] with m > n using O(1) extra space in O(n + m) time for the worst, O(m) time
for the average (uniformly distributed keys) and O(n) time for the best case.
Remark 4.1. Associative permutation sort technique is on-line in a manner that after each
iteration, the keys in the range [δ, δ + n − 1] are sorted at the beginning of the array and
ready to be used.
5 Conclusions
The technique of associative sorting is presented. In-place associative permutation sort
technique is proposed which solves the main difficulties of distributive sorting algorithms
by its inherent three basic steps namely (i) practicing, (ii) permutation and (iii) restoring.
It is very simple and straightforward and around 30 lines of C code is enough.
From time complexity point of view, both techniques show similar characteristics with
bucket sort and distribution counting sort. They sorts the integers/keys associatively
in O(m) time for the average (uniformly distributed keys) and O(n) time for the best
case. Although its worst case time complexity is O(n + m), it is upper bounded by O(n2)
29
for arrays where m > n2. On the other hand, it requires only O(1) additional space,
making it time-space efficient compared to bucket sort and distribution counting sort.
The ratio m
n defines its efficiency (time-space trade-offs) letting very large arrays to be
sorted in-place. Furthermore, the dependency of the efficiency on the distribution of the
keys is O(n) which means it replaces all the methods based on address calculation, that
are known to be very efficient when the keys have known (usually uniform) distribution
and require additional space more or less proportional to n. Hence, in-place associative
permutation sort asymptotically outperforms all content based sorting algorithms making
them attractive in almost every case even when space is not a critical resource.
The technique seems to be very flexible, efficient and applicable for other problems as
well, such as hashing, searching, succinct data structures, gaining space, etc.
The only drawback of associative permutation sort is that it is unstable. However,
as mentioned before, an imaginary subspace can create other subspaces and associations
using the idle integers that were already practiced by manipulating either their position
or value or both. Hence, different approaches can be developed to solve problems such as
stability.
6 Discussion
Associative permutation sort first finds the minimum of the array and starts with the
keys in [min(S), min(S) + n − 1]. However, instead of starting with this interval, omitting
the MSBs, if we consider a word as w − 1 bits and the most significant ⌈log n⌉ bits of a
word as the key and the remaining bits as the satellite information, the problem reduces
to sorting n integer keys S[0 . . . n − 1] each in the range [0, 2⌈log n⌉ − 1]. Since it is possible
that 2⌈log n⌉ − 1 > n − 1, the keys in [n, 2⌈log n⌉ − 1] become the keys that are out of the
30
practiced interval.
As a result, when the keys are sorted according to their most significant ⌈log n⌉ bits,
in-place associative most significant ⌈log n⌉ radix sort is obtained. After the array is
sorted according to their most significant ⌈log n⌉ bits, the idle keys are grouped and each
group is preceded by the corresponding node that has practiced them. Hence, each group
can be sorted sequentially or recursively assuming the satellite information as the key. If
itself is used, it becomes an algorithm based on hash-and-conquer paradigm in contrast
to divide-and-conquer. However, size of subgroups decreases and it may not be efficient
when the ratio of range of keys in each subgroup to size of that subgroup, i.e., m
n increases.
Hence, other strategies may need to be developed after the first pass.
References
[1] A.E. Cetin, "In-place associative integer sorting", arXiv:1209.0572v2 [cs.DS]
[2] A.E. Cetin, "Improved in-place associative integer sorting", arXiv:1209.3668v1 [cs.DS]
[3] A.E. Cetin,
"Sorting distinct
integer keys using in-place associative sort",
arXiv:1209.1942v2 [cs.DS]
[4] A.E. Cetin, "Sorting distinct integers using improved in-place associative sort",
arXiv:1209.4714v1 [cs.DS]
[5] A.E. Cetin, "In-place associative permutation sort", arXiv:arXiv:1210.1771v1 [cs.DS]
[6] K.S. Lashley, "The problem of serial order in behavior", in Cerebral Mechanisms in
Behavior, ed. LA Jeffress, John Wiley & Sons, 1966.
31
[7] K.S. Lashley, "In search of the engram", IEEE Transactions on Electronic Computer,
Vol. EC-15, no. 4, 1966.
[8] R.N.A. Henson, "Short-term memory for serial order: The start-end model", Cognitive
Psychology, Vol. 36, pp. 73 - 137, 1998.
[9] C.A.R Hoare, "Quicksort", Comput. J., Vol. 5, pp. 10 - 16, 1962.
[10] D.L. Shell, "A High Speed Sorting Procedure", Communications of ACM, Vol. 2, pp.
30 - 32, 1959.
[11] A. Burnetas, D. Solow, R. Agrawal, "An analysis and implementation of an efficient
in-place bucket sort", Acta Informatica, Vol. 34, pp. 687 - 700, 1997.
[12] J. Williams, "Heapsort", Communications of the ACM, Vol. 7, pp. 347 - 348.
[13] Anonymous, "Merge Sort", Wikipedia, 2012.
[14] A. Levitin, The Design and Analysis of Algorithms. Addison-Wesley, 2007.
[15] H.H. Seward, Information Sorting in the Application of Electronic Digital Computers
to Business Operations, Master's thesis, MIT Digital Computer Laboratory, Report R-
232, Cambridge, 1954.
[16] W. Feurzig, "Algorithm 23, mathsort", Commun. ACM, Vol. 3, pp. 601 - 602, 1960.
[17] E.J. Isaac, R.C. Singleton, "Sorting by address calculation", Journal of the ACM,
Vol. 3, pp. 169 - 174, 1956.
[18] M.E. Tarter, R.A. Kronmal, "Non-uniform key distribution and address calculation
sorting", Proc. ACM Nat'l Conf. 21, 1966.
32
[19] I. Flores, "Computer time for address calculation sorting", Journal of the ACM, Vol.
7, pp. 389 - 409, 1960.
[20] B. Jones, "A variation on sorting by address calculation", Communications of the
ACM , Vol. 13, pp. 105 - 107, 1970.
[21] G. Gupta, "Sorting by hashing and inserting", Proc. ACM Annual Computer Science
Conf. 17, pp. 409 - 409, 1989.
[22] F. Suraweera, J.M. Al-Anzy, "Analysis of a modified address calculation sorting
algorithm", Comput. J. Vol. 31, pp. 561 - 563, 1988.
[23] H.M. Mahmoud, Sorting, A Distribution Theory, John Wiley and Sons, 2000.
[24] T.H. Cormen, C.E. Leiserson, R.L. Rivest, C. Stein, Introduction to Algorithms, The
MIT Press, 2001.
[25] R. Sedgewick, Algorithms in C, Parts 1-4: Fundamentals, Data Structures, Sorting,
Searching, Addison-Wesley, 1997.
[26] D.E. Knuth, The Art of Computer Programming, Volume 3: Sorting and Searching,
Addison-Wesley, 1973.
[27] R. Shiffrin, J. Cook, "Short-term forgetting of item and order information", Journal
of Verbal Learning and Verbal Behavior, Vol. 17, pp. 189 - 218, 1978.
[28] W.K. Estes, A.W. Melton, E. Martin, "An associative basis for coding and organi-
zation in memory", Coding processes in human memory (pp. 161-190). Washington,
1972.
[29] A.G. Hanlon, "Content-addressable and associative memory systems", IEEE Trans-
actions on Electronic Computer, Vol. EC-15, pp. 509 - 521, 1966.
33
[30] J. Katajainen, T. Pasanen, "Stable minimum space partitioning in linear time", BIT
Numerical Mathematics, Vol. 32, pp. 580 - 585, 1992.
[31] J. Salowe, W. Steiger, "Simplifed stable merging tasks", Journal of Algorithms, Vol.
8, pp. 557 - 571, 1987.
[32] G. Franceschini, S. Muthukrishnan, M. Patrascu, "Radix sorting with no extra
space", ESA'07 Proc. 15th annual European conference on Algorithms, pp. 194 - 205,
2007.
[33] D. Belazzougui, P. Boldi, R. Pagh, S. Vigna, "Monotone Minimal Perfect Hashing:
Searching a Sorted Table with O(1) Accesses", SODA '09 Proc. of the twentieth Annual
ACM-SIAM Symposium on Discrete Algorithms , pp. 785-794.
[34] HK Rosen, Handbook of Discrete and Combinatorial Mathematics, CRC Press, 2000.
34
|
1802.06748 | 3 | 1802 | 2019-05-25T19:33:09 | Breaking the Linear-Memory Barrier in MPC: Fast MIS on Trees with Strongly Sublinear Memory | [
"cs.DS"
] | Recently, studying fundamental graph problems in the \emph{Massively Parallel Computation (MPC) framework, inspired by the MapReduce paradigm, has gained a lot of attention. An assumption common to a vast majority of approaches is to allow $\widetilde{\Omega}(n)$ memory per machine, where $n$ is the number of nodes in the graph and $\widetilde{\Omega}$ hides polylogarithmic factors. However, as pointed out by Karloff et al. [SODA'10] and Czumaj et al. [STOC'18], it might be unrealistic for a single machine to have linear or only slightly sublinear memory.
In this paper, we thus study a more practical variant of the MPC model which only requires substantially sublinear or even subpolynomial memory per machine. In contrast to the linear-memory MPC model and also to streaming algorithms, in this low-memory MPC setting, a single machine will only see a small number of nodes in the graph. We introduce a new and strikingly simple technique to cope with this imposed locality.
In particular, we show that the Maximal Independent Set (MIS) problem can be solved efficiently, that is, in $O(\log^3 \log n)$ rounds, when the input graph is a tree. This constitutes an almost exponential speed-up over the low-memory MPC algorithm in $O(\sqrt{\log n})$-algorithm in a concurrent work by Ghaffari and Uitto [SODA'19] and substantially reduces the local memory from $\widetilde{\Omega}(n)$ required by the recent $O(\log \log n)$-round MIS algorithm of Ghaffari et al. [PODC'18] to $n^{\alpha}$ for any $\alpha>0$, without incurring a significant loss in the round complexity. Moreover, it demonstrates how to make use of the all-to-all communication in the MPC model to almost exponentially improve on the corresponding bound in the $\mathsf{LOCAL}$ and $\mathsf{PRAM}$ models by Lenzen and Wattenhofer [PODC'11]. | cs.DS | cs | Breaking the Linear-Memory Barrier in MPC:
Fast MIS on Trees with Strongly Sublinear Memory
Sebastian Brandt
Manuela Fischer
Jara Uitto
ETH Zurich
ETH Zurich
ETH Zurich & Uni. Freiburg
[email protected]
[email protected]
[email protected]
Abstract
Recently, studying fundamental graph problems in the Massively Parallel Computation (MPC)
framework, inspired by the MapReduce paradigm, has gained a lot of attention. An assumption
common to a vast majority of approaches is to allow (cid:101)Ω(n) memory per machine, where n is the
number of nodes in the graph and (cid:101)Ω hides polylogarithmic factors. However, as pointed out by Karloff
et al. [SODA'10] and Czumaj et al. [STOC'18], it might be unrealistic for a single machine to have
linear or only slightly sublinear memory.
In this paper, we thus study a more practical variant of the MPC model which only requires
substantially sublinear or even subpolynomial memory per machine. In contrast to the linear-memory
MPC model and also to streaming algorithms, in this low-memory MPC setting, a single machine will
only see a small number of nodes in the graph. We introduce a new and strikingly simple technique
to cope with this imposed locality.
√
In particular, we show that the Maximal Independent Set (MIS) problem can be solved efficiently,
that is, in O(log3 log n) rounds, when the input graph is a tree. This constitutes an almost exponential
log n)-algorithm in a concurrent work by Ghaffari
speed-up over the low-memory MPC algorithm in (cid:101)O(
and Uitto [SODA'19] and substantially reduces the local memory from (cid:101)Ω(n) required by the recent
O(log log n)-round MIS algorithm of Ghaffari et al. [PODC'18] to nε for any ε > 0, without incurring
a significant loss in the round complexity. Moreover, it demonstrates how to make use of the all-to-all
communication in the MPC model to almost exponentially improve on the corresponding bound in the
LOCAL and PRAM models by Lenzen and Wattenhofer [PODC'11].
9
1
0
2
y
a
M
5
2
]
S
D
.
s
c
[
3
v
8
4
7
6
0
.
2
0
8
1
:
v
i
X
r
a
1
Introduction
Parallel Computation Paradigms for Massive Data: When confronted with huge data sets, purely
sequential approaches become untenably inefficient. To address this issue, several parallel computation
frameworks specially tailored for processing large scale data have been introduced.
Inspired by the
MapReduce paradigm [DG08], Karloff, Suri, and Vassilvitskii [KSV10] proposed the Massively Parallel
Computation (MPC) model, which was later refined in many works [GSZ11, BKS14, ANOY14, BKS17,
C(cid:32)LM+17].
Massively Parallel Computation Model: In the MPC model, an input instance of size N is dis-
tributed across M machines with local memory of size S each. The computation proceeds in rounds,
each round consisting of local computation at the machines interleaved with global communication (also
called shuffling, adopting the MapReduce terminology) between the machines.
In the shuffling step, every machine is allowed to send as many messages to as many machines as
it wants, as long as for every machine the total size of sent and received messages does not exceed its
local memory capacity. The quantity of main interest is the round complexity: the number of rounds
needed until the problem is solved, that is, until every machine outputs its part of the solution. This
measure constitutes a good estimate for the actual running time, as local computation is presumed to
be negligible compared to the cost-intensive shuffling, which requires a massive amount of data to be
transferred between machines.
Sublinear Memory Constraint: Note that S ≥ N leads to a degenerate case that allows for a trivial
solution. Indeed, as the data fits into the local memory of a single machine, the input can be loaded
there, and a solution can be computed locally. Due to the targeted application of MPC in the presence
of massive data sets, thus large N , it is often crucial that S is not only smaller than N but actually
substantially sublinear in N . The total memory M · S in the system has to be at least N , so that the
input actually fits onto the machines, but ideally not much larger. Summarized, one requires S = (cid:101)O (N ε)
memory on each of the M = (cid:101)O
need essentially linear in n -- for instance, (cid:101)Ω(n) or mildly sublinear like n1−o(1) -- memory per machine,
this assumption. Note that for sparse graphs with N = (cid:101)O(n) edges, this violates the sublinear memory
explicitly restricting the attention to dense graphs with N =(cid:101)Ω(n1+ε) edges, as to ensure sublinearity in
where n is the number of nodes in the input graph1. We refer to [BDH18] for a brief discussion of
Sublinear Memory for Graph Problems: Basically all known MPC techniques for graph problems
(cid:16)
N 1−ε(cid:48)(cid:17)
machines, for 0 < ε(cid:48) ≤ ε < 1.
constraint, getting close to the degenerate regime. This issue has been artificially circumvented by
N while still not having to relinquish the nice property that (essentially) all nodes fit into the memory
of a single machine [KSV10].
Besides being a stretch of the definition, this additionally imposed condition of denseness of the input
graph does not seem to be realistic. In fact, as recently also pointed out by [C(cid:32)LM+17], most practical
large graphs are sparse. For instance in the Internet, most of the nodes have a small degree. Even for
dense graphs, where in theory the sublinear memory constraint is met, practicability of the parameter
range does not need to be ensured; linear or slightly sublinear in n might be prohibitively large.
Furthermore, it is a very natural question to ask whether there is a fundamental reason why the
known techniques get stuck at the near-linear barrier. One important aspect of our work is, from the
theory perspective, that it breaks this threshold and thereby opens up a whole new unexplored domain
of research questions.
Low-Memory MPC Model: We study a more realistic regime of the parameters for problems on large
graphs, captured by the following low-memory MPC model.
1In the context of graph problems, it is typical to assume that all incident edges of a node are stored on the same machine,
resulting in two copies of an edge, one for each endpoint. We refer to [PRS16, Section 1.1] for a thorough discussion. Also
see the remark at the end of this section.
1
Low-Memory MPC Model for Graph Problems:
The input is a graph G = (V, E) with n nodes and m edges of size N = (cid:101)O(n + m). Given M =
(cid:16) N 1+α(cid:48)
machines with local memory S = (cid:101)O (nε) each, for arbitrary constants ε > 0 and α(cid:48) ≥ 0, we
(cid:101)O
(cid:17)
S
raise the question of what problems on G can be solved efficiently -- that is, in poly log log n rounds.
Note that for sparse graphs, this condition exactly matches the sublinear memory constraint, and
hence does not allow a trivial solution, as opposed to the setting with linear memory. We point out
that low memory variants of the MPC model have been studied before [PPR+12, CPPU15], resulting
in O(log n)-round algorithms for a variety of problems. For many of the fundamental graph problems,
however, O(log n) is often particularly easy to achieve, for instance by directly adopting LOCAL algo-
rithms. We thus restrict our attention to "efficient" algorithms, which we define to be a poly log log n
function, given that the state-of-the-art algorithms in the MPC model tend to end in this regime of round
complexities. Note that no general super-constant lower bounds are known [RVW16].
Concurrent Related Work: Until very recently, MPC research had focused on linear-memory algo-
rithms. After (a preliminary version of) this work, the low-memory setting gained a lot of attention. This
led to a variety of new results for graph problems in this model. We briefly outline recent developments
that have taken place after this work. In a follow-up work, [BBD+19] devise MIS and matching algo-
rithms in uniformly sparse graphs in O(log2 log n) rounds. In independent concurrent works, Ghaffari
and Uitto [GU19] and Onak [Ona18] provide algorithms for the problems of maximal independent set and
log log n)-
√
log n) rounds. In [CFG+18], Chang et al. develop an O(
matching in general graphs in (cid:101)O(
√
round low-memory MPC algorithm for (∆ + 1)-list coloring.
Remark 1.1. If a node cannot be stored on a single machine, as its degree is larger than S, one has to
introduce some sort of a workaround, e.g., have several smaller-degree copies of the same node on several
separate machines. In the end of Section 2, we argue how to get rid of this issue, in our problem setting,
by a clean-up phase in the very beginning. To make the statements and arguments more readable, we
throughout think of this clean-up as having taken place already. Instead, one could also work with the
simplifying assumption that every machine has S = (cid:101)O (nε + ∆) memory, so that this issue does not arise
in the first place.
1.1 Limitations of Linear-Memory MPC Techniques
In the following, we briefly overview recent techniques from the world of Massive Parallel Computation
algorithms, and give some indications as to why they are likely to fail in the low-memory setting. The
restriction to substantially sublinear memory, to the best of our knowledge, indeed rules out all the known
MPC techniques, which seem to hit a boundary at roughly S = (cid:101)Ω(n): moving from essentially linear to
significantly sublinear memory incurs an exponential overhead in their round complexity, regardless of
the density of the graph. This blow-up in the running time gives rise to the question of to what extent
this near-linear memory is necessary for efficient algorithms.
(Direct) PRAM/LOCAL Simulation: One easy way of devising MPC algorithms is by shoehorning
parallel or distributed algorithms into the MPC setting. For not too resource-heavy PRAM algorithms,
there is also a standard simulation technique [KSV10, GSZ11] that automatically transforms them into
MPC algorithms. This approach, however, suffers from several shortcomings. First and foremost, the
reduction leads to an Ω(log n) round complexity, which is exponentially above our efficiency threshold.
Round Compression: Another similar technique, called round compression, introduced by Assadi and
Khanna [AK17, Ass17], provides a generic way of compressing several rounds of a distributed algorithm
into fewer MPC rounds, resulting in an (almost) exponential speed-up. However, this method heavily
relies on storing intermediate values, leading to a blow-up of the memory. In particular, when requiring
the algorithm to run in poly log log n rounds, superlinear memory per machine seems inevitable.
2
Filtering: The idea of the filtering technique [LMSV11, KMVV15] is to reduce the size of the input
by carefully removing a large amount of edges from the input graph that do not contribute to the
(optimal) solution of the problem. This reduction is done by either randomly sampling the edges, or
by deterministically choosing sets of relevant edges, so that the resulting (partial) problem instance fits
on a single machine, and hence can be solved there. This requires significantly superlinear memory, or
logarithmically many rounds if memory is getting close to the linear regime. Moreover, the approach
seems to get stuck fundamentally at S =(cid:101)Ω(n), since it relies on one machine eventually seeing the whole
(filtered) graph.
Coresets: One very recent and promising direction for MPC graph algorithms is the one of (randomized
composable) coresets [AK17,ABB+19], in some sense building on the filtering approach. The idea is that
not all the information of the graph is needed to (approximately) solve the problem. One thus can get rid
of unimportant parts of the information. Solving the problem on this core, one then can derive a perfect
solution or a good approximation to it, at much lower cost. This solution, however, is found by loading
the coreset (or parts of it) on one machine, and then locally computing a solution, which again seems to
be stuck at S =(cid:101)Ω(n), for similar reasons as the filtering approach.
1.2 Local Techniques for Low-Memory MPC
In this section, we propose a direction that seems to be promising to pursue in order to devise efficient
MPC algorithms in the substantially sublinear memory regime.
Inherent Locality and Local Algorithms: The low-memory MPC model, as compared to the tradi-
tional MPC graph model and the streaming setting, suffers from inherent locality: Since the memory of a
single machine is too small to fit all the nodes simultaneously, it will never be able to have a global view
of all the nodes in the graph. When devising techniques, we thus need to deal with this intrinsic local
view of the machines. It seems natural to borrow ideas from local distributed graph algorithms, which
are designed exactly to cope with this locality restriction. A direct simulation, however, in most cases
only results in Ω(poly log n)-round algorithms. The problem is that these algorithms do not make use of
the additional power of the MPC model, the global all-to-all communication, as the communication in
those message-passing-based models is restricted to neighboring nodes.
Local Meets Global: We propose a strikingly simple technique to enhance local-inspired approaches
with global communication, in order to arrive at efficient algorithms in the world of low-memory MPC
which are exponentially faster than their local counterparts and whose memory requirements are polyno-
mially smaller per machine than their traditional MPC counterparts. We describe this technique in the
context of the MIS problem on trees, even though it is more general.
1.3 Our Results
In this paper, we focus on the Maximal Independent Set (MIS) problem, one of the most fundamental local
graph problems. We propose efficient and surprisingly simple algorithms for the case of trees, which is
particularly interesting for the following reason. While trees admit a trivial solution in the linear-memory
MPC model, this cheat will not work in our low-memory setting. In some sense, it thus is the easiest
non-trivial case, which makes it the most natural starting point for further studies. In fact, we strongly
believe that our techniques can be extended to more general graph families and problems2.
We provide two different efficient algorithms for MIS on trees. Our first algorithm in Theorem 1.2
is strikingly simple and intuitive, but comes with a small overhead in the total memory of the system,
meaning that M · S is superlinear in the input size n.
2Indeed, there is a follow-up work generalizing our approach from trees to uniformly sparse graphs and from MIS only
to MIS and maximal matching [BFU18b, BBD+19].
3
trees in the low-memory setting, that is, with S = (cid:101)O (nε) local memory on each of M = (cid:101)O(cid:0)n1−ε/3(cid:1)
Theorem 1.2. There is an O(log2 log n)-round MPC algorithm that w.h.p.3 computes an MIS on n-node
machines, for any 0 < ε < 1.
Our second algorithm in Theorem 1.3 gets rid of this overhead at the cost of a factor of log log n in
the running time.
trees in the low-memory setting, that is, with S = (cid:101)O (nε) local memory on each of M = (cid:101)O(cid:0)n1−ε(cid:1)
Theorem 1.3. There is an O(log3 log n)-round MPC algorithm that w.h.p. computes an MIS on n-node
machines, for any 0 < ε < 1.
The algorithms in Theorems 1.2 and 1.3 almost match the conditional lower bound of Ω(log log n)
for MIS (on general graphs) due to Ghaffari, Kuhn, and Uitto [GKU19], which holds unless there is an
o(log n)-round low-memory MPC algorithm for connected components. This, in turn, is believed to be
impossible under a popular conjecture [YV18].
Our algorithms improve almost exponentially on the (cid:101)O(
√
√
log n)-round low-memory MPC algorithms
in concurrent works -- for bounded-arboricity by Onak [Ona18] and for general graphs by Ghaffari and
Uitto [GU19] -- as well as on the algorithms directly adopted from the PRAM/LOCAL model: an O(log n)-
round algorithm for general graphs due to Luby [Lub86] and independently Alon, Babai, Itai [ABI86],
log n · log log n)-round algorithm for trees by Lenzen and Wattenhofer [LW11]. Note that
and the O(
for rooted trees, the PRAM/LOCAL algorithm by Cole and Vishkin [CV86] directly gives rise to an
O(log∗ n)-round low-memory MPC algorithm.
or even n1/poly log log n (see Corollary 1.4) while not incurring a significant loss in the round complexity,
compared to the recent O(log log n)-round MIS algorithm by Ghaffari et al. [GGK+18].
Moreover, our result shows that the local memory can be reduced substantially from (cid:101)Ω(n) to nε
range for S as given in Theorems 1.2 and 1.3, that is, S = (cid:101)O(nα), where α > 0 is an arbitrary constant.
(cid:101)O(cid:0)n1−ε/3(cid:1) machines with S = (cid:101)O (nε) local memory each in O(cid:0) 1
However, ε does not need to be a constant. Indeed, we can even go to subpolynomial memory S = no(1).
ε · log2 log n(cid:1) MPC rounds.
Corollary 1.4. For any ε = Ω (1/poly log log n), an MIS on an n-node tree can be computed on M =
Throughout the paper, when we mention the low-memory MPC setting, we refer to the parameter
1.4 Our Approach in a Nutshell
In the following, we give a short (and slightly imprecise) sketch of the steps of our algorithm. Our approach
is based on the shattering technique which recently has gained a lot of attention in the LOCAL model of
distributed computing [BEPS16] and goes back to the early nineties [Bec91]. The idea of shattering is to
randomly break the graph into several significantly smaller components by computing a partial solution.
The problem on the remaining components then is solved by a post-shattering algorithm.
Shattering
The goal of our shattering technique is to compute an independent set such that after the removal of these
independent set nodes and all their neighbors, the remaining graph, w.h.p., consists of components of size
at most poly log n. This is done in two steps: first, the maximum degree, w.h.p., is reduced to poly log n
using the iterated subsample-and-conquer method, and then a local shattering algorithm is applied to
this low-degree graph.
I) Degree Reduction via Iterated Subsample-and-Conquer
Our subsample-and-conquer method will w.h.p. reduce the maximum degree of a graph polynomially,
from ∆ to roughly ∆1/(1+ε), as long as ∆ = Ω(poly log n). After O(log1+ε log ∆) iterations, the degree of
our graph drops to poly log n.
3As usual, w.h.p. stands for with high probability, and means with probability at least 1 − n−c, for any c ≥ 1.
4
Subsample: We sample the nodes independently with probability roughly ∆
1+ε , where ∆ is an up-
per bound on the current maximum degree4 of the graph. This subsampling step guarantees, roughly
speaking, the following three very desirable properties of the graph G(cid:48) induced by the sampled nodes.
− 1
i) The diameter of each connected component of G(cid:48) is bounded by O(log∆ n).
ii) The number of nodes in each connected component of G(cid:48) is at most nε/3.
iii) Every node with degree ∆1/(1+ε) or higher in G has many neighbors in G(cid:48).
Conquer: We find a random MIS in all the connected components of G(cid:48) in parallel. This can be done
by gathering the connected components5, locally picking one of the two 2-colorings of this tree uniformly
at random, and adding the black, say, nodes to the MIS. We will see that properties i) and ii) are
crucial to ensure that the gathering can be done efficiently. In particular, storing the components on a
single machine is possible due to the small size of the components, and the gathering is fast due to the
small diameter. Because of property iii), the randomness in the choice of the MIS in every connected
component, as well as the tree structure, all high-degree nodes in the original graph (sampled or not),
w.h.p., will have an adjacent independent set node and thus, are removed from the graph for the next
iteration.
II) Low-Degree Local Shattering
Once the degree has dropped to ∆(cid:48) = poly log n, we apply the shattering part of the LOCAL MIS algo-
rithm of Ghaffari [Gha16], which runs in O(log ∆(cid:48)) = O(log log n) rounds and w.h.p. leads to connected
components of size poly∆(cid:48) · log n = poly log n in the remainder graph. Observe that the simulation of
this algorithm in the MPC model is straightforward.
Post-Shattering
We gather the connected components of size poly log n and solve the remaining problem locally.
2 Algorithm Overview and Roadmap
In this section, we give the formal statements we need to prove our main result, and provide an overview
of the structure of the remainder of the paper. We start with a result that is repeatedly used to gather
all nodes of a connected component onto one machine, provided that they fit there.
It will come in
two variants, which naturally give rise to Theorems 1.2 and 1.3, respectively. The proof is deferred to
Section 4 (part a)) and the full version [BFU18a] (part b)).
Lemma 2.1 (Gathering). Let G be an n-node graph and G(cid:48) any n(cid:48)-node subgraph of G consisting of
connected components of size at most k = O(cid:0)nε/3(cid:1) and diameter at most d. Then there are
a) an O(log d)-round low-memory MPC algorithm with M = (cid:101)O(cid:0)n1−ε/3(cid:1) machines and
b) an O(log d · log log n)-round low-memory MPC algorithm with M = (cid:101)O(cid:0)n1−ε(cid:1) machines, if n(cid:48) · d3 =
O(n),
that compute an assignment of nodes to machines so that all the nodes of a connected component of G(cid:48)
are on the same machine.
Next, we will provide the results corresponding to the two main parts of our algorithm, the shattering
and the post-shattering.
4Note that in the MPC model it is easy to keep track of the maximum degree.
5Gathering the connected components means loading all the nodes of a connected component onto the same machine.
5
Lemma 2.2 (Shattering). There are
a) an O(log log n · log log ∆)-round low-memory MPC algorithm that uses M = (cid:101)O(n1−α/3) machines
b) an O(log2 log n · log log ∆)-round low-memory MPC algorithm with M = (cid:101)O(n1−α) machines
and
that compute an independent set on an n-node tree with maximum degree ∆ so that the remainder graph,
after removal of the independent set nodes and their neighbors, w.h.p. has only components of size at
most poly log n.
The proof of this Shattering Lemma can be found in Section 3. The following Post-Shattering Lemma
is a direct consequence of the Gathering Lemma.
Lemma 2.3 (Post-Shattering). There are
a) an O(log k)-round low-memory MPC algorithm with M = (cid:101)O(cid:0)n1−ε/3(cid:1) machines and
b) an O(log k · log log n)-round low-memory MPC algorithm with M = (cid:101)O(cid:0)n1−ε(cid:1) machines
that find an MIS in an n-node graph consisting of connected components of size k = O(cid:0)nε/3(cid:1).
Proof. By Lemma 2.1, we can gather the connected components in O(log k) rounds. Then, an MIS of
each connected component can be computed locally. Note that Theorem 1.1 by Ghaffari [Gha16] certifies
that the number of nodes remaining after our shattering process can be made small enough to satisfy the
conditions required by Lemma 2.1.
Note that the naive simulation of the corresponding LOCAL post-shattering algorithm [Gha16, PS92]
√
would lead to a round complexity of 2O(
log log n).
We now put together the results to prove Theorems 1.2 and 1.3.
Proof of Theorems 1.2 and 1.3. We apply the shattering algorithm from Lemma 2.2 to get an indepen-
dent set, with connected components of size k = poly log n in the remainder graph. Then we run the
post-shattering algorithm from Lemma 2.3 to find an MIS in all these components. The combination of
the initial independent set found by the shattering and all the MIS found by the post-shattering results
in an MIS in the original tree.
Memory per Machine below ∆: If the degree of a node is larger than the local memory, one needs to
store several lower-degree copies of this node on different machines. Here, we give a short argument for
why one can assume without loss of generality that all incident edges of a node are stored on the same
machine. Notice that in a tree with n nodes, there can be at most n1−ε/2 nodes with degree at least
nε/2. If we now just ignore all these high-degree nodes and find an MIS among the remaining nodes, the
resulting graph, after removal of all MIS nodes and their neighbors, has at most n1−ε/2 nodes. Repeating
this argument roughly 2/ε times gives an MIS in the whole input graph.
3 Shattering
Lemma 3.1 (Iterated Subsample-and-Conquer). There are
a) an O(log1+ε log ∆)-round low-memory MPC algorithm with M = (cid:101)O(n1−α/3) machines and
b) an O(log1+ε log ∆ · log log n)-round low-memory MPC algorithm with M = (cid:101)O(n1−α) machines
that compute an independent set on an n-node tree with maximum degree ∆ such that the remainder graph,
after removal of the independent set nodes and their neighbors, w.h.p. has maximum degree poly log n.
6
The proof of this lemma can be found in Section 3.1.
Lemma 3.2 (Low-Degree Local Shattering [Gha16]). There is an O(log ∆)-round LOCAL algorithm that
computes an independent set on an n-node graph with maximum degree ∆ so that the remainder graph,
after removal of all nodes in the independent set and their neighbors, w.h.p. has connected components of
size poly∆ · log n.
We now combine these two results to prove Lemma 2.2.
Proof of Lemma 2.2. We apply the algorithm of Lemma 3.1, w.h.p. yielding an independent set with a
remainder graph that has maximum degree ∆(cid:48) = poly log n. On this low-degree graph, we simulate the
LOCAL algorithm of Lemma 3.2 in a straight-forward manner, which takes O(log ∆(cid:48)) = O(log log n)
rounds and w.h.p. leaves us with connected components of size poly∆(cid:48) · log n = poly log n.
3.1 Degree Reduction via Iterated Subsampling
We prove the following result, and then show how it can be used to prove Lemma 3.1. For the purposes
of the proof of Lemma 3.3 we assume that ∆ is a large enough poly log n in order to be able to apply
Lemma 2.1. Notice that from the perspective of the final runtime, the exponent of the logarithm turns
into a constant factor hidden in the O-notation.
Lemma 3.3. There are
a) an O(log log n)-round low-memory MPC algorithm with M = (cid:101)O(n1−α/3) machines and
b) an O(log2 log n)-round low-memory MPC algorithm with M = (cid:101)O(n1−α) machines
that compute an independent set on an n-node tree G with maximum degree ∆ = Ω(poly log n) such that
the remainder graph, after removal of the independent set nodes and their neighbors, w.h.p. has maximum
degree at most ∆(1+δ(cid:48))δ, for some δ = Θ (1/(1 + ε)) and any δ(cid:48) > 0.
Proof. We first outline the algorithm and then slowly go through the steps of the algorithm again while
proving its key properties.
Algorithm: Every node is sampled independently with probability ∆−δ into a set V (cid:48). The connected
components of G(cid:48) = G[V (cid:48)] are gathered by Lemma 2.1, and one of the two 2-colorings is picked uniformly
at random, independently for every connected component. This can be done locally. All the black nodes,
say, are added to the MIS, and are removed from the graph along with their neighbors. See Fig. 1 in
Section 4.2.
Subsampling: We first prove that the random subsampling leads to nice properties of the graph induced
by subsampled nodes.
Claim 3.4. After the subsampling, w.h.p., the following holds.
i) Every connected component of G(cid:48) has diameter O(cid:0) 1
δ · log∆ n(cid:1).
ii) Every connected component of G(cid:48) consists of nO((1−δ)/δ) nodes.
(cid:16)
∆(1+δ(cid:48))δ(cid:17)
iii) Every node with degree Ω
in G has degree Ω(poly log n) in G(cid:48).
Proof. Consider an arbitrary path of length (cid:96) = Ω(cid:0) 1
δ · log∆ n(cid:1) in G. This path is in G(cid:48) only if all its nodes
also the diameter of every connected component, in G(cid:48) is bounded by O(cid:0) 1
δ · log∆ n(cid:1). Since the degree
among the subsampled nodes is bounded by O(cid:0)∆1−δ(cid:1), w.h.p., which is a simple application of Chernoff
are subsampled into V (cid:48), which happens with probability at most ∆−δ·(cid:96) = 1
polyn . A union bound over all --
at most n2 many -- paths in the tree T shows that, w.h.p., the length of every path, and hence in particular
7
Figure 1: Consider a non-subsampled high-degree node v. While most of its neighbors will not be subsampled (gray
nodes), there will still be many, say k, subsampled neighbors. Since we are in a tree, all the subtrees G1, . . . Gk of
subsampled neighbors of v are disjoint and thus colored independently. Node v will thus, w.h.p., see at least one
black neighbor. This node will be added to the MIS, leading to the removal of v from the graph.
and union bound, it follows that every connected component consists of at most O(cid:0)∆(1−δ)·(cid:96)(cid:1) = nO((1−δ)/δ)
(cid:16)
∆(1+δ(cid:48))δ(cid:17)
nodes. Finally, another simple Chernoff and union bound argument shows that every node with degree
= Ω(poly log n) neighbors in G(cid:48), which concludes the
Ω
proof of Claim 3.4.
Gathering: Since G(cid:48) consists of components that have a low diameter by Claim 3.4 i) and that are small
enough to fit on a single machine by Claim 3.4 ii) -- provided that δ = Θ (1/(1 + ε)) is chosen such that
the components have size O(cid:0)nε/3(cid:1) -- we can gather them efficiently by Lemma 2.1, in either O(log log n)
in the graph G has at least Ω
(cid:16)
∆δ(cid:48)·δ(cid:17)
or O(log2 log n) rounds. The random MIS can then be easily computed locally.
Random MIS: It remains to show that every high-degree node in G, w.h.p., has at least one adjacent
node that joins the random MIS, which leads to the removal of this high-degree node from the graph.
Note that this is trivially true for all subsampled nodes, by maximality of an MIS.
Now consider an arbitrary non-subsampled node v with degree Ω
and its Ω(poly log n)
subsampled neighbors, by Claim 3.4 iii). Observe that, since we are in a tree and thus in particular in
a triangle-free graph, there cannot be edges between these neighbors. Therefore no two neighbors of a
non-subsampled node belong to the same connected component in G(cid:48), which means that all the neighbors
in V (cid:48) of v are colored independently, and hence are added to an MIS independently with probability 1/2.
By the Chernoff inequality, w.h.p. at least one of v's neighbors must have been added to an MIS, and a
union bound over all nodes concludes the proof of the degree reduction, and hence of Lemma 3.3.
(cid:16)
∆(1+δ(cid:48))δ(cid:17)
Proof of Lemma 3.1. Follows from log
1
(1+δ(cid:48))δ
log ∆ = log1+ε log ∆ many applications of Lemma 3.3.
4 Gathering Connected Components
In this section, we provide a proof of the Gathering Lemma in Lemma 2.1. Our approach is essentially a
tuned version of the Hash-to-Min algorithm by Chitnis et al. [CDSMR13] and the graph exponentiation
idea by Lenzen and Wattenhofer [LW10]. Notice that, however, Chitnis et al. only show an O(log n)
bound for the round complexity; it is not possible to just use their method as a black box. The section is
divided into two subsections, where we first give a simple and fast but memory-inefficient algorithm and
then present a slightly slower algorithm that only needs a constant space overhead.
In very recent works, independent of this paper, Andoni et al. [ASS+18] and Assadi et al. [ASW18]
studied, among other problems, finding connected components in the low-memory setting of MPC. In
8
······u1u2ukG2GkG1v (a)
(b)
Figure 2: We illustrate the gathering algorithm with help of the tree depicted in Fig. 2a. The edges added by node
u are illustrated in Fig. 2b by dashed arcs. Fig. 2c displays how the edges added by nodes v1 and v2, drawn as
dashed arcs, shortcut the shortest path between nodes s and t.
(c)
particular, Andoni et al. give algorithms to find connected components and to root a forest with constant
success probability, with O(m) total memory in time O(log d · log log n). While their results are more
general, ours have the advantages of being (arguably) much simpler and deterministic. Furthermore, to
turn their algorithm to work with high probability, the straightforward approach requires a logarithmic
overhead in the total memory.
We present the naive gathering algorithm in Section 4.1 and the in-space gathering in Section 4.2.
4.1 Naive Gathering
We first present the algorithm. The underlying idea of the algorithm is to find a minimum-ID6 node
within every component and to create a virtual graph that connects all the nodes of that component to
this minimum-ID node, the leader.
Gathering Algorithm: In every round, every node u completes its 1-hop neighborhood to a clique.
Once a round is reached in which there are no more edges to be added, u stops and selects its minimum-ID
neighbor as its leader. We refer to Fig. 2 in Section 4.2 for an illustration.
Observe that once there is a round in which u does not add any edges, the component of u forms a
clique, and thus all nodes in this component have the same leader, namely the minimum-ID node in this
clique. Next, we prove that this algorithm terminates quickly.
Claim 4.1. The gathering algorithm takes O(log d) rounds on a graph with diameter d.
Proof. Consider any shortest path u1, . . . , u(cid:96) of length 2 ≤ (cid:96) ≤ d. After the first round, every ui gets
connected to ui−2 and ui+2 for 2 < i < (cid:96)−1. Thus, the diameter of the new graph is at most (cid:100)2d/3(cid:101). After
O(log d) iterations, the diameter within each component has reduced to 1, and the algorithm halts.
It remains to show that not too many edges are added, so that the virtual graph of any component
still fits into the memory of a single machine.
Claim 4.2. The number of edges in the virtual graph created by the gathering algorithm in a component
of size k is O(k3).
Proof. During the execution of the algorithm, each node in a component may create an edge between
any other two nodes in the corresponding component, thus at most k3.
6We assume without loss of generality that every node has a unique identifier. If not, every node can draw an O(log n)-bit
identifier at random, which w.h.p. will be unique.
9
uv1v2st Since we require the components to be of size at most O(nε/3), the previous claim guarantees that
the virtual graph of any connected component indeed fits into the memory. So as to not overload any
machine with too many components, we assume that the shuffling distributes the components to the
machines in an arbitrary feasible way, e.g., greedily7.
Remark 4.3. A weakness of the gathering algorithm is that we need O(k3) memory to store a connected
component of size k, even if this component originally just consisted of as few as k − 1 edges. This is
because a single edge can exist on up to k machines. In the worst case, the required memory is blown up
by a power 3. This leads to a super-linear overall memory requirement, that is, we need roughly N 1+2ε/3
total memory in the system. Notice that this can be implemented either by adding more machines or by
adding more memory to the machines, since we do not care on which machines the resulting components
lie, as long as they fit the memory.
4.2 In-Space-Gathering in Trees
The simple and naive gathering algorithm can be very wasteful in terms of space usage over the whole
system. In this section, we provide a fine-tuned version of the gathering method that works, asymptoti-
cally, in space, thereby proving part b) of Lemma 2.1. In other words, the total space requirement drops
to O(n). Informally, our algorithm first turns every connected component into a rooted tree and then
determines which nodes are contained in the same tree component by making sure that each node learns
the ID of the root of its tree. For the latter part, we prove the following.
Lemma 4.4. There is an O(log d)-round low-memory MPC algorithm that works in an n-node forest of
rooted trees with maximum diameter d and, for every node, determines the root of the corresponding tree.
The algorithm requires M = O(cid:0)n1−ε(cid:1) machines.
Proof. Let parent(v) denote the parent of node v and define parent(r) = r for a root node r. Consider
the following pointer-forwarding algorithm that is run in parallel for every node v. In every round, for
every child u of v, we set parent(u) := parent(v). The process terminates once v points to a root, i.e., to
a node r for which parent(r) = r. Notice that after every step, following the parent pointers still leads
to the root node.
Let (v1, v2, . . . , vk) be the directed path from node v1 to the root r = vk of its subtree in round t.
After one round of the algorithm, every vi is connected to vmin{i+2,k}. Thus, the length of the path is at
most (cid:100)k/2(cid:101). After O(log k) = O(log d) rounds the algorithm terminates yielding the claim.
Root a Tree: Given Lemma 4.4, what remains to show for our algorithm is how to root a tree. The idea
is to once more use the graph exponentiation method to learn an (cid:96)-hop neighborhood of a node in log (cid:96)
steps. However, in order to prevent the space requirement from getting out of hand, each node performs
only a bounded number of exponentiation steps, after which all nodes that already know their parent in
the output orientation are removed from the graph. Then this process is iterated until at most one node
(per connected component) remains.
Tree-Rooting Algorithm A: In the following, we give a formal description of an algorithm A for
rooting a tree of diameter d. The algorithm takes an integer B as input parameter that describes the
initial memory budget for each node v, i.e., an upper bound on the number of edges that v may add
before the first node removal. The execution of A is subdivided in phases i = 0, 1, . . . which consist of
O(log d) rounds each. Set B0 = B.
Phase i of A: In phase i, each node v does the following:
In round 0, node v sets its local budget Bv to Bi.
In each following round j = 1, 2, . . ., node v
first connects its 1-hop neighborhood to a clique by adding edges between all its neighbors that are not
7An alternative and simple way to prevent overloading is to add an O(log n) factor of memory per machine and consider
a random assignment of components to machines as a balls-into-bins process.
10
connected yet, but it does so only if the number of added edges is at most Bv. Then v updates its local
budget by decreasing Bv by the number of edges that v added. If Bv was not large enough to connect v's
1-hop neighborhood to a clique, then v does not add any edges in round j. This concludes the description
of round j, of which there are O(log d) many.
Denote the tree at the beginning of phase i by Ti, and for each neighbor u of v, denote the set of
u(v). Phase i concludes with a number of special rounds:
nodes that are closer to u than v in Ti by Si
First, v checks whether it has a neighbor u(cid:48) in Ti with the following properties:
i) Si
u(v) is contained in the current 1-hop neighborhood of v, for each neighbor u of v in Ti satisfying
u (cid:54)= u(cid:48).
ii) Si
u(cid:48)(v) is not (entirely) contained in the current 1-hop neighborhood of v.
If such a neighbor u(cid:48) exists (which, by definition, is unique), then v sets parent(v) = u(cid:48). Second, v
removes all edges that it added during phase i (regardless of whether a parent is set). Third, v is removed
from Ti if it already chose its parent, i.e., if it set parent(v). Fourth, the budget per node is updated,
by setting Bi+1 = Bi · ni/ni+1, where ni and ni+1 are the numbers of nodes of Ti and Ti+1, respectively.
This concludes the description of phase i.
We execute this process until at most one node remains.
Termination of A: Since in each phase (at the very least) all leaves are removed, this process eventually
terminates.
It is straightforward to check that if a node v chooses its parent u(cid:48) = parent(v) in phase i, then any
neighbor u (cid:54)= u(cid:48) of v in Ti also chooses its parent in phase i, and, what is more, u chooses v as its parent
(which, combined with the following observations, shows that the orientation of the input tree induced
by the parent choices of the nodes yields indeed a rooted tree). Hence, given the above process, one of
two things happens in the end: either exactly one node remains, or all nodes are removed but there is
exactly one pair of nodes that chose each other as their parent. In the former case, no action has to be
taken, as the remaining node is simple the root of our rooted tree. In order to handle the latter case,
we add a simple fifth special round at the end of each phase i: Each node v removed in phase i checks
whether the node it chose as its parent chose v as its parent. If this is the case, then the node with the
higher id removes its choice of parent and becomes the root node of the input tree. See Fig. 3 for an
illustration of algorithm A.
11
(a) The leaves have only one neighbor, which be-
comes their parent. Nodes u1, u2, and v do not have
enough budget to add edges.
(b) Once the leaves are removed once, enough bud-
get is freed for nodes u1 and u2 to add edges that
connect their neighbors.
(c) Once nodes u1 and u2 know that v is their parent, node v can focus its budget to the remainder of the
tree illustrated by the gray area.
Figure 3: An illustration of the (local) steps performed by algorithm A. For the sake of this example, assume that
the initial budget is 2. As illustrated in Fig. 3a, leaves are always able to determine their parent. Assuming that
the tree has non-leaf minimum degree 3, removing all the leaves at least roughly doubles the budget of all nodes.
Thus, in the second step, illustrated in Fig. 3b, nodes with degree at most 4 are able to complete their 1-hop
neighborhoods into a clique. Small subtrees rooted at (or connected to) v are removed quickly in our process, as
illustrated in Fig. 3c, and therefore, node v requires large subtrees to survive for many phases.
Running Time of A: We present a number of lemmas in order to determine the runtime of algorithm
A. Here, a subtree T (v) rooted at some node v corresponds to the descendants of v in the rooted tree T
returned by A (or in the rooted subtree of T induced by the nodes of some Ti).
Lemma 4.5. Consider some arbitrary phase i, and let T (v) be the subtree of Ti rooted at v. If T (v) ≤
√
Bi, then v chooses its parent in phase i and is removed from the tree.
an edge between nodes from T (v). Since T (v) ≤ √
Proof. Let k be some arbitrary non-negative integer, and consider any node u in T (v) with distance at
least 2k to v. Observe that, according to A, the distance between any two nodes in Ti decreases by a factor
of at most 2 per round. Hence, after round k of phase i, all nodes contained in the 1-hop neighborhood
of u are actually also contained in T (v). Thus, each edge that u would have added if it had connected its
1-hop neighborhood to a clique in each of the rounds 1, . . . , k + 1, disregarding any budget constraints, is
Bi, the number of edges between nodes from T (v) is
bounded from above by Bi; it follows that u had enough budget to indeed connect its 1-hop neighborhood
to a clique in each round up to and including round k + 1.
Now consider any node w whose distance to v in Ti is at least 2k, but at most 2k+1− 1. Let w0, . . . , wk
be nodes on the unique path between v and w with distance 20, . . . , 2k to v. Due to the observations
above, it is straightforward to check that, in each round 1 ≤ h ≤ k, node wh−1 connects node wh to v,
while in round k + 1, node wk connects node u to v. Since the depth of T (v) is upper bounded by log d,
it follows that after log d rounds, all nodes from T (v) are contained in v's 1-hop neighborhood. Hence,
v will choose the only neighbor that is not contained in T (v) as its parent, and v is removed in phase i.
Since k was chosen arbitrarily, the lemma statement follows.
Lemma 4.6. Let T be a rooted tree with n nodes and diameter at most d. Let 1 ≤ α ≤ n, and let C be
the set of nodes v with the property that T (v) ≤ α. Then, C ≥ n · (α/(d + α)).
12
vu1u2···vu1u2···vu1u2··· Proof. Assign one dollar to each node that is not contained in a subtree of size at most α. Every such
node then distributes its dollar evenly among all of its descendants in C. Note that, for each leaf node w
of the tree obtained from T by deleting all nodes in C, the number of descendants of w in C is at least
α since otherwise w would be in C, by the definition of C. Hence, all nodes that are not contained in C
have at least α descendants in C.
Consider then any node v ∈ C. Since the diameter of the tree is d, node v can have at most d
ancestors in T . Every ancestor of v distributes at most 1/α dollars to v and therefore, v receives at most
d/α dollars. As the amount of dollars did not change during its redistribution from nodes not contained
in C to nodes in C, we can conclude that C·(d/α) ≥ n−C which implies that C ≥ n·(α/(d+α)).
Lemma 4.7. Assume that the input parameter B for our algorithm A satisfies B ≥ d3. Then the runtime
of A on trees with n nodes and diameter d is O(log d · log log n).
Proof. Observe that the sequence B0, B1, . . . of budgets at the beginning of phases 0, 1, . . . is monotonically
non-decreasing, by definition. Hence, Bi ≥ d3 for all phases i. Now consider some arbitrary phase i, and
let ni denote the number of nodes of Ti. By Lemma 4.5 and Lemma 4.6, the number of nodes that are
removed in phase i is at least ni · (
Bi)). Thus, for the new budget Bi+1, it holds by definition
that
Bi/(d +
√
√
(cid:18) d +
(cid:19)
√
Bi
Bi+1 ≥ Bi ·
1
1 − √
√
Bi
d+
Bi
= Bi ·
3
≥ B
2
i
d
.
d
, which implies Bi+5 ≥ B2
Since, as observed above, d ≤ B1/3
√
i . Recall that in
Bi/d)-fraction of nodes is removed. Thus, after O(log log n) phases, all nodes
each phase i, at least a (
(except possibly for one node) have been removed and the termination condition of A is satisfied. Since
every phase takes O(log d) time, the claim follows.
, we obtain Bi+1 ≥ B7/6
i
i
Now we have all the ingredients to prove the second part of Lemma 2.1. It is a simple corollary of
the following theorem.
Theorem 4.8. Consider a forest F of n nodes where every tree is of diameter at most d. There is an
MPC algorithm that finds the connected components of F in time O(log d·log log n) where M·S = O(n·d3).
Proof. Imagine that we run algorithm A in parallel on all trees of the input forest F , with input parameter
B = d3. There are only two parts of A that are of a global nature, i.e., where the actions of nodes do
not depend on their immediate neighborhood: the termination condition that all nodes, possibly except
for one, have been removed, and the part where the node's budgets are updated from Bi to Bi+1. The
former is easily adapted to the case of forests; each node simply terminates when itself or all its neighbors
are removed. Regarding the updating of the budget, we adapt the tree rooting algorithm as follows: we
still set the new budget Bi+1 to Bi · ni/ni+1, but now ni and ni+1 denote the total number of nodes (i.e.,
in all trees of the forest) that have not been removed yet at the beginning of phase i, resp. phase i + 1.
In the following, we verify that Lemmas 4.5, 4.6, and 4.7 also hold for forests instead of trees. In the
case of Lemma 4.5, this is obvious as the argumentation is local and thus also applies to forests. Lemma
4.6 trivially also holds for forests since the lemma statement holds for all trees in the forest. Finally, since
the argumentation of the proof of Lemma 4.7 does not make use of the fact that the input graph is a tree
except when applying Lemmas 4.5 and 4.6, it follows that Lemma 4.7 also holds for forests.
Hence, our adapted tree rooting algorithm actually transforms the forest into a rooted forest in time
O(log d · log log n). Now we can apply Lemma 4.4, and, e.g., color each component with the color of the
root node, thereby marking the connected components. Due to the runtime given in Lemma 4.4, our
total runtime is still O(log d · log log n).
It remains to show that the claimed memory constraints are satisfied. Due to the space guarantee
given in Lemma 4.4, it is sufficient to show that the memory overhead induced by adding edges during
13
the execution of out forest rooting algorithm does not exceed the allowed amount. Thus, consider the
number of edges added in an arbitrary phase i. Since each node adds at most as many edges as its budget
allows, i.e., at most Bi edges, the total number of edges added in phase i is upper bounded by ni · Bi. By
the definition of Bj+1, we have nj+1 · Bj+1 = nj · Bj, for any phase j. Hence, the value of ni · Bi is the
same for every phase i, and we obtain ni · Bi = n0 · B0 = n · d3. Therefore, the number of edges added in
any phase i does not exceed n· d3, and since all added edges are removed again at the end of each phase,
the lemma statement follows.
Remark 4.9. In the analysis, we implicitly assumed that edges incident on nodes are always added
only once. It could, however, be the case that some node is "unlucky" and many of its neighbors add a
copy of the same edge many times. This misfortune could potentially result in adding nε copies of the
same (virtual) edge, which could, in turn, overload the memory per machine constraint on the machines
containing these unlucky nodes. For the sake of simplicity, we decided to leave this problem to the
shuffling algorithm of the underlying MPC framework that can, for example, load the nodes onto the
machines greedily after each communication step. Since the total memory constraint is satisfied, this is
always feasible. Alternatively, the shuffling algorithm could simply drop duplicate messages.
14
5 Open Questions
In this paper, we introduced a variant of the MPC model in which the standard assumption of S =(cid:101)Ω(n)
memory per machine is removed.
General Graphs: We showed that in the case of the MIS problem on trees this assumption is not
necessary: Restricting the memory to nε per machine still allows an O(log2 log n)-round algorithm. The
first intriguing open problem follows.
P1 Devise a low-memory MPC algorithm that finds an MIS in general graphs in time poly log log n.
Other Fundamental Graph Problems: As an MIS of the line graph8 corresponds to a Maximal
Matching in the original graph, an MIS algorithm usually directly gives rise to a Maximal Matching
algorithm. In the MPC framework, however, it might not even be possible to store the line graph, which
seems to complicate the simulation of the MIS algorithm on the line graph. hard. the linear , i.e., a graph
where the edges of the input graph correspond to nodes in the line graph, in memory. To the best of our
knowledge, the field of Maximal Matching in the MPC world is wide open. Naturally, there are many
other standard graph problems some of which we list here.
P2 Devise an efficient low-memory MPC algorithm for the Maximal Matching problem.
P4 Devise an efficient low memory MPC algorithm for the (∆ + 1)-coloring problem.
P5 Devise an efficient low memory MPC algorithm for the (2∆ − 1)-edge-coloring problem.
P4 Devise an efficient low memory MPC algorithm that finds an O(1)-coloring in time O(log d)
in trees.
References
[ABB+19]
[ABI86]
[AK17]
Sepehr Assadi, MohammadHossein Bateni, Aaron Bernstein, Vahab S. Mirrokni, and Cliff
Stein. Coresets meet EDCS: algorithms for matching and vertex cover on massive graphs. In
Proceedings of the Thirtieth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2019, San Diego, California, USA, January 6-9, 2019, pages 1616 -- 1635, 2019.
Noga Alon, L´aszl´o Babai, and Alon Itai. A Fast and Simple Randomized Parallel Algorithm
for the Maximal Independent Set Problem. Journal of Algorithms, 7(4):567 -- 583, 1986.
Sepehr Assadi and Sanjeev Khanna. Randomized composable coresets for matching and
vertex cover. In the Proceedings of the Symposium on Parallel Algorithms and Architectures
(SPAA), pages 3 -- 12, 2017.
[ANOY14] Alexandr Andoni, Aleksandar Nikolov, Krzysztof Onak, and Grigory Yaroslavtsev. Parallel
algorithms for geometric graph problems. In Proceedings of the Symposium on Theory of
Computing (STOC), pages 574 -- 583, 2014.
[Ass17]
[ASS+18]
Sepehr Assadi.
1709.04599, 2017.
Simple round compression for parallel vertex cover.
arXiv preprint:
A. Andoni, C. Stein, Z. Song, Z. Wang, and P. Zhong. Parallel Graph Connectivity in Log
Diameter Rounds. the Proceedings of the Symposium on Foundations of Computer Science
(FOCS), pages 674 -- 685, 2018.
8A line graph is a graph with a node for every edge in the input graph, and an edge between two nodes if the corresponding
edges are incident.
15
[ASW18]
S. Assadi, X. Sun, and O. Weinstein. Massively Parallel Algorithms for Finding Well-
Connected Components in Sparse Graphs. ArXiv e-prints, 2018.
[BBD+19]
Soheil Behnezhad, Sebastian Brandt, Mahsa Derakhshan, Manuela Fischer, Mohammad-
Taghi Hajiaghayi, Richard M. Karp, and Jara Uitto. Massively parallel computation of
matching and mis in sparse graphs. Proceedings of the International Symposium on Princi-
ples of Distributed Computing (PODC), 2019.
[BDH18]
Soheil Behnezhad, Mahsa Derakhshan, and MohammadTaghi Hajiaghayi. Brief announce-
ment: Semi-mapreduce meets congested clique. arXiv preprint arXiv:1802.10297, 2018.
[Bec91]
J´ozsef Beck. An algorithmic approach to the Lov´asz local lemma. Random Structures &
Algorithms, 2(4):343 -- 365, 1991.
[BEPS16]
Leonid Barenboim, Michael Elkin, Seth Pettie, and Johannes Schneider. The Locality of
Distributed Symmetry Breaking. Journal of the ACM (JACM), 63(3):20, 2016.
[BFU18a]
Sebastian Brandt, Manuela Fischer, and Jara Uitto. Breaking the linear-memory barrier in
mpc: Fast mis on trees with strongly sublinear memory. arXiv:1802.06748, 2018.
[BFU18b]
Sebastian Brandt, Manuela Fischer, and Jara Uitto. Matching and MIS for uniformly sparse
graphs in the low-memory MPC model. CoRR, abs/1807.05374, 2018.
[BKS14]
Paul Beame, Paraschos Koutris, and Dan Suciu. Skew in parallel query processing. In Pro-
ceedings of the 33rd ACM SIGMOD-SIGACT-SIGART symposium on Principles of database
systems, pages 212 -- 223. ACM, 2014.
[BKS17]
Paul Beame, Paraschos Koutris, and Dan Suciu. Communication steps for parallel query
processing. Journal of the ACM (JACM), 64(6):40, 2017.
[CDSMR13] Laukik Chitnis, Anish Das Sarma, Ashwin Machanavajjhala, and Vibhor Rastogi. Finding
In ICDE '13, pages 50 -- 61.
connected components in map-reduce in logarithmic rounds.
IEEE Computer Society, 2013.
[CFG+18] Yi-Jun Chang, Manuela Fischer, Mohsen Ghaffari, Jara Uitto, and Zufan Zheng. The
complexity of (δ + 1) coloring in congested clique, massively parallel computation, and
centralized local computation. arXiv preprint arXiv:1808.08419, abs/1808.08419, 2018.
[C(cid:32)LM+17] Artur Czumaj, Jakub (cid:32)Lacki, Aleksander Madry, Slobodan Mitrovi´c, Krzysztof Onak, and
Piotr Sankowski. Round compression for parallel matching algorithms. arXiv preprint:
1707.03478, 2017.
[CPPU15] Matteo Ceccarello, Andrea Pietracaprina, Geppino Pucci, and Eli Upfal. Space and time
efficient parallel graph decomposition, clustering, and diameter approximation. In the Pro-
ceedings of the Symposium on Parallel Algorithms and Architectures (SPAA), pages 182 -- 191,
2015.
[CV86]
Richard Cole and Uzi Vishkin. Deterministic Coin Tossing and Accelerating Cascades:
Micro and Macro Techniques for Designing Parallel Algorithms. In Symposium on Theory
of Computing, pages 206 -- 219, 1986.
[DG08]
Jeffrey Dean and Sanjay Ghemawat. Mapreduce: simplified data processing on large clusters.
Communications of the ACM, 51(1):107 -- 113, 2008.
16
[GGK+18] Mohsen Ghaffari, Themis Gouleakis, Christian Konrad, Slobodan Mitrovic, and Ronitt Ru-
binfeld. Improved massively parallel computation algorithms for mis, matching, and vertex
cover. In Proceedings of the International Symposium on Principles of Distributed Comput-
ing (PODC), page to appear, 2018.
[Gha16]
Mohsen Ghaffari. An improved distributed algorithm for maximal independent set. In the
Proceedings of ACM-SIAM Symposium on Discrete Algorithms (SODA), 2016.
[GKU19]
Mohsen Ghaffari, Fabian Kuhn, and Jara Uitto. personal communication, 2019.
[GSZ11]
[GU19]
Michael T Goodrich, Nodari Sitchinava, and Qin Zhang. Sorting, searching, and simulation
in the mapreduce framework. In International Symposium on Algorithms and Computation,
pages 374 -- 383. Springer, 2011.
Mohsen Ghaffari and Jara Uitto. Sparsifying distributed algorithms with ramifications in
massively parallel computation and centralized local computation.
In the Proceedings of
ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1636 -- 1653, 2019.
[KMVV15] Ravi Kumar, Benjamin Moseley, Sergei Vassilvitskii, and Andrea Vattani. Fast greedy al-
gorithms in mapreduce and streaming. ACM Transactions on Parallel Computing (TOPC),
2(3):14, 2015.
[KSV10]
Howard Karloff, Siddharth Suri, and Sergei Vassilvitskii. A Model of Computation for
MapReduce. In the Proceedings of ACM-SIAM Symposium on Discrete Algorithms (SODA),
pages 938 -- 948, 2010.
[LMSV11]
Silvio Lattanzi, Benjamin Moseley, Siddharth Suri, and Sergei Vassilvitskii. Filtering: a
method for solving graph problems in mapreduce. In the Proceedings of the Symposium on
Parallel Algorithms and Architectures (SPAA), pages 85 -- 94, 2011.
[Lub86]
[LW10]
[LW11]
[Ona18]
Michael Luby. A Simple Parallel Algorithm for the Maximal Independent Set Problem.
SIAM Journal on Computing, 15(4):1036 -- 1053, 1986.
Christoph Lenzen and Roger Wattenhofer. Brief Announcement: Exponential Speed-Up of
Local Algorithms Using Non-Local Communication. In 29th Symposium on Principles of
Distributed Computing (PODC), Zurich, Switzerland, July 2010.
Christoph Lenzen and Roger Wattenhofer. MIS on Trees. In Proceedings of the International
Symposium on Principles of Distributed Computing (PODC), pages 41 -- 48, 2011.
Krzysztof Onak. Round compression for parallel graph algorithms in strongly sublinear
space. CoRR, abs/1807.08745, 2018.
[PPR+12] Andrea Pietracaprina, Geppino Pucci, Matteo Riondato, Francesco Silvestri, and Eli Upfal.
In Proceedings of the International
Space-round tradeoffs for mapreduce computations.
Conference on Supercomputing, pages 235 -- 244. ACM, 2012.
[PRS16]
[PS92]
Gopal Pandurangan, Peter Robinson, and Michele Scquizzato. Fast distributed algorithms
for connectivity and mst in large graphs. In Proceedings of the 28th ACM Symposium on
Parallelism in Algorithms and Architectures, pages 429 -- 438. ACM, 2016.
Alessandro Panconesi and Aravind Srinivasan. Improved distributed algorithms for color-
ing and network decomposition problems. In Proceedings of the Symposium on Theory of
Computing (STOC), pages 581 -- 592. ACM, 1992.
17
[RVW16]
[YV18]
Tim Roughgarden, Sergei Vassilvitskii, and Joshua R Wang. Shuffles and circuits:(on lower
bounds for modern parallel computation). In Proceedings of the 28th ACM Symposium on
Parallelism in Algorithms and Architectures, pages 1 -- 12. ACM, 2016.
Grigory Yaroslavtsev and Adithya Vadapalli. Massively parallel algorithms and hardness
for single-linkage clustering under lp distances.
In Proceedings of the 35th International
Conference on Machine Learning, ICML 2018, Stockholmsmassan, Stockholm, Sweden, July
10-15, 2018, pages 5596 -- 5605, 2018.
18
|
1810.10965 | 1 | 1810 | 2018-10-18T13:18:49 | Towards a compact representation of temporal rasters | [
"cs.DS"
] | Big research efforts have been devoted to efficiently manage spatio-temporal data. However, most works focused on vectorial data, and much less, on raster data. This work presents a new representation for raster data that evolve along time named Temporal k^2 raster. It faces the two main issues that arise when dealing with spatio-temporal data: the space consumption and the query response times. It extends a compact data structure for raster data in order to manage time and thus, it is possible to query it directly in compressed form, instead of the classical approach that requires a complete decompression before any manipulation. In addition, in the same compressed space, the new data structure includes two indexes: a spatial index and an index on the values of the cells, thus becoming a self-index for raster data. | cs.DS | cs |
Towards a compact representation of temporal
rasters (cid:63)
Ana Cerdeira-Pena1, Guillermo de Bernardo1,2, Antonio Farina1,
Jos´e R. Param´a1, and Fernando Silva-Coira1
1 Universidade da Coruna, Fac. Inform´atica, CITIC, Spain
2 Enxenio S.L.
Abstract. Big research efforts have been devoted to efficiently manage
spatio-temporal data. However, most works focused on vectorial data,
and much less, on raster data. This work presents a new representation
for raster data that evolve along time named Temporal k2raster. It faces
the two main issues that arise when dealing with spatio-temporal data:
the space consumption and the query response times. It extends a com-
pact data structure for raster data in order to manage time and thus, it is
possible to query it directly in compressed form, instead of the classical
approach that requires a complete decompression before any manipula-
tion. In addition, in the same compressed space, the new data structure
includes two indexes: a spatial index and an index on the values of the
cells, thus becoming a self-index for raster data.
1
Introduction
Spatial data can be represented using either a raster or a vector data model [6].
Basically, vector models represent the space using points and lines connecting
those points. They are used mainly to represent man-made features. Raster
models represent the space as a tessellation of disjoint fixed size tiles (usually
squares), each one storing a value. They are traditionally used in engineering,
modeling, and representations of real-word elements that were not made by
men, such as pollution levels, atmospheric and vapor pressure, temperature,
precipitations, wind speed, land elevation, satellite imagery, etc.
Temporal evolution of vectorial data has been extensively studied, with a
large number of data structures to index and/or store spatio-temporal data.
Examples are the 3DR-tree [14], HR-tree [10], the MVR-tree [13], or PIST [3].
In [9] the classical Map Algebra of Tomlin for managing raster data is ex-
tended to manage raster data with a temporal evolution. The conceptual solution
(cid:63) Funded in part by European Union's Horizon 2020 research and innovation pro-
gramme under the Marie Sklodowska-Curie grant agreement No 690941 (project
BIRDS); by Xunta de Galicia/FEDER-UE [CSI: ED431G/01 and GRC: ED431C
2017/58]; by MINECO-AEI/FEDER-UE [Datos 4.0: TIN2016-78011-C4-1-R; Ve-
locity: TIN2016-77158-C4-3-R; and ETOME-RDFD3: TIN2015-69951-R]; and by
MINECO-CDTI/FEDER-UE [INNTERCONECTA: uForest ITC-20161074].
is simple, instead of considering a matrix, it considers a cube, where each slice
of the temporal dimension is the raster corresponding to one time instant.
Most real systems capable of managing raster data, like Rasdaman, Grass, or
even R are also capable of managing time-series of raster data. These systems,
as well as raster representation formats such as NetCDF (standard format of
the OGC3) and GeoTiff, rely on classic compression methods such as run length
encoding, LZW, or Deflate to reduce the size of the data. The use of these
compression methods poses an important drawback to access a given datum or
portion of the data, since the dataset must be decompressed from the beginning.
Compact data structures [7, 11] are capable of storing data in compressed
form and enable us to access a given datum without the need for decompressing
from the beginning. In most cases, compact data structures are equipped with an
index that provides fast access to data. There are several compact data structures
designed to store raster data [2, 8]. In this work, we extend one of those compact
data structures, the k2raster [8], to support representing time-series of rasters.
2 Related work
In this section, we first revise the k2tree, a compact data structure that can be
used to represent binary matrices. Then, we also present several compact data
structures for representing raster data containing integers in the cells. We pay
special attention to discuss one of them, the k2raster, which is the base of our
proposal Temporal k2raster (T−k2raster).
k2tree: The k2tree [5] was initially designed to represent web graphs, but it also
allows to represent binary matrices, that is, rasters where the cells contain only
a bit value. It is conceptually a non-balanced k2-ary tree built from the binary
matrix by recursively dividing it into k2 submatrices of the same size. First, the
original matrix is divided into k2 submatrices of size n2/k2, being n×n the size
of the matrix. Each submatrix generates a child of the root whose value is 1 if it
contains at least one 1, and 0 otherwise. The subdivision continues recursively for
each node representing a submatrix that has at least one 1, until the submatrix
is full of 0s, or until the process reaches the cells of the original matrix (i.e.,
submatrices of size 1×1).
The k2tree is compactly stored using just two bitmaps T and L. T stores
all the bits of the conceptual k2tree, except the last level, following a level-wise
traversal: first the bit values of the children of the root, then those in the second
level, and so on. L stores the last level of the tree.
It is possible to obtain any cell, row, column, or window of the matrix very
efficiently, by running rank and select operations [7] over the bitmaps T and L.
k3tree: The k3tree [2] is obtained by simply adding a third dimension to the
k2tree, and thus, it conceptually represents a binary cube. This can be trivially
done by using the same space partitioning and representation techniques from
the k2tree, yet applied to cubes rather than to matrices.
3 http://www.opengeospatial.org/standards/netcdf
Thus, each 1 in the binary cube represents a tuple (cid:104)x, y, z(cid:105), where (x, y) are
the coordinates of the cell of the raster and z is the value stored in that cell.
k2acc: The k2acc [2] representation for a raster dataset is composed by as many
k2trees as different values can be found in the raster. Given t different values in
the raster: v1 < v2 < . . . < vt, k2acc contains K1, K2, . . . , Kt k2trees, where each
Ki has a value 1 in those cells whose value is v ≤ vi.
2D-1D mapping: In [12], it is presented a method that uses a space-filling
curve to reduce the raster matrix to an array, and the use of one dimensional
index (for example a B-tree) over that array to access the data.
k2raster: k2raster has proven to be superior in both space and query time [12,
8] to all the other compact data structures for storing rasters. In [8], it was also
compared with NetCDF. It drew slightly worse space needs than the compressed
version (that uses Deflate) of NetCDF, but queries performed noticeably faster.
k2raster is based in the same partitioning method of the k2tree, that is, it
recursively subdivides the matrix into k2 submatrices and builds a conceptual
tree representing these subdivisions. Now, in each node, instead of having a
single bit, it contains the minimum and maximum values of the corresponding
submatrix. The subdivision stops when the minimum and maximum values of
the submatrix are equal, or when the process reaches submatrices of size 1×1.
Again the conceptual tree is compactly represented using, in addition to binary
bitmaps, efficient encoding schemes for integer sequences.
More in detail, let n×n be the size of the input matrix. The process begins
by obtaining the minimum and maximum values of the matrix. If these values
are different, they are stored in the root of the tree, and the matrix is divided
into k2 submatrices of size n2/k2. Each submatrix produces a child node of the
root storing its minimum and maximum values. If these values are the same,
that node becomes a leaf, and the corresponding submatrix is not subdivided
anymore. Otherwise, this procedure continues recursively until the maximum
and minimum values are the same, or the process reaches a 1×1 submatrix.
Figure 1 shows an example of the recursive subdivision (top) and how the
conceptual tree is built (centre-top), where the minimum and maximum values
of each submatrix are stored at each node. The root node corresponds to the
original raster matrix, nodes at level 1 correspond to submatrices of size 4×4,
and so on. The last level of the tree corresponds to cells of the original matrix.
Note, for instance, that all the values of the bottom-right 4×4 submatrix are
equal; thus, its minimum and maximum values are equal, and it is not further
subdivided. This is the reason why the last child of the root node has no children.
The compact representation includes two main parts. The first one represents
the topology of the tree (T ) and the second one stores the maximum/minimum
values at the nodes (Lmin/Lmax). The topology is represented as in the k2tree,
except that the last level (L) is not needed. The maximum/minimum values are
differentially encoded with respect to the values stored at the parent node. Again,
these values are stored as arrays following the same method of the k2tree, that is,
following the same level-wise order of the conceptual tree. By using differential
Fig. 1. Example (using k = 2) of integer raster matrix (top), conceptual tree of the
k2raster, conceptual tree with differential encoding, and final representation of the raster
matrix. Lmax and Lmin contain the maximum and minimum values of each node
following a level-wise order and using differential encoding.
encoding, the numbers become smaller. Directly Addressable Codes (DACs) [4]
take advantage of this, and at the same time, provide direct access. The last two
steps to create the final representation of the example matrix are also illustrated
in Figure 1. In the center-bottom and bottom parts, we respectively show the
tree with the differences for both the maximum and minimum values, and the
data structures that compose the final representation of the k2raster. Therefore,
the original raster matrix is compactly stored using just a bitmap T , which
represents the tree topology, and a pair of integer arrays (Lmax and Lmin),
which contain the minimum and maximum values stored at the tree. Note that
when the raster matrix contains uniform areas, with large areas of equal or
similar values, this information can be stored very compactly using differential
encoding and DACs.
The maximum/minimum values provide indexation of the stored values, this
technique is usually known as lightweight indexation. It is possible to query
the structure only decompressing the affected areas. Queries can be efficiently
computed navigating the conceptual tree by running rank and select operations
on T and, in parallel, accessing the arrays Lmax and Lmin.
455335663266333333333345554444335544443335444444344444Step19:25:35:34:35:35:54:433456:26:63:33:33:21 1110 1010 0100 1011 Tree(T): 90345 0233 0333 0010 0122 0011 0100 1022 1100 2021Local maximumvalues(Lmax): 2301 2 0 0 0 00Local mínimum values(LMin): 98777777666655669:59:76:56:67:765569778333234434335663266333333333345554444335544443335444444344444Step29877777766665566663266333333333345554444335544443335444444344444Step39877777766665566663266333333333354444544443335444444344444Step49877777766665566Level0 (root)Level1Level2Level39:24:10:01:00:00:25:222103:00:43:13:13:00:30:23:03:12:201100221000110011220Level0 (root)Level1Level2Level3rootLevel1Level2Level3Conceptual treeof thek2-rasterConceptual treeof thek2-raster (usingdifferentialencoding)Final representation 3 T−k2raster: A temporal representation for raster data
Let M be a raster matrix of size n×n that evolves along time with a timeline
of size τ time instants. We can define M= (cid:104)M1, M2, . . . , Mτ(cid:105) as the sequence of
raster matrices Mi of size n×n for each time instant i ∈ [1, τ ].
A rather straightforward baseline representation for the temporal raster ma-
trix M can be obtained by simply representing each raster matrix Mi in a
compact way with a k2raster. In this section we use a different approach. The
idea is to use sampling at regular intervals of size tδ. That is, we represent with
a k2raster all the raster matrices Ms, s = 1 + i tδ, i ∈ [0, (τ − 1)/tδ]. We will
refer to those Mi rasters as snapshots of M at time i. The tδ − 1 raster matrices
Mt, t ∈ [s + 1, s + tδ − 1] that follow a snapshot Ms are encoded using Ms as
a reference. The idea is to create a modified k2raster(cid:48) to represent Mt where, at
each step of the construction process, the values in the submatrices are encoded
as differences with respect to the corresponding submatrices in Ms rather than
as differences with respect to the parent node as usual in a regular k2raster.
With this modification, we still expect to encode small gaps for the maximum
and minimum values in each node of the conceptual tree of Mt. Yet, in addition,
when a submatrix in Mt is identical to the same submatrix in Ms, or when
all the values in both submatrices differ only in a unique gap value α, we can
stop the recursive splitting process and simply have to keep a reference to the
corresponding submatrix of Ms and the gap α (when they are identical, we
simply set α = 0). In practice, keeping that reference is rather cheap as we only
have to mark, in the conceptual tree of Mt, that the subtree rooted at a given
node p has the same structure of the one from the conceptual tree of Ms. For
such purpose, in the final representation of k2raster(cid:48), we include a new bitmap
eqB, aligned to the zeroes in T . That is, if we have T [i] = 0 (node with no
children), we set eqB[rank0(T, i)] ← 1,4 and set Lmax[i] ← α. Also, if we have
T [i] = 0, we also can set eqB[rank0(T, i)] ← 0 and Lmax[i] ← β (where β is
the gap between the maximum values of both submatrices) to handle the case
in which the maximum and minimum values in the corresponding submatrix are
identical (as in a regular k2raster).
The overall construction process of the k2raster(cid:48) for the matrix Mt related
to the snapshot Ms can be summarized as follows. At each step of the recursive
process, we consider a submatrix of Mt and the related submatrix in Ms. Let the
corresponding maximum and minimum values of the submatrix of Mt be maxt
and mint, and those of Ms be maxs and mins respectively. Therefore:
-- If maxt and mint are identical (or if we reach a 1×1 submatrix), the recursive
process stops. Being zt the position in the final bitmap T , we set T [zt] ← 0,
eqB[rank0[T, zt]] ← 0, and Lmax[zt] ← (maxt − maxs).5
4 From now on, asume rankb(B, i) returns the number of bits set to b in B[0, i − 1],
and rankb(B, 0) = 0. Note that the first index of T , eqB, Lmax, and Lmin is 0.
5 Since in k2raster(cid:48) we have to deal both with positive and negative values, we actually
apply a zig-zag encoding for the gaps (maxt − maxs).
-- If all the values in Ms and Mt differ only in a unique value α (or if they
are identical, hence α = 0), we set T [zt] ← 0, eqB[rank0[T, zt]] ← 1, and
Lmax[zt] ← (maxt − maxs).
-- Otherwise, we split the submatrix Mt into k2 parts and continue recursively.
We set T [zt] ← 1, and, as in the regular k2raster, Lmax[zt] ← (maxt−maxs),
and Lmin[rank1(zt)] ← (mint − mins).
Fig. 2. Structures involved in the creation of a T−k2raster considering τ = 3.
Figure 2 includes an example of the structures involved in the construction of
a T−k2raster over a temporal raster of size 8×8, with τ = 3. The raster matrix
corresponding to the first time instant becomes a snapshot that is represented
exactly as the k2raster in Figure 1. The remaining raster matrices Ms+1 and
Ms+2 are represented with two k2raster(cid:48) that are built taking Ms as a refer-
ence. We have highlighted some particular nodes in the differential conceptual
trees corresponding to Ms+1 and Ms+2. (i) the shaded node labeled (cid:104)0: 0(cid:105) in
Ms+1 indicates that the first 4×4 submatrix of both Ms and Ms+1 are identical.
Therefore, node (cid:104)0: 0(cid:105) has no children, and we set: T [2] ← 0, eqB[1] ← 1, and
Lmax[2] ← 0. (ii) the shaded node labeled (cid:104)1: 1(cid:105) in Ms+2 illustrates the case in
which all the values of a given submatrix are increased by α ← 1. In this case
values (cid:104)6, 6, 5, 5(cid:105) in Ms become (cid:104)7, 7, 6, 6(cid:105) in Ms+2. Again, the recursive traversal
stops at that node, and we set: T [8] ← 0, eqB[3] ← 1, and Lmax[8] ← 1 (values
are increased by 1). (iii) the shaded node labeled (cid:104)1: 2(cid:105) in Ms+1 corresponds
663266333333333345554444335544443335444444344444Ms(snapshot)774477444333333345554444335544443345544444444444Ms+19877875577777755776643446666444445554444335544443345644444445444Ms+2987777776666556698777777666655660:10:00:00:00:10:01:00:01:00:00:0110010001:11:11:00:01:210000:09:25:35:34:35:35:54:433456:26:63:33:33:20:10:00:00:00:10:02:00:02:00:00:02:12:11:01:12:30:10:01:10:00:011001110211121009:59:76:56:67:7655697783332344343351 0111 0010 0001 00100 0101 1110 0000 0010 1000 1010 10001 100 0 1 0 1 00 0 101 00 0eqBestá alineado con 0s en T eqB= 0 si min y máxiguales en el nodoeqB= 1, mismos valores que snapshot1 1111 0000 1010 0001 00100 0202 0010 2211 0000 0020 1111 2111 1011 1010 20101 1100 1 0 1 0 1010 0 0 101 00 01 1110 1010 0100 1011 90345 0233 0333 0010 0122 0011 0100 1022 1100 20212301 2 0 0 0 00T: Lmax: LMin: rootL1L2L3T: Lmax: LMin: eqB: T: Lmax: LMin: eqB: rootL1L2L3L1L2L3root to the node labeled (cid:104)3: 2(cid:105) in Ms. In this case, when we sum the maximum and
minimum values of both nodes we obtain that that node in Ms+1 has the same
maximum and minimum values (set to 4). Consequently the recursive process
stops again. In this case, we set T [7] ← 0, eqB[3] ← 0, and Lmax[7] ← 1.
4 Querying temporal raster data
In this section, we show two basic queries over T−k2raster.
Obtaining a cell value in a time instant: This query retrieves the value of a
cell (r, c) of the raster at time instant t: v ← getCellV alue(r, c, t). For solving this
query, there are two cases: if t is represented by a snapshot, then the algorithm
to obtain a cell in the regular k2raster is used, otherwise, a synchronized top-
down traversal of the trees representing that time instant (Mt) and the closest
previous snapshot (Ms) is required.
Focusing on the second case, the synchronized traversal inspects the two
nodes at each level corresponding to the submatrix that contains the queried
cell. The problem is that due to parts of Mt or Ms having the same value, the
shape of the trees representing them can be different. Therefore, it is possible
that one of the two traversals reaches a leaf, whereas the other does not. In such
a case, the traversal that did not reach a leaf, continues, but the process must
remember the value in the reached leaf, since that is the value that will be added
or subtracted to the value found when the continued traversal reaches a leaf.
Indeed, we have three cases: (a) the processed submatrix of Mt is uniform,
(b) the original submatrix of Ms is uniform and, (c) the processed submatrix
after applying the differences with the snapshot has the same value in all cells.
case. To obtain the
value stored at cell
invoked as
getCell(n, r, c, 1, 1, Lmaxs[0], Lmaxt[0]), assuming that the cell at position (0,0)
of the raster is that in the upper-left corner.
the raster matrix Mt,
Algorithm 1 shows
(r, c) of
the pseudocode of
this
it
is
zs is used to store the current position in the bitmap T of Ms (Ts) during the
downward traversal at any given step of the algorithm, similarly, zt is the position
in T of Mt (Tt). When zs (zt) has a −1 value, it means that the traversal reached
a leaf and, in maxvals (maxvalt) the algorithm keeps the maximum value stored
at that leaf node. Note that, Ts, Tt, Lmaxs, Lmaxt, and k are global variables.
In lines 1-11, the algorithm obtains the child of the processed node that
contains the queried cell, provided that in a previous step, the algorithm did not
reach a leaf node (signaled with zs/zt set to −1). In maxvals (maxvalt), the
algorithm stores the maximum value stored in that node.
If the condition in line 12 is true, the algorithm has reached a leaf in both
trees, and thus the values stored in maxvals and maxvalt are added/subtracted
to obtain the final result. If the condition of line 15 is true, the algorithm reaches
a leaf in the snapshot. This is signaled by setting zs to −1 and then a recursive
call continues the process.
The If in line 19 treats the case of reaching a leaf in Mt. If the condition of
line 20 is true, the algorithm uses bitmap eqB to check if the uniformity is in
Algorithm 1: getCell(n, r, c, zs, zt, maxvals, maxvalt) returns the value
at cell (r, c)
1 if zs (cid:54)= −1 then
5
6 end
7 if zt (cid:54)= −1 then
zs ← (rank1(Ts, zs) − 1) · k2 + 1
zs ← zs + (cid:98)r/(n/k)(cid:99) · k + (cid:98)c/(n/k)(cid:99) +1
vals ← Lmaxs[zs − 1]
maxvals ← maxvals − vals
zt ← (rank1(Tt, zt) − 1) · k2+1
zt ← zt + (cid:98)r/(n/k)(cid:99) · k + (cid:98)c/(n/k)(cid:99) +1
maxvalt ← Lmaxt[zt − 1])
2
3
4
8
9
16
20
21
22
23
24
10
11 end
12 if (zs > Ts or zs = −1 or Ts[zs] = 0) and (zt > Tt or zt = −1 or Tt[zt] = 0) then
/* Both leafs */
return maxvals + ZigZag Decoded(maxvalt)
13
14 end
15 else if zs > Ts or zs = −1 or Ts[zs] = 0 then /* Leaf in Snapshot */
zs ← −1
return getCell(n/k, r mod (n/k), c mod (n/k), zs, zt, maxvals, maxvalt)
17
18 end
19 else if zt > Tt or zt = −1 or Tt[zt] = 0 then /* Leaf in time instant */
if zt (cid:54)= −1 and Tt[zt] = 0 then
eq ← eqB[rank0(Tt, zt)]
if eq = 1 then zt ← −1 ;
else return maxvals + ZigZag Decoded(maxvalt) ;
end
return getCell(n/k, r mod (n/k), c mod (n/k), zs, zt, maxvals, maxvalt)
25
26 end
27 else /* Both internal nodes */
28
29 end
return getCell(n/k, r mod (n/k), c mod (n/k), zs, zt, maxvals, maxvalt)
the original Mt submatrix or if it is in the submatrix resulting from applying
the differences between the corresponding submatrix in Ms and Mt. A 1 in eqB
implies the latter case, and this is solved by setting zt to −1 and performing
a recursive call. A 0 means that the treated original submatrix of Mt has the
same value in all cells, and that value can be obtained adding/subtracting the
values stored in maxvals and maxvalt, since the unique value in the submatrix
of Mt is encoded as a difference with respect to the maximum value of the same
submatrix of Ms, and thus the traversal ends.
The last case is that the treated nodes are not leaves, that simply requires a
recursive call.
Retrieving cells with range of values in a time instant: (cid:104)[ri, ci](cid:105) ←
getCells(vb, ve, r1, r2, c1, c2, t) obtains from the raster of the time instant t, the
positions of all cells within a region [r1, r2] × [c1, c2] containing values in the
range [vb, ve].
Again, if t is represented with a snapshot, the query is solved with the normal
algorithm of the k2raster. Otherwise, as in the previous query, the search involves
a synchronized top-down traversal of both trees. This time requires two main
changes: (i) the traversal probably requires following several branches of both
trees, since the queried region can overlap the submatrices corresponding to
several nodes of the tree, (ii) at each level, the algorithm has to check whether
the maximum and minimum values in those submatrices are compatible with
the queried range, discarding those that fall outside the range of values sought.
5 Experimental evaluation
In this section we provide experimental results to show how T−k2raster handles
a dataset of raster data that evolve along time. We discuss both the space
requirements of our representation and its performance at query time.
We used several synthetic and real datasets to test our representation, in or-
der to show its capabilities. All the datasets are obtained from the TerraClimate
collection [1], that contains high-resolution time series for different variables,
including temperature, precipitations, wind speed, vapor pressure, etc. All the
variables in this collection are taken in monthly snapshots, from 1958 to 2017.
Each snapshot is a 4320×8640 grid storing values with 1/24◦ spatial resolution.
From this collection we use data from two variables: TMAX (maximum temper-
ature) is used to build two synthetic datasets, and VAP (vapor pressure) is com-
pressed directly using our representation. Variable TMAX is a bad scenario for
our approach, since most of the cells change their value between two snapshots.
In this kind of dataset, our T−k2raster would not be able to obtain good com-
pression. Hence, we use TMAX to generate two synthetic datasets that simulate
a slow, and approximately constant, change rate, between two real snapshots.
We took the snapshots for January and February 2017 and built two synthetic
datasets called T 100 and T 1000, simulating 100 and 1000 intermediate steps
between both snapshots; however, note that to make comparisons easier we only
take the first 100 time steps in both datasets. We also use a real dataset, VAP,
that contains all the monthly snapshots of the variable VAP from 1998 to 2017.
Note that, although we choose a rather small number of time instants in our ex-
periments, the performance of our proposal is not affected by this value: it scales
linearly in space with the number of time instants, and query times should be
unaffected as long as the change rate is similar.
We compared our representation with two baseline implementations. The
first, called k2raster6 is a representation that stores just a full snapshot for
each time instant, without trying to take advantage of similarities between close
time instants. The second baseline implementation, NetCDF, stores the different
raster datasets in NetCDF format, using straightforward algorithms on top of
the NetCDF library7 (v.4.6.1) to implement the query operations. Note that
NetCDF is a classical representation designed mainly to provide compression,
through the use of Deflate compression over the data. Therefore, it is not
designed to efficiently answer indexed queries.
We tested cell value queries (getCellValue) and range queries (getCells). We
generated sets of 1000 random queries for each query type and configuration:
1000 random cell value queries per dataset, and sets of 1000 random range
6 https://gitlab.lbd.org.es/fsilva/k2-raster
7 https://www.unidata.ucar.edu/software/netcdf/
queries for different spatial window sizes (ranging from 4×4 windows to the
whole matrix), and different ranges of values (considering cells with 1 to 4
possible values). To achieve accurate results, when the total query time for a
query set was too small, we repeated the full query set a suitable number of
times (in practice, 100 or 1000 times) and measured the average time per query.
All tests were run on an Intel (R) Core TM i7-3820 CPU @ 3.60GHz (4 cores)
with 10MB of cache and 64GB of RAM, over Ubuntu 12.04.5 LTS with kernel
3.2.0-126 (64 bits). The code is compiled using gcc 4.7 with -O9 optimizations.
T−k2raster (varying tδ)
20
6
10
8
4
k2raster
50
NetCDF (varying deflate level)
0
2
5
9
T 100
398.2 407.0 429.6 456.7 584.4 820.8
769.3
14241.3 615.3 539.5 528.0
T 1000 170.4 152.5 151.2 154.6 196.2 304.6
Table 1. Space requirements (in MB) of T−k2raster, k2raster and NetCDF over synthetic
datasets.
496.6
14241.3 435.0 344.7 323.6
Table 1 displays the space requirements for the datasets T 100 and T 1000 in
all the representations. We tested our T−k2raster with several sampling intervals
tδ, and also show the results for NetCDF using several deflate levels, from level
0 (no compression) to level 9. Our representation achieves the best compression
results in both datasets, especially in T 1000, as expected, due to the slower
change rate. In T 100, our approach achieves the best results for tδ = 4, since
as the number of changes increases our differential approach becomes much less
efficient. In T 1000, the best results are also obtained for a relatively small tδ
(6-8), but our proposal is still smaller than k2raster for larger tδ. NetCDF is only
competitive when compression is applied, otherwise it requires roughly 20 times
the space of our representations. In both datasets, NetCDF with compression
enabled becomes smaller than the k2raster representation, but T−k2raster is
able to obtain even smaller sizes.
Fig. 3. Space/time trade-off on T 100 and T 1000 datasets for cell value queries.
Figure 3 shows the space/time trade-off for the datasets T 100 and T 1000
in cell value queries. We show the results only for NetCDF with compression
enabled (deflate level 2 and 5), and for T−k2raster with a sampling interval of 6
and 50. The T−k2raster is slower than the baseline k2raster, but is much smaller
if a good tδ is selected. Note that we use two extreme sampling intervals to show
0.1 1 10 100 1000 10000 100000 300 400 500 600 700 800 900 1000query time (µs/query)space (Mbytes)T_100T-k2rasterk2rasterNetCDF 0 100 200 300 400 500 600space (Mbytes)T_1000T-k2rasterk2rasterNetCDF the consistency of query times, since in practice only the best approach in space
would be used for a given dataset. In our experiments we work with a fixed tδ, but
straightforward heuristics could be used to obtain an space-efficient T−k2raster
without probing for different periods: for instance, the number of nodes in the
tree of differences and in the snapshot is known during construction, so a new
snapshot can be built whenever the size of the tree of differences increases above
a given threshold.
T−k2raster
50
6
3.6
5.1
3.8
5.5
222.9
429.3
248.1
489.4
T 100
k2raster
2.8
3.6
163.9
301.7
NetCDF
2
6130
6240
9610
9340
5
10070
10100
15330
14790
T−k2raster
50
6
3.3
3.5
3.4
3.5
207.1 228.9
213.4 234.3
T 1000
k2raster
2.5
2.6
167.6
172.7
NetCDF
2
6160
6160
9370
9510
5
10020
10100
15110
15240
wnd rng
16
256
1
4
1
4
ALL 1
111450 126220
78250 443830 580660
79650 89380
63350 436400 568730
Table 2. Range query times over T 100 and T 1000 datasets. Times shown in µs/query
for different spatial windows (wnd) and range of values (rng).
Table 2 shows an extract of the range query times for all the representations
in datasets T 100 and T 1000. We only include here the results for T−k2raster
with a tδ of 6 and 50, and for NetCDF with deflate level 2 and 5, since query times
with the other parameters report similar conclusions. We also show the results
for some relevant spatial window sizes and ranges of values. In all the cases,
T−k2raster is around 50% slower than k2raster, due to the need of querying two
trees to obtain the results. However, the much smaller space requirements of our
representation compensate for this query time overhead, especially in T 1000.
NetCDF, that is not designed for this kind of queries, cannot take advantage of
spatial windows or ranges of values, so it is several orders of magnitude slower
than the other approaches. The last query set (ALL) involves retrieving all the
cells in the raster that have a given value (i.e. the spatial window covers the
complete raster). In this context, NetCDF must traverse and decompress the
whole raster, but our representation cannot take advantage of its spatial indexing
capabilities, so this provides a fairer comparison. Nevertheless, both T−k2raster
and k2raster are still several times faster than NetCDF in this case, and our
proposal remains very close in query times to the k2raster baseline.
Figure 4 (left) shows the space/time trade-off for the real dataset VAP. Re-
sults are similar to those obtained for the previous datasets: our representation,
T−k2raster, is a bit slower in cell value queries than k2raster, but also requires
significantly less space. The NetCDF baseline is much slower, even if it becomes
competitive in space when deflate compression is applied.
Finally, Figure 4 (right) displays the query times for all the alternatives in
range queries over the VAP dataset. The k2raster is again a bit faster than the
T−k2raster, as expected, but the time overhead is within 50%. NetCDF is much
slower, especially in queries involving small windows, as it has to traverse and
decompress a large part of the dataset just to retrieve the values in the window.
T−k2raster
50
6
k2raster
wnd rng
4 1
2
3
4
16 1
2
3
4
64 1
2
3
4
2.2
2.0
1.9
1.7
4.3
3.7
3.3
2.9
26.4
21.1
16.8
16.2
2.2
2.0
1.8
1.7
4.2
3.7
3.4
3.0
26.1
21.6
17.3
16.6
256
1
2
3
4
239.6 242.5
207.2 218.7
181.7 187.9
142.2 146.5
1.6
1.4
1.3
1.3
3.1
2.7
2.4
2.2
19.5
16.1
12.9
12.6
179.4
161.0
140.3
112.9
NetCDF
2
5570
5530
5580
5550
5670
5630
5660
5720
6150
6140
6130
6220
8720
8660
8590
8300
5
9520
9430
9430
9470
9670
9730
9660
9740
10470
10440
10450
10660
14820
14640
14430
14020
ALL 1 60400 62900
46200 411700 552500
Fig. 4. Results for VAP dataset. Left plot shows space/time tradeoff for cell value
queries. Right table shows query times for range queries. Times in µs/query.
Note that even if the window covers the complete raster, T−k2raster and k2raster
achieve significantly better query times.
6 Conclusions and future work
In this work we introduce a new representation for time-evolving raster data. Our
representation, called T−k2raster, is based on a compact data structure for raster
data, the k2raster, that we extend to efficiently manage time series. Our proposal
takes advantage of similarities between consecutive snapshots in the series, so it is
especially efficient in datasets where few changes occur between consecutive time
instants. The T−k2raster provides spatial and temporal indexing capabilities,
and is also able to efficiently filter cells by value. Results show that, in datasets
where the number of changes is relatively small, our representation can compress
the raster and answer queries very efficiently. Even if its space efficiency depends
on the dataset change rate, the T−k2raster is a good alternative to compress
raster data with high temporal resolution, or slowly-changing datasets, in small
space.
As future work, we plan to apply to our representation some improvements
that have already been proposed for the k2raster, such as the use of specific
compression techniques in the last level of the tree. We also plan to develop
an adaptive construction algorithm, that selects an optimal, or near-optimal,
distribution of snapshots to maximize compression.
0.1 1 10 100 1000 10000 100000 50 100 150 200 250 300query time (µs/query)space (Mbytes)VAP cell value queriesT-k2rasterk2rasterNetCDF References
1. Abatzoglou, J.T., Dobrowski, S.Z., Parks, S.A., Hegewisch, K.C.: Terraclimate, a
high-resolution global dataset of monthly climate and climatic water balance from
1958 -- 2015. Scientific Data 5(170191) (2017)
2. de Bernardo, G., ´Alvarez-Garc´ıa, S., Brisaboa, N.R., Navarro, G., Pedreira, O.:
Compact Querieable Representations of Raster Data. In: Proceedings of the
International Symposium on String Processing and Information Retrieval (SPIRE).
pp. 96 -- 108 (2013)
3. Botea, V., Mallett, D., Nascimento, M.A., Sander, J.: Pist: An efficient and prac-
tical indexing technique for historical spatio-temporal point data. GeoInformatica
12(2), 143 -- 168 (2008)
4. Brisaboa, N.R., Ladra, S., Navarro, G.: DACs: Bringing direct access to variable-
length codes. Information Processing and Management 49(1), 392 -- 404 (2013)
5. Brisaboa, N.R., Ladra, S., Navarro, G.: Compact representation of web graphs with
extended functionality. Information Systems 39(1), 152 -- 174 (2014)
6. Couclelis, H.: People manipulate objects (but cultivate fields): Beyond the raster-
vector debate in GIS. In: Proceedings of GIS: from space to territory - theories
and methods of spatio-temporal reasoning. pp. 65 -- 77 (1992)
7. Jacobson, G.: Succinct static data structures. Ph.D. thesis, Carnegie-Mellon (1988)
8. Ladra, S., Param´a, J.R., Silva-Coira, F.: Scalable and queryable compressed storage
structure for raster data. Information Systems (72), 179 -- 204 (2017)
9. Mennis, J., Viger, R., Tomlin, C.D.: Cubic map algebra functions for spatio-
temporal analysis. Cartography and Geographic Information Science 32(1), 17 -- 32
(2005)
10. Nascimento, M.A., Silva, J.R.O.: Towards historical R-trees. In: Proceedings of the
1998 ACM symposium on Applied Computing, SAC'98. pp. 235 -- 240. ACM (1998)
11. Navarro, G.: Compact Data Structures -- A practical approach. Cambridge Uni-
versity Press (2016)
12. Pinto, A., Seco, D., Guti´errez, G.: Improved queryable representations of rasters.
In: Proceedings of the 2017 Data Compression Conference (DCC). pp. 320 -- 329
(2017)
13. Tao, Y., Papadias, D.: MV3R-tree: A spatio-temporal access method for timestamp
and interval queries. In: Proceedings of the 27th International Conference on Very
Large Data Bases (VLDB). pp. 431 -- 440 (2001)
14. Vazirgiannis, M., Theodoridis, Y., Sellis, T.K.: Spatio-temporal composition and
indexing for large multimedia applications. ACM Multimedia Systems Journal
6(4), 284 -- 298 (1998)
|
1611.06222 | 2 | 1611 | 2017-07-24T13:21:13 | Approximate Near Neighbors for General Symmetric Norms | [
"cs.DS",
"cs.CG",
"cs.LG",
"math.MG"
] | We show that every symmetric normed space admits an efficient nearest neighbor search data structure with doubly-logarithmic approximation. Specifically, for every $n$, $d = n^{o(1)}$, and every $d$-dimensional symmetric norm $\|\cdot\|$, there exists a data structure for $\mathrm{poly}(\log \log n)$-approximate nearest neighbor search over $\|\cdot\|$ for $n$-point datasets achieving $n^{o(1)}$ query time and $n^{1+o(1)}$ space. The main technical ingredient of the algorithm is a low-distortion embedding of a symmetric norm into a low-dimensional iterated product of top-$k$ norms.
We also show that our techniques cannot be extended to general norms. | cs.DS | cs | Approximate Near Neighbors for General Symmetric Norms
Alexandr Andoni
Columbia University
Huy L. Nguyen
Northeastern University
Aleksandar Nikolov
University of Toronto
[email protected]
[email protected]
[email protected]
Ilya Razenshteyn
MIT CSAIL
[email protected]
Erik Waingarten
Columbia University
[email protected]
July 25, 2017
Abstract
We show that every symmetric normed space admits an efficient nearest neighbor search data
structure with doubly-logarithmic approximation. Specifically, for every n, d = no(1), and every
d-dimensional symmetric norm k · k, there exists a data structure for poly(log log n)-approximate
nearest neighbor search over k·k for n-point datasets achieving no(1) query time and n1+o(1) space.
The main technical ingredient of the algorithm is a low-distortion embedding of a symmetric
norm into a low-dimensional iterated product of top-k norms.
We also show that our techniques cannot be extended to general norms.
7
1
0
2
l
u
J
4
2
]
S
D
.
s
c
[
2
v
2
2
2
6
0
.
1
1
6
1
:
v
i
X
r
a
Introduction
1
The Approximate Near Neighbor problem (ANN) is defined as follows. The input is a dataset P
lying in a metric space (X, dX), a distance threshold r > 0, and a desired approximation c > 1. The
goal is to preprocess P so that, given a query point q ∈ X, with the promise that at least one of the
data points is within distance r, output a data point within distance cr from q. The ANN problem
is an important tool in modern data analysis, and, at the same time, is a source of many exciting
theoretical developments, see, e.g., the survey in [And16].
In many applications, the metric is defined on d-dimensional real vectors Rd. Depending on
the relation between the dimension d and the number of data points n, two main regimes have
emerged:
low- and high-dimensional. The low-dimensional regime corresponds to d = o(log n);
hence algorithms can afford to be exponential in the dimension. In the low-dimensional regime,
efficient ANN algorithms are known for any metric space [Cla99, KR02, KL04, BKL06]. In this
paper, we focus on the high-dimensional regime, when ω(log n) ≤ d ≤ no(1), which is relevant for
many applications.
The best-studied metrics are the Hamming ('1) and the Euclidean ('2) distances. There are good
reasons for this: '1 and '2 are very common in applications and admit very efficient algorithms based
on hashing, in particular, Locality-Sensitive Hashing (LSH) [IM98, AI06] and its data-dependent
versions [AINR14, AR15]. Hashing-based algorithms for ANN over '1/'2 have now been the subject
of a two-decade-long line of work, leading to a very good understanding of algorithms and their
limitations. All such algorithms for c-approximate ANN obtain space n1+ρu+o(1) and query time
nρq+o(1) for some exponents ρu and ρq < 1 dependent on c; e.g., the most recent paper [ALRW17]
gives tight time–space trade-offs for every approximation factor c > 1.1 We point the reader to
[HIM12] and [ALRW17], which summarize the state of affairs of the high-dimensional ANN over
'1/'2. A practical perspective is presented in the surveys [WSSJ14, WLKC15].
Beyond '1 and '2, the landscape of ANN is much more mysterious, despite having received
significant attention. In 1998, [Ind01] showed an efficient data structure for '∞ for c = O(log log d)
approximation. There are a few extensions of this result to other metrics, some of which proceed via
embedding a metric into '∞ (see Section 1.3). However, we are still very far from having a general
recipe for ANN data structures for general metrics with a non-trivial approximation; this is in stark
contrast with the success of the low-dimensional regime. This state of affairs motivates the following
broad question.
Problem 1. For a given approximation c > 1, which metric spaces allow efficient ANN algorithms?
An algorithm for general metrics is highly desirable both in theory and in practice. From the
theoretical perspective, we are interested in a common theory of ANN algorithms for a wide class
of distances. Such a theory would yield data structures (or impossibility results) for a variety of
important distance measures for which we still do not know efficient ANN algorithms (e.g., matrix
norms, the Earth Mover's Distance (EMD), the edit distance, etc.). Perhaps even more tantalizing
is understanding what exactly makes some distances harder than others, and how to quantify that
hardness. From the practical perspective, it is also desirable to have a generic algorithm: one that
either uses the underlying distance measure as a black box, or provides a "knob" to easily specialize
to any desired distance. In practice, one must oftentimes tune the distance to the specifics of the
application, and hence algorithms that allow such tuning without major re-implementations are
preferred.
In this paper, we focus on the following important case of Problem 1.
1The exact dependence, for '2, is that one can achieve any ρu, ρq ≥ 0 satisfying c2√
ρq + (c2 − 1)√
ρu =
√
2c2 − 1.
1
Problem 2. Solve Problem 1 for high-dimensional normed spaces.
Norms are important for two reasons. First, most metric spaces arising in applications are actually
norms (e.g., the Earth-Mover Distance [NS07]). Second, norms are geometrically nicer than general
metrics, so there is hope for a coherent theory (e.g., for the problems of sketching and streaming
norms, see the generic results of [AKR15, BBC+15]). Using embeddings into '2 [Joh48, Bal97], one
can solve ANN for any norm with approximation O
0 < ε < 1/2 is a constant; however, no better results are known in general.
(cid:17), space n1+ε, and query time nε, where
(cid:16)pd/ε
1.1 Our main result
In this paper we nearly settle Problem 2 for symmetric norms, i.e., norms that are invariant under
all permutations and changes of signs of the coordinates of a vector. We show the following general
result:
Theorem 1.1. For every n, d = no(1), and every d-dimensional symmetric norm k · k, there exists
a data structure for ANN over k · k for n-point datasets with approximation (log log n)O(1) space
n1+o(1), and query time no(1).
We note that the techniques behind Theorem 1.1 cannot be extended to general norms; see
details in Section 1.6.
1.2 Why symmetric norms?
The class of symmetric norms is, in some sense, a sweet spot. On the one hand, symmetric norms
are mathematically nice and, as we show, allow for a clean characterization that leads to an efficient
ANN data structure (see the proof overview from Section 1.4). On the other hand, symmetric norms
vastly generalize 'p distances and enable many new interesting examples, some of which arise in
applications. We first consider the following two examples of symmetric norms, which are crucial
for the subsequent discussion.
The first important example is the top-k norm: the sum of k largest absolute values of the
coordinates of a vector; k = 1 corresponds to '∞, while k = d corresponds to '1. Another rich set
of examples is that of Orlicz norms: for any non-zero convex function G : R+ → R+ such that
G(0) = 0, we define the unit ball of a norm k · kG to be:
x ∈ Rd (cid:12)(cid:12)(cid:12)
n
dX
i=1
G(cid:0)xi(cid:1) ≤ 1o
.
Clearly, for 1 ≤ p < ∞ the 'p norm is Orlicz via G(t) = tp.
In statistics and machine learning, Orlicz norms are known as M-estimators (for the case of
convex losses) [CW15]. A specific example is the Huber loss. Even though non-convex losses do not
correspond to norms, our algorithm still can handle them (see Section 3).
Other examples of symmetric norms used in applications include:
• k-support norm [AFS12] used for the sparse regression problem; its unit ball is the convex hull of
• box-Θ norm [MPS14] (again, used for sparse regression), defined for 0 < a < b ≤ c and Θ = {θ ∈
{x x is k-sparse,kxk2 ≤ 1},
[a, b]d kθk1 ≤ c} as kxk = minθ∈Θ
, and its dual;
(cid:18)Pd
x2
i
θi
i=1
(cid:19)1/2
2
(cid:12)(cid:12)(cid:12) x1 + x2 = x
• K-functional [DMS93] used to show tight tail bounds, defined for t > 0 as kxk = minnkx1k1 + t ·
• k · k1,2,s norms [KW16] used for dimension reduction, defined as kxk =(cid:0)P
, where S1 is
the set of s largest absolute values of coordinates of x, S2 is the set of next s largest coordinates,
etc.
i kxSik2
1
(cid:1)1/2
kx2k2
o
,
Finally, we show two simple ways to construct many interesting examples of symmetric norms.
Let 0 = a0 ≤ a1 ≤ a2 ≤ . . . ≤ ad be a non-decreasing sub-additive2 sequence. We can define two
norms associated with it [BS88]: a minimal norm is defined as
and a maximal norm is equal to
kxk = max
ak · (average of the largest k absolute values of the coordinates of x) ,
1≤k≤d
dX
(ak − ak−1) · (k-th largest absolute value of a coordinate of x) .
kxk =
k=1
The minimal norm is the smallest norm such that for every k one has:
(cid:13)(cid:13)(cid:13)(1, 1, . . . , 1
}
{z
k
(cid:13)(cid:13)(cid:13) = ak.
, 0, 0, . . . , 0)
Similarly, the maximal norm is the largest such norm. Minimal norms will provide hard examples of
symmetric norms that preclude some simple(r) approaches to ANN (see Section B.1). We also note
that the dual (with respect to the standard dot product) of any symmetric norm is symmetric as
well.
d(cid:1).
1.3 Prior work: ANN for norms beyond '1 and '2
For norms beyond '1 and '2, the cornerstone result in ANN is a data structure for '∞ due to
Indyk [Ind01]. For every ε > 0, the data structure achieves space n1+ε, query time no(1), and
approximation Oε(log log d). This is a doubly-exponential improvement over embeddings of '∞ into
'1/'2 which require distortion Ω(cid:0)√
It is well-known [Woj91] that any d-dimensional normed space embeds into '∞ with distortion
(1 + ε), which raises the question: can we combine this embedding with the result from [Ind01]
to solve ANN for any norm? It turns out that the answer is negative: accommodating a norm of
interest may require embedding into a very high-dimensional '∞. In the worst case, we need 2Oε(d)
dimensions, and this bound is known to be tight [Bal97], even for spaces as simple as '2. Even
though this approach would give a non-trivial approximation of O(log log 2O(d)) = O(log d), the
resulting data structure has query time which is exponential in d; thus, this approach is interesting
only for the low-dimensional regime d = o(log n).
The result of [Ind01] has been extended as follows. In [Ind02, Ind04, AIK09, And09] it was
shown how to build data structures for ANN over arbitrary 'p-products of metrics given that there
exists an ANN data structure for every factor. Recall that the 'p-product of metric spaces M1,
M2, . . . , Mk is a metric space with the ground set M1 × M2 × . . . × Mk and the following distance
function:
d(cid:0)(x1, x2, . . . , xk), (y1, y2, . . . , yk)(cid:1) =
(cid:13)(cid:13)(cid:13)(cid:0)dM1(x1, y1), dM2(x2, y2), . . . , dMk(xk, yk)(cid:1)(cid:13)(cid:13)(cid:13)p
.
2For every n, m, one has an+m ≤ an + am.
3
In a nutshell, if we can build efficient ANN data structures for every Mi with approximation
c, there exist an efficient data structure for ANN over the product space with approximation
O(c · log log n). Note that the above also implies ANN for the standard 'p, though for this case a
better approximation O(log log d) is possible via randomized embeddings into '∞ [And09].
For small values of p, one can also get c = 2O(p) [NR06, BG15] using different techniques.
1.4 Overview of the proof of Theorem 1.1
We prove Theorem 1.1 in three steps.
• First, we build a data structure for d-dimensional top-k norms. We proceed by constructing
a randomized embedding into d-dimensional '∞ with constant distortion, and then invoke the data
structure for ANN over '∞ from [Ind01].
Our embedding is a refinement of the technique of max-p-stable distributions used in [And09] to
embed 'p into '∞. Surprisingly, the technique turns out to be very general, and can handle top-k
norms as well an arbitrary Orlicz norm.
While this technique can handle even arbitrary symmetric norms (see Appendix B), there exist
symmetric norms, for which this approach leads to merely a logΩ(1) d-approximation, which is
exponentially worse than the bound we are aiming at (see Section B.1).
• To bypass the above limitation and obtain the desired (log log n)O(1)-approximation, we show
the following structural result: any d-dimensional symmetric norm allows a constant-distortion
(deterministic) embedding into a low-dimensional iterated product of top-k norms. More specifically,
the host space Y is an '∞-product of dO(1) copies of the '1-product of X1, X2, . . . , Xd, where Xk is
Rd equipped with the top-k norm.
The dimension of Y is dO(1) which is significantly better than the bound 2Ω(d) necessary to embed
symmetric norms (even '2) into '∞. It is exactly this improvement over the naïve approach that
allows us to handle any dimension d = no(1) as opposed to the trivial o(log n).
• Finally, we use known results [Ind02, And09], which allow us to construct a data structure for ANN
over a product space if we have ANN data structures for the individual factors. Each such step
incurs an additional log log n factor in the resulting approximation. Since we have built a data
structure for top-k norms, and can embed a symmetric norm into an iterated product of top-k
norms, we are done!
Embeddings into iterated product spaces have been successfully used before for constructing data
structures for ANN over Fréchet distance [Ind02], edit distance [Ind04], and Ulam distance [AIK09].
Theorem 1.1 gives yet another confirmation of the power of the technique.
1.5 Optimality of Theorem 1.1
There remains one aspects of Theorem 1.1 that can potentially be improved: the approximation
factor (log log n)O(1).
One of the bottlenecks for our algorithm is the ANN data structure for '∞ from [Ind01], which
gives O(log log d) approximation. This bound is known to be tight [ACP08, KP12] for certain
models of computation (in particular, for decision trees, which captures the result of [Ind01]). Thus,
going beyond approximation Ω(log log d) in Theorem 1.1 might be hard; however, it remains entirely
possible to improve the approximation from (log log n)O(1) to O(log log d), which we leave as an
open question.
4
1.6 Lower bounds for general norms
The second step of the proof of Theorem 1.1 (see Section 1.4) shows how to embed any d-dimensional
symmetric norm into a universal normed space of dimension dO(1) with a constant distortion. In
contrast, we show that for general norms a similar universal construction is impossible. More
formally, for a fixed 0 < ε < 1/3, suppose U is a normed space such that for every d-dimensional
normed space X there exists a randomized linear embedding of X into U with distortion O(d1/2−ε).
dΩε(1)(cid:17). By John's theorem [Joh48], d-dimensional '2 is
Then, U must have dimension at least exp(cid:16)
d, so our lower bound is tight up to sub-polynomial factors. See
a universal space for distortion
Section 6 for details.
√
To take this a step further, it would be highly desirable to prove stronger hardness results for
ANN over general norms. One approach would be to show that such a norm X has high robust
expansion, which is a property used to deduce ANN lower bounds [PTW10, ALRW17]. There exist
metrics M that have high robust expansion, such as the shortest path metric of a spectral expander
(see Appendix C). To obtain a hard norm, it suffices to embed such an N-point metric M into
a logO(1) N-dimensional norm with a constant distortion. The result of [Mat96] shows that there
exist N-point metrics M which cannot be embedded into any norm of dimension N o(1). However,
these metrics are not expanders, and for expanders such a dimension reduction procedure might be
possible. 3
1.7 Other related work: dealing with general norms
The recent result of [BBC+15] completely characterizes the streaming complexity of any symmetric
norm. Even though many symmetric norms (including '∞) are hard in the streaming model, the
state of affairs with ANN is arguably much nicer. In particular, our results imply that all symmetric
norms have highly efficient ANN data structures. We also point out that streaming algorithms for
the special case of Orlicz norms have been studied earlier [BO10].
Another related work is [AKR15], which shows that for norms, the existence of good sketches
is equivalent to uniform embeddability into '2. Sketches are known to imply efficient ANN data
structures, but since many symmetric norms do not embed into '2 uniformly, we conclude that
ANN is provably easier than sketching for a large class of norms.
Finally, we also mention the work of [AV15], who study ANN under the class of high-dimensional
distances which are Bregman divergences. These results are somewhat disjoint since the Bregman
divergences are not norms.
2 Preliminaries
2.1 Norms and products
We denote non-negative real numbers by R+. For any subset A ⊆ R, we let χA : R → {0, 1} be the
indicator function of A. Let X be a normed space over Rd. We denote BX the unit ball of X, and
k · kX the norm of X. We denote X∗ the dual norm of X with respect to the standard dot product
h·,·i, i.e kxkX∗ = sup{hx, yi : y ∈ BX}. For a vector x ∈ Rd we define x = (x1,x2, . . . ,xd) to
be the vector of the absolute values of the coordinates of x. For a positive integer d and 1 ≤ p ≤ ∞,
we denote 'd
p the space Rd equipped with the standard 'p norm, which we denote by k · kp.
3In a recent work, Naor [Nao17] showed that this approach is impossible. He shows that embedding an N-point
spectral expander with constant distortion into any normed space requires poly(N) dimensions.
5
Definition 2.1. For any vector x ∈ Rd, we let x∗ = Px be the vector obtained by applying the
permutation matrix P to x so coordinates of x∗ are sorted in non-increasing absolute value.
Definition 2.2 (Symmetric norm). A norm k · kX : Rd → R is symmetric if for every x ∈ Rd,
kxkX =
= kx∗kX.
(cid:13)(cid:13)(cid:13)x(cid:13)(cid:13)(cid:13)X
See the introduction for examples of symmetric norms. We note once again that the dual norm
of a symmetric norm is also symmetric.
A natural way to combine norms is via product spaces, which we will heavily exploit in this
paper.
spaces. We define the 'p-product space, denoted L
Definition 2.3 (Product space). Let 1 ≤ p ≤ ∞. Let (X1, dX1), (X2, dX2), . . . , (Xk, dXk) be metric
'p Xi, to be a metric space whose ground set is
X1×X2×. . .×Xk, and the distance function is defined as follows: the distance between (x1, x2, . . . , xk)
and (x0
k) is defined as the 'p norm of the vector(cid:0)dX1(x1, x0
2), . . . , dXk(xk, x0
1), dX2(x1, x0
2, . . . , x0
k)(cid:1).
1, x0
Next we define the top-k norm:
Definition 2.4. For any k ∈ [d], the top-k norm, k · kT (k) : Rd → R, is the sum of the absolute
values of the top k coordinates. In other words,
kxkT (k) =
x∗
i,
kX
i=1
where x∗ is the vector obtained in Definition 2.1.
Definition 2.5. Given vectors x, y ∈ Rd, we say x weakly majorizes y if for all k ∈ [d],
kX
i=1
i ≥ kX
x∗
i=1
y∗
i .
Lemma 2.6 (Theorem B.2 in [MOA11]). If x, y ∈ Rd where x weakly majorizes y, then for any
symmetric norm k · kX,
kxkX ≥ kykX .
Definition 2.7. For i ∈ [d], let ξ(i) ∈ Rd be the vector
{z }
ξ(i) = (1, . . . , 1
, 0, . . . , 0
{z }
)
i
d−i
consisting of exactly i 1's, and d − i 0's.
2.2 ANN for '∞ and '∞-products
We will crucially use the following two powerful results of Indyk. The first result is for the standard
d-dimensional '∞ space.
Theorem 2.8 ([Ind01, Theorem 1]). For any ε ∈ (0, 1/2), there exists a data structure for ANN
for n-points datasets in the 'd∞ space with approximation O
time O(d · log n).
(cid:17), space O(d · n1+ε), and query
(cid:16) log log d
ε
6
The second is a generalization of the above theorem, which applies to an '∞-product of k metrics
X1, . . . Xk, and achieves approximation O(log log n). It only needs black-box ANN schemes for each
metric Xi.
Theorem 2.9 ([Ind02, Theorem 1]). Let X1, X2, . . . , Xk be metric space, and let c > 1 be a real
number. Suppose that for every 1 ≤ i ≤ k and every n there exists a data structure for ANN for
n-point datasets from Xi with approximation c, space S(n) ≥ n, query time Q(n), and probability of
success 0.99. Then, for every ε > 0, there exists ANN under Lk
'∞ M with:
• O(ε−1 log log n) approximation,
• O(Q(n) log n + dk log n) query time, where d is the time to compute distances in each Xi, and
• S(n) · O(knε) space/preprocessing.
Strictly speaking, we need to impose a technical condition on the ANN for each Xi - that it
reports the point with the smallest priority - which is satisfied in all our scenarios; see [Ind02,
Section 2] for details. Also, the original statement of [Ind02] gave a somewhat worse space bound.
The better space results simply from a better analysis of the algorithm, as was observed in [AIK09];
we include a proof in Appendix A.
3 ANN for Orlicz and top-k norms
Before showing a data structure for general symmetric norms, we give an algorithm for general
Orlicz norms. We then show how to apply these ideas to top-k norms. This restricted setting has a
simple analysis and illustrates one of the main techniques used in the rest of the paper. A similar
approach was used in prior work to construct randomized embeddings of 'p norms into '∞, and
solve the ANN search problem; here we show that these techniques are in fact applicable in much
greater generality.
Lemma 3.1. Let k · kG be an Orlicz norm. For every D, α > 1 and every µ ∈ (0, 1/2) there exists
a randomized linear map f : Rd → Rd such that for every x ∈ Rd:
• if kxkG ≤ 1, then Prf
• if kxkG > αD, then Prf
h(cid:13)(cid:13)f(x)(cid:13)(cid:13)∞ ≤ 1i ≥ µ;
h(cid:13)(cid:13)f(x)(cid:13)(cid:13)∞ > D
i ≥ 1 − µα.
Proof. Let the distribution D over R+ have the following CDF F : R+ → [0, 1]:
F(t) = Pr
u∼D[u ≤ t] = 1 − µG(t).
Consider the following randomized linear map f : Rd → Rd:
(cid:18) x1
x2
u2
(cid:19)
xd
ud
where u1, . . . , ud ∼ D are i.i.d. samples from D. Suppose that kxkG ≤ 1. Then, Pd
u1
, . . . ,
,
(x1, x2, . . . , xd) f7→
i=1 G(xi) ≤ 1.
This, in turn, implies:
hkf(x)k∞ ≤ 1i =
Pr
f
dY
i=1
Pr
ui∼D
(cid:20)(cid:12)(cid:12)(cid:12)(cid:12) xi
ui
(cid:21)
(cid:12)(cid:12)(cid:12)(cid:12) ≤ 1
dY
i=1
=
Pd
i=1 G(xi) ≥ µ.
µG(xi) = µ
7
Now suppose that kxkG > αD. This, together with the convexity of G(·), implies:
(cid:18)xi
(cid:19)
αD
G
≥ α.
(cid:18)xi
(cid:19)
dX
i=1
G
≥ (1 − α)G(0) + α · dX
(cid:20)(cid:12)(cid:12)(cid:12)(cid:12) xi
(cid:12)(cid:12)(cid:12)(cid:12) ≤ D
dY
(cid:21)
i=1
=
ui
i=1
Pr
ui∼D
D
dY
i=1
Thus, we have:
hkf(x)k∞ ≤ D
i =
Pr
f
Pd
i=1 G(xi/D) ≤ µα.
µG(xi/D) = µ
Theorem 3.2. For every d-dimensional Orlicz norm k · kG and every ε ∈ (0, 1/2), there exists a
data structure for ANN over k·kG, which achieves approximation O
and query time O (dnε).
Proof. Let P ⊂ Rd be a dataset of n points. Consider the data structure which does the following:
1. For all 1 ≤ i ≤ nε, we independently apply the randomized linear map f from Lemma 3.1
(cid:17) using space O(cid:0)dn1+ε(cid:1)
(cid:16) log log d
ε2
with parameters µ = n−ε, D = O
(cid:17), and α = 2
(cid:16) log log d
Pi = {fi(x) x ∈ P}
ε
ε. We define
to be the image of the dataset under fi, where fi is the i-th independent copy of f.
2. For each 1 ≤ i ≤ nε, we use Theorem 2.8 to build a data structure for ANN over '∞ with
approximation D for dataset Pi. We refer to the i-th data structure as Ti.
Each Ti occupies space O(dn1+ε) and achieves approximation D with query time O(d log n). To
answer a query q ∈ Rd, we query Ti with fi(q) for each i ∈ [nε]. Let xi be the point returned by
Ti, and let pi ∈ P be the pre-image of xi under fi, so that fi(pi) = xi. If for some Ti, the point
returned satisfies kpi − qkG ≤ αD, then we return pi.
• If there exists some p ∈ P with kp − qkG ≤ 1, then by Lemma 3.1, with probability 1 −
(1 − n−ε)nε ≥ 3
5, some fi has kfi(p− q)k∞ ≤ 1. Since fi is linear, kfi(p)− fi(q)k∞ ≤ 1 as well.
• Let i ∈ [nε] be an index where some p ∈ P with kp− qkG ≤ 1 has kfi(p)− fi(q)k∞ ≤ 1. Every
other p0 ∈ P with kp0 − qkG ≥ αD satisfies
Prhkfi(p0) − fi(q)k∞ ≤ D
i ≤ 1
n2 .
A union bound over at most n points with distance greater than αD to q shows that except
n, Ti returns some pi ∈ P with kpi − qkG ≤ αD. Thus, the total
with probability at most 1
The total query time is O (dnε · log n) and the total space used is O(cid:0)dn1+2ε(cid:1). This data structure
probability of success of the data structure is at least 3
(cid:17). Decreasing ε by a constant factor, we get the desired
(cid:16) log log d
5 − 1
n.
achieves approximation αD = O
guarantees.
ε2
8
Remark. The construction of the randomized embedding in Lemma 3.1 and the data structure
from Theorem 3.2 work in a somewhat more general setting, rather than just for Orlicz norms. For
a fixed norm k · k, we can build a randomized map f : Rd → Rd with the guarantees of Lemma 3.1
if there exists a non-decreasing G: R+ → R+ where G(0) = 0, G(t) → ∞ as t → ∞, and for every
x ∈ Rd:
• if kxk ≤ 1, thenPd
• if kxk ≥ αD, thenPd
i=1 G(xi) ≤ 1, and
(cid:16)xi
D
(cid:17) ≥ α.
i=1 G
The data structure itself just requires the existence of a randomized linear map satisfying the
conditions of Lemma 3.1.
We now describe how to obtain a data structure for ANN for any top-k norm.
Lemma 3.3. Fix any k ∈ [d]. For every D, α > 1 and every µ ∈ (0, 1/2), there exists a randomized
linear map f : Rd → Rd such that for every x ∈ Rd:
• if kxkT (k) ≤ 1, then Prf
• if kxkT (k) > αD, then Prf
hkf(x)k∞ ≤ 1i ≥ µ;
hkf(x)k∞ > D
i ≥ 1 − µα−1.
Proof. We define G: R+ → R+ where for every x ∈ Rd,
k ,∞)(t)
If kxkT (k) ≤ 1, there are at most k coordinates where xi ≥ 1
i=1 G
G(t) = t · χ[ 1
i ≥ αD. Therefore,Pd
1. If kxkT (k) ≥ αD, thenPk
i=1 x∗
proof now follows in the same way as Lemma 3.1.
k. Therefore,Pd
(cid:16)x∗
(cid:17) ≥Pk
i
D
i=1 G
(cid:16)x∗
i=1 G(xi) ≤ kxkT (k) ≤
(cid:17) ≥ α − 1. The
i
D
Lemma 3.3 gives us a data structure for any top-k norm with approximation O(log log d) applying
Theorem 3.2.
One could imagine using a similar argument to design an algorithm for general symmetric norms.
This idea indeed works and yields an algorithm with approximation eO(log d) for a general symmetric
norm (see Appendix B for a detailed analysis of this approach). However, we show this strategy
√
cannot achieve an approximation better than Ω(
log d) (see the end of the same Appendix B).
4 Embedding symmetric norms into product spaces
In this section, we construct an embedding of general symmetric norms into product spaces of top-k
norms. To state the main result of this section, we need the following definition.
Definition 4.1. For any c1, . . . , cd ≥ 0, let Ld
'1 T (c) ⊂ Rd2 be the space given by the seminorm
T,1 · Rd2 → R where for x = (x1, . . . , xd) ∈ Rd2 and x1, . . . , xd ∈ Rd:
dX
k · k(c)
kxk(c)
T,1 =
ckkxkkT (k).
k=1
9
Ld
Theorem 4.2 (Embedding into a product space). For any constant δ ∈ (0, 1/2), any symmetric
'1 T (c) where
norm k · kX : Rd → R can be embedded linearly with distortion 1 + δ into Lt
+ such that for every x ∈ Rd,
t = dO(log(1/δ)δ−1). In particular, there exists c ∈ Rt×d
'∞
dX
!
'∞
k=1
ci,kkxkT (k)
≤ (1 + δ)kxkX .
The vectors inLt
(1 − δ)kxkX ≤ max
i∈[t]
Ld
'1 T (c) ⊂ Rtd2 can be broken up into td blocks of d coordinates each. The
embedding referenced above will simply map x ∈ Rd into Rtd2 by making each of the td many blocks
equal to a copy of x. The non-trivial part of the above theorem is setting the constants ci,k for
i ∈ [t] and k ∈ [d] so (1) holds. Before going on to give the proof of Theorem 4.2, we establish
some definitions and propositions which will be used in the proof. For the subsequent sections, let
β ∈ (1, 2) be considered a constant close to 1.
Definition 4.3 (Levels and Level Vectors). For any fixed vector x ∈ Rd and any k ∈ Z, we
define level k with respect to x as Bk(x) = {i ∈ [d] β−k−1 < xi ≤ β−k}. Additionally, we let
bk(x) = Bk(x) be the size of level k with respect to x. The level vector of x, V (x) ∈ Rd is given by
(1)
V (x) = (βk, . . . , βk
b−k(x) times
}
}
, βk−1, . . . , βk−1
b−k+1(x) times
{z
}
, . . . , β−k, . . . , β−k
bk(x) times
{z
{z
, 0, . . . 0)
where k is some integer such that all non-zero coordinates lie in some level between −k and k. We
say the i-th level vector Vi(x) ∈ Rd is given by
}
Vi(x) = (β−i, . . . , β−i
bi(x) times
, 0, . . . , 0).
{z
The notation used for level vectors appears in [BBC+15]; however, we refer to level k as the
coordinates of x lying in (β−k−1, β−k]; whereas [BBC+15] refers to level k as the coordinates of x
lying in [βk−1, βk).
Definition 4.4. Fix some τ > 0. For any vector x ∈ Rd, let C(x) ∈ Rd be the vector where each
i ∈ [d] sets
(
C(x)i =
xi ≥ τ
xi < τ
.
xi
0
Proposition 4.5 (Proposition 3.4 in [BBC+15]). Let k · kX be any symmetric norm and x ∈ Rd be
any vector. Then
1
β
kV (x)kX ≤ kxkX ≤ kV (x)kX .
Proposition 4.6. Let k · kX be any symmetric norm. For any vector x ∈ Rd,
kxkX − τ d ≤ kC(x)kX ≤ kxkX .
Proof. Note that x weakly majorizes C(x), so kC(x)kX ≤ kxkX. For the other direction, let
v = x − C(x). Then v has all coordinates with absolute value at most τ, so τ dξ(1) weakly majorizes
v. Therefore,
kxkX ≤ kC(x)kX + kvkX ≤ kC(x)kX + τ d.
10
Intuitively, the above two propositions say that up to multiplicative loss β and additive loss τ d
in the norm of the vector, we may assume that all coordinates are exactly βj for j ≥ logβ(τ). Thus,
if x ∈ Rd, then
kxkX − τ d ≤ kV (C(x))kX ≤ βkxkX .
d2 , so when kxkX ≤ 1 there are at most 2 logβ d non-empty levels in
If additionally, we let τ = β
V (C(x)).
Definition 4.7 (Rounded counts vector). Fix any level vector x ∈ Rd. The rounded counts vector
of x, W(x) ∈ Rd is given by y where the y ∈ Rd is constructed using the following procedure:
1: Initialize y = (0, . . . , 0) ∈ Rd and c = d.
2: for k = −∞, . . . , 2 logβ(d) − 1 do
3:
4:
5:
6:
7:
8:
9: end for
10: Return y
Let j ∈ Z+ be the integer where βj−1 < bk(x) ≤ βj.
if c ≥ bβjc then
Set the first bβjc zero-coordinates of y with β−k. Update c ← c − bβ−kc.
if bk(x) ≥ 0 then
end if
end if
Intuitively, W(x) represents the level vector of x where we ignore coordinates smaller than β
d2 ,
and additionally, we round the counts of coordinates to powers of β.
Lemma 4.8. For every vector x ∈ Rd and any symmetric norm k · kX,
kxkX − τ d ≤ kW(V (C(x)))kX ≤ β2kxkX .
Proof. The bound kxkX − τ d ≤ kW(V (C(x)))kX follows by combining Proposition 4.5 and Propo-
sition 4.6, as well as the monotonicity of norms. The bound kW(V (C(x)))kX ≤ β2kxkX follows
from Proposition 4.5, Proposition 4.6, as well as Lemma 3.5 from [BBC+15].
In order to simplify notation, we let R: Rd → Rd given by R(x) = W(V (C(x))).
Definition 4.9. Let the set L ⊂ Rd+ be given by
L = {y ∈ Rd+ y1 ≥ . . . yd ≥ 0}.
Additionally, for an arbitrary symmetric norm k · kX with dual norm k · kX∗, we let the set R ⊂ L
be given by
R = {R(y) ∈ Rd+ y ∈ L ∩ BX∗}.
Definition 4.10. Fix a vector y ∈ L \ {0} (y has non-negative, non-increasing coordinates). Let
the maximal seminorm with respect to y, k · ky : Rd → R be the seminorm where for every x ∈ Rd,
kxky = hx∗, yi.
We first show there exists some setting of c ∈ Rd such that we may compute kxky as ⊕d
Lemma 4.11. For every vector y ∈ L \ {0}, there exists c1, . . . , cd ≥ 0 where for all x ∈ Rd,
'1T (c).
kxky = kxk(c)
T,1.
11
Proof. For k ∈ [d], we let ck = yk − yk+1, where yd+1 = 0.
dX
dX
dX
dX
!
hx∗, yi =
x∗
iyi =
x∗
i
i=1
i=1
k=i
ck
=
ck
k=1
!
i
x∗
kX
i=1
dX
k=1
=
ckkxkT (k)
Given Lemma 4.11, it suffices to show that for an arbitrary symmetric norm k · kX, we may
compute kxkX (with some distortion) as a maximum over many maximal seminorms.
In the
following lemma, we show that taking the maximum over maximal norms from R suffices, but gives
sub-optimal parameters. We then improve the parameters to prove Theorem 4.2.
Lemma 4.12. Let k · kX be an arbitrary symmetric norm and let k · kX∗ be its dual norm. Then
for any kxkX ≤ 1,
kxkX − τ d ≤ max
y∈R kxky ≤ β2kxkX .
Proof. Without loss of generality, we rescale the norm so that ke1kX∗ = 1, where e1 is the first
standard basis vector. Consider any x ∈ Rd with kxkX ≤ 1. Then since k · kX is symmetric, we may
assume without loss of generality that all coordinates of x are non-negative and in non-increasing
order. Thus for each y ∈ L ∩ {0}, we have kxky = hx, yi.
The lower bound simply follows from the fact that R(z), other than coordinates less than τ, is
monotonically above z, and all coordinates in x are non-negative. More specifically,
y∈R hx, yi + τ d,
hx, R(z)i + τ d = max
kxkX = sup
hx, zi ≤ sup
z∈L∩BX∗
z∈L∩BX∗
where τ d comes from the fact that because kxkX ≤ 1, every coordinate of x is at most 1. On the
other hand, we have
y∈R hx, yi = β2 max
max
y∈R hx,
y
β2i ≤ β2 sup
z∈BX∗
hx, zi = β2kxkX ,
where we used the fact that k y
β2kX∗ ≤ 1 by Lemma 4.8.
'∞ ⊕d
1−τ d ≤ β3 (where we used the fact τ = β
Given Lemma 4.12, it follows that we may linearly embed X into ⊕t
'1 T (c) where t = R,
d and that 1 + β/d ≤ β for a large enough
with distortion β2
d). The embedding follows by copying the vector x into the t spaces ⊕d
'1T (c) corresponding to
each vector y ∈ R given in Lemma 4.11. The one caveat is that this embedding requires t copies
of ⊕d
2 logβ d many levels, and each contains has number of coordinates being some value in {βi}logβ d
i=0 .
'1T (c), and t is as large as(cid:16)logβ d + 1(cid:17)2 logβ d = dO(log log d). This is because there are at most
Thus, our algorithm becomes inefficient once d ≥ 2ω(cid:0) log n
In order to avoid this problem, we will make the embedding more efficient by showing that we
do not need all of R, but rather a fine net of R. In addition, our net will be of polynomial size in
the dimension, which gives an efficient algorithm for all ω(log n) ≤ d ≤ no(1). We first show that it
suffices to consider fine nets of R, and then build a fine net of R of size poly(d).
Lemma 4.13. Fix an γ ∈ (0, 1/2). Let k · kX be an arbitrary symmetric norm and k · kX∗ be its
dual norm. If N is a γ-net of R with respect to distance given by k · kX∗, then
log log n
(cid:1)
.
(1 − γ − τ d)kxkX ≤ max
y∈N
kxky ≤ (β2 + γ)kxkX
12
Proof. Since the embedding we build is linear, it suffices to show that every vector x ∈ Rd with
kxkX = 1 has
1 − γ − τ d ≤ max
y∈N
kxky ≤ (β2 + γ).
Consider a fixed vector x ∈ Rd with kxkX = 1. Additionally, we may assume the coordinates of x
are non-negative and in non-increasing order. We simply follow the computation:
hx∗, yi + γkxkX + τ d,
kxkX = kx∗kX ≤ max
where we used Lemma 4.12 and the fact that kvkX∗ ≤ γ in a γ-net of R with respect to the distance
given by k · kX∗. On the other hand,
(hx∗, yi + hx∗, vi) + τ d ≤ max
y∈N
y∈R hx∗, yi + τ d = max
y∈N
max
y∈N
kxky = max
y∈R (hx∗, yi + hx∗, vi) ≤ max
y∈R kxky + γkxkX ≤ (β2 + γ)kxkX ,
where again, we used Lemma 4.12 and the fact that kvkX∗ ≤ γ.
Finally, we conclude the theorem by providing a γ-net for R of size dO(log(1/γ)γ−1).
Lemma 4.14. Fix any symmetric space X with dual X∗. There exists an 8(β − 1)-net of size
dO(log(1/(β−1))/ log β) for R with respect to distances given by k · kX∗.
We defer the proof of Lemma 4.14 to the next section. The proof of Theorem 4.2 follows by
combining Lemma 4.11, Lemma 4.13, and Lemma 4.14. In particular, given a β−1
8 -net of R, we get
an embedding with distortion at most (β2 + 8(β − 1))(1 + (8(β − 1) + τ d)2) from Lemma 4.13. We
let τ = β
proceed to proving Lemma 4.14, which gives the desired upper bound on the size of the net.
d2 and β =p1 + δ/100 to get the desired linear embedding with distortion 1 + δ. We now
4.1 Proof of Lemma 4.14: bounding the net size
We now give an upper bound on the size of a fine net of R. We proceed by constructing a further
simplification of R(x). Intuitively we show that one can ignore the higher levels if there are fewer
coordinates in the higher levels than some lower level.
Lemma 4.15. Let k · kX be a symmetric norm. Consider any nonnegative vector x ∈ Rd+ as well
as two indices u, v ∈ [d]. Let y ∈ Rd+ be the vector with:
0
xk
zk
yk =
xu + xv
k ∈ [d] \ {u, v}
k = u
k = v
.
z0
k =
max{zu, zv}
min{zu, zv}
k ∈ [d] \ {u, v}
k = u
k = v
Then kykX ≥ kxkX.
Proof. Consider the vector z ∈ BX∗ where hx, zi = kxkX. Now, we let z0 ∈ Rd be given by
Note that z0 is a permutation of z, so z0 ∈ BX∗. Now,
hy, z0i = (xu + xv) max{zu, zv} + X
xkzk ≥ X
k∈[d]\{u,v}
k∈[d]
xkzk = hx, zi = kxkX .
13
Definition 4.16. Consider a vector x ∈ R. We define the simplified rounded vector S(x) as the
vector returned by the following procedure.
1: Initialize z = x
2: for k = 0, 1, . . . , 2 logβ(d) − 1 do
3:
4:
5:
6: end for
7: Sort the coordinates of z in non-increasing order and return z.
if bk(z) ≤ maxj<k+3 logβ(β−1) bj(z) then
end if
Set all coordinates of z of value β−k to 0 i.e. set bk(z) = 0.
Next we show that the simplified rounded vector is close to the rounded counts vector.
Lemma 4.17. Let k · kX be a symmetric norm and let x ∈ R. Then kS(x) − xkX ≤ 2(β − 1)kxkX.
Proof. Consider some k ∈ [2 logβ d − 1] and let Ck(x) ⊂ [d] be set of coordinates where x is at level
k and does not equal S(x) = z, i.e.,
Ck(x) = {i ∈ [d] xi = β−k and xi 6= zi}.
Additionally, for each k ∈ [2 logβ d − 1], let Tk ⊂ [d] be the coordinates at level k in x which trigger
line 3 of S(x), and thus become 0's in z (we do not need to consider the case k = 0 since line 3
never triggers, in fact, we do not need to consider k ∈ [−3 logβ(b − 1)] either). In other words,
max
bj(z)}.
j<k+3 logβ(β−1)
Tk(x) = {i ∈ [d] xi = β−k and at iteration k of S(x), bk(z) ≤
Note that T1(x), . . . , T2 logβ d−1(x) are all disjoint, and Ck(x) ≤P
sort, causing xi 6= zi. Thus, we may consider an injection sk : Ck(x) →S
j∈[k] Tj(x), since whenever we
zero out coordinates in levels less than or equal to k, S(x) will shift the coordinates when we
j∈[k] Tj(x), which charges
coordinates in Ck(x) to coordinates which were zeroed out in line 3 of S(x).
Additionally, for each j ∈ [2 logβ d − 1] where Tj(x) 6= ∅, we let qj be the integer between 0 and
j + 3 logβ(β − 1) which triggered line 3 of S(x) at k = j. More specifically, 0 ≤ qj ≤ j + 3 logβ(β − 1)
is the integer for which bj(x) ≤ bqj(x).
Finally, for each j ∈ [2 logβ d − 1] where Tj(x) 6= ∅, we let gj : Tj(x) → Bqj(x) (recall that
Bqj(x) ⊂ [d] are the indices of x at level qj) be an arbitrary injection. Such an injection exists
k∈[2 logβ d−1] Ck(x) → [d] where
because bj(x) ≤ bqj(x). We may consider the mapping F : S
we "aggregate" coordinates of S
F(i) = gj(sk(i))
where k and j are such that i ∈ Ck(x) and sk(i) ∈ Tj(x).
See Figure 4.1 for an example of a coordinate being mapped by F. Let y be the vector where
k∈[2 logβ(d)−1] Ck(x) of x according to the map F according to
Lemma 4.15. In particular, we define y ∈ Rd where for i ∈ [d], we let
yi = X
Note that for each i ∈ [d], 0 ≤ (x − z)i ≤ xi, and S
i0∈F −1(i)
xi0.
k∈[2 logβ(d)−1] Ck(x) ⊂ [d] are the non-zero
coordinates of x − z. Thus, from Lemma 4.15, we conclude that kx − zkX ≤ kykX. We now turn to
upper-bounding kykX.
14
x =
i
Bqj(x)
gj
−3 logβ(β − 1)
sk
'
Tj(x)
i0
Ck(x)
Figure 1: Example mapping of particular levels of x with F showing aggregation of coordinates.
The coordinate i0 ∈ Ck(x) belongs to level k and will be non-zero in x − S(x).
In particular,
coordinate i0 ∈ Ck(x) is mapped via sk to coordinate ' ∈ Tj(x), which is β−j in x and 0 in S(x).
Then coordinate ' is mapped to i ∈ Bqj(x), where qj is the level below j with bqj ≥ bj. Thus, the
composed map F sends i0 to i.
Fix some i ∈ [d] where xi = β−j. Then
yi = X
k>j−3 logβ(β−1)
k (i)) as 0 when g−1
k0 (g−1
k (i)) = β−k0. Thus,
≤ X
β−k0
k0 (g−1
where we interpret xs−1
k0 (g−1
xs−1
k (i)) 6= 0, xs−1
yi ≤ X
X
k>j−3 logβ(β−1)
k0≥k
X
k0≥k
,
xs−1
k0 (g−1
k (i))
k (i) = ∅, or s−1
k0 (g−1
k (i)) = ∅. Note that whenever
β − 1 ≤ β1−j+3 logβ(β−1)
β1−k
(β − 1)2
≤ β(β − 1) · β−j.
k>j−3 logβ(β−1)
Recall that xi = β−j, so β(β − 1)x weakly majorizes y, and thus
kx − zkX ≤ kykX ≤ β(β − 1) · kxkX .
Hence, when β ≤ 2, we have kx − zkX ≤ 2(β − 1)kxkX.
Proof of Lemma 4.14. We now prove the theorem by showing that the set
N = {S(x) ∈ Rd x ∈ R}
is a γ-net of R, and we give an upper bound on the size. By Lemma 4.17, kx − S(x)kX ≤
2(β − 1)kxkX ≤ 8(β − 1). So it suffices to give an upper bound on the size of N.
We bound from above the number of net points by an encoding argument. Let z = S(x) and let
tk =
bk(z)
maxj<k+3 logβ(β−1) bj(z) .
Let k∗ ∈ {0, . . . , 2 logβ d−1} be the smallest level k with non-zero bk(z). For all j > k∗−3 logβ(β−1),
we either have tj(z) = 0 or tj(z) ≥ 1. Additionally, z has d coordinates, so
2 logβ d−1Y
j=k∗−3 logβ(β−1)
max(tj, 1) ≤ d−3 logβ(β−1),
since terms of the product "cancel" except for at most −3 logβ(β − 1), which are each at most d.
15
(( We will encode z ∈ N in three steps. In the first step, we use 2 logβ d − 1 bits in order to
encode whether bk(z) = 0 or bk(z) > 0. In the second step, we then encode the values bk∗+j(z) for
j ∈ {0, . . . , 3 logβ(1/(β − 1))}. Finally, we go through j > k∗ + 3 logβ
a prefix-free code, where writing ti uses at most O (log max(ti, 1)) many bits. Thus, in total, the
number of bits we use is
(cid:16) 1
β−1
O
(cid:19)
+
logβ d + log d logβ
(cid:18) 1
log d · log(cid:16) 1
(cid:17)
β − 1
(cid:17)
log d · log(cid:16) 1
log β
= O
= O
β−1
β−1
log β
.
+ log
2 logβ d−1X
2 logβ d−1Y
j=k∗−3 logβ(β−1)
j=k∗−3 logβ(β−1)
log max(tj, 1)
max(tj, 1)
(cid:17), and encode ti using
Thus, we obtain N is a 8(β − 1)-net, and the size of N is dO(log(1/(β−1))/ log β).
5 Proof of the main theorem: ANN for symmetric norms
We now prove our main result, Theorem 1.1. The algorithm here achieves approximation
log2 log n · log log d
!
O
We proceed by giving an algorithm for Lt
Lemma 5.1. Fix some c1, . . . , cd ≥ 0. Let L
[And09], and Theorem 2.9.
where for every x = (x1, . . . , xd) ∈ Rd2,
'∞
.
Ld
ε5
'1 T (c) using Theorem 2.8, Theorem 5.1.2 from
'∞ T (c) be the space with k · k(c)
T,∞ : Rd2 → R seminorm
kxk(c)
T,∞ = max
k∈[d]
ckkxkkT (k).
ε3
(cid:16) log log n·log log d
(cid:17) using space O(cid:0)d2 · n1+ε(cid:1) and query time O(cid:0)d2 · nε(cid:1).
For every ε ∈ (0, 1/2), there exists a data structure for ANN over k · k(c)
mation O
Proof. Given the randomized embedding from Lemma 3.3, we can build a data structure for ckk·kT (k)
achieving approximation O( log log d
) using space O(d2n1+ε/2) and query time O(d2nε/2). This data
structure works in the same way as in the proof of Theorem 3.2. We handle the constant ck by
rescaling the norm, and since the embeddings are linear, it does not affect the correctness of the
data structure. Then we apply Theorem 2.8.
T,∞ which achieves approxi-
ε2
Lemma 5.2. Fix some c1, . . . , cd ≥ 0. Let L
where x = (x1, . . . , xm) ∈ Rd2,
'1 T (c) be the space with k · k(c)
dX
ckkxkkT (k).
kxk(c)
T,1 =
k=1
T,1 : Rd2 → R seminorm
(cid:16) log log n·log log d
For every ε ∈ (0, 1/2), there exists a data structure for ANN over k·k(c)
O
(cid:17) using space O(d2 · n1+ε) and query time O(d2 · nε).
ε4
T,1 which achieves approximation
16
Proof. The proof follows from Theorem 5.1.2 in [And09] and Lemma 5.1.
Finally, we are combine the above results to get an improved algorithm for general symmetric
norms.
Theorem 5.3. For every d-dimensional symmetric norm k · kX and every ε ∈ (0, 1/2), there exists
a data structure for ANN over k · kX which achieves approximation O( log2 log n log log d
) using space
dO(1) · O(n1+ε) and query time dO(1) · O(nε).
Proof. Given Theorem 4.2, we embed k · kX intoL
'1 T (c) with approximation (1 ± 1
result from Lemma 5.2 allows we to apply Theorem 2.9 to obtain the desired data structure.
10). The
L
'∞
ε5
Theorem 5.3 implies our main result Theorem 1.1 stated in the introduction.
6 Lower bounds
In this section, we show that our techniques do not extend to general norms. In particular, we show
there does not exist a universal norm U for which any norm embeds (possibly randomized) with
constant distortion, unless the blow-up in dimension is exponential. Hence the result from below
applies to cases of U = '∞ as well as an (low-dimensional) product spaces.
Theorem 6.1. For any ε > 0, let U be a d0-dimensional normed space such that for any d-
dimensional normed space X, there exists a distribution D supported on linear embeddings f : Rd →
Rd0 where for every x ∈ Rd,
kxkX ≤ kf(x)kU ≤ DkxkU
3 over the draw of f ∼ D, for D = O(d1/2−ε). Then d0 = exp(cid:0)Ω(d2ε)(cid:1).
holds with probability at least 2
We will prove the above theorem by showing that if there exists a universal normed space U
satisfying the conditions of Theorem 6.1 above, then two parties, call them Alice and Bob, can
use the embeddings to solve the communication problem Index with only a few bits. Let U be
a proposed d0-dimensional normed space satisfying the conditions of Theorem 6.1. By the John's
theorem [Bal97], we may apply a linear transform so that:
d0B'2
Lemma 6.2. For any ε > 0, there exists a set of exp(cid:0)Ω(d2ε)(cid:1) many points on the unit sphere Sd−1
B'2 ⊂ BU ⊂
such that pairwise inner-products are at most
whose coordinates are ± 1√
Proof. Consider picking two random points x, y ∈ Sd−1 where each entry is ± 1√
Bernstein's inequality,
d1/2−ε . In fact, these points may consist of points
(cid:21)
We may pick exp(cid:0)Ω(d2ε)(cid:1) random points and union bound over the probability that some pair has
≤ 2 exp(cid:16)−Ω(d2ε)(cid:17)
(cid:20)
hx, yi ≥ 1
. Then by
d1/2−ε
Pr
x,y
√
.
1
d
d
large inner product.
Fix ε > 0 and C = d1/2−ε, and let P be set a set of unit vectors with pairwise inner-product at
C of size exp(Ω(d2ε)). For each a ∈ {0, 1}P consider the following norm:
most 1
kxka = C · max
y∈P :ay=1
hx, yi.
Assume there exists a randomized linear embedding f : Rd → Rd0 with the following guarantees:
17
consider rounding each entry of M by to the nearest integer multiple of
each x ∈ Sd−1, k(M − M0)xkU ≤ k(M − M0)xk2 ≤ 1
is an integer multiple of
vectors in B2.
• For every x ∈ Rd, kxka ≤ kf(x)kU ≤ Dkxka with probability at least 2
3.
Note the embedding f can be described by M, a d0 × d matrix of real numbers. Additionally, we
poly(d) to obtain M0. For
poly(d). Thus, we may assume each entry of M
poly(d)) factor in the distortion of the embedding for
We now show that the existence of the randomized embedding implies a one-way randomized
protocol for the communication problem Index. We first describe the problem. In an instance of
Index:
poly(d), and lose (1 ± 1
1
1
• Alice receives a string a ∈ {0, 1}n.
• Bob receives an index i ∈ [n].
• Alice communicates with Bob so that he can output ai.
Theorem 6.3 ([KNR99]). The randomized one-way communication complexity of Index is Ω(n).
We give a protocol for Index:
1. Suppose Alice has input a ∈ {0, 1}P . She will generate the norm k · ka described above. Note
that f ∼ D has that for each x ∈ Rd, the embedding preserves the norm of x up to D with
3. In particular, if Bob's input is i ∈ P, corresponding to point y, then an
probability 2
embedding f ∈ D, which we represent as a d0 × d matrix M, satisfies:
kyka ≤ kM ykU ≤ Dkyka
with probability 2
3. In particular, with probability 2
3:
• If ai = 0, then kyka ≤ 1, which implies kM ykU ≤ D.
• If ai = 1, then kyka ≥ C, which implies kM ykU ≥ C.
Alice computes the set Pc ⊂ P of vectors which satisfy the above property (i.e. the embedding
M preserves increases the norm by at most a factor D).
2. Alice finds a subset B ⊂ Pc of linearly independent vectors such that every x ∈ Pc we have
d0kM xkU ≤ C · D · √
x ∈ span(B). Note that B ≤ d and for all x ∈ B, kM xk2 ≤ √
d0.
Therefore, each M x ∈ Rd0 can be written with O(d0) bits. So Alice sends the set B, as well as
M x for each x ∈ B using O(dd0) bits.
3. In order for Bob to decode ai, he first checks whether y ∈ span(B), and if not, he guesses. If
y ∈ span(B), which happens with probability 2
3, then Bob writes
y = X
bi∈B
cibi
18
and M y =P
bi∈B ciM bi. If kM ykU ≤ D, then ai = 0 and if kM ykU ≥ C then ai = 1. Thus,
if D < C
2 , Bob can recover ai with probability 2
3.
dd0 ≥ Ω (P), which in turn implies d0 ≥ exp(cid:0)Ω(d2ε)(cid:1).
Alice communicates O(dd0) bits, and Bob is able to recover ai with probability 2
3. By Theorem 6.3,
Acknowledgments
We thank Piotr Indyk for suggesting the proof of Theorem 6.1. We also thank Assaf Naor for
discussions related to these research questions. Thanks to Clément Canonne for pointing us to
relevant literature about symmetric norms.
This work is supported in part by Simons Foundation, Google, and NSF (CCF-1617955), as well
as by the NSF Graduate Research Fellowship (Grant No. DGE-16-44869).
References
[ACP08]
Alexandr Andoni, Dorian Croitoru, and Mihai Patrascu. Hardness of nearest neighbor under
L-infinity. In Proceedings of the 49th Annual IEEE Symposium on Foundations of Computer
Science (FOCS '2008), pages 424–433, 2008.
Andreas Argyriou, Rina Foygel, and Nathan Srebro. Sparse prediction with the k-support norm.
In Proceedings of Advances in Neural Information Processing Systems 25 (NIPS '2012), pages
944–952, 2012.
Alexandr Andoni and Piotr Indyk. Near-optimal hashing algorithms for approximate nearest
neighbor in high dimensions. In Proceedings of the 47th Annual IEEE Symposium on Foundations
of Computer Science (FOCS '2006), pages 459–468, 2006.
Alexandr Andoni, Piotr Indyk, and Robert Krauthgamer. Overcoming the '1 non-embeddability
barrier: Algorithms for product metrics. In Proceedings of the 20th ACM-SIAM Symposium on
Discrete Algorithms (SODA '2009), pages 865–874, 2009.
[AFS12]
[AI06]
[AIK09]
[AINR14] Alexandr Andoni, Piotr Indyk, Huy L. Nguyen, and Ilya Razenshteyn. Beyond locality-sensitive
hashing. In Proceedings of the 25th ACM-SIAM Symposium on Discrete Algorithms (SODA '2014),
pages 1018–1028, 2014. Available as arXiv:1306.1547.
[AKR15] Alexandr Andoni, Robert Krauthgamer, and Ilya Razenshteyn. Sketching and embedding are
equivalent for norms. In Proceedings of the 47th ACM Symposium on the Theory of Computing
(STOC '2015), pages 479–488, 2015. Available as arXiv:1411.2577.
[And09]
[And16]
[ALRW17] Alexandr Andoni, Thijs Laarhoven, Ilya Razenshteyn, and Erik Waingarten. Optimal hashing-
based time–space trade-offs for approximate near neighbors. In Proceedings of the 28th ACM-SIAM
Symposium on Discrete Algorithms (SODA '2017), 2017. Available as arXiv:1608.03580.
Alexandr Andoni. Nearest Neighbor Search: the Old, the New, and the Impossible. PhD thesis,
MIT, 2009.
Alexandr Andoni. High-dimensional computational geometry. In Peter Bühlmann, Petros Drineas,
Michael Kane, and Mark J. van der Laan, editors, Handbook of Big Data., pages 105–123.
Chapman and Hall/CRC, 2016.
Alexandr Andoni and Ilya Razenshteyn. Optimal data-dependent hashing for approximate near
neighbors. In Proceedings of the 47th ACM Symposium on the Theory of Computing (STOC '2015),
pages 793–801, 2015. Available as arXiv:1501.01062.
Amirali Abdullah and Suresh Venkatasubramanian. A directed isoperimetric inequality with
application to Bregman near neighbor lower bounds. In Proceedings of the 47th ACM Symposium
on the Theory of Computing (STOC '2015), pages 509–518, 2015.
Keith Ball. An Elementary Introduction to Modern Convex Geometry, volume 31 of MSRI
Publications. Cambridge University Press, 1997.
[AR15]
[AV15]
[Bal97]
[BBC+15] Jarosław Błasiok, Vladimir Braverman, Stephen R. Chestnut, Robert Krauthgamer, and Lin F.
Yang. Streaming symmetric norms via measure concentration. Available as arXiv:1511.01111,
2015.
19
[BG15]
[BKL06]
[BO10]
[BS88]
[Cla99]
[CW15]
[DMS93]
[HIM12]
[IM98]
[Ind01]
[Ind02]
[Ind04]
[Joh48]
[KL04]
[KNR99]
[KP12]
[KR02]
[KW16]
[Mat96]
Yair Bartal and Lee-Ad Gottlieb. Approximate nearest neighbor search for 'p-spaces (2 < p < ∞)
via embeddings. Available as arXiv:1512.01775, 2015.
Alina Beygelzimer, Sham Kakade, and John Langford. Cover trees for nearest neighbor. In
Proceedings of the 23rd International Conference on Machine Learning (ICML '2006), pages
97–104, 2006.
Vladimir Braverman and Rafail Ostrovsky. Zero-one frequency laws. In Proceedings of the 42rd
ACM Symposium on the Theory of Computing (STOC '2010), pages 281–290, 2010.
Colin Bennett and Robert Sharpley. Interpolation of Operators, volume 129 of Pure and Applied
Mathematics. Academic Press, 1988.
Ken Clarkson. Nearest neighbor queries in metric spaces. Discrete Comput. Geom., 22(1):63–93,
1999. Previously in SoCG97.
Kenneth L. Clarkson and David P. Woodruff. Sketching for m-estimators: A unified approach to
robust regression. In Proceedings of the 26th ACM-SIAM Symposium on Discrete Algorithms
(SODA '2015), pages 921–939, 2015.
S. J. Dilworth and S. J. Montgomery-Smith. The distribution of vector-valued Rademacher series.
Annals of Probability, 21(4):2046–2052, 1993.
Sariel Har-Peled, Piotr Indyk, and Rajeev Motwani. Approximate nearest neighbor: Towards
removing the curse of dimensionality. Theory of Computing, 8(1):321–350, 2012.
Piotr Indyk and Rajeev Motwani. Approximate nearest neighbors: Towards removing the curse
of dimensionality. In Proceedings of the 30th ACM Symposium on the Theory of Computing
(STOC '1998), pages 604–613, 1998.
Piotr Indyk. On approximate nearest neighbors under '∞ norm. Journal of Computer and System
Sciences, 63(4):627–638, 2001.
Piotr Indyk. Approximate nearest neighbor algorithms for fréchet distance via product metrics.
In Proceedings of the 18th ACM Symposium on Computational Geometry (SoCG '2002), pages
102–106, 2002.
Piotr Indyk. Approximate nearest neighbor under edit distance via product metrics. In Proceedings
of the 15th ACM-SIAM Symposium on Discrete Algorithms (SODA '2004), pages 646–650, 2004.
Fritz John. Extremum problems with inequalities as subsidiary conditions. In Studies and Essays
Presented to R. Courant on his 60th Birthday, January 8, 1948, pages 187–204. Interscience
Publishers, Inc., New York, N. Y., 1948.
Robert Krauthgamer and James R. Lee. Navigating nets: simple algorithms for proximity search.
In Proceedings of the 15th ACM-SIAM Symposium on Discrete Algorithms (SODA '2004), pages
798–807, 2004.
Ilan Kremer, Noam Nisan, and Dana Ron. On randomized one-round communication complexity.
Computational Complexity, pages 21–49, 1999.
Michael Kapralov and Rina Panigrahy. NNS lower bounds via metric expansion for '∞ and EMD.
In Proceedings of the 39th International Colloquium on Automata, Languages and Programming
(ICALP '2012), pages 545–556, 2012.
David R. Karger and Matthias Ruhl. Finding nearest neighbors in growth-restricted metrics.
In Proceedings of the 34th ACM Symposium on the Theory of Computing (STOC '2002), pages
741–750, 2002.
Felix Krahmer and Rachel Ward. A unified framework for linear dimensionality reduction in L1.
RIM, 70(1):209–231, 2016.
Jiří Matoušek. On the distortion required for embedding finite metric spaces into normed spaces.
Israel Journal of Mathematics, 93:333–344, 1996.
20
[MOA11] Albert W. Marshall, Ingram Olkin, and Barry C. Arnold. Inequalities: theory of majorization
[MPS14]
[Nao17]
[NR06]
[NS07]
and its applications. Academic press, 2011.
Andrew M. McDonald, Massimiliano Pontil, and Dimitris Stamos. Spectral k-support norm regu-
larization. In Proceedings of Advances in Neural Information Processing Systems 25 (NIPS '2014),
pages 3644–3652, 2014.
Assaf Naor. A spectral gap precules low-dimensional embeddings. In Proceedings of the 33rd
International Symposium on Computational Geometry (SoCG '2017), 2017.
Assaf Naor and Yuval Rabani. On approximate nearest neighbor search in 'p, p > 2. Manuscript,
2006.
Assaf Naor and Gideon Schechtman. Planar earthmover is not in L1. SIAM Journal on Computing,
37(3):804–826, 2007. An extended abstract appeared in FOCS'06.
[PTW10] Rina Panigrahy, Kunal Talwar, and Udi Wieder. Lower bounds on near neighbor search via metric
expansion. In Proceedings of the 51st Annual IEEE Symposium on Foundations of Computer
Science (FOCS '2010), pages 805–814, 2010.
[WLKC15] Jun Wang, Wei Liu, Sanjiv Kumar, and Shih-Fu Chang. Learning to hash for indexing big data -
[Woj91]
[WSSJ14]
a survey. Available as arXiv:1509.05472, 2015.
Przemysław Wojtaszczyk. Banach Spaces for Analysts. Cambridge University Press, 1991.
Jingdong Wang, Heng Tao Shen, Kingkuan Song, and Jianqiu Ji. Hashing for similarity search: a
survey. Available as arXiv:1408.2927, 2014.
A Bounding space in Theorem 2.9
Here we justify the space bound of the algorithm from Theorem 2.9 (from [Ind02]). We note that
the improved bound was also claimed in [AIK09], albeit without a proof.
First of all, as suggested at the end of Section 3 of [Ind02], one modifies the algorithm to obtain
space of the form of n1+ε, at the expense of increasing the approximation to O(ε−1 log log n). This
is done by replacing the conditions in Case 2 and 3 by respectively:
#1+ε
"B(s, R(s) + c + 1) ∩ Si
#1+ε
"Si − B(p, R0)
Si
S
B(s, R(s)) ∩ Si
Si
,
<
Si − B(s, R0 + 2)
S
.
<
and
With the remaining algorithm being precisely the same, our only task here is to argue the space
bound. First of all we bound the sum of the number of points stored in all the leaves. For a tree with
m nodes, let L(m) be an upper bound on this count. We would like to prove that L(m) ≤ m1+ε. As
in [Ind02], we only need to focus on cases 2 and 3 of the construction, as case 1 does not replicate
the points. We will consider the case 2 (case 3 is exactly similar).
j = Si ∩ ∪s∈Sj B(s, c + 1), whereas S = m. By construction, we have
Let mj = Sj and m0
By induction, assume L(m0
thatP mj = m and mj/m > (m0
L(m) ≤X
j)1+ε for all children. Then, we have that:
mj = m1+ε.
j/m)1+ε for all j.
j) ≤ (m0
L(m0
j) ≤X
j)1+ε < mεX
(m0
j
j
j
21
We now argue the total space is O(S(n) · k log n · n). Since the depth of the tree is O(k log n),
we have that the total number of points stored in the ANN data structures is O(k log n · C(n)) =
O(k log n · n1+). Since each ANN is on at most n points, we have that, for each occurrence of a
point in the ANN data structure, we have an additional factor of S(n)/n.4 Hence the total space
occupied by all the ANN data structures is O(S(n)/n · k log n · n1+). Using a smaller ε (to hide the
log n factor), we obtain the stated space bound of O(S(n) · k · n).
B fO(log d)-ANN for symmetric norms
We provide a simple ANN algorithm for general symmetric norm achieving O(log d log log d) approx-
imation using near-linear space and sub-linear query time. The algorithm will leverage the results
in the previous section by relating general symmetric norms to Orlicz norms. Recall the definition
of level vectors in Definition 4.3.
Definition B.1. Let k · kX be any symmetric norm. Let Lk > 0 be the minimum number of
coordinates needed at level k to have norm at least 1. In other words,
Lk = min{j ∈ [d] kβ−iξ(j)kX > 1}.
At a high level, we will relate the norm of a vector x ∈ Rd to the norm of its level vectors Vk(x).
The definition above gives a way to measure the contribution of level k to the norm. For example,
if x ∈ Rd has norm kxkX ≥ D, and there are only 2 logβ d non-zero levels with respect to x, then
some level vector kVk(x)kX ≥ D
logβ d), since we may divide Vk(x) into a
sum of vectors with Lk coordinates at level k.
On the other hand, if x ∈ Rd has kxkX ≤ 1, then bk < Lk for each k. Since we consider only
2 logβ d relevant levels, for kxkS ≤ 1,
2 logβ d. This implies bk = Ω( DLk
2 logβ d−1X
k=0
≤ 2 logβ d.
bk
Lk
Additionally,P2 logβ d−1
(bk/Lk) can be decomposed as an additive contribution of coordinates. In
particular, coordinate xi contributes 1/Lk if i ∈ Bk. Therefore, we can hope to approximate the
symmetric norm by an Orlicz norms and apply the arguments from Lemma 3.1.
k=0
The lemma below formalizes the ideas discussed above.
Lemma B.2. Let k · kX be any symmetric norm. For any D, α > 1, there exists a non-decreasing
function G: R+ → R+ with G(0) = 0 and G(t) → ∞ as t → ∞, where every vector x ∈ Rd satisfies
the following:
(cid:16)xi
i=1 G(xi) ≤ 2 logβ d.
• If kxkX ≤ 1, then Pd
• If kxkX > αD · 7 logβ d, then Pd
2 logβ d−1X
i=1 G
D
(cid:17) ≥ α · 2 logβ d.
Proof. For i ≥ 0, let Ai = (β−i−1, β−i]. The function G: R+ → R+ is defined as
G(t) =
i=0
χAi(t)
Li
+ α · 2 logβ d · t · χ(1,∞)(t)
(2)
4Here we assume the natural condition that S(n) is increasing, which is, otherwise, easy to guarantee.
22
Recall the norm satisfies, kξ(1)kX = 1, so if kxkX ≤ 1, then xi ≤ 1 for all i ∈ [d]. This means
Note that G(0) = 0 and G(t) → ∞ as t → ∞.
χ(1,∞)(xi) = 0 so the second term of the RHS of (2) is zero. Therefore,
2 logβ d−1X
2 logβ d−1X
χAk(xi)
dX
dX
G(xi) =
i=1
i=1
k=0
Lk
k=0
=
bk
Lk
where bk is defined with respect to x. Since, bk < Lk for all 0 ≤ k < 2 logβ d,
If x ∈ Rd where kxkX > αD · 7 logβ d, then the vector k x
that for any vector x ∈ Rd with kxkX > α · 7 logβ d,
DkX > α · 7 logβ d. So it suffices to prove
dX
i=1
dX
i=1
G(xi) ≤ 2 logβ d.
G(xi) ≥ α · 2 logβ d
Additionally, for any vector x ∈ Rd, we may consider the vector C(x) ∈ Rd for τ = β
d2 from
Definition 4.4. By Proposition 4.6, kC(x)kX ≥ kxkX − β
d > α · 6 logβ d. Therefore, we may assume
x ∈ Rd has kxkX > α · 6 logβ d, and that all non-zero coordinates have absolute values greater than
d2 . Equivalently, bk = 0 for all k ≥ 2 logβ d. If for some i ∈ [d], xi ≥ 1, then the second term in
the RHS of (2) is non-zero, and G(xi) ≥ α · 2 logβ d. So we may further assume all coordinates of
x lie in levels k = 0, . . . , 2 logβ d − 1. Note that
β
2 logβ d−1X
dX
G(xi) =
G(Vk(x)i),
dX
and for each 0 ≤ k < 2 logβ d,Pd
(cid:26)
We partition the levels into two groups,
(cid:27)
i=1
k=0
i=1
i=1 G(Vk(x)i) = bk
Lk
A =
k bk
Lk
< 1
For all k ∈ B,
(cid:27)
≥ 1
.
k bk
Lk
.
(cid:25)
(cid:26)
B =
≤ 2bk
Lk
and
(cid:24) bk
Lk
kVk(x)kX ≤
since by the triangle inequality, we can break Vk(x) into at most
at level k each having norm at least 1.
Suppose for the sake of contradiction thatP
α · 4 logβ d ≥ X
α · 6 logβ d < kxkX ≤ X
Additionally, since kxkX > α · 6 logβ d, and
k∈B
bk
Lk
k∈B
2bk
Lk
≥ X
kVk(x)kX + X
k∈B
k∈B
≤ α · 2 logβ d. Then
kVk(x)kX .
kVk(x)kX ,
k∈A
23
(cid:24) bk
(cid:25)
Lk
vectors with Lk coordinates
it follows that
X
kVk(x)kX > α · 2 logβ d.
k∈A
However, this is a contradiction for since A ≤ 2 logβ d and kVk(x)kX ≤ 1.
Lemma B.3. For any ε ∈ (0, 1/2), there exists a data structure for ANN over any symmetric norm
k · kX which achieves approximation O
Proof. We fix β = 3
The one difference is that we rescale the '∞ norm by
(cid:17) using space O(dn1+ε) and query time O(dnε).
2. The proof of this lemma follows in the same way as the proof of Theorem 3.2.
(cid:16) log d log log d
2 logβ d after applying the embedding.
ε2
1
B.1 The logΩ(1) d-approximation is necessary
Let us remark that we cannot push the technique much further. Namely, any G(·) (even non-convex)
√
requires approximation Ω(
log d) for the following norm. Define the norm of a vector to be
This is the minimal norm for ak =
√
that an approximation with any G(·) ends up having a distortion Ω(
log d).
The idea is to consider the following vectors: for every 1 ≤ k ≤ d, we consider a vector
√
k (see Section 1.2 for the definition). It is not hard to check
(cid:18) x∗
1 + x∗
2 + . . . x∗
√
k
k
(cid:19)
.
kxk = max
1≤k≤d
(cid:16)1, 1, . . . , 1
{z
}
k
, 0, 0, . . . , 0(cid:17)
,
and besides, we consider a vector(cid:16)1,
3 − √
The remaining calculation is a simple exercise.
2 − 1,
√
√
√
d − √
2, . . . ,
d − 1(cid:17)
.
C Lower bound for arbitrary metrics: expander graphs
We give an example of a metric that is hard for current approaches to ANN search. The lower
bound is based on the notion of robust expansion, which implies all known lower bounds for ANN
[PTW10, ALRW17]. In what follows, we will refer to d = log N as the dimension of a finite metric
space of size N.
Our example of a hard metric will be the shortest path metric on any spectral expander graph. We
note that a similar theorem to the one below is also known for a finite subset of the high-dimensional
Earth-Mover Distance [KP12].
Fix M to be the metric induced by the shortest path distance on a 3-regular expander G on N
nodes. In particular, assume that 1 − λ(G) > c, where c is an absolute constant, and λ(G) ∈ (0, 1)
is the second-largest eigenvalue of the normalized adjacency matrix of G. Let d be the dimension
d = log N.
Theorem C.1. For any approximation α > 1, and data set size n ≥ 1 with dΩ(1) ≤ n ≤ N O(1),
any α-ANN data structure on n points which makes t cell probes (with cells of size at most
w ≤ (d log n)O(1)), and has success probability at least γ > n−1+o(1), must use space m =
γΩ(1/t)N Ω(1/(αt)) = γΩ(1/t)2Ω(d/(αt)).
24
We proceed by introducing a few definitions from [PTW10], and then prove lower bounds on
the robust expansion.
Definition C.2 ([PTW10]). In the Graphical Neighbor Search problem (GNS), we are given a
bipartite graph H = (U, V, E) where the dataset comes from U and the queries come from V . The
dataset consists of pairs P = {(pi, xi) pi ∈ U, xi ∈ {0, 1}, i ∈ [n]}. On query q ∈ V , if there exists a
unique pi with (pi, q) ∈ E, then we want to return xi.
One can use the GNS problem to prove lower bounds on c-ANN as follows: build a GNS graph
H by taking U = V = [N], and connecting two points u ∈ U, v ∈ V iff they are at a distance at
most r (see details in [PTW10]). We will also need to make sure that in our instances q is not closer
than cr to other points except the near neighbor.
We now introduce the notion of robust expansion, used in [PTW10] to prove lower bounds.
Definition C.3 (Robust Expansion [PTW10]). For a GNS graph H = (U, V, E), fix a distribution
e on E ⊂ U × V , and let µ be the marginal on U and η be the marginal on V . For δ, γ ∈ (0, 1], the
robust expansion Φr(δ, γ) is defined as follows:
Φr(δ, γ) =
min
A⊂V :η(A)≤δ
min
e(A×V )≥γ
B⊂U: e(A×B)
µ(B)
η(A) .
We now prove a lower bound on the robust expansion Φr(δ, γ) for a GNS graph arising from
the shortest path metric on the expander graph G. Fix r = d/α. The hard distribution e is
defined as follows: pick p at random from M and obtain q by running a random walk of length r
starting at p. Note that for n < N 1/4 and sufficiently high constant α, the distribution satisfies the
weak-independence condition required for applying the results in [PTW10].
Fix any sets A, B ⊂ M, where a = A/N and b = B/N. By the expander mixing lemma
applied to Gr, we obtain that:(cid:12)(cid:12)(cid:12)EGr(A, B) − A·B
3rN
(cid:12)(cid:12)(cid:12) ≤ λ33rqA · B.
Considering that Pr[q ∈ B p ∈ A] = EGr (A,B)
aN·3r
, we have that:
Pr[q ∈ B p ∈ A] ≤ b + λrq
γ ≤ Φr · a + λrpΦr.
b/a.
Restricting to sets A, B such that Pr[q ∈ B p ∈ A] ≥ γ, for which we must have that Φr =
Φr(a, γ) ≥ b/a (by definition), we conclude:
Hence, either Φr = Ω(γ/a) or Φr = Ω(γ2/λ2r).
Proof of Theorem C.1. Applying Theorem 1.5 from [PTW10], we have that, for t ≥ 1 cell probes,
either:
• mtw/n ≥ Ω(γ · mt), an impossibility;
• or mtw/n ≥ Ω(γ2/λ2r), or mt = Ω( n
w γ2/λ2r), implying m = γ2/tN Ω(1/(αt)).
25
To show how bad the situation for expander metrics is, we state a lower bound on for α-ANN on
the expander metric described above in the list-of-points model, which captures the hashing-based
algorithms of [ALRW17] and in the decision tree model of [Ind01]. The proofs follow from a simple
derivation using the robust expansion lower bounds in Section 7 of [ALRW17] and a reduction of
decision trees to O(log m)-cell-probe data structures similar to Appendix A in [ACP08].
Theorem C.4. Any list-of-points data structure for (c, r)-ANN for random instances of n points
in the expander metric of dimension d (described above) with query time t and space m has either
t = Ω(n), or m = exp (Ω(d)).
Theorem C.5. Let d = Ω(log1+ε n) for some ε > 0. Any decision tree of size m and depth t and
word size w succeeding with probability γ satisfies:
(cid:18)
mO(log m)tw
≥ Φr
In particular, for any ρ > 0, if w ≤ nρ, either t ≥ eΩ(n1−ρ) or m = exp(cid:16)Ω(dε/(1+ε))(cid:17) poly(n).
mO(log m) ,
O(log m)
n
1
γ
(cid:19)
.
26
|
1312.6677 | 3 | 1312 | 2015-03-05T19:51:27 | Path Finding I :Solving Linear Programs with \~O(sqrt(rank)) Linear System Solves | [
"cs.DS",
"math.NA",
"math.OC"
] | In this paper we present a new algorithm for solving linear programs that requires only $\tilde{O}(\sqrt{rank(A)}L)$ iterations to solve a linear program with $m$ constraints, $n$ variables, and constraint matrix $A$, and bit complexity $L$. Each iteration of our method consists of solving $\tilde{O}(1)$ linear systems and additional nearly linear time computation.
Our method improves upon the previous best iteration bound by factor of $\tilde{\Omega}((m/rank(A))^{1/4})$ for methods with polynomial time computable iterations and by $\tilde{\Omega}((m/rank(A))^{1/2})$ for methods which solve at most $\tilde{O}(1)$ linear systems in each iteration. Our method is parallelizable and amenable to linear algebraic techniques for accelerating the linear system solver. As such, up to polylogarithmic factors we either match or improve upon the best previous running times in both depth and work for different ratios of $m$ and $rank(A)$.
Moreover, our method matches up to polylogarithmic factors a theoretical limit established by Nesterov and Nemirovski in 1994 regarding the use of a "universal barrier" for interior point methods, thereby resolving a long-standing open question regarding the running time of polynomial time interior point methods for linear programming. | cs.DS | cs | Solving Linear Programs with (cid:101)O(
Path Finding I :
√
rank) Linear System Solves
Yin Tat Lee
MIT
[email protected]
Aaron Sidford
MIT
[email protected]
5
1
0
2
r
a
M
5
]
S
D
.
s
c
[
3
v
7
7
6
6
.
2
1
3
1
:
v
i
X
r
a
Abstract
In this paper we present a new algorithm for solving linear programs that requires only
(cid:101)O((cid:112)rank(A)L) iterations to solve a linear program with m constraints, n variables, and con-
straint matrix A, and bit complexity L. Each iteration of our method consists of solving (cid:101)O(1)
ods which solve at most (cid:101)O(1) linear systems in each iteration. Our method is parallelizable and
Our method improves upon the previous best iteration bound by factor of Ω((m/ rank(A))1/4)
for methods with polynomial time computable iterations and by Ω((m/ rank(A))1/2) for meth-
linear systems and additional nearly linear time computation.
amenable to linear algebraic techniques for accelerating the linear system solver. As such, up
to polylogarithmic factors we either match or improve upon the best previous running times for
solving linear programs in both depth and work for different ratios of m and rank(A).
Moreover, our method matches up to polylogarithmic factors a theoretical limit established
by Nesterov and Nemirovski in 1994 regarding the use of a “universal barrier” for interior point
methods, thereby resolving a long-standing open question regarding the running time of poly-
nomial time interior point methods for linear programming.
Introduction
1
Given a matrix, A ∈ Rm×n, and vectors, (cid:126)b ∈ Rm and (cid:126)c ∈ Rn, solving the linear program1
min
(cid:126)x∈Rn : A(cid:126)x≥(cid:126)b
(cid:126)cT (cid:126)x
(1.1)
is a core algorithmic task for both the theory and practice of computer science.
Since Karmarkar’s breakthrough result in 1984, proving that interior point methods can solve lin-
ear programs in polynomial time for a relatively small polynomial, interior point methods have been
an incredibly active area of research with over 1200 papers written just as of 1994 [29]. Currently,
the fastest asymptotic running times for solving (1.1) in many regimes are interior point methods.
√
Previously, state of the art interior point methods for solving (1.1) require either O(
mL) itera-
1This expression is the dual of a linear program written in standard form. It is well known that all linear programs
can be written as (1.1). Note that this notation of m and n differs from that in some papers. Here m denotes the
number of constraints and n denotes the number of variables. To avoid confusion we state many of our results in
terms of(cid:112)rank(A) instead of
√
n .
1
However,
tions of solving linear systems [31] or O((m rank(A))1/4L) iterations of a more complicated but still
polynomial time operation [36, 39, 41, 1].2
In this paper we present a new interior point method that solves general linear programs in
solutions to (1.1) and despite this existential result, the O((m rank(A))1/4L) iteration bound for
polynomial time linear programming methods has not been improved in over 20 years.
in a breakthrough result of Nesterov and Nemirovski in 1994, they showed that
there exists a universal barrier function that if computable would allow (1.1) to be solved in
O((cid:112)rank(A)L) iterations [28]. Unfortunately, this barrier is more difficult to compute than the
(cid:101)O((cid:112)rank(A)L) iterations thereby matching the theoretical limit proved by Nesterov and Nemirovski
only solving (cid:101)O(1) linear systems and performing additional (cid:101)O(nnz(A)) work in each iteration.4 Our
algorithm is parallelizable and we achieve the first (cid:101)O((cid:112)rank(A)L) depth polynomial work method
linear programming algorithm has a running time of (cid:101)O((nnz(A) + (rank(A))ω)(cid:112)rank(A)L) where
up to polylogarithmic factors.3 Furthermore, we show how to achieve this convergence rate while
for solving linear programs. Furthermore, using one of the regression algorithms in [24, 17], our
ω < 2.3729 is the matrix multiplication constant [42]. This is the first polynomial time algorithm
for linear programming to achieve a nearly linear dependence on nnz(A) for fixed n. Furthermore,
we show how to use acceleration techniques as in [37] to decrease the amortized per-iteration costs
of solving the requisite linear system and thereby achieve a linear programming algorithm with
running time faster than the previous fastest running time of O(m1.5nL) whenever m = Ω (n). This
is the first provable improvement on both running time and the number of iterations for general
interior point methods in over 20 years.
We achieve our results through an extension of standard path following techniques for linear
programming [31, 7] that we call weighted path finding. We study what we call the weighted central
path, an idea of adding weights to the standard logarithmic barrier function [35, 5, 21] that was
recently used implicitly by Mądry to make an important breakthrough improvement on the running
time for solving unit-capacity instances of the maximum flow problem [19]. We provide a general
analysis of properties of the weighted central path, discuss tools for manipulating points along the
path, and ultimately produce an efficiently computable path that converges in (cid:101)O((cid:112)rank(A)L) steps.
We hope that these results may be of independent interest and serve as tools for further improving
the running time of interior point methods in general. While the analysis in this paper is quite
technical, our linear programming method is straightforward and we hope that these techniques
may prove useful in practice.
1.1 Previous Work
Linear programming is an extremely well studied problem with a long history. There are numerous
algorithmic frameworks for solving linear programming problems, e.g. simplex methods [4], ellipsoid
methods [10], and interior point methods [8]. Each method has a rich history and an impressive body
of work analyzing the practical and theoretical guarantees of the methods. We couldn’t possibly
2Here and in the rest of the paper L denotes the standard “bit complexity” of the linear program. The parameter
L is at most the number of bits needed to represent (1.1). For integral A, (cid:126)b, and (cid:126)c the quantity L is often defined
to be the potentially smaller quantity L = log(m) + log(1 + dmax) + log(1 + max{(cid:13)(cid:13)(cid:126)c(cid:13)(cid:13)∞,(cid:13)(cid:13)(cid:126)b(cid:13)(cid:13)∞}) where dmax is the
3Here and in the remainder of the paper we use (cid:101)O(·) to hide polylog(n, m) factors.
largest absolute value of the determinant of a square sub-matrix of A [8].
4We assume that A has no rows or columns that are all zero as these can be remedied by trivially removing
constraints or variables respectively or immediately solving the linear program. Therefore nnz(A) ≥ min{m, n}.
2
cover the long line of beautiful work on this important problem in full, and we make no attempt.
Instead, here we present the major improvements on the number of iterations required to solve (1.1)
and discuss the asymptotic running times of these methods. For a more comprehensive history of
polynomial time algorithms for linear programming and interior point we refer the reader to one of
the many excellent references on the subject, e.g. [29, 43].
dominated by the time needed to solve a linear system of the form(cid:0)AT DA(cid:1) (cid:126)x = (cid:126)y for some positive
In 1984 Karmarkar [8] provided the first proof of an interior point method running in polyno-
mial time. This method required O(mL) iterations where the running time of each iteration was
diagonal matrix D ∈ Rm×m and some (cid:126)y ∈ Rn. Using low rank matrix updates and preconditioning
Karmarkar achieved a running time of O(m3.5L) for solving (1.1) inspiring a long line of research
into interior point methods.5
Karmarkar’s result sparked interest in a particular type of interior point methods, known as path
following methods. These methods solve (1.1) by minimizing a penalized objective function ft((cid:126)x),
ft((cid:126)x) where
min
(cid:126)x∈Rn
ft((cid:126)x) def= t · (cid:126)cT (cid:126)x + φ((cid:126)x)
and t is a parameter. Usually, the standard log barrier φ((cid:126)x) def= −(cid:80)
where φ : Rn → R is a barrier function such that φ((cid:126)x) → ∞ as (cid:126)x tends to boundary of the polytope
i∈[m] log([A(cid:126)x − (cid:126)b]i) is used.
Path following methods first approximately minimize ft for small t, then use this minimizer as an
initial point to minimize f(1+c)t for some constant c, and then repeat until the minimizer is close to
the optimal solution of (1.1).
Using this approach Renegar provided the first polynomial time interior point method which
√
solves (1.1) in O (
mL) iterations [31]. As with Karmarkar’s result the running time of each
iteration of this method was dominated by the time needed to solve a linear system of the form
(cid:0)AT DA(cid:1) (cid:126)x = (cid:126)y. Using a combination of techniques involving low rank updates, preconditioning
and fast matrix multiplication, the amortized complexity of each iteration was improved [38, 7, 29].
The previously fastest running time achieved by such techniques was O(m1.5nL) [37].
In a seminal work of Nesterov and Nemirovski [29], they showed that path-following methods can
in principle be applied to minimize any linear cost function over any convex set by using a suitable
barrier function. Using this technique they showed how various problems such as semidefinite
programming, finding extremal ellipsoids, and more can all be solved in polynomial time via path
following. In this general setting, the number of iterations required depended on the square root of
a quantity associated with the barrier called self-concordance. They showed that for any convex set
in Rn, there exists a barrier function, called the universal barrier function, with self-concordance
O(n). Therefore, in theory any such convex optimization problems with n variables can be solved in
nL) iterations. However, this result is generally considered to be only of theoretical interest as
O (
the universal barrier function is defined as the volume of certain polytopes, a problem which in full
generality is NP-hard and its derivatives can only approximated by solving O(nc) linear programs
for some large constant c [18].
√
Providing a barrier that enjoys a fast convergence rate and is easy minimize approximately is
an important theoretical question with numerous implications. Renegar’s path-following method
mL) linear systems. Exploiting the
effectively reduces solving a linear program to solving O(
√
5Here and in the remainder of the paper when we provide asymptotic running times for linear programming
algorithms, for simplicity we hide additional dependencies on L that may arise from the need to carry out arithmetic
operations to precision L.
3
structure of these systems yields the fastest known algorithms for combinatorial problems such as
minimum cost flow [3] and multicommodity flow [37]. Given recent breakthroughs in solving two
broad class of linear systems, symmetric diagonally dominant linear systems [33, 14, 9, 15] and
overdetermined system of linear equations [2, 24, 17] improving the convergence rate of barrier
methods while maintaining easy to compute iterations could have far reaching implications6
In 1989, Vaidya [41] made an important breakthrough in this direction. He proposed two barrier
functions related to the volume of certain ellipsoids which were shown to yield O((m rank(A))1/4 L)
and O(rank(A)L) iteration linear programming algorithms [39, 41, 36]. Unfortunately each iteration
of these methods required explicit computation of the projection matrix D1/2A(AT DA)−1AT D1/2
for a positive diagonal matrix D ∈ Rm×m. This was slightly improved by Anstreicher [1] who showed
it sufficed to compute the diagonal of this projection matrix. Unfortunately both these methods
do not yield faster running times than [37] unless m (cid:29) n and neither are immediately amenable to
take full advantage of improvements in solving structured linear system solvers.
Year Author
1984 Karmarkar [8]
1986 Renegar [31]
1989 Vaidya [40]
1994 Nesterov and Nemirovskii [29]
2013 This paper
Number of Iterations
√
O(mL)
mL)
O(
O((cid:112)rank(A)L)
O((cid:112)rank(A)L)
Nature of iterations
Linear system solve
Linear system solve
O((m rank(A))1/4 L)
Expensive linear algebra
(cid:101)O(1) Linear system solves
These results seem to suggest that you can solve linear programs closer to the O((cid:112)rank(A)L)
Volume computation
bound achieved by the universal barrier only if you pay more in each iteration. In this paper we
show that this is not the case. Up to polylogarithmic factors we achieve the convergence rate of the
universal barrier function while only having iterations of cost comparable to that of Karmarkar’s
and Renegar’s algorithms.
1.2 Our Approach
In this paper our central goal is to produce an algorithm to solve (1.1) in (cid:101)O((cid:112)rank(A)L) iterations
where each iteration solves (cid:101)O(1) linear systems of the form(cid:0)AT DA(cid:1) (cid:126)x = (cid:126)y. To achieve our goal ide-
ally we would produce a barrier function φ such that standard path following yields a (cid:101)O((cid:112)rank(A)L)
racy using (cid:101)O(1) linear system solves. Instead, we consider manipulating a barrier that we can easily
compute the gradient of, the standard logarithmic barrier, φ((cid:126)x) = −(cid:80)
iteration algorithm with low iterations costs. Unfortunately, we are unaware of a barrier function
that both yields a fast convergence rate and has a gradient that can be computed with high accu-
i∈[m] log[A(cid:126)x − (cid:126)b]i.
Note that the behavior of the logarithmic barrier is highly dependent on the representation of
(1.1). Just duplicating a constraint, i.e. a row of A and the corresponding entry in (cid:126)b, corresponds
to doubling the contribution of some log barrier term − log[A(cid:126)x − (cid:126)b]i to φ. It is not hard to see
that repeating a constraint many times can actually slow down the convergence of standard path
In other words, there is no intrinsic reason to weight all the − log[A(cid:126)x − (cid:126)b]i
following methods.
the same and the running time of path following methods do depend on the weighting of the
6Indeed, in Part II [16]we show how ideas in this paper can be used to yield the first general improvement to the
running time of solving the maximum flow problem on capacitated directed graphs since 1998 [6].
4
− log[A(cid:126)x − (cid:126)b]i. Recently, Mut and Terklaky proved that by duplicating constraints on Klee-Minty
√
cubes carefully, the standard logarithmic barrier really requires O(
m log(1/)) iterations [23].
To alleviate this issue, we add weights to the log barrier that we change during the course of
the algorithm. We show that by carefully manipulating these weights we can achieve a convergence
rate that depends on the dimension of the polytope, rank(A), rather than the number of constrains
m. In Section 4, we study this weighted log barrier function given by
gi(A(cid:126)x − (cid:126)b) · log([A(cid:126)x − (cid:126)b]i)
φ((cid:126)x) = − (cid:88)
i∈[m]
>0 → Rm
>0 is a weight function of the current point and we investigate what properties
where (cid:126)g : Rm
of (cid:126)g((cid:126)x) yield a faster convergence rate.
To illustrate the properties of the weighted logarithmic barrier, suppose for simplicity that we
normalize A and (cid:126)b so that A(cid:126)x − (cid:126)b = (cid:126)1 and let (cid:126)g def= (cid:126)g((cid:126)1). Under these assumptions, we show that
the rate of convergence of path following depends on(cid:13)(cid:13)(cid:126)g(cid:13)(cid:13)1 and
i A(cid:0)AT diag ((cid:126)g) A(cid:1)−1
AT (cid:126)1i.
(1.2)
(cid:126)1T
max
i∈[m]
To improve the convergence rate we would like to keep both these quantities small. For a general
matrix A, the quantity (1.2) is related to the leverage scores of the rows of A, a commonly used
measure for the importance of rows in a linear system [20].
i A(cid:0)AT diag ((cid:126)g) A(cid:1)−1 AT (cid:126)1i is the effective
resistance of the graph while keeping(cid:13)(cid:13)(cid:126)g(cid:13)(cid:13)1 small. Thus, if it exists, an optimal (cid:126)g would simply make
put a resistor of resistance 1/gi on the edge i. Then, (cid:126)1T
resistance of the edge i [32]. Hence, we wish to to find g to minimize the maximum effective
For illustration purposes, if we assume that A is the incidence matrix of a certain graph and
all effective resistances the same.
This electric network inverse problem is well studied [34] and motivates us to considering the
following weight function
(cid:126)g((cid:126)s) def= arg max
(cid:126)w∈Rm
−(cid:126)1T (cid:126)w +
1
α
log det(cid:0)AT S−1WαS−1A(cid:1) + β
(cid:88)
i∈[m]
log wi.
(1.3)
for carefully chosen constants α, β where S def= diag((cid:126)s((cid:126)x)) and W = diag( (cid:126)w). The optimality
conditions of this optimization problem imply that the effective resistances are small, the total
weight is small, no weight is too small, and every term in the logarithmic barrier is sufficiently
penalized. This barrier is related to the volumetric barrier function used by Vaidya [40] and can be
viewed as searching for the best function in a family of volumetric barrier function. This formulation
nL) iteration path-following algorithm by
with some careful analysis can be made to yield an O(
solving the following minimax problem
t(cid:126)cT (cid:126)x − (cid:126)1T (cid:126)w +
log det(cid:0)AT S−1WαS−1A(cid:1) + β
(cid:88)
(1.4)
√
log wi
min
(cid:126)x∈Rn
max
(cid:126)w∈Rm
1
α
i∈[m]
where (cid:126)s((cid:126)x) def= A(cid:126)x − (cid:126)b, S def= diag((cid:126)s((cid:126)x)) and W = diag( (cid:126)w).
Unfortunately, computing the derivative of the minimax formula still requires computing the
diagonal of the projection matrix as in Vaidya and Anstreicher’s work [37, 1] and is therefore too
5
inefficient for our purposes. In Section 6 we show how to compute (cid:126)w approximately up to certain
multiplicative coordinate-wise error using dimension reduction techniques. However, this error is
still too much for path following to handle the directly as multiplicatively changing weights can hurt
our measures of centrality too much.
Therefore, rather than using the weighted log barrier
where the weights (cid:126)g((cid:126)x) depends on the (cid:126)x directly, we maintain separate weights (cid:126)w and current point
(cid:126)x and use the barrier
φ((cid:126)x) = − (cid:88)
φ((cid:126)x, (cid:126)w) = − (cid:88)
i∈[m]
gi((cid:126)x) log(s((cid:126)x)i)
wi log(s((cid:126)x)i).
i∈[m]
We then maintain two invariants, (1) (cid:126)x is centered, i.e. (cid:126)x close to the minimum point of t · (cid:126)cT (cid:126)x +
φ((cid:126)x, (cid:126)w) and (2) (cid:126)w close to (cid:126)g((cid:126)x) multiplicatively.
We separate the problem of maintaining these invariants into two steps. First, we design a step
for changing (cid:126)x and (cid:126)w simultaneously that improves centrality without moving (cid:126)w too far away from
(cid:126)g((cid:126)x). We do this by decomposing a standard Newton step into a change in (cid:126)x and a change in (cid:126)w
with a ratio chosen using properties of the particular weight function. Second, we show that given
a multiplicative approximation to (cid:126)g((cid:126)x) and bounds for how much (cid:126)g((cid:126)x) may have changed, we can
maintain the invariant that (cid:126)g((cid:126)x) is close to (cid:126)w multiplicatively without moving (cid:126)w too much. We
formulate this as a general two player game and prove that there is an efficient strategy to maintain
our desired invariants. Combining these and standard techniques in path-following methods, we
obtain an O((cid:112)rank(A)L) iterations path-following algorithm where each iterations consists of O(1)
linear system solves.
We remark that a key component of our result is a better understanding of the effects of weighting
the logarithmic barrier and note that recently Mądry [19] has shown another way of using weighted
barrier functions to achieve a O(m10/7) time path-following method for the maximum flow problem
on unweighted graphs. We hope this provides further evidence of the utility of the weighted central
path discussed in later sections.
1.3 Geometric Interpretation of the Barrier
While to the best of our knowledge the specific weighted barrier, (1.3), presented in the previous
section is new, the minimax problem, (1.4), induced by the weight function is closely related to
fundamental problems in convex geometry. In particular, if we set α = 1, t = 0, and consider the
limit as β → 0 in (1.4) then we obtain the following minimax problem
−(cid:126)1T (cid:126)w + log det(cid:0)AT S−1WS−1A(cid:1) .
min
(cid:126)x∈Rn
max
(cid:126)w≥0
The maximization problem inside (1.5) is often referred to as D-optimal design and is directly
related to computing the John Ellipsoid of the polytope {(cid:126)y ∈ Rn : [A ((cid:126)y − (cid:126)x)]i ≤ s((cid:126)x)i} [11]. In
and
particular, (1.5) is directly computing the John Ellipsoid of the polytope
hence, one can view our linear programming algorithm as using approximate John Ellipsoids to
improve the convergence rate of interior point methods.
(cid:126)x ∈ Rn : A(cid:126)x ≥ (cid:126)b
(cid:110)
(1.5)
(cid:111)
6
Our algorithm is not the first instance of using John Ellipsoids in convex optimization or linear
programming. In a seminal work of Tarasov, Khachiyan and Erlikh in 1988 [12], they showed that
a general convex problem can be solved in O(n) steps of computing John Ellipsoid and querying a
separating hyperplane oracle. Furthermore, in 2008 Nesterov [26] also demonstrated how to use a
√
John ellipsoid to compute approximate solutions for certain classes of linear programs in O(
n/)
iterations and O(n2m + n1.5m/) time.
From this geometric perspective, there are two major contributions of this paper. First, we
show that the logarithmic volume of an approximate John Ellipsoid is an almost optimal barrier
function for linear programming and second, that computing approximate John Ellipsoids can be
streamlined such that the cost of these operations is comparable to pert-iteration cost of using the
standard logarithmic barrier function.
1.4 Overview
The rest of the paper is structured as follows. In Section 2 we provide details regarding the mathe-
matical notation we use throughout the paper. In Section 3 we provide some preliminary information
on linear programming and interior point methods. In Section 4 we formally introduce the weighted
path and analyze this path assuming access to weight function. In Section 5 we present our weight
function. In Section 6 we showed approximate weights suffices and in Section 7 we put everything
together to present a (cid:101)O((cid:112)rank(A)L) iteration algorithm for linear programming where in each
iteration we solve (cid:101)O(1) linear systems. Finally, in the Appendix we provide some additional mathe-
matical tools we use throughout the paper. Note that throughout this paper we make little attempt
to reduce polylogarithmic factors in our running time.
2 Notation
Here we introduce various notation that we will use throughout the paper. This section should
be used primarily for reference as we reintroduce notation as needed later in the paper. (For a
summary of linear programming specific notation we use, see Appendix A.)
Variables: We use the vector symbol, e.g. (cid:126)x, to denote a vector and we omit the symbol when
we denote the vectors entries, e.g. (cid:126)x = (x1, x2, . . .). We use bold, e.g. A, to denote a matrix. For
integers z ∈ Z we use [z] ⊆ Z to denote the set of integers from 1 to z. We let (cid:126)1i denote the vector
that has value 1 in coordinate i and is 0 elsewhere.
Vector Operations: We frequently apply scalar operations to vectors with the interpretation
that these operations should be applied coordinate-wise. For example, for vectors (cid:126)x, (cid:126)y ∈ Rn we let
(cid:126)x/(cid:126)y ∈ Rn with [(cid:126)x/(cid:126)y]i
def= (xi/yi) and log((cid:126)x) ∈ Rn with [log((cid:126)x)]i = log(xi) for all i ∈ [n] .
Matrix Operations: We call a symmetric matrix A ∈ Rn×n positive semidefinite (PSD) if
(cid:126)xT A(cid:126)x ≥ 0 for all (cid:126)x ∈ Rn and we call A positive definite (PD) if (cid:126)xT A(cid:126)x > 0 for all (cid:126)x ∈ Rn. For
a positive definite matrix A ∈ Rn×n we denote let (cid:107) · (cid:107)A : Rn → R denote the norm such that for
√
all (cid:126)x ∈ Rn we have (cid:107)(cid:126)x(cid:107)A
(cid:126)xT A(cid:126)x. For symmetric matrices A, B ∈ Rn×n we write A (cid:22) B to
indicate that B − A is PSD (i.e. (cid:126)xT A(cid:126)x ≤ (cid:126)xT B(cid:126)x for all (cid:126)x ∈ Rn) and we write A ≺ B to indicate
that B − A is PD (i.e. that (cid:126)xT A(cid:126)x < (cid:126)xT B(cid:126)x for all (cid:126)x ∈ Rn). We define (cid:31) and (cid:23) analogously. For
A, B ∈ Rn×m, we let A ◦ B denote the Schur product, i.e. [A ◦ B]ij
def= Aij · Bij for all i ∈ [n] and
j ∈ [m], and we let A(2) def= A ◦ A. We use nnz(A) to denote the number of nonzero entries in A.
def=
7
Diagonal Matrices: For A ∈ Rn×n we let diag(A) ∈ Rn denote the vector such that diag(A)i =
Aii for all i ∈ [n]. For (cid:126)x ∈ Rn we let diag((cid:126)x) ∈ Rn×n be the diagonal matrix such that
diag(diag((cid:126)x)) = (cid:126)x. For A ∈ Rn×n we let diag(A) be the diagonal matrix such that diag(diag(A)) =
diag(A). For a vector (cid:126)x ∈ Rn when the meaning is clear from context we use X ∈ Rn×n to denote
X def= diag((cid:126)x).
(cid:126)x and (cid:126)y are close multiplicatively. We often write (cid:13)(cid:13)X−1((cid:126)y − (cid:126)x)(cid:13)(cid:13)∞ ≤ to convey the equivalent
more convenient to write (cid:13)(cid:13) log (cid:126)x − log (cid:126)y(cid:13)(cid:13)∞ ≤ which is approximately equivalent for small . In
Multiplicative Approximations: Frequently in this paper we need to convey that two vectors
facts that yi ∈ [(1 − )xi, (1 + )xi] for all i or (1 − )X (cid:22) Y (cid:22) (1 + )X. At times we find it
Lemma 33, we bound the quality of this approximation.
Matrices: We use Rm
matrix A ∈ Rm×n and vector (cid:126)x ∈ Rm
>0 to denote the vectors in Rm where each coordinate is positive and for a
>0 we define the following matrices and vectors
• Projection matrix PA((cid:126)x) ∈ Rm×m: PA((cid:126)x) def= X1/2A(AT XA)−1AT X1/2.
• Leverage scores (cid:126)σA((cid:126)x) ∈ Rm: (cid:126)σA((cid:126)x) def= diag(PA((cid:126)x)).
• Leverage matrix ΣA((cid:126)x) ∈ Rm×m: ΣA((cid:126)x) def= diag(PA((cid:126)x)).
• Projection Laplacian ΛA((cid:126)x) ∈ Rm×m: ΛA((cid:126)x) def= ΣA((cid:126)x) − PA((cid:126)x)(2).
The definitions of projection matrix and leverage scores are standard when the rows of A are
reweighed by the values in vector (cid:126)x.
Convex Sets: We call a set U ⊆ Rk convex if for all (cid:126)x, (cid:126)y ∈ Rk and all t ∈ [0, 1] it holds that
t · (cid:126)x + (1 − t) · (cid:126)y ∈ U. We call U symmetric if (cid:126)x ∈ Rk ⇔ −(cid:126)x ∈ Rk. For any α > 0 and convex set
U ⊆ Rk we let αU def= {(cid:126)x ∈ Rkα−1(cid:126)x ∈ U}. For any p ∈ [1,∞] and r ∈ R≥0 the (cid:96)p ball of radius r is
given by {(cid:126)x ∈ Rk(cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)p ≤ r}.
∂xi
Calculus: For a function f : Rn → R differentiable at x ∈ Rn, we denote the gradient of f at (cid:126)x by
∇f ((cid:126)x) ∈ Rn where we have [∇f ((cid:126)x)]i = ∂
f ((cid:126)x) for all i ∈ [n]. If f ∈ Rn → R is twice differentiable
at x ∈ Rn, we denote the Hessian of f at x by∇2f ((cid:126)x) where we have [∇f ((cid:126)x)]ij = ∂2
f ((cid:126)x) for all
i, j ∈ [n]. Often we will consider functions of two vectors, g : Rn1×n2 → R, and wish to compute the
gradient and Hessian of g restricted to one of the two vectors. For (cid:126)x ∈ Rn and (cid:126)y ∈ Rm then we let
∇(cid:126)x(cid:126)g((cid:126)a,(cid:126)b) ∈ Rn1 denote the gradient of (cid:126)g for fixed (cid:126)y at point {(cid:126)a,(cid:126)b} ∈ Rn1×n2. We define ∇(cid:126)y, ∇2
(cid:126)x(cid:126)x,
(cid:126)y(cid:126)y similarly. Furthermore for h : Rn → Rm differentiable at (cid:126)x ∈ Rn we let J((cid:126)h((cid:126)x)) ∈ Rm×n
and ∇2
denote the Jacobian of (cid:126)h at (cid:126)x where for all i ∈ [m] and j ∈ [n] we let [J((cid:126)h((cid:126)x))]ij
h((cid:126)x)i. For
functions of multiple vectors we use subscripts, e.g. J(cid:126)x, to denote the Jacobian of the function
restricted to the (cid:126)x variable.
def= ∂
∂xj
∂xi∂xj
3 Preliminaries
Here we provide a brief introduction to path following methods for linear programming. The purpose
of this section is to formally introduce interior point terminology and methodology that we build
upon to obtain (cid:101)O((cid:112)rank(A)L) iteration solver. The algorithm and the analysis discussed in this
section can be viewed as a special case of the framework presented in Section 4. The reader well
8
versed in path following methods can likely skip this section and to the more curious reader we
encourage them to consider some of the many wonderful expositions on this subject [28, 43, 7] for
further reading.
3.1 The Setup
Given a matrix, A ∈ Rm×n, and vectors, (cid:126)b ∈ Rm and (cid:126)c ∈ Rn, the central goal of this paper is to
efficiently compute a solution to the following linear program
min
(cid:126)x∈Rn : A(cid:126)x≥(cid:126)b
(cid:126)cT (cid:126)x
(3.1)
It is well known that this is the dual of the standard form of a linear program and hence all linear
programs can be expressed by (3.1). We call a vector (cid:126)x ∈ Rm feasible if A(cid:126)x ≥ (cid:126)b, we call (cid:126)cT (cid:126)x the
cost of such a vector. therefore our goal is to either compute a minimum cost feasible vector or
determine that none exists.
We assume that A is full rank, i.e. rank(A) = n, and that m ≥ n. Nevertheless, we still write
many of our results using rank(A) rather than n for two reasons. First, this notation makes clear
that rank(A) is referring to the smaller of the two quantities m and n. Second, if rank(A) < n,
then we can reduce the number of variables to rank(A) by a change of basis.7 Hence, we only need
to solve linear programs in the full rank version.
3.2 Path Following Interior Point
Interior point methods solve (3.1) by maintaining a point (cid:126)x that is in the interior of the feasible
region, i.e. (cid:126)x ∈ S0 where
S0 def= {(cid:126)x ∈ Rn : A(cid:126)x > (cid:126)b}.
These methods attempt to iteratively decrease the cost of (cid:126)x while maintaining strict feasibility. This
is often done by considering some measurement of the distance to feasibility such as (cid:126)s((cid:126)x) def= A(cid:126)x−(cid:126)b,
called the slacks, and creating some penalty for these distances approaching 0. Since (cid:126)s((cid:126)x) > 0 if
and only if (cid:126)x ∈ S0 by carefully balancing penalties for small (cid:126)s((cid:126)x) and penalties for large (cid:126)cT (cid:126)x these
methods eventually compute a point close enough to the optimum solution that it can be computed
exactly.
Path following methods fix ratios between the the penalty for large (cid:126)cT (cid:126)x and the penalty for
small (cid:126)s((cid:126)x) and alternate between steps of optimizing with respect to this ratio and changing the
>0 → R such
ratio. These methods typically encode the penalties through a barrier function φ : Rm
that φ((cid:126)s((cid:126)x)) → ∞ as s((cid:126)x)i → 0 for any i ∈ [m] and they encode the ratio through some parameter
t > 0. Formally, they attempt to solve optimization problems of the following form for increasing
values of t
ft((cid:126)x) def= t · (cid:126)cT (cid:126)x + φ((cid:126)s((cid:126)x))
min
(cid:126)x∈Rm
ft((cid:126)x) where
(3.2)
Since φ((cid:126)s((cid:126)x)) → ∞ as s((cid:126)x)i → 0 the minimizer of ft((cid:126)x), denoted (cid:126)x∗(t), is in S0 for all t. As t
increases the effect of the cost vector on (cid:126)x∗(t) increases and the distance from the boundary of the
feasible region as measured by (cid:126)s((cid:126)x) decreases. One can think of the points {(cid:126)x∗(t) t > 0} as a path
minished by using a subspace embedding [24] to replace (cid:126)x with Π(cid:126)y for subspace embedding Π and (cid:101)O(rank(A))
7In general, computing this change of basis may be computationally expensive. However, this cost can be di-
dimensional (cid:126)y. Then using the reduction in Appendix E we only need to work with an (cid:101)O(rank(A)) rank matrix.
9
in Rn, called the central path, where (cid:126)x∗(t) approaches a solution to (3.1) as t → ∞. A standard
choice of barrier is the standard log barrier, φ((cid:126)s((cid:126)x)) def= −(cid:80){i} log(s((cid:126)x)i) and for this choice of barrier
we refer to {(cid:126)x∗(t) t > 0} as the standard central path.
Path following methods typically follow the following framework:
(1) Compute Initial Point: Compute an approximation (cid:126)x∗(t) for some t.
(2) Follow the central path: Repeatedly increase t and compute an approximation to (cid:126)x∗(t).
(3) Round to optimal solution: Use the approximation to (cid:126)x∗(t) to compute the solution to (3.1).
Steps (1) and (3) are typically carried out by standard interior point techniques. These techniques
are fairly general and covered briefly in Section 7 and Appendix E. However, the manner in which (2)
is performed varies greatly from method to method. In the following subsection we provide a simple
technique for performing (2) that yields reasonable running times and serves as the foundation for
the algorithms considered in the remainder of the paper.
3.3 Following the Path
There are numerous techniques to follow the central path, i.e. approximately compute (cid:126)x∗(t) for
increasing values of t. Even with the barrier fixed there are numerous schemes to balance maintaining
a point close to a central path point, advancing to a further central path point, and performing the
numerical linear algebra needed for these operations [37, 7, 22, 29].
In this section we present a simple and common method whereby we simply alternate between
improving our distance to (cid:126)x∗(t) for some fixed t, and increasing t by some fixed multiplicative factor.
This method reduces the analysis of path following to bounding the computational complexity of
centering, i.e. improve the distance to (cid:126)x∗(t), and bounding how much increasing t hurts centrality,
i.e. increases the distance to (cid:126)x∗(t). In the remainder of this section we show how to perform this
analysis for the standard central path, φ((cid:126)x) def= −(cid:80)
Typically path following methods center, i.e. minimize ft((cid:126)x), using Newton’s method or some
variant thereof. While for an arbitrary current point (cid:126)x ∈ S0 and t > 0 the function ft((cid:126)x) can be ill-
behaved, in a region near (cid:126)x∗(t) the Hessian of ft((cid:126)x) given by ∇2ft((cid:126)x) = AT S−2A for S def= diag((cid:126)s((cid:126)x))
changes fairly slowly. More precisely, if one considers the second order approximation of ft((cid:126)z) around
some point (cid:126)x ∈ S0 “close enough” to (cid:126)x∗(t) ,
i∈[m] log(s((cid:126)x)i).
ft((cid:126)z) ≈ ft((cid:126)x) + (cid:104)∇ft((cid:126)x), (cid:126)z − (cid:126)x(cid:105) +
((cid:126)z − (cid:126)x)T (∇2ft((cid:126)x)) ((cid:126)z − (cid:126)x) ,
1
2
and applies one step of Newton’s method, i.e. minimizes this quadratic approximation to compute
(cid:126)x(new) := (cid:126)x − (∇2ft((cid:126)x))−1∇ft((cid:126)x)
= (cid:126)x − (AT S−2A)−1(t(cid:126)c − AT (cid:126)s)
for (cid:126)s def= (cid:126)s((cid:126)x) then this procedure rapidly converges to (cid:126)x∗(t).
To quantify this, we measure centrality, i.e. how close the current point (cid:126)x ∈ S0 is to (cid:126)x∗(t), by
the size of this Newton step in the Hessian induced norm. For (cid:126)x ∈ S0 and Newton step (cid:126)ht((cid:126)x) def=
(∇2ft((cid:126)x))−1∇ft((cid:126)x) we denote centrality by δt((cid:126)x) def=(cid:13)(cid:13)(cid:126)ht((cid:126)x)(cid:13)(cid:13)∇2ft((cid:126)x). Standard analysis of Newton’s
10
√
Combining these facts yields that in O(
method shows that if δt((cid:126)x) is less than some constant then for (cid:126)x(new) := (cid:126)x−(cid:126)h((cid:126)x) we have δt((cid:126)x(new)) =
O(δt((cid:126)x)2) (See Lemma 5). Furthermore, under these conditions it is not hard to show that for
t(cid:48) = t(1 + (m)−1/2) we have δt(cid:48)((cid:126)x(new)) ≤ O (δt((cid:126)x)) (See Lemma 1).
m) iterations we can double t while maintaining a
nearly centered (cid:126)x, i.e. δt((cid:126)x) at most a constant. With some additional work discussed briefly in
Section 7 it can be shown that by maintaining a nearly centered (cid:126)x and changing t by a constant
factor at most (cid:101)O(L) times one can compute a solution to (3.1). Therefore, this method solves (3.1)
mL) iterations where the cost of each iteration is O(nnz(A)) plus the time need to solve a
√
in O(
linear system in the matrix AT S−2A.
4 Weighted Path Finding
In this section we introduce the optimization framework we use to solve the linear programs, the
weighted central path. After formally defining the path (Section 4.1), we prove properties of the path
(Section 4.2) and show how to center along the path (Section 4.3). We show that the performance
of path following methods using a weighted central path depends crucially on how the weights are
computed and in Section 4.4 we characterize the properties we require of such a weight function
in order to ensure that our weighted path following scheme converges efficiently. In Section 4.2 we
analyze the convergence rate of our weighted path following scheme assuming the ability to compute
these weights exactly. In the following section we then show how it suffices to compute the weights
approximately (Section 6), we show how to compute these weights efficiently (Section 5), and we
show how this yields an efficient linear program solver (Section 7).
Section 3.2. Rather than keeping the barrier function φ((cid:126)x) = −(cid:80)
4.1 The Weighted Path
Our weighted path following method is a generalization of the path following scheme presented in
i∈[m] log s((cid:126)x)i fixed we allow for
greater flexibility in how we penalize slack variables and adaptively modify the barrier function in
order to take larger steps. In addition to maintaining a feasible point (cid:126)x and a path parameter t we
maintain a set of positive weights (cid:126)w ∈ Rm
>0 and attempt to minimize the penalized objective function
ft : S0 × Rm
>0 → R given for all (cid:126)x ∈ S0 and (cid:126)w ∈ Rm
wi log s((cid:126)x)i.
(4.1)
ft((cid:126)x, (cid:126)w) def= t · (cid:126)cT (cid:126)x − (cid:88)
>0 by
i∈[m]
>0} and our goal is to compute a sequence of feasible
Note that trivially any (cid:126)x ∈ S0 can be expressed as arg min(cid:126)y∈Rn ft((cid:126)y, (cid:126)w) for some (cid:126)w ∈ Rm
We maintain a feasible point {(cid:126)x, (cid:126)w} ∈ {S0 × Rm
points for increasing t and changing (cid:126)w such that ft((cid:126)x, (cid:126)w) is nearly minimized with respect to (cid:126)x.
>0 and
therefore, every (cid:126)x ∈ S0 is a weighted central path point for some choice of weights. However, in
>0} to a solution for (1.1) we
order to to convert a weighted central path point {(cid:126)x, (cid:126)w} ∈ {S0 × Rm
will need to have t large and(cid:13)(cid:13) (cid:126)w(cid:13)(cid:13)1 small which precludes this trivial choice of t and (cid:126)w.
the weighted central path while making large increases in t and maintaining(cid:13)(cid:13) (cid:126)w(cid:13)(cid:13)1 small. Ultimately,
this will allow us to solve linear programs in (cid:101)O((cid:112)rank(A)L) iterations while only solving (cid:101)O(1) linear
In the remainder of the paper, we show that by careful updating (cid:126)x, (cid:126)w, and t we can stay close to
systems in each iteration.
11
4.2 Properties of the Weighted Path
As in Section 3.3 for a feasible {(cid:126)x, (cid:126)w} ∈ {S0 × Rm
>0} we measure the centrality of this point by the
size of the Newton step on (cid:126)x in the Hessian norm, denoted by δt((cid:126)x, (cid:126)w) and we call {(cid:126)x, (cid:126)w} a central
path point if δt((cid:126)x, (cid:126)w) = 0. For the penalized objective function ft, we see that the Newton step,
(cid:126)ht((cid:126)x, (cid:126)w), is given by
(cid:126)ht((cid:126)x, (cid:126)w) = (∇2
(cid:126)x(cid:126)xft((cid:126)x, (cid:126)w))−1∇(cid:126)xft((cid:126)x, (cid:126)w)
= (AT S−1WS−1A)−1(t(cid:126)c − AT S−1 (cid:126)w)
and the centrality, δt((cid:126)x, (cid:126)w), is given by for all {(cid:126)x, (cid:126)w} ∈ {S0 × Rm
δt((cid:126)x, (cid:126)w) def=(cid:13)(cid:13)(cid:126)ht((cid:126)x, (cid:126)w)(cid:13)(cid:13)∇2
(cid:126)x(cid:126)xft((cid:126)x, (cid:126)w) =(cid:13)(cid:13)t(cid:126)c − AT S−1 (cid:126)w(cid:13)(cid:13)(AT S−1WS−1A)
>0} by
−1
(4.2)
(4.3)
√
(cid:13)(cid:13) (cid:126)w(cid:13)(cid:13)1 =(cid:80)
i∈[m] wi.
Whereas in the standard central path we saw that the centrality increased at a rate of
m as
t increased, here we show that in this more general case, the m is replaced by the total weight
Lemma 1 (Weighted Path Step). For all {(cid:126)x, (cid:126)w} ∈ {S0 × Rm
δ(1+α)t((cid:126)x, (cid:126)w) ≤ (1 + α)δt((cid:126)x, (cid:126)w) + α
(cid:113)(cid:13)(cid:13) (cid:126)w(cid:13)(cid:13)1
>0} and t, α ≥ 0, we have
δt((cid:126)x, (cid:126)w) yields
Proof. Let (cid:126)s def= (cid:126)s((cid:126)x). By (4.3) we have
Now,(cid:13)(cid:13) ·(cid:13)(cid:13)(AT S−1WS−1A)−1 is a norm and therefore by the triangle inequality and the definition of
δ(1+α)t((cid:126)x, (cid:126)w) =(cid:13)(cid:13)(1 + α)t(cid:126)c − AT S−1 (cid:126)w(cid:13)(cid:13)(AT S−1WS−1A)−1.
δ(1+α)t((cid:126)x, (cid:126)w) ≤ (1 + α)δt((cid:126)x, (cid:126)w) + α(cid:13)(cid:13)AT S−1 (cid:126)w(cid:13)(cid:13)(AT S−1WS−1A)−1.
(cid:115)(cid:88)
(cid:13)(cid:13)AT S−1 (cid:126)w(cid:13)(cid:13)(AT S−1WS−1A)−1 =(cid:13)(cid:13)W−1/2 (cid:126)w(cid:13)(cid:13)PS−1A( (cid:126)w) ≤(cid:13)(cid:13)W−1/2 (cid:126)w(cid:13)(cid:13)2 =
(4.4)
Recall that PS−1A ( (cid:126)w) = W1/2S−1A(AT S−1WS−1A)−1AT S−1W1/2 is a projection matrix. Con-
sequently PS−1A ( (cid:126)w) (cid:22) I and we have
wi.
(4.5)
i∈[m]
Combining (4.4) and (4.5) yields the result.
Now to see how well a Newton step on (cid:126)x can center, i.e. decrease δt((cid:126)x, (cid:126)w), we need to bound how
fast the second order approximation of ft((cid:126)x, (cid:126)w) can change, i.e. how much the Hessian, ∇2
(cid:126)x(cid:126)xft((cid:126)x, (cid:126)w),
changes as we change (cid:126)x. We do this by bounding how much the slacks can change as we change (cid:126)x.
As ∇2
(cid:126)x(cid:126)xft((cid:126)x, (cid:126)w) = AT S−1WS−1A this immediately bounds how much the Hessian can change as
we change (cid:126)x. The following lemma is motivated by similar results in [40, 1].
12
Lemma 2 (Relative Change of Slacks). Let (cid:126)x(new) = (cid:126)x + (cid:126)∆ for some (cid:126)x ∈ S0 and (cid:126)∆ ∈ Rn. Let
(cid:126)s(new) and (cid:126)s denote the slacks associated with (cid:126)x(new) and (cid:126)x respectively. If (cid:13)(cid:13)S−1A(cid:126)∆(cid:13)(cid:13)∞ < 1 then
(4.6)
Proof. Clearly (cid:126)s(new) = (cid:126)s + A(cid:126)∆ and therefore the multiplicative change in slacks is given by
i∈[m]
i∈[m]
(cid:126)x(new) ∈ S0 and
(cid:13)(cid:13)W−1/2(cid:126)1i
In particular, choosing (cid:126)∆ = −(cid:126)ht((cid:126)x, (cid:126)w) yields
(cid:13)(cid:13)PS−1A( (cid:126)w).
(cid:13)(cid:13)W−1/2(cid:126)1i
(cid:13)(cid:13)PS−1A( (cid:126)w).
(cid:13)(cid:13)S−1A(cid:126)∆(cid:13)(cid:13)∞ ≤(cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13)AT S−1WS−1A · max
(cid:13)(cid:13)S−1A(cid:126)∆(cid:13)(cid:13)∞ ≤ δt((cid:126)s, (cid:126)w) · max
(cid:13)(cid:13)S−1((cid:126)s(new) − (cid:126)s)(cid:13)(cid:13)∞ =(cid:13)(cid:13)S−1A(cid:126)∆(cid:13)(cid:13)∞. Consequently (cid:126)x(new) ∈ S0 if and only if(cid:13)(cid:13)S−1A(cid:126)∆(cid:13)(cid:13)∞ < 1.
To prove (4.6) we note that by definition of(cid:13)(cid:13) ·(cid:13)(cid:13)∞
(cid:12)(cid:12)(cid:12)(cid:68)
(cid:13)(cid:13)S−1A(cid:126)∆(cid:13)(cid:13)∞ = max
(cid:12)(cid:12)(cid:12)(cid:68)(cid:0)AT S−1WS−1A(cid:1)1/2 (cid:126)∆,(cid:0)AT S−1WS−1A(cid:1)−1/2
(cid:13)(cid:13)(AT S−1WS−1A)
(cid:13)(cid:13)S−1A(cid:126)∆(cid:13)(cid:13)∞ ≤(cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13)AT S−1WS−1A · max
Using that A is full rank and therefore AT S−1WS−1A (cid:31) 0 then yields
(cid:13)(cid:13)S−1A(cid:126)∆(cid:13)(cid:13)∞ = max
(cid:13)(cid:13)AT S−1(cid:126)1i
Applying Cauchy Schwarz we have
S−1A(cid:126)∆, (cid:126)1i
i∈[m]
(cid:69)(cid:12)(cid:12)(cid:12) .
AT S−1(cid:126)1i
(cid:69)(cid:12)(cid:12)(cid:12)i
.
i∈[m[
Recalling the definition PS−1A ( (cid:126)w) = W1/2S−1A(cid:0)AS−1WS−1A(cid:1)−1 AT S−1W1/2 yields the result.
(cid:13)(cid:13)PS−1A( (cid:126)w) decreases, the region over which Newton steps do
Lemma 2 implies that as(cid:13)(cid:13)W−1/2(cid:126)1i
not change the Hessian too much increases. We call this quantity, (cid:13)(cid:13)W−1/2(cid:126)1i
(cid:13)(cid:13)PS−1A( (cid:126)w), the slack
i∈[m[
−1.
sensitivity as it measures how much slack changes during a Newton step.
Definition 3 (Slack Sensitivity). For (cid:126)s, (cid:126)w ∈ Rm
>0 the slack sensitivity8, γ((cid:126)s, (cid:126)w) is given by
(cid:13)(cid:13)W−1/2(cid:126)1i
(cid:13)(cid:13)PS−1A( (cid:126)w).
γ((cid:126)s, (cid:126)w) def= max
i∈[m]
Geometrically, slack sensitivity indicates how much a relative slack can change during a Newton
step, equivalently, how small is the Newton step region compared to the original polytope. From
Lemmas 1 and 2 our goal in using the weighted central path is clear. We wish to keep(cid:13)(cid:13) (cid:126)w(cid:13)(cid:13)1 small so
that we can make large increases to t without increasing centrality and we wish to keep γ((cid:126)s((cid:126)x), (cid:126)w)
small so that over a large region we can improve centrality quickly. Unfortunately, while it is not
too difficult to produce weights that meet these criterion, changing the weights can also increase
δt. Therefore, we also need to choose weights in such a way that they do not change too drastically
as we take Newton steps.
In the next subsection we introduce the step that we use to improve
centrality and account for possible changes in the weights.
8In the previous version in ArXiv, we called it weighted condition number which is confusing. We are indebted to
an anonymous reviewer for suggesting this name.
13
4.3 Centering Steps
Here we define the centering step we use to decrease δt((cid:126)x, (cid:126)w). There are two ways to decrease δt,
one is to perform a Newton step on (cid:126)x which corresponds to move (cid:126)x closer to the central path., one is
to set (cid:126)w such that δt((cid:126)x, (cid:126)w) = 0 which corresponds to move the path itself to closer to (cid:126)x. By mixing
two steps, we can slow down progress along a specific weighted path as much as we want but still
obtaining the guarantee of Newton method. We call this r-step where r controls the ratio of how
much we change (cid:126)w and (cid:126)x. Setting r = 0 corresponds to a standard Newton step on (cid:126)x where the
weights are not updated. Setting r = ∞ coresponds to changing (cid:126)w to make (cid:126)x completely centered.
There are two reasons we do thisinstead of a standard Newton step:
1. When we change (cid:126)x, we need to change the weights (cid:126)w accordingly to maintain the the properties
we want. However, when we change the weights (cid:126)w, we need to update (cid:126)x again, and so on. For
the weight function we consider in Section 5 the change of (cid:126)w required is large. Consequently,
after updating the weights we need to move (cid:126)x even more and it is not clear how to maintain
good weights and good centrality at the same time if we neglect the direction in which the
weights change. However, the weights we use actual change in a direction which partial helps
improve centrality. Considering a r-step helps us account for this progress directly.
2. We cannot compute the weights we want to use exactly. Instead we only know how to compute
them approximately up to 1/polylog(m) multiplicative error using Johnson–Lindenstrauss. .
Therefore, if we take a full Newton step on (cid:126)x and update the weights using the weight function,
the error in our approximation is possibly so large that the step in full would not help centrality.
To control this error and center when we cannot compute the weights exactly, we exploit that
the r-step gives us part of the change in the weights that we can compute precisely.
Definition 4 (r-step). Given a feasible point {(cid:126)x(old), (cid:126)w(old)} ∈ {S0 × Rm
a r-step
>0}, a path parameter t, and
{(cid:126)x(new), (cid:126)w(new)} = stept((cid:126)x(old), (cid:126)w(old), r)
is defined as follows
where we recall that
(cid:126)x(new) def= (cid:126)x(old) − 1
1 + r
r
(cid:126)w(new) def= (cid:126)w(old) +
(cid:126)ht((cid:126)x(old), (cid:126)w(old)),
W(old)S−1
1 + r
(old)A(cid:126)ht((cid:126)x(old), (cid:126)w(old))
(cid:126)ht((cid:126)x(old), (cid:126)w(old)) def= (AT S−1
(old)W(old)S−1
(old)A)−1(t(cid:126)c − AT S−1
(old) (cid:126)w(old))
and we let (cid:126)s(old) and (cid:126)s(new) denote the slacks with (cid:126)x(old) and (cid:126)x(new) respectively.
Note that for a r-step we have
(cid:126)s(new) = (cid:126)s(old) − 1
1 + r
A(cid:126)h((cid:126)x(old), (cid:126)w(old))
and therefore
W−1
(old)( (cid:126)w(new) − (cid:126)w(old)) = −rS−1
(old)((cid:126)s(new) − (cid:126)s(old)).
(4.7)
(4.8)
14
In other words, a r-step performs a multiplicative update on the weights that is exactly r times
larger than the update on the slacks.
Using Lemma 2 we now show that so long as δt((cid:126)x(old), (cid:126)w(old)) is reasonably small with respect
to the slack sensitivity, any r-step produces a feasible {(cid:126)x(new), (cid:126)w(new)} and does not change the
Hessian too much.
Lemma 5 (Stability of r-step). Let {(cid:126)x(new), (cid:126)w(new)} = stept((cid:126)s(old), (cid:126)w(old), r) where
γ def= γ((cid:126)x(old), (cid:126)w(old))
and
δt
def= δt((cid:126)x(old), (cid:126)w(old)) ≤ 1
8γ
.
(old)((cid:126)s(new) − (cid:126)s(old))(cid:13)(cid:13)W(old)
Under these conditions we have(cid:13)(cid:13)S−1
(old)((cid:126)s(new) − (cid:126)s(old))(cid:13)(cid:13)∞ ≤ 1
(cid:13)(cid:13)S−1
(old)( (cid:126)w(new) − (cid:126)w(old))(cid:13)(cid:13)∞ ≤ r
(cid:13)(cid:13)W−1
≤ 1
1 + r
1 + r
1 + r
· δt,
· δt · γ,
· δt · γ.
Consequently {(cid:126)x(new), (cid:126)w(new)} is feasible and
(1 − 3δtγ)∇2
(cid:126)x(cid:126)xft((cid:126)x(old), (cid:126)w(old)) (cid:22) ∇2
(cid:126)x(cid:126)xft((cid:126)x(new), (cid:126)w(new)) (cid:22) (1 + 3δtγ)∇2
(cid:126)x(cid:126)xft((cid:126)x(old), (cid:126)w(old))
(4.9)
(4.10)
(4.11)
(4.12)
Proof. Equation (4.9) follows from the definition of δt and (4.7). Equations (4.10) and (4.11) follow
from Lemma 2, the definition of γ, (4.7), and (4.8). Since δt ≤ 1
8γ this implies that slack or weight
>0}.
changes by more than a multiplicative factor of 1
8 and therefore clearly {(cid:126)s(new), (cid:126)w(new)} ∈ {S0×Rm
To prove (4.12) note that (4.10) and (4.11) imply that
(cid:16)
(cid:16)
1 − r
1+r δtγ
1 − 1
1+r δtγ
(cid:17)
W(old) (cid:22) W(new) (cid:22)(cid:16)
(cid:17)
S(old) (cid:22) S(new) (cid:22)(cid:16)
1 + r
1+r δtγ
(cid:17)
(cid:17)
1 + 1
S(old).
>0} we have that
1+r δtγ
W(old),
(cid:126)x(cid:126)xft((cid:126)x, (cid:126)w) = AT S−1WS−1A for (cid:126)x, (cid:126)w ∈ {S0 × Rm
(cid:126)x(cid:126)xft((cid:126)x(new), (cid:126)w(new)) (cid:22)
(cid:126)x(cid:126)xft((cid:126)x(old), (cid:126)w(old)).
Since ∇2
1 − r
(cid:16)
(cid:16)
1+r δtγ
1 + 1
1+r δtγ
(cid:17)
(cid:17)2∇2
(cid:126)x(cid:126)xft((cid:126)x(old), (cid:126)w(old)) (cid:22) ∇2
(cid:16)
(cid:16)
(cid:17)
(cid:17)2 ≤ 1 + 3δtγ
1+r δtγ
1 + r
1 − 1
1+r δtγ
(cid:16)
(cid:16)
and
1 − r
1+r δtγ
1 + 1
1+r δtγ
(cid:16)
(cid:16)
(cid:17)
(cid:17)2∇2
1+r δtγ
1 + r
1 − 1
1+r δtγ
(cid:17)
(cid:17)2 ≥ 1 − 3δtγ.
Using that 0 ≤ δtγ ≤ 1
8 and computing the Taylor series expansions9 yields that
9Throughout this paper, when we use taylor series expansions we may use more than just the second order
approximation to the function.
15
Using Lemma 5 we now bound how much a r-step improves centrality.
Lemma 6 (Centrality Improvement of r-step). Let {(cid:126)x(new), (cid:126)w(new)} = stept((cid:126)x(old), (cid:126)w(old), r) where
γ def= γ((cid:126)x(old), (cid:126)w(old))
and
δt
def= δt((cid:126)x(old), (cid:126)w(old)) ≤ 1
8γ
.
We have the following bound on the change in centrality
δt((cid:126)x(new), (cid:126)w(new)) ≤ 2
1 + r
· γ · δ2
t .
Proof. Let (cid:126)ht
definition of stept, we see that
def= (cid:126)ht((cid:126)x(old), (cid:126)w(old)) and let (cid:126)∆ def= S−1
(old)((cid:126)s(new) − (cid:126)s(old)) = −1
(old)A(cid:126)ht. Recalling the
(cid:32)
(cid:33)
(cid:32)
·
1+r S−1
(cid:33)
1 − r (cid:126)∆i
1 + (cid:126)∆i
(cid:126)w(new)
i
(cid:126)s(new)
i
=
=
(cid:32)
(cid:126)w(old)
i
(cid:126)s(old)
i
(cid:126)w(old)
i
(cid:126)s(old)
i
(cid:126)w(old)
i
(cid:126)s(old)
i
(cid:33)
i
i
=
− r (cid:126)w(old)
(cid:126)∆i
(cid:33)(cid:32)
+ (cid:126)s(old)
(cid:126)∆i
1 − (1 + r)(cid:126)∆i
1 + (cid:126)∆i
(cid:16)
(old) (cid:126)w(old) =
∇xft((cid:126)x(old), (cid:126)w(old)) = t(cid:126)c − AT S−1
= −(1 + r)AT S−1
(old)W(old)
Using the definition of (cid:126)ht we have
(4.13)
(4.14)
(4.15)
(cid:17) (cid:126)ht
(old)A
AT S−1
(cid:126)∆
(old)W(old)S−1
(cid:17)
.
and therefore
t(cid:126)c = AT S−1
(old)W(old)
Combining (4.13) and (4.14) and using the definition of (cid:126)∆ then yields
∇xft((cid:126)x(new), (cid:126)w(new)) = t(cid:126)c − AT S−1
(new) (cid:126)w(new)
(cid:16)(cid:126)1 − (1 + r)(cid:126)∆
(cid:32)
= AT S−1
(old)W(old)
(cid:126)1 − (1 + r)(cid:126)∆ − (cid:126)1 +
(cid:33)
(1 + r)(cid:126)∆
(cid:126)1 + (cid:126)∆
= −(1 + r)AT S−1
= AT S−1
(old)W(old)S−1
(old)W(old)
(cid:126)∆2
(cid:126)1 + (cid:126)∆
(old)diag((cid:126)∆)(I + diag((cid:126)∆))−1A(cid:126)ht
Now by Lemma 5 we know that
(new)W(new)S−1
Therefore by (4.15) and the fact that
AT S−1
(cid:16)
(cid:126)w(old)(cid:17)
PS−1
(old)A
(new)A (cid:23) (1 − 3δtγ)AT S−1
(old)A.
(old)W(old)S−1
(cid:17)−1
AT S−1
= W1/2
(old)S−1
(old)A
AT S−1
(old)WS−1
(old)A
(old)W1/2
(old) (cid:22) I,
(cid:16)
16
S−1
(old)
A( (cid:126)w(old))
AT S−1
(new)W(new)S−1
(new)A
we have
δt((cid:126)x(new), (cid:126)w(new)) =(cid:13)(cid:13)∇xft((cid:126)x(new), (cid:126)w(new))(cid:13)(cid:13)(cid:16)
(old)A(cid:126)h(cid:13)(cid:13)P
(old)A(cid:126)h(cid:13)(cid:13)2
(cid:17)−1
≤ (1 − 3δtγ)−1/2(cid:13)(cid:13)diag((cid:126)∆)(I + diag((cid:126)∆))−1W1/2
(old)S−1
≤ (1 − 3δtγ)−1/2(cid:13)(cid:13)diag((cid:126)∆)(I + diag((cid:126)∆))−1W1/2
(old)S−1
(old)A(cid:126)h(cid:13)(cid:13)2
≤ (1 − 3δtγ)−1/2
(cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13)∞
(cid:13)(cid:13)W1/2
1 −(cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13)∞
(old)S−1
(cid:32) (cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13)∞
(cid:33)
1 −(cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13)∞
where in the last step we use that by Lemma 5,(cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13)∞ ≤ 1
quickly we need to maintain weights such that (cid:13)(cid:13) (cid:126)w(cid:13)(cid:13)1, γ((cid:126)s, (cid:126)w), and δt((cid:126)x, (cid:126)w) are small. Rather
4.4 Weight Functions
In Sections 4.1, 4.2, and 4.3 we saw that to make our weighted path following schemes to converge
1+r δtγ and that δ ≤ 1
than showing how to do this directly, here we assume we have access to some fixed differentiable
function for computing the weights and we characterize when such a weight function yields an
efficient weighted path following scheme. This allows us to decouple the problems of using weights
effectively and computing these weights efficiently.
8γ by assumption.
= (1 − 3δtγ)
≤ 2
1 + r
−1/2 ·
δt
γδ2
t
>0 → Rm
For the remainder of this paper we assume that we have a fixed differentiable weight function
>0 from slacks to positive weights (see Section 5 for a description of the function we
(cid:126)g : Rm
use). For slacks (cid:126)s ∈ Rm
>0 we let G((cid:126)s) def= diag((cid:126)g((cid:126)s)) denote the diagonal matrix associated with the
slacks and we let G(cid:48)((cid:126)s) def= J(cid:126)s((cid:126)g((cid:126)s)) denote the Jacobian of the weight function with respect to the
slacks.
bounded, and good slack sensitivity, i.e. γ((cid:126)x, (cid:126)g((cid:126)s((cid:126)x))) small, we need to ensure that the weights do
not change too much as we change (cid:126)x. For this, we use the operator norm of I + r−1G((cid:126)s)−1G(cid:48)((cid:126)s)S
to measure for how much the weight function can diverge from the change in weights induced by
a r-step, i.e. how consistent (cid:126)g is to the central path. Lastly, to simplify the analysis we make a
For the weight function to be useful, in addition to yielding weights of small size, i.e. (cid:13)(cid:13)(cid:126)g((cid:126)s)(cid:13)(cid:13)1
uniformity assumption that none of the weights are two big, i.e. (cid:13)(cid:13)(cid:126)g((cid:126)s)(cid:13)(cid:13)∞ is bounded. Formally we
• Size : The size c1((cid:126)g) satisfies c1((cid:126)g) ≥(cid:13)(cid:13)(cid:126)g((cid:126)s)(cid:13)(cid:13)1
(cid:13)(cid:13)I + r−1G((cid:126)s)−1G(cid:48)((cid:126)s)S(cid:13)(cid:13)G((cid:126)s) ≤ 1 and (cid:13)(cid:13)(cid:0)I + r−1G((cid:126)s)−1G(cid:48)((cid:126)s)S(cid:1) (cid:126)y(cid:13)(cid:13)∞ ≤(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)∞ + cr
• Slack Sensitivity: The slack sensitivity cγ((cid:126)g) satisfies cγ((cid:126)g) ≥ 1 and γ((cid:126)s, (cid:126)g((cid:126)s)) ≤ cγ((cid:126)g).
• Step Consistency : The step consistency cr((cid:126)g) satisfies cr((cid:126)g) ≥ 1 and ∀r ≥ cr((cid:126)g) and ∀(cid:126)y ∈ Rm
Definition 7 (Weight Function). A weight function is a differentiable function from (cid:126)g : Rm
such that for constants c1((cid:126)g), cγ((cid:126)g), and cr((cid:126)g), we have the following for all (cid:126)s ∈ Rm
>0:
define a weight function as follows.
(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)G((cid:126)s).
>0 → Rm
>0
17
• Uniformity : The weight function satisfies(cid:13)(cid:13)(cid:126)g((cid:126)s)(cid:13)(cid:13)∞ ≤ 2
When the weight function (cid:126)g is clear from context we often write c1, cγ, and cr.
To get a sense of the magnitude of these parameters, in Theorem 12 we prove that there is a
rank A), slack sensitivity O(1) and step consistency O(cid:0)log(cid:0) m
weight function with size O(
hence lemmas with polynomial dependence of slack sensitivity and step consistency suffice for our
purposes. However, for the remainder of this section and Section 6 we let the weight function be
fixed but arbitrary.
(cid:1)(cid:1);
rank A
√
Ideally, in our weighted path following schemes we would just set (cid:126)w = (cid:126)g((cid:126)s) for any slacks (cid:126)s we
compute. However, actually computing (cid:126)g((cid:126)s) may be expensive to compute exactly and therefore we
analyze schemes that maintain separate weights, (cid:126)w ∈ Rm
>0 with the invariant that (cid:126)w is close to (cid:126)g((cid:126)s)
multiplicatively. Formally, we define (cid:126)Ψ((cid:126)s, (cid:126)w) for all (cid:126)s, (cid:126)w ∈ Rm
and attempt to keep(cid:13)(cid:13)(cid:126)Ψ((cid:126)s, (cid:126)w)(cid:13)(cid:13)∞ small despite changes that occur due to r-steps.
>0 by
(cid:126)Ψ((cid:126)s, (cid:126)w) def= log((cid:126)g((cid:126)s)) − log( (cid:126)w)
Now we wish to show that a r-step does not increase (cid:126)Ψ((cid:126)s, (cid:126)w) by too much. To do this, we first
(4.16)
prove the following helper lemma.
Lemma 8. For a weight function (cid:126)g and (cid:126)s0, (cid:126)s1 ∈ S0 such that
0 ((cid:126)s1 − (cid:126)s0)(cid:13)(cid:13)∞ ≤ 1
∞ def=(cid:13)(cid:13)S−1
(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)∞
(cid:18)(cid:126)g((cid:126)s1)
(cid:126)g((cid:126)s0)
log
32cr
def=(cid:13)(cid:13)S−1
(cid:19)
0 ((cid:126)s1 − (cid:126)s0)(cid:13)(cid:13)G((cid:126)s0) ≤ ∞
(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)G((cid:126)s0)
(cid:18)(cid:126)g((cid:126)s1)
(cid:126)g((cid:126)s0)
log
cr
+
.
and
g
(cid:18)(cid:126)s1
(cid:13)(cid:13)(cid:13)(cid:13)log
+
1
cr
1
cr
Proof. Let (cid:126)p : Rm → Rm be defined for all i ∈ [m] and s ∈ Rm
≤ 3∞ and
(cid:126)s0
(cid:126)s0
>0 by
we have
(cid:18)(cid:126)s1
(cid:19)
(cid:13)(cid:13)(cid:13)(cid:13)log
≤ (1 + 6cr∞) g.
(cid:126)p((cid:126)s)i
def= log((cid:126)si) +
1
cr
log((cid:126)g((cid:126)si)).
Clearly J(cid:126)s[(cid:126)p((cid:126)s)] = S−1 + c−1
we see that for all i ∈ [0, 1] ,
r G−1((cid:126)s)G(cid:48)((cid:126)s). Therefore, letting (cid:126)st
def= (cid:126)s0 + t((cid:126)s1 − (cid:126)s0) for all t ∈ [0, 1]
(cid:126)p((cid:126)si) = (cid:126)p((cid:126)s0) +
S−1
t +
1
cr
G−1((cid:126)st)G(cid:48)((cid:126)st)
((cid:126)s1 − (cid:126)s0)dt.
Applying Jensen’s inequality and the definition of (cid:126)p then yields that for all i ∈ [0, 1] and any norm
(cid:20)
i
0
(cid:21)
(cid:21)
(cid:13)(cid:13) ·(cid:13)(cid:13) , (cid:13)(cid:13)(cid:13)(cid:13)log
(cid:18) (cid:126)si
(cid:19)
(cid:126)s0
(cid:18) (cid:126)g((cid:126)si)
(cid:126)g((cid:126)s0)
(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) ≤
0
+
1
cr
log
Now for all t ∈ [0, 1] define (cid:126)at ∈ Rm
>0 by
(cid:126)at
def= log
i
(cid:13)(cid:13)(cid:13)(cid:13)(cid:20)
(cid:18) (cid:126)st
(cid:126)s0
I +
G−1((cid:126)st)G(cid:48)((cid:126)st)St
1
cr
S−1
t ((cid:126)s1 − (cid:126)s0)
(4.17)
(cid:13)(cid:13)(cid:13)(cid:13) dt.
(cid:18) (cid:126)g((cid:126)st)
(cid:19)
(cid:126)g((cid:126)s0)
(cid:19)
− 1
cr
log
18
t ∈ [0, i]. By Lemma 33 and the fact that ∞ ≤ 1
(4.17) and applying Definition 7 yields that ∀i ∈ [0, M ] ,
and let M be the supremum over all i ∈ [0, 1] such that(cid:13)(cid:13) log (cid:126)g((cid:126)st) − log (cid:126)g((cid:126)s0)(cid:13)(cid:13)∞ ≤ 3.5cr∞ for all
this implies that(cid:13)(cid:13)G((cid:126)si)−1((cid:126)g((cid:126)si) − (cid:126)g((cid:126)s0))(cid:13)(cid:13)∞ ≤
4cr∞ and(cid:13)(cid:13)G((cid:126)s0)−1((cid:126)g((cid:126)s0) − (cid:126)g((cid:126)si))(cid:13)(cid:13)∞ ≤ 4cr∞ for all i ∈ [0, M ]. Therefore, choosing(cid:13)(cid:13) ·(cid:13)(cid:13)G((cid:126)s0) in
t ((cid:126)s1 − (cid:126)s0)(cid:13)(cid:13)G((cid:126)st)dt ≤ (1 + 4cr∞)
(cid:13)(cid:13)S−1
Similarly, by choosing(cid:13)(cid:13) ·(cid:13)(cid:13)∞ in (4.17), we have ∀i ∈ [0, M ] that
(cid:16)(cid:13)(cid:13)S−1
t ((cid:126)s1 − (cid:126)s0)(cid:13)(cid:13)G((cid:126)st)
(cid:13)(cid:13)S−1
t ((cid:126)s1 − (cid:126)s0)(cid:13)(cid:13)∞ + cr
(cid:13)(cid:13)G((cid:126)s0) < (1 + 4cr∞)1/2
(cid:13)(cid:13)∞ <
g ≤ (1 + 6cr∞) g.
(cid:13)(cid:13)(cid:126)ai
(cid:13)(cid:13)(cid:126)ai
1 − ∞
(cid:17)
32cr
dt
0
i
i
√
0
∞
1 − ∞
<
+
1 + 4cr∞
1 − ∞
crg ≤ 2.2∞
8, we have
(cid:114) 8
7
≤ 1
2
19
By the definition of (cid:126)ai, the triangle inequality, and Lemma 33 we then have that
(cid:13)(cid:13) log((cid:126)g((cid:126)si)) − log((cid:126)g((cid:126)s0))(cid:13)(cid:13)∞ < cr
(cid:0)2.2∞ +(cid:13)(cid:13) log((cid:126)si) − log((cid:126)s0)(cid:13)(cid:13)∞
(cid:1) < 3.5cr∞.
Since (cid:126)g is continuous we have that M = 1 and the result follows.
Using this lemma we bound on how much a r-step increases (cid:126)Ψ as follows
Lemma 9. Let {(cid:126)x(new), (cid:126)w(new)} = stept((cid:126)x(old), (cid:126)w(old), cr) where
δt
def= δt((cid:126)x(old), (cid:126)w(old)) ≤ 1
8cγ
Letting
− log
we have
(cid:126)∆ def= log
(cid:126)g((cid:126)s(new))
(cid:126)g((cid:126)s(old))
(cid:32)
(cid:33)
(cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13)∞ ≤ 4cγδt
(cid:13)(cid:13)W−1/2(cid:126)1i
Since(cid:13)(cid:13) log((cid:126)g((cid:126)s(old))) − log( (cid:126)w(old))(cid:13)(cid:13)∞ ≤ 1
γ((cid:126)s, (cid:126)w) = max
i∈[m]
.
8
and
(cid:32)
(cid:126)w(new)
(cid:126)w(old)
def=(cid:13)(cid:13) log((cid:126)g((cid:126)s(old))) − log( (cid:126)w(old))(cid:13)(cid:13)∞ ≤ 1
(cid:33)
(cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13)W(new)
(cid:13)(cid:13)(cid:126)1i
(cid:13)(cid:13)PS−1A( (cid:126)w) = max
(cid:13)(cid:13)S−1A(AT S−1WS−1A)
= (cid:126)Ψ((cid:126)s(new), (cid:126)w(new)) − (cid:126)Ψ((cid:126)s(old), (cid:126)w(old)),
≤ ecr
1 + cr
δt + 13cγδ2
t .
−1AT S−1.
and
i∈[m]
Proof. Recall the following definition of slack sensitivity
γ((cid:126)s(old), (cid:126)w(old)) ≤
γ((cid:126)s(old), (cid:126)g((cid:126)s(old))) ≤ 1.1cγ.
Therefore, since δt ≤ 1
(cid:13)(cid:13)W−1
(old)( (cid:126)w(new) − (cid:126)w(old))(cid:13)(cid:13)∞ ≤ 1.1crcγδt
64cγ cr
1 + cr
and (cid:13)(cid:13)S−1
(old)((cid:126)s(new) − (cid:126)s(old))(cid:13)(cid:13)∞ ≤ 1.1cγδt
1 + cr
, by Lemma 5 and (4.18) we have
(4.18)
.
≤ 1
2
(4.19)
(old)((cid:126)s(new)−(cid:126)s(old)) and using that cr ≥ 1 and − 2 ≤
i
i
1 +
s(old)
(cid:32)
(4.20)
+ cr log
s(new)
i
− s(old)
(cid:33)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(old)((cid:126)s(new) − (cid:126)s(old))(cid:13)(cid:13)∞ ≤ 1.1cγ δt
(cid:33)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:33)
(cid:13)(cid:13)S−1
(old)((cid:126)s(new) − (cid:126)s(old))(cid:13)(cid:13). (4.21)
(cid:126)s(new)
(cid:126)s(old)
+ cr log
(cid:32)
≤
1+cr
(cid:126)w(new)
(cid:126)w(old)
i
i
i
r
i
i
s(old)
(cid:33)
(cid:32)
(cid:33)
+ cr log
s(new)
i
≤ 2c2
− s(old)
− (cid:126)s(old)
s(new)
i
s(old)
i
w(new)
i
w(old)
(cid:33)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =
Recalling that W−1
log(1 + ) ≤ for < 1
(cid:32)
2 we have that for all i ∈ [m]
1 − cr
(old)( (cid:126)w(new)− (cid:126)w(old)) = −crS−1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)log
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)log
(cid:32)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:126)s(new)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2
Letting(cid:13)(cid:13) ·(cid:13)(cid:13) denote either(cid:13)(cid:13) ·(cid:13)(cid:13)∞ or(cid:13)(cid:13) ·(cid:13)(cid:13)W(old)
, recalling that(cid:13)(cid:13)S−1
(cid:33)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) +
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)log
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)cr log
(cid:32)
(cid:32)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)log
(cid:33)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + 2.2crcγδt
. Therefore, letting(cid:13)(cid:13) ·(cid:13)(cid:13) be(cid:13)(cid:13) ·(cid:13)(cid:13)∞ in (4.21), we have
(cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13) ≤
, and applying (4.20) yields
(cid:126)g((cid:126)s(new))
(cid:126)g((cid:126)s(old))
(cid:126)g((cid:126)s(new))
(cid:126)g((cid:126)s(old))
(cid:126)s(new)
(cid:126)s(old)
(cid:126)s(new)
(cid:126)s(old)
(cid:32)
(cid:32)
(cid:33)
(cid:33)
≤ cr
(cid:126)s(old)
i
(cid:32)
+ log
1
cr
1.1cγ δt
+
log
cr
1+cr
(cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13)∞ ≤ 3cr∞ + 2.2crcγδt
Similarly, letting(cid:13)(cid:13) ·(cid:13)(cid:13) be(cid:13)(cid:13) ·(cid:13)(cid:13)W(old)
(cid:13)(cid:13)(cid:126)∆(cid:13)(cid:13)W(old)
(old)( (cid:126)w(new) − (cid:126)w(old))(cid:13)(cid:13)∞ ≤ 1.1cγδt yields the result.
≤ e/2crg(1 + 6cr∞) + 2.2crcγδt
≤ e
Finally, noting that(cid:13)(cid:13)W−1
1.1cγδt
1 + cr
δt + 10cγδ2
t .
≤ 4cγδt.
1 + cr
1 + cr
cr
δt
By Lemma 5 and (4.19), (cid:126)s(old) and (cid:126)s(new) meet the conditions of Lemma 8 with ∞ ≤ 1.1cγ δt
g ≤ e/2δt
1+cr
and
in (4.19) and noting that by definition of yields
4.5 Centering Using Exact Weights
Here we bound the rate of convergence rate of path following assuming that we can compute the
weight function (cid:126)g exactly. We start by providing a basic lemma regarding how the Newton step size
changes as we change (cid:126)w.
Lemma 10 (Effect of Weight Change). Let (cid:126)x ∈ S0 and let (cid:126)w(old), (cid:126)w(new) ∈ Rm
>0 with
∞ def=(cid:13)(cid:13) log( (cid:126)w(new)) − log( (cid:126)w(old))(cid:13)(cid:13)∞ ≤ 1
2
it follows that
δt((cid:126)x, (cid:126)w(new)) ≤ (1 + ∞)
(cid:104)
δt((cid:126)x, (cid:126)w(old)) +(cid:13)(cid:13) log( (cid:126)w(new)) − log( (cid:126)w(old))(cid:13)(cid:13)W(old)
(cid:105)
,
(4.22)
20
def= AT S−1W(new)S−1A. By the definition
(4.23)
(4.24)
(4.25)
Proof. Let H(old)
of δt and the triangle inequality we have
def= AT S−1W(old)S−1A and let H(new)
(new)
δt((cid:126)x, (cid:126)w(new)) =(cid:13)(cid:13)t(cid:126)c − AT S−1 (cid:126)w(new)(cid:13)(cid:13)H−1
≤(cid:13)(cid:13)t(cid:126)c − AT S−1 (cid:126)w(old)(cid:13)(cid:13)H−1
(cid:13)(cid:13)t(cid:126)c − AS−1 (cid:126)w(old)(cid:13)(cid:13)H−1
(cid:13)(cid:13)AT S−1 (cid:126)w(new) − AT S−1 (cid:126)w(old)(cid:13)(cid:13)H−1
(new)
Furthermore, since PAS−1( (cid:126)w(new)) (cid:22) I we have
By definition of ∞ and Lemma 33 H−1
(new) (cid:22) (1 + ∞)2H−1
Using that (ex−1)2
ex ≤ (1 + x)2x2 for x ≤ 1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:126)w(new) − (cid:126)w(old)
√
(cid:126)w(new) (cid:126)w(old)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)W(old)
+(cid:13)(cid:13)AT S−1 (cid:126)w(new) − AT S−1 (cid:126)w(old)(cid:13)(cid:13)H−1
(new)
(old) and therefore
≤ (1 + ∞)δt((cid:126)x, (cid:126)w(old)).
(new)
(new)
≤
−1/2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)W(old)
=(cid:13)(cid:13)W
(new)( (cid:126)w(new) − (cid:126)w(old))(cid:13)(cid:13)PAS−1 ( (cid:126)w(new))
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:126)w(new) − (cid:126)w(old)
2 and letting x =(cid:2)log( (cid:126)w(new)) − log( (cid:126)w(old))(cid:3)
≤ (1 + ∞)(cid:13)(cid:13) log( (cid:126)w(new)) − log( (cid:126)w(old))(cid:13)(cid:13)W(old)
√
(cid:126)w(new) (cid:126)w(old)
i we have
(4.26)
Combining (4.23), (4.24), (4.25), and (4.26) completes the proof.
(cid:126)x(new) = centeringExact((cid:126)x(old))
1. (cid:126)x(new) = (cid:126)x(old) − 1
(cid:126)h((cid:126)x(old), (cid:126)g((cid:126)s(old))).
1+cr
With this lemma we can now show how much centering progress we make by just updating (cid:126)x
and using the weight function. Note that in this proof we are just using the r-step in the proof, not
the algorithm itself. We will need to use the r-step itself only later when we drop the assumption
that we can compute (cid:126)g exactly.
Theorem 11 (Centering with Exact Weights). Fix a weight function (cid:126)g, let (cid:126)x(old) ∈ S0, and let
(cid:126)x(new) = centeringExact((cid:126)x(old))
If
then
def= δt((cid:126)x(old), (cid:126)g((cid:126)s(old))) ≤ 1
δt
80cγcr
(cid:18)
(cid:19)
δt((cid:126)x(new), (cid:126)g((cid:126)s(new))) ≤
1 − 1
4cr
δt((cid:126)x(old), (cid:126)g((cid:126)s(old))).
21
Proof. Let {(cid:126)x(new), (cid:126)w(new)} ∈ {S0×Rm
Note that this (cid:126)s(new) is the same as the (cid:126)s(new) in the theorem statement.
>0} be the result of a cr step from {(cid:126)x(old), (cid:126)w(old)} ∈ {S0×Rm
>0}.
Now by Lemma 6 we have that
Furthermore, defining (cid:126)∆ as in Lemma 9 and noting that (cid:126)w(old) = (cid:126)g((cid:126)s(old)) we have
δt((cid:126)s(new), (cid:126)w(new)) ≤ cγδ2
t
.
(cid:33)
(cid:32)
(cid:33)
(cid:32)
(cid:32)
(cid:126)∆ def= log
(cid:126)g((cid:126)s(new))
(cid:126)g((cid:126)s(old))
− log
(cid:126)w(new)
(cid:126)w(old)
= log
(cid:126)g((cid:126)s(new))
(cid:126)w(new)
(cid:13)(cid:13) log((cid:126)g((cid:126)s(new))/ (cid:126)w(new))(cid:13)(cid:13)∞ ≤ 4cγδt ≤ 1/2
(cid:13)(cid:13) log((cid:126)g((cid:126)s(new))/ (cid:126)w(new))(cid:13)(cid:13) (cid:126)w(new) ≤ ecr
δt + 13cγδ2
t
1 + cr
we see by Lemma 9 that
and
(4.27)
(4.28)
(4.29)
(cid:33)
.
(cid:21)
with = 0 because we are using exact weight computation. Applying Lemma 10 to (4.27), (4.28),
and (4.29) we have
(cid:20)
δt((cid:126)x(new), (cid:126)g((cid:126)s(new))) ≤ (1 + 4cγδt)
cγδ2
t +
cr
1 + cr
δt + 13cγδ2
t
(cid:18)
≤ cr
1 + cr
1 − 1
2cr
≤
δt + 20cγcrδ2
t
(cid:19)
(cid:18)
+
1
4cr
δt ≤
1 − 1
4cr
(cid:19)
δt
r ) iterations of CenteringExact
r ) then in Θ(c−1
)) and maintain δt((cid:126)x, (cid:126)g((cid:126)s)) = O(c−1
γ c−1
r ) using O(c−1
−1/2
γ c−2
r c
1
γ c−1
From this lemma we have that if δt((cid:126)x, (cid:126)g((cid:126)s)) is O(c−1
−1/2
γ c−1
r c
1
we can decrease δt((cid:126)x, (cid:126)g((cid:126)s)) by a multiplicative constant. Furthermore by Lemma 1 we see that we
can increase t by a multiplicative (1 + O(c−1
γ c−1
r ).
Thus we can double t and maintain δt((cid:126)x, (cid:126)g((cid:126)s)) = O(c−1
) iterations of
CenteringExact. In Section 7 we make this argument rigorously in the more general setting. In
the following sections, we show how to relax this requirement that (cid:126)g is computed exactly.
5 A Weight Function for (cid:101)O((cid:112)rank(A)L) Convergence
Section 4 yields an (cid:101)O((cid:112)rank(A)L) iteration interior point method. In Section 5.1 we motivate and
>0 that when used in the framework proposed in
Here, we present the weight function (cid:126)g : Rm
>0 → Rm
describe the weight function (cid:126)g, in Section 5.2 we prove that (cid:126)g satisfies Definition 7 with nearly opti-
mal c1((cid:126)g), cγ((cid:126)g), and cr((cid:126)g), and in Section 5.3 we show how to compute and correct approximations
to (cid:126)g efficiently.
22
barrier, φ((cid:126)s) = −(cid:80)
5.1 The Weight Function
Our weight function was inspired by the volumetric barrier methods of [40, 1].10 These papers
considered using the volumetric barrier, φ((cid:126)s) = − log det(AT S−2A) , in addition to the standard log
i∈[m] log si. In some sense the standard log barrier has a good slack sensitivity,
1, but a large size, m, and the volumetric barrier has a worse slack sensitivity,
m, but better total
weight, n. By carefully applying a weighted combination of these two barriers [40] and [1] achieved
an O((m rank(A))1/4L) iteration interior point method at the cost more expensive linear algebra in
each iteration.
√
Instead of using a fixed barrier, our weight function (cid:126)g : Rm
>0 is computed by solving
a convex optimization problem whose optimality conditions imply both good size and good slack
sensitivity. We define (cid:126)g for all (cid:126)s ∈ Rm
>0 → Rm
(cid:126)g((cid:126)s) def= arg min
(cid:126)w∈Rm
>0
f ((cid:126)s, (cid:126)w) where
>0 by
f ((cid:126)s, (cid:126)w) def= (cid:126)1T (cid:126)w − 1
α
log det(AT
s WαAs) − β
log wi
(5.1)
(cid:88)
i∈[m]
where here and in the remainder of this section we let As
chosen later such that the following hold
def= S−1A and the parameters α, β ∈ R are
α ∈ (0, 1)
,
β ∈ (0, 1)
, and
β1−α ≥ 1
2
.
(5.2)
To get a sense for why (cid:126)g has the desired properties, , suppose for illustration purposes that
>0. Using Lemma 34 and setting the gradient of (5.1) to (cid:126)0 we see
α = 1 and β = 0 and fix (cid:126)s ∈ Rm
that if (cid:126)g exists then
(cid:126)g((cid:126)s) = (cid:126)σAs((cid:126)g((cid:126)s)) def= diag
(G((cid:126)s))1/2As(AT
s G((cid:126)s)As)−1AT
s (G((cid:126)s))1/2(cid:17)
where we use the definition of (cid:126)σAs from Section 2. Consequently,
and
γ((cid:126)s, (cid:126)g((cid:126)s)) = 1 .
s (G((cid:126)s))1/2 is a projection matrix,(cid:13)(cid:13)(cid:126)σAs((cid:126)g((cid:126)s))(cid:13)(cid:13)1 =
Furthermore, since (G((cid:126)s))1/2As(AT
rank(A). Therefore, this would yield a weight function with good cγ and c1.
s G((cid:126)s)As)−1AT
Unfortunately picking α = 1 and β = 0 makes the optimization problem for computing (cid:126)g
degenerate. In particular for this choice of α and β, (cid:126)g((cid:126)s) could be undefined. In the follow sections
we will see that by picking better values for α and β we can trade off how well (cid:126)g performs as a
weight function and how difficult it is to compute approximations to (cid:126)g.
(cid:16)
(cid:13)(cid:13)PAs ((cid:126)g) = 1
(cid:13)(cid:13)G−1/2(cid:126)1i
max
i
5.2 Weight Function Properties
Here, we show that (cid:126)g : R≥0 → R≥0 as given by (5.1) is a weight function with respect to Definition 7
and we bound the values of c1((cid:126)g), cγ((cid:126)g), and cr((cid:126)g). The goal of this section is to prove the following.
Theorem 12 (Properties of Weight Function). Let us define α and β by
α = 1 −
log2
(cid:16) 2m
1
rank(A)
(cid:17)
and
β =
rank(A)
2m
For this choice of parameters (cid:126)g is a weight function meeting the criterion of Definition 7 with
10 See Section 1.2 for further intuition.
23
Next, using Lemma 36 and chain rule, we compute the following for all i, j ∈ [m],
∇ (cid:126)w
f ((cid:126)s, (cid:126)w) = (cid:126)1 − 1
α
ΣW−α(cid:0)αWα−1(cid:1) − βW−1(cid:126)1
= (cid:0)I − ΣW−1 − βW−1(cid:1) (cid:126)1
(cid:17) − ΣijIij + βIij
(cid:16)
∂(∇ (cid:126)w
f ((cid:126)s, (cid:126)w))i
∂wj
j
= − wiΛij (cid:126)w−α
− α
=
Σij
wiwj
Λij
wiwj
j
α (cid:126)wα−1
(cid:126)w2
i
βIi=j
wiwj
+
• Size : c1((cid:126)g) = 2 rank(A).
• Slack Sensitivity: cγ((cid:126)g) = 2.
• Step Consistency : cr((cid:126)g) = 2 log2
(cid:16) 2m
rank(A)
(cid:17)
.
We break the proof into several parts. In Lemma 13, we prove basic properties of f. In Lemma 14
we prove that (cid:126)g is a weight function and bound its size. In Lemma 15 we bound the slack sensitivity
of (cid:126)g and in Lemma 16 we show that (cid:126)g is consistent.
We start by computing the gradient and Hessian of f ((cid:126)s, (cid:126)w) with respect to (cid:126)w.
Lemma 13. For all (cid:126)s ∈ Rm
f ((cid:126)s, (cid:126)w) =(cid:0)I − ΣW−1 − βW−1(cid:1) (cid:126)1
>0 and (cid:126)w ∈ Rm
∇ (cid:126)w
>0, we have
and ∇2
(cid:126)w (cid:126)w
f ((cid:126)s, (cid:126)w) = W−1 (Σ + βI − αΛ) W−1
where Σ def= ΣAs(Wα(cid:126)1) and Λ def= ΛAs(Wα(cid:126)1).
Proof. Using Lemma 34 and the chain rule we compute the gradient of ∇w f ((cid:126)s, (cid:126)w) as follows
.
(Using that Σis diagonal)
Consequently ∇2
(cid:126)w (cid:126)w
f ((cid:126)s, (cid:126)w) = W−1 (Σ + βI − αΛ) W−1 as desired.
Using this lemma, we prove that (cid:126)g is a weight function with good size.
Lemma 14. The function (cid:126)g is a weight function meeting the criterion of Definition 7. For all
(cid:126)s ∈ Rm
>0 and i ∈ [m] we have
β ≤ gi((cid:126)s) ≤ 1 + β
and
(cid:13)(cid:13)(cid:126)g((cid:126)s)(cid:13)(cid:13)1 = rank(A) + β · m.
Furthermore, for all (cid:126)s ∈ Rm
>0, the weight function obeys the following equations
G((cid:126)s) = (Σg + βI) (cid:126)1
and G(cid:48)((cid:126)s) = −2G((cid:126)s) (G((cid:126)s) − αΛg)
−1 ΛgS−1
where Σg
def= ΣAs(G((cid:126)s)α(cid:126)1), Λg
def= ΛAs(G((cid:126)s)α(cid:126)1), and G(cid:48)((cid:126)s) is the Jacobian matrix of (cid:126)g at (cid:126)s.
24
Proof. By Lemma 32 and (5.2) we have that for all (cid:126)w, (cid:126)s ∈ Rm
>0 ,
ΣAs( (cid:126)w) (cid:23) ΛAs( (cid:126)w) (cid:23) αΛAs( (cid:126)w).
f ((cid:126)s, (cid:126)w) (cid:23) βW−2 and f ((cid:126)s, (cid:126)w) is convex for (cid:126)w, (cid:126)s ∈ Rm
>0. Using
Therefore, by Lemma 13, ∇2
Lemma 13, we see that that for all i ∈ [m] it is the case that
(cid:126)w (cid:126)w
(cid:104)∇ (cid:126)w
(cid:105)
f ((cid:126)s, (cid:126)w)
=
i
1
wi
(wi − σi − β)
i
(cid:105)
f ((cid:126)s, (cid:126)w)
(cid:104)∇ (cid:126)w
is strictly negative and if (cid:126)wi ∈ (1 + β,∞) then
Since 0 ≤ σi ≤ 1 for all i by Lemma 32 and β ∈ (0, 1) by (5.2), we see that if (cid:126)wi ∈ (0, β) then
is strictly positive.
>0 , the (cid:126)w that minimizes this convex function f ((cid:126)s, (cid:126)w) lies in the box
(cid:104)∇ (cid:126)w
that(cid:13)(cid:13)(cid:126)σ(cid:13)(cid:13)1 = tr (PAs(G((cid:126)s)α)). Since PAs(G((cid:126)s)α(cid:126)1) is a projection onto the image of G((cid:126)s)α/2As and
Therefore, for any (cid:126)s ∈ Rm
between β and 1 + β. Since f is strongly convex in this region, the minimizer is unique.
The formula for G((cid:126)s) follows by setting ∇ (cid:126)w
f ((cid:126)s, (cid:126)w) = (cid:126)0 and the size of (cid:126)g follows from the fact
since (cid:126)g((cid:126)s) > 0 and (cid:126)s > 0, the dimension of the image of G((cid:126)s)α/2As is the rank of A. Hence, we
have that
(cid:13)(cid:13)(cid:126)g((cid:126)s)(cid:13)(cid:13)1 ≤(cid:13)(cid:13)(cid:126)σ(cid:13)(cid:13)1 +(cid:13)(cid:13)β(cid:126)1(cid:13)(cid:13)1 = rank(A) + β · m.
f ((cid:126)s, (cid:126)w)
(cid:105)
i
To compute G(cid:48)((cid:126)s), we note that for (cid:126)w ∈ Rm
rule, we get the following for all i, j ∈ [m],
>0 and Λw
def= ΛWαA(S−2(cid:126)1), by Lemma 36 and chain
∂(∇ (cid:126)w
f ((cid:126)s, (cid:126)w))i
∂sj
= −w−1
i Λijs2
j
(cid:16)−2s−3
j
(cid:17)
= 2w−1
i Λijs−1
j
.
f ((cid:126)s, (cid:126)w) with respect to (cid:126)s. Since we have already shown that J (cid:126)w(∇ (cid:126)w
Consequently, J(cid:126)s(∇ (cid:126)w
f ((cid:126)s, (cid:126)w)) = 2W−1ΛwS−1 where J(cid:126)s denotes the Jacobian matrix of the function
∇ (cid:126)w
(cid:126)w (cid:126)wft((cid:126)s, (cid:126)w) =
W−1 (Σw + βI − αΛw) W−1 is positive definite (and hence invertible), by applying the implicit
function theorem to the specification of (cid:126)g((cid:126)s) as the solution to ∇ (cid:126)w
f ((cid:126)s, (cid:126)w) = (cid:126)0, we have
f ((cid:126)s, (cid:126)w)) = ∇2
G(cid:48)((cid:126)s) = −(cid:16)
(cid:17)−1(cid:16)
(cid:17)
J (cid:126)w(∇ (cid:126)w
f ((cid:126)s, (cid:126)w))
J(cid:126)s(∇ (cid:126)w
f ((cid:126)s, (cid:126)w))
= −2G((cid:126)s) (G((cid:126)s) − αΛg)
−1 ΛgS−1
Using Lemma 14 we now show that (cid:126)g has a good slack sensitivity.
Lemma 15 (Weight Function Slack Sensitivity). For all (cid:126)s ∈ Rm
Proof. Fix an arbitrary (cid:126)s ∈ Rm
know that (cid:126)g = (Σ + βI) (cid:126)1 and β ≤ gi ≤ 1 + β ≤ 2 for all i ∈ [m]. Furthermore, since β1−α ≥ 1
α ∈ (0, 1) by (5.1) and clearly G = G1−αGα we have
>0 and let (cid:126)g def= (cid:126)g((cid:126)s), and Σ def= ΣAs((cid:126)gα). Recall that by Lemma 14 we
2 and
>0, we have γ((cid:126)s, (cid:126)g((cid:126)s)) ≤ 2.
Gα (cid:22) β1−αGα (cid:22) G (cid:22) (2)1−αGα (cid:22) 2Gα
1
2
Applying this and using the definition of PAs((cid:126)g) yields
s GαAs)−1AT
s GAs)−1AT
s (cid:22) 2As(AT
As(AT
s = 2G−α/2PAs((cid:126)gα)G−α/2
(5.3)
.
(5.4)
25
Hence, by definition of the weight slack sensitivity we have
γ((cid:126)s, (cid:126)g) = max
i
i
= max
≤ max
i
= max
i
(cid:13)(cid:13)G−1/2(cid:126)1i
(cid:113)
(cid:113)
(cid:114)
2(cid:126)1T
(cid:13)(cid:13)PAs ((cid:126)g)
(cid:114) σi
s
(cid:126)1i
s GAs)−1AT
(cid:126)1T
i As(AT
i G−α/2PAs((cid:126)gα)G−α/2(cid:126)1i
σi
gα
i
≤ 2 max
≤ 2
2
i
gi
2 and gi ≥ σi.
where the last line due to the fact g1−α
i
≥ β1−α ≥ 1
Finally, we bound the step consistency of (cid:126)g.
Lemma 16 (Weight Function Step Consistency). For all (cid:126)s ∈ Rm
>0, (cid:126)y ∈ Rm, r ≥ 2
1−α, and
we have
(cid:107)B(cid:126)y(cid:107)G((cid:126)s) ≤(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)G((cid:126)s)
B def= I +
1
r
G((cid:126)s)−1G(cid:48)((cid:126)s)S,
(cid:107)B(cid:126)y(cid:107)∞ ≤(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)∞ +
(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)G((cid:126)s).
2
1 − α
and
Proof. Fix an arbitrary (cid:126)s ∈ Rm
Λ def= ΛAs((cid:126)gα). Also, fix an arbitrary (cid:126)y ∈ Rm and let (cid:126)z def= B(cid:126)y.
−1 ΛS−1 and therefore
By Lemma 14, G(cid:48) = −2G (G − αΛ)
>0 and let (cid:126)g def= (cid:126)g((cid:126)s), (cid:126)σ def= (cid:126)σAs((cid:126)gα), Σ def= ΣAs((cid:126)gα), P def= PAs((cid:126)gα),
B = I + r−1G−1(cid:16)−2G (G − αΛ)
−1(cid:0)G − (α + 2r−1)Λ(cid:1)
= (G − αΛ)
= (G − αΛ)
S
−1 (G − αΛ) − 2r−1 (G − αΛ)
−1 Λ
−1 ΛS−1(cid:17)
.
By Lemma 14, we have G (cid:23) Σ. By the definition of Λ = Σ − P(2), we have Σ (cid:23) Λ and Lemma 32
shows that Λ (cid:23) 0. Hence, we have
0 (cid:22) Λ (cid:22) Σ ≺ G.
Using this and 0 < 2r−1 ≤ 1 − α, we have that
0 ≺ G − (α + 2r−1)Λ (cid:22) G − αΛ .
Thus, G − αΛ is positive definite and therefore (cid:126)z is the unique vector such that
(G − αΛ) (cid:126)z =(cid:0)G − (α + 2r−1)Λ(cid:1) (cid:126)y
To bound(cid:13)(cid:13)(cid:126)z(cid:13)(cid:13)G, we note that since G (cid:31) 0 we have
(cid:16)
I − αG−1/2ΛG−1/2(cid:17)
G1/2(cid:126)z =
(cid:16)
I − (α + 2r−1)G−1/2ΛG−1/2(cid:17)
(5.5)
G1/2(cid:126)y
Furthermore, since 0 (cid:22) G−1/2ΛG−1/2 (cid:22) I, we have that
0 (cid:22) I − (α + 2r−1)G−1/2ΛG−1/2 (cid:22) I − αG−1/2ΛG−1/2
26
(5.6)
for arbitrary i ∈ [m] and using that (cid:126)σi ≥ 0 then yields that
−2
−2
(1 − α) Σ(cid:126)z + β(cid:126)z + αP(2)(cid:126)z = G(cid:126)z − αΛ(cid:126)z
G1/2(cid:126)z(cid:13)(cid:13)(I−αG−1/2ΛG−1/2)
G1/2(cid:126)z(cid:13)(cid:13)(I−(α+2k−1)G−1/2ΛG−1/2)
I − αG−1/2ΛG−1/2(cid:17)
and consequently(cid:13)(cid:13)(cid:126)z(cid:13)(cid:13)G =(cid:13)(cid:13)(cid:16)
I − αG−1/2ΛG−1/2(cid:17)
≤(cid:13)(cid:13)(cid:16)
=(cid:13)(cid:13)G1/2(cid:126)y(cid:13)(cid:13) =(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)G .
Therefore,(cid:13)(cid:13)B(cid:126)y(cid:13)(cid:13)G ≤(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)G as desired.
Next, to bound(cid:13)(cid:13)(cid:126)z(cid:13)(cid:13)∞, we use that Λ = Σ − P(2) and (cid:126)g = (cid:126)σ + β(cid:126)1 and (5.5) to derive
= (cid:0)G − (α + 2r−1)Λ(cid:1) (cid:126)y
= (cid:0)1 − α − 2r−1(cid:1) Σ(cid:126)y + β(cid:126)y +(cid:0)α + 2r−1(cid:1) P(2)(cid:126)y
((1 − α)(cid:126)σi + β)(cid:126)zi ≤(cid:12)(cid:12)(cid:12)α(cid:126)1T
(cid:12)(cid:12)(cid:12) +
(cid:12)(cid:12)(cid:12)(cid:0)(1 − α − 2r−1)(cid:126)σi + β(cid:1) (cid:126)yi +(cid:0)α + 2r−1(cid:1) (cid:126)1T
(cid:12)(cid:12)(cid:12)[P(2)(cid:126)z]i
(cid:12)(cid:12)(cid:12) + ((1 − α)(cid:126)σi + β)(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)∞ +
(cid:12)(cid:12)(cid:12)[P(2)(cid:126)y]i
(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)Σ
(cid:13)(cid:13)(cid:126)z(cid:13)(cid:13)Σ + ((1 − α)(cid:126)σi + β)(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)∞ + (cid:126)σi
≤ ((1 − α)(cid:126)σi + β)(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)∞ + (1 + α)(cid:126)σi
(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)G5.6
(cid:126)zi ≤ (cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)∞ +
(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)G
≤ (cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)∞ +
and therefore(cid:13)(cid:13)B(cid:126)y(cid:13)(cid:13)∞ =(cid:13)(cid:13)(cid:126)z(cid:13)(cid:13)∞ ≤(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)∞ + 2
(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)G.
toLeft multiplying this equation by (cid:126)1T
i
(1 + α) (cid:126)σi
(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)G
((1 − α)(cid:126)σi + β)
1 − α
≤ α
≤ α(cid:126)σi
Consequently,
i P(2)(cid:126)y
i P(2)(cid:126)z
.
(cid:12)(cid:12)(cid:12)
1−α
(cid:12)(cid:12)(cid:12)
2
(0 < 2r−1 ≤ (1 − α) < 1)
(Lemma 32)
(Σ (cid:22) Gand)
From Lemmas 14, 15 and 16, the proof of Theorem 12 is immediate. Since m ≥ rank(A) we
have log2 (2m/ rank(A)) ≥ 1 and α ∈ (0, 1). Furthermore β ∈ (0, 1) and
(cid:18) rank(A)
(cid:19)(cid:16)
2m
β1−α =
1
log2(2m/ rank(A))
(cid:17)
>0 we have (cid:13)(cid:13)(cid:126)g((cid:126)s)(cid:13)(cid:13)1 ≤ 2 · rank(A) by
1
2
=
and therefore (5.2) is satisfied. Furthermore, for all (cid:126)s ∈ Rm
Lemma 14. The bounds on cγ((cid:126)g) and cr((cid:126)g) then follow from Lemma 15 and Lemma 16 respectively.
5.3 Computing and Correcting The Weights
Here, we describe how to efficiently compute approximations to the weight function (cid:126)g : Rm≥0 → Rm≥0
as given by (5.1). The two main technical tools we use towards this end are the gradient descent
method, Theorem 17, a standard result in convex optimization, and fast numerical methods for
27
estimating leverage scores using the Johnson-Lindenstrauss Lemma, Theorem 20, a powerful tool
in randomized numerical linear algebra.
compute (cid:126)g((cid:126)s) to high accuracy in (cid:101)O(1) iterations. Unfortunately, this result makes two assumptions
Since the weight function, (cid:126)g, is defined as the minimizer of a convex optimization problem (5.1),
we could use the gradient descent method directly to minimize f and hence compute (cid:126)g. Indeed, in
Lemma 19 we show how applying the gradient descent method in a carefully scaled space allows us to
to compute (cid:126)g((cid:126)s): (1) we are given a weight (cid:126)w ∈ R≥0 that is not too far from (cid:126)g((cid:126)s) and (2) we compute
the gradient of f exactly.
Assumption (1) is not an issue as we always ensure that (cid:126)g does not change too much between
calls to compute (cid:126)g and therefore can always use our previous weights as the approximation to (cid:126)g((cid:126)s).
However, naively computing the gradient of f is computationally expensive and hence assumption
(2) is problematic. To deal with this issue we use the fact that by careful application of Johnson-
Lindenstrauss one can compute a multiplicative approximation to the gradient efficiently and in
Theorem 21 we show that this suffices to compute an approximation to (cid:126)g that suffices to use in our
weighted path following scheme.
First we prove the theorem regarding gradient descent method we use in our analysis. This the-
orem shows that if we take repeated projected gradient steps then we can achieve linear convergence
up to bounds on how much the hessian of the function changes over the domain of interest. 11
Theorem 17 (Simple Constrained Minimization for Twice Differentiable Function [25]). Let H be
a positive definite matrix and Q ⊆ Rm be a convex set. Let f ((cid:126)x) : Q → R be a twice differentiable
function such that there are constants L ≥ µ ≥ 0 such that for all (cid:126)x ∈ Q we have µH (cid:22) ∇2f ((cid:126)x) (cid:22)
LH. If for some (cid:126)x(0) ∈ Q and all k ≥ 0 we apply the update rule
(cid:126)x(k+1) = arg min
(cid:126)x∈Q
(cid:68)∇f ((cid:126)x(k)), (cid:126)x − (cid:126)x(k)(cid:69)
H ≤(cid:16)
(cid:13)(cid:13)(cid:126)x(k) − (cid:126)x∗(cid:13)(cid:13)2
+
L
2
(cid:13)(cid:13)(cid:126)x − (cid:126)x(k)(cid:13)(cid:13)2
H
then for all k ≥ 0 we have
(cid:17)k(cid:13)(cid:13)(cid:126)x(0) − (cid:126)x∗(cid:13)(cid:13)2
Lemma 18 (Hessian Approximation). For(cid:13)(cid:13)W−1((cid:126)g((cid:126)s) − (cid:126)w)(cid:13)(cid:13)∞ ≤ 1
1 − µ
L
the optimal point (cid:126)g((cid:126)s) such that the Hessian of f does not change too much.
12 we have
H.
To apply this Theorem 17 to compute (cid:126)g((cid:126)s) we first need to show that there is a region around
2(1 − α)
3
W−1 (cid:22) ∇2
(cid:126)w (cid:126)w
f ((cid:126)s, (cid:126)w) (cid:22) 3
2
W−1.
Proof. From Lemma 13, we know that
f ((cid:126)s, (cid:126)w) = W−1 (Σ + βI − αΛ) W−1
where Σ = ΣAs( (cid:126)wα) and Λ = ΛAs( (cid:126)wα). Using 0 (cid:22) Λ (cid:22) Σ, we have
∇2
(cid:126)w (cid:126)w
(1 − α)W−1 (Σ + βI) W−1 (cid:22) ∇2
ww
f ((cid:126)s, (cid:126)w) (cid:22) W−1 (Σ + βI) W−1
11Note that this theorem is usually stated with H = I, i.e. the standard Euclidean norm rather than the one
induced by H. However, Theorem 17 can be proved by these standard results just by a change of variables.
28
Using that(cid:13)(cid:13)W−1((cid:126)g((cid:126)s) − (cid:126)w)(cid:13)(cid:13)∞ ≤ 1
12 and applying Lemma 14 we have
(cid:18)
(cid:18)
(cid:19)−2
(cid:19)2
Σ + βI (cid:22)
1 − 1
12
Σ + βI (cid:23)
1 − 1
12
ΣAs((cid:126)gα) + βI (cid:22)
ΣAs((cid:126)gα) + βI (cid:23)
(cid:18)
(cid:18)
(cid:19)−2
(cid:19)2
1 − 1
12
G (cid:22) 3
2
W
1 − 1
12
G (cid:23) 2
3
W.
and
Combining Theorem 17 and Lemma 18, we get the following algorithm to compute the weight
function using the exact computation of the gradient of f. Note that this algorithm applies Theorem
17 multiple times as in each iteration we are taking a gradient step with respect to a different norm.
Lemma 19 (Exact Weight Computation). Given (cid:126)w(0) ∈ Rm
≤
1−α
24 . Let
(cid:13)(cid:13)(cid:13)W−1
(0)((cid:126)g((cid:126)s) − (cid:126)w(0))
(cid:13)(cid:13)(cid:13)∞
.
(cid:27)
(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)2
>0 such that
(cid:13)(cid:13)(cid:13)∞
≤ 1 − α
(0)( (cid:126)w − (cid:126)w(0))
(cid:126)w(k)(cid:17)α(cid:17)
(cid:16)(cid:16)
+ β
24
Q =
(cid:126)w∈Q
(cid:126)w(k+1) = arg min
For all k ≥ 0 let
(cid:26)
(cid:126)w ∈ Rm (cid:13)(cid:13)(cid:13)W−1
(cid:13)(cid:13)(cid:13)(cid:13) (cid:126)w − 1
(cid:16)
This implies that for all k ,(cid:13)(cid:13)(cid:13)G((cid:126)s)−1((cid:126)g((cid:126)s) − (cid:126)w(k))
(cid:13)(cid:13)(cid:13)2
(cid:126)w(k)(cid:17)α(cid:17)
(cid:16)(cid:16)
(cid:126)w(k)(cid:17)α(cid:17)
(cid:16)(cid:16)
Proof. Note that iterations of Theorem 17 can be rewritten as
(k) − βW−1
I − ΣAs
(cid:16)
(cid:126)w(k+1) = arg min
(cid:126)w(k) + (cid:126)σAs
(cid:68)(cid:16)
(cid:13)(cid:13)(cid:13)(cid:13) (cid:126)w − 1
(cid:126)w(k) + (cid:126)σAs
= arg min
≤ 4m2
W−1
(cid:126)w∈Q
+ β
12
(cid:18)
(cid:19)k
1 − 1 − α
(cid:17) (cid:126)1, (cid:126)w − (cid:126)w(k)(cid:69)
(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)2
W−1
(cid:126)w∈Q
∞
(k)
2
2
.
(k)
W−1
(k)
(cid:13)(cid:13)(cid:13) (cid:126)w − (cid:126)w(k)(cid:13)(cid:13)(cid:13)2
W−1
(k)
+
where the last line simply comes for expanding the quadratic function and ignoring the constant
term. Hence, we see that the iteration on (cid:126)w(k+1) is in fact a gradient descent step. To apply
Theorem 17 we note that for any (cid:126)w ∈ Q the definition of Q and the fact that α ∈ (0, 1) implies that
(1 − 1
24 )W(0) (cid:22) W (cid:22) (1 + 1
24 )W(0). Therefore Lemma 18 shows that for all (cid:126)w(k) ∈ Q,
(k) (cid:22) 2(1 − α)
W−1
(cid:13)(cid:13)(cid:13) (cid:126)w(k+1) − (cid:126)g((cid:126)s)
(cid:13)(cid:13)(cid:13)2
(0) (cid:22) ∇2
(cid:18)
1 − 1 − α
(cid:19)(cid:13)(cid:13)(cid:13) (cid:126)w(k) − (cid:126)g((cid:126)s)
where the left most and right most inequality comes from the fact they lies in a small region Q.
Hence, Theorem 17 and inequality (5.7) shows that
f ((cid:126)s, (cid:126)w) (cid:22) 3
2
(0) (cid:22) 2W−1
(k).
(cid:13)(cid:13)(cid:13)2
1 − α
W−1
W−1
(5.7)
≤
(cid:126)w (cid:126)w
2
3
.
W−1
(k)
W−1
(k)
4
29
Since
we have
(0)((cid:126)g((cid:126)s) − (cid:126)w(0))
(cid:13)(cid:13)(cid:13)∞
(cid:13)(cid:13)(cid:13)W−1
(cid:13)(cid:13) (cid:126)w(k) − (cid:126)g((cid:126)s)(cid:13)(cid:13)2
≤ 1−α
G−1((cid:126)s) ≤
≤
(cid:1)2 W(k). Hence,
G−1((cid:126)s)
24
24
1 − 1 − α
24 and (cid:126)w(k) ∈ Q, we know that G((cid:126)s) (cid:23)(cid:0)1 − 1−α
(cid:19)(cid:13)(cid:13) (cid:126)w(k−1) − (cid:126)g((cid:126)s)(cid:13)(cid:13)2
(cid:18)
1 − 1 − α
(cid:18)
1 − 1 − α
(cid:18)
1 − 1 − α
(cid:13)(cid:13)G−1((cid:126)s)((cid:126)g((cid:126)s) − (cid:126)w(0))(cid:13)(cid:13)2
(cid:19)−2(cid:18)
(cid:19)(cid:13)(cid:13) (cid:126)w(k−1) − (cid:126)g((cid:126)s)(cid:13)(cid:13)2
(cid:19)k(cid:13)(cid:13) (cid:126)w(0) − (cid:126)g((cid:126)s)(cid:13)(cid:13)2
(cid:1)2
(cid:0)1 − 1−α
∞ ≤ m(1 + β)
G−1((cid:126)s)
G−1((cid:126)s)
12
12
4
≤
∞ ≤ β−1(cid:13)(cid:13) (cid:126)w(k) − (cid:126)g((cid:126)s)(cid:13)(cid:13)2
24
The result follows from the facts that
G−1((cid:126)s) ≤ m(cid:13)(cid:13)G((cid:126)s)(cid:13)(cid:13)∞
(cid:13)(cid:13) (cid:126)w(0) − (cid:126)g((cid:126)s)(cid:13)(cid:13)2
and lemma 14 that(cid:13)(cid:13)G−1((cid:126)s)( (cid:126)w(k) − (cid:126)g((cid:126)s))(cid:13)(cid:13)2
(0)((cid:126)g((cid:126)s) − (cid:126)w(0))(cid:13)(cid:13)2
(cid:13)(cid:13)W−1
∞
G−1((cid:126)s) where β = rank(A)
2m .
Unfortunately, we cannot use the previous lemma directly as computing (cid:126)σAs exactly is too
expensive for our purposes. However, in [32] they showed that we can compute leverage scores, (cid:126)σAs,
approximately by solving only polylogarithmically many regression problems (See [20] for more
details). These results use the fact that the leverage scores of the the ith constraint, i.e. [(cid:126)σAs]i is
the (cid:96)2 length of vector PA((cid:126)x)(cid:126)1i and that by the Johnson-Lindenstrauss lemma these lengths are
persevered up to multiplicative error if we project these vectors onto certain random low dimensional
subspace. Consequently, to approximate the (cid:126)σAs we first compute the projected vectors and then
use it to approximate (cid:126)σAs and hence only need to solve (cid:101)O(1) regression problems. For completeness,
we provide the algorithm and theorem here:
(cid:126)σ(apx) = computeLeverageScores(A, (cid:126)x, )
1. Let k =(cid:6)24 log(m)/2(cid:7).
(cid:16)
=(cid:80)k
2. Let (cid:126)q(j) be k random ±1/
3. Compute (cid:126)p(j) = X1/2A(AT XA)−1AT X1/2(cid:126)q(j).
4. Return (cid:126)σ(apx)
(cid:17)2
k vectors of length m.
√
.
(cid:126)p(j)
i
j=1
i
Theorem 20 ([32]). For 0 < < 1 with probability at least 1− 1
returns (cid:126)σ(apx) such that for all i ∈ [m] ,
m the algorithm computeLeverageScores
(1 − ) (cid:126)σA((cid:126)x)i ≤ (cid:126)σ(apx)
i
≤ (1 + ) (cid:126)σA((cid:126)x)i.
by solving only O(−2 · log m) linear systems.
Now, we show that we can modify Lemma 19 to use computeLeverageScores and we prove that
this still provides adequate error guarantees. Our weight computation and the analysis is as follows.
30
rank(A)
2. Q =
(cid:126)w = computeWeight((cid:126)s, (cid:126)w(0), K)
1. Let cr = 2 log2
(cid:16) 2m
(cid:126)w ∈ Rm (cid:13)(cid:13)(cid:13)W−1
(cid:110)
(cid:13)(cid:13)(cid:13) (cid:126)w − 1
(cid:16) 2m
(cid:17), β = rank(A)
(cid:13)(cid:13)(cid:13)∞
(cid:111)
3. For j = 1 to k where k = (cid:100)12cr log(cid:0) 4m
(cid:1)(cid:101)
≤ 1
(cid:16)
S−1A,(cid:0) (cid:126)w(j)(cid:1)α
(cid:17)(cid:13)(cid:13)(cid:13)2
(cid:17)
, α = 1 −
(0)( (cid:126)w − (cid:126)w(0))
(cid:16)
3a. (cid:126)σ(j) = computeLeverageScores
(cid:126)w(j−1) + (cid:126)σ(j) + β(cid:126)1
(cid:17)
rank(A)
log2
12cr
,
K
2
1
W−1
(j−1)
3b. (cid:126)w(j) = arg min (cid:126)w∈Q
4. Output (cid:126)w(j).
2m , =
K
48cr log( 2m
K )
.
Note that the convex set Q is aligned with standard basis and hence the step 3b can be computed
by explicit formula (5.9).
Theorem 21 (Approximate Weight Computation). Let (cid:126)s ∈ Rm
and K ∈ (0, 1). The algorithm computeWeight((cid:126)s, (cid:126)w(0), K) returns (cid:126)w such that
>0,(cid:13)(cid:13)W−1
(0)((cid:126)g((cid:126)s) − (cid:126)w(0))(cid:13)(cid:13)∞ ≤ 1
12cr
12,
(cid:13)(cid:13)G((cid:126)s)−1((cid:126)g((cid:126)s) − (cid:126)w)(cid:13)(cid:13)∞ ≤ K
with probability(cid:0)1 − 1
(cid:1)(cid:100)12cr log( 4m
K )(cid:101).
m
systems.
The running time is dominated by the time needed to solve O(c3
r log3(m/K) log(m)/K2) linear
Proof. Consider an execution of computeWeight((cid:126)s, (cid:126)w(0), K) where each computeLeverageScores
computes (cid:126)σAs
during this idealized execution of computeWeight.
(cid:0)( (cid:126)w(j))α(cid:1), and let (cid:126)v(j)denote the (cid:126)w(j) computed
(cid:0)( (cid:126)w(j))α(cid:1) exactly, i.e. (cid:126)σ(j) = (cid:126)σAs
Now suppose that for all i ∈ [m] we have
(1 − )M(cid:126)v(j)
(cid:126)w(j+1) = median
(5.8)
for some M ≥ 0 and j ∈ [k − 1]. Since the objective function and the constraints for step 3b. are
axis-aligned we can compute (cid:126)w(j) coordinate-wise and we see that
i
i ≤ (cid:126)w(j)
i ≤ (1 + )M(cid:126)v(j)
(cid:16)
(cid:18)
(cid:17)
(cid:126)w(j)(cid:17)α(cid:17)
(cid:16)(cid:16)
(cid:16)(cid:0) (cid:126)w(j+1)(cid:1)α(cid:17)
+ β
,
(cid:126)w(0), (cid:126)w(j) +
(cid:126)σAs
1
2
(cid:18)(cid:18)
1 − 1
12cr
(cid:19)
(cid:16)(cid:0) (cid:126)w(j+1)(cid:1)α(cid:17)
(cid:19)
(cid:19)
1 +
1
12cr
(cid:126)w(0)
(5.9)
where [median ((cid:126)x, (cid:126)y, (cid:126)z)]i is equal to the median of xi, yi and zi for all i ∈ [m]. By (5.8), (5.9), and
the fact that (1 − ) σAs
for all i ∈ [m], we have
that
≤ (1 + ) σAs
≤ (cid:126)σ(j+1)
i
i
i
(1 − )M +1(cid:126)v(j+1)
≤ (cid:126)w(j+1)
≤ (1 + )M +1(cid:126)v(j+1)
.
i
i
i
Since (cid:126)v(0) = (cid:126)w(0) and since j ∈ [k − 1] was arbitrary we can apply induction and we have that for
all j ∈ [k]
12Recall that cr = 2
1−α = 2 log
rank(A)
(1 − )j(cid:126)v(j)
i ≤ (cid:126)w(j)
i ≤ (1 + )j(cid:126)v(j)
i
.
(cid:16) 2m
(cid:17) ≥ 2.
31
1 − 1
Note that k ≤ 1
8 and therefore by Taylor series expansion we have(cid:13)(cid:13)V−1
(cid:13)(cid:13)G((cid:126)s)−1((cid:126)g((cid:126)s) − (cid:126)w(k))(cid:13)(cid:13)∞ ≤(cid:13)(cid:13)G((cid:126)s)−1((cid:126)g((cid:126)s) − (cid:126)v(k))(cid:13)(cid:13)∞ +(cid:13)(cid:13)G((cid:126)s)−1(cid:16)
(cid:19)−2(cid:13)(cid:13)V−1
Furthermore since (cid:126)v(k) ∈ Q we know that G((cid:126)s) (cid:23)(cid:16)
(cid:17)2
Lemma 19, and recalling that k = (cid:100)12cr log(cid:0) 4m
(cid:1)(cid:101) we have
(cid:18)
(cid:19) k
(cid:19)
(cid:18) 4m
1 − 1
6cr
− k
12cr
+ 1.5(cid:100)12cr log
1 − 1
12cr
≤ 2m · exp
(cid:101) ≤ K
≤ 2m
+ 1.5k
(cid:18)
(cid:18)
(cid:19)
12cr
+
K
2
≤ K
2
K
8 k.
V(k). Putting these together, applying
(k)
(cid:0) (cid:126)w(k) − (cid:126)v(k)(cid:1)(cid:13)(cid:13)∞ ≤ 9
(cid:126)v(k) − (cid:126)w(k)(cid:17)(cid:13)(cid:13)∞
(k)((cid:126)v(k) − (cid:126)w(k))(cid:13)(cid:13)∞
(cid:16)√
(cid:17)
Finally, we show how to compute an initial weight without having an approximate weight to
help the computation. The algorithm computeInitialWeight((cid:126)s, K) computes an initial weight in
iterations of computeWeight by computing (cid:126)g for a large enough value of β and then
O
decreasing β gradually.
rank A
(cid:16) 2m
(cid:17)
, α = 1 −
(cid:126)w = computeInitialWeight((cid:126)s, K)
1. Let cr = 2 log2
rank(A)
2. Loop until β = rank(A)
2a. (cid:126)w = computeWeight((cid:126)s, (cid:126)w,
√
2b. β = max
rank(A)
3. Output computeWeight((cid:126)s, (cid:126)w, K).
(cid:26)(cid:18)
(1−α)3/2
1 −
1000c2
r
(cid:19)
50cr
).
2m
1
(cid:16) 2m
1
log2
rank(A)
(cid:17), β = 12cr and (cid:126)w = β(cid:126)1.
(cid:27)
β, rank(A)
2m
.
Theorem 22 (Computating Initial Weights). For (cid:126)s ∈ Rm
the algorithm computeInitialWeight((cid:126)s, K) returns (cid:126)w ∈ Rm
>0 and K > 0, with constant probability
>0 such that
(cid:13)(cid:13)G((cid:126)s)−1((cid:126)g((cid:126)s) − (cid:126)w)(cid:13)(cid:13)∞ ≤ K.
O((cid:112)rank (A) log(1/K)/K2) linear systems.
The total running time of computeInitialWeight((cid:126)s, K) is dominated by the time needed to solve
Proof. Fix (cid:126)s ∈ Rm
>0 and let As
def= S−1A. For all β > 0 let (cid:126)g : R>0 → Rm be defined by13
(cid:126)g(β) def= arg min
(cid:126)w∈Rm
>0
(cid:126)1T (cid:126)w − 1
α
log det(AT
s WαAs) − β
log wi
(cid:88)
i∈[m]
The algorithm computeInitialWeight((cid:126)s, K) maintains the invariant that before step 2a
(cid:13)(cid:13)W−1((cid:126)g(β) − (cid:126)w)(cid:13)(cid:13)∞ ≤ 1
12cr
.
(5.10)
13Note that early we assumed that β < 1 and here we use much larger values of β. However, this bound on β was
primarily to assist in bounding c1 and does not affect this proof.
32
Since (cid:126)g(β) = (cid:126)σ(β) + β where (cid:126)σ(β) def= (cid:126)σAs((cid:126)gα(β)), we have that for all i ∈ [m]
β ≤ g(β)i ≤ 1 + β.
Therefore, in the step 1, the initial weight, (cid:126)w = β(cid:126)1 ∈ Rm
2a, by Theorem 21 we have
>0 satisfies the invariant (5.10). After step
.
(5.11)
Therefore, it suffices to prove that (cid:126)g(β) is close to (cid:126)g(β − θ) for small θ.
To bound how much (cid:126)g(β) changes for small changes in β we proceed similarly to Lemma 14.
First by the implicit function theorem and direct calculation we know that
J (cid:126)w(∇ (cid:126)w
f ((cid:126)s, (cid:126)w))
= G(β) (G(β) − αΛg)
−1 (cid:126)1
(5.12)
= −(cid:16)
d(cid:126)g
dβ
(cid:13)(cid:13)G(β)−1((cid:126)g(β) − (cid:126)w)(cid:13)(cid:13)∞ ≤ 1
(cid:17)
(cid:17)−1(cid:16)
Jβ(∇ (cid:126)w
f ((cid:126)s, (cid:126)w))
50cr
where Λg
(5.12) in a similar manner to Lemma 16. Note that
def= ΛAs(G(β)α(cid:126)1). Next to estimate how fast (cid:126)g can change as a function of β we estimate
where Σ(β) def= ΣAs((cid:126)gα(β)). Consequently,
Using this estimate of how much (cid:126)g changes in the Σ(β) norm, we now estimate how much (cid:126)g changes
in the (cid:96)∞ norm. Let (cid:126)z def= (G(β) − αΛg)
−1 (cid:126)1. Then, we have
.
(5.13)
G(β) − αΛg (cid:23) (1 − α)G(β) (cid:23) (1 − α)Σ(β)
1
=
Σ(β)
Σ(β)
−1 (cid:126)1
dβ
Σ(β)
(cid:13)(cid:13)(cid:13)2
rank (A)
1 − α
≤
1 − α
(cid:13)(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)(cid:13)G(β)−1 d(cid:126)g
≤ (cid:13)(cid:13)(cid:13)(G(β) − αΛg)
(cid:13)(cid:13)(cid:13)(cid:126)1
(cid:13)(cid:13)(cid:13)2
((1 − α) (cid:126)σi(β) + β)(cid:126)zi ≤ (cid:12)(cid:12)(cid:12)α(cid:126)1T
(cid:33)
=(cid:13)(cid:13)(cid:126)z(cid:13)(cid:13)∞ ≤ max
(cid:13)(cid:13)G(β − θ)−1((cid:126)g(β − θ) − (cid:126)w)(cid:13)(cid:13)∞ ≤ 1
(cid:32) α(cid:13)(cid:13)(cid:126)z(cid:13)(cid:13)Σ(β)
≤ max
1 − α
i P(2)(cid:126)z
(cid:12)(cid:12)(cid:12) + 1
≤ α(cid:126)σi(β)(cid:13)(cid:13)(cid:126)z(cid:13)(cid:13)Σ(β) + 1.
(cid:32)(cid:112)rank (A)
(1 − α)3/2
,
1
β
12cr
(cid:33)
.
,
1
β
Using (5.13) and α < 1, we have
(cid:13)(cid:13)(cid:13)(cid:13) d ln (cid:126)g
dβ
(cid:13)(cid:13)(cid:13)(cid:13)∞
Using (5.11), integrating, and applying Lemma 33 we have that
√
(1−α)3/2β
rank(A)
1000c2
r
for θ ≤
. Hence, this proves that step 2a preserves the invariant (5.10) at step 2a.
Hence, the algorithm satisfies the assumptions needed for Theorem 21 throughout and computeWeight
portion and
ins step 2a works as desired. Since each iteration β decreased by O
iterations. Using
the initial β is O(1) we see that the algorithm requires only O
Theorem 21 to bound the total number of linear systems solved then yields the result.
(cid:16)
(cid:17)
1/(cid:112)rank (A)
(cid:17)
(cid:16)(cid:112)rank (A)
33
6 Approximate Weights Suffice
In the previous sections, we analyzed a weighted path following strategy assuming oracle access
to a weight function we could compute exactly and showed how to compute a weight function
approximately. In this section we show why it suffices to compute multiplicative approximations
to the weight function. Ultimately, having access to this “noisy oracle” will only cause us to lose
polylogarithmic factors in the running time as compared to the “exact oracle” case.
This is a non-trivial statement as the weight function serves several roles in our weighted path
following scheme. First, it ensures a good ratio between total weight c1 and slack sensitivity cγ.
This allows us to take make large increases to the path parameter t after which we can improve
centrality. Second, the weight function is consistent and does not differ too much from the cr-update
step direction. This allows us to change the weights between cr-update steps without moving too
far away from the central path. Given a multiplicative approximation to the weight function, this
first property is preserved up to an approximation constant however this second property is not.
To effectively use multiplicative approximations to the weight function we cannot simply use
the weight function directly. Rather we need to smooth out changes to the weights by using some
slowly changing approximation to the weight function. In this section we show how this can be
achieved in general. First, in Section 6.1, we present the smoothing problem in a general form that
we call the chasing 0 game and we provide an effective strategy for playing this game. Then in
Section 6.2 we show how to use this strategy to produce a weighted path following scheme that uses
multiplicative approximations to the weight function.
6.1 The Chasing 0 Game
The chasing 0 game is as follows. There is a player, an adversary, and a point (cid:126)x ∈ Rm. The goal of
the player is to keep the point close to (cid:126)0 in (cid:96)∞ norm and the goal of the adversary tries to move (cid:126)x
away from (cid:126)0 ∈ Rm. The game proceeds for an infinite number of iterations where in each iteration
the adversary moves the current point (cid:126)x(k) ∈ Rm to some new point (cid:126)y(k) ∈ Rm and the player needs
to respond. The player does not know (cid:126)x(k), (cid:126)y(k), or the move of the adversary. All the player knows
is that the adversary moved the point within some convex set U (k) and the player knows some
(cid:126)z(k) ∈ Rn that is close to (cid:126)y(k) in (cid:96)∞ norm.14 With this information the player is allowed to move
the point a little more than the adversary. Formally, the player is allowed to set the next point to
(cid:126)x(k+1) ∈ Rm such that (cid:126)∆(k) def= (cid:126)x(k+1) − (cid:126)y(k) ∈ (1 + )U for some fixed > 0.
In particular, we would like an efficient strategy for computing (cid:126)∆(k) such that(cid:13)(cid:13)(cid:126)x(k)(cid:13)(cid:13)∞ is bounded
The question we would like to address is, how close the player can keep (cid:126)x(k+1) to (cid:126)0 in (cid:96)∞ norm?
for all k ≥ 0.
14To apply this result to weighted central path following we let the current points (cid:126)x(k) denote the difference between
log( (cid:126)w) and log((cid:126)g ((cid:126)x)). The sets U (k) are then related to the cr-update steps and the steps of the player are related
to the weights the path following strategy picks.
34
Chasing 0 Game:
1. Given R > 0, > 0, (cid:126)x(0) ∈ Rm.
2. For k = 1, 2,···
2a. The adversary announces symmetric convex set U (k) ⊆ Rn and (cid:126)u(k) ∈ U (k).
2b. The adversary sets (cid:126)y(k) := (cid:126)x(k) + (cid:126)u(k).
2c. The adversary announces (cid:126)z(k) such that(cid:13)(cid:13)(cid:126)z(k) − (cid:126)y(k)(cid:13)(cid:13)∞ ≤ R.
2d. The player chooses (cid:126)∆(k) ∈ (1 + ) U (k).
2e. The player sets (cid:126)x(k+1) = (cid:126)y(k) + (cid:126)∆(k).
player can follow to ensure that that (cid:13)(cid:13)(cid:126)x(k)(cid:13)(cid:13)∞ is never too large. Our strategy simply consists of
We show that assuming that the U (k) are sufficiently bounded then there is strategy that the
taking “gradient steps” using the following potential function.
Definition 23. For any µ ≥ 0 let pµ : R → R and Φµ : Rm → R be given by
∀x ∈ R
:
pµ(x) def= eµx + e−µx
and Φµ((cid:126)x) def=
pµ(xi).
(cid:88)
i∈[m]
In other words, for all k we simply set (cid:126)∆(k) to be the vector in (1 + )U (k) that best minimizes
the potential function of the observed position, i.e. Φµ((cid:126)z(k)) for an appropriate choice of µ. In the
following theorem we show that this suffices to keep Φµ((cid:126)x(k)) small and that small Φµ((cid:126)x(k)) implies
Theorem 24. Suppose that each U (k) is a symmetric convex set that contains an (cid:96)∞ ball of radius
rk and is contained in a (cid:96)∞ ball of radius Rk ≤ R.15 Let 0 < < 1
where
5 and consider the strategy
(cid:69)
µ =
.
small(cid:13)(cid:13)(cid:126)x(k)(cid:13)(cid:13)∞ and hence has the desired properties.
(cid:68)∇Φµ((cid:126)z(k)), (cid:126)∆
(cid:19)
(cid:18)
log(cid:0) 12mτ
In particular, we have(cid:13)(cid:13)(cid:126)x(k)(cid:13)(cid:13)∞ ≤ 12R
and suppose Φµ((cid:126)x(0)) ≤ 12mτ
∀k ≥ 0
(cid:126)∆(k) = (1 + ) arg min
(cid:126)∆∈U (k)
: Φµ((cid:126)x(k+1)) ≤
1 − 2rk
24R
Let τ def= maxk
Rk
rk
12R
(or more specifically(cid:13)(cid:13)(cid:126)x(0)(cid:13)(cid:13)∞ ≤ 12R
(cid:1).
Φµ((cid:126)x(k)) + m
≤ 12mτ
Rk
2R
.
log(cid:0) 6τ
(cid:1) ) then
µΦµ((cid:126)x) − 2µm ≤(cid:13)(cid:13)∇Φµ((cid:126)x)(cid:13)(cid:13)1
To prove Theorem 24 we first provide the following lemma regarding properties of the potential
function Φµ.
Lemma 25 (Properties of the Potential Function). For all (cid:126)x ∈ Rm, we have
and
eµ(cid:107)(cid:126)x(cid:107)∞ ≤ Φµ((cid:126)x) ≤ 2meµ(cid:107)(cid:126)x(cid:107)∞
(6.1)
Furthermore, for any symmetric convex set U ⊆ Rm and any (cid:126)x ∈ Rm, let (cid:126)x(cid:91) def= arg max(cid:126)y∈U (cid:104)(cid:126)x, (cid:126)y(cid:105)16
def= max(cid:126)y∈U (cid:104)(cid:126)x, (cid:126)y(cid:105). Then for all (cid:126)x, (cid:126)y ∈ Rm with(cid:13)(cid:13)(cid:126)x − (cid:126)y(cid:13)(cid:13)∞ ≤ δ ≤ 1
and(cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)U
e−µδ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:13)(cid:13)U − µ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:91)(cid:13)(cid:13)1 ≤(cid:68)∇Φµ((cid:126)x),∇Φµ((cid:126)y)(cid:91)(cid:69) ≤ eµδ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:13)(cid:13)U + µeµδ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:91)(cid:13)(cid:13)1.
15Formally we assume that if (cid:126)x ∈ U (k) then(cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)∞ ≤ R and we assume that if(cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)∞ ≤ r then (cid:126)x ∈ U (k).
5µ we have
(6.2)
16This is a scaled version of # operator in [27] and hence we name it differently.
35
eµx ≤ pµ(x) ≤ 2eµx
If additionally U is contained in a (cid:96)∞ ball of radius R then
Proof. First we note that for all x ∈ R we have
e−µδ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:13)(cid:13)U − µmR ≤(cid:13)(cid:13)∇Φµ((cid:126)x)(cid:13)(cid:13)U ≤ eµδ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:13)(cid:13)U + µeµδmR.
eµx − e−µx(cid:17)
µ(x) = µ(cid:0)eµx − e−µx(cid:1) and
eµδ−µy − e−µy−µδ(cid:17)
and therefore we have (6.1).
since x − y ≤ δ we have that x = y + z for some z ∈ (−δ, δ). Using that p(cid:48)(x) is monotonic in
x we then have
p(cid:48)
µ(x) = p(cid:48)
= µ
Next let x, y ∈ R such that x − y ≤ δ. Note that (cid:12)(cid:12)p(cid:48)
(cid:16)
µ(x)(cid:12)(cid:12) = p(cid:48)
(cid:16)
eµy+µδ − e−µy−µδ(cid:17)
µ(y + δ)
= eµδp(cid:48)(y) + µ
µ(y + z) ≤ p(cid:48)
µ(x) = µ sign(x)
and p(cid:48)
(6.3)
(cid:16)
µ(x) = p(cid:48)
≤ eµδ(cid:12)(cid:12)p(cid:48)(y)(cid:12)(cid:12) + µeµδ.
(6.4)
By symmetry (i.e. replacing x and y) this implies that
(6.5)
Since U is symmetric this implies that for all i ∈ [m] we have sign(∇Φµ((cid:126)y)(cid:91))i = ∇Φµ((cid:126)y)i =
sign(yi) . Therefore, if for all i ∈ [n] we have sign(xi) = sign(yi), by (6.4), we see that
(cid:68)∇Φµ((cid:126)x),∇Φµ((cid:126)y)(cid:91)(cid:69)
i
=
µ(x) ≥ e−µδp(cid:48)(y) − µ
p(cid:48)
(cid:88)
µ(yi) + µeµδ(cid:17)∇Φµ((cid:126)y)(cid:91)
≤ (cid:88)
≤ eµδ(cid:68)∇Φµ((cid:126)y),∇Φµ((cid:126)y)(cid:91)(cid:69)
= eµδ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:13)(cid:13)U + µeµδ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:91)(cid:13)(cid:13)1.
µ(xi)∇Φµ((cid:126)y)(cid:91)
p(cid:48)
(cid:16)
eµδp(cid:48)
i
i
i
+ µeµδ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:91)(cid:13)(cid:13)1
2 since δ ≤ 1
e−µδ(cid:12)(cid:12)p(cid:48)
5µ. Thus, we have
≤ sign (yi) p(cid:48)
µ(yi)(cid:12)(cid:12) − µ ≤ − µ
Similarly, using (6.5), we have e−µδ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:13)(cid:13)U − µ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:91)(cid:13)(cid:13)1 ≤ (cid:10)∇Φµ((cid:126)x),∇Φµ((cid:126)y)(cid:91)(cid:11) and hence
Taking inner product on both sides with ∇Φµ((cid:126)y)(cid:91)
Thus, (6.2) holds in general.
(6.2) holds. On the other hand if sign(xi) (cid:54)= sign(yi) then we know that xi ≤ δ and consequently
p(cid:48)
µ(xi) ≤ µ(eµδ − e−µδ) ≤ µ
µ(xi) ≤ 0 ≤ eµδ(cid:12)(cid:12)p(cid:48)
i and using definition of(cid:13)(cid:13)·(cid:13)(cid:13)U and ·(cid:91), we get (6.2).
Finally we note that since U is contained in a (cid:96)∞ ball of radius R, we have(cid:13)(cid:13)(cid:126)y(cid:91)(cid:13)(cid:13)1 ≤ mR for all
(cid:126)y. Using this fact, (6.2), and the definition of(cid:13)(cid:13) ·(cid:13)(cid:13)U, we obtain
where the last line comes from the fact ∇Φµ((cid:126)y)(cid:91) ∈ U and the definition of (cid:13)(cid:13) ·(cid:13)(cid:13)U. By symmetry
e−µδ(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:13)(cid:13)U − µmR ≤(cid:68)∇Φµ((cid:126)x),∇Φµ((cid:126)y)(cid:91)(cid:69) ≤(cid:13)(cid:13)∇Φµ((cid:126)x)(cid:13)(cid:13)U
µ(yi)(cid:12)(cid:12) + µeµδ.
2
(6.3) follows.
36
Using Lemma 25 we prove Theorem 24.
Proof. [Theorem 24] For the remainder of the proof, let (cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)U (k) = max(cid:126)y∈U (k) (cid:104)(cid:126)x, (cid:126)y(cid:105) and (cid:126)x(cid:91)(k) =
arg max(cid:126)y∈U (k) (cid:104)(cid:126)x, (cid:126)y(cid:105). Since U (k) is symmetric, we know that (cid:126)∆(k) = − (1 + )(cid:0)∇Φµ((cid:126)z(k))(cid:1)(cid:91)(k) and
therefore by applying the mean value theorem twice we have that
Φµ((cid:126)x(k+1)) = Φµ((cid:126)y(k)) +
= Φµ((cid:126)x(k)) +
(cid:68)∇Φµ((cid:126)z), (cid:126)x(k+1) − (cid:126)y(k)(cid:69)
(cid:68)∇Φµ((cid:126)y), (cid:126)y(k) − (cid:126)x(k)(cid:69)
(cid:28)
+
(cid:68)∇Φµ((cid:126)z), (cid:126)x(k+1) − (cid:126)y(k)(cid:69)
(cid:17)(cid:91)(k)(cid:29)
(cid:16)∇Φµ((cid:126)z(k))
∇Φµ((cid:126)z),
for some (cid:126)y between (cid:126)y(k) and (cid:126)x(k) and some (cid:126)z between (cid:126)x(k+1) and (cid:126)y(k). Now, using that (cid:126)y(k) − (cid:126)x(k) ∈
U (k) and that (cid:126)x(k+1) − (cid:126)y(k) = (cid:126)∆(k) we have
Φµ((cid:126)x(k+1)) ≤ Φµ((cid:126)x(k)) +(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:13)(cid:13)U (k) − (1 + )
.
(6.6)
Furthermore, since < 1
Since U k is contained within the (cid:96)∞ ball of radius Rk Lemma 25 shows that
(cid:13)(cid:13)∇Φµ((cid:126)y)(cid:13)(cid:13)U (k) ≤ eµRk(cid:13)(cid:13)∇Φµ((cid:126)x(k))(cid:13)(cid:13)U (k) + mµRkeµRk .
5 and Rk ≤ R, by triangle inequality we have(cid:13)(cid:13)(cid:126)z−(cid:126)z(k)(cid:13)(cid:13)∞ ≤ (1 + )Rk + R ≤
3R and(cid:13)(cid:13)(cid:126)z(k) − (cid:126)x(k)(cid:13)(cid:13)∞ ≤ 2R. Therefore, applying Lemma 25 twice yields that
(cid:69) ≥ e−3µR(cid:13)(cid:13)∇Φµ((cid:126)z(k))(cid:13)(cid:13)U (k) − µmRk
(cid:68)∇Φµ((cid:126)z),∇Φµ((cid:126)z(k))(cid:91)(k)
≥ e−5µR(cid:13)(cid:13)∇Φµ((cid:126)x(k))(cid:13)(cid:13)U (k) − 2µmRk.
(6.8)
(6.7)
Combining (6.6), (6.7), and (6.8) then yields that
Φµ((cid:126)x(k+1)) ≤ Φµ((cid:126)x(k)) −(cid:0)(1 + )e−5µR − eµR(cid:1)(cid:13)(cid:13)∇Φµ((cid:126)x(k))(cid:13)(cid:13)U (k) + mµRkeµR + 2(1 + )mµRk.
Since we chose µ =
12R, we have
1 + ≤
2
Hence, we have (1 + )e−5µR − eµR ≤
2. Also, since 0 < < 1
+ (1 + 6µR) ≤
2
e5µR + e6µR.
5 we have
mµRkeµR + 2(1 + )mµRk ≤(cid:0)eµR + 2(1 + )(cid:1) mµRk ≤ m
(cid:13)(cid:13)∇Φµ((cid:126)x(k))(cid:13)(cid:13)U (k) + m
Φµ((cid:126)x(k+1)) ≤ Φµ((cid:126)x(k)) −
2
7Rk
24R
.
7Rk
24R
.
Using Lemma 25 and the fact that Uk contains a (cid:96)∞ ball of radius rk, we have
Φµ((cid:126)x(k)) − 2m
(cid:13)(cid:13)∇Φµ((cid:126)x(k))(cid:13)(cid:13)U (k) ≥ rk
(cid:13)(cid:13)∇Φµ((cid:126)x(k))(cid:13)(cid:13)1 ≥ rk
(cid:16)
(cid:17)
.
12R
Thus, we have
37
Therefore, we have that
Φµ((cid:126)x(k+1)) ≤
≤
(cid:19)
(cid:19)
(cid:18)
(cid:18)
1 − 2rk
24R
1 − 2rk
24R
Φµ((cid:126)x(k)) +
rk
12R
m + m
7Rk
24R
Φµ((cid:126)x(k)) + m
Rk
2R
.
Hence, if Φµ((cid:126)x(k)) ≤ 12mτ
have by induction that Φµ((cid:126)x(k)) ≤ 12mτ
immediately from Lemma 25.
, we have Φµ((cid:126)x(k+1)) ≤ 12mτ
for all k. The necessary bound on (cid:13)(cid:13)(cid:126)x(k)(cid:13)(cid:13)∞ then follows
. SinceΦµ((cid:126)x(0)) ≤ 12mτ
by assumption we
6.2 Centering Step With Noisy Weight
Here we show how to use the results of the previous section to perform weighted path following
given access only to a multiplicative approximation of the weight function. In particular, we show
how to use Theorem 24 to improve the centrality of (cid:126)x while maintaining the invariant that (cid:126)w is
close to (cid:126)g((cid:126)x) multiplicatively.
>0}, we measure the distance between the
>0, in log scale (cid:126)Ψ((cid:126)s, (cid:126)w) def= log((cid:126)g((cid:126)s)) −
log( (cid:126)w). Our goal is to keep(cid:13)(cid:13)(cid:126)Ψ((cid:126)s, (cid:126)w)(cid:13)(cid:13)∞ ≤ K for some error threshold K. We choose K to be just
As in Section 4 given a feasible point, {(cid:126)x, (cid:126)w} ∈ {S0×Rm
>0, and the weight function, (cid:126)g((cid:126)s) ∈ Rm
current weights, (cid:126)w ∈ Rm
small enough that we can still decrease δt((cid:126)x, (cid:126)w) linearly and still approximate (cid:126)g((cid:126)s), as in general it
may be difficult to compute (cid:126)g((cid:126)s) when (cid:126)w is far from (cid:126)g((cid:126)s). Furthermore, we ensure that (cid:126)Ψ doesn’t
change too much in either (cid:107) · (cid:107)∞ or (cid:107) · (cid:107)W(new) and thereby ensure that the centrality does not
increase too much as we move (cid:126)w towards (cid:126)g((cid:126)s).
We meet these goals by playing the chasing 0 game where the vector we wish to keep near (cid:126)0 is
(cid:126)Ψ((cid:126)s, (cid:126)w), the adversaries moves are cr-steps, and our moves change log( (cid:126)w). The cr-step decreases δt
and since we are playing the chasing 0 game we keep (cid:126)Ψ((cid:126)s, (cid:126)w) small. Finally, since by the rules of the
chasing 0 game we do not move (cid:126)w much more than (cid:126)g((cid:126)s) has moved, we have by similar reasoning
to the exact weight computation case, Theorem 11 that changing (cid:126)w does not increase δt too much.
This inexact centering operation and the analysis are formally defined and analyzed below.
Most of the parameter balancing involved in this paper lies in the theorem below. Due to the
step consistency, we know know that after a cr-steps, the weight does not move too far away that
we can move it back without hurting centrality too much if we can compute the weight exactly. The
Chasing 0 game shows that we can mimic this if we compute the weight accurate enough. Therefore,
the balancing is simply about how accurate we need to do.
K
60cr log(960crcγ m3/2)
, δt = δt((cid:126)x(old), (cid:126)w(old)), = 1
5cr
((cid:126)x(new), (cid:126)w(apx)) = centeringInexact((cid:126)x(old), (cid:126)w(old), K, approxWeight)
1. R =
and µ =
12R .
2. {(cid:126)x(new), (cid:126)w(new)} = stept((cid:126)x(old), (cid:126)w(old), cr) as in Definition 4.
3. Let U = {(cid:126)y ∈ Rm (cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)W(new)
5. (cid:126)w(apx) := exp(cid:0)log( (cid:126)w(new)) + (1 + ) arg min(cid:126)u∈U
cr+1 δt and(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)∞ ≤ 4cγδt}
4. Compute (cid:126)z = approxWeight((cid:126)s, (cid:126)w(new), R).
(cid:10)∇Φµ
≤ cr+0.14
(cid:0)log((cid:126)z) − log(cid:0) (cid:126)w(new)(cid:1)(cid:1) , (cid:126)u(cid:11)(cid:1)
Note that in step 5 in centeringInexact, we need to project a certain vector onto the intersection
, and box,(cid:13)(cid:13) ·(cid:13)(cid:13)∞. In Section C we show that this can be computed in parallel in
of ball,(cid:13)(cid:13) ·(cid:13)(cid:13)W(new)
38
12R. Let ((cid:126)x(new), (cid:126)w(apx)) = centeringInexact((cid:126)x(old), (cid:126)w(old), K), then
and Φµ
def= Φµ((cid:126)Ψ((cid:126)x(old), (cid:126)w(old))) ≤ 960crcγm3/2
(cid:19)
1 − 0.5
1 + cr
δt
Φµ((cid:126)x(k)) +
2mcγδt
5R
≤ 960crcγm3/2.
depth O(1) and work O(m) and therefore this step is not a bottleneck in the computational cost
of our weighted path following schemes.
Theorem 26 (Centering with Inexact Weights). Given current point {(cid:126)x(old), (cid:126)w(old)} ∈ {S0 × Rm
error parameter K ≤ 1
>0},
, and approximate weight computation oracle, approxWeight, such that
(cid:13)(cid:13) log(approxWeight((cid:126)s, (cid:126)w, R))−log ((cid:126)g((cid:126)s))(cid:13)(cid:13)∞ ≤ R for (cid:126)s, (cid:126)w ∈ Rm
>0 with(cid:13)(cid:13) log( (cid:126)w)−log ((cid:126)g((cid:126)s))(cid:13)(cid:13)∞ ≤ 2K,
8cr
assume that
def= δt((cid:126)x(old), (cid:126)w(old)) ≤
δt
240crcγ log(cid:0)960crcγm3/2(cid:1)
K
and
(cid:18)
where µ =
δt((cid:126)x(new), (cid:126)w(apx)) ≤
(cid:18)
(cid:19)
Also, we have(cid:13)(cid:13) log((cid:126)g((cid:126)s(new))) − log( (cid:126)w(apx))(cid:13)(cid:13)∞ ≤ K.
U def= {(cid:126)y ∈ Rm (cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)W(new)
Φµ((cid:126)Ψ((cid:126)x(new), (cid:126)w(apx))) ≤
δt
600c2
rR
≤ Cw
1 −
√
m
Proof. By Lemma 9, we know that for a cr-update step, we have (cid:126)Ψ((cid:126)x(new), (cid:126)w(new))−(cid:126)Ψ((cid:126)x(old), (cid:126)w(old)) ∈
U where U is the symmetric convex set given by
(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)∞ ≤ C∞}
and
where
Note that since δt ≤ K(cid:0)240crcγ log(cid:0)960crcγm3/2(cid:1)(cid:1)−1 we have
and Cw =
cr + 1/8
cr + 1
C∞ = 4cγδt
δt + 13cγδ2
t .
(cid:32)
240crcγ log(cid:0)960crcγm3/2(cid:1)(cid:33)
K
C∞ ≤ 4cγ
60cr log(cid:0)960crcγm3/2(cid:1) = R
K
≤
Therefore U is contained in a (cid:96)∞ ball of radius R. Again using the bound on δt we have
Cw =
cr + 1
8
cr + 1
δt + 13cγδ2
t ≤ cr + 1
8
cr + 1
δt +
0.008
cr
δt
≤ cr + 0.14
cr + 1
δt.
(6.9)
Consequently, U ⊆ U where we recall that U is the symmetric convex set defined by
U = {(cid:126)y ∈ Rm (cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)W(new)
≤ cr + 0.14
cr + 1
δt
and
(cid:13)(cid:13)(cid:126)y(cid:13)(cid:13)∞ ≤ 4cγδt}.
that(cid:13)(cid:13)(cid:126)Ψ((cid:126)s(old), (cid:126)w(old))(cid:13)(cid:13)∞ ≤ K ≤ 1
8cr
Therefore, we can play the chasing 0 game on (cid:126)Ψ((cid:126)s(old), (cid:126)w(old)) attempting to maintain the invariant
without taking steps that are more than 1 + times the size of
U. We pick = 1
5cr
so to not interfere with our ability to decrease δt linearly.
39
log
12R
(cid:19)
≤ K
satisfies the following
(cid:18) 12mτ
To use the chasing 0 game to maintain (cid:13)(cid:13)(cid:126)Ψ((cid:126)s(old), (cid:126)w(old))(cid:13)(cid:13)∞ ≤ K we need to ensure that R
(cid:96)∞ ball that U contains. Since(cid:13)(cid:13)(cid:126)g((cid:126)s(old))(cid:13)(cid:13)∞ ≤ 2 by Definition 7 and since(cid:13)(cid:13)(cid:126)Ψ((cid:126)x(old), (cid:126)w(old))(cid:13)(cid:13)∞ ≤ 1
by assumption we have that(cid:13)(cid:13) (cid:126)w(old)(cid:13)(cid:13)∞ ≤ 3. By Lemma 5 we know that(cid:13)(cid:13) (cid:126)w(new)(cid:13)(cid:13)∞ ≤ 4 if δtcγ ≤ 1
Consequently, if(cid:13)(cid:13)(cid:126)u(cid:13)(cid:13)∞ ≤ δt
where here τ is as defined in Theorem 24. To bound τ we need to lower bound the radius of the
∀u ∈ Rm :
m, then (cid:126)u ∈ U. Thus, U contains a a box of radius
(cid:18) 12mτ
m and consequently
contained in a box of radius 4cγδt, we have that τ ≤ 16cγ
m and since U is
and consequently
(cid:13)(cid:13)(cid:126)u(cid:13)(cid:13)2
√
4
4m
√
δt
8
8
4
12R
log
≤ 60crR log
This proves that we meet the conditions of Theorem 24. Therefore, we have
.
√
W(new)
(cid:13)(cid:13)(cid:126)u(cid:13)(cid:13)2
960crcγm3/2(cid:17) ≤ K.
(cid:19)(cid:19)
(cid:19)
Φµ((cid:126)x(k)) + m
Φµ((cid:126)x(k)) +
∞ ≥ 1
(cid:16)
(cid:18) δt
√
m
4
√
m
1
2R
2mcγδt
5R
Φµ((cid:126)Ψ((cid:126)x(new), (cid:126)w(apx))) ≤
1 − 2
24R
(4cγδt)
where we do not need to re-derive the last line because it follows from Theorem 24.
Consequently, (cid:13)(cid:13)(cid:126)Ψ((cid:126)x(old), (cid:126)w(old))(cid:13)(cid:13)∞ ≤ K and Φµ((cid:126)Ψ((cid:126)x(new), (cid:126)w(apx))) ≤ 960crcγm3/2. Since K ≤
8, we have (cid:13)(cid:13)G((cid:126)s(old))−1( (cid:126)w(old) − (cid:126)g((cid:126)s(old)))(cid:13)(cid:13)∞ ≤ 1.2 and γ((cid:126)s(old), (cid:126)w(old)) ≤ 2cγ. Consequently, by
1
Lemma 6 we have
δt((cid:126)x(new), (cid:126)w(new)) ≤ γ((cid:126)x(old), (cid:126)w(old)) · δ2
t ≤ 2 · cγ · δ2
t
1 −
δt
=
600c2
rR
≤ 960crcγm3/2.
(cid:19)
(cid:18)
(cid:18)
(cid:104)
Let
∞ def=(cid:13)(cid:13) log( (cid:126)w(apx)) − log( (cid:126)w(new))(cid:13)(cid:13)∞ and
w
By our bounds on U, we have
∞ ≤ (1 + )R ≤ 1
100cr
Using Lemma 10, we have that
and w = (1 + )
def=(cid:13)(cid:13) log( (cid:126)w(apx)) − log( (cid:126)w(new))(cid:13)(cid:13)W(new)
(cid:20) cr + 0.14
.
≤ cr + 0.37
cr + 1
δt.
cr + 1
δt
(cid:21)
(cid:105) ≤ 3cγδ2
δt((cid:126)x(new), (cid:126)w(apx)) ≤ (1 + ∞)
1
≤
1 +
(cid:18)
δt((cid:126)x(new), (cid:126)w(new)) + w
(cid:19)(cid:18) cr + 0.34
(cid:19)
δt + 3cγδ2
t ≤
t + (1 + ∞)w
(cid:18) cr + 0.5
(cid:19)
δt
cr + 1
100cr
cr + 1
40
7 The Algorithm
In this section we show how to put together the results of the previous sections to solve a linear
program. First, in Section 7.1 we provide a path following routine that allows us to move quickly
from one approximate central path point to another. Using this subroutine, in Section 7.2 we show
how to obtain an algorithm for solving a linear program in (cid:101)O((cid:112)rank(A)L) iterations that consist of
solving linear systems in the original constraint matrix. In the Appendix we provide additional proof
details such as how these algorithm only require approximate linear system solvers (Appendix D)
and how to initialize our interior point technique and round approximate solutions to optimal ones
(Appendix E).
7.1 Path Following
We start by analyzing the running time of pathFollowing a subroutine for following the weighted
central path.
(cid:17)
(cid:16) 2m
(cid:18)
((cid:126)x(new), (cid:126)w(new)) = pathFollowing((cid:126)x(old), (cid:126)w(old), tstart, tend)
1. cr = 2 log2
2. While t < tend
2a. ((cid:126)x(new), (cid:126)w(apx)) = centeringInexact((cid:126)x(old), (cid:126)w(old), K, computeWeight)
2b.
, t = tstart, K = 1
24cr
t(new) := t
(cid:19)
rank(A)
1 +
.
.
√
r log(crm)
1
1010c3
rank(A)
1
r log(cid:0)1920crm3/2(cid:1)
r log(cid:0)1920crm3/2(cid:1)
2c. (cid:126)x(old) := (cid:126)x(new), (cid:126)w(old) := (cid:126)w(apx), t := t(new)
2d. For every
100cr log(crm) steps, check if the current (cid:126)x, (cid:126)w satisfies the δ and Φ invariants.
m
If it does not satisfies, roll back to the last time the invariants were met.
3. Output ((cid:126)x(old), (cid:126)w(old)).
Theorem 27 (Main Result). Given {(cid:126)x(old), (cid:126)w(old)} ∈ {S0 × Rm
>0} and tstart ≤ tend. Suppose that
and Φµ((cid:126)Ψ((cid:126)x(old), (cid:126)w(old))) ≤ 1920crm3/2
δtstart((cid:126)x(old), (cid:126)w(old)) ≤
11520c2
(cid:16)(cid:112)rank(A) log
(cid:16) tend
(cid:17)(cid:17)
where µ = 2 log (52crm) /K. Let ((cid:126)x(new), (cid:126)w(apx)) = pathFollowing((cid:126)x(old), (cid:126)w(old), tstart, tend), then
and Φµ((cid:126)Ψ((cid:126)x(new), (cid:126)w(new))) ≤ 1920crm3/2.
δtend((cid:126)x(new), (cid:126)w(new)) ≤
1
11520c2
Furthermore, computing ((cid:126)x(new), (cid:126)w(new)) takes O
where the cost of each iteration is dominated by the time need to solve O(1) linear system solves.
iterations in expectation
tstart
Proof. This algorithm maintains the invariant that
δt((cid:126)x(old), (cid:126)w(old)) ≤
r log(cid:0)1920crm3/2(cid:1) and Φµ((cid:126)Ψ((cid:126)x(old), (cid:126)w(old))) ≤ 1920crm3/2
1
11520c2
in each iteration in the beginning of the step (2a). Note that our oracle computeWeight satisfies
the assumption of Theorem 26 since 2K ≤ 1
. Hence, centeringInexact can use computeWeight
12cr
41
(cid:18)
(cid:19)
to find the approximations of (cid:126)g((cid:126)s(new)). Hence, Theorem 26 shows that we have
δt((cid:126)x(new), (cid:126)w(apx)) ≤
1 − 0.5
1 + cr
and Φµ((cid:126)Ψ((cid:126)x(new), (cid:126)w(apx))) ≤ 1920crm3/2.
δt
Using the fact c1((cid:126)g) ≤ 2 rank(A) and that (cid:126)w(new) is within a multiplicative factor of two of (cid:126)g((cid:126)s(new))
by Lemma 1 we have
δt(new)((cid:126)x(new), (cid:126)w(apx))
(cid:32)
1 +
≤
≤
1010c3
1
r log (crm)(cid:112)rank(A)
r log(cid:0)1920crm3/2(cid:1)
1
(cid:33)(cid:18)
1 − 0.5
1 + cr
(cid:19)
δt +
(cid:113)(cid:13)(cid:13) (cid:126)w(new)(cid:13)(cid:13)1
r log (crm)(cid:112)rank(A)
1010c3
(cid:1)(cid:100)12cr log( 4m
11520c2
Theorem 21 shows that with probability(cid:0)1 − 1
K )(cid:101), computeWeight outputs a correct
answer. Therefore, for each
100cr log(crm) iterations there is constant probability that the whole
procedure runs correctly. Hence, we only need to know how long it takes to check the current state
satisfies δt and Φµ invariants. We can check the δt easily using only 1 linear system solve. To check
Φµ, we need to compute the weight function exactly. To do this, we use lemma 19 and note that
computing the leverage scores exactly takes m linear system solve. Therefore, the averaged cost of
step 2d is just O(1) linear system solves and this justified the total running time.
m
m
7.2 Solving a Linear Program
Here we show how to use the properties of pathFollowing proved in Theorem 27 to obtain a linear
program solver. Given the previous theorem all that remains is to show how to get the initial central
point and round the optimal point to a vertex. We defer much of the proof of how to obtain an
initial point, deal with unbounded solutions, and round to an optimal vertex to Lemma 40 proved
in Appendix E.
Theorem 28. Consider a linear programming problem of the form
min
(cid:126)x∈Rn : A(cid:126)x≥(cid:126)b
(cid:126)cT (cid:126)x
(7.1)
where A ∈ Rm×n, (cid:126)b ∈ Rm, and (cid:126)c ∈ Rn have integer coefficients. Let L denote the bit complexity of
(7.1) and suppose that for any positive definite diagonal matrix D ∈ Rm×m with condition number
2 O(L) there is an algorithm solve(A,(cid:126)b, D, ) such that
(cid:13)(cid:13)solve(A,(cid:126)b, D, ) − (DA)+(cid:126)b(cid:13)(cid:13)AT D2A ≤ (cid:13)(cid:13) (DA)+ (cid:126)b(cid:13)(cid:13)AT D2A
(cid:16)(cid:112)rank(A) (T + nnz(A)) L
(cid:17)
in time O (T log(1/)) for any > 0 with success probability greater than 1 − 1
an algorithm to solve (7.1) in expected time O
constraints of an optimal solution or prove that the program is unfeasible or unbounded.
m. Then, there is
, i.e, find the active
(7.2)
42
Using [24] as the Solve algorithm, we obtain an algorithm that solves (7.1) in time
O
(cid:16)(cid:112)rank(A) (nnz(A) + (rank(A))ω) L
(cid:17)
(cid:68)
(cid:69)
.
(7.3)
where ω < 2.3729 [42] is the matrix multiplication constant.
Proof. Applying the Lemma 40 we obtain a modified linear program
given A(new)(cid:126)x ≥ (cid:126)b(new)
(cid:126)c(new), (cid:126)x
min
which is bounded and feasible with O(n) variables, O(m) constraints, O(rank(A)) rank and O(L)
bit complexity. Also, we are given an explicit interior point (cid:126)x0.
To obtain an initial weighted central path point, we can use Theorem 22. However, (cid:126)x may not be
close to central path, i.e. δt could be large. To fix this, we can temporarily change the cost function
such that δt = 0. In particular, we can set (cid:126)cmodified = AT S−1
x (cid:126)w and get δt = 0 for this modified cost
function. One can think of Theorem 27 as showing that we can get the central path point from a
certain cost function tstart(cid:126)c to another cost function tend(cid:126)c in time that depends only logarithmically
on the multiplicative difference between these two vectors. Clearly, instead of increasing t we can
decrease t similarly. Hence, we can decrease t such that we get the central path point (cid:126)xcenter for
the cost function 2− Θ(L)(cid:126)cmodified. Since 2− Θ(L) is close enough to zero, it can be shown that δt is
small also for the cost function 2− Θ(L)(cid:126)c. Then, we could use Theorem 27 to increase t and obtain
the central path point for t = 2 Θ(L).
standard duality gap theorem,17 we know that the duality gap of (cid:126)xt is less than (cid:13)(cid:13) (cid:126)w(cid:13)(cid:13)1/t and in
this case it is less than 2− Θ(L) because(cid:13)(cid:13) (cid:126)w(cid:13)(cid:13)1 ≤ 2 rank (A). Now, we can use the conclusion of the
Then, we can use centeringInexact to make δt becomes and hence (cid:126)cT (cid:126)xt close to (cid:126)cT (cid:126)x. By a
Lemma 40 to find the active constraints of an optimal solution of the original linear program or
prove that it is infeasible or unbounded.
During the algorithm, we only called the function centeringInexact O(L) times and hence the
algorithm only executes O(L) linear system solves. In Section D, we show that these linear systems
do not need to be solved exactly and that inexact linear algebra suffices. Using this observation and
letting using [24] as the solve routine yields the total running time of
(cid:16)(cid:112)rank(A) (nnz(A) + (rank(A))ω) L
(cid:17)
.
O
In Section C, we show that the projection problem in centeringInexact can be computed in
(cid:101)O(1) depth and O(m) work and other operations are standard parallelizable linear algebra opera-
tions. Therefore, we achieve the first (cid:101)O((cid:112)rank (A)L) depth polynomial work method for solving
Theorem 29. There is an (cid:101)O((cid:112)rank (A)L) depth polynomial work algorithm to solve linear program
linear programs.
of the form
where L denote the bit complexity of the linear program.
17See [16] or [25] for a more detailed treatment of this fact in a more general regime.
min
(cid:126)x∈Rn : A(cid:126)x≥(cid:126)b
(cid:126)cT (cid:126)x
43
7.3 Accelerating the Solver
In this section, we show that how we can apply acceleration methods for decreasing the iterations
of interior point techniques can be applied to our algorithm to yield a faster method. In particular
we show how to adapt techniques of Vaidya [37] for using fast matrix multiplication to obtain a
faster running time. Our goal here is to provide a simple exposition of how the iteration costs of
our method can be decrease. We make no attempt to explore the running time of our algorithm in
all regimes and we note that since our algorithm only needs to solve linear systems in scalings of
the original constraint matrix there may be techniques to improve our algorithm further in specific
regimes by exploiting structure in A.
To accelerate our path following method, we note that we solve systems of two forms: we solve
systems in AT S−1WS−1A to update (cid:126)x and we solve systems in AT S−1WαS−1A to update (cid:126)w.
Since we have proven in Lemma 15 that two system are spectrally similar, we only need to know
how to solve system of the form AT S−1WS−1A and then we can use preconditioning to solve either
system. Furthermore, we note similarly to Vaidya [37] that the S and W matrices do not change
too much from iteration to iteration and therefore a sequence of the necessary linear system can be
solved faster than considering them individually. Below we state and Appendix F we prove a slight
improvement of a result in [37] formally analyzing one way of solving these systems faster.
Theorem 30. Let (cid:126)d(i) ∈ Rm
that d(i)
j
given the (cid:126)d(i) in a sequence, in each iteration i we can compute (cid:0)AT DiA(cid:1)−1 (cid:126)xi for Di = diag((cid:126)di)
>0 be a sequence of r positive vectors. Suppose that the number of times
for any i ∈ [r] and j ∈ [m] is bounded by Cr2 for some C ≥ 1. Then if we are
(cid:54)= d(i+1)
j
and arbitrary (cid:126)xi ∈ Rn with average cost per iteration
(cid:18) mnω−1
r
(cid:101)O
(cid:19)
+ n2 + Cωr2ω + Cω−1nrω
where ω < 2.3729 [42] is the matrix multiplication constant.
Using Theorem 30 we simply need to estimate how much the diagonal entries S−1WS−1 to
obtain a faster linear program solver. We prove the following.
Theorem 31. For any n
(cid:18)(cid:112)mβ
O
(cid:18)
m ≤ β ≤ 1 and r > 1, there is an
mnω−1
nnz(A) + n2 +
r
+ β−ωr2ω + β−(ω−1)nrω
(cid:19)
(cid:19)
L
(7.4)
time algorithm for solving linear programming problems of the form
min (cid:126)cT (cid:126)x given A(cid:126)x ≥ (cid:126)b
where A ∈ Rm×n.
Proof. Instead of using β = n
m , 1] be arbitrary as in the
theorem statement. Looking at the analysis in Section 5 we see that this yields a weight function
with c1 = O (βm), cγ = O(1) and cr = O(1). Consequently, it takes O(cid:0)√
βmL(cid:1) iterations to solve
m in the weight function we let β ∈ [ n
the linear program.
44
theorem to compute(cid:0)AT DjA(cid:1)−1 (cid:126)x for each group of operations, we need to estimate the change of
We separate the sequence of the linear systems involved into groups of size r. To use the previous
the diagonal entries S−1WS−1. For the change of S, Lemma 5 shows that
= O(1).
Since we have added β in the weight function, we have (cid:126)wi ≥ β and
(cid:13)(cid:13) log ((cid:126)sj) − log ((cid:126)sj+1)(cid:13)(cid:13)Wj
(cid:13)(cid:13) log ((cid:126)sj) − log ((cid:126)sj+1)(cid:13)(cid:13)2 = O(β−1/2).
Therefore, in a period of r operations, at most O(cid:0)β−1r2(cid:1) coordinates can change multiplicatively
only changing (cid:126)d a total of O(cid:0)β−1r2(cid:1) over a sequence of r operations. Using Theorem 30 and using
Therefore, we can maintain a vector (cid:126)d such that D is spectrally similar to S−1WS−1 while
by a constant factor. Similarly, we can use inequality (6.9) to analyze the change of W.
ADA as pre-conditioner for the necessary linear system solves, we can solve the linear system with
average cost
(cid:18)
Using that the total number of iterations is O(cid:0)√
nnz(A) + n2 +
O
r
+ β−ωr2ω + β−(ω−1)nrω
βmL(cid:1) then yields (7.4).
mnω−1
(cid:19)
.
8 Acknowledgments
We thank Yan Kit Chim, Andreea Gane, Jonathan A. Kelner, Lap Chi Lau, Aleksander Mądry,
Cameron Musco, Christopher Musco, Lorenzo Orecchia, Ka Yu Tam and Nisheeth Vishnoi for many
helpful conversations. This work was partially supported by NSF awards 0843915 and 1111109, NSF
Graduate Research Fellowship (grant no. 1122374) and Hong Kong RGC grant 2150701. Finally,
we thank the referees for extraordinary efforts and many helpful suggestions.
References
[1] Kurt M. Anstreicher. Volumetric path following algorithms for linear programming. Math.
Program., 76:245–263, 1996.
[2] Kenneth L Clarkson and David P Woodruff. Low rank approximation and regression in input
sparsity time. In Proceedings of the 45th annual ACM symposium on Symposium on theory of
computing, pages 81–90. ACM, 2013.
[3] Samuel I Daitch and Daniel A Spielman. Faster approximate lossy generalized flow via interior
point algorithms. In Proceedings of the 40th annual ACM symposium on Theory of computing,
pages 451–460. ACM, 2008.
[4] George B Dantzig. Maximization of a linear function of variables subject to linear inequalities.
New York, 1951.
[5] RobertM. Freund. Projective transformations for interior-point algorithms, and a superlinearly
convergent algorithm for the w-center problem. Mathematical Programming, 58(1-3):385–414,
1993.
45
[6] Andrew V. Goldberg and Satish Rao. Beyond the flow decomposition barrier. J. ACM,
45(5):783–797, 1998.
[7] Clovis C Gonzaga. Path-following methods for linear programming. SIAM review, 34(2):167–
224, 1992.
[8] Narendra Karmarkar. A new polynomial-time algorithm for linear programming. In Proceedings
of the sixteenth annual ACM symposium on Theory of computing, pages 302–311. ACM, 1984.
[9] Jonathan A. Kelner, Lorenzo Orecchia, Aaron Sidford, and Zeyuan Allen Zhu. A Simple,
Combinatorial Algorithm for Solving SDD Systems in Nearly-Linear Time. January 2013.
[10] Leonid G Khachiyan. Polynomial algorithms in linear programming. USSR Computational
Mathematics and Mathematical Physics, 20(1):53–72, 1980.
[11] Leonid G Khachiyan. Rounding of polytopes in the real number model of computation. Math-
ematics of Operations Research, 21(2):307–320, 1996.
[12] LG Khachiyan, SP Tarasov, and II Erlikh. The method of inscribed ellipsoids. In Soviet Math.
Dokl, volume 37, pages 226–230, 1988.
[13] Adam R Klivans and Daniel Spielman. Randomness efficient identity testing of multivariate
polynomials. In Proceedings of the thirty-third annual ACM symposium on Theory of computing,
pages 216–223. ACM, 2001.
[14] Ioannis Koutis, Gary L. Miller, and Richard Peng. A nearly-m log n time solver for sdd linear
systems. In Foundations of Computer Science (FOCS), 2011 IEEE 52nd Annual Symposium
on, pages 590 –598, oct. 2011.
[15] Yin Tat Lee and Aaron Sidford. Efficient accelerated coordinate descent methods and faster
In The 54th Annual Symposium on Foundations of
algorithms for solving linear systems.
Computer Science (FOCS), 2013.
[16] Yin Tat Lee and Aaron Sidford. Path finding ii: An\ o (m sqrt (n)) algorithm for the minimum
cost flow problem. arXiv preprint arXiv:1312.6713, 2013.
[17] Mu Li, Gary L Miller, and Richard Peng. Iterative row sampling. 2012.
[18] László Lovász and Santosh Vempala. Simulated annealing in convex bodies and an o*(n4)
volume algorithm. J. Comput. Syst. Sci., 72(2):392–417, 2006.
[19] Aleksander Madry. Navigating central path with electrical flows:
from flows to matchings,
and back. In Proceedings of the 54th Annual Symposium on Foundations of Computer Science,
2013.
[20] Michael W. Mahoney. Randomized algorithms for matrices and data. Foundations and Trends
in Machine Learning, 3(2):123–224, 2011.
[21] Nimrod Megiddo. Pathways to the optimal set in linear programming. In Nimrod Megiddo,
editor, Progress in Mathematical Programming, pages 131–158. Springer New York, 1989.
46
[22] Shinji Mizuno, Michael J Todd, and Yinyu Ye. On adaptive-step primal-dual interior-point
algorithms for linear programming. Mathematics of Operations research, 18(4):964–981, 1993.
[23] Murat Mut and Tamás Terlaky. A tight iteration-complexity upper bound for the mty predictor-
corrector algorithm via redundant klee-minty cubes. 2013.
[24] Jelani Nelson and Huy L Nguyên. Osnap: Faster numerical linear algebra algorithms via sparser
subspace embeddings. arXiv preprint arXiv:1211.1002, 2012.
[25] Yu Nesterov. Introductory Lectures on Convex Optimization: A Basic Course, volume I. 2003.
[26] Yu. Nesterov. Rounding of convex sets and efficient gradient methods for linear programming
problems. Optimization Methods Software, 23(1):109–128, February 2008.
[27] Yu Nesterov. Efficiency of coordinate descent methods on huge-scale optimization problems.
SIAM Journal on Optimization, 22(2):341–362, 2012.
[28] Yu E Nesterov and Michael J Todd. Self-scaled barriers and interior-point methods for convex
programming. Mathematics of Operations research, 22(1):1–42, 1997.
[29] Yurii Nesterov and Arkadii Semenovich Nemirovskii. Interior-point polynomial algorithms in
convex programming, volume 13. Society for Industrial and Applied Mathematics, 1994.
[30] Christos H Papadimitriou and Kenneth Steiglitz. Combinatorial optimization: algorithms and
complexity. Courier Dover Publications, 1998.
[31] James Renegar. A polynomial-time algorithm, based on newton’s method, for linear program-
ming. Mathematical Programming, 40(1-3):59–93, 1988.
[32] Daniel A Spielman and Nikhil Srivastava. Graph sparsification by effective resistances. SIAM
Journal on Computing, 40(6):1913–1926, 2011.
[33] Daniel A Spielman and Shang-Hua Teng. Nearly-linear time algorithms for graph partitioning,
graph sparsification, and solving linear systems. In Proceedings of the thirty-sixth annual ACM
symposium on Theory of computing, pages 81–90. ACM, 2004.
[34] Gilbert Strang. Inverse problems and derivatives of determinants. Archive for Rational Me-
chanics and Analysis, 114(3):255–265, 1991.
[35] Michael J Todd. Scaling, shifting and weighting in interior-point methods. Computational
Optimization and Applications, 3(4):305–315, 1994.
[36] Pravin M. Vaidya. A new algorithm for minimizing convex functions over convex sets (extended
abstract). In FOCS, pages 338–343, 1989.
[37] Pravin M Vaidya. Speeding-up linear programming using fast matrix multiplication. In Foun-
dations of Computer Science, 1989., 30th Annual Symposium on, pages 332–337. IEEE, 1989.
[38] Pravin M Vaidya. An algorithm for linear programming which requires o (((m+ n) n 2+(m+
n) 1.5 n) l) arithmetic operations. Mathematical Programming, 47(1-3):175–201, 1990.
47
[39] Pravin M. Vaidya. Reducing the parallel complexity of certain linear programming problems
(extended abstract). In FOCS, pages 583–589, 1990.
[40] Pravin M Vaidya. A new algorithm for minimizing convex functions over convex sets. Mathe-
matical Programming, 73(3):291–341, 1996.
[41] Pravin M Vaidya and David S Atkinson. A technique for bounding the number of iterations in
path following algorithms. Complexity in Numerical Optimization, pages 462–489, 1993.
[42] Virginia Vassilevska Williams. Multiplying matrices faster than coppersmith-winograd.
In
Proceedings of the forty-fourth annual ACM symposium on Theory of computing, pages 887–
898. ACM, 2012.
[43] Yinyu Ye. Interior point algorithms: theory and analysis, volume 44. John Wiley & Sons, 2011.
A Glossary
Here we summarize problem specific notation we use throughout the paper. For many quantities
we included the typical order of magnitude as they appear during our algorithms.
• Linear program related: constraint matrix A ∈ Rm×n , cost vector (cid:126)c ∈ Rn, constraint vector
(cid:126)b ∈ Rm, solution (cid:126)x ∈ Rn, weights of constraints (cid:126)w ∈ Rm where m is the number of constraints
and n is the number of variables.
• Bit complexity: L = log(m) + log(1 + dmax) + log(1 + max{(cid:13)(cid:13)(cid:126)c(cid:13)(cid:13)∞,(cid:13)(cid:13)(cid:126)b(cid:13)(cid:13)∞}) where dmax is the
largest absolute value of the determinant of a square sub-matrix of A.
G corresponds to (cid:126)g.
i∈[m] wi log s((cid:126)x)i.
• Slacks: (cid:126)s((cid:126)x) = A(cid:126)x − (cid:126)b.
• Matrix version of variables: S is the diagonal matrix corresponds to (cid:126)s, W corresponds to (cid:126)w,
• Newton step (4.2): (cid:126)ht((cid:126)x, (cid:126)w) = (∇2
(cid:126)x(cid:126)xft((cid:126)x, (cid:126)w))−1∇(cid:126)xft((cid:126)x, (cid:126)w) =(cid:0)AT S−1WS−1A(cid:1)−1(cid:0)t(cid:126)c − AT S−1 (cid:126)w(cid:1) .
• Penalized objective function (4.1): ft((cid:126)x, (cid:126)w) = t · (cid:126)cT (cid:126)x −(cid:80)
• Centrality (4.3): δt((cid:126)x, (cid:126)w) =(cid:13)(cid:13)(cid:126)ht((cid:126)x, (cid:126)w)(cid:13)(cid:13)∇2
(cid:13)(cid:13)PS−1A( (cid:126)w) ≈ 1.
• Properties of weight function (Def 7): size c1((cid:126)g) = (cid:13)(cid:13)(cid:126)g((cid:126)s)(cid:13)(cid:13)1 ≈ rank (A), slack sensitivity
(cid:1).
cγ((cid:126)g) = sup(cid:126)s γ((cid:126)s, (cid:126)g((cid:126)s)) ≈ 1, step consistency cr((cid:126)g) ≈ log(cid:0) m
• Slack Sensitivity(4.3): γ((cid:126)s, (cid:126)w) = maxi∈[m]
1
polylog(m) .
(cid:126)x(cid:126)xft((cid:126)x, (cid:126)w) ≈
(cid:13)(cid:13)W−1/2(cid:126)1i
rank A
• Difference between (cid:126)g and (cid:126)w (4.16): (cid:126)Ψ((cid:126)s, (cid:126)w) = log((cid:126)g((cid:126)s)) − log( (cid:126)w).
• Potential function for tracing 0 (Def 23): Φµ((cid:126)x) = eµx + e−µx ≈ poly(m).
48
• The weight function proposed (5.1):
(cid:126)g((cid:126)s) = arg min
>0
(cid:126)w∈Rm
f ((cid:126)s, (cid:126)w) where
where As = S−1A, α ≈ 1 − 1/ log2
B Technical Tools
log det(AT
s WαAs) − β
(cid:88)
i
log wi
f ((cid:126)s, (cid:126)w) = (cid:126)1T (cid:126)w − 1
α
(cid:16) m
rank(A)
(cid:17)
, β ≈ rank(A)/m.
In this section, we provide and prove various mathematical facts that we use throughout the paper.
B.1 Matrix Properties
First, we prove various properties regarding projection matrices that we use throughout the paper.
Lemma 32 (Projection Matrices). Let P ∈ Rn×n be an arbitrary projection matrix and let Σ =
diag(P). For all i, j ∈ [n] and (cid:126)x ∈ Rn we have the following
(1) Σii =(cid:80)
j∈[n] P(2)
ij ,
(2) 0 (cid:22) P(2) (cid:22) Σ (cid:22) I,
ij ≤ ΣiiΣjj,
i P(2)(cid:126)x ≤ Σii
(3) P(2)
(4) (cid:126)1T
Proof. To prove (1), we simply note that by definition of a projection matrix P = PP and therefore
(cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)Σ.
Σii = Pii = (cid:126)1T
i P(cid:126)1i = (cid:126)1T
i PP(cid:126)1i =
P2
ij =
P(2)
ij
(cid:88)
j∈[n]
(cid:88)
j∈[n]
To prove (2), we observe that since P is a projection matrix, all its eigenvectors are either 0
or 1. Therefore, Σ (cid:22) I and by (1) Σ − P(2) is diagonally dominant. Consequently, Σ − P(2) (cid:23) 0.
Rearranging terms and using the well known fact that the shur product of two positive semi-definite
matrices is positive semi-definite yields (2).
To prove (3), we use P = PP, Cauchy-Schwarz, and (1) to derive
(cid:88)
k∈[n]
Pij =
PikPkj ≤
(cid:12)(cid:12)(cid:12)(cid:126)1T
(cid:12)(cid:12)(cid:12) =
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:88)
j∈[n]
(cid:118)(cid:117)(cid:117)(cid:117)(cid:116)
P2
ik
k∈[n]
(cid:88)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤
ij (cid:126)xj
k∈[n]
(cid:88)
(cid:118)(cid:117)(cid:117)(cid:117)(cid:116)
(cid:88)
j∈[n]
49
=(cid:112)ΣiiΣjj
.
P2
kj
· (cid:88)
j∈[n]
Squaring then yields (3).
To prove (4), we note that by the definition of P(2) and Cauchy-Schwarz, we have
i P(2)(cid:126)x
P(2)
Σjj(cid:126)x2
j
P(4)
ij
Σjj
(B.1)
Now, by (1) and (3), we know that(cid:88)
Since(cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)Σ
(cid:113)(cid:80)
j∈[n] Σjj(cid:126)x2
P4
ij
Σjj
j∈[n]
def=
≤ (cid:88)
j∈[n]
P2
ijΣiiΣjj
Σjj
= Σii
(cid:88)
j∈[n]
j, combining (B.1) and (B.2) yields
i P(2)(cid:126)x
P2
ij = Σ2
ii
(cid:12)(cid:12)(cid:12)(cid:126)1T
(cid:12)(cid:12)(cid:12) ≤ Σii
(B.2)
(cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)Σ as desired.
B.2 Taylor Expansions and Multiplicative Approximations
Throughout this paper we use log((cid:126)a) − log((cid:126)b) as a convenient way of working with B−1((cid:126)a − (cid:126)b) or
A−1((cid:126)b − (cid:126)a). In this section we make this connection rigorous by providing several helper lemmas
used throughout the paper.
Lemma 33 (Log Notation). Suppose(cid:13)(cid:13) log((cid:126)a) − log((cid:126)b)(cid:13)(cid:13)∞ = ≤ 1/2 then
(cid:13)(cid:13)B−1((cid:126)a − (cid:126)b)(cid:13)(cid:13)∞ ≤ + 2.
If(cid:13)(cid:13)B−1((cid:126)a − (cid:126)b)(cid:13)(cid:13)∞ = ≤ 1/2, then(cid:13)(cid:13) log((cid:126)a) − log((cid:126)b)(cid:13)(cid:13)∞ ≤ + 2.
Proof. Using the Taylor expansion of ex and log(1 + x), we get the following two inequalities which
prove the claim
1 + x ≤ ex ≤ 1 + x + x2 for x ≤ 1
2
x − x2 ≤ log(1 + x) ≤ x for x ≤ 1
2
.
,
B.3 Matrix Calculus
Here, we derive various matrix calculus formulas used in Section 5. These are now somewhat
standard and also discussed in [40, 1] but we derive them here for completeness. In this section, we
define
RA( (cid:126)w)ij
def= (cid:126)aT
i (AT WA)−1(cid:126)aj.
We start by computing the derivative of the volumetric barrier function, f ( (cid:126)w) def= log det(AT WA).
Lemma 34 (Derivative of Volumetric Barrier). For A ∈ Rn×m, let f : Rm
>0 → R be given by
f ( (cid:126)w) def= log det(AT WA). Then the following holds
∀ (cid:126)w ∈ Rm
>0
: ∇f ( (cid:126)w) = diag(RA( (cid:126)w)) def= ΣA( (cid:126)w)W−1(cid:126)1.
50
(cid:104)
Proof. For all i ∈ [m] and (cid:126)w ∈ Rm, we know that
1
α
∂
∂ (cid:126)wi
Applying the matrix determinant lemma then yields that
f ( (cid:126)w + α(cid:126)1i) − f ( (cid:126)w)
f ( (cid:126)w) = lim
α→0
= lim
α→0
1
α
(cid:105)
(cid:2)log(cid:0)det(AT WA) · (1 + α(cid:126)aT
i ) − log det(AT WA)(cid:3) .
(cid:2)log det(AT WA + α(cid:126)ai(cid:126)aT
i (AT WA)−1(cid:126)ai)(cid:1) − log(cid:0)det(AT WA)(cid:1)(cid:3) .
f ( (cid:126)w) = lim
α→0
1
α
∂
∂ (cid:126)wi
Therefore,
∂
∂ (cid:126)wi
f ( (cid:126)w) = lim
α→0
log(1 + αR( (cid:126)w)ii)
α
= R( (cid:126)w)ii.
Next we bound the rate of change of entries of the resistance matrix.
Lemma 35 (Derivative of Effective Resistance). For all A ∈ Rm×n, (cid:126)w ∈ Rm
have
>0, and i, j, k ∈ [m] we
∂
∂ (cid:126)wk
[RA( (cid:126)w)]ij = −RA( (cid:126)w)ikRA( (cid:126)w)kj
where diag(RA( (cid:126)w)) def= ΣA( (cid:126)w)W−1(cid:126)1.
Proof. By definition, we have that
∂
∂ (cid:126)wk
RA( (cid:126)w)ij = lim
α→0
(cid:105)
R( (cid:126)w + α(cid:126)1k)ij − R( (cid:126)w)ij
(cid:104)
1
α
and
R( (cid:126)w + α(cid:126)1k)ij = (cid:126)1T
i A(AT WA + αAT (cid:126)1k(cid:126)1T
k A)−1AT (cid:126)1j
.
Furthermore, by applying the Sherman-Morrison formula, we know that
(AT WA + αAT (cid:126)1k(cid:126)1T
k A)−1 = (AT WA)+ − α(AT WA)−1AT (cid:126)1k(cid:126)1T
k A(AT WA)−1
1 + α(cid:126)1T
k A(AT WA)−1AT (cid:126)1k
(B.3)
(B.4)
.
(B.5)
Combining (B.3), (B.4), and (B.5) yields the result.
Finally, we use this to derive the Jacobian of leverage scores.
Lemma 36 (Derivative of Leverage Scores). For all A ∈ Rm×n, (cid:126)w ∈ Rm
>0 we have the following
J (cid:126)w((cid:126)σA( (cid:126)w)) = ΛA( (cid:126)w)W−1.
Proof. Since by definition (cid:126)σA( (cid:126)w)i = (cid:126)wiRA( (cid:126)w)ii by the previous lemma, we have that
∂
∂ (cid:126)wj
(cid:126)σA( (cid:126)w)i = (cid:126)1i=jR( (cid:126)w)ii − (cid:126)wiR( (cid:126)w)(2)
ij .
Writing this in matrix form and recalling the definition of the Jacobian then yields
J (cid:126)w((cid:126)σA( (cid:126)w)) = diag(RA( (cid:126)w)) − WRA( (cid:126)w)(2).
Right multiplying by I = WW−1 and recalling the definition of ΛA then yields the result.
51
C Projecting Onto Ball Intersect Box
In the algorithm centeringInexact, we need to compute
(cid:104)(cid:126)a, (cid:126)u(cid:105)
arg min
(cid:126)u∈U
we show how this can be computed in nearly linear time and in particular it can be computed in
where U = {(cid:126)x ∈ Rm (cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)W ≤ b and(cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)∞ ≤ c} for some (cid:126)w ≥ (cid:126)0, i.e. we need to project (cid:126)a onto the
intersection of the ball, (cid:8)(cid:126)x ∈ Rm (cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)W ≤ b(cid:9), and the box (cid:8)(cid:126)x ∈ Rm (cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)∞ ≤ c(cid:9). In this section
parallel in depth (cid:101)O(1) and work (cid:101)O(m).
(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:126)x
Note that by rescaling we can rewrite (C.1) as
arg max
≤1,−li≤xi≤li
(cid:104)(cid:126)a, (cid:126)x(cid:105)
(C.2)
(C.1)
for some li. Let us consider a simple algorithm which first ignore the box constraint and find the
best vector (cid:126)a.
If (cid:126)a does not violate any box constraint, then it is the solution. Otherwise, we
pick a most violated constraint i, i.e. the coordinate with highest ai /li. Then, we threshold this
coordinates and repeat the procedure on the remaining coordinate.
(cid:126)x = projectOntoBallBox((cid:126)a)
1. Set (cid:126)a = (cid:126)a/(cid:13)(cid:13)(cid:126)a(cid:13)(cid:13)2.
2. Sort the coordinate such that ai /li is in descending order.
sign ((cid:126)aj) lj
3. For i = 0,··· , m
(cid:114)
1−(cid:80)i
1−(cid:80)i
if j ∈ {1, 2,··· , i}
otherwise
3b. If (cid:126)x is a feasible solution, output (cid:126)x.
3a. Set (cid:126)x =
k=0 l2
k
k=0 a2
k
(cid:126)aj
.
Lemma 37. The algorithm projectOntoBallBox outputs a solution of the problem (C.2).
Proof. We claim that for all k ≤ i where i is the last step in the algorithm, we have
where Ω = {x : (cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)2 ≤ 1,−li ≤ xi ≤ li} and Ωk = Ω ∩ {x :
(cid:104)(cid:126)a, (cid:126)x(cid:105) = max
(cid:126)x∈Ωk
(cid:104)(cid:126)a, (cid:126)x(cid:105)
max
(cid:126)x∈Ω
(cid:126)x is feasible at the last step, we have
xi = li for i ∈ {1, 2,··· , k}}. Since
(cid:126)xlast = arg max
(cid:126)x∈Ωk
= arg max
(cid:126)x∈Ω
(cid:104)(cid:126)a, (cid:126)x(cid:105)
(cid:104)(cid:126)a, (cid:126)x(cid:105) .
Therefore, the correctness of the algorithm follows from the claim.
Now, we prove the claim by induction. The base case is trivial because Ω = Ω0. Now proceed
by contradiction and suppose that
(cid:104)(cid:126)a, (cid:126)x(cid:105) > max
(cid:126)x∈Ωk+1
max
(cid:126)x∈Ωk
(cid:104)(cid:126)a, (cid:126)x(cid:105) .
52
(C.3)
(cid:104)(cid:126)a, (cid:126)x(cid:105). If for all j > k, we have yj < lj. Then, the (cid:126)x found in the (k + 1)th
Let (cid:126)y = arg max(cid:126)x∈Ωk
iteration is exactly (cid:126)y and it is feasible and hence the algorithm outputs (cid:126)y. Otherwise, there is j
such that yj = lj. Since (cid:126)y /∈ Ωk+1, we have yk+1 < lk+1 and hence j > k + 1.
Consider
(cid:126)z(t) = (cid:126)y +
sign (yk+1) t
yk+1 +
where is a very small positive number. Note that d
dt
(cid:13)(cid:13)(cid:126)z(t)(cid:13)(cid:13)2 ≤ 1 for t > 0 but close to 0. Also, we have
(cid:126)1j
(cid:126)1k+1 − sign (yj) t
lj
(cid:13)(cid:13)(cid:126)z(t)(cid:13)(cid:13)2(cid:12)(cid:12)(cid:12)t=0
= 2
yk+1
yk+1+ − 2 < 0 and hence
d
dt
Take = lk+1 − yk+1, then we have
d
dt
(cid:104)(cid:126)a, (cid:126)z(cid:105) =
ak+1
yk+1 +
− aj
lj
.
(cid:104)(cid:126)a, (cid:126)z(cid:105) =
ak+1
lk+1
− aj
lj
> 0
because j > k + 1 and ai /li is in descending order. Therefore, (cid:126)z(t) is a feasible and better solution
for small positive t. Hence, it proves (cid:126)y is not the optimal solution of max(cid:126)x∈Ωk (cid:104)(cid:126)a, (cid:126)x(cid:105) that contradicts
to the definition of (cid:126)y.
Hence, max(cid:126)x∈Ω (cid:104)(cid:126)a, (cid:126)x(cid:105) = max(cid:126)x∈Ωk (cid:104)(cid:126)a, (cid:126)x(cid:105) and the algorithm outputs an optimal solution.
(cid:126)x = projectOntoBallBoxParallel((cid:126)a)
1. Set (cid:126)a = (cid:126)a/(cid:13)(cid:13)(cid:126)a(cid:13)(cid:13)2.
k and(cid:80)i
3. Precompute(cid:80)i
2. Sort the coordinate such that ai /li is in descending order.
4. Find the first i such that 1−(cid:80)i
1−(cid:80)i
sign ((cid:126)aj) lj
(cid:114)
1−(cid:80)i
1−(cid:80)i
k=0 a2
k=0 l2
k
k=0 a2
if j ∈ {1, 2,··· , i}
k
otherwise
k for all i.
.
5. Output (cid:126)x =
k=0 l2
(cid:126)aj
≤ l2
i+1
a2
i+1
.
k=0 l2
k
k=0 a2
k
The algorithm projectOntoBallBoxParallel is a parallel and more efficient version projectOntoBallBox.
All other operations in our algorithm are standard linear algebra and hence the following theorem
shows that our linear programming solver is indeed parallelizable.
Lemma 38. The algorithm projectOntoBallBoxParallel outputs an solution of the optimization
problem (C.2) in depth O(1) and work O(m).
Proof. Note that in the algorithm projectOntoBallBox, the value
is increasing through the algorithm. To see this, note that in step 3b, if (cid:126)x is not feasible, that means
there is j such that
k=0 l2
k
k=0 a2
k
1 −(cid:80)i
1 −(cid:80)i
1 −(cid:80)i
1 −(cid:80)i
53
k=0 l2
k
k=0 a2
k
>
l2
j
a2
j
.
Since ai/li is in descending order, j = i + 1. Therefore, we have
k=0 l2
k
k=0 a2
k
1 −(cid:80)i
1 −(cid:80)i
1 −(cid:80)i+1
1 −(cid:80)i+1
k=0 l2
k
k=0 a2
k
l2
i+1
a2
i+1
.
>
1 −(cid:80)i
1 −(cid:80)i
k=0 l2
k
k=0 a2
k
.
>
Hence, we have
Using this fact, it is easy to see the algorithm projectOntoBallBoxParallel and the algorithm
projectOntoBallBox outputs the same vector. Obviously, all steps can be computed in depth O(1)
and work O(m).
D Inexact Linear Algebra
Throughout much of our analysis of weighted path following we assumed that linear systems in
A could be solved exactly.
In this section we relax this assumption and discuss the effect of
using inexact linear algebra in our linear programming algorithms. We show that rather than
computing(cid:0)AT DA(cid:1)−1 (cid:126)x precisely for positive diagonal matrix D it suffices to solve these systems
approximately.
Throughout this section we assume that for any matrix A ∈ Rn×m and vector (cid:126)b ∈ Rm there is
an algorithm solve(A,(cid:126)b) which outputs an vector (cid:126)x such that
(D.1)
Since A is full rank, we can write (cid:126)c = AT (cid:126)d for some (cid:126)d. From equation (4.2), the Newton step is
(cid:13)(cid:13)(cid:126)x − A+(cid:126)b(cid:13)(cid:13)AT A ≤ (cid:13)(cid:13)A+(cid:126)b(cid:13)(cid:13)AT A.
(cid:32)
√
(cid:33)
√
(cid:126)w
−
(cid:126)s(cid:126)d√
(cid:126)w
t
(cid:126)ht((cid:126)x, (cid:126)w) = (AT S−1WS−1A)−1AT S−1
√
(cid:126)w
WS−1A
(cid:16)√
(cid:32)
=
t
(cid:17)+
(cid:33)
.
W
−
(cid:126)s(cid:126)d√
(cid:126)w
(cid:33)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)AT S−1WS−1A
(cid:32)√
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)solve
WS−1A, t
√
(cid:126)w
−
(cid:126)s(cid:126)d√
(cid:126)w
− (cid:126)ht
Suppose that we compute (cid:126)ht by the algorithm solve above, then we have
≤ (cid:13)(cid:13)(cid:126)ht
(cid:13)(cid:13)AT S−1WS−1A
= δt ((cid:126)x, (cid:126)w) .
Hence, the outcome of solve differs from the Newton step (cid:126)ht by a relative small amount in
(cid:107)·(cid:107)AT S−1WS−1A. Hence, it suffices to prove that δt is stable under this small amount in (cid:107)·(cid:107)AT S−1WS−1A
and hence is the algorithm solve will only increase δ by a little compared with using exact linear
algebra.
(cid:13)(cid:13)(cid:13)(cid:126)∆
(cid:13)(cid:13)(cid:13)AT S−1WS−1A
≤ 1
8γ . Then, we have
Lemma 39. Let γ def= γ((cid:126)x, (cid:126)w) and (cid:126)x(new) = (cid:126)x + (cid:126)∆. Let η =
(cid:16)
(cid:17) ≤ (1 − γη)
δt
(cid:126)x(new), (cid:126)w
−1 (δt ((cid:126)x, (cid:126)w) + η) .
54
Proof. By the same proof in Lemma 5, we have that
(cid:16)
Therefore, we have
δt
(cid:126)x(new), (cid:126)w
=
(cid:17)
(new)A
(new) (cid:126)w
AT S−1
(new)WS−1
(new) (cid:126)w
≤ (1 + γη)
(cid:13)(cid:13)S−1((cid:126)s(new) − (cid:126)s)(cid:13)(cid:13)∞ ≤ γη.
(cid:13)(cid:13)(cid:13)t(cid:126)c − AT S−1
(cid:13)(cid:13)(cid:13)(cid:16)
(cid:17)−1
(cid:13)(cid:13)(cid:13)t(cid:126)c − AT S−1
(cid:13)(cid:13)(cid:13)(cid:16)
(cid:13)(cid:13)(cid:13)(cid:13)AT
≤ (1 + γη)(cid:13)(cid:13)t(cid:126)c − AT S−1 (cid:126)w(cid:13)(cid:13)(AT S−1WS−1A)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)W
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:126)s(new) − (cid:126)s
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:126)s(new) − (cid:126)s
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)W
(cid:13)(cid:13)S−1((cid:126)s(new) − (cid:126)s)(cid:13)(cid:13)W ≤ η.
≤ (1 + γη) δt ((cid:126)x, (cid:126)w) + (1 − γη)−1
= (1 + γη) δt ((cid:126)x, (cid:126)w) +
(new)WS−1
(cid:17)−1
AT S−1
(cid:126)s(new)
(cid:126)s
(new)A
−1 +
.
By the same proof in Lemma 5, we have that
Thus, we have the result.
(cid:18) (cid:126)w
(cid:126)s
(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)(AT S−1WS−1A)
−1
− (cid:126)w
(cid:126)s(new)
Therefore, as long as we choose small enough, the algorithm solve gives an accurate enough
(cid:126)x(new) for the centering step. Similarly, it is easy to see that it also gives accurate enough (cid:126)w(new)
because the error of (cid:126)w(new) due to solve is small in(cid:13)(cid:13) ·(cid:13)(cid:13)W norm and the tracing 0 game can afford
for this error.
At last, we need to check solve gives us a way to compute weight function. Since the weight
function computation relies on the function computeLeverageScores, we only need to know if we
can compute (cid:126)l in the computeLeverageScores with high enough accuracy. Now, we use the notation
is the computeLeverageScores. Without loss of generality, we can assume X = I. Let (cid:126)l(apx) and
(cid:126)p(apx) be the approximate (cid:126)l and (cid:126)p computed by the algorithm Solve. Then, we have
(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)(cid:126)l(j)(cid:17)(apx) − (AT A)+AT (cid:126)q(j)
(cid:13)(cid:13)(cid:13)(cid:13)AT A
Hence, for any i, j, we have(cid:13)(cid:13)(cid:13)(cid:13)(cid:126)p(j)
i −(cid:16)
(cid:126)p(apx)
i
(cid:17)(j)(cid:13)(cid:13)(cid:13)(cid:13)∞
≤
=
(cid:13)(cid:13)(cid:13)(cid:13)AT A
=
−1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)(cid:126)l(j)(cid:17)(apx) − A+(cid:126)q(j)
≤ (cid:13)(cid:13)A+(cid:126)q(j)(cid:13)(cid:13)AT A
= (cid:13)(cid:13)AT (cid:126)q(j)(cid:13)(cid:13)(AT A)
(cid:114) n
≤ (cid:13)(cid:13)(cid:126)q(j)(cid:13)(cid:13)2 ≤
(cid:126)p(apx)(cid:17)(j)(cid:13)(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:126)p(j) −(cid:16)
(cid:13)(cid:13)(cid:13)(cid:13)A
(cid:18)(cid:16)(cid:126)l(j)(cid:17)(apx) − (cid:126)l(j)
(cid:114) n
(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)2
k
.
≤
.
k
55
Therefore, we have(cid:118)(cid:117)(cid:117)(cid:116) k(cid:88)
Therefore, if ≤(cid:113)
j=1
(cid:16)
(cid:126)p(j)
i
(cid:17)2 −
(cid:118)(cid:117)(cid:117)(cid:116) k(cid:88)
j=1
(cid:18)(cid:16)
(cid:126)p(apx)
i
(cid:17)(j)(cid:19)2 ≤
(cid:118)(cid:117)(cid:117)(cid:116) k(cid:88)
√
≤
j=1
nk.
(cid:18)
i −(cid:16)
(cid:126)p(j)
(cid:126)p(apx)
i
(cid:17)(j)(cid:19)2
1
mpolylog(m), the error is small enough for computeLeverageScores.
E Bit Complexity and Linear Program Reductions
In this section, we show how to reduce solving an arbitrary linear program to finding a low cost
solution in a bounded linear program for which we have an explicit interior point. Throughout this
section letA ∈ Rm×n, (cid:126)b ∈ Rm, (cid:126)c ∈ Rn, and consider the following general linear program
min
(cid:126)x∈Rn : A(cid:126)x≥(cid:126)b
(cid:126)cT (cid:126)x
(E.1)
We assume that the entries of A, (cid:126)b, and (cid:126)c are integers and we let OP T denote the optimal value of
(E.1) and we let L denote the bit complexity of (E.1) where
L def= log(m) + log(1 + dmax(A)) + log(1 + max{(cid:13)(cid:13)(cid:126)c(cid:13)(cid:13)∞,(cid:13)(cid:13)(cid:126)b(cid:13)(cid:13)∞})
and dmax(A) denotes the largest absolute value of the determinant of a square sub-matrix of A.
Our goal is to efficiently transform (E.1) to a linear program of the same form
(cid:126)x∈Rn(cid:48)
min
: A(cid:48)(cid:126)x≥(cid:126)b(cid:48)
(cid:126)c(cid:48)T (cid:126)x
(E.2)
where A(cid:48) ∈ Rm(cid:48)×n(cid:48)
, (cid:126)b(cid:48) ∈ Rm(cid:48), and (cid:126)c(cid:48) ∈ Rn(cid:48) are integer, and nnz(A(cid:48)), n(cid:48), m(cid:48), and the bit complexity
of (E.2) denoted, L(cid:48), are comparable to nnz(A), n, m, and L. Furthermore, we require that (E.2)
is bounded, has an explicit efficiently computable interior point, and that we can convert any low
cost feasible solution to a solution of (E.1) in linear time.
While there are standard tools to perform reductions to ensure that (E.1) is bounded and has an
explicit initial feasible point or to ensure that the optimal integral solution can be easily computed
explicitly, we need to particularly careful when using these reductions to ensure that nnz(A), n,
and m are not increased significantly. As the running times of our path following techniques in
Section (7) depend crucially on these parameters in this section we prove the following Lemma
claiming that such an efficient reduction is possible.
, (cid:126)b(cid:48) ∈ Rm(cid:48), (cid:126)c(cid:48) ∈ Rn(cid:48),
Lemma 40. In O(nnz(A) + n + m) time we can compute integer A(cid:48) ∈ Rm(cid:48)×n(cid:48)
(cid:126)x(cid:48) ∈ Rm(cid:48). Such that nnz(A(cid:48)) = O(nnz(A) + n + m), n(cid:48) = O(n), m(cid:48) = O(m), A(cid:48)(cid:126)x(cid:48) ≥ (cid:126)b(cid:48), and (E.2)
is bounded and has bit complexity at most 12L1 + 7 log(20n). Furthermore, if we can find a feasible
point in (E.2) such that the cost of that point is at most the OP T + 2−12(L+log(20n)) where OP T is
the value of (E.2) then we can either
1. Find the active constraints of a basic feasible optimal solution (E.1) using only one matrix
vector multiplication by A; or
56
2. Prove that (E.1) is infeasible or unbounded.
We break this proof into two parts. First in Lemma 41 we show how to transform (E.1) so that
the linear program is bounded and has an explicit feasible point. Then in Lemma 43 we follow the
approach of [3] and show that we can perturb the cost of a linear program to make the optimal
solution unique and thereby make it easy to compute an exact integral solution.
Lemma 41. Consider the following modified linear program
min (cid:126)cT (cid:126)x + n23L+4z such that A(cid:126)x + z(cid:126)1 ≥ (cid:126)b, 2L+1 ≥ z ≥ 0, 2L+1(cid:126)1 ≥ (cid:126)x ≥ −2L+1(cid:126)1
(E.3)
where A, (cid:126)b, and (cid:126)c are as in (E.1) and L is the bit complexity of (E.1). (E.3) is bounded with
an explicit interior point (cid:126)x = 0, z = 2L + 1. Furthermore, (E.1) is bounded and feasible with an
optimal solution (cid:126)x if and only if ((cid:126)x, 0) is an optimal solution of (E.3) with 2L ≥ xi ≥ −2L, (E.1)
is unbounded if and only if there is a basic feasible solution, ((cid:126)x, z), of (E.3) with xi > 2L for
some i, and (E.1) is infeasible if and only if there is basic feasible solution, ((cid:126)x, z), of (E.3) with
(cid:126)z (cid:54)= 0. Furthermore, (E.3) can be written in the form (E.2) such that all these properties hold with
nnz(A(cid:48)) = O(nnz(A) + n + m), n(cid:48) = O(n), m(cid:48) = O(m), and L(cid:48) ≤ 4L + 2 log(16n).
Proof. Case 1: Suppose (E.1) is bounded and feasible. It is known that any basic feasible solution
of (E.1) is a vector of rational numbers with both absolute value of numerator and denominator
are bounded by 2L [30]. Therefore, −n22L ≤ OP T ≤ n22L. Given any feasible solution (cid:126)x of (E.1),
the point ((cid:126)x, z = 0) is a feasible solution of (E.3) with same cost value. Hence, the linear program
(E.3) is feasible and the optimal value of (E.3) is at most n22L.
On the other hand, clearly (E.3) is feasible because (cid:126)x = (cid:126)0, z = 2L + 1 is an interior point.
Furthermore, (E.3) is bounded and therefore has some optimal value. Consider any optimal basic
feasible solution ((cid:126)x, z) of (E.3), we have (cid:126)cT (cid:126)x is between −n22L+1 and n22L+1. Also, z is a rational
number with the absolute value of denominator are bounded by 2L using Cramer’s rule. Therefore,
we have z ≥ 2−L1 or z = 0. If z ≥ 2−L, then the total cost is at least n23L+42−L − n22L+1 > n22L.
However, as we argued above, the optimal value of (E.3) is at most n22L. Therefore, optimal
solution has z = 0 and 2L ≥ xi ≥ −2L for all i.
Case 2: Suppose (E.1) is not feasible. In this case, any feasible point ((cid:126)x, z) in (E.3) has z (cid:54)= 0
and by the reasoning in the previous section any basic feasible solution has cost greater than n22L.
Case 3: Suppose (E.1) is not bounded. Let OPTk = min (cid:126)cT (cid:126)x such that A(cid:126)x ≥ (cid:126)b, k + 2L ≥ xi ≥
−2L−k. Thus, we have OPT1 < OPT0 and any optimal point of the case k = 1 has some coordinate
larger than 2L or smaller −2L. By similar argument as above, we have that the optimal point of
(E.3) is of the form ((cid:126)x, 0) and some coordinate of (cid:126)x is larger than 2L or smaller −2L.
To compute the bit complexity of (E.3) note that we can write (E.3) in the form of (E.2) by
where I ∈ Rm×m and(cid:126)0 ∈ Rm (E.4)
choosing
A(cid:48) =
A (cid:126)1
(cid:126)0
I
−I
(cid:126)0
(cid:126)0T
1
(cid:126)0T −1
, (cid:126)b(cid:48) =
(cid:126)b
−2L1+1
2L1+1
0
2L1+1
(cid:126)c
(cid:19)
(cid:18)
, (cid:126)c(cid:48) =
(cid:0)(cid:2) A (cid:126)1 (cid:3)(cid:1) ≤ n · dmax (A) .
n23L+4
Thus n(cid:48) = n + 1, m(cid:48) = 3m + 2, and it is easy to see that
dmax(A(cid:48)) = dmax
57
Therefore, the bit complexity of (E.3) is at most log(1 + ndmax(A)) + log(1 + n23L+4) ≤ 4L +
2 log(16n) as desired.
Following the approach in [3] to use the following isolation lemma, we show that it is possible
to transform the linear program into one with unique optimal solution by randomly perturbing the
cost function.
Lemma 42 ([13]). Given any collection of linear functions on n variables c1, c2,··· , cn with coeffi-
cients in the range {−K,−K − 1,··· , K − 1, K}. If c1,··· , cn are independently chosen uniformly
at random in {−2Kn,··· , 2Kn}. Then, with probability greater than 1
2, there is a unique linear
function of minimum value at c1, c2,··· , cn.
Note that for we can think every vertex (cid:126)x is a linear function (cid:126)cT (cid:126)x on the cost variables (cid:126)c.
Although there are exponentially many vertices, the above lemma shows that the minimizer is
attained at a unique vertex (linear function).
Lemma 43. Suppose that (E.1) is feasible and bounded and consider the following modified linear
program
min(cid:0)22L+3n(cid:126)c + (cid:126)r(cid:1)T
(cid:126)x given A(cid:126)x ≥ (cid:126)b.
(E.5)
where each coordinate in (cid:126)r ∈ Rm is chosen uniformly at random from the integers {−2L+1n,··· , 2L+1n}.
Let OP T ’ denote the optimal value of the linear program (E.5). Given any feasible solution for
the linear program (E.5) with cost less than OP T + n−12−3L−2, we can find the active constraints
of a basic feasible optimal solution of (E.1) by using only one matrix vector multiplication with A.
Furthermore, the bit complexity of (E.5) is at most 3L + log(8n).
Proof. Since the set of basic solutions to (E.5) and (E.1) are the same, we know that any basic
feasible solution of (E.5) is a vector of rational numbers with absolute value of numerator and
denominator both bounded by 2L. Consequently our perturbation of the cost function maintains
that an optimum solution to (E.5) is an optimal solution to (E.1). Hence, the Isolation Lemma
shows that with probability greater than 1
Now consider the polytope Pt = {(cid:126)x such that A(cid:126)x ≥ (cid:126)b and (cid:0)22L+3n(cid:126)c + (cid:126)r(cid:1)T (cid:126)x ≤ OP T +t2−2L−1}
(cid:126)x is in the polytope of {A(cid:126)x ≥ (cid:126)b} and hence (cid:13)(cid:13)(cid:126)x(cid:13)(cid:13)∞ ≤ 2L. Therefore, for any (cid:126)x ∈ Pt, we have
(cid:13)(cid:13)(cid:126)x − (cid:126)x∗(cid:13)(cid:13)∞ ≤ t · 2L+1 for any t ≤ 1. Therefore, for any (cid:126)x ∈ Pt, (cid:13)(cid:13)A(cid:126)x − A(cid:126)x∗(cid:13)(cid:13)∞ ≤ nt22L+1. Since
for t > 0. Since (E.5) has a unique solution, by a similar argument as before, P1 contains only one
basic feasible solution of (E.5) and hence Pt − (cid:126)x∗ = t (P1 − (cid:126)x∗) for any t ≤ 1. Also, for any (cid:126)x ∈ P1,
2, the linear program (E.5) has a unique solution (cid:126)x∗.
A(cid:126)x∗ is a vector of rational numbers with the absolute value of denominator are bounded by 2L, we
can distinguish if a constraint is satisfied or not when nt22L+1 < 2−L−1.
Combining Lemma 41 and Lemma 43 proves Lemma 40.
F Numerical Linear Algebra for Acceleration
Here we prove Theorem 30 needed for the accelerated linear program solver. Below we restate the
theorem for convenience.
58
Theorem 44. Let (cid:126)d(i) ∈ Rm
that d(i)
j
given the (cid:126)d(i) in a sequence, in each iteration i we can compute (cid:0)AT DiA(cid:1)−1 (cid:126)xi for Di = diag((cid:126)di)
>0 be a sequence of r positive vectors. Suppose that the number of times
for any i ∈ [r] and j ∈ [m] is bounded by Cr2 for some C ≥ 1. Then if we are
(cid:54)= d(i+1)
j
and arbitrary (cid:126)xi ∈ Rn with the average cost per iteration
(cid:18) mnω−1
r
(cid:101)O
(cid:19)
+ n2 + Cωr2ω + Cω−1nrω
where ω < 2.3729 [42] is the matrix multiplication constant.
Proof. For all i ∈ [r] let Bi = AT DiA. Since D1 ∈ Rm×m is diagonal and A ∈ Rn×m we can
compute D1A trivially in O(mn) time. Furthermore from this we can compute B1 = AT D1A in
O(mnω−1) time using fast matrix multiplication by splitting A into m
n blocks of size n and using that
m > n. Furthermore, using fast matrix multiplication we can then compute B−1
in O(nω) time and
similarly we can compute B−1
1 AT in O(mnω−1) time. Now, we show how to use this computation
of B−1
1 AT in O(mnω−1) time to decrease the running time of future iterations.
1 and B−1
For all k > 1, let Dk = D1 + ∆k for some diagonal ∆k ∈ Rm×m and let rk
def= nnz(∆k). Let
Pk ∈ Rrk×n be the 1 − 0 matrix that selects the rows of A for which the diagonal entry in ∆k is
nonzero, let Sk ∈ Rrk×rk be the diagonal matrix whose diagonal entries are the non-zero diagonal
entries of ∆k and Ak
1
def= PkA.
k SkPk and hence by the Woodbury matrix identity, we have
B−1
Note that ∆k = PT
= B−1
Assume we have computed AkB−1
compute(cid:0)S−1
1 AT
k
(F.1)
k ∈ Rrk×rk explicitly, we can use fast matrix multiplication to
i (cid:126)xi in just
k ). Then, we can use (F.1) to compute B−1
1
AkB−1
1 AT
k
k + AkB−1
1 − B−1
k AT
k SkPkA(cid:1)−1
i =(cid:0)AT D1A + AT PT
(cid:0)S−1
(cid:1)−1 in time O(rω
(cid:16)(cid:0)S−1
(cid:1) + nnz(Ak) + nnz
k + AkB−1
(cid:18) mnω−1
(cid:19)
i (cid:126)xi is
k AT
k
1
(cid:18) mnω−1
+ n2 + nrk + rω
k
= (cid:101)O
r
(cid:101)O
(cid:1)−1
(cid:1)−1(cid:17)(cid:17)
O
1 AT
k
time. Consequently, not counting the time to compute AkB−1
average cost of computing B−1
= O(nrk + n2)
k AT
k ∈ Rrk×rk, we have that the
(cid:16)
k + AkB−1
nnz(cid:0)B−1
(cid:19)
+ n2 + Cωr2ω
r
(F.2)
because rk ≤ Cr2 and nrk ≤ 2n2 + 2r2
k.
All that remains is to estimate the cost of computing AkB−1
order the rows of A such that AT
this, to compute AkB−1
k AT
k we see that it suffices to compute
k = [AT
k−1 RT
(cid:18) AkB−1
UkB−1
k AkB−1
1 AT
1 Ak UkB−1
1 UT
k
1 UT
k
(cid:19)
k . For notational simplicity, we
k ] where Rk ∈ Ruk×n where uk = rk − rk−1. From
k AT
Now, since we precomputed B−1
compute B−1
k by extracting columns from B−1
1 AT and U is just a subset of the rows of A, we see that we can
1 AT . Thus, we see that the time to compute
1 UT
59
k Ak is dominated by the time to multiply a matrix of size at most rk × n and n × uk. We
matrices of size uk × uk which can be done in O(rknuω−2
· n
)
k
k B−1
AT
can do this by multiplying O
time. Thus the average cost of computing AT
(cid:17)
(cid:16) rk
(cid:19)
·(cid:0)rknuω−2
where we used the fact that since (cid:80)
(cid:88)
(cid:18) 1
1≤k<r
O
uk
uk
r
achieve when each uk = Cr.
k B−1
k Ak is
(cid:1) ≤ O(Crn · r · (Cr)ω−2) = O(Cω−1nrω)
k uk = rk, rk ≤ Cr2 and the minimum value of (cid:80)
k
k uω−2
k
is
60
|
1703.07964 | 1 | 1703 | 2017-03-23T08:37:54 | Minimum Cuts and Shortest Cycles in Directed Planar Graphs via Noncrossing Shortest Paths | [
"cs.DS"
] | Let $G$ be an $n$-node simple directed planar graph with nonnegative edge weights. We study the fundamental problems of computing (1) a global cut of $G$ with minimum weight and (2) a~cycle of $G$ with minimum weight. The best previously known algorithm for the former problem, running in $O(n\log^3 n)$ time, can be obtained from the algorithm of \Lacki, Nussbaum, Sankowski, and Wulff-Nilsen for single-source all-sinks maximum flows. The best previously known result for the latter problem is the $O(n\log^3 n)$-time algorithm of Wulff-Nilsen. By exploiting duality between the two problems in planar graphs, we solve both problems in $O(n\log n\log\log n)$ time via a divide-and-conquer algorithm that finds a shortest non-degenerate cycle. The kernel of our result is an $O(n\log\log n)$-time algorithm for computing noncrossing shortest paths among nodes well ordered on a common face of a directed plane graph, which is extended from the algorithm of Italiano, Nussbaum, Sankowski, and Wulff-Nilsen for an undirected plane graph. | cs.DS | cs |
Minimum Cuts and Shortest Cycles in Directed Planar
Graphs via Noncrossing Shortest Paths∗
Hung-Chun Liang†
Hsueh-I Lu‡
April 27, 2018
Abstract
Let G be an n-node simple directed planar graph with nonnegative edge weights. We
study the fundamental problems of computing (1) a global cut of G with minimum weight
and (2) a cycle of G with minimum weight. The best previously known algorithm for the
former problem, running in O(n log3 n) time, can be obtained from the algorithm of Ł acki,
Nussbaum, Sankowski, and Wulff-Nilsen for single-source all-sinks maximum flows. The
best previously known result for the latter problem is the O(n log3 n)-time algorithm of
Wulff-Nilsen. By exploiting duality between the two problems in planar graphs, we solve
both problems in O(n log n log log n) time via a divide-and-conquer algorithm that finds a
shortest non-degenerate cycle. The kernel of our result is an O(n log log n)-time algorithm
for computing noncrossing shortest paths among nodes well ordered on a common face
of a directed plane graph, which is extended from the algorithm of Italiano, Nussbaum,
Sankowski, and Wulff-Nilsen for an undirected plane graph.
1 Introduction
Let G be an n-node m-edge simple graph with nonnegative edge weights. G is unweighted if
the weights of all edges of G are identical. Let C be a subgraph of G. The weight w(C) of C
is the sum of edge weights of C. Let G \ C denote the graph obtained from G by deleting the
edges of C. Paths are allowed to repeat nodes throughout the paper. For nodes s and t, an
st-path of G is a path of G from s to t and an st-cut of G is a subgraph C of G such that there
are no st-paths in G \ C. A (global) cut of G is an st-cut of G for some nodes s and t of G. A
cycle of G is an ss-path of G for some node s of G.
• The minimum-cut problem on G seeks a cut of G with minimum weight. For instance,
the v1v3-cut consisting of edge v2v3 is the minimum cut of the graph in Figure 1(a). The
best known algorithm on directed G, due to Hao and Orlin [30], runs in O(mn log n2
m )
time. On undirected G, Nagamochi and Ibaraki [54] and Stoer and Wagner [61] solved
the problem in O(mn + n2 log n) time and Karger [37] solved the problem in expected
∗A preliminary version of this paper appeared as the master's thesis of the first author [45]. The journal version
appeared in SIAM Journal on Discrete Mathematics [46].
†Graduate Institute of Computer Science and Information Engineering, National Taiwan University. Email:
[email protected].
‡Corresponding author. Department of Computer Science and Information Engineering, National Taiwan Uni-
versity. This author also holds joint appointments in the Graduate Institute of Networking and Multimedia and the
Graduate Institute of Biomedical Electronics and Bioinformatics, National Taiwan University. Address: 1 Roosevelt
Road, Section 4, Taipei 106, Taiwan, ROC. Research of this author is supported in part by MOST grant 104–2221–
E–002–044–MY3. Email: [email protected]. Web: www.csie.ntu.edu.tw/~hil.
1
v4
6
v2
9
2
7
8
5
v1
v3
1
v4
6
⋆
⋆
⋆
9
f2
⋆
v1
f1
2
⋆
v2
7
f3
8
v3
(a) G
(b) G△
f2
7
⋆
6
5
1
f4
⋆
⋆
f3
⋆
f1
8
2
(c) G∗
△
5
1
9
⋆
f4
Figure 1: (a) A simple planar graph G. (b) A simple bidirected plane graph G△ obtained from
G by adding edges with weights ⋆ = 0 (respectively, ⋆ = ∞) if we are seeking a minimum cut
(respectively, shortest cycle) of G. (c) The dual of G△.
O(m log3 n) time. Kawarabayashi and Thorup [38] recently announced the first known
o(mn)-time algorithm on undirected unweighted G, improving upon the algorithm of
Gabow [24] designed twenty years ago.
• The shortest-cycle problem on G seeks a cycle of G with minimum weight. For instance,
cycle v2v3v2 with weight 6 is the shortest cycle of the graph in Figure 1(a). Since a short-
est directed cycle containing edge ts is obtainable from a shortest st-path, the problem
on directed graphs can be reduced to computing all-pairs shortest paths in, e.g., O(mn +
n2 log n) time [10]. Vassilevska Williams and Williams [65] argued that finding a truly
subcubic algorithm for the problem might be hard. For directed (respectively, undi-
rected) unweighted G, Itai and Rodeh [32] solved the problem in O(µ(n) log n) (respec-
tively, O(min(mn, µ(n)))) time, where µ(n) = O(n2.373) [64] is the time for multiplying
two n × n matrices.
If G is undirected and planar, Chalermsook, Fakcharoenphol, and Nanongkai [8] showed that
the time complexity of both aforementioned problems on G is O(log n) times that of finding an
st-cut of G with minimum weight for any given nodes s and t. Plugging in the O(n log n)-time
algorithms, e.g., of Frederickson [23], Borradaile and Klein [1], and Erickson [14], the reduction
of Chalermsook et al. solved both problems in O(n log2 n) time. Plugging in the O(n log log n)-
time algorithm of Italiano, Nussbaum, Sankowski, and Wulff-Nilsen [33], the reduction of
Chalermsook et al. solved both problems in O(n log n log log n) time. The best known result for
both problems on G is the O(n log log n)-time algorithm of Ł acki and Sankowski [44], relying
upon the st-cut oracle of Italiano et al. [33].
This paper addresses both problems for the case that G is directed and planar. While the
minimum-cut problem has been thoroughly studied for undirected planar graphs, surprisingly
no prior work is specifically for directed planar graphs. Djidjev [12] claimed that his technique
for unweighted undirected planar graphs solves the shortest-cycle problem on unweighted
directed planar G in O(n3/2) time and left open the problem of finding a shortest cycle in un-
weighted directed planar G in o(n3/2) time. Weimann and Yuster [67] gave an O(n3/2)-time al-
gorithm for the shortest-cycle problem, which should be adjustable to solve the minimum-cut
problem also in O(n3/2) time (via similar techniques to our proof for Lemma 4.2 in §4 to handle
degeneracy in shortest cycles). Wulff-Nilsen [68] reduced the time for the shortest-cycle prob-
lem on G to O(n log3 n), but it is unclear how to adjust his algorithm to solve the minimum-cut
problem without increasing the required time by too much. The algorithm of Ł acki, Nuss-
2
baum, Sankowski, and Wulff-Nilsen [43] for single-source all-sinks maximum flows solves the
minimum-cut problem on directed planar G in O(n log3 n) time. Below is our result:
Theorem 1.1. It takes O(n log n log log n) time to solve the minimum-cut and shortest-cycle problems
on an n-node simple directed planar graph with nonnegative edge weights.
As pointed out by anonymous reviewers, Mozes, Nikolaev, Nussbaum, and Weimann [51]
recently announced an O(n log log n)-time algorithms for the minimum-cut problem. How-
ever, unlike our Theorem 1.1, their algorithm requires the condition that there is a unique
shortest path between any two nodes. For general directed planar graphs with nonnegative
edge weights, they apply an isolation lemma [50, 53] to perturb the edge weights to meet the
condition with high probability. Thus, their results are Monte Carlo randomized algorithms.
Related work
The only known nontrivial linear-time algorithm for the shortest-cycle problem, due to Chang
and Lu [9], works on undirected unweighted planar graphs. For undirected G, if G is embed-
ded on an orientable surface of genus g, Erickson, Fox, and Nayyeri [15] solved the problem in
gO(g)n log log n time, based on the algorithm of Ł acki and Sankowski [44] for undirected planar
graphs. If G is undirected and unweighted and is 2-cell embedded on an orientable surface of
genus g = O(nα) with 0 < α < 1, Djidjev [12] solved the problem in O(g3/4n5/4 log n) time.
On undirected unweighted O(1)-genus G, Weimann and Yuster [67] solved the problem in
O(n log n) time. For directed planar G, even if G is unweighted, our Theorem 1.1 remains
the best known algorithm. If G is unweighted and embedded on a genus-g surface, the tech-
nique of Djidjev [12] solved the problem in O(g1/2n3/2) time. The shortest-cycle problem on G
with negative edge weights can be reduced to one with nonnegative edge weights using the
standard reweighting technique via a shortest-path tree in G (e.g., [20, 25, 27, 42, 52]). Cygan,
Gabow, and Sankowski [11] studied the problem on graphs whose edge weights are bounded
integers. Yuster [69] studied the version on undirected G asking for each node a shortest cycle
containing the node. See e.g., [5, 6, 7, 16, 19, 21, 22] for algorithms that compute shortest cy-
cles with prescribed topological properties. See, e.g., [32, 47, 49, 59, 60, 70] for approximation
algorithms of the shortest-cycle problem.
The closely related problem that seeks a minimum st-cut for given nodes s and t and its
dual problem that seeks a maximum st-flow have been extensively studied even for only pla-
nar graphs (see, e.g., [1, 14, 39, 66]). A minimum st-cut of G can be obtained in O(m + n) time
from a maximum st-flow f of G by identifying the edges from the nodes of G reachable from s
to the nodes of G not reachable from s in the residual graph of G with respect to f. No efficient
reductions for the other direction are known. Orlin [55] gave the only known O(mn)-time
algorithms for the maximum st-flow problem on general graphs with integral edge weights.
For undirected planar G, Reif [58] gave an O(n log2 n)-time algorithm for the minimum st-cut
problem. Frederickson [23] improved the time complexity of Reif's algorithm to O(n log n).
The best known algorithms for both problems, due to Italiano et al. [33], run in O(n log log n)
time. The attempt of Janiga and Koubek [34] to generalize Reif's algorithm to directed planar G
turned out to be flawed [17, 36, 51]. Borradaile and Klein [1] and Erickson [14] gave O(n log n)-
time algorithms for both problems on directed planar graphs. On directed planar unweighted
G, Brandes and Wagner [4] and Eisenstat and Klein [13] solved both problems in O(n) time.
The algorithm of Kaplan and Nussbaum [36] is capable of exploiting the condition that nodes s
and t are close. For directed planar G, the O(n log3 n)-time algorithm of Ł acki et al. [43] obtains
the minimum weights of st-cuts for any given s and all nodes t of G. For any given node sub-
sets S and T of directed planar G, the O(n log3 n)-time algorithm of Borradaile, Klein, Mozes,
3
Nussbaum, and Wulff-Nilsen [2] computes a subgraph C of G with minimum weight such
that there is no st-path in G \ C for any s ∈ S and t ∈ T . On undirected planar G, Borradaile,
Sankowski, and Wulff-Nilsen [3] gave an O(n log4 n)-time algorithm to compute a Gomory-
Hu cut-equivalent tree [28], a compact representation of st-cuts with minimum weights for all
nodes s and t.
The kernel of our result is an O(n log log n)-time algorithm for computing noncrossing
shortest paths among nodes well ordered on a common face of a directed plane graph, which
is extended from the algorithm of Italiano et al. [33] for an undirected plane graph. A closely
related NP-hard shortest-noncrossing-paths problem seeks noncrossing paths between k given
terminal pairs on h faces with minimum total weight in a plane graph. Takahashi, Suzuki, and
Nishizeki [63] solved the problem for undirected plane graphs with h ≤ 2 in O(n log k) time.
Papadopoulou [56] addressed the geometric version of the problem, where the terminal pairs
are on the boundaries of h polygonal obstacles in the plane with complexity n and gave an
O(n)-time algorithm for the case h ≤ 2. Erickson and Nayyeri [18] generalized the result of
Takahashi et al., solving the problem for undirected planar graphs in 2O(h2)n log k time. They
also generalized the result of Papadopoulou to solve the geometric version in 2O(h2)n time.
Each of these algorithms computes an implicit representation of the answers, which may have
total size Ω(kn). Polishchuk and Mitchell [57] addressed the problem of finding noncrossing
thick paths with minimum total weight. Takahashi, Suzuki, and Nishizeki [62] also considered
the rectilinear version of the problem.
Technical overview and outline
Our proof for Theorem 1.1 consists of a series of reductions. Based upon the duality between
simple cycles and minimal cuts in plane graphs, Section 2 gives an O(n)-time reduction from
the minimum-cut and shortest-cycle problems in an n-node planar graph to the problem of
finding a shortest non-degenerate cycle in an n-node O(1)-degree plane graph G (Lemma 2.1).
Let C be a balanced separator of G that corresponds to a fundamental cycle with respect to
a shortest-path tree of G. A shortest non-degenerate cycle that does not cross C can be re-
cursively computed from the subgraphs of G separated by C. Although we cannot afford to
compute a shortest non-degenerate cycle that crosses C, Section 3 reduces the problem of find-
ing a shortest non-degenerate cycle to finding a C-short cycle, i.e., a non-degenerate cycle that
crosses C with the property that if it is not shortest, then a shortest non-degenerate cycle that
does not cross C has to be a shortest non-degenerate cycle of G (Lemma 3.1). This reduction
is a divide-and-conquer recursive algorithm using the balanced separator C and thus intro-
duces an O(log n)-factor overhead in the running time. A cycle of G that crosses a shortest
path P of G can be shortcutted into a non-degenerate cycle that crosses P at most once. Sec-
tion 4 reduces the problem of finding a C-short cycle to that of finding a (C, P )-short cycle, i.e.,
a non-degenerate cycle whose weight is no more than that of any non-degenerate cycle that
crosses a shortest path P of G in C exactly once (Lemma 4.2). By the technique of Reif [58] that
incises G along P , Section 4 further reduces the problem of finding a (C, P )-short cycle to that
of finding shortest noncrossing paths among nodes well ordered on the boundary of external
face (Lemma 4.1). As a matter of fact, this shortest-noncrossing-paths problem can be solved
by the O(n log n)-time algorithm of Klein [40], already yielding improved O(n log2 n)-time al-
gorithms for the minimum-cut and shortest-cycle problems. (Mozes et al. [51] also mentioned
that O(n log2 n)-time algorithms can be obtained by plugging in the O(n log n)-time minimum
st-cut algorithm of Borradaile and Klein [1] into a directed version of the reduction algorithm
of Chalermsook et al. [8].) To achieve the time complexity of Theorem 1.1, Section 5 solves the
problem in O(n log log n) time by extending the algorithm of Italiano et al. [33] for an undi-
4
rected plane graph. Section 6 concludes the paper.
2 Reduction to finding shortest non-degenerate cycles
Directed graph G is bidirected if, for any two nodes s and t of G, st is an edge of G if and only if
ts is an edge of G. The graph in Figure 1(a) is not bidirected. The degree of node v in bidirected
G is the number of neighbors of v in G. The degree of bidirected G is the maximum degree of
the nodes in G. A bidirected plane graph is a bidirected planar graph equipped with a plane
embedding in which edges between two adjacent nodes are bundled together. Figures 1(b)
and 1(c) show two bidirected plane graphs G△ and G∗
△. A cycle passing each node at most once
is simple. A cycle is degenerate if it is a node or passes both edges st and ts for two nodes s and t.
A cycle not simple (respectively, degenerate) is non-simple (respectively, non-degenerate). Cycle
C1 in Figure 2(a) is non-degenerate and non-simple. In the graph G of Figure 1(a), cycle v2v3v2
is degenerate and simple, cycle v2v3v4v2 is non-degenerate and simple, and cycle v1v2v4v2v3v1
is degenerate and non-simple. The shortest degenerate cycle of G is v2v3v2 with weight 6. The
shortest non-degenerate cycle of G is v2v3v4v2 with weight 16. Theorem 1.1 can be proved by
the following lemma:
Lemma 2.1. It takes O(n log n log log n) time to compute a shortest non-degenerate cycle in an n-node
O(1)-degree simple bidirected plane graph with nonnegative edge weights.
Proof of Theorem 1.1. Adding edges with weights 0 (respectively, ∞) to the input graph does
not affect the weight of minimum cuts (respectively, shortest cycles). Hence, we may assume
without loss of generality that the input graph G△ has at least four nodes and is a simple
bidirected plane graph such that each face of G△ is a triangle. See Figures 1(a) and 1(b) for
examples. Let the dual G∗
△ of G△ be the simple bidirected plane graph on the 2n − 4 faces of
G△ sharing the same set of 6n − 12 edges with G△ that is obtainable in O(n) time from G△ as
follows: For any two adjacent nodes s and t of G△, there are directed edges f g = st and gf = ts
△, where f and g are the two faces of G△ incident with the bundled edges between s and
in G∗
t such that face g immediately succeeds face f in clockwise order around node s of G△. See
Figure 1(c) for an example. Observe that C is a minimal cut of G△ if and only if C is a simple
△. By nonnegativity of edge weights, a shortest non-degenerate cycle
non-degenerate cycle of G∗
△ is a minimum cut of G△. For instance, the shortest non-degenerate cycle of the graph G∗
of G∗
in Figure 1(c) is f1f4f3f1 with weight 5. It corresponds to the v1v3-cut {v1v3, v2v3, v4v3} of G△,
which in turn corresponds to the minimum cut {v2v3} of G. Although the degenerate cycle
f1f4f1 is a shortest cycle of G∗
△, it does not correspond to a cut of G in the above manner. Since
each node of G∗
△ has exactly three neighbors, the statement of the theorem for the minimum-cut
problem follows from applying Lemma 2.1 on G∗
△.
△
By nonnegativity of edge weights, it takes O(n) time to obtain a shortest degenerate cycle
of G△ by examining the O(n) degenerate cycles of G△ on exactly two nodes. By Lemma 2.1,
the statement of the theorem for the shortest-cycle problem is immediate from the following
claim:
It takes O(n) time to obtain from G△ an O(n)-node O(1)-degree simple bidirected plane
graph G such that a shortest non-degenerate cycle of G△ can be computed from a shortest
non-degenerate cycle of G in O(n) time.
Let G1 and G2 be simple bidirected plane graphs with nonnegative edge weights such that
G2 is obtained from G1 by the following O(d)-time operation on a degree-d node v of G1 with
d ≥ 4: If u1, . . . , ud are the neighbors of v in G1 in clockwise order around v, then SPLIT(v)
5
u1
u2
u1
0
0
5
0
u4
2
4
0
0
1
3
v
7
6
0
8
5
u3
u4
0
0
0
1
3
v1
v4
u2
0
2
v2
0
0
0
8
v3
4
0
u3
0
0
0
0
7
6
(a) G1, C1, and C ∗
1
(b) G2, C ∗
2, and C2
Figure 2: Bidirected plane graphs G1 and G2 and their edge weights are in black solid lines.
2 = u1v1v2u2u3v3v4u4u1 are in blue
Shortest non-degenerate cycles C1 = u1vu2u3vu4u1 and C ∗
1 = u1vu4u1 and C2 = u1v1v2v3v4u4u1 are in
dashed lines. Shortest non-degenerate cycles C ∗
red dotted lines.
• adds zero-weight path v1v2 ··· vd−1vdvd−1 ··· v2v1 with new nodes v1, . . . , vd,
• replaces edge uiv by edge uivi with the same weight for each i = 1, . . . , d,
• replaces edge vui by edge viui with the same weight for each i = 1, . . . , d, and
• deletes v.
See Figure 2 for an example of G1 and G2. An O(n)-node O(1)-degree simple bidirected plane
graph G can be obtained in O(n) time from G△ by iteratively applying SPLIT on each node v
of G△ with degree d ≥ 4. To prove the claim, it suffices to ensure the following statement:
A shortest non-degenerate cycle C1 of G1 is obtainable in O(d) time from a shortest non-
degenerate cycle C ∗
2 of G2.
2 and C1.
2 ). Since C ∗
2 with 1 ≤ i 6= j ≤ d such that P has at least two edges and all
For each uiuj-path P of C ∗
internal nodes of P are in {v1, . . . , vd}, we replace P by path uivuj. By w(P ) = w(uivuj), we
2 is non-degenerate, so is the resulting O(d)-time obtainable
have w(C1) = w(C ∗
cycle C1 of G1. Since C1 may pass v more than once, C1 could be non-simple. See Figure 2
1 ) for any shortest simple
for an example of C ∗
1 of G1. By nonnegativity of edge weights, we have w(C1) ≥ w(C ∗
non-degenerate cycle C ∗
1 )
even if C1 is non-simple. Let C2 be the cycle of G2 that is obtained from C ∗
1 as follows: If
1, then replace it by path uivi ··· vjuj. By
there is a path uivuj with 1 ≤ i 6= j ≤ d in C ∗
1 ). Otherwise, let C2 = C ∗
w(uivuj) = w(uivi ··· vjuj), we have w(C2) = w(C ∗
1 is
1 is non-degenerate, so is C2. See Figure 2
1. Since C ∗
simple, there is at most one path uivuj in C ∗
1 ), C1 is a shortest non-
for an example of C ∗
degenerate cycle of G1.
It remains to show w(C1) = w(C ∗
1 and C2. By w(C1) = w(C ∗
2 ) ≤ w(C2) = w(C ∗
1. Since C ∗
The rest of the paper proves Lemma 2.1.
3 Divide-and-conquer via balanced separating cycles
Let C be a simple non-degenerate cycle of a bidirected plane graph G with nonnegative edge
weights. Let intG(C) (respectively, extG(C)) denote the subgraph of G consisting of the nodes
and edges on the boundary of the faces of G inside (respectively, outside) of C. A non-
degenerate cycle C3 of G is C-short if one of C1, C2, and C3 is a shortest non-degenerate cycle
6
1
2
1
v4
4
1
2
v3
v1
1
2
2
4
2
0
v2
1
1
3
9
9
v6
v5
v2
3
1
1
1
2
9
9
v7
v6
3
9
9
v1
1
2
2
4
v8
4
1
2
v3
1
0
1
2
1
v4
1
0
v5
3
1
2
9
9
v7
(a) G, C, and C ∗
(b) H
Figure 3: (a) The bidirected plane graph G and its edge weights are in black. The blue dashed
cycle C = v1v2v3v4v5v1 is a segmented cycle of G whose segments are shortest paths P1 =
v1v2v3 and P2 = v1v5v4 of G. The shortest non-degenerate cycle of intG(C) is v1v2v5v1 with
weight 5. The shortest non-degenerate cycle of extG(C) is v2v6v7v5v4v3v2 with weight 7. The
red dotted cycle C ∗ = v2v6v7v5v2 with weight 4 is the unique C-short non-degenerate cycle of
G and the unique shortest non-degenerate cycle of G. (b) A bidirected plane graph H.
of G, where C1 (respectively, C2) is a shortest non-degenerate cycle of intG(C) (respectively,
extG(C)). We say that C is segmented if it consists of the following three paths in order: (1) a
shortest path P1, (2) an edge, and (3) the reverse of a shortest path P2, where one of P1 and P2
is allowed to be a node. Let shortest paths P1 and P2 be the segments of C. See Figure 3(a) for
an example. This section proves Lemma 2.1 using Lemmas 3.1, 3.2, and 3.3. Section 4 proves
Lemma 3.1.
Lemma 3.1. Let G be an n-node O(1)-degree simple bidirected plane graph with nonnegative edge
weights. Given a segmented simple non-degenerate cycle C of G together with its segments, it takes
O(n log log n) time to compute a C-short non-degenerate cycle of G.
Lemma 3.2 (Henzinger, Klein, Rao, and Subramanian [31]). It takes O(n) time to compute a
shortest-path tree rooted at any given node of an n-node connected simple directed planar graph with
nonnegative edge weights.
Lemma 3.3 (Lipton and Tarjan [48], Goodrich [29], Klein, Mozes, and Sommer [41, Lemma 1]).
Let G△ be an n-node simple undirected plane triangulation with nonnegative face weights summing to
1 such that the weight of each face of G△ is at most 1
4 . Given any spanning tree T of G△, it takes O(n)
time to obtain an edge e of G△ \ T such that the total weight of the faces of G△ inside (respectively,
outside) of the simple cycle in T ∪ {e} is no more than 3
4 .
Proof of Lemma 2.1. We give a divide-and-conquer recursive algorithm on the input graph H,
which can be assumed to be connected without loss of generality. For each degree-2 node y
whose neighbors x and z are non-adjacent, we replace y and its incident edges by edges xz
and zx with weights w(xy) + w(yz) and w(zy) + w(yx), respectively. The resulting graph G
obtainable in O(n) time from H remains an O(1)-degree simple connected bidirected plane
graph. See Figure 3 for an example of H and G. Let ℓ be the number of faces in G. Since each
maximal simple path on the degree-2 nodes of G has O(1) edges, G has O(ℓ) nodes. A shortest
non-degenerate cycle of H can be obtained in O(n) time from a shortest non-degenerate cycle
of G, which can be found in O(1) time for the case with ℓ ≤ 4.
7
4
1
2
v3
v1
1
2
2
4
2
0
v2
1
1
3
9
9
v6
1
2
1
v4
v5
3
1
2
9
9
v1
1
6
0
1
6
v5
0
1
6
v4
1
6
v2
0
1
6
1
6
v3
0
v7
v6
v7
(a) G and T
(a) G△ and T0
Figure 4: (a) The bidirected plane graph G having 6 faces and its edge weights are in black.
A shortest-path tree T rooted at v1 is in blue dashed lines.
(b) The plane triangulation G△
consists of all edges. The numbers are weights of the faces of G△. The undirected version G0
of G consists of the black solid edges and the blue dashed edges. The undirected version T0 of
T is in blue dashed lines. The edges in G△ \ G0 are in red dotted lines.
To obtain a shortest non-degenerate cycle of G for the case with ℓ ≥ 5, let T be an O(n)-time
obtainable shortest-path tree of G rooted at an arbitrary node as ensured by Lemma 3.2. For
each face f of the simple undirected unweighted version G0 of G having size k ≥ 3, (1) let f be
triangulated into k − 2 faces via adding k − 3 edges without introducing multiple edges, (2) let
an arbitrary one of the k − 2 faces be assigned weight 1
ℓ , and (3) let the remaining k − 3 faces
be assigned weights 0. Let G△ be the resulting simple plane triangulation. The undirected
version T0 of T is a spanning tree of G0 and G△. See Figure 4 for an example. Lemma 3.3
ensures an edge xy of G△ \ T0 obtainable in O(n) time such that the face weights of G△ inside
(respectively, outside) of the simple cycle of T0 ∪ {xy} sum to at most 3
4. For instance, such
an edge xy in the example in Figure 4 is v3v4. If x and y are adjacent in G, then let E = ∅;
otherwise, let E consist of edges xy and yx with weights ∞. We union G and E to obtain a
simple bidirected plane graph G∗. Let s be the least common ancestor of x and y in T . Let
C be the segmented simple non-degenerate cycle of G∗ consisting of (1) the sx-path of T , (2)
edge xy, and (3) the reverse of the sy-path of T . By Lemma 3.1, it takes O(n log log n) time
to compute a C-short cycle C3 of G∗. Let H1 = intG∗(C) \ E and H2 = extG∗(C) \ E. No
matter E = ∅ or not, H1 and H2 are subgraphs of G . We recursively compute a shortest
non-degenerate cycle C1 (respectively, C2) in H1 (respectively, H2), which is also a shortest
non-degenerate cycle of intG∗(C) (respectively, extG∗(C)). By definition of C3, a cycle C ∗ in
{C1, C2, C3} with minimum weight is a shortest non-degenerate cycle of G∗. If C ∗ passes an
edge in E 6= ∅, then the weight of each non-degenerate cycle of G∗ and G is ∞. Otherwise, we
return C ∗ as a shortest non-degenerate cycle of G. The algorithm runs in O(n log log n) time
without accounting for the time for its subsequent recursive calls. By ℓ ≥ 5, the number ℓ1
(respectively, ℓ2) of faces in H1 (respectively, H2) is at most 3
20 ℓ, implying that there
are O(log n) levels of recursion. By ℓ1 + ℓ2 ≤ ℓ + 2, the overall number of faces in each recursion
level is O(n), implying that the overall number of nodes in each recursion level is O(n). The
algorithm runs in O(n log n log log n) time.
4 ℓ + 1 ≤ 19
8
3
3
5
5
1
0
u
6
6
0
0
s
1
1
u1
1
1
u2
1
1
t
(a) G
1
1
0
1
8
8
v
9
9
u1
1
1
u2
1
5
0
1
1
5
1
1
u
3
6
∞
1
8
∞
∞
v1
0
8
∞∞
v2
∞
v
1
9
s
3
1
0
0
6
9
t
(b) H
Figure 5: (a) With P = su1u2t and C = su1u2tvs, the red dotted cycle u1u2uvu1 and the blue
dashed cycle u1vuu2u1 are (C, P )-cycles of G with minimum weight 2. (b) H is obtained from
incising G along P .
4 Non-degenerate cycles that cross the separating cycle
This section proves Lemma 3.1 by Lemma 4.2, which is proved by Lemmas 3.2 and 4.1. Sec-
tion 5 proves Lemma 4.1. If graph G has uv-paths, then let dG(u, v) denote the weight of a
shortest uv-path of G. If G has no uv-paths, then let dG(u, v) = ∞.
Lemma 4.1. Let G be an n-node simple connected bidirected plane graph with nonnegative edge
weights. Let u1, . . . , uℓ, vℓ, . . . , v1 be O(n) nodes on the boundary of the external face of G in order. It
takes overall O(n log log n) time to compute dG(ui, vi) for each i = 1, . . . , ℓ.
Let G be a simple bidirected plane graph. A simple path Q of G aligns with subgraph H
of G if Q or the reverse of Q is a path of H. A simple path Q of G passing at least one edge
deviates from subgraph H of G if the edges and the internal nodes of Q are not in H. For any
simple path P of G, a non-degenerate cycle of G is a P -cycle if it consists of a path aligning
with P and a path deviating from P . For any simple non-degenerate cycle of G and any path
P of G aligning with C, a P -cycle is a (C, P )-cycle if the first edge of its path deviating from P
is in intG(C) if and only if the last edge of its path deviating from P is in extG(C). For instance,
the C ∗ in Figure 3(a) is a P1-cycle of G whose path aligning with P1 is node v2. The first
edge v2v6 (respectively, last edge v5v2) of its path deviating from P1 is in extG(C) (respectively,
intG(C)), so C ∗ is a (C, P1)-cycle of G. C ∗ is also a (C, P2)-cycle. A non-degenerate cycle of G
is (C, P )-short if its weight is no more than that of any (C, P )-cycle of G.
Lemma 4.2. Let G be an n-node O(1)-degree simple bidirected plane graph with nonnegative edge
weights. Let C be a simple non-degenerate cycle of G. Given a path P of G aligning with C, it takes
O(n log log n) time to compute a (C, P )-short cycle of G.
Proof. Let C ∗ be a (C, P )-cycle of G with minimum weight. For instance, the red and blue
cycles in Figure 5(a) are two (C, P )-cycles with minimum weight 2. Let C0 be a shortest non-
degenerate cycle of G passing at least one endpoint of P , which can be obtained in O(n) time
via examining shortest uv-paths in G\{uv, vu} by Lemma 3.2 for all O(1) edges uv of G incident
to at least one endpoint of P . If C ∗ passes some endpoint of P , then w(C0) ≤ w(C ∗), implying
9
R2
Pi
u1
P1
P2
P1
P2
R1
Q1
Q2
(a)
(b)
v2
v1
u2
(c)
Figure 6: (a) The red dotted P1-cycle not intersecting P2 is in intG(C). The blue dashed P2-
cycle not intersecting P1 is in extG(C). (b) The red dotted cycle consists of Q1, R1, Q2, and R2
in order is a (C, P1)-cycle. (c) The degenerate cycle C ′ is obtained from the non-degenerate red
dotted cycle C ∗ by replacing the u1v1-path of C ∗ with the blue dashed u1v1-path of Pi. The
green dash-dotted cycle C ′′ is a non-degenerate cycle contained by C ′.
that C0 is a cycle ensured by the lemma. The rest of the proof assumes that C ∗ does not pass
any endpoint of P . Thus, P has internal nodes. Let H be an O(n)-node O(1)-degree simple
bidirected plane graph obtainable in O(n) time as follows: Suppose that u0, . . . , uℓ+1 with ℓ ≥ 1
are the nodes of P in order. Let s = v0 = u0 and t = vℓ+1 = uℓ+1. We incise G along P by
be replaced by vvi (respectively, viv) with the same weight,
• adding new nodes v1, . . . , vℓ, a new path P ′ = sv1 ··· vℓt and the reverse of P ′,
• for each i = 1, . . . , ℓ, letting each edge vui (respectively, uiv) incident to ui in intG(C) \ P
• letting the weight of each edge in P ′ and the reverse of P ′ be ∞, and
• embedding the resulting graph H such that P and P ′ are on the external face.
See Figure 5 for an example. By Lemma 4.1, it takes overall O(n log log n) time to compute
dH (ui, vi) and dH(vi, ui) for each i ∈ {1, . . . , ℓ}. Let i1 (respectively, i2) be an i ∈ {1, . . . , ℓ} that
minimizes dH(ui, vi) (respectively, dH (vi, ui)). By Lemma 3.2, it takes O(n) time to obtain a
simple shortest ui1vi1-path P1 of H and a simple shortest vi2ui2-path P2 of H. The weight of P1
(respectively, P2) is minimum over all uivi-paths (respectively, viui-paths) of H with 1 ≤ i ≤ ℓ.
Let C1 (respectively, C2) be the non-degenerate cycle of G corresponding to P1 (respectively,
P2). Let Q be the path of C ∗ that deviates from P . Let ui and uj with 1 ≤ i, j ≤ ℓ be the
first and last nodes of Q, respectively. If the first edge of Q is in intG(C), then C ∗ corresponds
to a viui-path of H, implying w(C2) ≤ w(C ∗).
If the last edge of Q is in intG(C), then C ∗
corresponds to a ujvj-path of H, implying w(C1) ≤ w(C ∗). For instance, the red (respectively,
blue) cycle of G in Figure 5(a) corresponds to the red u1v1-path (respectively, blue v1u1-path)
of H in Figure 5(b). Thus, one of C0, C1, and C2 with minimum weight is a cycle ensured by
the lemma.
Proof of Lemma 3.1. Let G1 = intG(C) and G2 = extG(C). Let P1 and P2 be the given segments
of C. Let C ∗ be a shortest non-degenerate cycle of G whose number of edges not in P1 ∪ P2
is minimized over all shortest non-degenerate cycles of G. If C ∗ is a cycle of G1 or G2, then
10
any cycle of G is C-short, including the one ensured by Lemma 4.2. The rest of the proof
assumes that neither G1 nor G2 contains C ∗. By Lemma 4.2, it suffices to ensure that C ∗ is a
(C, P1)-cycle. We need the following claim:
For each i ∈ {1, 2}, if C ∗ ∩ Pi 6= ∅, then C ∗ is a Pi-cycle of G.
By the claim, C ∗ intersects both P1 and P2 or else C ∗ would be a cycle of G1 or G2, as illustrated
by Figure 6(a), contradicting the assumption. Since C ∗ is a P1-cycle and a P2-cycle, C ∗ consists
of four paths Q1, R1, Q2, and R2 in order such that Qi aligns with Pi and Ri deviates from
P1 ∪ P2 for each i ∈ {1, 2}. By the assumption, if R1 ⊆ Gi and R2 ⊆ Gj, then {i, j} = {1, 2}.
Thus, C ∗ is a (C, P1)-cycle. See Figure 6(b) for an illustration. It remains to prove the claim.
Assume for contradiction that C ∗ intersects Pi but is not a Pi-cycle for an index i ∈ {1, 2}.
There are nodes u1, v1, u2, v2 of Pi with u1 6= v1 and u2 6= v2 such that
• u1 precedes v1 in Pi,
• u2 succeeds v2 in Pi,
• the u1v1-path and the u2v2-path of C ∗ deviate from Pi, and
• the u1v1-path of C ∗ deviates from the u2v2-path of C ∗.
Let C ′ be the cycle of G obtained from C ∗ by replacing the u1v1-path of C ∗ with the u1v1-
path of Pi. Since Pi is a shortest path of G, w(C ′) ≤ w(C ∗). Since C ∗ is non-degenerate, the
reverse of each of the u2v2-path of C ′ is not in C ′. Thus, even if C ′ is degenerate, there is a non-
degenerate cycle C ′′ in C ′. See Figure 6(c) for an illustration. By nonnegativity of edge weights,
w(C ′′) ≤ w(C ′). By w(C ′′) ≤ w(C ∗), C ′′ is a shortest non-degenerate cycle of G whose number
of edges not in P1 ∪ P2 is fewer than the number of edges of C ∗ not in P1 ∪ P2, contradicting
the definition of C ∗.
5 Noncrossing shortest paths among nodes on external face
This section proves Lemma 4.1 via extending techniques of Reif [58] and Italiano et al. [33] for
undirected planar graphs. Algorithms for r-divisions (Lemma 5.1) and dense-distance graphs
(Lemma 5.2) are reviewed in §5.1. Data structures for fast-Dijkstra algorithm (Lemma 5.5)
are given in §5.2. Data structures that enables efficient partition of boundary nodes via non-
crossing paths (Lemma 5.6) are given in §5.3. Tools involving noncrossing shortest paths
(Lemma 5.8) are given in §5.4. Lemma 4.1 is proved by Lemmas 3.2, 5.1, 5.2, 5.5, 5.6, and
5.8 in §5.5.
5.1 Dense-distance graph
Let G be a simple bidirected plane graph. A division D of G is an edge-disjoint partition of G
into bidirected plane subgraphs, each of which is a piece of D. The multiplicity of node v of G
in D is the number of pieces of D containing v. A node of G with multiplicity two or more in
D is a boundary node of D. A face of a piece of D is a hole of the piece if it is not a face of G. For
any r > 0, an r-division (see, e.g., [23, 31, 33, 41, 44]) of G is a division of G with O(n/r) pieces,
each having O(r) nodes, O(√r) boundary nodes, and O(1) holes.
Lemma 5.1 (Klein, Mozes, and Sommer [41]). For any r > 0, it takes O(n) time to compute an
r-division for an n-node simple bidirected plane graph each of whose faces contains at most three nodes.
11
y1
1
9
1
1
1
1
x1
9
9
x2
1
1
1
1
1
1
1
1
9
x3
1
y2
1
9
9
3
4
2
x1
5
2
x2
3
4
2
3
5
3
2
2
4
4
2
x3
2
2
y1
y2
4
4
(a) H
(b) K(H)
Figure 7: (a) A piece H in which x1, x2, and x3 are the boundary nodes in one hole and y1 and
y2 the boundary nodes in the other hole. (b) K(H).
Let D be an r-division of G. For any connected component H of any piece of D, let K(H)
denote the complete directed graph on the boundary nodes of D in H in which w(uv) =
dH (u, v). See Figure 7 for an example. The dense distance graph (see, e.g., [33]) K(D) of D is
the O(n)-edge simple directed graph on the O(n/√r) boundary nodes of D simplified from the
union of K(H) over all connected components H of all pieces of D by keeping exactly one
copy of parallel edges with minimum weight. For any edge uv of K(D), an underlying uv-path
is a uv-path in some connected component H of some piece of D with weight equal to w(uv)
in K(D). For any path Π of K(D), an underlying path of Π consists of an underlying uv-path
for each edge uv of Π.
Lemma 5.2 (Klein [40]). For any given r-division D of an n-node simple bidirected plane graph with
nonnegative edge weights, it takes O(n log r) time to compute K(D) and a data structure from which,
for any path Π of K(D), the first c edges of an underlying path of Π can be obtained in O(c log log r)
time.
5.2 Fast-Dijkstra algorithm
Consider the following equation
w(u1v1) + w(u2v2) ≤ w(u1v2) + w(u2v1)
(1)
for distinct nodes u1, u2, v1, v2 of a simple directed graph H with edge weights. A type-1 Monge
unit is a complete H equipped with a cyclic ordering for the nodes of H such that Equation (1)
holds for any distinct nodes u1, u2, v2, v1 of H in order. A type-2 Monge unit is a complete
bipartite H equipped with an ordering for each of the two maximal independent sets of H
such that Equation (1) holds for any distinct nodes u1 and u2 of one independent set in order
and any distinct nodes v1 and v2 of the other independent set in order.
A Monge decomposition of a simple directed graph K with edge weights is a set M of Monge
units on node subsets of K such that K is the graph simplified from the union of the Monge
units in M. The multiplicity of a node v of K in M is the number of Monge units in M that
contain v. The size of M is the sum of the multiplicities of all nodes of K in M. An equivalent
12
x2
x2
x1
2
5
2
y1
2
3
5
6
2
2
7
y2
x3
x1
2
5
2
2
2
6
5
3
2
2
2
y1
x3
2
2
2
y2
(a)
(b)
Figure 8: Each of the two graphs can be simplified from the union of two type-2 Monge units.
form of the following lemma is proved by Mozes and Wulff-Nilsen [52, §4.4] using the algo-
rithm of Klein [40] and used by Kaplan, Mozes, Nussbaum, and Sharir [35, §5.2]. Specifically,
for any hole C of a piece H of D, the complete graph on the nodes of C with w(uv) = dH(u, v)
for any nodes u and v in C equipped with the cyclic ordering of C is a type-1 Monge unit. For
instance, the subgraphs of K(H) in Figure 7(b) induced by {x1, x2, x3} and {y1, y2} equipped
with their cyclic orders on the holes are two type-1 Monge units. For any two holes C1 and C2
of a piece H of D, Mozes et al. showed that the complete bipartite graph on the nodes of C1
and C2 with w(uv) = dH (u, v) for nodes u and v such that each of C1 and C2 contains exactly
one of u and v can be simplified from the union of O(1) type-2 Monge units. For instance,
the subgraph of K(H) in Figure 7 consisting of edges between {x1, x2, x3} and {y1, y2} can be
simplified from the union of the graphs in Figures 8(a) and 8(b). The edges of the graph in
Figure 8(a) from hx1, x2, x3i (respectively, hy2, y1i) to hy1, y2i (respectively, hx3, x1, x2i) form a
type-2 Monge unit. The edges of the graph in Figure 8(b) from hx3, x1, x2i (respectively, hy1, y2i)
to hy1, y2i (respectively, hx3, x1, x2i) form a type-2 Monge unit.
Lemma 5.3. For any given r-division D of an n-node simple bidirected plane graph with nonnegative
edge weights, it takes O(n log r) time to obtain a Monge decomposition M (D) of K(D) such that the
multiplicity of a node of K(D) in M (D) is O(1) times its multiplicity in D.
As summarized in the following lemma, given a size-m Monge decomposition of graph
K, there are O(m log2 m)-time obtainable data structures for range minimum queries (see, e.g.,
Kaplan et al. [35] and Gawrychowski, Mozes, and Weimann [26]) with which the fast-Dijkstra
algorithm of Fakcharoenphol and Rao [20] outputs a shortest-path tree of K in O(m log2 m)
time.
Lemma 5.4. Given a size-m Monge decomposition of a simple strongly connected directed graph K
with nonnegative edge weights, it takes O(m log2 m) time to compute a shortest-path tree of K rooted
at any given node.
Lemma 5.5. Let D be a given r-division of an n-node simple plane graph with nonnegative edge
It takes O(n log r) time to compute a data structure from which, for any subset X of the
weights.
boundary nodes of D such that the subgraph K of K(D) induced by X is strongly connected, it takes
O(m log2 m) time to compute a shortest-path tree of K rooted at any given node, where m is the sum of
the multiplicities of the nodes of X in D.
Proof. Let M (D) be a Monge decomposition of K(D) as ensured by Lemma 5.3. Let M consist
of the subgraph H[X] of H induced by X for each Monge unit H in M (D). Each H[X] remains
13
u1
1
2
1
1
x
1
1
u2
2
1
4
6
z
3
5
1
0
v1
u1
1
1
1
2
y
1
1
1
1
x
2
0
1
1
v2
u2
2
1
z
1
0
v1
1
1
y
1
1
2
0
v2
(a) G, P1, and P2
(b) G[P1, P2]
Figure 9: (a) P1 = u1xyzv1 and P2 = u2xyzv2 are noncrossing shortest paths of G. (b) G[P1, P2].
a Monge unit with the induced cyclic ordering (respectively, orderings) of the nodes in H[X]
for the first (respectively, second) type. Thus, M is a Monge decomposition of K preserving the
property that the multiplicity of a node of K in M is O(1) times its multiplicity in D, implying
that the size of M is O(m). It takes overall O(m) time to obtain the induced cyclic ordering
or the two induced orderings of the nodes of H[X] from H for each Monge unit H in M (D).
Since the weight of each edge of H[X] can be obtained in O(1) time from its weight in H, we
have an implicit representation of M in O(m) time. The lemma follows from Lemma 5.4.
5.3 Noncrossing paths
Let G be a simple connected bidirected plane graph. Let u1, u2, v2, v1 be distinct nodes on the
boundary of the external face of connected plane graph G in order. A simple u1v1-path P1 and
a simple u2v2-path P2 of G are noncrossing if P1 ∩ P2 is empty or is a path. For instance, in
Figure 9, P1 in red and P2 in blue are noncrossing. For noncrossing P1 and P2, let G[P1, P2]
denote the connected bidirected plane subgraph of G enclosed by P1, P2, and the u1u2-path
and v2v1-path on the boundary of the external face of G following the order of u1, u2, v2, v1.
See Figure 9 for an example.
Let D be an r-division of G. Our proof of Lemma 4.1 needs a data structure B(D) with the
following property: For distinct nodes u1, u2, u3, v3, v2, v1 on the external face of G in order, any
disjoint simple u1v1-path P1 and u3v3-path P3 of G, and any simple u2v2-path P2 of G[P1, P3]
such that P1 and P2 are noncrossing, given X(1, 3) and P2 \ P1, it takes O((m1 + m2) log r) time
to obtain X(1, 2) and X(2, 3), where X(i, j) with 1 ≤ i < j ≤ 3 consists of the boundary nodes
of D in G[Pi, Pj], m1 is the sum of multiplicities of the nodes of X(1, 3) in D, and m2 is the
number of edges in P2 \ P1. See Figure 10 for an illustration.
Lemma 5.6. It takes O(n) time to compute a data structure B(D) for any given r-division D of any
n-node simple connected bidirected plane graph.
Proof. Given X(1, 3) and the edge set E of P2 \ P1, it takes O(m1 + m2) time to obtain the nodes
of X(1, 3) in E, which belongs to X(1, 2) ∩ X(2, 3). Let X consist of the nodes of X(1, 3) not
in E. If X = ∅, then X(1, 2) = X(2, 3) = X(1, 3). The rest of the proof assumes X 6= ∅.
Let 0 (respectively, 1) consist of the pieces H of D such that H contains nodes of X and no
(respectively, some) edges of E. We have 1 6= ∅, since G[P1, P3] is connected and E 6= ∅. Let
A be the O(m1 + m2)-time obtainable undirected bipartite graph on the nodes x in X and the
pieces H of D in 0 such that H and x are adjacent in A if and only if H contains x in G. The
nodes of X in the same connected component of A either all belong to X(1, 2) or all belong
to X(2, 3). Since G[P1, P3] is connected, each connected component of A contains a node of X
14
P1
u1
u2
P2
u3
P3
(a) G[P1, P3]
v1
v2
v3
P1
u1
u2
P2
(b) G[P1, P2]
u2
P2
u3
P3
(c) G[P2, P3]
v1
v2
v2
v3
Figure 10: An illustration for the definition of B(D), where P1 is the blue solid u1v1-path, P3
is the green dash-dotted u3v3-path, and P2 is the red dotted u2v2-path. P1 and P3 are disjoint.
P1 and P2 are noncrossing. (a) G[P1, P3], in which the boundary nodes of D form X(1, 3). (b)
G[P1, P2], in which the boundary nodes of D form X(1, 2). (c) G[P2, P3], in which the boundary
nodes of D form X(2, 3).
that belongs to a piece of H ∈ 1 in G. It takes overall O(m1 + m2) time to obtain H ∩ E, C ∩ E,
and C ∩ X for each hole C of each piece H of D in 1. Since each piece of D has O(1) holes,
it remains to show that with the B(D) defined below, for each hole C of each piece H of D in
1, it takes O(m log r) time to determine the nodes of C ∩ X in X(1, 2), where m is the number
of nodes in H ∩ X plus the number of edges in H ∩ E.
Assume without loss of generality that the external face of each piece H of D is a hole of H.
The O(n)-time obtainable data structure B(D) consists of (1) the cyclic ordering of the incident
edges around each node of G and (2) the following items for each hole C of each piece H of D:
• An arbitrary simple path Q of H from a node of C to a node q on the external face of H.
• The ordering indices of the nodes on Q.
• The cyclic ordering indices of the nodes on C.
It takes overall O(m1 + m2) time to obtain Q ∩ E for each hole C of each piece H of D in 1.
With the first part of B(D), if uv is an edge of G[P1, P3] with u ∈ P2 and v /∈ P2, then it takes
O(1) time to determine whether v ∈ G[P1, P2]. With the second part of B(D), for any k-node
subset U of any piece H of D and any hole C of H, it takes O(k) time to determine the ordering
indices of the nodes of U ∩ Q in Q and the cyclic ordering indices of the nodes of U ∩ C in C.
Case 1: C ∩ E 6= ∅. As illustrated by Figure 11(a), it takes overall O(m log r) time via sorting
their ordering indices to compute, for each node x of C∩X, the first node u ∈ E in the traversal
of C starting from x following the order of u1, u3, v3, v1 and the node v of C preceding u in the
traversal. We have x ∈ X(1, 2) if and only if v ∈ G[P1, P2], which can be determined in O(1)
time.
Case 2: C∩E = ∅. As illustrated by Figure 11(b), if Q∩E 6= ∅, then let v be the node preceding
the first node u of Q in E. Let C ′ be the boundary of the external face of H. As illustrated by
Figure 11(c), if Q ∩ E = ∅, then let v be the node of C ′ preceding the first node u of C ′ in E on
the traversal of C ′ starting from q following the order of u1, u3, v3, v1. Either way, it takes O(m)
15
v
u
E
x
E
C
(a)
E
C
Q
E
v
u
q
Q
(b)
q
v
u
C
Q
E
C ′
(c)
Figure 11: Illustrations for the proof of Lemma 5.6.
P ′
2
P1
u1
u
u2
v
P2
v1
v2
Figure 12: An illustration for the proof of Lemma 5.7.
time to obtain v and determine whether v ∈ G[P1, P2]. If v ∈ G[P1, P2], then C ∩ X ⊆ X(1, 2).
Otherwise, C ∩ X ⊆ X(2, 3).
5.4 Noncrossing shortest paths
Lemma 5.7. Let G be a simple connected bidirected plane graph with nonnegative edge weights. If
nodes u1, u2, v2, v1 are on the boundary of the external face of G in order, then for any shortest u1v1-
path P1 of G, there is a shortest u2v2-path P2 of G such that P1 and P2 are noncrossing.
2 6= ∅.
Proof. As illustrated by Figure 12, suppose that P ′
2 in P1. Let P2 be obtained from
Let u (respectively, v) be the first (respectively, last) node of P ′
2 by replacing its uv-path with the uv-path of P1. By the order of u1, u2, v2, v1 on the boundary
P ′
of the external face of G, P2 is well defined and is a shortest u2v2-path of G such that P1 and
P2 are noncrossing.
2 is a shortest u2v2-path of G with P1∩P ′
Lemma 5.8. Let G be an n-node simple connected bidirected plane graph with nonnegative edge
weights. Let u1, . . . , uk, vk, . . . , v1 be 2k distinct nodes on the boundary of the external face of G in
16
order. For each i ∈ {1, k}, let Pi be a simple shortest uivi-path of G such that P1 and Pk are noncross-
ing. Let h be the number of nodes of G[P1, Pk] not in P1 ∩ Pk. Given P1 \ Pk and Pk \ P1, consider the
problem of computing dG(ui, vi) for all i = 1, . . . , k.
1. If P1 ∩ Pk = ∅, then the problem can be solved in O(h log k) time.
2. If P1 ∩ Pk = ∅ and we are given a set Z of O(1) nodes such that for each i = 1, . . . , k at least
one shortest uivi-path passes at least one node of Z, then the problem can be solved in O(h) time.
3. If P1 ∩ Pk 6= ∅ and we are given w(P1 ∩ Pk), then the problem can be solved in O(h) time.
Proof. Since P1 \ Pk and Pk \ P1 are given, it takes O(h) time to obtain G[P1, Pk] excluding the
edges and internal nodes of P1∩Pk. Statements 2 and 3 follow from Lemmas 3.2 and 5.7. As for
Statement 1, under the assumption that a simple shortest uava-path Pa and a simple shortest
ubvb-path Pb of G are given and disjoint, below is the recursive algorithm MEASURE(a, b) with
1 ≤ a < b ≤ k for solving the (a, b)-subproblem of computing dG(ui, vi) for all indices i with
a < i < b:
Let i = ⌊(a + b)/2⌋. By Lemma 3.2, it takes time linear in the number of nodes in
G[Pa, Pb] to obtain dG(ui, vi) and a simple shortest uivi-path Pi of G[Pa, Pb] that is
noncrossing with both Pa and Pb. For the (a, i)-subproblem, if Pa ∩ Pi = ∅, then
call MEASURE(a, i); otherwise, apply Statement 2 with Z consisting of an arbitrary
node in Pa ∩ Pi. For the (i, b)-subproblem, if Pi ∩ Pb = ∅, then call MEASURE(i, b);
otherwise, apply Statement 2 with Z consisting of an arbitrary node in Pi ∩ Pb.
The algorithm for the statement obtains dG(u1, v1) and dG(uk, vk) from P1 and Pk and calls
MEASURE(1, k). Since each dG(ui, vi) with 1 < i < k is computed by Lemma 3.2 or Statement 2,
the correctness holds trivially. By the choice of i, MEASURE(1, k) runs in O(log k) levels of
recursion. Since Pa ∩ Pb = ∅ holds for each call to MEASURE(a, b), each node of G[P1, Pk]
appears in at most two subgraphs G[Pa, Pb] in the same level of recursion. Thus, the overall
running time for each level of recursion is O(h). The algorithm runs in O(h log k) time.
5.5 Proving Lemma 4.1
Proof of Lemma 4.1. For each i = 1, . . . , ℓ, let di = dG(ui, vi). With the modification below, each
di with 1 ≤ i ≤ ℓ equals the weight of a shortest u′
i-path in the resulting G, which remains an
iv′
O(n)-node simple connected bidirected plane graph: (1) for each i = 1, . . . , ℓ, add new nodes u′
i
and v′
i and
i in the external face, zero-weighted edges u′
ivi, (2) contract each zero-weighted strongly connected subgraph into a single node, (3) delete
v′
all self-loops, and (4) delete all except one copy of each set of multiple edges with minimum
weight. Thus, the rest of the proof assumes that u1, . . . , uℓ, vℓ, . . . , v1 are distinct and G does
not have any zero-weighted cycles, implying that all shortest paths of G are simple.
i, and ∞-weighted edges uiu′
iui and viv′
Let G△ be an O(n)-node bidirected plane graph obtainable in O(n) time from G by identi-
fying nodes ui and vi into a new node zi for each i = 1, . . . , ℓ and then triangulating each face
of size larger than 3. Let
(2)
r = max(1,⌈log6
2 n⌉).
By Lemma 5.1, an r-division D0 for G△ can be computed in O(n) time. Let D1 be the division of
G induced by D0: Each piece of D1 is obtained from a piece of D0 by deleting the edges added
to triangulate faces of size larger than 3. Each piece of D0 has O(r) nodes, O(√r) boundary
nodes, and O(1) holes, so does each piece of D1. Let I consist of indices 1 and ℓ and the indices
i such that at least one of ui and vi is a boundary node of D1. Since each zi with i ∈ I \ {1, ℓ}
is a boundary node in D0, the cardinality of I is O(n/√r). To turn both of ui and vi with i ∈ I
17
Subroutine SOLVE(a, b)
If I(a, b) = ∅, then solve the (a, b)-subproblem by Lemma 5.8(1) and return.
If I(a, b) 6= ∅, then let i be a median of I(a, b) and let P (respectively, P ′) be a shortest uivi-path whose
first (respectively, last) c edges can be obtained in O(c log log r) time.
Case 1: P ∩ (Pa ∪ Pb) = ∅. Let Pi = P . Call LABEL(Pi), SOLVE(a, i), and SOLVE(i, b). Return.
Case 2: P ∩ (Pa ∪ Pb) 6= ∅.
• Call LABEL(P [ui, x]), where x is the first node of P in Pa ∪ Pb.
• Call LABEL(P ′[y, vi]), where y is the last node of P ′ in P [ui, x] ∪ Pa ∪ Pb.
Case 2(1): y ∈ Pa ∪ Pb. Let j be the index in {a, b} with x ∈ Pj.
• If y /∈ Pj, then solve the (a, b)-subproblem by Lemma 5.8(2) with Z = {x, y}. Return.
• If y ∈ Pj, then let Pi = P [ui, x]Pj [x, y]P ′[y, vi], implying w(Pi ∩ Pj) = φ(y) − φ(x).
– If x ∈ Pa, then solve the (a, i)-subproblem by Lemma 5.8(3) and call SOLVE(i, b). Return.
– If x ∈ Pb, then solve the (i, b)-subproblem by Lemma 5.8(3) and call SOLVE(a, i). Return.
Case 2(2): y /∈ Pa ∪ Pb, implying y ∈ P [ui, x] and y 6= x. Let Pi = P [ui, y]P ′[y, vi]. Let Z = {x}.
• If x ∈ Pa, then solve the (a, i)-subproblem by Lemma 5.8(2) and call SOLVE(i, b). Return.
• If x ∈ Pb, then solve the (i, b)-subproblem by Lemma 5.8(2) and call SOLVE(a, i). Return.
Figure 13: Subroutine SOLVE(a, b).
into boundary nodes, we introduce O(n/r) new O(√r)-node pieces, which form a partition of
the nodes ui and vi with i ∈ I. Let D be the resulting division of G. Each new piece of D has
O(√r) nodes and no edges, so it has O(√r) boundary nodes and O(1) holes. Thus, D is an
r-division of G such that each ui with 1 ≤ i ≤ ℓ is a boundary node in D if and only if so is vi.
Let G′ be the simple bidirected plane graph with edge weights obtained from G by reversing
the direction of each edge. Let D′ be the r-division of G′ corresponding to D. By Equation (2),
it takes O(n log log n) time to compute K(D) and K(D′) and the data structures ensured by
Lemmas 5.2 and 5.5.
For any nodes x and y in a shortest path P of G, let P [x, y] denote the xy-path of P . We
need a subroutine LABEL(P ) to compute label φ(z) for each node z of a shortest path P of G
under the assumption that φ(z) for at most one node of P is pre-computed:
Let z∗ be the node with pre-computed φ(z∗). If there is no such a node, then let
z∗ be an arbitrary node of P and let φ(z∗) = 0. For each node z that precedes z∗
in P , let φ(z) = φ(z∗) − w(P [z, z∗]). For each node z that succeeds z∗ in P , let
φ(z) = φ(z∗) + w(P [z, z∗]).
Subroutine LABEL(P ) runs in O(1) time per node of P and does not overwrite φ(z) for any z
with pre-computed φ(z). After running LABEL(P ), for any nodes x and y of P , w(P [x, y]) can
be obtained from φ(y) − φ(x) in O(1) time.
For any indices a and b, let set I(a, b) consist of the indices i ∈ I with a < i < b. For
each i ∈ {1, ℓ}, let Pi be a shortest uivi-path of G obtainable in O(n) time by Lemma 3.2. If
P1 ∩ Pℓ 6= ∅, then the lemma follows from Lemma 5.8(2) with Z = {x} for an arbitrary node
x ∈ P1 ∩ Pℓ. The rest of the proof assumes P1 ∩ Pℓ = ∅. The algorithm proving the lemma calls
LABEL(P1), LABEL(Pk), and SOLVE(1, ℓ), where the main subroutine SOLVE(a, b), as defined in
Figure 13 and elaborated below, solves the (a, b)-subproblem of computing di for all indices i
with a ≤ i ≤ b under the condition that
18
ua
ui
ub
va
vi
vb
x
y
(a)
ua
x
y
ui
ub
va
vi
vb
ua
ui
ub
x
y
va
vi
vb
(b)
(c)
Figure 14: Illustrations for the proof of Lemma 4.1. All Pa and Pb are in black. Each P [ui, x] is
in red dots. Each P ′[y, vi] is in blue dashes.
• shortest uava-path Pa of G and shortest ubvb-path Pb of G are disjoint,
• φ(z) is pre-computed for each node z ∈ Pa ∪ Pb, and
• the set X(a, b) of boundary nodes of D in G[Pa, Pb] is given.
By Equation (2), it remains to prove that SOLVE(1, ℓ) correctly solves the (1, ℓ)-subproblem in
O(n log r) time. If I(a, b) = ∅, then all ui with a < i < b are not boundary nodes in D. Since
these ui induce a connected subgraph of G, they belong to a common piece of D, implying
b − a = O(r). The (a, b)-subproblem can be solved by Lemma 5.8(1) in O(h(a, b) log r) time,
where h(a, b) is the number of nodes in G[Pa, Pb] that are not in Pa ∩ Pb.
For the case with I(a, b) 6= ∅, we cannot afford to directly compute a shortest uivi-path
Pi of G for a median i of I(a, b) by Lemma 3.2. Instead, in the subgraph of K(D) induced by
the given set X(a, b) of boundary nodes of D in G[Pa, Pb], we compute a shortest uivi-path Π
(respectively, Π′) of K(D) (respectively, K(D′)), the first (respectively, last) c edges of whose
underlying path P (respectively, P ′) can be obtained in O(c log log r) time by Lemma 5.2. By
Lemma 5.7, G[Pa, Pb] contains at least one shortest uivi-path of G, implying that the subgraph
of K(D) induced by X(a, b) contains at least one shortest uivi-path of K(D). Therefore, P
and P ′ are shortest uivi-paths of G in G[Pa, Pb]. If P does not intersect Pa ∪ Pb, then it takes
O(log log r) time per node to obtain P . As in Case 1 of Figure 13, the subroutine lets Pi = P
and calls LABEL(Pi), SOLVE(a, i), and SOLVE(i, b). If P intersects Pa ∪ Pb, it takes O(log log r)
time per node to obtain P [ui, x] and P ′[y, vi], where x is the first node of P in Pa ∪ Pb and y is
the last node of P ′ in P [ui, x] ∪ Pa ∪ Pb, as stated by the first two bullets in Case 2 of Figure 13.
The subroutine calls LABEL(P [ui, x]) and LABEL(P ′[y, vi]).
• As illustrated by Figure 14(a), if each of Pa and Pb contains exactly one of x and y, then
the (a, b)-subproblem is solved in O(h(a, b)) time by Lemma 5.8(2) with Z = {x, y}, as
stated by the first bullet in Case 2(1) of Figure 13.
• As illustrated by Figure 14(b), if x, y ∈ Pa, then let Pi = P [ui, x]Pa[x, y]P ′[y, vi]. The
(i, b)-subproblem is solved by calling SOLVE(i, b). The (a, i)-subproblem is solved by
Lemma 5.8(3) with w(Pa ∩ Pi) = φ(y) − φ(x) in O(h(a, b)) time. The case with x, y ∈ Pb is
similar. The second bullet of Case 2(1) in Figure 13 states these two cases.
• As illustrated by Figure 14(c), if x ∈ Pa and y /∈ Pa ∪ Pb, then the shortest uivi-path
Pi = P [ui, y]P ′[y, vi] is disjoint with Pa ∪ Pb. The (i, b)-subproblem is solved by call-
ing SOLVE(i, b). Since at least one shortest uivi-path of G[Pa, Pi] passes x, the (a, i)-
subproblem can be solved in O(h(a, b)) time by Lemma 5.8(2) with Z = {x}. The case
with x ∈ Pb and y /∈ Pa ∪ Pb is similar. Case 2(2) in Figure 13 states these two cases.
19
The correctness holds trivially, since each di with 1 ≤ i ≤ ℓ is computed somewhere during
the execution of SOLVE(1, ℓ) by Lemma 5.8. Since i is chosen to be a median of I(a, b) in each
subroutine call to SOLVE(a, b), there are O(log n) levels of recursion in executing SOLVE(1, ℓ).
Let m(a, b) be the sum of the multiplicities of the nodes of X(a, b) in D. By Lemma 5.5, the
time for computing Π and Π′ is O(m(a, b) log2 m(a, b)). In order to maintain the condition that
X(a, b) is given whenever SOLVE(a, b) is called, we apply Lemma 5.6 to obtain X(a, i) and
X(i, b) in O((m(a, b) + mi) log r) time before calling SOLVE(a, i) or SOLVE(i, b), where mi is the
number of edges in Pi \ (Pa ∪ Pb). Since Pa and Pb are disjoint, each boundary node of D
is contained by one or two subgraphs G[Pa, Pb] of the same recursion level. Since there are
O(n/r) pieces of D and each piece of D has O(√r) boundary nodes, the sum of m(a, b) over all
subgraphs G[Pa, Pb] at the same recursion level is O(n/√r). Since each edge of G appears in at
most one Pi \ (Pa ∪ Pb) for all subroutine calls to SOLVE(a, b), the sum of all mi throughout the
execution of SOLVE(1, ℓ) is O(n). By Equation (2), the overall time for computing Π and Π′ is
O(cid:18)log n ·
n
√r
log2 n(cid:19) = O(n).
The overall time of finding all paths P , P [ui, x], and P ′[y, vi] is O(n log log r), since their edges
are disjoint and all of them are obtainable in O(log log r) time per node. Therefore, the running
time of SOLVE(1, ℓ) is dominated by the sum of the O(h(a, b) log r) time for solving the (a, b)-
subproblems by Lemmas 5.8(1), 5.8(2), and 5.8(3) at the bottom of recursion. Since the sum of
h(a, b) over all these (a, b)-subproblems is O(n), the running time of SOLVE(1, ℓ) is O(n log r).
The lemma is proved.
6 Concluding remarks
We give the first known O(n log n log log n)-time algorithms for finding a minimum cut and a
shortest cycle in an n-node simple directed planar graph G with nonnegative edge weights.
For the case that G is restricted to be unweighted, our shortest-cycle algorithm remains the
best known result for the shortest-cycle problem. The best algorithm for the minimum-cut
problem, running in O(n log n) time, is obtained by plugging in the O(n)-time minimum st-
cut algorithm of, e.g., Brandes and Wagner [4] and Eisenstat and Klein [13] to a directed version
of the reduction algorithm of Chalermsook et al. [8]. Thus, an interesting future direction is
to further reduce the running time of our algorithms on both problems for this special case.
Extending our results to bounded-genus graphs is also of interest.
Acknowledgment
We thank the anonymous reviewers for helpful comments.
References
[1] G. Borradaile and P. N. Klein. An O(n log n) algorithm for maximum st-flow in a directed
planar graph. Journal of the ACM, 56(2):9.1–9.30, 2009.
[2] G. Borradaile, P. N. Klein, S. Mozes, Y. Nussbaum, and C. Wulff-Nilsen. Multiple-source
multiple-sink maximum flow in directed planar graphs in near-linear time. In Proceedings
of the 52nd Annual IEEE Symposium on Foundations of Computer Science, pages 170–179,
2011.
20
[3] G. Borradaile, P. Sankowski, and C. Wulff-Nilsen. Min st-cut oracle for planar graphs with
near-linear preprocessing time. ACM Transactions on Algorithms, 11(3):16.1–16.29, 2015.
[4] U. Brandes and D. Wagner. A linear time algorithm for the arc disjoint Menger problem
in planar directed graphs. Algorithmica, 28(1):16–36, 2000.
[5] S. Cabello. Finding shortest contractible and shortest separating cycles in embedded
graphs. ACM Transactions on Algorithms, 6(2):24.1–24.18, 2010.
[6] S. Cabello, E. W. Chambers, and J. Erickson. Multiple-source shortest paths in embedded
graphs. SIAM Journal Computing, 42(4):1542–1571, 2013.
[7] S. Cabello, É. Colin de Verdière, and F. Lazarus. Finding shortest non-trivial cycles in
directed graphs on surfaces. In Proceedings of the 26th ACM Symposium on Computational
Geometry, pages 156–165, 2010.
[8] P. Chalermsook, J. Fakcharoenphol, and D. Nanongkai. A deterministic near-linear time
algorithm for finding minimum cuts in planar graphs. In Proceedings of the 15th Annual
ACM-SIAM Symposium on Discrete Algorithms, pages 828–829, 2004.
[9] H.-C. Chang and H.-I. Lu. Computing the girth of a planar graph in linear time. SIAM
Journal on Computing, 42(3):1077–1094, 2013.
[10] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to Algorithms. MIT
Press, 3rd edition, 2009.
[11] M. Cygan, H. N. Gabow, and P. Sankowski. Algorithmic applications of Baur-Strassen's
theorem: Shortest cycles, diameter and matchings. In Proceedings of the 53rd Annual IEEE
Symposium on Foundations of Computer Science, pages 531–540, 2012.
[12] H. Djidjev. A faster algorithm for computing the girth of planar and bounded genus
graphs. ACM Transactions on Algorithms, 7(1):3.1–3.16, 2010.
[13] D. Eisenstat and P. N. Klein. Linear-time algorithms for max flow and multiple-source
shortest paths in unit-weight planar graphs. In Proceedings of the 45th ACM Symposium on
Theory of Computing, pages 735–744, 2013.
[14] J. Erickson. Maximum flows and parametric shortest paths in planar graphs. In Proceed-
ings of the 21st Annual ACM-SIAM Symposium on Discrete Algorithms, pages 794–804, 2010.
[15] J. Erickson, K. Fox, and A. Nayyeri. Global minimum cuts in surface embedded graphs. In
Proceedings of the 23rd Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1309–
1318, 2012.
[16] J. Erickson and S. Har-Peled. Optimally cutting a surface into a disk. Discrete & Computa-
tional Geometry, 31(1):37–59, 2004.
[17] J. Erickson and A. Nayyeri. Minimum cuts and shortest non-separating cycles via homol-
ogy covers. In Proceedings of the 22nd Annual ACM-SIAM Symposium on Discrete Algorithms,
pages 1166–1176, 2011.
[18] J. Erickson and A. Nayyeri. Shortest non-crossing walks in the plane. In Proceedings of the
22nd Annual ACM-SIAM Symposium on Discrete Algorithms, pages 297–208, 2011.
21
[19] J. Erickson and P. Worah. Computing the shortest essential cycle. Discrete & Computational
Geometry, 44(4):912–930, 2010.
[20] J. Fakcharoenphol and S. Rao. Planar graphs, negative weight edges, shortest paths, and
near linear time. Journal of Computer and System Sciences, 72(5):868–889, 2006.
[21] K. Fox. Shortest non-trivial cycles in directed and undirected surface graphs.
In Pro-
ceedings of the 24th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 352–364,
2013.
[22] K. Fox. Fast algorithms for surface embedded graphs via homology. PhD thesis, University of
Illinois at Urbana-Champaign, 2014.
[23] G. N. Frederickson. Fast algorithms for shortest paths in planar graphs, with applications.
SIAM Journal on Computing, 16(6):1004–1022, 1987.
[24] H. N. Gabow. A matroid approach to finding edge connectivity and packing arbores-
cences. Journal of Computer and System Sciences, 50(2):259–273, 1995.
[25] H. N. Gabow and R. E. Tarjan. Faster scaling algorithms for network problems. SIAM
Journal on Computing, 18(5):1013–1036, 1989.
[26] P. Gawrychowski, S. Mozes, and O. Weimann. Submatrix maximum queries in Monge
matrices are equivalent to predecessor search. In B. Speckmann, editor, Proceedings of the
42nd International Colloquium on Automata, Languages, and Programming, pages 580-592,
2015.
[27] A. V. Goldberg. Scaling algorithms for the shortest paths problem. SIAM Journal on Com-
puting, 24(3):494–504, 1995.
[28] R. E. Gomory and T. C. Hu. Multi-terminal network flows. Journal of the SIAM, 9(4):551–
570, 1961.
[29] M. T. Goodrich. Planar separators and parallel polygon triangulation. Journal of Computer
and System Sciences, 51(3):374–389, 1995.
[30] J. Hao and J. B. Orlin. A faster algorithm for finding the minimum cut in a directed graph.
Journal of Algorithms, 17(3):424–446, 1994.
[31] M. R. Henzinger, P. N. Klein, S. Rao, and S. Subramanian. Faster shortest-path algorithms
for planar graphs. Journal of Computer and System Sciences, 55(1):3–23, 1997.
[32] A. Itai and M. Rodeh. Finding a minimum circuit in a graph. SIAM Journal on Computing,
7(4):413–423, 1978.
[33] G. F. Italiano, Y. Nussbaum, P. Sankowski, and C. Wulff-Nilsen.
Improved algorithms
for min cut and max flow in undirected planar graphs. In Proceedings of the 43rd ACM
Symposium on Theory of Computing, pages 313–322, 2011.
[34] L. Janiga and V. Koubek. Minimum cut in directed planar networks. Kybernetika, 28(1):37–
49, 1992.
[35] H. Kaplan, S. Mozes, Y. Nussbaum, and M. Sharir. Submatrix maximum queries in Monge
In Proceedings of the 23rd
matrices and Monge partial matrices, and their applications.
Annual ACM-SIAM Symposium on Discrete Algorithms, pages 338–355, 2012.
22
[36] H. Kaplan and Y. Nussbaum. Minimum s-t cut in undirected planar graphs when the
source and the sink are close. In T. Schwentick and C. Dürr, editors, Proceedings of the 28th
International Symposium on Theoretical Aspects of Computer Science, pages 117–128, 2011.
[37] D. R. Karger. Minimum cuts in near-linear time. Journal of the ACM, 47(1):46–76, 2000.
[38] K.-i. Kawarabayashi and M. Thorup. Deterministic global minimum cut of a simple graph
In Proceedings of the 47th ACM Symposium on Theory of Computing,
in near-linear time.
pages 665–674, 2015.
[39] S. Khuller and J. Naor. Flow in planar graphs: A survey of recent results. In Planar Graphs,
DIMACS Series in Discrete Math and Theoretical Computer Science 9, pages 59–84. AMS,
1993.
[40] P. N. Klein. Multiple-source shortest paths in planar graphs.
In Proceedings of the 16th
Annual ACM-SIAM Symposium on Discrete Algorithms, pages 146–155, 2005.
[41] P. N. Klein, S. Mozes, and C. Sommer. Structured recursive separator decompositions
for planar graphs in linear time. In Proceedings of the 45th ACM Symposium on Theory of
Computing, pages 505–514, 2013.
[42] P. N. Klein, S. Mozes, and O. Weimann. Shortest paths in directed planar graphs with neg-
ative lengths: A linear-space O(n log2 n)-time algorithm. ACM Transactions on Algorithms,
6(2):30.1–30.18, 2010.
[43] J. Ł acki, Y. Nussbaum, P. Sankowski, and C. Wulff-Nilsen. Single source - all sinks max
flows in planar digraphs. In Proceedings of the 53rd Annual IEEE Symposium on Foundations
of Computer Science, pages 599–608, 2012.
[44] J. Ł acki and P. Sankowski. Min-cuts and shortest cycles in planar graphs in O(n log log n)
time. In Proceedings of the 19th Annual European Symposium on Algorithms, pages 155–166,
2011.
[45] H.-C. Liang. Minimum cuts and shortest cycles in directed planar graphs via shortest
non-crossing paths. Master's thesis, National Taiwan University, July 2015.
[46] H.-C. Liang and H.-I. Lu. Minimum cuts and shortest cycles in directed planar graphs via
noncrossing shortest paths. SIAM Journal on Discrete Mathematics, 31(1):454–476, 2017.
[47] A. Lingas and E.-M. Lundell. Efficient approximation algorithms for shortest cycles in
undirected graphs. Information Processing Letters, 109(10):493–498, 2009.
[48] R. J. Lipton and R. E. Tarjan. A separator theorem for planar graphs. SIAM Journal on
Applied Mathematics, 36:177–189, 1979.
[49] B. Monien. The complexity of determining a shortest cycle of even length. Computing,
31(4):355–369, 1983.
[50] R. Motwani and P. Raghavan. Randomized Algorithms. Cambridge University Press, 1995.
[51] S. Mozes, C. Nikolaev, Y. Nussbaum, and O. Weimann. Minimum cut of directed
planar graphs in O(n log log n) time. Computing Research Repository, December 2015.
http://arxiv.org/abs/1512.02068.
23
[52] S. Mozes and C. Wulff-Nilsen. Shortest paths in planar graphs with real lengths in
O(n log2 n/ log log n) time. In M. de Berg and U. Meyer, editors, Proceedings of the 18th An-
nual European Symposium on Algorithms, Lecture Notes in Computer Science 6347, pages
206–217. Springer, 2010.
[53] K. Mulmuley, U. V. Vazirani, and V. V. Vazirani. Matching is as easy as matrix inversion.
Combinatorica, 7(1):105–113, 1987.
[54] H. Nagamochi and T. Ibaraki. Computing edge-connectivity in multigraphs and capaci-
tated graphs. SIAM Journal on Discrete Mathematics, 5(1):54–66, 1992.
[55] J. B. Orlin. Max flows in O(nm) time, or better. In Proceedings of the 45th ACM Symposium
on Theory of Computing, pages 765–774, 2013.
[56] E. Papadopoulou. k-pairs non-crossing shortest paths in a simple polygon. International
Journal of Computational Geometry and Applications, 9(6):533–552, 1999.
[57] V. Polishchuk and J. S. B. Mitchell. Thick non-crossing paths and minimum-cost flows in
polygonal domains. In Proceedings of the 23rd ACM Symposium on Computational Geometry,
pages 56–65, 2007.
[58] J. H. Reif. Minimum s-t cut of a planar undirected network in O(n log2 n) time. SIAM
Journal on Computing, 12(1):71–81, 1983.
[59] L. Roditty and R. Tov. Approximating the girth. ACM Transactions on Algorithms, 9(2):15.1–
15.13, 2013.
[60] L. Roditty and V. Vassilevska Williams. Subquadratic time approximation algorithms for
the girth. In Proceedings of the 23rd Annual ACM-SIAM Symposium on Discrete Algorithms,
pages 833–845, 2012.
[61] M. Stoer and F. Wagner. A simple min-cut algorithm. Journal of the ACM, 44(4):585–591,
1997.
[62] J. Takahashi, H. Suzuki, and T. Nishizeki. Finding shortest non-crossing rectilinear paths
in plane regions. In Proceedings of the 4th International Symposium on Algorithms and Com-
putation, pages 98–107, 1993.
[63] J.-y. Takahashi, H. Suzuki, and T. Nishizeki. Shortest noncrossing paths in plane graphs.
Algorithmica, 16(3):339–357, 1996.
[64] V. Vassilevska Williams. Multiplying matrices faster than Coppersmith-Winograd.
In
Proceedings of the 44th ACM Symposium on Theory of Computing, pages 887–898, 2012.
[65] V. Vassilevska Williams and R. Williams. Subcubic equivalences between path, matrix
and triangle problems. In Proceedings of the 51st Annual IEEE Symposium on Foundations of
Computer Science, pages 645–654, 2010.
[66] K. Weihe. Edge-disjoint (s, t)-paths in undirected planar graphs in linear time. Journal of
Algorithms, 23(1):121–138, 1997.
[67] O. Weimann and R. Yuster. Computing the girth of a planar graph in O(n log n) time.
SIAM Journal on Discrete Mathematics, 24(2):609–616, 2010.
24
[68] C. Wulff-Nilsen. Algorithms for planar graphs and graphs in metric spaces. PhD thesis, Uni-
versity of Copenhagen, 2010.
[69] R. Yuster. A shortest cycle for each vertex of a graph. Information Processing Letters, 111(21-
22):1057–1061, 2011.
[70] R. Yuster and U. Zwick. Finding even cycles even faster. SIAM Journal on Discrete Mathe-
matics, 10(2):209–222, 1997.
25
|
1903.02999 | 2 | 1903 | 2019-04-23T17:26:24 | Tiers for peers: a practical algorithm for discovering hierarchy in weighted networks | [
"cs.DS"
] | Interactions in many real-world phenomena can be explained by a strong hierarchical structure. Typically, this structure or ranking is not known; instead we only have observed outcomes of the interactions, and the goal is to infer the hierarchy from these observations. Discovering a hierarchy in the context of directed networks can be formulated as follows: given a graph, partition vertices into levels such that, ideally, there are only edges from upper levels to lower levels. The ideal case can only happen if the graph is acyclic. Consequently, in practice we have to introduce a penalty function that penalizes edges violating the hierarchy. A practical variant for such penalty is agony, where each violating edge is penalized based on the severity of the violation. Hierarchy minimizing agony can be discovered in $O(m^2)$ time, and much faster in practice. In this paper we introduce several extensions to agony. We extend the definition for weighted graphs and allow a cardinality constraint that limits the number of levels. While, these are conceptually trivial extensions, current algorithms cannot handle them, nor they can be easily extended. We solve the problem by showing the connection to the capacitated circulation problem, and we demonstrate that we can compute the exact solution fast in practice for large datasets. We also introduce a provably fast heuristic algorithm that produces rankings with competitive scores. In addition, we show that we can compute agony in polynomial time for any convex penalty, and, to complete the picture, we show that minimizing hierarchy with any concave penalty is an NP-hard problem. | cs.DS | cs | Noname manuscript No.
(will be inserted by the editor)
Tiers for peers
a practical algorithm for discovering hierarchy in weighted
networks
Nikolaj Tatti
9
1
0
2
r
p
A
3
2
]
S
D
.
s
c
[
2
v
9
9
9
2
0
.
3
0
9
1
:
v
i
X
r
a
the date of receipt and acceptance should be inserted later
Abstract Interactions in many real-world phenomena can be explained by a
strong hierarchical structure. Typically, this structure or ranking is not known;
instead we only have observed outcomes of the interactions, and the goal is
to infer the hierarchy from these observations. Discovering a hierarchy in the
context of directed networks can be formulated as follows: given a graph, par-
tition vertices into levels such that, ideally, there are only edges from upper
levels to lower levels. The ideal case can only happen if the graph is acyclic.
Consequently, in practice we have to introduce a penalty function that penal-
izes edges violating the hierarchy. A practical variant for such penalty is agony,
where each violating edge is penalized based on the severity of the violation.
Hierarchy minimizing agony can be discovered in O(cid:0)m2(cid:1) time, and much faster
in practice. In this paper we introduce several extensions to agony. We extend
the definition for weighted graphs and allow a cardinality constraint that limits
the number of levels. While, these are conceptually trivial extensions, current
algorithms cannot handle them, nor they can be easily extended. We solve the
problem by showing the connection to the capacitated circulation problem,
and we demonstrate that we can compute the exact solution fast in practice
for large datasets. We also introduce a provably fast heuristic algorithm that
produces rankings with competitive scores. In addition, we show that we can
compute agony in polynomial time for any convex penalty, and, to complete
the picture, we show that minimizing hierarchy with any concave penalty is
an NP-hard problem.
Keywords Hierarchy discovery; agony; capacitated circulation; weighed
graphs
The research described in this paper builds upon and extends the work appearing in ICDM15
as [20].
Nikolaj Tatti
Helsinki Institute for Information Technology (HIIT) and
Department of Information and Computer Science, Aalto University, Finland
E-mail: [email protected]
2
1 Introduction
Nikolaj Tatti
Interactions in many real-world phenomena can be explained by a strong hi-
erarchical structure. As an example, it is more likely that a line manager in
a large, conservative company will write emails to her employees than the
other way around. Typically, this structure or ranking is not known; instead
we only have observed outcomes of the interactions, and the goal is to infer
the hierarchy from these observations. Discovering hierarchies or ranking has
applications in various domains: (i) ranking individual players or teams based
on how well they play against each other [4], (ii) discovering dominant animals
within a single herd, or ranking species based on who-eats-who networks [9],
(iii) inferring hierarchy in work-places, such as, U.S. administration [12], (iv )
summarizing browsing behaviour [11], (v ) discovering hierarchy in social net-
works [7], for example, if we were to rank twitter users, the top-tier users
would be the content-providers, middle-tiers would spread the content, while
the bottom-tier are the consumers.
We consider the following problem of discovering hierarchy in the context
of directed networks: given a directed graph, partition vertices into ranked
groups such that there are only edges from upper groups to lower groups.
Unfortunately, such a partitioning is only possible when the input graph
has no cycles. Consequently, a more useful problem definition is to define a
penalty function p on the edges. This function should penalize edges that are
violating a hierarchy. Given a penalty function, we are then asked to find the
hierarchy that minimizes the total penalty.
The feasibility of the optimization problem depends drastically on the
choice of the penalty function. If we attach a constant penalty to any edge
that violates the hierarchy, that is, the target vertex is ranked higher or equal
than the source vertex, then this problem corresponds to a feedback arc set
problem, a well-known NP-hard problem [2], even without a known constant-
time approximation algorithm [5].
A more practical variant is to penalize the violating edges by the severity
of their violation. That is, given an edge (u, v) we compare the ranks of the
vertices r(u) and r(v) and assign a penalty of max(r(u) − r(v) + 1, 0). Here,
the edges that respect the hierarchy receive a penalty of 0, edges that are in
the same group receive a penalty of 1, and penalty increases linearly as the
violation becomes more severe, see Figure 1. This particular score is referred as
agony. Minimizing agony was introduced by Gupte et al. [7] where the authors
is the number of edges. A faster discovery algorithm with the computational
provide an exact O(cid:0)nm2(cid:1) algorithm, where n is the number of vertices and m
complexity of O(cid:0)m2(cid:1) was introduced by Tatti [19]. In practice, the bound
O(cid:0)m2(cid:1) is very pessimistic and we can compute agony for large graphs in
reasonable time.
In this paper we specifically focus on agony, and provide the following main
extensions for discovering hierarchies in graphs.
weighted graphs: We extend the notion of the agony to graphs with
weighted edges. Despite being a conceptually trivial extension, current algo-
Tiers for peers
3
rithms [7, 19] for computing agony are specifically design to work with unit
weights, and cannot be used directly or extended trivially. Consequently, we
need a new approach to minimize the agony, and in order to do so, we demon-
strate that we can transform the problem into a capacitated circulation, a
classic graph task known to have a polynomial-time algorithm.
cardinality constraint: The original definition of agony does not restrict
the number of groups in the resulting partition. Here, we introduce a cardinal-
ity constraint k and we are asking to find the optimal hierarchy with at most
k groups. This constraint works both with weighted and non-weighted graphs.
Current algorithms for solving agony cannot handle cardinality constraints.
Luckily, we can enforce the constraint when we transform the problem into a
capacitated circulation problem.
fast heuristic: We introduce a fast divide-and-conquer heuristic. This
heuristic is provably fast, see Table 1, and -- in our experiments -- produces
competitive scores when compared to the optimal agony.
convex edge penalties: Minimizing agony uses linear penalty for edges.
We show that if we replace the linear penalty with a convex penalty, see Fig-
ure 1, we can still solve the problem in polynomial time by the capacitated
circulation solver. However, this extension increases the computational com-
plexity.
concave edge penalties: To complete the picture, we also study concave
edge penalties, see Figure 1. We show that in this case discovering the optimal
hierarchy is an NP-hard problem. This provides a stark difference between
concave and convex edge penalties.
canonical solution: A hierarchy minimizing agony may not be unique.
For example, given a DAG any topological sorting of vertices will give you an
optimal agony of 0. To address this issue we propose to compute a canonical
solution, where, roughly speaking, the vertices are ranked as high as possible
without compromising the optimality of the solution. We demonstrate that
this solution is unique, it creates a hierarchy with the least amount of groups,
and that we can compute it in O(n log n + m) time, if we are provided with the
optimal solution and the flow resulted from solving the capacitated circulation.
y
t
l
a
n
e
p
e
g
d
e
6
4
2
0
−3
−2
−1
3
rank difference, r(u) − r(v)
0
1
2
convex (polynomial)
linear = agony (polynomial)
concave (NP-hard)
constant = FAS (NP-hard)
4
5
Fig. 1 A toy example of edge penalties as a function of the rank difference between the
vertices.
4
Nikolaj Tatti
Table 1 Summary of running times of different algorithms for computing agony: n is the
number of vertices, m is the number of edges, k is the number of allowed ranks.
Algorithm
variant
input type
running time
Exact
Exact
Exact
Canonical
Heuristic
Heuristic
Heuristic
Heuristic
plain
speed-up
speed-up
--
plain
plain
SCC
SCC
unweighted
weighted
optimal rank and the flow
no cardinality constraint
cardinality constraint
no cardinality constraint
cardinality constraint
O(m log n(m + n log n))
O(m(min(kn, m) + n log n))
O(m log n(m + n log n))
O(m + n log n)
O(m log n)
O(cid:0)m log n + k2n(cid:1)
O(m log n)
O(cid:0)m log n + k2n + km log n(cid:1)
This paper is an extension of a conference paper [20]. In this extension we
significantly speed-up the exact algorithm, propose a provably fast heuristic,
and provide a technique for selecting unique canonical solutions among the
optimal rankings.
The rest of the paper is organized as follows. We introduce the notation
and formally state the optimization problem in Section 2. In Section 3 we
transform the optimization problem into a capacitated circulation problem,
allowing us a polynomial-time algorithm, and provide a speed-up in Section 4.
In Section 5 we discuss alternative edge penalties. We demonstrate how to
extract a canonical optimal solution in Section 6. We discuss the related work
in Section 8 and present experimental evaluation in Section 9. Finally, we
conclude the paper with remarks in Section 10.
2 Preliminaries and problem definition
We begin with establishing preliminary notation and then defining the main
problem.
The main input to our problem is a weighted directed graph which we will
denote by G = (V, E, w), where w is a function mapping an edge to a real
positive number. If w is not provided, we assume that each edge has a weight
of 1. We will often denote n = V and m = E.
As mentioned in the introduction, our goal is to partition vertices V . We
express this partition with a rank assignment r, a function mapping a vertex
to an integer. To obtain the groups from the rank assignment we simply group
the vertices having the same rank.
Given a graph G = (V, E) and a rank assignment r, we will say that an edge
(u, v) is forward if r(u) < r(v), otherwise edge is backward, even if r(u) = r(v).
Ideally, rank assignment r should not have backward edges, that is, for any
(u, v) ∈ E we should have r(u) < r(v). However, this is only possible when
G is a DAG. For a more general case, we assume that we are given a penalty
function p, mapping an integer to a real number. The penalty for a single edge
(u, v) is then equal to p(d), where d = r(u)−r(v). If p(d) = 0, whenever d < 0,
then the forward edges will receive 0 penalty.
Tiers for peers
5
We highlight two penalty functions. The first one assigns a constant penalty
to each backward edge,
pc(d) =(1
0
if d ≥ 0
otherwise
.
The second penalty function assigns a linear penalty to each backward edge,
pl (d) = max(0, d + 1)
.
For example, an edge (u, v) with r(u) = r(v) is penalized by pl (r(u) − r(v)) =
1, the penalty is equal to 2 if r(u) = r(v) + 1, and so on.
Given a penalty function and a rank assignment we can now define the the
score for the ranking to be the sum of the weighted penalties.
Definition 1 Assume a weighted directed graph G = (V, E, w) and a rank
assignment r. Assume also a cost function p mapping an integer to a real
number. We define a score for a rank assignment to be
q(G, r, p) = Xe=(u,v)∈E
w(e)p(r(u) − r(v))
.
We will refer the score q(G, r, pl ) as agony.
Example 1 Consider the left ranking r1 of a graph G given in Figure 2. This
ranking has 5 backward edges, consequently, the penalty is q(G, r1, pc) = 5.
On the other hand, there are 2 edges, (i, a) and (e, g), having the agony of 1.
Moreover, 2 edges has agony of 2 and (d, b) has agony of 3. Hence, agony is
equal to
q(G, r1, pl ) = 2 × 1 + 2 × 2 + 1 × 3 = 10 .
The agony for the right ranking r2 is q(G, r2, pl ) = 7. Consequently, r2 yields
a better ranking in terms of agony.
a
c
e
f
h
g
i
b
d
i
b
d
f
g
a
c
e
h
Fig. 2 Toy graphs. Backward edges are represented by dotted lines, while the forward edges
are represented by solid lines. Ranks are represented by dashed grey horizontal lines.
We can now state our main optimization problem.
Problem 1 Given a graph G = (V, E, w), a cost function p, and an integer
k, find a rank assignment r minimizing q(r, G) such that 0 ≤ r(v) ≤ k − 1 for
every v ∈ V . We will denote the optimal score by q(G, k, p).
6
Nikolaj Tatti
We should point out that we have an additional constraint by demanding
that the rank assignment may have only k distinct values, that is, we want to
find at most k groups. Note that if we assume that the penalty function is non-
decreasing and does not penalize the forward edges, then setting k = V is
equivalent of ignoring the constraint. This is the case since there are at most V
groups and we can safely assume that these groups obtain consecutive ranks.
However, an optimal solution may have less than k groups, for example, if G
has no edges and we use pl (or pc), then a rank assigning each vertex to 0
yields the optimal score of 0. We should also point out that if using pc, there
is always an optimal solution where each vertex has its own rank. This is not
the case for agony.
It is easy to see that minimizing q(G, pc) is equivalent to finding a directed
acyclic subgraph with as many edges as possible. This is known as Feedback
Arc Set (FAS) problem, which is NP-complete [2].
On the other hand, if we assume that G has unit weights, and set k = V ,
then minimizing agony has a polynomial-time O(cid:0)m2(cid:1) algorithm [7, 19].
3 Computing agony
In this section we present a technique for minimizing agony, that is, solving
Problem 1 using pl as a penalty. In order to do this we show that this problem
is in fact a dual problem of the known graph problem, closely related to the
minimum cost max-flow problem.
3.1 Agony with shifts
We begin with an extension to our optimization problem.
Problem 2 (Agony-with-shifts) Given a graph G = (V, E, w, s), where w
maps an edge to a, possibly infinite, non-negative value, and s maps an edge
to a possibly negative integer, find a rank assignment r minimizing
Xe=(u,v)∈E
w(e) × max(r(u) − r(v) + s(e), 0)
.
We denote the optimal sum with q(G).
In order to transform the problem of minimizing agony to Agony-with-
shifts, assume a graph G = (V, E, w) and an integer k. We define a graph
H = (W, F, w, s) as follows. The vertex set W consists of 2 groups: (i) V
vertices, each vertex corresponding to a vertex in G (ii) 2 additional vertices
α and ω. For each edge e = (u, v) ∈ E, we add an edge f = (u, v) to F .
We set w(f ) = w(e) and s(f ) = 1. We add edges (v, ω) and (α, v) for every
v ∈ V with s(v, ω) = s(α, v) = 0 and w(v, ω) = w(α, v) = ∞. Finally we add
(ω, α) with s(ω, α) = 1 − k and w(ω, α) = ∞. We will denote this graph by
H (G, k) = H.
Tiers for peers
7
a
1
b
2
1
1
G
c
d
α
1(1)
a
b
1
(
2
)
1(1)
d
1
(
2
)
c
H (G, 4)
ω
−3(∞)
Fig. 3 Toy graph G and the related circulation graph H (G, 4). Edge costs and shifts for
(α, v) and (v, ω) are omitted to avoid clutter.
Example 2 Consider G = (V, E), a graph with 4 vertices and 4 edges, given in
Figure 3. Set cardinality constraint k = 4. In order to construct H (G, k) we
add two additional vertices α and ω to enforce the cardinality constraint k.
We set edge costs to −1 and edges capacities to be the weights of the input
graph. We connect α and ω with a, b, c, and d, and finally we connect ω to α.
The resulting graph is given in Figure 3.
3.2 Agony is a dual problem of Circulation
Minimizing agony is closely related to a circulation problem, where the goal is
to find a circulation with a minimal cost satisfying certain balance equations.
Problem 3 (Capacitated circulation) Given a graph G = (V, E, c, s),
where c maps an edge to a, possibly infinite, non-negative value, and s maps
an edge to a possibly negative integer, find a flow f such that 0 ≤ f (e) ≤ c(e)
for every e ∈ E and
Xe=(v,u)∈E
f (e) = Xe=(u,v)∈E
f (e),
for every v ∈ V
maximizing
We denote the above sum as circ(G).
s(e)f (e)
.
Xe∈E
This problem is known as capacitated circulation problem, and can be
solved in O(m log n(m + n log n)) time with an algorithm presented by Orlin
[14]. We should stress that we allow s to be negative. We also allow capacities
for certain edges to be infinite, which simply means that f (e) ≤ c(e) is not
enforced, if c(e) = ∞.
The following proposition shows the connection between the agony and the
capacitated circulation problem.
8
Nikolaj Tatti
Proposition 1 Assume a weighted directed graph with shifts, G = (V, E, w, s).
Then q(G) = circ(G).
Proof Let G = (V, E, w, s). To prove this result we will show that computing
circ(G) is a linear program, whose dual corresponds to optimizing Agony-
with-shifts. In order to do this, we first express a general Capacitated
circulation problem as a linear program,
s(u, v)f (u, v)
f (u, v),
maximize X(u,v)∈E
f (v, u) = X(u,v)∈V
w(u, v) ≥ f (u, v) ≥ 0,
X(v,u)∈V
such that
for every v ∈ V ,
for every (u, v) ∈ E .
This program has the following dual program,
η(u, v)w(u, v)
minimize X(u,v)∈E
such that for every (u, v) ∈ E
π(v) − π(u) + η(u, v) ≥ s(u, v),
π(v) − π(u) ≥ s(u, v),
η(u, v) ≥ 0,
if w(u, v) < ∞,
if w(u, v) = ∞,
(1)
which is optimized over the variables π and η.
If π are integers, then they correspond to the ranking r. Moreover, η(u, v) =
max(π(u)−π(v)+s(u, v), 0). So that, w(u, v)η(u, v) corresponds to the penalty
term in the sum of Agony-with-shifts, and the objective function of the
dual program corresponds exactly to the objective of Agony-with-shifts.
To complete the proof we need to show that there is an optimal integer-
valued dual solution π and η. This result follows from the fact that the con-
straints of the dual form an arc-vertex incidence matrix, which is known to
be totally unimodular [15, Corollary of Theorem 13.3], Since s(u, v) are in-
tegers, Theorem 13.2 in [15] implies that there is an optimal solution with
integer-valued π, completing the proof. ⊓⊔
3.3 Algorithm for minimizing agony
Proposition 1 states that we can compute agony but it does not provide direct
means to discover an optimal rank assignment. However, a closer look at the
proof reveals that minimizing agony is a dual problem of Capacitated cir-
culation. That is, if we were to solve the dual optimization problem given in
Equation 1, then we can extract the optimal ranking from the dual parameters
π by setting r(v) = π(v)− π(α) for v ∈ V , where α is the special vertex added
during the construction of H.
Tiers for peers
9
Luckily, the algorithms for solving Capacitated circulation by Ed-
monds and Karp [3] or by Orlin [14] in fact solve Equation 1 and are guar-
anteed to have integer-valued solution as long as the capacities s(u, v) are
integers, which is the case for us.
If we are not enforcing the cardinality constraint, that is, we are solving
q(G, k) with k = V , we can obtain a significant speed-up by decomposing G to
strongly connected components, and solve ranking for individual components.
Proposition 2 Assume a graph G, and set k = V . Let {Ci} be the strongly
connected components of G, ordered in a topological order. Let ri be the rank-
ing minimizing q(G(Ci),Ci). Let bi = Pi−1
j=1 Cj. Then the ranking r(v) =
ri(v) + bi, where Ci is the component containing v, yields the optimal score
q(G, k).
Proof Note that max r(v) ≤ k, hence r is a valid ranking. Let r′ be the ranking
minimizing q(G, k). Let r′
i be the projection of the ranking to Ci. Then
q(G, r′) ≥Xi=1
q(G(Ci), r′
i) ≥Xi=1
q(G(Ci), ri) = q(G, r) ,
where the last equality holds because any cross-edge between the components
is a forward edge. ⊓⊔
4 Speeding up the circulation solver
In this section we propose a modification to the circulation solver. This modi-
fication provides us with a modest improvement in computational complexity,
and -- according to our experimental evaluation -- significant improvement in
running time in practice.
Before explaining the modification, we first need to revisit the original
Orlin's algorithm. We refer the reader to [14] for a complete expose.
The solver actually solves a slightly different problem, namely, an unca-
pacitated circulation.
Problem 4 (Circulation) Given a directed graph F = (W, A, t, b) with
weights on edges and biases on vertices, find a flow f such that 0 ≤ f (e)
for every e ∈ A and
X(v,u)∈A
f (v, u) − X(u,v)∈A
(2)
f (u, v) = b(v),
for every v ∈ W
minimizing
t(u, v)f (u, v)
.
X(u,v)∈A
10
Nikolaj Tatti
To map our problem to Circulation, we us the trick described by Orlin
[14]: we replace each capacitated edge e = (v, w) with a vertex u and two
edges (v, u) and (w, u). We set b(u) = −c(e), and add c(v, w) to b(w). The
costs are set to t(v, u) = max(−s(e), 0) and t(w, u) = max(s(e), 0). For each
uncapacitated edge (v, w), we connect v to w with t(v, w) = −s(v, w).1
From now on, we will write H = H (G, k), and F = (W, A, s, b) to be the
graph modified as above. We split W to W1 and W2: W1 are the original
vertices in H, while W2 are the vertices rising from the capacitated edges.
We also write n and m to be the number of vertices and edges in H,
respectively, and n′ and m′ to be the number of vertices and edges in F ,
respectively. Note that n′, m′ ∈ O(m).
Consider the dual of uncapacitated circulation.
Problem 5 (dual to Circulation) Given a directed graph F = (W, A, s, b)
with weights on edges and biases on vertices, find dual variables π on vertices
maximizing
such that
b(v)π(v)
X(u,v)∈A
t(e) + π(w) − π(v) ≥ 0,
for every e = (v, w) ∈ A .
(3)
The standard linear programming theory states that f and π satisfying
Eq. 2 -- 3 are optimal solutions to their respective problems if and only if the
slackness conditions hold,
(t(e) + π(w) − π(v))f (e) = 0,
for every e = (v, w) ∈ A .
(4)
The main idea behind Orlin's algorithm is to maintain a flow f and a dual
π satisfying Eqs. 3 -- 4, and then iteratively enforce Eq. 2. More specifically, we
first define an excess of a vertex to be
e(v) = b(v) + X(w,v)∈A
f (w, v) − X(v,w)∈A
f (v, w)
.
Our goal is to force e(v) = 0 for every v. This is done in gradually in multiple
iterations. Assume that we are given ∆, a granularity which we will use to
modify the flow. The following steps are taken: (i) We first construct a residual
graph R which consists of all the original edges, and reversed edges for all
edges with positive flow. (ii) We then select a source s with e(s) ≥ α∆2, and
construct a shortest path tree T in R, weighted by t(e) + π(w) − π(v), for
e = (v, w). (iii) The dual variables are updated to π(v) − d(v), where d is the
shortest distance from s to v. (iv) We select a sink r with e(r) ≤ −α∆, and
augment the flow along the path in T from r to s. This is repeated until there
are no longer viable options for s or r. After that we half ∆, and repeat.
1 The reason for the minus sign is that we expressed Circulation as a minimization
problem and Capacitated circulation as a maximization problem.
2 Here α is a fixed parameter 1/2 < α < 1, we use α = 3/4.
Tiers for peers
11
To guarantee polynomial convergence, we also must contract edges for
which f (e) ≥ 3n′∆, where n′ is the number of vertices in the (original) input
graph. Assume that we contract (v, w) into a new vertex u. We set π(u) = π(v),
b(u) = b(v)+b(w). We delete the edge (v, w), and edges adjacent to v and w are
migrated to u; the cost of an edge t(w, x) must be changed to t(w, x) + t(v, w),
and similarly the cost of an edge t(x, w) must be changed to t(x, w) − t(v, w).
A high-level pseudo-code is given in Algorithm 1.
Algorithm 1: Orlin's algorithm for solving Circulation.
1 ∆ ← min(max e(v), max −e(v));
2 while there is excess do
3
4
5
6
7
8
9
10
11
contract any edges with f (e) ≥ 3n∆;
while max e(v) ≥ α∆ and min e(v) ≤ −α∆ do
s ← a vertex with e(s) ≥ α∆;
r ← a vertex with e(r) ≤ −α∆;
T ← shortest path tree from s in residual graph, weighted by
t(e) + π(w) − π(v);
update π using T ;
P ← path in T from r to s;
augment flow by ∆ along P ;
∆ ← ∆/2;
The bottleneck of this algorithm is computing the shortest path tree. This
is the step that we will modify. In order to do this we first point out that
Orlin's algorithm relies on two things that inner loop should do: (i) Eqs. 3 -- 4
must be maintained, and (ii) path augmentations are of granularity ∆, after
the augmentations there should not be a viable vertex for a source or a viable
vertex for a sink. As long as these two conditions are met, the correctness
proof given by Orlin [14] holds.
Our first modification is instead of selecting one source s, we select all
possible sources, S ← {v ∈ V e(v) ≥ α∆}, and compute the shortest path
tree using S as roots. Once this tree is computed, we subtract the shortest
distance from π, select a sink t, and augment flow along the path from t to
some root s ∈ S.
π are modified.
The following lemma guarantees that Eqs. 3 -- 4 are maintained when f and
Lemma 1 Let f and π be flow and dual variables satisfying the slackness
conditions given in Eq. 4. Let S be a set of vertices. Define d(v) be the shortest
distance from S to v in the residual graph with weighted edges t(e)+π(w)−π(v).
Let π′ = π − d. Then π′ satisfy Eq. 3, and f and π′ respect the slackness
conditions in Eq. 4. Moreover, t(e) + π′(w) − π′(v) = 0 for every edge in the
shortest path tree.
Note that since we modify f only along the edges of the shortest path tree,
this lemma guarantees that Eq. 4 is also maintained when we augment f . The
12
Nikolaj Tatti
proof of this lemma is essentially the same as the single-source version given
by Orlin [14].
Proof Let e = (v, w) ∈ E. Then e is also in residual graph, and d(w) ≤
d(v) + t(e) + π(w) − π(v). This implies
t(e) + π′(w) − π′(v) = t(e) + π(w) − d(w) − π(v) + d(v) ≥ 0,
proving the first claim. If e is in the shortest path tree, then d(w) = d(v) +
t(e) + π(w) − π(v), which implies the third claim, t(e) + π(w) − π(v) = 0.
(w, v) = e′ is also in residual graph, we must have d(v) = d(w). Thus,
To prove the second claim, if f (e) > 0, then t(e) + π(w) − π(v) = 0. Since
t(e) + π′(w) − π′(v) = t(e) + π(w) − d(w) − π(v) + d(v) = 0 .
This completes the proof. ⊓⊔
Once we augment f , we need to update the shortest path tree. There are
three possible updates: (i) adding a flow may result in a new backward edge in
the residual graph, (ii) reducing a flow may result in a removing a backward
edge in the residual graph, and (iii) deleting a source from S requires that the
tree is updated.
In order to update the tree we will use an algorithm by Ramalingam and
Reps [16] to which we will refer as rr. The pseudo-code for the modified solver
is given in Algorithm 2.
Algorithm 2: A modified algorithm for Circulation.
1 ∆ ← min(max e(v), max −e(v));
2 while there is excess do
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
contract any edges with f (e) ≥ 3n′∆;
S ← {v ∈ V e(v) ≥ α∆};
Q ← {v ∈ V e(v) ≤ −α∆};
T ← shortest path tree from S in residual graph, weighted by t(e) + π(w) − π(v);
update π using T , see Lemma 1;
while S 6= ∅ and Q 6= ∅ do
select r ∈ Q;
P ← path in T from r to some s ∈ S;
augment flow by ∆ along P ;
update residual graph;
if e(s) < α∆ then delete s from S;
if e(r) > −α∆ then delete r from Q;
update T using [16];
update π using T , see Lemma 1;
∆ ← ∆/2;
Before going further, we need to address one technical issue. rr requires
that edge weights are positive, whereas we can have weights equal to 0. We
solve this issue by adding ǫ = 1/n′ to each edge. Since the original weights
Tiers for peers
13
are integers and a single path may have n′ − 1 edges, at most, the obtained
shortest path tree is a valid shortest path tree for the original weights. We use
ǫ only for computing and updating the tree; we will not use it when we update
the dual variables.
In order to update the tree, first note that the deleting the source s from S
is essentially the same as deleting an edge: computing the tree using S as roots
is equivalent to having one auxiliary root, say σ, with only edges connecting
σ to S. Removing s from S is then equivalent to deleting an edge (σ, s).
The update is done by first adding the new edges, and then deleting the
necessary edges. We first note that the edge additions do not require any
updates by rr. This is because the internal structure of rr is a subgraph
of all edges that can be used to form the shortest path. Any edge that is
added will be from a child to a parent, implying that it cannot participate in
a shortest path.3
Proposition 3 Algorithm 2 runs in O(m(min(kn, m) + n log n)) time, as-
suming G is not weighted.
To prove the result we need the following lemma.
Lemma 2 At any point of the algorithm, the dual variables π satisfy π(v) −
π(u) ≤ k for any u, v.
Proof Let us first prove that this result holds if we have done no edge contrac-
tions. Let α and ω be the vertices in H, enforcing the cardinality constraint.
Assume that v and w are both in W1. Then Eq. 3 guarantees that π(u) ≥ π(α)
and π(ω) ≥ π(v) implying
π(v) − π(u) ≤ π(ω) − π(α) ≤ t(ω, α) = k − 1
.
Now assume that v (and/or u) is in W2. Then the shortest path tree connects
it to a vertex x ∈ W1, and either π(u) = π(x) or π(u) = π(x) − 1. This leads
to that the difference π(v) − π(u) can be at most k.
To see why the lemma holds despite edge contractions, note that we can
always unroll the contractions to the original graph, and obtain π′ that sat-
isfies Eq. 3. Moreover, if x is a new vertex resulted from a contraction, after
unrolling, there is a vertex u ∈ W such that π(x) = π′(u). This is because
when we create x, we initialize π(x) to be dual of one contracted vertices.
Consequently, the general case reduces to the first case. ⊓⊔
Proof (of Proposition 3) Since b(v) = −1, for v ∈ W2 and b(v) ≥ 0 for v ∈ W1,
we have ∆ = 1, and after a single iteration e(v) = 0, for v ∈ W . So, we need
only one outer iteration. Consequently, we only need to show that the inner
loop needs O(m(min(kn, m) + n log n)) time.
Let us write O to be the vertices who are either in W1, or, due to a contrac-
tion, contain a vertex in W1. Let Pi be the path selected during the ith itera-
i to be the number of vertices whose distance is changed4
tion. Let us write n′
3 The edge can participate later when we delete edges.
4 taking into account the ǫ trick
14
Nikolaj Tatti
during the ith iteration of the inner loop; let mi be the number of edges adja-
cent to these vertices. Finally, let us write ni to be the number of vertices in
O whose distance is changed.
i log n′
i log n′
i) time.
Ramalingam and Reps [16] showed that updating a tree during the ith
iteration requires O(mi + n′
i) time. More specifically, the update algo-
rithm first detects the affected vertices in O(mi) time, and then computes the
new distances using a standard Dijkstra algorithm with a binomial heap in
O(mi + n′
We can optimize this to O(mi + n′
i + ni log ni) by performing a trick sug-
gested by Orlin: Let X be the vertices counted towards ni and let Y be the
remaining vertices counted towards n′
i. A vertex in Y is either in a path be-
tween two vertices in X, or is a leaf. In the latter case it may be only connected
to only two (known) vertices. We can first compute the distances for X by frog-
leaping the vertices in Y in O(mi + ni log ni) time. This gives us the updated
distances for X and for vertices in Y that are part of some path. Then we can
proceed to update the leaf vertices in Y in O(mi + ni) time.
The total running time of an inner loop is then
Pi + mi + n′
i + ni log ni! ⊆ O Xi
i + ni log n! .
O Xi
First note that we can have at most O(m) terms in the sum. This is because
we either have P max(e(i), 0) ≤ 2∆αn′ or P max(−e(i), 0) ≤ 2∆αn′ due to
to the previous outer loop iteration, and since the contractions can only reduce
these terms.
Pi + mi + n′
A path from a leaf to a root in T cannot contain two consecutive vertices
that are outside O. Hence, the length of a path is at most O(n).
Let us now bound the number of times a single vertex, say v, needs to be
updated. Assume that we have changed the distance but the dual π(v) has
not changed. In other words, we have increased the ǫ part of the distance.
This effectively means that π(v) remained constant but we have increased the
number of edges from the vertex to the root. Since we can have at most O(n)
long path, we can have at most O(n) updates without before updating π(v).
Note that at least one root, say s ∈ S, will not have its dual updated until the
very last iteration. Lemma 2 now implies that we can update, that is, decrease,
π(v) only O(k) times. Consequently, we can only update v O(nk) times.
i ∈ O(mnk), Pi ni ∈ O(cid:0)n2k(cid:1), and
This immediately implies that Pi n′
Pi mi ∈ O(mnk). Since path lengths are O(n), we also havePi Pi ∈ O(mn).
This gives us a total running time of
O(cid:0)mn + mnk + n2k log n(cid:1) = O(nk(m + n log n))
.
We obtain the final bound by alternatively bounding rr with O(m + n log n),
and observing that you need only O(m) updates. ⊓⊔
The theoretical improvement is modest: we essentially replaced m with
min(nk, m). However, in practice this is a very pessimistic bound, and we
Tiers for peers
15
will see that this approach provides a significant speed-up. Moreover, this
result suggests -- backed up by our experiments -- that the problem is easier
for smaller values of k. This is opposite to the behavior of the original solver
presented in [20]. Note also, that we assumed that G has no weights. If we
have integral weights of at most ℓ, then the running time increases by O(log ℓ)
time.
5 Alternative penalty functions
We have shown that we can find ranking minimizing edge penalties pl in poly-
nomial time. In this section we consider alternative penalties. More specifically,
we consider convex penalties which are solvable in polynomial time, and show
that concave penalties are NP-hard.
5.1 Convex penalty function
We say that the penalty function is convex if p(x) ≤ (p(x − 1) + p(x + 1))/2
for every x ∈ Z.
Let us consider a penalty function that can be written as
ℓ
ps (x) =
max(0, αi(x − βi)),
Xi=1
where αi > 0 and βi ∈ Z for 1 ≤ i ≤ ℓ. This penalty function is convex. On
the other hand, if we are given a convex penalty function p such that p(x) = 0
for x < 0, then we can safely assume that an optimal rank assignment will
have values between 0 and V − 1. We can define a penalty function ps with
ℓ ≤ V terms such that ps (x) = p(x) for x < V . Consequently, finding an
optimal rank assignment using ps will also yield an optimal rank assignment
with respect to p.
Note that pl is a special case of ps . This hints that we can solve q(G, k, ps )
with a technique similar to the one given in Section 3. In fact, we can map
this problem to Agony-with-shifts. In order to do this, assume a graph
G = (V, E, w) and an integer k. Set n = V and m = E. We define a graph
H = (W, F, w, s) as follows. The vertex set W consists of 2 groups: (i) n
vertices, each vertex corresponding to a vertex in G (ii) 2 additional vertices
α and ω. For each edge e = (v, w) ∈ E, we add ℓ edges fi = (u, v) to F . We
set s(fi) = −βi and w(fi) = αiw(e). We add edges to α and ω to enforce
the cardinality constraint, as we did in Section 3.1. We denote this graph by
H (G, k, ps ) = H.
Example 3 Consider a graph G given in Figure 4 and a penalty function
ps (d) = max(0, d + 1) + 2 max(0, d − 3). The graph H = H (G, 3, ps ) has 5
vertices, the original vertices and the two additional vertices. Each edge in G
results in two edges in H. This gives us 6 edges plus the 7 edges adjacent to
α or ω. The graph H without α and ω is given in Figure 4.
16
Nikolaj Tatti
a
1
b
2
1
6
a
1(1)
−3(2)
b 3
-
3
(
4
)
1
(
2
)
3(2)
−
1(1)
G
c
H (G, 3, ps )
c
3
Fig. 4 Toy graph G and the related circulation graph H (G, 3, ps ). To avoid clutter the
vertices α and ω and the adjacent edges are omitted.
Finally, let us address the computational complexity of the problem. The
circulation graph H (G, k, ps ) will have n + 2 vertices and ℓm + n edges. If the
penalty function p is convex, then we need at most ℓ = n functions to represent
p between the range of [0, n − 1]. Moreover, if we enforce the cardinality con-
straint k, we need only ℓ = k components. Consequently, we will have at most
dm + n, edges where d = min(k, ℓ, n) for ps , and d = min(k, n) for a convex
penalty p. This gives us computational time of O(dm log n(dm + n log n)).
5.2 Concave penalty function
We have shown that we can solve Problem 1 for any convex penalty. Let us
consider concave penalties, that is penalties for which p(x) ≥ (p(x − 1) +
p(x + 1))/2. There is a stark difference compared to the convex penalties as
the minimization problem becomes computationally intractable.
Proposition 4 Assume a monotonic penalty function p : Z → R such that
p(x) = 0 for x < 0, p(2) > p(1), and there is an integer t such that
p(t) >
p(t − 1) + p(t + 1)
2
(5)
and
p(s)
s + 1 ≥
p(y)
y + 1
,
for every 0 ≤ s ≤ y and y ∈ [t − 1, t, t + 1]. Then, determining whether
q(G, k, p) ≤ σ for a given graph G, integer k, and threshold σ is an NP-hard
problem.
We provide the proof in Appendix.
While the conditions in Proposition 4 seem overly complicated, they are
quite easy to satisfy. Assume that we are given a penalty function that is
concave in [−1,∞], and p(−1) = 0. Then due to concavity we have
p(x)
x + 1 ≥
p(x + 1)
x + 2
,
for x ≥ 0
.
This leads to the following corollary.
Tiers for peers
17
Corollary 1 Assume a monotonic penalty function p : Z → R such that
p(x) = 0 for x < 0, p(2) > p(1), and p is concave and non-linear in [−1, ℓ]
for some ℓ ≥ 1. Then, determining whether q(G, k, p) ≤ σ for a given graph
G, integer k, and threshold σ is NP-hard problem.
Note that we require p to be non-linear. This is needed so that the proper
inequality in Equation 5 is satisfied. This condition is needed since pl satisfies
every other requirement. Corollary 1 covers many penalty functions such as
p(x) = √x + 1 or p(x) = log(x + 2), for x ≥ 0. Note that the function needs
to be convex only in [−1, ℓ] for some ℓ ≥ 1. At extreme, ℓ = 1 in which case
t = 0 satisfies the conditions in Proposition 4.
6 Selecting canonical solution
A rank assignment minimizing agony may not be unique. In fact, consider
a graph G with no edges, then any ranking will have the optimal score of
0. Moreover, if the input graph G is a DAG, then any topological sorting of
vertices will yield the optimal score of 0.
In this section we introduce a technique to select a unique optimal solution.
The idea here is to make the ranks as small as possible without compromising
the optimality of the solution. More specifically, let us define the following
relationship between to rankings.
Definition 2 Given two rank assignments r and r′, we write r (cid:22) r′ if r(x) ≤
r′(x) for every x.
The following proposition states that there exists exactly one ranking with
the optimal score that is minimal with respect to the (cid:22) relation. We will refer
to this ranking as canonical ranking.
Proposition 5 Given a graph G and an integer k, there exists a unique op-
timal rank assignment r such that r (cid:22) r′ for every optimal rank assignment
r′.
The proof of this proposition is given in Appendix.
Canonical ranking has many nice properties. The canonical solution for a
graph without edges assigns rank 0 to all vertices. More generally, if G = (V, E)
is a DAG, then the source vertices S of G will receive a rank of 0, the source
vertices of G(V \ S) will receive a rank of 1, and so on. For general graphs we
have the following proposition.
Proposition 6 Let r be the canonical ranking. Then r has the least distinct
rank values among all optimal solutions.
In other words, the partition of V corresponding to the canonical ranking
has the smallest number of groups.
Our next step is to provide an algorithm for discovering canonical ranking.
In order to do so we assume that we use Orlin's algorithm and obtain the flow
18
Nikolaj Tatti
f and the dual π, described in Problem 4 and 5. We construct the residual
graph R, as described in Section 4, edges weighted by t(e) + π(w) − π(v). We
then compute, d(v) which is the shortest path in R from α to v. Finally, we
set r∗(v) = r(v) − d(v).
Once, we have computed the residual graph, we simply compute the short-
est path distance from q and subtract the distance from the optimal ranking,
see Algorithm 3.
Algorithm 3: canon(G), computes canonical optimal solution
1 f, π ← optimal flow and dual of Circulation;
2 R ← residual graph;
3 d(v) ← shortest weighted distance from α to v in R;
4 foreach v ∈ V do r∗(v) ← r(v) − d(v);
5 return r∗;
Proposition 7 Algorithm canon returns canonical solution with optimal score.
We give the proof of this proposition in Appendix.
Proposition states that to compute the canonical ranking it is enough to
form the residual graph, compute the shortest edge distances d(v) from the
vertex q, and subtract them from the input ranking. The computational com-
plexity of these steps is O(m + n log n). Moveover, this proposition holds for
a more general convex penalty function, described in Section 5.1.
7 A fast divide-and-conquer heuristic
In this section we propose a simple and fast divide-and-conquer approach. The
main idea is as follows: We begin with the full set of vertices and we split them
into two halves: the left half will have smaller ranks than the right half. We then
continue splitting the smaller sets recursively, and obtain a tree. We show that
this can be done in O(m log n) time. If we are given a cardinality constraint k,
then we prune the tree using dynamic program that runs in O(cid:0)k2n(cid:1) time. We
also propose a variant, where we perform SCC decomposition, and perform
then divide-and-conquer on individual components. To enforce the cardinality
constraint in this case, we need additional O(cid:0)km log n + k2n(cid:1) time.
7.1 Constructing a tree by splitting vertices
As mentioned above, our goal is to construct a tree T . This tree is binary and
ordered, that is, each non-leaf vertex has a left child and a right child.
Each leaf α5 in this tree T is associated with a set of vertices that we
denote by Vα. Every vertex of the input graph should belong to some leaf, and
5 we will systematically denote the vertices in T with Greek letters
Tiers for peers
19
no two leaves share a vertex. If α is a non-leaf, then we define Vα to be the
union of vertices related to each descendant leaf of α. We also define Eα to be
the edges in E that have both endpoints in Vα.
Since the tree is ordered, we can sort the leaves, left first. Using this order,
we define a rank r(v) to be the rank of the leaf in which v is included. We
define q(T ) = q(r).
Our goal is to construct T with good q(T ). We do this by splitting Vα of
a leaf α to two leaves such that the agony is minimized.
Luckily, we can find the optimal split efficiently. Let us first express the
gain in agony due to a split. In order to do so, assume a tree T , and let α
be a leaf. Let X be the vertices in leaves that are left to α, and let Z be the
vertices in leaves that are right to α.
We define b(α) to be the total weight of the edges from Z to X,
b(α) = X(z,x)∈E
x∈X,z∈Z
w(z, x)
.
Let y be a vertex in Vα. We define
ib(y; α) = X(z,y)∈E
z∈Z
w(z, y) and ob(y; α) = X(y,x)∈E
x∈X
w(y, x)
to be the total weight of the backward edges adjacent to y and Z or X. We
also define the total weights
Let
ib(α) = Xy∈Vα
ib(y; α)
flux (y; α) = X(x,y)∈Eα
and ob(α) = Xy∈Vα
w(x, y) − X(y,x)∈Eα
ob(y; α)
.
w(y, x)
to be the total weight of incoming edges minus the total weight of the outgoing
edges.
Finally, let us define
d (y; α) = flux (y; α) + ib(y; α) − ob(y; α)
.
We can now use these quantities to express how a split changes the score.
Proposition 8 Let α be a leaf of a tree T . Assume a new tree T ′, where we
have split α to two leaves. Let Y1 be the vertex set of the new left leaf, and Y2
the vertex set of the new right leaf. Then the score difference is
q(T ′) − q(T ) = b(α) + ib(α) − Xy∈Y2
d (y; α)
that can be rewritten as
q(T ′) − q(T ) = b(α) + ob(α) + Xy∈Y1
d (y; α)
.
20
Nikolaj Tatti
Proof We will show that
q(T ′) − q(T ) = b(α) + Xy∈Y1
ib(y; α) + Xy∈Y2
ob(y; α) + Xy∈Y1
flux (y; α)
.
(6)
Equation 6 can be then rewritten to the forms given in the proposition.
Let Y0 be the set of all vertices to the left of α. Let Y3 be the set of all
vertices to the right of α. Note that Y0 ∪ Y1 ∪ Y2 ∪ Y3 = V . Write t(i, j) to
be the total weight of edges from Yi to Yj. Also, write c(i, j) to be the total
change in the penalty of edges from Yi to Yj due to a split.
Note that c(0, 1) = c(0, 2) = c(0, 3) = c(1, 3) = c(2, 3) = 0 since these are
forward edges that remain forward. Also, c(0, 0) = c(1, 1) = c(2, 2) = c(3, 3) =
0 since the rank difference of these edges has not changed. For the same reason,
c(3, 2) = c(1, 0) = 0.
Case (i): Since a split shifts Y3 by one rank, c(3, 0) = t(3, 0) and c(3, 1) =
t(3, 1). Case (ii): Since a split shifts Y2 by one rank, c(2, 0) = t(2, 0). Case
(iii): The penalty of an edge from Y2 to Y1 increases by w(e). Summing over
these edges leads to c(2, 1) = t(2, 1). Case (iv ): The penalty of an edge from Y1
to Y2 decreases by w(e). Summing over these edges leads to c(1, 2) = −t(1, 2).
q(T ′) − q(T ) =Xi,j
c(i, j) = t(3, 0) + t(3, 1) + t(2, 0) + t(2, 1) − t(1, 2)
This leads to
.
First, note that
t(3, 0) = b(α) ,
t(3, 1) = Xy∈Y1
ib(y; α) ,
t(2, 0) = Xy∈Y2
ob(y; α)
.
To express t(2, 1) − t(1, 2), we can write
Xy∈Y1
w(x, y) − X(y,x)∈Eα
flux (y; α) = X(x,y)∈Eα
y∈Y1
y∈Y1
w(y, x)
= t(1, 1) + t(2, 1) − t(1, 1) − t(1, 2) = t(2, 1) − t(1, 2)
.
This proves Eq. 6, and the proposition. ⊓⊔
Proposition 8 gives us a very simple algorithm for finding an optimal split:
A vertex y for which d (y) ≥ 0 should be in the right child, while the rest ver-
tices should be in the left child. If the gain is negative, then we have improved
the score by splitting. However, it is possible to have positive gain, in which
case we should not do a split at all. Note that the gain does not change if we
do a split in a different leaf. This allows to treat each leaf independently, and
not care about the order in which leaves are tested.
The difficulty with this approach is that if we simply recompute the quan-
tities every time from the scratch, we cannot guarantee a fast computation
time. This is because if there are many uneven splits, we will enumerate over
some edges too many times. In order to make the algorithm provably fast, we
Tiers for peers
21
argue that we can detect which of the new leaves has fewer adjacent edges,
and we only enumerate over these edges.
Let us describe the algorithm in more details. We start with the full graph,
but as we split the vertices among leaves, we only keep the edges that are intra-
leaf; we delete any cross-edges between different leaves. As we delete edges, we
also maintain 4 counters for each vertex, flux (y; α), ib(y; α), ob(y; α), and the
unweighted degree, deg(y), where α is the leaf containing y.
For each leaf α, we maintain four sets of vertices,
Nα = {y ∈ Vα deg(y; α) > 0, d (y; α) < 0} ,
Pα = {y ∈ Vα deg(y; α) > 0, d (y; α) ≥ 0} ,
N ∗
α = {y ∈ Vα deg(y; α) = 0, d (y; α) < 0} ,
P ∗
α = {y ∈ Vα deg(y; α) = 0, d (y; α) ≥ 0}
.
The reason why we treat vertices with zero degree differently is so that we can
bound Nα or Pα by the number of adjacent edges.
Note that we maintain these sets only for leaves. To save computational
time, when a leaf is split, its sets are reused by the new leaves, and in the
process are modified.
In addition, we maintain the following counters
1. the total weights b(α), ib(α), ob(α), and
2. in order to avoid enumerating over N ∗
α and P ∗
α when computing the gain,
we also maintain the counters
dbN (α) = Xy∈Nα
ib(y) − ob(y) ,
dbP (α) = Xy∈Pα
ib(y) − ob(y)
.
We also maintain gain(α) for non-leaves, which is the agony gain of splitting
α. We will use this quantity when we prune the tree to enforce the cardinality
constraint.
If we decide to split, then we can do this trivially: according to Proposition 8
Nα and N ∗
α should be in the right
child. Our task is to compute the gain, and see whether we should split the
leaf, and compute the structures for the new leaves.
α should be in the left child while Pα and P ∗
Given a leaf α, our first step is to determine whether Nα or Pα has fewer
edges. More formally, we define adj (X) to be the edges that have at least one
end point in X. We then need to compute whether adj (Nα) ≤ adj (Pα). This
is done by cleverly enumerating over elements of Nα and Pα simultaneously.
The pseudo-code is given in Algorithm 4.
Proposition 9 Let m1 = adj (Nα) and m2 = adj (Pα). Then LeftS-
maller returns true if and only if m1 ≤ m2 in O(min(m1, m2)) time.
Proof Assume that the algorithm returns true, so Y1 = ∅ and c1 ≤ c2. Since
Y1 = ∅, then c1 = m1, which leads to m1 = c1 ≤ c2 ≤ m2. Assume that the
algorithm returns false. Then the while loop condition guarantees that Y2 = ∅
and c1 ≥ c2. Since Y2 = ∅, then c2 = m2. Either Y1 6= ∅ or c1 > c2. If latter,
22
Nikolaj Tatti
Algorithm 4: LeftSmaller(α), tests whether adj (Nα) ≤ adj (Pα).
1 Y1 ← Nα, Y2 ← Pα;
2 c1 ← 0; c2 ← 0;
3 until (Y1 = ∅ and c1 ≤ c2) or (Y2 = ∅ and c1 ≥ c2) do
4
5
6
7
8
9
if c1 ≤ c2 then
y ← vertex in Y1; delete y from Y1;
c1 ← c1 + deg(y);
else
y ← vertex in Y2; delete y from Y2;
c2 ← c2 + deg(y);
10 return Y1 = ∅ and c1 ≤ c2;
then m2 = c2 < c1 ≤ m1. If former, then m2 = c2 ≤ c1 < m1. This proves the
correctness.
To prove the running time, first note, since there are no singletons, each
iterations will increase either c1 or c2. Assume that m1 ≤ m2. If we have not
terminated after 2m1 iterations, then we must have m1 < c2. Since c1 ≤ m1 <
c2, we will then only increase c1. This requires at most m1 iterations (actually,
we can show that we only need 1 more iteration). In conclusion, the algorithm
runs in O(m1) time. The case for m1 ≥ m2 is similar. ⊓⊔
We can now describe our main algorithm, given in Algorithms 5, 6, and 7.
Split is given a leaf α. As a first step, Split determines which side has
fewer edges using LeftSmaller. After that it computes the gain, and checks
whether a split is profitable. If it is, then it calls either ConstructLeft
or ConstructRight, depending which one is faster. These two algorithms
perform the actual split and updating the structures, and then recurse on the
new leaves.
Algorithm 5: Split(α), checks if we can improve by splitting α,
and decides which side is more economical to split. Calls either Con-
structLeft or ConstructRight to update the structures.
1 if LeftSmaller(α) then
2
3
g ← b(α) + ob(α) + dbN (α) + Py∈Nα
if g < 0 then ConstructLeft(α); gain(α) ← g ;
d(y; α);
4 else
5
6
g ← b(α) + ib(α) − dbP (α) − Py∈Pα
if g < 0 then ConstructRight(α); gain(α) ← g ;
d(y; α);
Let us next establish the correctness of the algorithm. We only need to
show that during the split the necessary structures are maintained properly.
We only show it for ConstructLeft, as the argument is exactly the same
for ConstructRight.
Proposition 10 ConstructLeft maintains the counters and the vertex sets.
Tiers for peers
23
Algorithm 6: ConstructLeft(α), performs a single split using Nα.
Recurses to Split for further splits.
1 create a new leaf β with sets Nβ = Nα, Pβ = ∅, N ∗
2 b(β) ← b(α) + ob(α);
3 create a new leaf γ with sets Nγ = ∅, Pγ = Pγ , N ∗
4 b(γ) ← b(α);
5 foreach x ∈ Nα do
γ = ∅, and P ∗
α, and P ∗
γ = P ∗
γ ;
α = P ∗
α = ∅;
6
7
8
b(γ) ← b(γ) + ib(x);
b(β) ← b(β) − ob(x);
delete edges (x, z) or (z, x) for any z ∈ Pα, and update flux , deg , ib, ob ;
β , dbN (β) and dbP (β);
γ , dbN (γ) and dbP (γ);
9 check the affected vertices and update Pβ, Nβ, P ∗
10 check the affected vertices and update Pγ, Nγ , P ∗
11 Split(β); Split(γ);
β , N ∗
γ , N ∗
Algorithm 7: ConstructRight(α), performs a single split using Pα.
Recurses to Split for further splits.
1 create a new leaf β with sets Nβ = Nα, Pβ = ∅, N ∗
2 b(β) ← b(α);
3 create a new leaf γ with sets Nγ = ∅, Pγ = Pγ , N ∗
4 b(γ) ← b(α) + ib(α);
5 foreach x ∈ Pα do
γ = ∅, and P ∗
α, and P ∗
γ = P ∗
γ ;
α = P ∗
α = ∅;
6
7
8
b(γ) ← b(γ) − ib(x);
b(β) ← b(β) + ob(x);
delete edges (x, z) or (z, x) for any z ∈ Pα, and update flux , deg , ib, ob ;
9 check the affected vertices and update Pβ, Nβ, P ∗
10 check the affected vertices and update Pγ, Nγ , P ∗
11 Split(β); Split(γ);
β , N ∗
γ , N ∗
β , dbN (β) and dbP (β);
γ , dbN (γ) and dbP (γ);
Proof During a split, our main task is to remove the cross edges between Nα
and Pα and make sure that all the counters and the vertex sets in the new
leaves are correct.
Let y ∈ Vβ. If there is no cross edge attached to a vertex y in Eα, then
d(y; β) = d(y; α) and deg(y; β) = deg(y; α). This means that we only need to
check vertices that are adjacent to a cross edge, and possibly move them to
a different set, depending on deg(y; β) and d(y; β). This is exactly what the
algorithm does. The case for y ∈ Vγ is similar.
of b(α) as well as additional edges to Nα, namely b(γ) = b(α) +Py∈Nα
which is exactly what algorithm computes. Also, b(β) = b(α)+Py∈Pα
b(α) + ob(α) −Py∈Nα
as we delete edges or move vertices from one set to another. ⊓⊔
The only non-trivial counters are b(γ) and b(β). Note that b(γ) consists
ib(y),
ob(y) =
ob(y). The remaining counters are trivial to maintain
We conclude this section with the computational complexity analysis.
Proposition 11 Constructing the tree can be done in O(m log n) time, where
m is the number of edges and n is the number of vertices in the input graph.
To prove the proposition, we need the following lemmas.
24
Nikolaj Tatti
β = N ∗
γ = P ∗
Lemma 3 Let m = adj (Nα). Updating the new leaves in ConstructLeft
can be done in O(m) time.
Proof The assignments Nβ = Nα, N ∗
α are done by
reference, so they can be done in constant time. Since there are no singletons
in Nα, there are most 2m vertices in Nα. Deleting an edge is done in con-
stant time, so the for-loop requires O(m) time. There are at most 2m affected
vertices, thus updating the sets also can be done in O(m) time. ⊓⊔
Lemma 4 Let m = adj (Pα). Updating the new leaves in ConstructRight
can be done in O(m) time.
α, Pγ = Pα, P ∗
The proof for the lemma is the same as the proof for Lemma 3.
Proof Let us write mα = min(adj (Nα),adj (Pα)) to be the smaller of the
two adjacent edges.
Lemmas 3 -- 4 implies that the running time is O(Pα mα), where α runs
over every vertex in the final tree.
We can express the sum differently: given an edge e, write
1 e ∈ adj (Nα) ,
1 e ∈ adj (Pα) ,
.
0 otherwise
and adj (Nα) ≤ adj (Pα),
and adj (Nα) > adj (Pα),
ieα =
That is, mα =Pe ieα. Write ie =Pα ieα. To prove the proposition, we show
that ie ∈ O(log n).
Fix e, and let α and β be two vertices in a tree for which ieα = ieβ =
1. Either α is a descendant of β, or β is a descendant of α. Assume the
latter, without the loss of generality. We will show that 2Eβ ≤ Eα, and this
immediately proves that ie ∈ O(log n).
To prove this, let us define cα to be the number of cross edges between Nα
and Pα, when splitting α. Assume, for simplicity, that β is the left descendant
of α. Then Eβ ≤ adj (Nα) − cα. Also, e ∈ adj (Nα), and by definition of ieα,
mα = adj (Nα). This gives us,
2Eβ ≤ 2mα− 2cα ≤ (adj (Nα)− cα) + (adj (Pα)− cα) = Eα− cα ≤ Eα
The case when β is the right descendant is similar, proving the result. ⊓⊔
.
7.2 Enforcing the cardinality constraint by pruning the tree
If we did not specify the cardinality constraint, then once we have obtained
the tree, we can now assign individual ranks to the leaves, and consequently
to the vertices. If k is specified, then we may violate the cardinality constraint
by having too many leaves.
In such case, we need to reduce the number of leaves, which we do by
pruning some branches. Luckily, we can do this optimally by using dynamic
Tiers for peers
25
programming. To see this, let T ′ be a subtree of T obtained by merging some
of the branches, making them into leaves. Then Proposition 8 implies that
q(T ′) is equal to
q(T ′) = W +
Xα is a non-leaf in T ′
gain(α) ,
where W is the total weight of edges.
This allows us to define the following dynamic program. Let opt (α; h) be
the optimal gain achieved in branch starting from α using only h ranks. If α
is the root of T , then opt (α; k) is the optimal agony that can be obtained by
pruning T to have only k leaves.
To compute opt (α; h), we first set opt (α; 1) = 0 for any α, and opt (α; h) = 0
if α is a leaf in T . If α is a non-leaf and k > 1, then we need to distribute the
budget among the two children, that is, we compute
opt (α; h) = gain(α) + min
1≤ℓ≤h−1
opt (β; ℓ) + opt (γ; h − ℓ)
.
We also record the optimal index ℓ, that allows us to recover the optimal tree.
Computing a single opt (α; h) requires O(k) time, and we need to compute at
most O(nk) entries, leading to O(cid:0)nk2(cid:1) running time.
7.3 Strongly connected component decomposition
If the input graph has no cycles and there is no cardinality constraint, then the
optimal agony is 0. However, the heuristic is not guaranteed to produce such a
ranking. To guarantee this, we add an additional -- and optional -- step. First,
we perform the SCC decomposition. Secondly, we pack strongly connected
components in the minimal number of layers: source components are in the
first layer, second layer consists of components having edges edge only from
the first layer, and so on. We then run the heuristic on each individual layer.
If k is not set, we can now create a global ranking, where the ranks of the
ith layer are larger than the ranks of the (i− 1)th layer. In other words, edges
between the SCCs are all forward.
If k is set, then we need to decide how many individual ranks each com-
ponent should receive. Moreover, we may have more than k layers, so some of
the layers must be merged. In such a case, we will demand that the merged
layers must use exactly 1 rank, together. The reason for this restriction is
that it allows us to compute the optimal distribution quickly using dynamic
programming.
The gain in agony comes from two different sources. The first source is the
improvement of edges within a single layer. Let us adopt the notation from the
previous section, and write opt (i; h) to be the optimal gain for ith layer using
h ranks. We can compute this using the dynamic program in the previous
section. The second source of gain is making the inter-layer edges forward.
Instead of computing the total weight of such edges, we compute how many
26
Nikolaj Tatti
edges are not made forward. These are exactly the edges that are between
the layers that have been merged together. In order to express this we write
w(j, i) to be the total weight of inter-layer edges having both end points in
layers j, . . . , i.
To express, the agony of the tree, let ki be the budget of individual layers.
We also, write [aj, bj] to mean that layers aj, . . . , bj have been merged, and
must share a single rank. We can show that the score of the tree T ′ that uses
this budget distribution of is then equal to
q(T ′) = W +Xi
opt (i, ki) +Xj
w(aj , bj),
where W is the total weight of the intra-layer edges. Note that W is a constant
and so we can ignore it.
To find the optimal ki and [aj, bj], we use the following dynamic program.
Let us write o(i; h) to be the gain of 1, . . . , i layers using h ranks. We can
express o(i; h) as
o(i; h) = min(min
j
w(j, i) + o(j − 1; h − 1) , min
ℓ
opt (i, ℓ) + o(i − 1, h − ℓ))
.
The first part represents merging j, . . . , i layers, while the second part repre-
sents spending ℓ ranks on the ith layer. By recording the optimal j and ℓ we
can recover the optimal budget distribution for each i and h.
running time, which is too expensive.
Computing the second part can be done in O(k) time, and computing
the first part can be done in O(n) time, naively. This leads to O(cid:0)n2k + nk2(cid:1)
Luckily we can speed-up the computation of the first term. To simplify
notation, fix h, and let us write f (j, i) = w(j, i) + o(j − 1; h − 1). We wish
to find j(i) such that f (j(i), i) is minimal for each i. Luckily, f satisfies the
condition,
f (j1, i1) − f (j2, i1) ≤ f (j1, i2) − f (j2, i2),
where j1 ≤ j2 ≤ i1 ≤ i2. Aggarwal et al. [1] now guarantees that j(i1) ≤ j(i2),
for i1 ≤ i2. Moreover, Aggarwal et al. [1] provides an algorithm that computes
j(i) in O(n) time. Unfortunately, we cannot use it since it assumes that f (j, i)
can be computed in constant time, which is not the case due to w(j, i).
Fortunately, we can still use the monotonicity of j(·) to speed-up the al-
gorithm. We do this by computing j(i) in an interleaved manner. In order to
do so, let ℓ be the number of layers, and let t be the largest integer such that
s = 2t ≤ ℓ. We first compute j(s). We then proceed to compute j(s/2) and
j(3s/2), and so on. We use the previously computed values of j(·) as sentinels:
when computing j(s/2) we do not test j > j(s) or when computing j(3s/2)
we do not test j < j(s). The pseudo-code is given in Algorithm 8.
To analyze the complexity, note that for a fixed s, the variables i, a and b
are only moving to the right. This allows us to compute w(j, i) incrementally:
whenever we increase i, we add the weights of new edges to the total weight,
whenever we increase j, we delete the weights of expiring edges from the
Tiers for peers
27
Algorithm 8: Fast algorithm for computing j(i) minimizing f (j(i), i)
1 ℓ ← largest possible i;
2 s ← max (cid:8)2t ≤ ℓ, t ∈ N(cid:9);
3 while s ≥ 1 do
4
5
6
7
8
9
foreach i = s, 3s, 5s, . . ., i ≤ ℓ do
a ← 1; b ← i;
if i − s ≥ 1 then a ← j(i − s);
if i + s ≤ ℓ then b ← min(b, j(i + s));
j(i) ← mina≤j≤b f (j, i);
s ← s/2;
total weight. Each edge is visited twice, and this gives us O(m) time for a
fixed s. Since s is halved during each outer iteration, there can be at most
O(log n) iterations. We need to do this for each h, so the total running time
is O(cid:0)km log n + nk2(cid:1).
As our final remark, we should point out that using this decomposition may
not necessarily result in a better ranking. If k is not specified, then the optimal
solution will have inter-layer edges as forward, so we expect this decomposition
to improve the quality. However, if k is small, we may have a better solution
if we allow to inter-layer edges go backward. At extreme, k = 2, we are guar-
anteed that the heuristic without the SCC decomposition will give an optimal
solution, so the SCC decomposition can only harm the solution. We will see
this behaviour in the experimental section. Luckily, since both algorithms are
fast, we can simply run both approaches and select the better one.
8 Related work
The problem of discovering the rank of an object based on its dominating
relationships to other objects is a classic problem. Perhaps the most known
ranking method is Elo rating devised by Elo [4], used to rank chess players.
In similar fashion, Jameson et al. [9] introduced a statistical model, where the
likelihood of the the vertex dominating other is based on the difference of their
ranks, to animal dominance data.
Maiya and Berger-Wolf [12] suggested an approach for discovering hierar-
chies, directed trees from weighted graphs such that parent vertices tend to
dominate the children. To score such a hierarchy the authors propose a statis-
tical model where the probability of an edge is high between a parent and a
child. To find a good hierarchy the authors employ a greedy heuristic.
The technical relationship between our approach and the previous studies
on agony by Gupte et al. [7] and Tatti [19] is a very natural one. The authors
of both papers demonstrate that minimizing agony in a unweighted graph is a
dual problem to finding a maximal eulerian subgraph, a subgraph in which, for
each vertex, the same number of outdoing edges and the number of incoming
edges is the same. Discovering the maximum eulerian subgraph is a special
28
Nikolaj Tatti
case of the capacitated circulation problem, where the capacities are set to
1. However, the algorithms in [7, 19] are specifically designed to work with
unweighted edges. Consequently, if our input graph edges or we wish to enforce
the cardinality constraint, we need to solve the problem using the capacitated
circulation solver.
The stark difference of computational complexities for different edge penal-
ties is intriguing: while we can compute agony and any other convex score in
polynomial-time, minimizing the concave penalties is NP-hard. Minimizing
the score q(G, k, pc) is equivalent to feedback arc set (FAS), which is
known to be APX-hard with a coefficient of c = 1.3606 [2]. Moreover, there
is no known constant-ratio approximation algorithm for FAS, and the best
known approximation algorithm has ratio O(log n log log n) [5]. In this paper
we have shown that minimizing concave penalty is NP-hard. An interesting
theoretical question is whether this optimization problem is also APX-hard,
and is it possible to develop an approximation algorithm.
Role mining, where vertices are assigned different roles based on their ad-
jacent edges, and other features, has received some attention. Henderson et al.
[8] studied assigning roles to vertices based on its features while McCallum
et al. [13] assigned topic distributions to individual vertices. A potential direc-
tion for a future work is to study whether the rank obtained from minimizing
agony can be applied as a feature in role discovery.
9 Experiments
In this section we present our experimental evaluation. Our main focus of the
experiments is practical computability of the weighted agony.
9.1 Datasets and setup
For our experiments we took 10 large networks from SNAP repository [10]. In
addition, for illustrative purposes, we used two small datasets: Nfl, consisting
of National Football League teams. We created an edge (x, y) if team x has
scored more points against team y during 2014 regular season, we assign the
weight to be the difference between the points. Since not every team plays
against every team, the graph is not a tournament. Reef, a food web of guilds
of species [17], available at [18]. The dataset consisted of 3 food webs of coral
reef systems: The Cayman Islands, Jamaica, and Cuba. An edge (x, y) appears
if a guild x is known to prey on a guild y. Since the guilds are common among
all 3 graphs, we combined the food webs into one graph, and weighted the
edges accordingly, that is, each edge received a weight between 1 and 3.
The sizes of the graphs, along with the sizes of the largest strongly con-
nected component, are given in the first 4 columns of Table 2.
The 3 Higgs and Nfl graphs had weighted edges, and for the remaining
graphs we assigned a weight of 1 for each edge. We removed any self-loops as
they have no effect on the ranking, as well as any singleton vertices.
Tiers for peers
29
For each dataset we computed the agony using Algorithm 2. We compared
the algorithm to the baseline given by Tatti [20]. For the unweighted graphs we
also computed the agony using Relief, an algorithm suggested by Tatti [19].
Note that this algorithm, nor the algorithm by Gupte et al. [7], does not work
for weighted graphs nor when the cardinality constraint k given in Problem 1
is enforced. We implemented algorithms in C++ and performed experiments
using a Linux-desktop equipped with a Opteron 2220 SE processor.6
Table 2 Basic characteristics of the datasets and the experiments. The 6th is the number
of groups in the optimal ranking.
Name
Amazon
Gnutella
EmailEU
Epinions
Slashdot
WebGoogle
WikiVote
largest SCC
time
baseline
V
E
V ′
E′
k
SCC
plain
[20]
[19]
62 586
265 214
75 879
82 168
403 394 3 387 388 395 234 3 301 092 17
50 916 24
151 132
9
443 506 10
9
841 201
25m6s 6h24m 4h27m
20s
45s
29s
2m
44s
20m
61s 1h38m 1h5m
875 713 5 105 039 434 818 3 419 124 31 10m31s 25m22s 8h50m 2h32m
7s
24m7s
4s
10s
33s
38s
147 892
418 956
508 837
870 161
7 115
103 689
1 300
39 456 12
14 149
34 203
32 223
71 307
8s
10m
49m
2s
6s
43s
32
Nfl
258
Reef
HiggsReply
37 145
HiggsRetweet 425 008
HiggsMention 302 523
205
4232
30 517
733 610
445 147
32
1
263
13 086
4 786
205
6
0 19
569 11
63 505 22
19 848 21
4ms
8ms
0.3s
12s
6s
5ms
100ms
5s
2m10s
1m34s
22ms
10ms
0.2s
10m
2m
--
--
--
--
--
9.2 Results
Let us begin by studying running times given in Table 2. We report the run-
ning times of our approach with and without the strongly connected compo-
nent decomposition as suggested by Proposition 2, and compare it against the
baselines, whenever possible. Note that we can use the decomposition only if
we do not enforce the cardinality constraint.
Our first observation is that the decomposition always helps to speed up
the algorithm. In fact, this speed-up may be dramatic, if the size of the strongly
connected component is significantly smaller than the size of the input graph,
for example, with HiggsRetweet. The running times are practical despite the
unappealing theoretical bound. This is due to several factors. First, note that
the theoretical bound of O(min(nk, m)m log n) given in Section 4 only holds
for unweighted graphs, and it is needed to bound the number of outer-loop
iterations. In practice, however, the number of these iterations is small, even
for weighted graphs. The other, and the main, reason is the pessimistic n in
the min(nk, m) factor; we spend nk inner-loop iterations only if the dual π(v)
6 The source code is available at http://users.ics.aalto.fi/ntatti/agony.zip
30
Nikolaj Tatti
of each vertex v increases by O(k), and between the increases the shortest path
from sources to v changes O(n) times. The latter change seems highly unlikely
in practice, leading to a faster computational time.
We see that our algorithm beats consistently both baselines. What is more
important: the running times remain practical, even if we do not use strongly
connected components. This allows us to limit the number of groups for large
graphs. This is a significant improvement over [20], where solving HiggsRetweet
without the SCC decomposition required 31 hours.
Our next step is to study the effect of the constraint k, the maximum
number of different rank values. We see in the 6th column in Table 2 that
despite having large number of vertices, that the optimal rank assignment has
low number of groups, typically around 10 -- 20 groups, even if the cardinality
constraint is not enforced.
Let us now consider agony as a function of k, which we have plotted in
Figure 5 for Gnutella and WikiVote graphs. We see that for these datasets
that agony remains relatively constant as we decrease k, and starts to increase
more prominently once we consider assignments with k ≤ 5.
×104
Gnutella
×104
WikiVote
4
)
k
,
G
(
q
2
0
2
5
)
k
,
G
(
q
3
2
1
0
10
constraint k
15
20
24
2
4
6
8
10
12
constraint k
Fig. 5 Agony as a function of the constraint k for Gnutella and WikiVote datasets.
Enforcing the constraint k has an impact on running time. As implied by
Proposition 3, low values of k should speed-up the computation. In Figure 6
we plotted the running time as a function of k, compared to the plain version
without the speed-up.
As we can see lower values of k are computationally easier to solve. This
is an opposite behavior of [20], where lowering k increased the computational
time. To explain this behaviour, note that when we decrease k we increase the
agony score, which is equivalent to the capacitated circulation. Both solvers
increase incrementally the flow until we have reached the solution. As we
lower k, we increase the amount of optimal circulation, and we need more
iterations to reach the optimal solution. The difference between the algorithm
is that for lower k updating the residual graph becomes significantly faster
than computing the tree from scratch. This largely overpowers the effect of
Tiers for peers
31
Gnutella
WikiVote
)
s
d
n
o
c
e
s
n
i
(
e
m
i
t
103
102
101
baseline [20]
speed-up
)
s
d
n
o
c
e
s
n
i
(
e
m
i
t
103
102
101
100
2
5
10
constraint k
15
20
24
2
4
6
8
10
12
constraint k
Fig. 6 Execution time as a function of the constraint k for Gnutella and WikiVote datasets.
Note that the y-axis is logarithmic.
Table 3 Scores, compared to the optimal, and running times of the heuristic. Here, SCC
is the heuristic with SCC decomposition, while plain is the plain version, opt is the optimal
agony.
Time (sec.)
Name
Amazon
Gnutella
EmailEU
Epinions
Slashdot
WebGoogle
WikiVote
Nfl
Reef
HiggsReply
HiggsRetweet
HiggsMention
q(SCC)
q(opt)
q(plain)
q(opt)
q(SCC)
q(plain)
q(opt)
SCC
plain
1.036
1.256
1.008
1.024
1.001
1.051
1.043
1.047
--
1.007
1.259
1.078
1.037
1.350
1.012
1.030
1.003
1.079
1.091
1.047
--
1.103
1.606
1.322
2 044 609
23 820
121 820
271 419
749 448
1 935 476
18 430
1172
0
5 499
19 264
24 165
2 046 344
25 603
122 362
273 016
750 760
1 985 831
19 276
1172
452
6 022
24 579
29 632
1 973 965
18 964
120 874
264 995
748 582
1 841 215
17 676
1119
0
5459
15 302
22 418
9.24
0.35
0.47
0.40
0.68
6.80
0.05
0.002
0.008
2.29
0.12
1.82
8.49
0.34
0.45
0.37
0.64
6.64
0.05
0.002
0.006
1.03
0.05
1.65
needing many more iterations to converge. However, there are exceptions: for
example, computing agony for WikiVote with k = 8 is slower than k = 9.
Let us now consider the performance of the heuristic algorithm. We report
the obtained scores and the running times in Table 3. We tested both vari-
ants: with and without SCC decomposition, and we do not enforce k. We first
observe that both variants are expectedly fast: processing the largest graphs,
Amazon and WebGoogle, required less than 10 seconds, while the exact version
needed 10 -- 25 minutes. The plain version is cosmetically faster. Heuristic also
produces competitive scores but the performance depends on the dataset: for
Gnutella and HiggsRetweet the SCC variant produced 25% increase to agony,
while for the remaining datasets the increase was lower than 8%. Note that,
Reef has agony of 0, that is, the network is a DAG but the plain variant was
not able to detect this. This highlights the benefit of doing the SCC decom-
position. In general, the SCC variant outperforms the plain variant when we
do not enforce the cardinality constraint.
32
Nikolaj Tatti
As we lower the cardinality constraint k, the plain variant starts to out-
perform the SCC variant, as shown in Figure 7. The reason for this is that the
SCC variant requires that a edge (u, v) between two SCCs is either forward or
r(u) = r(v). This restriction becomes too severe as we lower k and it becomes
more profitable to allow r(u) < r(v). At extreme k = 2, the plain version is
guaranteed to find the optimal solution, so the SCC variant can only harm the
solution.
Gnutella
WikiVote
)
l
a
m
i
t
p
o
(
q
/
)
c
i
t
s
i
r
u
e
h
(
q
1.6
1.4
1.2
1
plain
SCC
)
l
a
m
i
t
p
o
(
q
/
)
c
i
t
s
i
r
u
e
h
(
q
1.6
1.4
1.2
1
2
5
10
constraint k
15
20
24
2
4
6
8
10
12
constraint k
Fig. 7 Ratio of the agony given by the heuristic and the optimal agony as a function of the
constraint k for Gnutella and WikiVote datasets.
Table 4 Rank assignment discovered for Nfl dataset with k = 3 groups
Rank Teams
1.
2.
3.
den bal ne dal sea phi kc gb pit
stl nyg mia car no sd min cin buf det ind hou sf ari
wsh oak tb jax ten cle atl nyj chi
Let us look at the ranking that we obtained from Nfl dataset using k = 3
groups, given in Table 4. We see from the results that the obtained ranking is
very sensible. 7 of 8 teams in the top group consists of playoff teams of 2014
season, while the bottom group consists of teams that have a significant losing
record.
Finally, let us look at the rankings obtained Reef dataset. The graph is in
fact a DAG with 19 groups. To reduce the number of groups we rank the guilds
into k = 4 groups. The condensed results are given in Table 5. We see that
the top group consists of large fishes and sharks, the second group contains
mostly smaller fishes, a large portion of the third group are crustacea, while
the last group contains the bottom of the food chain, planktons and algae.
We should point out that this ranking is done purely on food web, and not
on type of species. For example, cleaner crustacea is obviously very different
than plankton. Yet cleaner crustacea only eats planktonic bacteria and micro-
Tiers for peers
33
Table 5 Ranked guilds of Reef dataset, with k = 4. For simplicity, we removed the duplicate
guilds in the same group, and grouped similar guilds (marked as italic, the number in
parentheses indicating the number of guilds).
Sharks (6), Amberjack, Barracuda, Bigeye, Coney grouper, Flounder, Frogfish, Grouper,
Grunt, Hind, Lizardfish, Mackerel, Margate, Palometa, Red hind, Red snapper, Remora,
Scorpionfish, Sheepshead, Snapper, Spotted eagle ray
Angelfish, Atlantic spadefish, Ballyhoo, Barracuda, Bass Batfish, Blenny Butterflyfish,
Caribbean Reef Octopus, Caribbean Reef Squid, Carnivorous fish II-V, Cornetfish, Cow-
fish, Damselfish, Filefish, Flamefish, Flounder, Goatfish, Grunt, Halfbeak, Hamlet, Hawk-
fish, Hawksbill turtle, Herring, Hogfish, Jack, Jacknife fish, Jawfish, Loggerhead sea
turtle, Margate, Moray, Needlefish Porcupinefish I-II, Porkfish, Pufferfish, Scorpionfish,
Seabream, Sergeant major, Sharptail eel, Slender Inshore Squid, Slippery dick, Snap-
per, Soldierfish, Spotted drum, Squirrelfish, Stomatopods II, Triggerfish, Trumpetfish,
Trunkfish, Wrasse, Yellowfin mojarra
Crustacea (31), Ahermatypic benthic corals, Ahermatypic gorgonians, Anchovy, An-
gelfish, Benthic carnivores II, Blenny, Carnivorous fish I, Common Octopus, Corallivorous
gastropods IV, Deep infaunal soft substrate suspension feeders, Diadema, Echinometra,
Goby, Green sea turtle, Herbivorous fish I-IV, Herbivorous gastropods I, Hermatypic
benthic carnivores I, Hermatypic corals, Hermatypic gorgonians, Herring, Infaunal hard
substrate suspension feeders, Lytechinus, Macroplanktonic carnivores II-IV, Macroplank-
tonic herbivores I, Molluscivores I, Omnivorous gastropod, Parrotfish, Pilotfish, Silver-
side, Stomatopods I, Tripneustes, Zooplanktivorous fish I-II,
Planktons (7), Algae (6), Sponges (2), Feeders (11), Benthic carnivores I, Carnivorous
ophiuroids, Cleaner crustacea I, Corallivorous polychaetes, Detritivorous gastropods I,
Echinoid carnivores I, Endolithic polychaetes, Epiphyte grazer I, Epiphytic autotrophs,
Eucidaris, Gorgonian carnivores I, Herbivorous gastropod carnivores I, Herbivorous gas-
tropods II-IV, Holothurian detritivores, Macroplanktonic carnivores I, Micro-detritivores,
Molluscivores II-III, Planktonic bacteria, Polychaete predators (gastropods), Seagrasses,
Sponge-anemone carnivores I, Spongivorous nudibranchs
detritivores while being eaten by many other guilds. Consequently, it is ranked
in the bottom group.
10 Concluding remarks
In this paper we studied the problem of discovering a hierarchy in a directed
graph that minimizes agony. We introduced several natural extensions: (i) we
demonstrated how to compute the agony for weighted edges, and (ii) how to
limit the number of groups in a hierarchy. Both extensions cannot be handled
with current algorithms, hence we provide a new technique by demonstrat-
ing that minimizing agony can be solved by solving a capacitated circulation
problem, a well-known graph problem with a polynomial solution.
We also introduced a fast divide-and-conquer heuristic that produces the
rankings with competitive scores.
We should point out that we can further generalize the setup by allowing
each edge to have its own individual penalty function. As long as the penalty
functions are convex, the construction done in Section 5.1 can still be used
to solve the optimization problem. Moreover, we can further generalize cardi-
34
Nikolaj Tatti
nality constraint by requiring that only a subset of vertices must have ranks
within some range. We can have multiple such constraints.
There are several interesting directions for future work. As pointed out
in Section 5.1 minimizing convex penalty increases the number of edges when
solving the corresponding circulation problem. However, these edges have very
specific structure, and we conjecture that it is possible to solve the convex case
without the additional computational burden.
References
1. A. Aggarwal, M. Klawe, S. Moran, P. Shor, and R. Wilber. Geometric
applications of a matrix-searching algorithm. Algorithmica, 2(1 -- 4):195 --
208, 1987.
2. I. Dinur and S. Safra. On the hardness of approximating vertex cover.
Annals of Mathematics, 162(1):439 -- 485, 2005.
3. J. Edmonds and R. M. Karp. Theoretical improvements in algorithmic
efficiency for network flow problems. Journal of ACM, 19(2):248 -- 264,
1972.
4. A. E. Elo. The rating of chessplayers, past and present. Arco Pub., 1978.
5. G. Even, J. (Seffi) Naor, B. Schieber, and M. Sudan. Approximating
minimum feedback sets and multicuts in directed graphs. Algorithmica,
20(2):151 -- 174, 1998.
6. M. Garey and D. Johnson. Computers and intractability: a guide to the
theory of NP-completeness. WH Freeman & Co., 1979.
7. M. Gupte, P. Shankar, J. Li, S. Muthukrishnan, and L. Iftode. Finding
hierarchy in directed online social networks. In Proceedings of the 20th
International Conference on World Wide Web, pages 557 -- 566, 2011.
8. K. Henderson, B. Gallagher, T. Eliassi-Rad, H. Tong, S. Basu, L. Akoglu,
D. Koutra, C. Faloutsos, and L. Li. Rolx: Structural role extraction &
mining in large graphs.
In Proceedings of the 18th ACM SIGKDD In-
ternational Conference on Knowledge Discovery and Data Mining, pages
1231 -- 1239, 2012.
9. K. A. Jameson, M. C. Appleby, and L. C. Freeman. Finding an appropriate
order for a hierarchy based on probabilistic dominance. Animal Behaviour,
57:991 -- 998, 1999.
10. J. Leskovec and A. Krevl. SNAP Datasets: Stanford large network dataset
collection. http://snap.stanford.edu/data, Jan. 2015.
11. L. Macchia, F. Bonchi, F. Gullo, and L. Chiarandini. Mining summaries
of propagations. In IEEE 13th International Conference on Data Mining,
pages 498 -- 507, 2013.
12. A. S. Maiya and T. Y. Berger-Wolf. Inferring the maximum likelihood
hierarchy in social networks. In Proceedings IEEE CSE'09, 12th IEEE In-
ternational Conference on Computational Science and Engineering, pages
245 -- 250, 2009.
Tiers for peers
35
13. A. McCallum, X. Wang, and A. Corrada-Emmanuel. Topic and role dis-
covery in social networks with experiments on enron and academic email.
J. Artif. Int. Res., 30(1):249 -- 272, 2007.
14. J. B. Orlin. A faster strongly polynomial minimum cost flow algorithm.
Operations Research, 41(2), 1993.
15. C. H. Papadimitriou and K. Steiglitz. Combinatorial Optimization: Algo-
rithms and Complexity. Prentice-Hall, Inc., 1982.
16. G. Ramalingam and T. Reps. On the computational complexity of dy-
namic graph problems. Theoretical Computer Science, 158:233 -- 277, 1996.
17. P. D. Roopnarine and R. Hertog. Detailed food web networks of three
Greater Antillean Coral Reef systems: The Cayman Islands, Cuba, and
Jamaica. Dataset Papers in Ecology, 2013, 2013.
18. H. R. Roopnarine PD.
Data from: Detailed food web net-
works of
systems: The Cay-
man Islands, Cuba, and Jamaica, 2012. Dryad Digital Repository,
http://dx.doi.org/10.5061/dryad.c213h.
three Greater Antillean Coral Reef
19. N. Tatti. Faster way to agony -- discovering hierarchies in directed graphs.
In Prooceding of European Conference of Machine Learning and Knowl-
edge Discovery in Databases, ECML PKDD 2014, pages 163 -- 178, 2014.
20. N. Tatti. Hierarchies in directed networks.
In Proceedings of the 15th
IEEE International Conference on Data Mining (ICDM 2015), 2015.
A Proof of Proposition 4
Proof To prove the completeness we will provide reduction from Maximum Cut [6]. An
instance of Maximum Cut consists of an undirected graph, and we are asked to partition
vertices into two sets such that the number of cross edges is larger or equal than the given
threshold σ.
Note that the conditions of the proposition guarantee that p(0) > 0.
Assume that we are given an instance of Maximum Cut, that is, an undirected graph
G = (V, E) and a threshold σ. Let m = E. Define a weighted directed graph H = (W, F, w)
as follows. Add V to W . For each edge (u, v) ∈ E, add a path with t intermediate vertices
from u to v, the length of the path is t + 2. Add also a path in reversed direction, from v
to u. Set edge weights to be 1. Add 4 special vertices α1, . . . , α4. Add edges (αi+1, αi), for
i = 1, . . . , 3 with a weight of
C = 2B
p(0)
p(2) − p(1)
, where B = 2(t + 1)m .
Add edges (αi, αi+1), for i = 1, . . . , 3 with a weight of
D = 4Cp(1) /p(0) + B .
Add edges (α0, v) and (v, α4), for each v ∈ V , with a weight of D.
Let r be the optimal ranking for H. We can safely assume that r(α1) = 0. We claim
that r(αi) = i − 1, and r(v) = 1, 2 for each v ∈ V . To see this, consider a ranking r′ such
that r′(αi) = i − 1 and the rank for the remaining vertices is 2. The score of this rank is
q(cid:0)H, r′(cid:1) = 3Cp(1) + 2(t + 1)mp(0) = 3Cp(1) + Bp(0)
.
Let (u, v) ∈ F with the weight of D. If r(u) ≥ r(v), then the score of r is at least Dp(0) =
4Cp(1) + Bp(0) which is more than q(H, r′). Hence, r(u) < r(v). Let (u, v) ∈ F with the
36
Nikolaj Tatti
weight of C. Note that r(u) ≥ r(v) + 1. Assume that r(u) ≥ r(v) + 2. Then the score is at
least
3Cp(1) + C(p(2) − p(1)) = 3Cp(1) + 2Bp(0) ,
which is a contradiction. This guarantees that r(αi) = i − 1, and r(v) = 1, 2 for each v ∈ V .
Consider (u, v) ∈ E and let u = x0, . . . , xt+1 = v be the corresponding path in H. Let
di = r(xi) − r(xi+1) and set ℓ = t + r(u) − r(v). Let P = P p(di) be the penalty contributed
by this path. Note that P ≤ p(ℓ), a penalty that we achieve by setting r(xi) = r(xi−1) + 1
for i = 1, . . . , t. This implies that di ≤ ℓ. The condition of the proposition now implies
P =
≥
≥
=
=
t
X
i=0
t
X
i=0
t
X
i=0
p(di) =
t
X
p(di)
i=0,di ≥0
max(di + 1, 0)
ℓ + 1
p(ℓ)
di + 1
ℓ + 1
p(ℓ)
t
X
i=0
1 + r(xi) − r(xi+1)
(t + 1 + r(u) − r(v)) = p(ℓ)
.
p(ℓ)
ℓ + 1
p(ℓ)
ℓ + 1
This guarantees that P = p(ℓ).
Partition edges E into two groups,
and
X = {(u, v) ∈ E ∈ r(u) = r(v)}
Y = {(u, v) ∈ E ∈ r(u) 6= r(v)}
.
Let ∆ = p(t − 1) + p(t + 1) − 2p(t). Note that concavity implies that ∆ < 0. Then
q(H, r) = 3C + X2p(t) + Y (p(t − 1) + p(t + 1))
= 3C + m2p(t) + Y (p(t − 1) + p(t + 1) − 2p(t))
= 3C + m2p(t) + Y ∆ .
The first two terms are constant. Consequently, q(H, r) is optimal if and only if Y , the
number of cross-edges is maximal.
Given a threshold σ, define σ′ = 3C + m2p(t) + ∆σ. Then q(H, r) ≤ σ′ if and only if
there is a cut of G with at least σ cross-edges, which completes the reduction. ⊓⊔
B Proof of Proposition 5 and 7
We will prove both Propositions 5 and 7 with the same proof.
Proof Let r∗ be the ranking returned by canon, and let π∗ = π − d be the corresponding
dual. Lemma 1 states that π∗ satisfies the slackness conditions, so it remains a solution to
Problem 5. This implies also that r∗ is an optimal ranking.
To complete the proof we need to show that for any r′, we have r∗ (cid:22) r′. Note that this
also proves that r∗ is a unique ranking having such property.
Let r′ be any optimal ranking, and let π′ be the corresponding dual. We can assume
that π′(α) = π∗(α) = 0. To prove the result we need to show that π′(v) ≥ π∗(v). We will
prove this by induction over the shortest path tree T from α. This certainly holds for α.
Tiers for peers
37
Let u be a vertex and let v be its parent in T , and let e ∈ E(T ) be the connecting edge.
Note that, by definition, −t(e) = π∗(u)−π∗(v). By the induction assumption, π∗(v) ≤ π′(v).
If e is forward, then due to Eq. 3
π′(u) − π′(v) ≥ −t(e) = π∗(u) − π∗(v) ≥ π∗(u) − π′(v)
If e is backward, then f (v, u) > 0. and Eq. 4 implies
π′(u) − π′(v) = −t(e) = π∗(u) − π∗(v) ≥ π∗(u) − π′(v)
This completes the induction step, and the proves the proposition. ⊓⊔
.
.
|
1612.03343 | 1 | 1612 | 2016-12-10T21:04:55 | Oblivious Sorting and Queues | [
"cs.DS"
] | We present a deterministic oblivious LIFO (Stack), FIFO, double-ended and double-ended priority queue as well as an oblivious mergesort and quicksort algorithm. Our techniques and ideas include concatenating queues end-to-end, size balancing of multiple arrays, several multi-level partitionings of an array. Our queues are the first to enable executions of pop and push operations without any change of the data structure (controlled by a parameter). This enables interesting applications in computing on encrypted data such as hiding confidential expressions. Mergesort becomes practical using our LIFO queue, ie. it improves prior work (STOC '14) by a factor of (more than) 1000 in terms of comparisons for all practically relevant queue sizes. We are the first to present double-ended (priority) and LIFO queues as well as oblivious quicksort which is asymptotically optimal. Aside from theortical analysis, we also provide an empirical evaluation of all queues. | cs.DS | cs | Oblivious Sorting and Queues
Johannes Schneider
University of Liechtenstein, Vaduz, Liechtenstein
6
1
0
2
c
e
D
0
1
]
S
D
.
s
c
[
1
v
3
4
3
3
0
.
2
1
6
1
:
v
i
X
r
a
Abstract
We present a deterministic oblivious LIFO (Stack), FIFO, double-ended and double-ended priority queue as well as
an oblivious mergesort and quicksort algorithm. Our techniques and ideas include concatenating queues end-to-end,
size balancing of multiple arrays, several multi-level partitionings of an array. Our queues are the first to enable
executions of pop and push operations without any change of the data structure (controlled by a parameter). This
enables interesting applications in computing on encrypted data such as hiding confidential expressions. Mergesort
becomes practical using our LIFO queue, ie. it improves prior work (STOC '14) by a factor of (more than) 1000 in
terms of comparisons for all practically relevant queue sizes. We are the first to present double-ended (priority) and
LIFO queues as well as oblivious quicksort which is asymptotically optimal. Aside from theortical analysis, we also
provide an empirical evaluation of all queues.
Keywords: sorting, queues, complexity, oblivious algorithms, privacy preserving, computation on encrypted data,
secure computing, fully homomorphic encryption, secure multi-party computation
1. Introduction
Advances in computing on encrypted data such as Fully Homomorphic Encryption (FHE) and secure multi-party
computation (SMC) might make outsourcing computation securely practically feasible. Memory access must also be
secured. For example, accessing the i-th element of an array of length n needs O(1) operations on RAM machines.
But for a program running on encrypted data, the same access mechanism reveals access patterns. The knowledge
of seemingly simple access patterns can help to disclose sensitive information such as stock trading patterns [17] or
encryption keys [10]. A simple solution requires to access all array elements requiring O(n) instead of O(1) time.
Oblivious RAM (ORAM) secures memory access more efficiently using multiple parties. Often relying on more than
one party is not desirable. Current solutions for oblivious data structures also do not hide (high level) operations,
which makes them unsuitable for omnipresent 'if-then-else' statements with private conditions and queue access in
branches. Evaluating a confidential expression, keeping data as well as the expression itself secret, becomes straight
forward using our LIFO queue and known techniques for computing on encrypted data. Such a scenario is important
for cloud computing, ie. a cloud provider might host data for customers, which run their own analytics functionality.
The customers wish to keep their data and algorithms private – in case of industrial automation an algorithm often
means a mathematical expression on time-series sensor data.1 To summarize, the main contributions are:
1. We present oblivious LIFO, FIFO and double-ended (priority) queues. The amortized overhead of an operation
on the LIFO queue is O(log n) in the maximal length n of the queue. Prior LIFO queues (based on priority
queues [21]) required O(log2 n). For a wide range of applications such as the producer-consumer problem in
a streaming context our FIFO queue has only O(log n) overhead which improves prior work [21] by a factor
log n. We are the first to introduce double-ended queues. Our double-ended queue needs O(log2 n).
2. We are the first to derive oblivious data structures to support push and pop operations that might not alter the
5
10
15
20
stored elements (depending on a parameter).
Email address: [email protected] (Johannes Schneider)
1In fact, a request from industry motivated this feature.
Preprint submitted to Journal of Theoretical Computer Science
June 25, 2018
25
30
35
40
45
50
55
60
65
3. Our deterministic mergesort algorithm improves on [9] for all relevant list sizes, eg. by two orders of magnitude
for sorting of 10 billion elements.
4. We state the first oblivious quicksort algorithm.
It is asymptotically optimal. The Monte Carlo algorithm
succeeds with high probability, ie. 1 − 1/nc for an arbitrary constant c.
1.1. Overview of Technique
We structure the array representing the queue in subarrays (SA) of increasing size. A SA might be itself a queue.
SAs are organized into parts that are merged and split if they are shifted between different SAs. Moving of elements
between SAs can cause some of the push and pop operations to require linear run-time in the maximal queue length.
But the time is amortized across many operations so that the average overhead is only (poly)logarithmic. Moving of
parts between SAs happens based on the number of pops and pushes. It is not dependent on the data held in the queue.
We develop a deterministic calling pattern that does not require knowing the number of stored elements in a queue.
This allows to hide the number of operations together with another idea: We permit the pop and push of a special
(empty) element that does not alter the number of stored elements in the data structure. Put differently, this disguises
whether an operation on the data structure changed the stored elements or not. Furthermore, to ensure efficient access
to both ends of a queue, eg. as needed for FIFO and double ended queues, we concatenate two ends of a (LIFO)
queue.
1.2. Outline
We first discuss our model and some notation (Section 2). The main data structures are given in Section 3 (Stack)
with a detailed explanation of core ideas and analysis, Section 4 (FIFO) and Section 5 (double-ended queue). Detailed
case studies are given in Section 9 after explaining the technique thoroughly. This includes an explanation how
obliviousness (and operation hiding) helps in securing code. Performance evaluation can be found in Section 11.
2. Preliminaries and Limitations
We assume knowledge of an upper bound on the maximal number of elements n that can be kept in the data
structure, ie. a queue is represented by an array of fixed size n. This assumption is common for oblivious data
structures. Adjusting the size of the data structure exactly to the actual number of elements is impossible since our
goal is to conceal the number of elements contained in the queue. Our queues support two operations: Push (allowing
empty elements) and Pop (allowing conditional popping). For obliviousness we proved an analogous definition as
[6]. Essentially, obliviousness implies that memory access patterns are the same for any input.
Definition 1. A data structure is oblivious if the sequence of memory access only depends on the number of push and
pop operations. A sorting algorithm is oblivious if the sequence of memory accesses is the same regardless of the
input.
We use a special (empty) element "∅" also denoted by a dash '−' indicating that an element in the queue is
unused. Its bit representation must be different from any data item stored in the queue. We use variants of compare
and exchange operations. The simplest form takes as input a binary bit b and two variables A and B. It assigns A := B
if the bit b is 1, otherwise A is not changed, ie. it computes A := b · B + (1 − b) · A. The compare-exchange-and-
erase CmpExEr(b, A, B) performs a compare and exchange as described and, additionally, it might erase B, ie.
it
sets variable B to ∅ if b is 1 and leaves it unchanged otherwise (see PseudoCode CmpExEr in Algorithm 1). For
the analysis we distinguish between input sensitive operations involving parameters of the push and pop elements as
well as data of the queue and operations that do not directly depend on any input data (but potentially on the number
of operations). The motivation is that for secure computation these distinctions are meaningful, since the former
correspond to (slower) operations on encrypted data. For our algorithms input sensitive operations always dominate
the time complexity – even when using non-encrypted data. They are split into elementary operations (+,-,·), called
E-Ops, and comparisons C-Ops, which are composed of elementary operation. The distinction is motivated since
comparisons are used to measure performance of sorting algorithms. For encrypted operations, comparisons might
have different time complexities, eg. for SMC such as [19] it is not clear how to perform a comparison in less than
Ω(nb · E − Ops) time, where nb is the number of bits of a compared number.
2
3. LIFO (Stack)
70
75
For a Last-In-First-Out (LIFO) queue (also called Stack) a pop operation returns the most recently pushed ele-
ment onto the data structure. To ensure obliviousness we access the same array elements independent upon the data
contained in the queue. Our queue always accesses the first element. A newly pushed element is stored in the first
position of the array. This implies that upon every insertion, we must shift elements to the right to avoid overwriting
of a previously inserted element. It is easy to shift the entire queue to the right but this requires linear run-time. To
improve efficiency, we logically split the array representing the queue into subarrays (SAs) of exponentially growing
size. We only shift parts of size at most 2k of a SA after every 2k push or pop operations.
LIFO Queue A with 3 subarrays S(i): A = [S(0) S(1) S(2)]
Subarray S(i) with 4 parts P(i,j): S(i) = P(i,0) P(i,1) P(i,2) P(i,3)
Part P(i,j) with 2i elements: P(i,j) = E(i,j,0) E(i,j,1) E(i,j,2)...E(i,j,2i-1)
Legend: "" separates subarrays, "" parts and " " elements
Elements 1-12 pushed to (=>) LIFO Queue:
1 => [ 1 - - - - - - - - - - - - - - - - - - - - - - - - - - -]
2 => [ 2 1 - - - - - - - - - - - - - - - - - - - - - - - - - -]
3 => [ 3 2 1 - - - - - - - - - - - - - - - - - - - - - - - - -]
4 => [ 4 3 - - 2 1 - - - - - - - - - - - - - - - - - - - - - -]
5 => [ 5 4 3 - 2 1 - - - - - - - - - - - - - - - - - - - - - -]
6 => [ 6 5 - - 4 3 2 1 - - - - - - - - - - - - - - - - - - - -]
7 => [ 7 6 5 - 4 3 2 1 - - - - - - - - - - - - - - - - - - - -]
8 => [ 8 7 - - 6 5 4 3 2 1 - - - - - - - - - - - - - - - - - -]
9 => [ 9 8 7 - 6 5 4 3 2 1 - - - - - - - - - - - - - - - - - -]
10 => [10 9 - - 8 7 6 5 4 3 2 1 - - - - - - - - - - - - - - - -]
11 => [1110 9 - 8 7 6 5 4 3 2 1 - - - - - - - - - - - - - - - -]
12 => [1211 - -10 9 8 7 6 5 - - 4 3 2 1 - - - - - - - - - - - -]
Elements popped from (<=) LIFO Queue:
12 <= [ -11 - -10 9 8 7 6 5 - - 4 3 2 1 - - - - - - - - - - - -]
11 <= [10 9 - - - - 8 7 6 5 - - 4 3 2 1 - - - - - - - - - - - -]
10 <= [ - 9 - - 8 7 6 5 - - - - 4 3 2 1 - - - - - - - - - - - -]
9 <= [ 8 7 - - - - 6 5 - - - - 4 3 2 1 - - - - - - - - - - - -]
8 <= [ - 7 - - 6 5 - - - - - - 4 3 2 1 - - - - - - - - - - - -]
7 <= [ 6 5 - - - - - - - - - - 4 3 2 1 - - - - - - - - - - - -]
6 <= [ - 5 - - - - - - - - - - 4 3 2 1 - - - - - - - - - - - -]
5 <= [ 4 3 - - - - 2 1 - - - - - - - - - - - - - - - - - - - -]
4 <= [ - 3 - - 2 1 - - - - - - - - - - - - - - - - - - - - - -]
3 <= [ 2 1 - - - - - - - - - - - - - - - - - - - - - - - - - -]
2 <= [ - 1 - - - - - - - - - - - - - - - - - - - - - - - - - -]
1 <= [ - - - - - - - - - - - - - - - - - - - - - - - - - - - -]
Figure 1: Pushes and pops onto a LIFO queue
More formally, a queue is implemented as an array A that is split into s subarrays (SA) S i growing exponen-
tially in size with i. The total length n of the array is n := Ps−1
i=0 S i. Each SA S i itself is partitioned into q parts
Pi,0, Pi,1, ..., Pi,q−1 of equal size Pi, j = S i/q. The size of a part varies for different SAs. We denote the k-th element
in Pi, j by Ei, j,k. Figure 1 shows the structure of a queue.
80
3.1. Push, Pop and Shifting
We explain the shifting procedure shown in Figure 1 for a sequence of push operations. We always push an
element onto the first position in the array A (or pop an element from there). After every modification of the queue,
3
Steps for a Push of 6 onto LIFO Queue:
Original queue:
[ 5 4 3 - - - 2 1 - - - - - - - - - - - - - - - - - - - -]
after ShiftPartsRight(Sublist 0):
[ - 5 4 3 - - 2 1 - - - - - - - - - - - - - - - - - - - -]
after E(0,0,0)=6
[ 6 5 4 3 - - 2 1 - - - - - - - - - - - - - - - - - - - -]
final queue after executing emptyTwoParts(Sublist 0):
[ 6 5 - - 4 3 - - 2 1 - - - - - - - - - - - - - - - - - -]
Executing emptyTwoParts(Sublist 0):
Original queue:
[ 6 5 4 3 - - 2 1 - - - - - - - - - - - - - - - - - - - -]
after ShiftPartsRight(Sublist 1):
[ 6 5 4 3 - - - - 2 1 - - - - - - - - - - - - - - - - - -]
after shifting P(0,2) to P(1,0):
[ 6 5 - 3 4 - - - 2 1 - - - - - - - - - - - - - - - - - -]
final queue after shifting of parts P(0,3) to P(2,0):
[ 6 5 - - 4 3 - - 2 1 - - - - - - - - - - - - - - - - - -]
Figure 2: Steps for pushing an element onto a LIFO queue
Algorithm 1 LIFO
Initialization(Number of SAs s with s ≥ 1)
q := 4 {number of parts per SA}
Ei, j,k := ∅, ∀i ∈ [0, s − 1], j ∈ [0, q − 1], k ∈ [0, 2i − 1]
npu := npo := 0 {counter for pushes and pops}
CmpExEr(b,A,B)
A := b · B + (1 − b) · A {Exchange A, B based on b}
B := (1 − b) · B + b · ∅ {Delete B based on b}
ShiftPartsRight(SA i,doOp)
empty&DoOp := doOp if Ei,0,0 , ∅ else 0
for SA j := q − 1 to 1 do
doS hi f t := empty&DoOp if Ei, j,0 = ∅ else 0
for Element k = 0 to S i/q − 1 do
CmpExEr(doShift, Ei, j,k ,Ei, j−1,k)
EmptyTwoParts(SA i)
q−1
j=0 (Ei, j,0 = ∅) else 0
isFull := 1 if ∧
ShiftPartsRight(SA i+1,isFull)
for SA j := q − 2 to q − 1 do
o := ( j − q + 2) · S i/q {offset for last 2 parts}
for Element k := 0 to S i/q − 1 do
CmpExEr(isFull, Ei+1,0,k+o,Ei, j,k)
MoveBetweenSAs(nOps,Operation Op)
mi := 0 {Find maximal SA to empty/refill}
while (nOps + 1) mod 2mi+1 = 0 do
mi := mi + 1
for SA i := min(mi, s − 2) to 0 do
Apply Operation Op on SA i
return (nOps + 1) mod 2max(0,s−2)
ShiftPartsLeft(SA i)
f ull&DoOp := 1 if Ei,0,q−1 , ∅ else 0
for SA j := to q − 2 do
doS hi f t := f ull&DoOp if Ei, j,0 , ∅ else 0
for Element k := 0 to S i/q − 1 do
CmpExEr(doShift, Ei, j,k ,Ei, j+1,k)
RefillTwoParts(SA i)
q−1
isEmpty := 1 if ∧
j=0 (Ei, j,0 , ∅) else 0
ShiftPartsLeft(SA i+1)
for SA j := q − 2 to q − 1 do
o := ( j − q + 2) · S i/q {offset for last 2 parts}
for Element k := 0 to S i/q − 1 do
CmpExEr(doShift, Ei, j,k,Ei+1,0,k+o)
Push(Element x)
npu := MoveBetweenSAs(npu,EmptyTwoParts)
b := 1 if x , ∅ ∧ Ei,0,0 , ∅ else 0
E0,0,0 := x if x , ∅ else E0,0,0
Pop(doPop)
result := E0,0,0 if doPop else ∅
E0,0,0 := ∅ if doPop else E0,0,0
npo := MoveBetweenSAs(npo,refillTwoParts)
return result
4
85
90
95
100
105
110
115
120
125
130
we modify (some) SAs to ensure that there is space for further pushes in the first SA. We shift elements to the right.
Shifting is only done on a part level, ie. either we shift all elements of a part or none. We perform frequent shifts
to overwrite empty small parts near the beginning of the array and less frequent shifts are conducted for larger parts
situated towards the end of the array. We shift parts within a SA but also move parts between SAs, ie. either we merge
two parts into one or we split a part into two parts. The subroutines for a push shown in Algorithm 1 are discussed
next.
ShiftPartsRight and EmptyTwoParts: ShiftPartsRight shifts elements from one part to the next part (on the right)
within a SA. It avoids overwriting of filled parts by checking if the part to be overwritten is indeed empty. To this
end, we only check if the first position of a part is empty. No parts are moved, if the first part of the SA is empty.
A parameter indicates whether shifting should take place or not. This is necessary to enable executions of push
operations that do not modify the queue. If the parameter is false, ie. zero, then no elements are moved. The order
of shifting is from back to front, ie. elements of the second to last part are shifted to the last part (given it is empty),
then the third to last part is shifted to the second to last (if empty) and so on. EmptyTwoParts empties the last two
parts of a SA i by merging them to form the first part of SA i + 1. It first empties the first part in SA i + 1 by doing a
ShiftPartsRight. Emptying only takes place if all parts of SA i are full and SA i + 1 is not completely full. Without this
condition for q > 2 a full part would be (continuously) shifted towards the right for repeated insertions of the empty
element ∅. This would lead to empty SAs followed by (partially) full SAs. As a consequence for pop operations we
would have to undo the shifting (or search the entire array).
Push: A push operation first ensures that the first position of the array is empty. Then, it inserts the pushed element at
the first position. A push and its suboperations are illustrated in Figure 2.
MoveBetweenSAs: Restructuring is done after every operation starting from some initial SA (down) to the very first
SA in the beginning of the queue. The (index of the) initial SA depends on the number of operations and not the
number of actual elements in the queue, which we wish to disguise. Parts of a SA are moved to the next SA, once a
SA is full. It might seem reasonable to move all parts of a full SA to the next. However, for alternating pushes and
pops this might trigger large performance penalties since parts are continuously moved back and forth between SAs.
To disguise the number of elements in the queue (and thus parts), we access all parts in the same deterministic manner
for any sequence of pushes of fixed length. Since we allow pushes of a special (empty) element that has no impact on
the number of stored elements, the number of operations (as an indicator for the actual number of elements contains)
is not exact. We assume that the array grows at a maximal rate, ie. every push is done using a non-empty element.
Since we always empty two parts of a SA, we must create space in a SA by moving elements, whenever a sequence of
operations could have resulted in the filling of two parts of that SA. For example, every push potentially fills one part
in the first SA, since they are of size one. Thus, we would empty the first SA after every second push operation. For
SA i with parts of size 2i, we would move two parts to the next SA after every 2i+1 operations. But this approach fails
for an arbitrary interleaving of operations pops and pushes of empty and non-empty elements. For example, for the
following sequence of pushed elements 1, 2, 3, ∅, 4, 5, the algorithm would attempt to empty the first SA after having
pushed 1, 2 and again after 1, 2, 3, ∅. The first SA contains 1,2,3 and misses one element to be full. Thus, the SA
would not be emptied and two more elements could be (attempted) to be pushed onto the SA before trying to empty
it again, but the SA becomes full after pushing one more element. Therefore, we perform restructuring operations
more frequently, ie. for SA i we execute EmptyTwoParts after every 2i operations (rather than after 2i+1). The last
SA that can be emptied is the second to last, ie. the one with index s − 2. The restructuring is done in Algorithm
MoveBetweenSAs which executes for a push operation EmptyTwoParts on all parts as described. It takes as input the
counter of the current operations and returns the next value for the counter, which is (usually) the counter incremented
by 1. However, once the maximal possible SA has been shifted the operation counter is reset to zero, eg. for s = 5 the
counter is reset after 2s−2 = 8 operations. The sequence of maximal SA indexes where parts might be moved to the
next SA is a repetition of the sequence 0,1,0,2,0,1,0,3.
Pop: The pop operation (and its subroutines) behave analogously to push but reverse. ShiftPartsLeft shifts parts of a
SA within the SA towards the beginning. In contrast to ShiftPartsRight, we do not need a parameter to indicate whether
we actually perform the operation or not. ShiftPartsLeft only shifts a part, if the first part is empty. RefillTwoParts
moves the first part of SA i + 1 to the beginning of SA i. One full part in SA i + 1 corresponds to two full parts in SA
i. As for emptying of parts and right shifts, no non-empty parts are overwritten.
5
3.2. Analysis
Theorem 1. The LIFO queue is oblivious.
135
140
145
150
155
160
Proof. According to Definition 1 we require that memory accesses are independent of the input. (They are allowed to
be dependent on the number of operations.) None of the procedures in Algorithm 1 accesses memory cell dependent
on an input value, ie. all loop-conditions do not depend on the input and any conditional access to memory cells of
the form 'cell0:=a if cell1=x else b' can be expressed as multiplications (Section 2).
We analyze push and pop operations with respect to time complexity (Theorem 2) and correctness (Theorem 3).
In the worst case a single operation might take Ω(n), where n is the maximal length of the queue. We prove that on
average, the time is only logarithmic in n.
log n
i=0 2i+1/2i = O(log n).
Theorem 2. For the LIFO queue a pop and push operation requires amortized O(log n) time, ie. 14q log(n/q) E-Ops
and 8q + 2 C-Ops.
The proof uses that two parts of SA i of length 2 · 2i are refilled (emptied) after every 2i push (pop) operations. Since
there are O(log n) SAs we get timeP
Proof. SA i is refilled (emptied) after every 2i pop (push) operations. After refilling (emptying) all SAs from index s−2
to 0, ie. after 2s−2 pop (push) operations, we start over by considering SA 0 only. The average run-time increases up to
the point, where SA s − 2 is considered. Thus, it suffices to compute the average number of operations for a sequence
of 2s−2 pop (push) operations. We analyze pop operations by counting of E-Ops followed by C-Ops. CmpExEr needs
7 E-OPs (2 additions, 1 subtraction, 4 multiplications). ShiftsPartsLeft for SA i needs 2i · 7(q − 1). RefillTwoParts
on SA i performs one shift in list i + 1 and moves one part of it, yielding 2i · 7(q − 1) + 2i · 7q = 7 · 2i · q. Since
RefillTwoParts on SA i is called after every 2i−1 pops, on average refilling of SA i contributes by 7 · 2i · q/2i−1 = 14q
E-Ops. By definition we have n =Ps−1
i=0 q · 2i = q · (2s − 1) yielding s = (log(n/q)) + 1. Summing over all
SAs gives
i=0 S i =Ps−1
Xi=0
s−2
14q = 14q(s − 1) = 14q log(n/q)
The analysis of C-Ops is analogous. CmpExEr contains zero comparisons. In ShiftPartsLeft we perform one com-
parison (line 1) and one in each of the 2i · (q − 1) iterations. A refill of SA i takes 2q comparisons (q to compute
isEmpty in RefillTwoParts (line 1) and q within ShiftPartsLeft. Therefore, the number of comparisons becomes
Ps−2
i=0 2q/2i−1 ≤ 8q. Adding two C-Ops due to lines 1-2 in Algorithm Pop completes the proof for pop. The push
operation is analyzed in the same manner.
Lemma 1. Each part Pi, j can only be in one of two states: empty (all elements being ∅) or full (no elements being
∅).
This follows since we modify either all or none of the elements of a part.
Proof. Initially, all parts are empty. Parts of the first SA can only be full or empty, since they contain at most one
element. Parts of SA i > 0 are only modified due to mergers, splits and shifting. Right or left shifting of a part within
a SA is done for entire parts. The part overwritten is an exact copy of the part being shifted. The part being shifted
becomes empty, ie. all elements are set to the empty element. When all parts of SA i are full, the last two parts of a
SA, ie. Pq−2,i and Pq−1,i, each of size 2i are shifted to the next SA, ie. to become the first P0,i+1 of size 2i+1. This part
is filled completely. A filled part in SA i (see procedure RefillTwoParts) split into two parts of the same size, yields
two full parts in SA i − 1.
Theorem 3. The LIFO queue works correctly.
We show that no elements are overwritten and no empty elements are returned if the array is non-empty since we refill
and empty parts of SAs sufficiently often.
6
165
170
175
180
185
190
195
200
205
210
Proof. In Algorithm 1 no parts are overwritten if the first element of a part is non-empty – see definition and usage
of variables f ull&doOp, empty&doOp in ShiftPartsLeft/Right; isFull, isEmpty in Empty/RefillTwoParts; line 2 of
push with Ei,0,0 , ∅. Since all elements of a part are either the empty element or differ from it (Lemma 1), checking
the first element suffices to avoid overwriting of non-empty parts.
We first show that there is no interleaving of empty and non-empty SAs. Let the t-th SA be the largest SA such that
at least one part is full. All SAs i < t contain at least one non-empty part. An arbitrary sequence of pushes cannot
reduce the number of full parts in a SA below two. This follows since we only empty two parts of a SA if all four
parts are full. An arbitrary sequence of pops cannot completely empty a SA except the last, since SA i being of size
q · 2i is refilled with elements from SA i + 1 after every 2i pops (see MoveBetweenSAs in Algorithm 1).
Next we show that there is no interleaving of SAs with some non-empty parts and SAs with only full parts. EmptyT-
woParts executes on SA i before it executes on SA j < i. Upon execution there are two possibilities: Either no or two
parts are moved to SA i + 1. In the first case at most 3 parts are full in SA i and thus, we could insert one more part,
in the second case the SA is full and two parts are emptied. Either way, it suffices to empty SA i after two parts in SA
i − 1 (might) have been filled. Since this corresponds to 2i−1 elements, our choice of calling EmptyTwoParts i after
every two 2i−1 operations suffices (see MoveBetweenSAs in Algorithm 1). Therefore, not all parts of a SA can be full,
if there is space in a larger SA. For refilling parts an analogous argument applies.
4. FIFO
A First-In-First-Out (FIFO) queue needs fast access to the first and the last element. We use an array of LIFO
queue variants of increasing lengths, ie. each SA of the FIFO queue is itself a LIFO queue. Each LIFO queue stores
elements in 'reverse' order, meaning the first element to be popped in the LIFO queue is the oldest element the LIFO
queue contains. In this way we can efficiently access the oldest element of each LIFO queue. The array structure is
visualized in Figure 3. Each LIFO queue matches on SA.
For a pop operation the LIFO queue with largest index that is non-empty is identified. Then an element is popped
from that queue. To make the algorithm oblivious we execute a pop operation on every LIFO queue within the FIFO
queue. We start from the back and pop an element from each LIFO queue, ie. SA, until the first non-empty LIFO
queue has been identified. For the remaining queues we execute pops using a parameter to indicate that, in fact, no
element should be popped. The key point is that indepedent of the value of the parameter the same memory cells are
accessed.
PopperQueue: A LIFO queue offers more functionality than is needed, since we do not push elements in the
front but only pop them except for the first queue, which is just a single element. Opposed to a LIFO queue, we can
therefore refill a SA completely. We reduce the number of parts from four to two. Using more parts per SA is slower
since we must shift the same elements multiple times rather than moving them less often in bigger chunks, ie. larger
SAs. We can reuse most LIFO procedures (Algorithm 1) without modification, ie. ShiftsPartLeft, RefillTwoParts and
Pop. We call this LIFO variant "PopperQueue". It is a special case of the LIFO queue from Section 3. It has the same
(asymptotic) properties, but it is roughly a factor of two faster, since it uses less parts and therefore requires less shifts
within a SA, ie. compare Theorem 2 for q = 4 (LIFO) and q = 2 (PopperQueue).
Due to the more involved array organization of a FIFO queue, the emptying of parts and refilling of parts needs
careful attention. It is not possible to concatenate two parts to get a larger part without extra processing, ie. two arrays
(of PopperQueues) placed after each other generally do not yield an array representing a larger PopperQueue with a
valid structure. The concatenation could give partially filled parts. For example, assume that there are two queues with
one SA and two parts, eg. [1−] and [−4], naive concatenation yields [1 − − 4 − −] having the partially filled part
− 4. Furthermore, we have to ensure a correct ordering of the elements within LIFO queues when moving elements
between them.
If one last part of the PopperQueue stored in SA i is full, we move 2i elements from queue i to the very last part
of queue i + 1. We pop one element after the other from queue i and put it directly into the last part of queue i + 1,
ie. the element of the j-th pop is put at the j-th position of the last part. At this point the whole queue i + 1 (except
the last part that was just inserted) might be empty which would cause subsequent calls of pop on queue i + 1 to fail.
Therefore, we attempt to shift elements from the last part of the last SA of LIFO queue i + 1 consisting of newly
inserted elements up to the first SA of queue i + 1.
7
FIFO Queue F with 3 LIFO queues A(m)
=
L u
Subarray S(i) with 2 parts P(i
t
(
S(i) = P(i,0) P(i,1)
! "#. $% &
(m) wi
subarrays S
(m) = [S(0)
(m-1)]
:
> )
2
with 2i elements: P(
i456 7 89;<=
,0) E(
?@ABCD EFGHIJKMNOPQRSTUVW
i-1)
Part P
-/013
Elements pus
,'*+
ueue
hXY Z[ \]^_ `abc d
e fg
jk l mno
pq r
¡ ¢£¤
¥¦ §
´ µ¶
·¸ ¹º »¼½
¾¿ À
Í ÎÏ
ÐÑ ÒÓ ÔÕÖ
×Ø Ù
ae çè
éê ëì íîï
ðñ ò
ÿ
! "
0 12
34 56 789
:; <
MN OP
QR STUVWX
YZ [
- - - -
- - - -
2 - - -
2 - - -
2 3 4 -
2 3 4 -
2 3 4 5
6 - - -
svw
©ª
ÁÂÃ
ÚÛÜ
óôõ
#$%
6
6
= >? @ ABC
\ ]^ _ `ab
Elements pop
plm noqr stuv wxyz {}~
¡
¢£ ¤¥¦§©
ª« ¬
½ ¾¿
ÀÁ ÂÃÄÅAEÇ
ÈÉ Ê
Û ÜÝ
Þss àáâãäå
aeç è
ù úû
üý þÿ
6
6
6
6
6
!" #$%
3 45
68 9:;<=>
?@ AB CDE
- ®¯ ° ±²³
Ë ÌÍ Î ÏÐÑ
é êë ì íîï
7
- - -
- - -
R ST
UV WXYZ[\
]^ _
l mn
opqrs tuv
wx y
- - - -
- - - -
- - - -
&'(
FGH
`ab
z{}
xy z{
«¬ -®
ÄÅ AEÇ
ÝÞ ßà
ö÷ øù
&' ()
DE FG
cd ef
´µ ¶·
ÒÓ ÔÕ
ðñ òó
)* +,
IJ KL
cd ef
~
- - - - - - - - - -
- - - - - - - - - -
- - - - - - - - - -
- - - - - - - - - -
- - - - - - - - - -
- - - - - - - - - -
- - - - - - - - - -
2 - - 3 4 - - - - -
2 - - 3 4 - - - - -
2 3 4 - - - - - - -
} ~
¯ ° ±²³
È É ÊËÌ
á â ãäå
ú û üýþ
-
* + ,./
H I JKL
g h ijk
2 3 4 - - - - - - -
4 - - - - - - - - -
4 - - - - - - - - -
- - - - - - - - - -
- - - - - - - - - -
- - - - - - - - - -
- - - - - - - - - -
- - - - - - - - - -
- - - - - - - - - -
- - - - - - - - - -
¸ ¹ º»¼
Ö × ØÙÚ
ô õ ö÷ø
-
. / 012
M N OPQ
g h ijk
Figure 3: Sequence of pushes and pops onto a FIFO queue
8
The push operation for the FIFO queue appends elements to the end of the very first LIFO queue. Since it is of length
two, we shift the second element of it to the left and then set the second position to the newly inserted element.
215
Corollary 1. For the FIFO queue a pop operation requires O(log2 n) and a push O(log n) time on average.
i=0 O(i) = O(s2) = O(log2 n).
For a push we move blocks of size 2i from SA i, ie. PopperQueue i, to SA i + 1 after every 2i operations, which needs
Proof. For a pop of the FIFO queue we do a pop for each of the PopperQueues givingPs
time linear in the queue length. Summation givesPs
i=0 O(2i/2i) = O(s) = O(log n).
4.1. Fast FIFO (and Double-Ended Queues)
FIFO queues are often used as buffers to distribute peak loads across a longer timespan. Commonly, a producer
pushes elements onto the queue continuously (as a stream), while a consumer repeatedly takes an element and
processes it. Buffering always introduces some delay in processing. Thus, usually an additional delay is tolerable. A
pop on the fast FIFO queue only returns an element given the queue has been filled partially, ie. it is at least half full.
Our FIFO queue that has only amortized O(log n) overhead rather than O(log2 n). The idea is to use two queues
"back to back": one for popping and one for pushing. The two queues share the last part, ie. both treat this part
as belonging to them. Thus, elements are pushed onto one of the queues and are continuously shifted to the right
with newly inserted elements until they reach the queue for popping. A pop only returns an element after its last
part of the last SA (shared with the pushing queue) has been filled. The same ideas also apply to double-ended queues.
For the Fast FIFO Queue (B2B Queue) the time complexity of a push and pop matches the corresponding operations
for the LIFO queue.
Corollary 2. For the B2B-FIFO Queue a pop and push operation require O(log n).
5. Double-Ended Queue
A double-ended queue supports popping elements at the head and tail as well as prepending elements at the
beginning and appending them at the end. We combine ideas for LIFO and FIFO queues. We use an array of queues
(as for FIFO queues) to address the need to push elements to the head of the array and pop them from the tail. Since
elements can also be pushed at the back, we use LIFO queues, ie. SA i of the double-ended queue is given by a LIFO
queue with i + 1 SAs (rather than PopperQueues). Pushing to the back requires identifying the last non-empty SA, ie.
queue, as for popping from the back in the FIFO queue. However, we can only push the element onto the queue, if it
is non-full, otherwise we push it onto the next queue. Popping elements from the front might trigger refilling of SAs.
In turn, we have to move the newest elements of one SA to another. Identifying the newest elements of a LIFO queue
(with elements sorted by age, ie. ascending insertion order) is cumbersome, since there is only efficient access to the
oldest element. To reverse order, we remove all elements from the array (using a sequence of pops) and insert them
into a temporary LIFO queue. This yields a queue sorted by newest to oldest elements. Then we move elements by
popping them from the temporary queue to the queue to refill, ie. for queue i we move 2i+1 elements. The remaining
elements are pushed back onto the emptied queue (used to create the temporary LIFO queue).
Theorem 4. Any operation on the double-ended queue has amortized time O(log2 n).
220
225
230
235
240
245
Operations are similar to the LIFO queue, except for refilling and emptying that needs an additional logarithmic factor
due to the popping and pushing of elements rather than direct access in O(1).
250
Proof. Pushing and popping to the front works the same as for LIFO queues except for the refilling and emptying
of full SA. We require an additional logarithmic factor, since we cannot just copy elements of one SA, ie. queue, to
another but we first pop them from the LIFO queue onto a temporary queue. More precisely, each element access
using a pop requires amortized O(log n) as shown in Theorem 2 rather than O(1). Pushing and popping to the back
requires executing a constant number of push and pop operations for all parts constituting LIFO queues. Since we
have O(log n) queues and each operation on a LIFO queue requires O(log n) (see Theorem 2), a single push and pop
operation requires O(log2 n).
255
9
260
265
270
275
280
285
290
295
300
6. Double-Ended Priority Queue
In this scenario, each data item has a priority. A double-ended priority queue can return either the data element
with the smallest or largest priority. The queue structure is the same as for double-ended queue. We ensure that each
SA, ie. LIFO queue, contains elements sorted in descending priority. When moving elements from one queue to
another, ie. to empty a full queue or refill a queue, we first create one single sorted array containing all elements from
both queues and then refill the smaller queue up to half of its capacity with the elements of smallest priority and put
the other elements in the larger queue. The sorting can be done by merging both arrays.
Popping the element of minimum priority requires finding the smallest element in SA 0. Popping the element
of maximum priority requires checking all parts, since we do not know which parts contain elements and which do
not as well as which part contains the element with largest priority. More precisely, we first (peek) all parts and find
the element and part with the maximum element. After that we perform a pop on the (first) queue containing the
maximum element. This is done by executing a pop for all parts. The parameter of the pop operation, determining
whether the operation indeed removes an element from the queue, must be set accordingly, ie. it is ∅ for all but the
queue containing the maximum element.
The restructuring is somewhat more involved. Upon a push that requires restructuring, eg. either refilling or
emptying queue i we first create one sorted array in increasing order by merging both queues as done for ordinary
mergesort (see also Section 7). We then refill SA i until it is half full with the smallest elements (in reversed order)
and insert the remaining to the next SA (in reversed order).
Theorem 5. Any operation on the double-ended priority queue has amortized time O(log2 n).
Proof. We discuss time followed by correctness. Pushing an element to the front (or popping the element of minimum
priority) works the same as for LIFO queues except for the emptying and refilling of full SA. We require an additional
logarithmic factor to move elements from queue i to queue i + 1 (or the other way around), we create a temporary
queue by repeatedly popping the element of maximum priority and adding it to the temporary queue. Each element
access using a pop requires amortized O(log n) as shown in Theorem 2 rather than O(1). Moving the elements from
the temporary queue onto the (new) queues i and i + 1 has the same asymptotic time complexity. Therefore, push to
the front need O(log2 n) time. Popping the maximum priority element requires executing a pop operation for all LIFO
queues (plus restructuring). Since we have O(log n) queues and each operation on a LIFO queue requires O(log n)
(see Theorem 2), a single push and pop operation requires O(log2 n). Popping the maximum priority element requires
executing a pop operation for all LIFO queues (plus restructuring). This requires O(log2 n).
Correctness of a pop of maximum priority follows, since we maintain all queues in descending order of priority. Thus,
the element of maximum priority is the first element in one of the queues. Since we consider the first elements of all
queues and return the one of maximum priority, correctness follows. For the minimum we only investigate the first
queue. Since upon every restructuring operation on queue i we keep the smallest half of both queues i and i + 1 in
queue i, it holds that after a restructuring all elements in SA i are smaller than any element in SA i + 1 > i. Using
induction, we have that the smallest element is in SA 0.
7. Oblivious Mergesort
Our oblivious mergesort algorithm (O-Mergesort) divides an unsorted array (or list) into SAs of one element. It
repeatedly merges two arrays of equal length to obtain a new sorted array of double the length until there is only
one array remaining. This array is sorted. To make the sorting procedure oblivious requires a queue that supports a
conditional pop, ie. we pop the element of the array if it is smaller than another element. For short arrays (of length
1), we use a naive sort. Otherwise, two PopperQueues are merged by repeatedly comparing the first element of each
queue A and B and appending the smaller one to the result array C. Note, that since A and B are sorted the element
put into C is the smallest element in both arrays. We pop an element from the array which element we just appended
to C – see Algorithm O-Merge 2.
Theorem 6. Sorting of an array of n elements requires at most 85n log n C-Ops and a total of n · (3 + 560(logn − 1) +
28 log2 n) E-Ops.
10
Algorithm 2 O-Merge
Input: Sorted PopperQueue A and B of length l
Output: Merged LIFO Queue C
b := 1 if A[0] ≤ B[0] else 0
C[0] := A[0] · b + B[0] · (1 − b)
C[1] := B[0] · b + A[0] · (1 − b)
eleA := A.pop(1) {Returns smallest element in A}
eleB := B.pop(1)
for k = 0 to 2 · l − 2 do
1: if l = 1 then
2:
3:
4:
5: else
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16: end if
{Set C[k] to the smallest element in A union B and remove the element}
b := 1 if eleA ≤ eleB else 0
C[k] := eleA · b + (1 − b) · eleB
eleA := A.pop(b) · b + (1 − b) · eleA
eleB := A.pop(1 − b) · (1 − b) + b · eleB
end for
C[2 · l − 1] := eleA · b + (1 − b) · eleB
305
310
315
320
Proof. The merger of two arrays of size l each requires 4l pop operations, each requiring 18 comparisons using
Theorem 2 with q = 2. Additionally, we need one more comparison per iteration. This gives a total of 85l S-Ops for
merging two arrays. In total we get the following boundP
The naive sort of two arrays of size one comparing the two elements requires 5 E-Ops. In total there are n/2
queues of length 1 yielding a total of 5n/2. The merger of two arrays of size l > 1 requires each 4l pop operations,
each requiring 28(log(l/2) + 1) = 28 log l E-Ops.2, giving a total of 112l log l. Additionally, we need 5 E-Ops for each
of the 2l operations, giving a total of 10l E-Ops. Overall we get l(10 + 112 log l). Overall, we get
2log n− j · 85 · 2 j ≤ 85n log n S-Ops.
log n−1
j=0
5n/2 +
2log n− j−1 · 112 · 2 j · (log(2 j) + 10)
log n−1
Xj=1
Xj=1
log n−1
= 5n/2 +
n · 56 · (log(2 j) + 10)
≤ 3n + 56n(log2 n/2 + 10(log n − 1))
= n · (3 + 560(log n − 1) + 28 log2 n)
The analysis uses that we merge n
2i arrays of length 2i and Theorem 2 to bound the time two merge two arrays. We
improve on [9] by a factor of more than 1000 in terms of the number of comparisons, ie. C-Ops. Comparisons are
often used to analyze sorting algorithms, since typically the total operations involved is proportional to the number of
comparisons. In our case, this does not necessarily hold, since we only require one comparison for shifting a large
number of elements. Therefore, the costs for shifting might dominate the costs for comparisons. To ensure a fair and
objective comparison among algorithms we also analyzed the number of other operations, ie. E-Ops, since they are
the dominant factor in our algorithm. With respect to the total number of operations O-Mergesort is asymptotically
worse by a factor log n. However, due to the extremely large constants used in the state-of-the-art [9] we use less
operations for all practically relevant scenarios, ie. for arrays of length up to roughly 25300. For illustration, when
sorting 10 billion elements we need more than 100x less E-Ops. Furthermore, E-Ops (or XORs, ANDs) are generally
less complex than comparisons, therefore in practice the speed-up might be even larger.
2We have log(l/2) + 1 rather than log(l/2) using Theorem 2 with q = 2 since we merge arrays of length l = 2x but we only support mergers of
arrays of length 2y − 1, thus we need y = x + 1 = log(l/2) + 1
11
325
330
335
340
345
350
355
8. Quicksort
Our oblivious quicksort algorithm (O-Quicksort) is a comparison-based divide and conquer algorithm. Small
arrays of size at most 4 log2 n are sorted using O-Mergesort. Larger arrays are recursively split into two smaller
(sub)arrays. An array is split using a pivot element. All elements less or equal to the pivot are put in one array and
all larger elements in the other array. Ideally, both arrays are of the same size. However, naive splitting likely leads
to badly balanced arrays leading to O(n2) run-time since an oblivious algorithm must treat both parts as potentially
large. However, when choosing the median as pivot, it is possible to ensure that both arrays are of equal size. We
compute an approximate median for all elements (Section 8.1). Unfortunately, choosing an approximate median still
leaves some uncertainty with respect to the exact array lengths after the splitting. Therefore, in the partition process
(Section 8.1), rather than swapping elements within one array, we create two arrays of fixed length, one for elements
larger than the pivot and one for all other elements. Since the length of each of the two arrays must be fixed using
conservative upper bounds, their sum of lengths exceeds the length of the array to be split. To get a single sorted array
requires a special reunification of both arrays (Section 8.2). For simplicity, we assume that all elements to be sorted
are distinct. This assumption is removed in Section 8.3.
8.1. Random Pivot Choice and Partition
Algorithm RandomPivot chooses several elements uniformly at random, sorts them and picks the median. By
choosing the median of a sufficiently large sample of elements we ensure that the chances of a split resulting in very
unbalanced arrays is small. We pick a fixed number of samples np, sort them, eg. using the O-MergeSort algorithm,
and then pick the middle element l/2 of the sorted array of length l as pivot.
Algorithm 3 RandomPivot
Input: Array A of length l, number of samples np
Output: Pivot p
1: P := Set of np elements of A chosen uniformly at random
2: S P := Sorted samples P {eg. using O-MergeSort}
3: p := S P[l/2] {Choose middle element (= Median) as pivot}
For the partitioning the entire array is split into two arrays, one with all elements being smaller than the pivot and
one with all elements being larger. The two arrays are given by two LIFO queues. We push elements that are smaller
than the approximated median on one of the queues and the larger elements on the other queue. We discuss the case
of duplicates in Section 8.3.
1
1 + p13c(log n)/np ≥ 1/4
Theorem 7. Algorithm RandomPivot returns a pivot p such that at least a fraction c f = 1
2 ·
of elements of an array of length n are larger than p and the same fraction is smaller than p with probability 1 − 1/nc
for np ≥ 13 · c · log n.
We obtain tail estimates using carefully applied Chernoff bounds.
Theorem 8 (Chernoff Bound). The probability that the number X of occurred independent events Xi ∈ {0, 1}, i.e.
X := Pi Xi, is not in [(1 − c0)E[X], (1 + c1)E[X]] with c0 ∈]0, 1] and c1 ∈]0, 1[ can be bounded by p(X ≤ (1 −
c0)E[X] ∨ X ≥ (1 + c1)E[X]) < 2e−E[X]·min(c0 ,c1)2/3.
Proof of Theorem 7:
Proof. The value of c f is minimized, when np is smallest. Thus, the bound c f ≥ 1/4 follows from substitution of the
lower bound, ie. np = 13 · c · log n, into c f = 1
= 1/4. The theorem holds if the
2 ·
pivot does not stem from the c f · n smallest or largest elements. If we pick less than c f · np < np/2 elements S ⊆ A
from the c f ·n smallest and less than c f ·np < np/2 elements L ⊆ A from the c f ·n largest elements this will be the case.
The reason being that the pivot p is the element at position np/2 in the sorted sequence and, thus, it will not be from
1 + p13c(log n)/np
1
= 1/2 ·
1
1+ √1
12
360
the set of c f · n smallest or largest elements. We expect to pick c f · np elements S out of the c f · n smallest elements
(and analogously for the largest), ie. E[S] = c f · np. We seek the smallest factor f > 1 such that when exceeding the
expectation by factor f the pivot is not chosen correctly. We have f · c f · np = np/2, if f = 1/(2 · c f ). The probability
that the expectation is exceeded by a factor f > 1 or more is given using a Chernoff bound (see Theorem 8) by
prob(S > f · E[S ]) < 1/2( f−1)2 /3·c f ·np ≤ 1/21/12·( f−1)2·np (Using c f ≥ 1/4)
= 1/21/12·(1/(2·c f )−1)2·np (Using f = 1/(2 · c f ))
= 1/21/12·((1+√13c log n/np)−1)2·np = 1/213/12·c log n = 1/n13/12·c
In the same manner we can compute prob(L > f · E[L]). Therefore the probability of both events becomes for n
sufficiently large:
prob((cid:0)L ≤ f · E[L](cid:1) ∧(cid:0)S ≤ f · E[S ](cid:1)) ≥ 1 −(cid:0)prob(L > f · E[L]) + prob(S > f · E[S ](cid:1) ≥ 1 − 1/nc
365
Expected Length
Min. Length
B
11 17 29 32 33 39 42 44 46 - -
Max. Length
C
91 87 84 71 61 51 48 - - - -
Copying of elements up to min. Length
A
11 17 29 32 33 - - - - - - 61 71 84 87 91
Copying of remaining elements (step by step)
A
11 17 29 32 33 39 - - - - 51 61 71 84 87 91
11 17 29 32 33 39 42 - - 48 51 61 71 84 87 91
11 17 29 32 33 39 42 44 - 48 51 61 71 84 87 91
11 17 29 32 33 39 42 44 46 48 51 61 71 84 87 91
Figure 4: Merger of two subarrays within O-Quicksort
8.2. Sorting and End-to-End Array Merger
370
So far we have obtained well-balanced, but unsorted SAs. Since we do not have access to their exact lengths
we use a conservative bound on their lengths given by the analysis of the partition process. O-Quicksort recurses
on two separate arrays stemming from the partitioning. We sort the array of smaller elements B than the pivot in
ascending order and the array of larger elements C in descending order. At the end, both arrays must be merged to
get a single final array of sorted elements. This requires some care since we do not know their exact lengths. Due to
the partitioning process we can bound the minimum length of B and C to be l/2 · (1 + ǫ) with ǫ := p13 · c · (log n)/l.
We copy the elements (up to the guaranteed minimum length bound) to the final array, so that these elements appear
sorted. This means we fill the final array with B from the left end towards the right and with C from the right end. The
13
Algorithm 4 O-Quicksort
Input: Array A of length l, Sort ascending: asc
Output: Sorted array A
B, C := Partition(A)
O-Quicksort(B, l/2 · (1 + ǫ), T rue))
O-Quicksort(C, l/2 · (1 + ǫ), False))
for k = 0 to l/2 · (1 − 2ǫ) do
B[k] := B[k] {Copy elements from B to A}
B[l − k] := C[k] {Copy elements from B to A}
end for
for k = l/2 · (1 − 2ǫ) to l/2 · (1 + ǫ) do
1: ǫ := p13 · c · (log n)/l
2: if l ≥ 4 log2 n then
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14: else
15:
16: end if
end for
if B[k] , v∅ then B[k] := B[k] {Copy elements from B to A}
if C[k] , v∅ then B[l − k] := C[k] {Copy elements from C to A}
B:= Sort using O-MergeSort or other alg. either ascending or desc. depending on asc
375
entire process is illustrated in Figure 4. The remaining elements are handled in the same fashion, but before setting an
array element in A to be an element from B or C, we check whether the element in A is still empty.
Theorem 9. O-Quicksort needs O(n log n) C-Ops and E-Ops. It produces a correct sorting with probability 1 − 1/nc
for an arbitrary constant c.
The recurrences are somewhat involved, since the lengths of both arrays used for recursion exceeds the original length
of the array being split. We conduct a staged analysis to obtain (asymptotically tight) bounds.
380
Proof. The complexity T (n) of O-Quicksort in terms of comparisons can be stated using a recurrence. For one call
(ignoring recursion) to an array of length l > 4 log2 n we have that the complexity is given by partitioning the array
being O(l) plus the reunification of both sorted arrays, ie. the copying of elements being also O(l). Thus, we get a
total of O(l) = c0 · l for some constant c0. We obtain the following recurrence for an array of length l using ǫ = √a/l
with a := 13 · c · (log n):
385
First call: T (l) = 2T (l/2 · (1 + pa/l)) + c0 · l
Second call: T (l/2 · (1 + pa/l)) ≤ 2T (l/4 · (1 + pa/l) · (1 + pa/(l/2))) + c0 · l/2 · (1 + pa/l)
≤ 2T (l/4 ·(cid:0)1 + pa/(l/2)(cid:1)2) + c0 · l/2 · (1 + pa/l)
Third call: T (l/4 · (1 + pa/(l/2))2) ≤ 2T (l/8 · (1 + pa/(l/4))3) + c0 · n/8 · (1 + pa/(l/4))3
i-th call: T (l/2i ·(cid:0)1 + pa/(l/2i−1)(cid:1)i) ≤ 2T (l/2i+1 ·(cid:0)1 + pa/(l/2i)(cid:1)i+1) + c0 · l/2i ·(cid:0)1 + pa/(l/2i)(cid:1)i+1
(1)
Assume we start splitting the entire array A with l = n. The total number of operations (C-Ops and E-Ops) at
recursion depth i is given by the additive term in Equation (1) multiplied by the number of calls to O-Quicksort being
2i, ie.
390
= c0 · n ·(cid:0)1 + pa/(n/2i)(cid:1)i+1
The total operations for the first r := log n − 8 log log n recursions is given by:
Xi=0 (cid:0)1 + pa/(log n)8(cid:1)log n
2ic0 · n/2i ·(cid:0)1 + pa/(n/2i)(cid:1)i+1
≤ c0 · n ·
c0 · n ·(cid:0)1 + pa/(n/2i)(cid:1)i+1
r−1
r−1
Xi=0
≤ c0 · n · log n
14
After r recursions the size of the input sequence for the recursive calls is at most n/2r · (1 + pa/(n/2r−1))r ≤ 2· log8 n
(for n sufficiently large). For another 6 log log n recursions on an array of length 2 log8 n the number of operations is
bounded by:
395
400
405
410
c0 · 2 log8 n ·
6 log log n−1
Xi=0
(1 + s 4 log n
(log n)2 )6 log log n = c0 · 2 log8 n ·
6 log log n−1
(1 +
Xi=0
)6 log log n ≤ 4c0 log8 n
2
plog n
The size of the remaining arrays is 4 log2 n using the same derivation as above using r recursions. To sort such an
array using O-MergeSort requires O(log2 n log log n) C-Ops and O(log2 n(log log n)2) E-Ops (see Theorem 6). There
are 2log n−2 log log n = n/ log2 n such arrays, giving a total of O(n log log n) C-Ops and O(n(log log n)2) E-Ops. To obtain
a correctly sorted queue all executions of RandomPivot must be successful. We perform at most log n − 2 log log n
recursions. Thus, in total there are at most n calls to RandomPivot, each succeeding with probability at least 1 − 1/nc′
for an arbitrary constant c′. The probability that all succeed is at least (1 − 1/nc′)n ≥ 1 − 1/nc′−2. Choosing c′ = c + 2
concludes the proof.
8.3. Equal or Duplicate Elements
So far we focused on arrays of distinct elements. For non-distinct elements our algorithm can fail to compute
balanced arrays in case the chosen median is not unique. In the most extreme case all elements are the same and
the split would result in one empty array and one array being the same as the array to be split. Elements can always
be made distinct by appending a unique number at the end, eg. by appending a counter to a array of elements
(0, 0, 0), we get (00, 01, 02). We assign elements that are equal to the pivot p to both arrays such that their lengths
maintain balanced. In a first phase we create two arrays B, C and maintain a counter lp for the elements equal to p by
distinguishing three cases for an element x that is compared to the pivot p, ie. x < p, x > p and x = p. In the first
case, we assign x to array B and increment the length counter of lB. In the second case we assign x to C and increment
the length counter lC of C. In the third case, we increment just the counter lp. In the second phase we distribute lp
copies of p to the arrays B and C such that their difference in length is as small as possible. We perform l iterations,
where l is the number of elements in the array to be partitioned, ie. A. In each iteration we subtract one from lp. If lp
is zero the arrays remain the same. Otherwise, if the lengths of lB is less than lC, we append a copy of the pivot to B
and increment the length counter lB otherwise we do the same for C. The (asymptotic) complexity remains the same.
415
9. Applications
9.1. Confidential Expressions
Confidential expressions, hiding data as well as operations on the data are a rather straight forward application of
our LIFO queue with conditional operations as well as basic operations, e.g. addition and multiplication, from secure
computing schemes such as fully homomorphic encryption (FHE) or secure multi-party computation (MPC). We first
discuss the evaluation of non-confidential expressions. For brevity we only discuss the evaluation of expressions
involving numbers, additions and multiplications. We focus on evaluating postfix expressions. When using encrypted
values, the expression remains confidential. That is to say, despite computing the expression we do not learn anything
about the expression except its length. The key to achieve this is the conditional push, ie. we execute a push operation
but it only has an impact given that the element to be pushed is different from the special element ∅ that is not
appended to the stack. Our algorithm 5 requires linear run-time in terms of the number of elements in the expression
(or, more precisely, in the bound we get for the length of the expression).
To evaluate confidential expressions, all array elements of the input A must be encrypted as well as variables that
depend on the array elements. Variables that indicate array lengths or the number of operations do not have to be
encrypted.3 To this end, one can use any of the known scheme for computing on encrypted data such as FHE or MPC,
eg. [5] or [19]. These schemes provide basic operations such as addition and multiplication that allow to construct
other operations such as comparisons, subtractions and more. However, certain operations like accessing an array
420
425
430
3In Algorithm 5 the values of n and s do not have to be encrypted. In the LIFO queue and its sub-procedures, npu, npo,q,o,mi and s remain
unencrypted. All other variables are encrypted.
15
Algorithm 5 Case Study: PostFix expressions
Input: LIFO Queue A of postfix symbols of length at most n
Output: Result of evaluation
1: st := LIFO(s) {Choose number of SAs s such that array can hold at least n elements}
2: for i := 0 to n − 1 do
symb := A.pop(1)
3:
toPush := symb if symb is a number else ∅
4:
st.push(toPush)
5:
isAdd := 1 if symb = " + " else 0
6:
resAdd := st.pop(isAdd) + st.pop(isAdd)
7:
isMul := 1 if symb = " ∗ " else 0
8:
resMul := st.pop(isMul) · st.pop(isMul)
9:
toPush := isMul · resMul + (1 − isMul) · ∅
10:
toPush := isAdd · resAdd + (1 − isAdd) · toPush
11:
st.push(toPush)
12:
13: end for
14: return st.pop(1)
element using an encrypted index might occur linear overhead in the length of the array. In our case, we manage to
keep the asymptotic running time, since we only have to directly substitute additions, subtractions, multiplication and
comparisons operations.
435
9.2. Stock Span Analysis
440
445
450
455
The Stock Span Problem is motivated by financial analysis of stocks. The span of a stock's price on day d is
the maximum number of consecutive days until day d, where the price of the stock has been at most its price on d.
The well-known text book solution is given in Algorithm 6 taking linear time in the number of prices n.4 A straight
forward solution gives a quadratic run-time algorithm due to the nested loops, ie. due to the worst case behavior of
the inner loop. This renders the solution impractical for larger datasets. The total number of iterations (when being
summed across all outer loop iterations) of the inner loop is only n. A single iteration of the inner loop could perform
all n iterations for some inputs. To ensure obliviousness we would have to execute (asymptotically) n iterations of the
inner loop for every execution of the outer loop. Furthermore, the code contains direct array access, eg. price[i]. In the
obvious manner, this would also incur linear run-time overhead. However, it is possible to transform the nested loop
by essentially changing the inner loop to an 'if'-conditional first without changing the number of iterations of the outer
loop. Then we make the loop oblivious using a conditional expression if-then-else. Essentially, in Algorithm 6
we replace the while and do keyword in line 5 by an if and then. Lines 6 to 8 form the else part. We only show
the final pseudocode after the translation of the 'if' into oblivious code in Algorithm 7. Since we must execute both
branches of the i f to keep the condition confidential, the algorithm requires that we can execute the 'pop' operation
without impacting the data structure, ie. without actually performing a pop. This is supported by our data structure
by using a special element in case the condition evaluates to true. The algorithm uses a peek operation, which returns
the first element without removing it. It can be implemented using a combination of pop and push operation, eg.
x := pop(1), push(x).
10. Related Work
In 2009 Gentry [5] introduced a fully homomorphic encryption(FHE) scheme based on lattices. Since then the
field of computing on encrypted data (and circuits) has evolved rapidly, as summarized in [15]. All approaches for
FHE are based on either addition and multiplication or XOR and AND. Secure computations can also be carried out
using multiple parties, such that no party learns a secret (if it does not collude with other parties). Secure multi-
party computation was introduced three decades ago [24, 7] and is still subject to extensive research, eg. [19]. Both
4We adjusted it from http://www.geeksforgeeks.org/the-stock-span-problem/
16
Algorithm 6 Case Study: Stock Span
Input: LIFO Queue price of prices of length at most n
Output: LIFO S with spans (in reverse order)
1: st := LIFO(s) {Choose number of SAs s such that array can hold at least n elements}
2: st.push(0)
3: S .push(1) {Span first element is 1}
4: for i = 0 to n − 1 do
5:
6:
7:
8:
9: end for
while st.peek() , ∅ ∧ price[st[0]] ≤ price[i] do st.pop()
span := i + 1 if st.peek() , ∅ else i − st[0]
S.push(span)
st.append(i)
Algorithm 7 Case Study: Oblivious Stock Span
Input: LIFO Queue price of prices of length at most n
Output: LIFO S with spans (in reverse order)
1: st := LIFO(s) {Choose number of SAs s such that array can hold at least n elements}
2: pi := price.pop(1)
3: st.push((0, pi))
4: S .push(1) {Span first element is 1}
5: i := 0
6: for k := 0 to n − 1 do
(sti,stp) := st.pop(1)
7:
popNext := 1 if sti = ∅ or sti ≤ pi else 0
8:
pi := price.pop(popNext) · popNext + (1 − popNext) · pi
9:
i := i + (1 − popNext)
10:
span := i + 1 if st.peek() , ∅ else i − sti
11:
span := ∅ if popNext else span
12:
S .push(span)
13:
pushi := (i, pi) if (1-popNext) else (∅, ∅)
14:
st.push(pushi)
15:
16: end for
17
460
465
470
475
480
485
490
495
500
505
510
SMC and FHE can compute expressions beyond addition and multiplication such as comparisons. These operations
could be used as black boxes to make our algoritms work on encrypted data. [5, 22] mentioned that circuit privacy
is achievable by adding a (large) noise vector of the encrypted 0. The original work on SMC [24] already allowed to
hide circuits using garbled circuits. Our approach also allows to achieve circuit privacy in a novel manner by hiding
whether a certain operation really impacted the computation. Our work is not limited to circuits. "Oblivious RAM"
(ORAM) [6] disguises access patterns by a client from a server using randomization. The original solution [6] is based
on a hierarchy of buffers such that each level of the hierarchy consists of several blocks. One block per level is read
and always written on the first level. For each level, some blocks are real and some are dummy containing random
data. The original algorithm has been improved, eg.
in a recent tree-based scheme [20] each data block is stored
somewhere in the tree, ie. following a specific path from the root. After an access minor reordering involving the
accessed elements takes place, potentially, resulting in some data being sent to the client. Some schemes, eg. [20, 23],
trade-off performance (or memory consumption) and the risk for memory access errors. Oblivious data structures
for ORAM covering arrays and queues are discussed in [23]. They make use of data access locality encountered for
arrays and queues. A single access to a parent node in the position map returns pointers to multiple children. They
hide whether the operation is a read or write to a memory cell. However, assuming one knows that a write must
occur in a function, one knows that some memory cell is modified. We do not use traditional ORAM techniques.
Furthermore, in our scenario, knowing that a certain operation is performed, ie. a pop or push, still gives no hint
whether the data structure was modified or not.
Other work designed oblivious data structures particularly for SMC, eg. [12, 21]. The work [12] uses ORAM
structures and secret sharing among parties to achieve obliviousness. In contrast, [21] presents a deterministic scheme
for priority queues using the bucket heap concept for priority queues [3] coming with O(log2 n) overhead. Bucket
heaps partition memory into buffers of size 22i+1 and signal blocks of size 22i [3]. Buckets store actual elements,
whereas buffers store overflowing elements. Once a buffer is full it is merged into a bucket. [21] adjusted this setting
to use blocks of equal size. Our queue shares the common idea of organizing data in blocks of increasing size that is
also found in other work, eg. [14]. We differ from prior work [14, 21, 3] in several aspects, eg. we perform a more
fine-grained partitioning using multiple blocks, eg. in the view of [21] we introduce one more level of partitioning
for buffer and data blocks. We have also come up with a deterministic oblivious sequence of restructuring operations
to handle empty and full blocks rather than counting the number of elements in the queue, eg. [21]. In contrast to
our work, prior work also does not hide the impact of an operation (ie. they do not hide the number of elements in a
bucket), which is essential for securing control structures. Our fast B2B-FIFO queue introduces novel ideas such as
block sharing not found in prior work.
The paper [1] shows how to compute the k-th ranked element for SMC. The paper [13] discusses median computa-
tion. Such operations might prove valuable also for sorting, eg. selecting the median element for quicksort. However,
both protocols [1, 13] disclose the outcome of secure comparisons, which might require non-desirable client interac-
tion and is not considered secure. The SMC protocol for sorting in [25] runs in constant rounds but needs to know
the product of the range of numbers R and it has communication and computational complexity that is proportional
to product of the range of numbers times the number of elements, ie. O(n · R) (an improved version has O(n2)). To
achieve constant rounds it relies on the evaluation of unbounded fan-in gates.
Sorting networks are naturally oblivious, since they use a fixed sequence of comparisons among elements in an
array that is not related to the data stored in the array. They have been used for secure sorting in [11, 9]. The work [9]
is based on a network with 19600 · n log n comparators. A comparator can be implemented by a comparison yielding
a bit b followed by an exchange of two elements A, B, ie. A := b · B + (1 − b) · A and B := b · A + (1 − b) · B.
Therefore a comparator needs 7 E-Ops in addition to the comparison, yielding 156800 · n log n operations. Though
this is asymptotically optimal, it is of little practical relevance due to the number of comparators needed. Additionally,
the depth (of the sorting network) is of order n log n, which makes it non-parallelizable. Our algorithms improve on
it for all relevant scenarios (see Section 7 for a detailed comparison). The oblivious randomized Shellshort algorithm
[8] is asymptotically optimal in terms of the number of comparisons using several techniques such as permutation of
the array as well as shaker and brick passes of the array.
Oblivious algorithms for geometric problems are presented in [4]. Algorithms for graphs incurring overhead up to
linear factor (in the number of nodes) are given in [2]. Other work [18] based on ORAM designed data structures for
maps. They allow for history independence, ie. different sequences of operations lead to indistinguishable (memory
layouts of the physical) data structures.
18
11. Evaluation
515
520
525
We shed light on two aspects that are not immediate from the asymptotic analysis. First, on the one hand our
oblivious data structures are more involved than using a naive oblivious implementation traversing the entire array
for each operation, on the other hand we have shown that asymptotically it outperforms a naive implementation. The
key question is, whether our oblivious queues outperform already for queues of small capacity or only for those with
large capacity. Therefore, we compared our implementation against a simple 'linear' oblivious queue that accesses
all elements (that could be stored) for each operation. Thus, the run-time is linear in the capacity. Second, how
much slower is our array compared to a non-oblivious queue. We have shown that the asymptotic factors are of order
O(log n) and O(log2 n) depending on the queue type. Here, we give more precise insights.
We implemented the benchmarks in Python. The evaluation was run on a machine equipped with an Intel 2.5 GHz
quad-core CPU with 8 GB RAM on Windows 7. For the non-oblivious queue we ran 1 Mio. operations. For the
oblivous linear queue, ie. the naive oblivious queue traversing the entire array for each operation, we attempted to run
1 Mio operations, but stopped after 1 hour if the computation was still ongoing and estimated the time it would take
to compute 1 Mio. operations. For the oblivious data structures we executed max(100000, 2 · capacity) operations,
since the maximal run-time is achieved if we execute a multiple of the capacity. Each operation was chosen randomly
among a push and pop operation. Due to the obliviousness it does not matter what parameters we use for the push and
pop operation.
ObliviousFIFO
ObliviousFastFIFO
ObliviousLinearQueue
NonObliviousQueue
105
104
103
102
101
]
s
[
e
m
T
i
ObliviousLIFO
ObliviousLinearQueue
NonObliviousQueue
104
103
]
s
[
e
m
T
i
102
101
100
27
55
111 223 447 895 1791 3583 7167~2^13~2^14~2^15
100
12
28
60
124
252
508
1020
2044
4092
8188
Queue Capacity
(a) FIFO Queues
ObliviousDoubleEndedQueue
ObliviousLinearQueue
NonObliviousQueue
10 6
10 5
10 4
]
s
[
e
m
T
i
10 3
10 2
10 1
Queue Capacity
(b) LIFO Queues
ObliviousDoubleEndedPriorityQueue
ObliviousLinearQueue
NonObliviousQueue
106
105
104
]
s
[
e
m
T
i
103
102
101
10 0
3068
6140
12284 24572 49148 98300 196604 393212 7864281572860
100
3070
6142 ~2^13 ~2^14 ~2^15 ~2^16 ~2^17 ~2^18 ~2^19 ~2^20
Queue Capacit y
Queue Capacity
(c) Double-Ended Queues
(d) Double-Ended Priority Queues
Figure 5: Running Times results for 1 Mio. operations for our oblivious queues compared to linear oblivious and
non-oblivious queues
19
530
535
540
545
550
The plots in Figures 5a, 5b,5c and 5d show the run-times comparing all queue variants for increasing maximum
queue sizes for FIFO and LIFO queues. Qualitatively all queues behave similarly as predicted by the asymptotic
analysis. For small queue sizes (LIFO and FastFIFO up to about 60, FIFO up to about 500) a simple linear oblivious
queue has an edge over our more complex queues. For double-ended queues performance is somewhat worse, but
simple linear queues are also outperformed for moderate queue sizes. With growing queue sizes the exponential gap
becomes clearly visible between the linear oblivious queue and our implementations. The LIFO and fast FastFIFO
queue are more than 100x faster for queues of capacity about 10000. For FIFO queues we reach the boundary of
100x performance improvement for queues of capacity about 100000. Note, that it is not uncommon that a simple
algorithm with bad asymptotic behavior outperforms more complex algorithms. For example, a naive bubble sort
might also be faster than mergesort for small data sizes although the later is asymptotically much faster.
When it comes to overhead of our oblivious queue compared to the built-in Python queue (surrounded by a
wrapper), which uses memory proportional to the actual array size, the asymptotic behavior is well-visible. Our
LIFO and FastFIFO queue both have an asymptotic overhead of log n compared to non-oblivious queues that directly
accesses queue elements. This results in close to parallel lines in Figures 5b and 5b. The overhead is roughly a factor
40 for queues of size 10000. For FIFO queues the asymptotic overhead is larger, ie. log2 n. The overhead is a factor
of 200 for arrays of the same size. In the light of overhead that typically comes with secure computation, eg. FHE or
SMC that can reach more than 5-6 orders of magnitude [16], our overhead is very modest. We also want to emphasize
that to the best of our knowledge no prior work has compared against a non-oblivious implementation. We believe
this is a very important benchmark.
12. Conclusions
We have presented oblivious queues accompanied by theoretical and practical investigation having only
(poly)logarithmic overhead. Since queues are an essential part of everyday programming, we believe that they will
play a major role for enabling computation using encrypted data, in particular with focus on expression hiding. Still,
many more common data structures and operations have to be realized efficiently before any of the existing technolo-
gies such as FHE and MPC become practical for a large range of applications
555
References
[1] G. Aggarwal, N. Mishra, and B. Pinkas. Secure computation of the median (and other elements of specified ranks). Journal of cryptology,
23(3):373–401, 2010.
[2] M. Blanton, A. Steele, and M. Alisagari. Data-oblivious graph algorithms for secure computation and outsourcing. In Proceedings of the 8th
ACM SIGSAC symposium on Information, computer and communications security, pages 207–218. ACM, 2013.
560
[3] G. S. Brodal, R. Fagerberg, U. Meyer, and N. Zeh. Cache-oblivious data structures and algorithms for undirected breadth-first search and
shortest paths. In Algorithm Theory-SWAT 2004, pages 480–492. 2004.
[4] D. Eppstein, M. T. Goodrich, and R. Tamassia. Privacy-preserving data-oblivious geometric algorithms for geographic data. In Proceedings
of the 18th SIGSPATIAL International Conference on Advances in Geographic Information Systems, pages 13–22. ACM, 2010.
[5] C. Gentry. Fully Homomorphic Encryption Using Ideal Lattices. In Proc. 41st ACM Symposium on Theory of Computing, pages 169–178,
565
2009.
[6] O. Goldreich. Towards a theory of software protection and simulation by oblivious rams. In Proc. of 19th Symposium on Theory of comput-
ing(STOC), pages 182–194, 1987.
[7] O. Goldreich, S. Micali, and A. Wigderson. How to play any mental game. In Proc. of 19th Symp. on Theory of computing, pages 218–229,
1987.
570
[8] M. T. Goodrich. Randomized shellsort: A simple oblivious sorting algorithm.
In Proceedings of the 21st Symposium on Discrete Algo-
rithms(SODA), pages 1262–1277, 2010.
[9] M. T. Goodrich. Zig-zag sort: A simple deterministic data-oblivious sorting algorithm running in o (n log n) time. In Proc. of 46th Symp. on
Theory of Computing, pages 684–693, 2014.
[10] M. S. Islam, M. Kuzu, and M. Kantarcioglu. Access pattern disclosure on searchable encryption: Ramification, attack and mitigation. In
575
NDSS, volume 20, page 12, 2012.
[11] K. V. Jonsson, G. Kreitz, and M. Uddin. Secure multi-party sorting and applications. IACR Cryptology ePrint Archive, 2011:122, 2011.
[12] M. Keller and P. Scholl. Efficient, oblivious data structures for MPC. pages 506–525, 2014.
[13] Y. Lindell and B. Pinkas. Secure multiparty computation for privacy-preserving data mining. Journal of Privacy and Confidentiality, 1(1):5,
2009.
580
[14] J. C. Mitchell and J. Zimmerman. Data-oblivious data structures. In LIPIcs-Leibniz International Proceedings in Informatics, volume 25.
Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik, 2014.
20
[15] C. Moore, M. O'Neill, E. O'Sullivan, Y. Doroz, and B. Sunar. Practical homomorphic encryption: A survey. In Int. Symp.on Circuits and
Systems (ISCAS), pages 2792–2795, 2014.
[16] M. Naehrig, K. Lauter, and V. Vaikuntanathan. Can homomorphic encryption be practical? In Proceedings of the 3rd ACM workshop on
585
Cloud computing security workshop, pages 113–124. ACM, 2011.
[17] B. Pinkas and T. Reinman. Oblivious RAM revisited. In Advances in Cryptology (CRYPTO), pages 502–519. 2010.
[18] D. S. Roche, A. J. Aviv, and S. G. Choi. A practical oblivious map data structure with secure deletion and history independence. arXiv
preprint arXiv:1505.07391, 2015.
[19] J. Schneider. Lean and fast secure multi-party computation: Minimizing communication and local computation using a helper. SECRYPT,
590
2016, extended version: arXiv:1508.07690, https://arxiv.org/abs/1508.07690.
[20] E. Stefanov, M. Van Dijk, E. Shi, C. Fletcher, L. Ren, X. Yu, and S. Devadas. Path oram: An extremely simple oblivious RAM protocol. In
Proc. of the SIGSAC conference on Computer & communications security, pages 299–310, 2013.
[21] T. Toft. Secure data structures based on multi-party computation. In Proceedings of the 30th annual symposium on Principles of distributed
computing, pages 291–292, 2011.
595
[22] M. Van Dijk, C. Gentry, S. Halevi, and V. Vaikuntanathan. Fully homomorphic encryption over the integers. In Advances in cryptology–
EUROCRYPT 2010, pages 24–43. 2010.
[23] X. S. Wang, K. Nayak, C. Liu, T. Chan, E. Shi, E. Stefanov, and Y. Huang. Oblivious data structures. In Proc. of the Conference on Computer
and Communications Security, pages 215–226, 2014.
[24] A. C.-C. Yao. How to generate and exchange secrets. In Foundations of Computer Science(FOCS), 1986.
[25] B. Zhang. Generic constant-round oblivious sorting algorithm for MPC. In Provable Security, pages 240–256. 2011.
600
21
|
1904.08150 | 1 | 1904 | 2019-04-17T09:22:20 | A Brief Note on Single Source Fault Tolerant Reachability | [
"cs.DS"
] | Let $G$ be a directed graph with $n$ vertices and $m$ edges, and let $s \in V(G)$ be a designated source vertex. We consider the problem of single source reachability (SSR) from $s$ in presence of failures of edges (or vertices). Formally, a spanning subgraph $H$ of $G$ is a {\em $k$-Fault Tolerant Reachability Subgraph ($k$-FTRS)} if it has the following property. For any set $F$ of at most $k$ edges (or vertices) in $G$, and for any vertex $v\in V(G)$, the vertex $v$ is reachable from $s$ in $G-F$ if and only if it is reachable from $s$ in $H - F$. Baswana et.al. [STOC 2016, SICOMP 2018] showed that in the setting above, for any positive integer $k$, we can compute a $k$-FTRS with $2^k n$ edges. In this paper, we give a much simpler algorithm for computing a $k$-FTRS, and observe that it extends to higher connectivity as well. Our results follow from a simple application of \emph{important separators}, a well known technique in Parameterized Complexity. | cs.DS | cs |
A Brief Note on Single Source Fault Tolerant Reachability
Daniel Lokshtanov∗
Pranabendu Misra†
Saket Saurabh‡
Meirav Zehavi§
Abstract
Let G be a directed graph with n vertices and m edges, and let s ∈ V (G) be a designated
source vertex. We consider the problem of single source reachability (SSR) from s in presence
of failures of edges (or vertices). Formally, a spanning subgraph H of G is a k-Fault Tolerant
Reachability Subgraph (k-FTRS) if it has the following property. For any set F of at most
k edges (or vertices) in G, and for any vertex v ∈ V (G), the vertex v is reachable from s in
G − F if and only if it is reachable from s in H − F . Baswana et.al. [STOC 2016, SICOMP
2018] showed that in the setting above, for any positive integer k, we can compute a k-FTRS
with 2kn edges. In this paper, we give a much simpler algorithm for computing a k-FTRS,
and observe that it extends to higher connectivity as well. Our results follow from a simple
application of important separators, a well known technique in Parameterized Complexity.
1
Introduction
Fault tolerant data structures aim to capture properties of real world networks, which are often
prone to a small number of failures. Such data structures allow us to test various properties of
the network after failures have occurred, and the repairs are awaited. The problem is modeled
as a directed graph (digraph) G where a small number of edges (or vertices) have failed, and a
parameter k is used as a bound on the maximum number of failures that may occur at a time.
In this paper, we consider the problem of deciding the reachability of all vertices v ∈ V (G)
from a designated source vertex s ∈ V (G) upon the failure of any k edges (or vertices) in the
input graph G. Specifically, our objective is to construct a sparse spanning subgraph H of G
that preserves all reachability relationships from the source vertex s upon the failure of any k
edges (or vertices) in G. More formally, we seek a spanning subgraph H of G with the following
property: For any set F of at most k edges (or vertices) in G, and for any vertex v ∈ V (G),
there is a path from s to v in G − F if and only if there is a path from s to v in H − F . Such
a graph H is called a k-Fault Tolerant Reachability Subgraph (k-FTRS). Observe that, beyond
the question of deciding the reachability of a vertex v, the graph H may also be used to find an
alternate route from s to v, if one exists, upon the failure of the edges in F . Now, the problem
is formally defined as follows. Given as input a digraph G, a designated source vertex s ∈ V (G)
and an integer k, we must output a spanning subgraph H of G that is a k-FTRS.
Recently, Baswana et al. [3] presented an algorithm for computing a k-FTRS. Specifically,
their algorithm runs in time O(2kmn) for a digraph G of n vertices and m edges, and produces
a k-FTRS where the in-degree of any vertex is upper bounded by 2k. Their algorithm is based
on the notion of farthest min-cut that was introduced by Ford and Fulkerson [14]. They suggest
that their methods may be of independent interest in other problems. This is indeed so, for the
notion of important separators, which generalizes the notion of furthest cuts, is a well-known
technique in Parameterized Complexity [9].
∗University of California Santa Barbara, USA. [email protected]
†Max-Planck Institute for Informatics, Saarbrucken, Germany. [email protected]
‡Institute of Mathematical Sciences, Chennai, India. [email protected]
§Ben-Gurion University, Beersheba, Israel. [email protected]
1
The notion of important separators was introduced by Marx [20] to give an FPT algorithm
for Multiway Cut. Subsequently, important separators and techniques based on them have
been used to resolve the complexity of several important problems such as Directed Feedback
Vertex Set [6], Multicut [21], Directed Multiway Cut [7], Almost 2-SAT [24], Parity
Multiway Cut [18] and a linear-time FPT algorithm for Directed Feedback Vertex
Set [19]. Informally speaking, important separators capture the entire collection of furthest
cuts in a graph that have a bounded cardinality. We refer the reader to the textbook of Cygan
et al. [9] for an introduction to important separators, and more generally to the various tools
and techniques in Parameterized Complexity.
Using the notion of important separators, we give a very simple and conceptually appealing
algorithm for computing a k-FTRS. Indeed, we generalize the problem slightly. Given a digraph
G, a designated source vertex s ∈ V (G), an integer λ and an integer k, we output a spanning
subgraph H of G such that for any set F of at most k edges (or vertices) and for any vertex
v ∈ V (G), there are λ edge-disjoint paths from s to v in G − F if and only if there are λ such
paths in H − F . The graph H is called a (λ, k)-Fault Tolerant Reachability Subgraph ((λ, k)-
FTRS). As before, H may be used for both testing the λ-connectivity of a vertex v from s, as
well as for obtaining an alternate collection of λ edge-disjoint paths from s to v, if they exist,
after the failure of up to k edges (or vertices). In particular, we obtain the following theorem.
Theorem 1. There is a (λ, k)-FTRS where each vertex has in-degree at most (k + λ)4k+λ.
Further, this graph can be constructed in time O(4k+λ(k + λ)2(m + n)m).
Our bound and running time, obtained by a direct application of known results on important
separators, are slightly worse than those given by Baswana et al. [3]. Our intent here is to provide
a conceptual exposition of the results of Baswana et al [3] and further, introduce important
separators as an algorithmic tool for problems on fault tolerant networks, and more generally
for distributed network problems. Due to space constraints the discussion of other related works
has been postponed to the appendix.
Related Work. While a number of results are known about graph reachability for undirected
graphs starting from the work of Nagamochi and Ibaraki [22], and more generally in the dy-
namic model1 [12], relatively little is known in the case of digraphs. Patrascu and Thorup [23]
considered the problem of edge failures in undirected graphs. They constructed a data structure
that processes a bath of k edge failures in O(k log2 n log log n) time, and then answers connec-
tivity queries for any pair of vertices in O(log log n) time. Subsequent results have led to a
randomized data structure that uses almost-linear space (in the number of vertices) and cor-
rectly answers the queries with high probability [16]. For vertex failures in undirected graphs,
Duan and Pettie [11] gave a data structure that can process a batch of k vertex failures in
O(k2c+4 log2 n log log n) time and thereafter answer connectivity queries in O(k) time. Here c is
a parameter of the algorithm offering a tradeoff between the running time and the space used.
Recently, they improved this algorithm to process the k failures in O(k3 log3 n) time and with
the same query time as before, while using O(km log n) space [12, 13]. We refer to [12] for
further details.
In digraphs, other than [3], an optimal oracle for dual fault tolerant reachability was proposed
by Choudhary [8]. Furthermore, fault tolerance in digraphs for shortest paths [10, 4, 17, 5, 1]
and strongly connected components [15, 2] were also studied. The various problems in the fault
tolerant model for digraphs are subject to extensive ongoing research. We refer to [3, 8, 12, 2, 1]
for a detailed discussion and further details.
1Here we have a sequence of edge insertions and deletions, as well as reachability queries, and we must
efficiently update the data structure and answer the queries.
2
2 Preliminaries
Let G be a digraph on n vertices and m edges. For a subset of edges X ⊆ E(G), we let G − X
denote the subgraph of G with vertex set V (G) and edge set E(G) \ X. We omit the braces
when the set contains only a single edge, i.e. G− {e} is denoted by G− e. For an edge e = (u, v),
u is called the tail of e and v is called the head of v. We denote these vertices by tail(e) and
head(e), respectively. For R ⊆ V (G), δ−(R) denotes the set of in-coming edges to R, i.e. the set
of edges of G such that tail(e) ∈ V (G) \ R and head(e) ∈ R. Similarly, δ+(R) denotes the set
of out-going edges from R. Let G be a digraph and S and T be two disjoint subsets of V (G).
A (directed) (S, T )-cut is a subset X of edges of G such that there is no path from a vertex in
S to a vertex in T in G − X. Any minimal (S, T )-cut X can be expressed as δ+(R) = X where
S ⊆ R ⊆ V (G) \ T is the set of vertices that are reachable from some vertex of S in G − X.
Hence, for any (S, T )-cut X, let RX denote the set of vertices that are reachable from S in
G − X. The set RX is called the reachability set of X, and if X is minimal, then X = δ+(RX ).
An (S, T )-cut X is an important (S, T )-separator if there is no other (S, T )-cut X ′ such
that X ′ ≤ X and RX ⊆ RX ′. (Observe that if X is an important (S, T )-separator, then it
is a minimal (S, T )-cut.) Let Xk(S, T ) denote the collection of all important (S, T )-separators
in G of size at most k. The following observation follows immediately from the definition of
important separators.
Observation 1. Let Y be any (S, T )-cut of size at most k in G. Then, there is an important
(S, T )-separator X ∈ Xk(S, T ) such that X ≤ Y and RY ⊆ RX.
We have the following result on the collection Xk(S, T ).
Lemma 2 (Theorem 8.36 [9]). The cardinality of Xk(S, T ) is upper bounded by 4k. Furthermore,
Xk(S, T ) can be computed in time O(Xk(S, T )k2(m + n)).
In this paper, we describe the construction of a (λ, k)-FTRS with respect to edge failures
only, because any vertex failure can be modeled by an edge failure: Split every vertex v into an
edge (vin, vout), where the incoming and outgoing edges of v are respectively directed into vin
and directed out of vout [3].
3 A simple algorithm for (λ, k)-FTRS
Let us first consider the case where λ = 1, i.e. the construction of a k-FTRS. Let v ∈ V (G) \ {s}
be an arbitrary vertex, and let X(v) denote the collection of all important (s, v)-separators in
G of size at most k + 1. By Lemma 2, there are at most (k + 1)4k+1 edges in the union of all
such important separators. We have the following claim.
Lemma 3. Let e ∈ δ−(v) \ (SX∈X(v) X). Then, G − e is a k-FTRS of G.
Proof. Suppose not, and consider a set F of at most k edges and a vertex w ∈ V (G) such that
w is unreachable from s in (G − e) − F , but there is a path from s to w in G − F . Since
e ∈ δ−(v), it follows that v is unreachable from s in (G − e) − F , but there is a path from s to
v in G − F . Thus, Y = F ∪ {e} is an (s, v)-cut in G. Since G and G − e only differ on e, this
implies that e lies on every path from s to v in G − F . We may conclude the following: (a) Y
is an (s, v)-cut in G of size at most k + 1; (b) the vertex u := tail(e) belongs to the reachability
set of Y , i.e. u ∈ RY .
Now, consider the (s, v)-cut Y in the graph G and the collection X(v). Since e /∈ SX∈X(v) X,
i.e. e /∈ X for any X ∈ X(v), and e = (u, v), it follows that u /∈ RX for any X ∈ X(v). However,
u ∈ RY and Y is (s, v)-cut of cardinality k + 1 in G. This is a contradiction to Observation 1.
Hence, G − e is a k-FTRS of G.
3
The above lemma can be turned into an iterative algorithm that gradually bounds the in-
degree of each vertex in the graph. Let α = (k + 1)4k+1 denote the upper-bound on δ−(v) ∩
(SX∈X(v) X) for any vertex v ∈ V (G) \ {s}.
Algorithm 1 An algorithm to compute a k-FTRS of a digraph G with a source vertex s.
1: procedure FTRS(G, s, k)
2:
Delete all incoming edges of the source vertex s in G.
while there exists v ∈ V (G) \ {s} such that δ−(v) > α do
3:
4:
5:
Compute X(v) via Lemma 2.
Pick an edge e ∈ δ−(v) \ (SX∈X(v) X) and delete it.
end while
6:
7: end procedure
The correctness of the above algorithm follows from Lemma 3. Furthermore, it is clear that
this algorithm terminates once the in-degree of every vertex is upper bounded by α. Hence, it
runs in time O(4kk2m(m + n)). This gives us the following corollary.
Corollary 4. There is a k-FTRS where each vertex has in-degree at most (k + 1)4k+1. Further,
this graph can be constructed in time O(4kk2(m + n)m).
The following simple lemma extends the above construction (and also the construction of
Baswana et al. [3]) to any value of λ.
Lemma 5. Let H be a (k + λ − 1)-FTRS of G. Then, H is also a (k, λ)-FTRS of G.
Proof. Suppose not. Consider a vertex v ∈ V (G) \ {s} and a set of F of at most k edges such
that the following holds: there are λ edge-disjoint paths from s to v in G − F , but there is no
such collection of paths in H − F . Consider a minimum (s, v)-cut X in H − F , and observe that
X ≤ λ − 1. It follows that Y = F ∪ X is an (s, v)-cut in H of size at most k + λ − 1. However,
Y is not an (s, v)-cut in G. This is a contradiction to the fact that H is a (k + λ − 1)-FTRS of
G. Therefore, H must be a (λ, k)-FTRS of G.
The proof of Theorem 1 follows from Corollary 4 and Lemma 5.
References
[1] Surender Baswana, Keerti Choudhary, Moazzam Hussain, and Liam Roditty. Approximate
single source fault tolerant shortest path. In Proceedings of the Twenty-Ninth Annual ACM-
SIAM Symposium on Discrete Algorithms, SODA 2018, New Orleans, LA, USA, January
7-10, 2018, pages 1901 -- 1915, 2018.
[2] Surender Baswana, Keerti Choudhary, and Liam Roditty. An efficient strongly connected
components algorithm in the fault tolerant model. In 44th International Colloquium on
Automata, Languages, and Programming, ICALP 2017, July 10-14, 2017, Warsaw, Poland,
pages 72:1 -- 72:15, 2017.
[3] Surender Baswana, Keerti Choudhary, and Liam Roditty. Fault-tolerant subgraph for
single-source reachability: General and optimal. SIAM J. Comput., 47(1):80 -- 95, 2018.
[4] Davide Bil`o, Luciano Gual`a, Stefano Leucci, and Guido Proietti. Multiple-edge-fault-
In 33rd Symposium on Theoretical Aspects of
tolerant approximate shortest-path trees.
Computer Science, STACS 2016, February 17-20, 2016, Orl´eans, France, pages 18:1 -- 18:14,
2016.
4
[5] Greg Bodwin, Fabrizio Grandoni, Merav Parter, and Virginia Vassilevska Williams. Pre-
serving distances in very faulty graphs. In 44th International Colloquium on Automata,
Languages, and Programming, ICALP 2017, July 10-14, 2017, Warsaw, Poland, pages
73:1 -- 73:14, 2017.
[6] Jianer Chen, Yang Liu, Songjian Lu, Barry O'Sullivan, and Igor Razgon. A fixed-parameter
algorithm for the directed feedback vertex set problem. J. ACM, 55(5):21:1 -- 21:19, 2008.
[7] Rajesh Hemant Chitnis, MohammadTaghi Hajiaghayi, and D´aniel Marx. Fixed-parameter
tractability of directed multiway cut parameterized by the size of the cutset. In Proceedings
of the Twenty-Third Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2012,
Kyoto, Japan, January 17-19, 2012, pages 1713 -- 1725, 2012.
[8] Keerti Choudhary. An optimal dual fault tolerant reachability oracle. In 43rd International
Colloquium on Automata, Languages, and Programming, ICALP 2016, July 11-15, 2016,
Rome, Italy, pages 130:1 -- 130:13, 2016.
[9] Marek Cygan, Fedor V. Fomin, Lukasz Kowalik, Daniel Lokshtanov, D´aniel Marx, Marcin
Pilipczuk, Michal Pilipczuk, and Saket Saurabh. Parameterized Algorithms. Springer,
2015.
[10] Ran Duan and Seth Pettie. Dual-failure distance and connectivity oracles. In Proceedings
of the Twentieth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2009,
New York, NY, USA, January 4-6, 2009, pages 506 -- 515, 2009.
[11] Ran Duan and Seth Pettie. Connectivity oracles for failure prone graphs. In Proceedings
of the 42nd ACM Symposium on Theory of Computing, STOC 2010, Cambridge, Mas-
sachusetts, USA, 5-8 June 2010, pages 465 -- 474, 2010.
[12] Ran Duan and Seth Pettie. Connectivity oracles for graphs subject to vertex failures.
CoRR, abs/1607.06865, 2016.
[13] Ran Duan and Seth Pettie. Connectivity oracles for graphs subject to vertex failures. In
Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA 2017, Barcelona, Spain, Hotel Porta Fira, January 16-19, pages 490 -- 509, 2017.
[14] Lester Randolph Ford Jr and Delbert Ray Fulkerson. Flows in networks. Princeton uni-
versity press, 1962.
[15] Loukas Georgiadis, Giuseppe F. Italiano, and Nikos Parotsidis. Strong connectivity in
In Proceedings of the Twenty-Eighth
directed graphs under failures, with applications.
Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2017, Barcelona, Spain,
Hotel Porta Fira, January 16-19, pages 1880 -- 1899, 2017.
[16] David Gibb, Bruce M. Kapron, Valerie King, and Nolan Thorn. Dynamic graph connec-
tivity with improved worst case update time and sublinear space. CoRR, abs/1509.06464,
2015.
[17] Manoj Gupta and Shahbaz Khan. Multiple source dual fault tolerant BFS trees. In 44th
International Colloquium on Automata, Languages, and Programming, ICALP 2017, July
10-14, 2017, Warsaw, Poland, pages 127:1 -- 127:15, 2017.
[18] Daniel Lokshtanov and M. S. Ramanujan. Parameterized tractability of multiway cut
with parity constraints. In Automata, Languages, and Programming - 39th International
Colloquium, ICALP 2012, Warwick, UK, July 9-13, 2012, Proceedings, Part I, pages 750 --
761, 2012.
5
[19] Daniel Lokshtanov, M. S. Ramanujan, and Saket Saurabh. When recursion is better than
iteration: A linear-time algorithm for acyclicity with few error vertices. In Proceedings of
the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2018,
New Orleans, LA, USA, January 7-10, 2018, pages 1916 -- 1933, 2018.
[20] D´aniel Marx. Parameterized graph separation problems.
In Parameterized and Exact
Computation, First International Workshop, IWPEC 2004, Bergen, Norway, September
14-17, 2004, Proceedings, pages 71 -- 82, 2004.
[21] D´aniel Marx and Igor Razgon. Fixed-parameter tractability of multicut parameterized by
the size of the cutset. In Proceedings of the 43rd ACM Symposium on Theory of Computing,
STOC 2011, San Jose, CA, USA, 6-8 June 2011, pages 469 -- 478, 2011.
[22] Hiroshi Nagamochi and Toshihide Ibaraki. A linear-time algorithm for finding a sparsek-
connected spanning subgraph of ak-connected graph. Algorithmica, 7(1):583 -- 596, Jun 1992.
[23] Mihai Patrascu and Mikkel Thorup. Planning for fast connectivity updates. In 48th Annual
IEEE Symposium on Foundations of Computer Science (FOCS 2007), October 20-23, 2007,
Providence, RI, USA, Proceedings, pages 263 -- 271, 2007.
[24] Igor Razgon and Barry O'Sullivan. Almost 2-sat is fixed-parameter tractable. J. Comput.
Syst. Sci., 75(8):435 -- 450, 2009.
6
|
1603.03490 | 2 | 1603 | 2016-06-14T06:16:28 | A Unifying Formalism for Shortest Path Problems with Expensive Edge Evaluations via Lazy Best-First Search over Paths with Edge Selectors | [
"cs.DS",
"cs.RO"
] | While the shortest path problem has myriad applications, the computational efficiency of suitable algorithms depends intimately on the underlying problem domain. In this paper, we focus on domains where evaluating the edge weight function dominates algorithm running time. Inspired by approaches in robotic motion planning, we define and investigate the Lazy Shortest Path class of algorithms which is differentiated by the choice of an edge selector function. We show that several algorithms in the literature are equivalent to this lazy algorithm for appropriate choice of this selector. Further, we propose various novel selectors inspired by sampling and statistical mechanics, and find that these selectors outperform existing algorithms on a set of example problems. | cs.DS | cs | A Unifying Formalism for Shortest Path Problems with Expensive
Edge Evaluations via Lazy Best-First Search over Paths with Edge Selectors
Christopher M. Dellin and Siddhartha S. Srinivasa
The Robotics Institute, Carnegie Mellon University
{cdellin, siddh}@cs.cmu.edu
6
1
0
2
n
u
J
4
1
]
S
D
.
s
c
[
2
v
0
9
4
3
0
.
3
0
6
1
:
v
i
X
r
a
Abstract
While the shortest path problem has myriad applications, the
computational efficiency of suitable algorithms depends inti-
mately on the underlying problem domain. In this paper, we
focus on domains where evaluating the edge weight function
dominates algorithm running time. Inspired by approaches in
robotic motion planning, we define and investigate the Lazy
Shortest Path class of algorithms which is differentiated by
the choice of an edge selector function. We show that sev-
eral algorithms in the literature are equivalent to this lazy al-
gorithm for appropriate choice of this selector. Further, we
propose various novel selectors inspired by sampling and sta-
tistical mechanics, and find that these selectors outperform
existing algorithms on a set of example problems.
1
Introduction
Graphs provide a powerful abstraction capable of represent-
ing problems in a wide variety of domains from computer
networking to puzzle solving to robotic motion planning.
In particular, many important problems can be captured as
shortest path problems (Figure 1), wherein a path p∗ of min-
imal length is desired between two query vertices through a
graph G with respect to an edge weight function w.
Despite the expansive applicability of this single abstrac-
tion, there exist a wide variety of algorithms in the literature
for solving the shortest path problem efficiently. This is be-
cause the measure of computational efficiency, and therefore
the correct choice of algorithm, is inextricably tied to the un-
derlying problem domain.
The computational costs incurred by an algorithm can be
broadly categorized into three sources corresponding to the
blocks in Figure 1. One such source consists of queries on
the structure of the graph G itself. The most commonly dis-
cussed such operation, expanding a vertex (determining its
successors), is especially fundamental when the graph is rep-
resented implicitly, e.g. for domains with large graphs such
as the 15-puzzle or Rubik's cube. It is with respect to ver-
tex expansions that A* (Hart, Nilsson, and Raphael 1968) is
optimally efficient.
A second source of computational cost consists of main-
taining ordered data structures inside the algorithm it-
self, which is especially important for problems with large
Presented at ICAPS 2016, London. This extended version includes
proofs and timing results in the appendix.
Graph
G = (V, E)
Weight Function
w : E → [0, +∞]
Query u
Shortest Path
Algorithm
Path p∗
Figure 1: While solving a shortest path query, a shortest path
algorithm incurs computation cost from three sources: ex-
amining the structure of the graph G, evaluating the edge
weight function w, and maintaining internal data structures.
branching factors. For such domains, approaches such as
partial expansion (Yoshizumi, Miura, and Ishida 2000) or it-
erative deepening (Korf 1985) significantly reduce the num-
ber of vertices generated and stored by either selectively fil-
tering surplus vertices from the frontier, or by not storing the
frontier at all.
The third source of computational cost arises not from
reasoning over the structure of G, but instead from evaluat-
ing the edge weight function w (i.e. we treat discovering an
out-edge and determining its weight separately). Consider
for example the problem of articulated robotic motion plan-
ning using roadmap methods (Kavraki et al. 1996). While
these graphs are often quite small (fewer than 105 vertices),
determining the weight of each edge requires performing
many collision and distance computations for the complex
geometry of the robot and environment, resulting in plan-
ning times of multiple seconds to find a path.
In this paper, we consider problem domains in which eval-
uating the edge weight function w dominates algorithm run-
ning time and investigate the following research question:
How can we minimize the number of edges we need to
evaluate to answer shortest-path queries?
We make three primary contributions. First, inspired by
lazy collision checking techniques from robotic motion
planning (Bohlin and Kavraki 2000), we formulate a class of
shortest-path algorithms that is well-suited to problem do-
mains with expensive edge evaluations. Second, we show
that several existing algorithms in the literature can be ex-
pressed as special cases of this algorithm. Third, we show
that the extensibility afforded by the algorithm allows for
Eeval ← ∅
wlazy(e) ← west(e) ∀e ∈ E
loop
Algorithm 1 Lazy Shortest Path (LazySP)
1: function LAZYSHORTESTPATH(G, u, w, west)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
pcandidate ← SHORTESTPATH(G, u, wlazy)
if pcandidate ⊆ Eeval then
Eselected ← SELECTOR(G, pcandidate)
for e ∈ Eselected \ Eeval do
wlazy(e) ← w(e)
Eeval ← Eeval ∪ e
return pcandidate
(cid:46) Evaluate (expensive)
novel edge evaluation strategies, which can outperform ex-
isting algorithms over a set of example problems.
2 Lazy Shortest Path Algorithm
We describe a lazy approach to finding short paths which is
well-suited to domains with expensive edge evaluations.
Problem Definition
A path p in a graph G = (V, E) is composed of a sequence
of adjacent edges connecting two endpoint vertices. Given
an edge weight function w : E → [0, +∞], the length of the
path with respect to w is then:
(cid:88)
e∈p
len(p, w) =
w(e).
(1)
Given a single-pair planning query u : (vstart, vgoal) inducing
a set of satisfying paths Pu, the shortest-path problem is:
p∗ = arg min
p ∈ Pu
len(p, w).
(2)
A shortest-path algorithm computes p∗ given (G, u, w).
Many such algorithms have been proposed to efficiently ac-
commodate a wide array of underlying problem domains.
The well-known principle of best-first search (BFS) is com-
monly employed to select vertices for expansion so as to
minimize such expansions while guaranteeing optimality.
Since we seek to minimize edge evaluations, we apply BFS
to the question of selecting candidate paths in G for evalua-
tion. The resulting algorithm, Lazy Shortest Path (LazySP),
is presented in Algorithm 1, and can be applied to graphs
defined implicitly or explicitly.
The Algorithm
We track evaluated edges with the set Eeval. We are given
an estimator function west of the true edge weight w. This
estimator is inexpensive to compute (e.g. edge length or even
0). We then define a lazy weight function wlazy which returns
the true weight of an evaluated edge and otherwise uses the
inexpensive estimator west.
At each iteration of the search, the algorithm uses wlazy
to compute a candidate path pcandidate by calling an existing
solver SHORTESTPATH (note that this invocation requires no
evaluations of w). Once a candidate path has been found, it is
Algorithm 2 Various Simple LazySP Edge Selectors
1: function SELECTEXPAND(G, pcandidate)
efirst ← first unevaluated e ∈ pcandidate
2:
vfrontier ← G.source(efirst)
3:
Eselected ← G.out edges(vfrontier)
4:
return Eselected
5:
6: function SELECTFORWARD(G, pcandidate)
return {first unevaluated e ∈ pcandidate}
7:
8: function SELECTREVERSE(G, pcandidate)
return {last unevaluated e ∈ pcandidate}
9:
10: function SELECTALTERNATE(G, pcandidate)
if LazySP iteration number is odd then
11:
12:
13:
14:
15: function SELECTBISECTION(G, pcandidate)
16:
return {first unevaluated e ∈ pcandidate}
return {last unevaluated e ∈ pcandidate}
(cid:26) unevaluated e ∈ pcandidate
return
else
furthest from nearest evaluated edge
(cid:27)
returned if it is fully evaluated. Otherwise, an edge selector
is employed which selects graph edge(s) for evaluation. The
true weights of these edges are then evaluated (incurring the
requisite computational cost), and the algorithm repeats.
LazySP is complete and optimal:
Theorem 1 (Completeness of LazySP) If the graph G is
finite, SHORTESTPATH is complete, and the set Eselected re-
turned by SELECTOR returns at least one unevaluated edge
on pcandidate, then LAZYSHORTESTPATH is complete.
Theorem 2 (Optimality of LazySP) If west is chosen such
that west(e) ≤ w(e) for some parameter ≥ 1 and
LAZYSHORTESTPATH terminates with some path pret, then
len(pret, w) ≤ (cid:96)∗ with (cid:96)∗ the length of an optimal path.
The optimality of LazySP depends on the admissibility of
west in the same way that the optimality of A* depends on
the admissibility of its goal heuristic h. Theorem 2 estab-
lishes the general bounded suboptimality of LazySP w.r.t.
the inflation parameter . While our theoretical results (e.g.
equivalences) hold for any choice of , for clarity our ex-
amples and experimental results focus on cases with = 1.
Proofs are available in the appendix.
The Edge Selector: Key to Efficiency
The LazySP algorithm exhibits a rough similarity to optimal
replanning algorithms such as D* (Stentz 1994) which plan
a sequence of shortest paths for a mobile robot as new edge
weights are discovered during its traverse. D* treats edge
changes passively as an aspect of the problem setting (e.g. a
sensor with limited range).
The key difference is that our problem setting treats edge
evaluations as an active choice that can be exploited. While
any choice of edge selector that meets the conditions above
will lead to an algorithm that is complete and optimal, its
efficiency is dictated by the choice of this selector. This mo-
tivates the theoretical and empirical investigation of different
edge selectors in this paper.
(a) Expand[77]
(b) Forward[34]
(c) Reverse[24]
(d) Alternate[23]
(e) Bisection[25]
(f) WeightSamp[22]
(g) Partition[22]
Figure 2: Snapshots of the LazySP algorithm using each edge selector discussed in this paper on the same obstacle roadmap
graph problem, with start ( ) and goal ( ). At top, the algorithms after evaluating five edges (evaluated edges labeled as
valid
invalid). At middle, the final set of evaluated edges. At bottom, for each unique path considered from left to right, the
or
number of edges on the path that are
unevaluated.
The total number of edges evaluated is noted in brackets. Note that the scale on the Expand plot has been adjusted because the
selector evaluates many edges not on the candidate path at each iteration.
evaluated and invalid, and
already evaluated,
evaluated and valid,
Simple selectors. We codify five common strategies in Al-
gorithm 2. The Expand selector captures the edge weights
that are evaluated during a conventional vertex expansion.
The selector identifies the first unevaluated edge efirst on the
candidate path, and considers the source vertex of this edge
a frontier vertex. It then selects all out-edges of this frontier
vertex for evaluation. The Forward and Reverse selectors se-
lect the first and last unevaluated edge on the candidate path,
respectively (note that Forward returns a subset of Expand).
The Alternate selector simply alternates between Forward
and Reverse on each iteration. This can be motivated by
both bidirectional search algorithms as well as motion plan-
ning algorithms such as RRT-Connect (Kuffner and LaValle
2000) which tend to perform well w.r.t. state evaluations.
The Bisection selector chooses among those unevaluated
edges the one furthest from an evaluated edge on the candi-
date path. This selector is roughly analogous to the collision
checking strategy employed by the Lazy PRM (Bohlin and
Kavraki 2000) as applied to our problem on abstract graphs.
In the following section, we demonstrate that instances of
LazySP using simple selectors yield equivalent results to ex-
isting vertex algorithms. We then discuss two more sophis-
ticated selectors motivated by weight function sampling and
statistical mechanics.
3 Edge Equivalence to A* Variants
In the previous section, we introduced LazySP as the path-
selection analogue to BFS vertex-selection algorithms. In
this section, we make this analogy more precise. In partic-
ular, we show that LazySP-Expand is edge-equivalent to a
LazySP
Selector
Expand
Existing
Algorithm
(Weighted) A*
Forward
Lazy Weighted A*
Alternate Bidirectional Heuristic
Front-to-Front Algorithm
Result
Edge-equivalent
(Theorems 3, 4)
Edge-equivalent
(Theorems 5, 6)
Conjectured
Table 1: LazySP equivalence results. The A*, LWA*, and
BHFFA algorithms use reopening and the dynamic hlazy
heuristic (4).
variant of A* (and Weighted A*), and that LazySP-Forward
is edge-equivalent to a variant of Lazy Weighted A* (see
Table 1). It is important to be specific about the conditions
under which these equivalences arise, which we detail here.
Proofs are available in the appendix.
Edge equivalence. We say that two algorithms are edge-
equivalent if they evaluate the same edges in the same order.
We consider an algorithm to have evaluated an edge the first
time the edge's true weight is requested.
Arbitrary tiebreaking. For some graphs, an algorithm
may have multiple allowable choices at each iteration (e.g.
LazySP with multiple candidate shortest paths, or A* with
multiple vertices in OPEN with lowest f-value). We will
say that algorithm A is equivalent to algorithm B if for any
302010101010101010 Pcandidate
(LazySP)
Vfrontier
Scandidate
(A*)
Figure 3: Illustration of the equivalence between A* and
LazySP-Expand. After evaluating the same set of edges, the
next edges to be evaluated by each algorithm can both be
expressed as a surjective mapping onto a common set of un-
expanded frontier vertices.
choice available to A, there exists an allowable choice avail-
able to B such that the same edge(s) are evaluated by each.
A* with reopening. We show equivalence to variants of
A* and Lazy Weighted A* that do not use a CLOSED list to
prevent vertices from being visited more than once.
A* with a dynamic heuristic.
In order to apply A* and
Lazy Weighted A* to our problem, we need a goal heuristic
over vertices. The most simple may be
hest(v) = min
p:v→vg
len(p, west).
(3)
Note that the value of this heuristic could be computed as
a pre-processing step using Dijkstra's algorithm (Dijkstra
1959) before iterations begin. However, in order for the
equivalences to hold, we require the use of the lazy heuristic
hlazy(v) = min
p:v→vg
len(p, wlazy).
(4)
This heuristic is dynamic in that it depends on wlazy which
changes as edges are evaluated. Therefore, heuristic values
must be recomputed for all affected vertices on OPEN after
each iteration.
Equivalence to A*
We show that
the LazySP-Expand algorithm is edge-
equivalent to a variant of the A* shortest-path algorithm.
We make use of two invariants that are maintained during
the progression of A*.
Invariant 1 If v is discovered by A* and v(cid:48) is undiscovered,
with v(cid:48) a successor of v, then v is on OPEN.
Invariant 2 If v and v(cid:48) are discovered by A*, with v(cid:48) a suc-
cessor of v, and g[v] + w(v, v(cid:48)) < g[v(cid:48)], then v is on OPEN.
When we say a vertex is discovered, we mean that it is either
on OPEN or CLOSED. Note that Invariant 2 holds because
we allow vertices to be reopened; without reopening (and
with an inconsistent heuristic), later finding a cheaper path
to v (and not reopening v(cid:48)) would invalidate the invariant.
We will use the goal heuristic hlazy from (4). Note that if
an admissible edge weight estimator w exists (that is, w ≤
w), then our A* can approximate the Weighted A* algorithm
(Pohl 1970) with parameter by using west = w, and the
suboptimality bound from Theorem 2 holds.
Equivalence.
In order to show edge-equivalence, we con-
sider the case where both algorithms are beginning a new
iteration having so far evaluated the same set of edges.
LazySP-Expand has some set Pcandidate of allowable can-
didate paths minimizing len(p, wlazy); the Expand selector
will then identify a vertex on the chosen path for expansion.
A* will iteratively select a set of vertices from OPEN to
expand. Because it is possible that a vertex is expanded mul-
tiple times (and only the first expansion results in edge eval-
uations), we group iterations of A* into sequences, where
each sequence s consists of (a) zero or more vertices from
OPEN that have already been expanded, followed by (b) one
vertex from OPEN that is to be expanded for the first time.
We show that both the set of allowable candidate paths
Pcandidate available to LazySP-Expand and the set of allow-
able candidate vertex sequences Scandidate available to A*
map surjectively to the same set of unexpanded frontier ver-
tices Vfrontier as illustrated in Figure 3. This is established by
way of Theorems 3 and 4 below.
Theorem 3 If LazySP-Expand and A* have evaluated the
same set of edges, then for any candidate path pcandidate cho-
sen by LazySP yielding frontier vertex vfrontier, there exists an
allowable A* sequence scandidate which also yields vfrontier.
Theorem 4 If LazySP-Expand and A* have evaluated the
same set of edges, then for any candidate sequence scandidate
chosen by A* yielding frontier vertex vfrontier, there exists an
allowable LazySP path pcandidate which also yields vfrontier.
Equivalence to Lazy Weighted A*
In a conventional vertex expansion algorithm, determining
a successor's cost is a function of both the cost of the edge
and the value of the heuristic. If either of these components
is expensive to evaluate, an algorithm can defer its compu-
tation by maintaining the successor on the frontier with an
approximate cost until it is expanded. The Fast Downward
algorithm (Helmert 2006) is motivated by expensive heuris-
tic evaluations in planning, whereas the Lazy Weighted A*
(LWA*) algorithm (Cohen, Phillips, and Likhachev 2014) is
motivated by expensive edge evaluations in robotics.
We show that the LazySP-Forward algorithm is edge-
equivalent to a variant of the Lazy Weighted A* shortest-
path algorithm. For a given candidate path, the Forward se-
lector returns the first unevaluated edge.
Variant of Lazy Weighted A*. We reproduce a variant
of LWA* without a CLOSED list in Algorithm 3. For the
purposes of our analysis, the reproduction differs from the
original presentation, and we detail those differences here.
With the exception of the lack of CLOSED, the differences
do not affect the behavior of the algorithm.
The most obvious difference is that we present the orig-
inal OPEN list as separate vertex (Qv) and edge (Qe) pri-
ority queues, with sorting keys shown on lines 3 and 4. A
vertex v in the original OPEN with trueCost(v) = true
corresponds to a vertex v in Qv, whereas a vertex v(cid:48) in the
original OPEN with trueCost(v(cid:48)) = f alse (and parent v)
corresponds to an edge (v, v(cid:48)) in Qe. Use of the edge queue
if Qv.TopKey ≤ Qe.TopKey then
v ← Qv.Pop()
for v(cid:48) ∈ G.GetSuccessors(v) do
(cid:46) For uninitialized, g[v] = ∞
g[vstart] ← 0
Qv ← {vstart}
(cid:46) Key: g[v] + h(v)
Qe ← ∅
(cid:46) Key: g[v] + w(v, v(cid:48)) + h(v(cid:48))
while min(Qv.TopKey, Qe.TopKey) < g[vgoal] do
Algorithm 3 Lazy Weighted A* (without CLOSED list)
1: function LAZYWEIGHTEDA*(G, w, w, h)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
Qe.Insert((v, v(cid:48)))
(v, v(cid:48)) ← Qe.Pop()
if g[v(cid:48)] ≤ g[v] + w(v, v(cid:48)) then
gnew ← g[v] + w(v, v(cid:48))
if gnew < g[v(cid:48)] then
(cid:46) evaluate
else
continue
g[v(cid:48)] = gnew
Qv.Insert(v(cid:48))
obviates the need for duplicate vertices on OPEN with dif-
ferent parents and the conf (v) test for identifying such du-
plicates. This presentation also highlights the similarity be-
tween LWA* and the inner loop of the Batch Informed Trees
(BIT*) algorithm (Gammell, Srinivasa, and Barfoot 2015).
The second difference is that the edge usefulness test (line
12 of the original algorithm) has been moved from before
inserting into OPEN to after being popped from OPEN, but
before being evaluated (line 12 of Algorithm 3). This change
is partially in compensation for removing the CLOSED list.
This adjustment does not affect the edges evaluated.
We make use of an invariant that is maintained during the
progression of Lazy Weighted A*.
Invariant 3 For all vertex pairs v and v(cid:48), with v(cid:48) a succes-
sor of v, if g[v] + max(w(v, v(cid:48)), w(v, v(cid:48))) < g[v(cid:48)], then ei-
ther vertex v is on Qv or edge (v, v(cid:48)) is on Qe.
We will use h(v) = hlazy(v) from (4) and w = wlazy. Note
that the use of these dynamic heuristics requires that the Qv
and Qe be resorted after every edge is evaluated.
Equivalence. The equivalence follows similarly to that for
A* above. Given the same set of edges evaluated, the set of
allowable next evaluations is identical for each algorithm.
Theorem 5 If LazySP-Forward and LWA* have evaluated
the same set of edges, then for any allowable candidate path
pcandidate chosen by LazySP yielding first unevaluated edge
eab, there exists an allowable LWA* sequence scandidate which
also yields eab.
Theorem 6 If LazySP-Forward and LWA* have evaluated
the same set of edges, then for any allowable sequence of
vertices and edges scandidate considered by LWA* yielding
evaluated edge eab, there exists an allowable LazySP can-
didate path pcandidate which also yields eab.
Relation to Bidirectional Heuristic Search
LazySP-Alternate chooses unevaluated edges from either
the beginning or the end of the candidate path at each it-
known
edges
path
distribution
edge indicator
distributions
?
. . .
Figure 4: Illustration of maximum edge probability selec-
tors. A distribution over paths (usually conditioned on the
known edge evaluations) induces on each edge e a Bernoulli
distribution with parameter p(e) giving the probability that
it belongs to the path. The selector chooses the edge with the
largest such probability.
Algorithm 4 Maximum Edge Probability Selector
(for WeightSamp and Partition path distributions)
1: function SELECTMAXEDGEPROB(G, pcandidate,Dp)
p(e) ← Pr( e ∈ P ) for P ∼ Dp
2:
emax ← unevaluated e ∈ pcandidate maximizing p(e)
3:
return {emax}
4:
eration. We conjecture that an alternating version of the Ex-
pand selector is edge-equivalent to the Bidirectional Heuris-
tic Front-to-Front Algorithm (Sint and de Champeaux 1977)
for appropriate lazy vertex pair heuristic, and that LazySP-
Alternate is edge-equivalent to a bidirectional LWA*.
4 Novel Edge Selectors
Because we are conducting a search over paths, we are free
to implement selectors which are not constrained to evalu-
ate edges in any particular order (i.e. to maintain evaluated
trees rooted at the start and goal vertices). In this section, we
describe a novel class of edge selectors which is designed
to reduce the total number of edges evaluated during the
course of the LazySP algorithm. These selectors operate by
maintaining a distribution over the set of potential paths P
at each iteration of the algorithm (see Figure 4). Such a path
distribution induces a Bernoulli distribution for each edge
e which indicates its probability p(e) to lie on the potential
path; at each iteration, the selectors then choose the unevalu-
ated edge that maximizes this edge indicator probability (Al-
gorithm 4). The two selectors described in this section differ
with respect to how they maintain this distribution over po-
tential paths.
Weight Function Sampling Selector
The first selector, WeightSamp, is motivated by the intuition
that it is preferable to evaluate edges that are most likely
to lie on the true shortest path. Therefore, it computes its
path distribution Dp by performing shortest path queries on
sampled edge weight functions drawn from a distribution
Dw. This edge weight distribution is conditioned on the the
known weights of all previously evaluated edges Eeval:
Dp : SP(w) for w ∼ Dw(Eeval).
(5)
obstacle
distribution
weight fn
distribution
path
distribution
known
edges
. . .
. . .
. . .
Figure 5: The WeightSamp selector uses the path distribu-
tion induced by solving the shortest path problem on a dis-
tribution over possible edge weight functions Dw. In this ex-
ample, samples from Dw are computed by drawing samples
from DO, the distribution of obstacles that are consistent
with the known edge evaluations.
For example, the distribution Dw might consist of the
edge weights induced by a model of the distribution of en-
vironment obstacles (Figure 5). Since this obstacle distribu-
tion is conditioned on the results of known edge evaluations,
we consider the subset of worlds which are consistent with
the edges we have evaluated so far. However, depending on
the fidelity of this model, solving the corresponding short-
est path problem for a given sampled obstacle arrangement
might require as much computation as solving the origi-
nal problem, since it requires computing the resulting edge
weights. In practice, we can approximate Dw by assuming
that each edge is independently distributed.
Partition Function Selector
While the WeightSamp selector captures the intuition that it
is preferable to focus edge evaluations in areas that are use-
ful for many potential paths, the computational cost required
to calculate it at each iteration may render it intractable. One
candidate path distribution that is more efficient to compute
has a density function which follows an exponential form:
Dp : fP (p) ∝ exp(−β len(p, wlazy)).
(6)
In other words, we consider all potential paths P between
the start and goal vertices, with shorter paths assigned more
probability than longer ones (with positive parameter β).
We call this the Partition selector because this distribution
is closely related to calculating partition functions from sta-
tistical mechanics. The corresponding partition function is:
(7)
exp(−β len(p, wlazy)).
(cid:88)
p∈P
Z(P ) =
Note that the edge indicator probability required in Algo-
rithm 4 can then be written:
(a) Initial p(e) scores on a constant-weight grid with β: 50, 33, 28
(b) Initial p(e) scores with ∞-weight obstacles with β: 50, 33, 28
(c) Initial p(e) scores
(d) Scores after five evaluations
Figure 6: Examples of the Partition selector's p(e) edge
score function. (a) With no known obstacles, a high β as-
signs near-unity score to only edges on the shortest path; as
β decreases and more paths are considered, edges immedi-
ately adjacent to the roots score highest. (b) Since all paths
must pass through the narrow passage, edges within score
highly. (c) For a problem with two a-priori known obstacles
(dark gray), the score first prioritizes evaluations between
the two. (d) Upon finding these edges are blocked, the next
edges that are prioritized lie along the top of the world.
Here, P \ e denotes paths in P that do not contain edge e.
It may appear advantageous to restrict P to only sim-
ple paths, since all optimal paths are simple. Unfortunately,
an algorithm for computing (7) efficiently is not currently
known in this case. However, in the case that P consists of
all paths, there exists an efficient incremental calculation of
(7) via a recursive formulation which we detail here.
We use the notation Zxy = Z(Pxy), with Pxy the set of
paths from x to y. Suppose the values Zxy are known be-
tween all pairs of vertices x, y for a graph G. (For a graph
with no edges, Zxy is 1 if x = y and 0 otherwise.) Consider a
modified graph G(cid:48) with one additional edge eab with weight
wab. All additional paths use the new edge eab a non-zero
number of times; the value Z(cid:48)
xy can be shown to be
p(e) = 1 − Z(P \ e)
Z(P )
.
(8)
Z(cid:48)
xy = Zxy +
ZxaZby
exp(βwab) − Zba
if exp(βwab) > Zba. (9)
Figure 7: Visualization of the first of three articulated motion planning problems in which the HERB robot must move its right
arm from the start configuration (pictured) to any of seven grasp configurations for a mug. Shown is the progression of the
Alternate selector on one of the randomly generated roadmaps; approximately 2% of the 7D roadmap is shown in gray by
projecting onto the space of end-effector positions.
This form is derived from simplifying the induced geomet-
ric series; note that if exp(βwab) ≤ Zba, the value Z(cid:48)
xy is
infinite. One can also derive the inverse: given values Z(cid:48),
calculate the values Z if an edge were removed.
This incremental formulation of (7) allows for the corre-
sponding score p(e) for edges to be updated efficiently dur-
ing each iteration of LazySP as the wlazy value for edges
chosen for evaluation are updated. In fact, if the values Z
are stored in a square matrix, the update for all pairs after an
edge weight change consists of a single vector outer product.
5 Experiments
We compared the seven edge selectors on three classes of
shortest path problems. The average number of edges evalu-
ated by each, as well as timing results from our implementa-
tions, are shown in Figure 8. In each case, the estimate was
chosen so that west ≤ w, so that all runs produced optimal
paths. The experimental results serve primarily to illustrate
that the A* and LWA* algorithms (i.e. Expand and Forward)
are not optimally edge-efficient, but they also expose dif-
ferences in behavior and prompt future research directions.
All experiments were conducted using an open-source im-
plementation.1 Motion planning results were implemented
using OMPL (S¸ucan, Moll, and Kavraki 2012).
Random partially-connected graphs. We tested on a set
of 1000 randomly-generated undirected graphs with V =
100, with each pair of vertices sharing an edge with prob-
ability 0.05. Edges have an independent 0.5 probability of
having infinite weight, else the weight is uniformly dis-
tributed on [1, 2]; the estimated weight was unity for all
edges. For the WeightSamp selector, we drew 1000 w sam-
ples at each iteration from the above edge weight distribu-
tion. For the Partition selector, we used β = 2.
Roadmap graphs on the unit square. We considered
roadmap graphs formed via the first 100 points of the (2, 3)-
Halton sequence on the unit square with a connection radius
of 0.15, with 30 pairs of start and goal vertices chosen ran-
domly. The edge weight function was derived from 30 sam-
pled obstacle fields consisting of 10 randomly placed axis-
1https://github.com/personalrobotics/lemur
aligned boxes with dimensions uniform on [0.1, 0.3], with
each edge having infinite weight on collision, and weight
equal to its Euclidean length otherwise. One of the resulting
900 example problems is shown in Figure 2. For the Weight-
Samp selector, we drew 1000 w samples with a naıve edge
weight distribution with each having an independent 0.1 col-
lision probability. For the Partition selector, we used β = 21.
Roadmap graphs for robot arm motion planning. We
considered roadmap graphs in the configuration space corre-
sponding to the 7-DOF right arm of the HERB home robot
(Srinivasa et al. 2012) across three motion planning prob-
lems inspired by a table clearing scenario (see Figure 7). The
problems consisted of first moving from the robot's home
configuration to one of 7 feasible grasp configurations for
a mug (pictured), second transferring the mug to one of 72
feasible configurations with the mug above the blue bin, and
third returning to the home configuration. Each problem was
solved independently. This common scenario spans various
numbers of starts/goals and allows a comparison w.r.t. diffi-
culty at different problem stages as discussed later.
For each problem, 50 random graphs were constructed by
applying a random offset to the 7D Halton sequence with
N = 1000, with additional vertices for each problem start
and goal configuration. We used an edge connection radius
of 3 radians, resulting E ranging from 23404 to 28109.
Each edge took infinite weight on collision, and weight
equal to its Euclidean length otherwise. For the WeightSamp
selector, we drew 1000 w samples with a naıve edge weight
distribution in which each edge had an independent 0.1 prob-
ability of collision. For the Partition selector, we used β = 3.
6 Discussion
The first observation that is evident from the experimen-
tal results is that lazy evaluation -- whether using Forward
(LWA*) or one of the other selectors -- grossly outperforms
Expand (A*). The relative penalty that Expand incurs by
evaluating all edges from each expanded vertex is a func-
tion of the graph's branching factor.
Since the Forward and Reverse selectors are simply mir-
rors of each other, they exhibit similar performance averaged
across the PartConn and UnitSquare problem classes, which
F
R
B
A
PartConn
online†(ms)
sel (ms)
UnitSquare
online†(ms)
sel (ms)
P‡
W
E
20.66 20.39
87.10 35.86 34.84 22.23 44.81
1.22 1.96 1.86 1.20 2.41 4807.19 3.32
0.02 0.01 0.01 0.01 0.03 4805.64 2.07
15.58 14.08
69.21 27.29 27.69 17.82 32.62
0.91 1.47 1.49 0.94 1.71 3864.95 1.72
0.01 0.01 0.01 0.01 0.02 3863.49 0.87
56.93 48.07
269.82 5.90 8.22 5.96 7.34 3402.21 5.80
0.00 0.00 0.00 0.00 0.00 3392.76 1.54
269.78 5.87 8.20 5.94 7.31
9.39 4.21
344.74 49.72 95.58 59.44 58.90
73.72 50.66
109.09 4.81 14.81 7.03 7.91 3375.35 7.25
0.00 0.00 0.00 0.00 0.00 3358.82 1.61
16.47 5.59
109.07 4.78 14.77 7.01 7.88
657.02 62.24 98.54 69.96 75.88
66.24 62.16
166.19 3.27 7.36 5.95 5.63 4758.04 5.99
0.00 0.00 0.00 0.00 0.00 4750.16 2.03
7.82 3.91
166.17 3.26 7.34 5.93 5.61
1845.38 78.90 30.70 37.04 69.26
30.82 31.38
534.16 9.61 2.50 4.91 8.47 2073.23 4.17
0.00 0.00 0.00 0.00 0.00 2069.29 0.98
3.90 3.15
ArmPlan(avg) 949.05 63.62 74.94 55.48 68.01
online (s)
sel (s)
eval (s)
ArmPlan1
online (s)
sel (s)
eval (s)
ArmPlan2
online (s)
sel (s)
eval (s)
ArmPlan3
online (s)
sel (s)
eval (s)
534.10 9.58 2.48 4.89 8.44
(a) Average number of edges evaluated for each problem class and
selector. The minimum selector, along with any selector within one
unit of its standard error, is shown in bold. The ArmPlan class is
split into its three constituent problems. Online timing results are
also shown, including the components from the invoking the selec-
tor and evaluating edges. †PartConn and UnitSquare involve trivial
edge evaluation time. ‡Timing for the Partition selector does not
include pre-computation time.
80
60
40
↑
80
60
40
100
80
60
E F R A B W P
E F R A B W P
E F R A B W P
(b) PartConn
(c) UnitSquare
(d) ArmPlan
Figure 8: Experimental results for the three problem classes
across each of the seven selectors, E:Expand, F:Forward,
R:Reverse, A:Alternate, B:Bisection, W:WeightSamp, and
P:Partition. In addition to the summary table (a), the plots
(b-d) show summary statistics for each problem class. The
means and standard errors in (b-c) are across the 1000 and
900 problem instances, respectively. The means and stan-
dard errors in (d) are for the average across the three con-
stituent problems for each of the 50 sampled roadmaps.
are symmetric. However, this need not the case for a particu-
lar instance. For example, the start of ArmPlan1 and the goal
of ArmPlan3 consist of the arm's single home configuration
in a relatively confined space. As shown in the table in Fig-
ure 8a, it appears that the better selector on these problems
attempts to solve the more constrained side of the problem
first. While it may be difficult to determine a priori which
part of the problem will be the most constrained, the simple
Alternate selector's respectable performance suggests that it
may be a reasonable compromise.
The per-path plots at the bottom of Figure 2 allow us to
characterize the selectors' behavior. For example, Alternate
often evaluates several edges on each path before finding an
obstacle. Its early evaluations also tend to be useful later,
and it terminates after considering 10 paths on the illustrated
problem. In contrast, Bisection exhibits a fail-fast strategy,
quickly invalidating most paths after a single evaluation, but
needing 16 such paths (with very little reuse) before it termi-
nates. In general, the Bisection selector did not outperform
any of the lazy selectors in terms of number of edges evalu-
ated. However, it may be well suited to problem domains in
which evaluations that fail tend be less costly.
The novel selectors based on path distributions tend to
minimize edge evaluations on the problems we considered.
While the WeightSamp selector performs similarly to Par-
tition on the simpler problems, it performs less well in the
ArmPlan domain. This may be because many more samples
are needed to approximate the requisite path distribution.
The path distribution selectors are motivated by focusing
evaluation effort in areas that are useful for many distinct
candidate paths, as illustrated in Figure 6. Note that in the
absence of a priori knowledge, the edges nearest to the start
and goal tend to have the highest p(e) score, since they are
members of many potential paths. Because it tends to focus
evaluations in a similar way, the Alternate selector may serve
as a simple proxy for the more complex selectors.
We note that an optimal edge selector could be theoret-
ically achieved by posing the edge selection problem as a
POMDP, given a probabilistic model of the true costs. While
likely intractable in complex domains, exploring this solu-
tion may yield useful approximations or insights.
Timing results. Figure 8a shows that the five simple se-
lectors incur a negligible proportion of the algorithm's run-
time. The WeightSamp and Partition selectors both require
additional time (especially the former) in order to reduce
the time spent evaluating edges. This tradeoff depends in-
timately on the problem domain considered. In the ArmPlan
problem class, the Partition selector was able to reduce aver-
age total online runtime slightly despite an additional 1.54s
of selector time. Note that Partition requires an expensive
computation of the graph's initial Z-values, which are in-
dependent of the true weights and start/goal vertices (and
can therefore be pre-computed, e.g. across all ArmPlan in-
stances). Full timing results are available in the appendix
(Figure 9).
Optimizations. While we have focused on edge evalua-
tions as the dominant source of computational cost, other
considerations may also be important. There are a number
of optimizations that allow for efficient implementation of
LazySP.
The first relates to the repeated invocations of the inner
shortest path algorithm (line 5 of Algorithm 1). Because
only a small number of edges change weights between in-
vocations, an incremental search algorithm such as SSSP
(Ramalingam and Reps 1996) or LPA* (Koenig, Likhachev,
and Furcy 2004) can be used to greatly improve the speed of
the inner searches. Since the edge selector determines where
on the graph edges are evaluated, the choices of the selector
and the inner search algorithm are related. For example, us-
ing the Forward selector with an incremental inner search
rooted at the goal results in behavior similar to D* (Stentz
1994) (albeit without the need to handle a moving start lo-
cation) since a large portion of the inner tree can be reused.
An optimization commonly applied to vertex searches
called immediate expansion is also applicable to LazySP. If
an edge is evaluated with weight w ≤ west, the inner search
need not be run again before the next edge is evaluated.
A third optimization is applicable to domains with infi-
nite edge costs (e.g. to represent infeasible transitions). If
the length of the path returned by the inner shortest path al-
gorithm is infinite, LazySP can return this path immediately
even if some of its edges remain unevaluated without af-
fecting its (sub)optimality. This reduces the number of edge
evaluations needed in the case that no feasible path exists.
Other methods for expensive edge evaluations. An al-
ternative to lazy evaluations is based on the observation that
when solved by vertex expansion algorithms, such problems
are characterized by slow vertex expansions. To mitigate
this, approaches such as Parallel A* (Irani and Shih 1986)
and Parallel A* for Slow Expansions (Phillips, Koenig, and
Likhachev 2014) aim to parallelize such expansions. We be-
lieve that a similar approach can be applied to LazySP.
Another approach to finding short paths quickly is to re-
lax the optimization objective (2) itself. While LazySP al-
ready supports a bounded-suboptimal objective via an in-
flated edge weight estimate (Theorem 2), it may also be pos-
sible to adapt the algorithm to address bounded-cost prob-
lems (Stern, Puzis, and Felner 2011).
Acknowledgements
We would like to thank Aaron Johnson and Michael Ko-
val for their comments on drafts of this work. This work
was (partially) funded by the National Science Foundation
IIS (#1409003), Toyota Motor Engineering & Manufactur-
ing (TEMA), and the Office of Naval Research.
References
Bohlin, R., and Kavraki, E. 2000. Path planning using Lazy
PRM. In IEEE International Conference on Robotics and
Automation, volume 1, 521 -- 528.
Cohen, B.; Phillips, M.; and Likhachev, M. 2014. Planning
In Robotics:
single-arm manipulations with n-arm robots.
Science and Systems.
Dijkstra, E. W. 1959. A note on two problems in connexion
with graphs. Numerishe Mathematik 1(1):269 -- 271.
Gammell, J.; Srinivasa, S.; and Barfoot, T. 2015. Batch In-
formed Trees (BIT*): Sampling-based optimal planning via
the heuristically guided search of implicit random geometric
graphs. In IEEE International Conference on Robotics and
Automation, 3067 -- 3074.
Hart, P.; Nilsson, N.; and Raphael, B. 1968. A formal basis
for the heuristic determination of minimum cost paths. IEEE
Transactions on Systems Science and Cybernetics 4(2):100 --
107.
Helmert, M. 2006. The fast downward planning system.
Artificial Intelligence Research 26:191 -- 246.
Irani, K. B., and Shih, Y. 1986. Parallel A* and AO* al-
gorithms: An optimality criterion and performance evalua-
In International Conference on Parallel Processing,
tion.
274 -- 277.
Kavraki, L.; Svestka, P.; Latombe, J.-C.; and Overmars, M.
1996. Probabilistic roadmaps for path planning in high-
IEEE Transactions on
dimensional configuration spaces.
Robotics and Automation 12(4):566 -- 580.
Koenig, S.; Likhachev, M.; and Furcy, D. 2004. Lifelong
planning A*. Artificial Intelligence 155(1 -- 2):93 -- 146.
Korf, R. E. 1985. Depth-first iterative-deepening: An op-
timal admissible tree search. Artificial Intelligence 27:97 --
109.
Kuffner, J., and LaValle, S. 2000. RRT-Connect: An effi-
cient approach to single-query path planning. In IEEE Inter-
national Conference on Robotics and Automation, volume 2,
995 -- 1001.
Phillips, M.; Koenig, S.; and Likhachev, M. 2014. Paral-
lel A* for planning with time-consuming state expansions.
In International Conference on Automated Planning and
Scheduling.
Pohl, I. 1970. Heuristic search viewed as path finding in a
graph. Artificial Intelligence 1(34):193 -- 204.
Ramalingam, G., and Reps, T. 1996. An incremental al-
gorithm for a generalization of the shortest-path problem.
Journal of Algorithms 21(2):267 -- 305.
Sint, L., and de Champeaux, D. 1977. An improved bidirec-
tional heuristic search algorithm. J. ACM 24(2):177 -- 191.
Srinivasa, S.; Berenson, D.; Cakmak, M.; Collet, A.; Dogar,
M.; Dragan, A.; Knepper, R.; Niemueller, T.; Strabala, K.;
Vande Weghe, M.; and Ziegler, J. 2012. HERB 2.0: Lessons
learned from developing a mobile manipulator for the home.
Proceedings of the IEEE 100(8):2410 -- 2428.
Stentz, A. 1994. Optimal and efficient path planning for
partially-known environments. In IEEE International Con-
ference on Robotics and Automation, volume 4, 3310 -- 3317.
Stern, R.; Puzis, R.; and Felner, A. 2011. Potential search: A
bounded-cost search algorithm. In International Conference
on Automated Planning and Scheduling.
S¸ucan, I. A.; Moll, M.; and Kavraki, L. E. 2012. The Open
IEEE Robotics & Automation
Motion Planning Library.
Magazine 19(4):72 -- 82. http://ompl.kavrakilab.org.
Yoshizumi, T.; Miura, T.; and Ishida, T. 2000. A* with
partial expansion for large branching factor problems.
In
AAAI Conference on Artificial Intelligence, 923 -- 929.
A Appendix: Proofs
LazySP
Proof of Theorem 2 Let p∗ be an optimal path w.r.t. w,
with (cid:96)∗ = len(p∗, w). Since west(e) ≤ w(e) and ≥ 1,
it follows that regardless of which edges are stored in Weval,
wlazy(e) ≤ w(e), and therefore len(p∗, wlazy) ≤ (cid:96)∗. Now,
since the inner SHORTESTPATH algorithm terminated with
pret, we know that len(pret, wlazy) ≤ len(p∗, wlazy). Further,
since the algorithm terminated with pret, each edge on pret
has been evaluated; therefore, len(pret, w) = len(pret, wlazy).
Therefore, len(pret, w) ≤ (cid:96)∗.
(cid:3)
Proof of Theorem 1 In this case, the algorithm will evalu-
ate at least unevaluated edge at each iteration. Since there
are a finite number of edges, eventually the algorithm will
(cid:3)
terminate.
A* Equivalence
Proof of Invariant 1 If v is discovered, then it must either
be on OPEN or CLOSED. v can be on CLOSED only after
it has been expanded, in which case v(cid:48) would be discovered
(cid:3)
(which it is not). Therefore, v must be on OPEN.
Proof of Invariant 2 Clearly the invariant holds at the be-
ginning of the algorithm, with only vstart on OPEN. If the
invariant were to no longer hold after some iteration, then
there must exist some pair of discovered vertices v and v(cid:48)
with v on CLOSED and g[v]+w(v, v(cid:48)) < g[v(cid:48)]. Since v is on
CLOSED, it must have been expanded at some previous iter-
ation, immediately after which the inequality could not have
held because g[v(cid:48)] is updated upon expansion of v. There-
fore, the inequality must have newly held after some inter-
vening iteration, with v remaining on CLOSED. Since the
values g are monotonically non-increasing and w is fixed,
this implies that g[v] must have been updated (lower). How-
ever, if this had happened, then v would have been removed
from CLOSED and placed on OPEN. This contradiction im-
(cid:3)
plies that the invariant holds at every iteration.
Proof of Theorem 3 Consider path p∗
lazy with length (cid:96)∗
lazy
yielding frontier vertex vfrontier via SELECTEXPAND. Con-
struct a vertex sequence s as follows. Initialize s with the
vertices on p∗
lazy from vstart to vfrontier, inclusive. Let N be the
number of consecutive vertices at the start of s for which
f (v) = (cid:96)∗
lazy, vstart, must
lazy, so N ≥ 1.) Remove from the start of s
have f (vstart) = (cid:96)∗
the first N − 1 vertices. Note that at most the first vertex on
s has f (v) = (cid:96)∗
lazy, and the last vertex on s must be vfrontier.
Now we show that each vertex in this sequence s, consid-
ered by A* in turn, exists on OPEN with minimal f-value.
Iteratively consider the following procedure for sequence s.
Throughout, we know that there must not be any vertex with
f (v) < (cid:96)∗
lazy; that would imply that a different path through
vb shorter than (cid:96)∗
lazy could not have
been chosen.
If the sequence has length > 1, then consider the first two
vertices on s, va and vb. By construction, f (va) = (cid:96)∗
lazy and
f (vb) (cid:54)= (cid:96)∗
lazy. In fact, from above we know that f (vb) >
(cid:96)∗
lazy. Therefore, we have that f (va) < f (vb), therefore and
lazy (Note that the first vertex on p∗
lazy exists, in which case p∗
g[va] + w(va, vb) < g[vb]. By Invariant 2, va must be on
OPEN, and with f (va) = (cid:96)∗
lazy, it can therefore be consid-
ered by A*. After it is expanded, f (vb) = (cid:96)∗
lazy, and we can
repeat the above procedure with the sequence formed by re-
moving the va from s.
If instead the sequence has length 1, then it must be ex-
actly (vfrontier), with f (vfrontier) = (cid:96)∗
lazy. Since the edge after
f (vfrontier) is not evaluated, then by Invariant 1, vfrontier must
(cid:3)
be on OPEN, and will therefore be expanded next.
Proof of Theorem 4 Given that all vertices in scandidate be-
sides the last are re-expansions, they can be expanded with
no edge evaluations. Once the last vertex, vfrontier, is to be
expanded by A*, suppose it has f-value (cid:96).
First, we will show that there exists a path with length (cid:96)
w.r.t. wlazy wherein all edges before vfrontier have been eval-
uated, and the first edge after vfrontier has not. Let pa be a
shortest path from vstart to vfrontier consisting of only evalu-
ated edges. The length of this pa must be equal to g[vfrontier];
if it were not, there would be some previous vertex on pa
with lower f-value than vfrontier, which would necessarily
have been expanded first. Let pb be the a shortest path from
vfrontier to vgoal. The length of pb must be hlazy(vfrontier) by
definition. Therefore, the path (pa, pb) must have length (cid:96),
and since vfrontier is a new expansion, the first edge on pb
must be unevaluated.
Second, we will show that there does not exist any path
shorter than (cid:96) w.r.t. wlazy. Suppose p(cid:48) were such a path, with
length (cid:96)(cid:48) < (cid:96). Clearly, vstart would have f-value (cid:96)(cid:48) (although
it may not be on OPEN). Consider each pair of vertices
(va, vb) along p(cid:48) in turn. In each case, if vb were either undis-
covered, or if g[va]+w(va, vb) < g[vb], then va would be on
OPEN (via Invariants 1 and 2, respectively) with f (va) = (cid:96)(cid:48),
and would therefore have been expanded before vfrontier. Oth-
erwise, we know that f (vb) = (cid:96)(cid:48), and we can continue to the
(cid:3)
next pair on p(cid:48).
LWA* Equivalence
Proof of Invariant 3 Clearly the invariant holds at the be-
ginning of the algorithm with only g[vstart] = 0, since the
inequality holds only for the out-edges of vstart, with vstart
on Qv. Consider each subsequent iteration. If a vertex v is
popped from Qv, then this may invalidate the invariant for
all successors of v; however, since all out-edges are immedi-
ately added to Qe, the invariant must hold. Consider instead
if an edge (v, v(cid:48)) which satisfies the inequality is popped
from Qe. Due to the inequality, we know that g[v(cid:48)] will be
recalculated as g[v(cid:48)] = g[v] + w(v, v(cid:48)), so that the inequality
is no longer satisfied for edge (v, v(cid:48)). However, reducing the
value g[v(cid:48)] may introduce satisfied inequalities across sub-
sequent out-edges of v(cid:48), but since v(cid:48) is added to Qv, the in-
(cid:3)
variant continues to hold.
Proof of Theorem 5 In the first component of the equiva-
lence, we will show that for any path p minimizing wlazy
allowable to LazySP-Forward, with (va, vb) the first uneval-
uated edge on p, there exists a sequence of vertices and edges
on Qv and Qe allowable to LWA* such that edge (va, vb) is
the first to be newly evaluated. Let the length of p w.r.t. wlazy
be (cid:96).'
We will first show that no vertex on Qv or edge on Qe
can have f (·) < (cid:96). Suppose such a vertex v, or edge e with
source vertex v, exists. Then g[v] + h(v) < (cid:96), and there
must be some path p(cid:48) consisting of an evaluated segment
from vstart to v of length g[v], followed a segment from v
to vgoal of length h(v). But then this path should have been
chosen by LazySP.
Next, we will show a procedure for generating an allow-
able sequence for LWA*. We will iteratively consider a se-
quence of path segments, starting with the segment from
vstart to va, and becoming progressively shorter at each it-
eration by removing the first vertex and edge on the path.
We will show that the first vertex on each segment vf has
g[vf ] = (cid:96)− h(vf ). By definition, this is true of the first such
segment, since g[vstart] = 0. For each but the last such seg-
ment, consider the first edge, (vf , vs). If vs has the correct
g[·], we can continue to the next segment immediately. Oth-
erwise, either vf is on Qv or (vf , vs) is on Qe by Invariant 3.
If the former is true, then vf can be popped from Qv with
f = (cid:96), thereby adding (vf , vs) to Qe. Then, (vf , vs) can be
popped from Qe with f = (cid:96), resulting in g[vs] = (cid:96) − h(vs).
We can then move on to the next segment.
At the end of this process, we have the trivial segment
(va), with g[va] = (cid:96) − h(va). If va is on Qv, then pop it
(with f (va) = (cid:96)), placing eab on Qe; otherwise, since eab is
unevaluated, it must already be on Qe. Since f (eab) = (cid:96), we
(cid:3)
can pop and evaluate it.
Proof of Theorem 6 Given that all vertices in scandidate en-
tail no edge evaluations, and all edges therein are re-
expansions, they can be considered with no edge evalua-
tions. Once the last edge exy is to be expanded by LWA*,
suppose it has f-value (cid:96).
First, we will show that there exists a path with length (cid:96)
w.r.t. wlazy which traverses unevaluated edge exy wherein all
edges before vx have been evaluated. Let px be a shortest
path segment from vstart to vx consisting of only evaluated
edges. The length of px must be equal to g[vx]; if it were not,
there would be some previous vertex on px with lower f-
value than vx, which would necessarily have been expanded
first. Let py be the a shortest path from vy to vgoal. The length
of py must be hlazy(vy) by definition. Therefore, the path
(px, exy, py) must have length (cid:96).
Second, we will show that there does not exist any path
shorter than (cid:96) w.r.t. wlazy. Suppose p(cid:48) were such a path, with
length (cid:96)(cid:48) < (cid:96), and with first unevaluated edge e(cid:48)
xy. Clearly,
vstart has g[vstart] = (cid:96)(cid:48) − h(vstart). Consider each evaluated
b] (cid:54)= (cid:96)(cid:48) − h(v(cid:48)
edge e(cid:48)
b),
ab would be on Qv or Qe with f (·) = (cid:96)(cid:48),
then either v(cid:48)
and would therefore be expanded before exy. Therefore, e(cid:48)
would then be popped from Qe with f (e(cid:48)
xy) = (cid:96)(cid:48), and it
would have been evaluated before exy with f (exy) = (cid:96). (cid:3)
ab along p(cid:48) in turn. In each case, if g[v(cid:48)
a or e(cid:48)
xy
B Appendix: Timing Results
We include an accounting of the cumulative computation
time taken by each component of LazySP for each of the
seven selectors for each problem class (Figure 9).
PartConn
total (ms)
sel-init (ms)
online (ms)
search (ms)
sel (ms)
eval (ms)
eval (edges)
UnitSquare
total (ms)
sel-init (ms)
online (ms)
search (ms)
sel (ms)
eval (ms)
eval (edges)
ArmPlan (avg)
total (s)
sel-init (s)
online (s)
search (s)
sel (s)
eval (s)
eval (edges)
ArmPlan1
total (s)
sel-init (s)
online (s)
search (s)
sel (s)
eval (s)
eval (edges)
ArmPlan2
total (s)
sel-init (s)
online (s)
search (s)
sel (s)
eval (s)
eval (edges)
ArmPlan3
total (s)
sel-init (s)
online (s)
search (s)
sel (s)
eval (s)
eval (edges)
Expand
Forward
Reverse
Alternate
Bisect
WeightSamp
Partition
1.22 ± 0.04
--
1.22 ± 0.04
0.48 ± 0.01
0.02 ± 0.00
--
87.10 ± 2.39 35.86 ± 1.04 34.84 ± 1.04 22.23 ±0.60 44.81 ±1.11
2.41 ±0.06 4807.19 ±135.22
--
--
2.41 ±0.06 4807.19 ±135.22
1.38 ±0.04
0.70 ± 0.02
0.03 ±0.00 4805.64 ±135.18
--
20.66 ± 0.57
1.86 ± 0.06
--
1.86 ± 0.06
1.05 ± 0.03
0.01 ± 0.00
--
1.96 ± 0.06
--
1.96 ± 0.06
1.12 ± 0.03
0.01 ± 0.00
--
1.20 ±0.03
--
1.20 ±0.03
0.68 ±0.02
0.01 ±0.00
--
--
0.91 ± 0.03
--
0.91 ± 0.03
0.35 ± 0.01
0.01 ± 0.00
--
69.21 ± 2.55 27.29 ± 1.03 27.69 ± 1.02 17.82 ±0.60 32.62 ±0.72
1.71 ±0.04 3864.95 ±117.66
--
--
1.71 ±0.04 3864.95 ±117.66
0.92 ±0.02
0.75 ± 0.02
0.02 ±0.00 3863.49 ±117.62
--
15.58 ± 0.47
1.49 ± 0.06
--
1.49 ± 0.06
0.82 ± 0.03
0.01 ± 0.00
--
1.47 ± 0.06
--
1.47 ± 0.06
0.79 ± 0.03
0.01 ± 0.00
--
0.94 ±0.03
--
0.94 ±0.03
0.51 ±0.02
0.01 ±0.00
--
--
7.03 ±0.63
--
7.03 ±0.63
0.02 ±0.00
0.00 ±0.00
7.01 ±0.63
5.96 ±0.31
--
5.96 ±0.31
0.02 ±0.00
0.00 ±0.00
5.94 ±0.31
5.90 ± 0.46
--
5.90 ± 0.46
0.02 ± 0.00
0.00 ± 0.00
5.87 ± 0.45
8.22 ± 0.53
--
8.22 ± 0.53
0.02 ± 0.00
0.00 ± 0.00
8.20 ± 0.52
4.81 ± 0.49 14.81 ± 1.45
--
--
4.81 ± 0.49 14.81 ± 1.45
0.02 ± 0.00
0.03 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
4.78 ± 0.49 14.77 ± 1.44
269.82 ± 17.95
7.34 ±0.43 3402.21 ±172.20
--
--
--
269.82 ± 17.95
7.34 ±0.43 3402.21 ±172.20
0.02 ± 0.00
0.02 ±0.00
0.02 ± 0.00
0.00 ± 0.00
0.00 ±0.00 3392.76 ±171.74
269.78 ± 17.95
7.31 ±0.43
9.39 ± 0.57
949.05 ± 63.46 63.62 ± 4.15 74.94 ± 5.07 55.48 ±2.95 68.01 ±3.86
56.93 ± 3.37
7.91 ±0.70 3375.35 ±319.81
109.09 ± 14.15
--
--
--
109.09 ± 14.15
7.91 ±0.70 3375.35 ±319.81
0.02 ± 0.00
0.02 ±0.00
0.02 ± 0.00
0.00 ±0.00 3358.82 ±318.17
0.00 ± 0.00
109.07 ± 14.15
7.88 ±0.70
16.47 ± 1.68
73.72 ± 7.63
344.74 ± 39.63 49.72 ± 4.25 95.58 ± 9.67 59.44 ±5.06 58.90 ±4.74
5.63 ±0.45 4758.04 ±407.56
166.19 ± 9.29
7.36 ± 0.69
5.95 ±0.52
--
--
- -
--
--
5.63 ±0.45 4758.04 ±407.56
166.19 ± 9.29
7.36 ± 0.69
5.95 ±0.52
0.01 ± 0.00
0.01 ±0.00
0.02 ± 0.00
0.02 ± 0.00
0.01 ±0.00
0.00 ± 0.00
0.00 ±0.00 4750.16 ±406.98
0.00 ± 0.00
0.00 ±0.00
7.82 ± 0.61
5.93 ±0.52
166.17 ± 9.28
5.61 ±0.45
7.34 ± 0.69
69.96 ±6.98 75.88 ±7.47
657.02 ± 29.24 62.24 ± 6.12 98.54 ±10.89
66.24 ± 6.36
8.47 ±0.99 2073.23 ±198.75
534.16 ± 55.64
2.50 ± 0.23
4.91 ±0.56
9.61 ± 1.33
--
--
--
--
--
--
8.47 ±0.99 2073.23 ±198.75
4.91 ±0.56
2.50 ± 0.23
534.16 ± 55.64
9.61 ± 1.33
0.03 ± 0.01
0.02 ±0.00
0.02 ±0.00
0.02 ± 0.00
0.02 ± 0.00
0.02 ± 0.01
0.00 ± 0.00
0.00 ± 0.00
0.00 ±0.00
0.00 ±0.00 2069.29 ±198.53
0.00 ± 0.00
534.10 ± 55.63
2.48 ± 0.23
4.89 ±0.56
8.44 ±0.99
3.90 ± 0.31
9.58 ± 1.33
30.82 ± 3.60
30.70 ± 3.62 37.04 ±4.59 69.26 ±7.97
1845.38 ±195.57 78.90 ±10.36
3.27 ± 0.25
--
3.27 ± 0.25
0.01 ± 0.00
0.00 ± 0.00
3.26 ± 0.25
15.81 ± 0.16
12.49 ± 0.11
3.32 ± 0.10
0.68 ± 0.02
2.07 ± 0.06
--
20.39 ± 0.56
15.13 ± 0.14
13.41 ± 0.12
1.72 ± 0.06
0.45 ± 0.01
0.87 ± 0.03
--
14.08 ± 0.46
496.57 ± 5.53
490.77 ± 5.51
5.80 ± 0.28
0.04 ± 0.00
1.54 ± 0.09
4.21 ± 0.22
48.07 ± 2.44
496.74 ± 8.22
489.49 ± 8.18
7.25 ± 0.66
0.04 ± 0.00
1.61 ± 0.16
5.59 ± 0.51
50.66 ± 4.43
495.21 ±12.65
489.22 ±12.64
5.99 ± 0.48
0.03 ± 0.00
2.03 ± 0.22
3.91 ± 0.27
62.16 ± 6.10
497.76 ±10.27
493.59 ±10.21
4.17 ± 0.43
0.04 ± 0.00
0.98 ± 0.13
3.15 ± 0.32
31.38 ± 3.80
Figure 9: Detailed timing results for each selector. The actual edge weights for the illustrative PartConn and UnitSquare prob-
lems were pre-computed, and therefore their timings are not included. The Partition selector requires initialization of the Z-
values (7) for the graph using only the estimated edge weights. Since this is not particular to either the actual edge weights
(e.g. from the obstacle distribution) or the start/goal vertices from a particular instance, this initialization (sel-init) is considered
separately. The online running time (online) is broken into LazySP's three primary steps: the inner search (search), invoking
the edge selector (sel), and evaluating edges (eval). We also show the number of edges evaluated.
|
1110.3850 | 1 | 1110 | 2011-10-17T23:35:11 | On the Power of Adaptivity in Sparse Recovery | [
"cs.DS"
] | The goal of (stable) sparse recovery is to recover a $k$-sparse approximation $x*$ of a vector $x$ from linear measurements of $x$. Specifically, the goal is to recover $x*$ such that
||x-x*||_p <= C min_{k-sparse x'} ||x-x'||_q
for some constant $C$ and norm parameters $p$ and $q$. It is known that, for $p=q=1$ or $p=q=2$, this task can be accomplished using $m=O(k \log (n/k))$ non-adaptive measurements [CRT06] and that this bound is tight [DIPW10,FPRU10,PW11].
In this paper we show that if one is allowed to perform measurements that are adaptive, then the number of measurements can be considerably reduced. Specifically, for $C=1+eps$ and $p=q=2$ we show
- A scheme with $m=O((1/eps)k log log (n eps/k))$ measurements that uses $O(log* k \log \log (n eps/k))$ rounds. This is a significant improvement over the best possible non-adaptive bound.
- A scheme with $m=O((1/eps) k log (k/eps) + k \log (n/k))$ measurements that uses /two/ rounds. This improves over the best possible non-adaptive bound.
To the best of our knowledge, these are the first results of this type.
As an independent application, we show how to solve the problem of finding a duplicate in a data stream of $n$ items drawn from ${1, 2, ..., n-1}$ using $O(log n)$ bits of space and $O(log log n)$ passes, improving over the best possible space complexity achievable using a single pass. | cs.DS | cs | On the Power of Adaptivity in Sparse Recovery
Piotr Indyk
Eric Price
David P. Woodruff
October 11, 2018
Abstract
The goal of (stable) sparse recovery is to recover a k-sparse approximation x∗ of a vector x from
linear measurements of x. Specifically, the goal is to recover x∗ such that
kx − x∗kp ≤ C min
k-sparse x′ kx − x′kq
for some constant C and norm parameters p and q. It is known that, for p = q = 1 or p = q = 2, this
task can be accomplished using m = O(k log(n/k)) non-adaptive measurements [CRT06] and that this
bound is tight [DIPW10, FPRU10, PW11].
In this paper we show that if one is allowed to perform measurements that are adaptive , then the
number of measurements can be considerably reduced. Specifically, for C = 1 + ǫ and p = q = 2 we
show
• A scheme with m = O( 1
ǫ k log log(nǫ/k)) measurements that uses O(log∗ k · log log(nǫ/k))
rounds. This is a significant improvement over the best possible non-adaptive bound.
• A scheme with m = O( 1
improves over the best possible non-adaptive bound.
ǫ k log(k/ǫ) + k log(n/k)) measurements that uses two rounds. This
To the best of our knowledge, these are the first results of this type.
As an independent application, we show how to solve the problem of finding a duplicate in a data
stream of n items drawn from {1, 2, . . . , n − 1} using O(log n) bits of space and O(log log n) passes,
improving over the best possible space complexity achievable using a single pass.
1
1
0
2
t
c
O
7
1
]
S
D
.
s
c
[
1
v
0
5
8
3
.
0
1
1
1
:
v
i
X
r
a
1 Introduction
In recent years, a new "linear" approach for obtaining a succinct approximate representation of n-dimensional
vectors (or signals) has been discovered. For any signal x, the representation is equal to Ax, where A is an
m × n matrix, or possibly a random variable chosen from some distribution over such matrices. The vector
Ax is often referred to as the measurement vector or linear sketch of x. Although m is typically much
smaller than n, the sketch Ax often contains plenty of useful information about the signal x.
A particularly useful and well-studied problem is that of stable sparse recovery. We say that a vector
x′ is k-sparse if it has at most k non-zero coordinates. The sparse recovery problem is typically defined
as follows: for some norm parameters p and q and an approximation factor C > 0, given Ax, recover an
"approximation" vector x∗ such that
kx − x∗kp ≤ C min
(this inequality is often referred to as ℓp/ℓq guarantee).
If the matrix A is random, then Equation (1)
should hold for each x with some probability (say, 2/3). Sparse recovery has a tremendous number of
applications in areas such as compressive sensing of signals [CRT06, Don06], genetic data acquisition and
analysis [SAZ10, BGK+10] and data stream algorithms1 [Mut05, Ind07]; the latter includes applications to
network monitoring and data analysis.
k-sparse x′(cid:13)(cid:13)x − x′(cid:13)(cid:13)q
(1)
It is known [CRT06] that there exist matrices A and associated recovery algorithms that produce ap-
proximations x∗ satisfying Equation (1) with p = q = 1, constant approximation factor C, and sketch
length
m = O(k log(n/k))
(2)
A similar bound, albeit using random matrices A, was later obtained for p = q = 2 [GLPS10] (building
on [CCF02, CM04, CM06]). Specifically, for C = 1 + ǫ, they provide a distribution over matrices A with
m = O(
1
ǫ
k log(n/k))
(3)
rows, together with an associated recovery algorithm.
It is also known that the bound in Equation (2) is asymptotically optimal for some constant C and
p = q = 1, see [DIPW10] and [FPRU10] (building on [GG84, Glu84, Kas77]). The bound of [DIPW10]
also extends to the randomized case and p = q = 2. For C = 1 + ǫ, a lower bound of m = Ω( 1
ǫ k log(n/k))
was recently shown [PW11] for the randomized case and p = q = 2, improving upon the earlier work of
[DIPW10] and showing the dependence on ǫ is optimal. The necessity of the "extra" logarithmic factor
multiplying k is quite unfortunate: the sketch length determines the "compression rate", and for large n any
logarithmic factor can worsen that rate tenfold.
In this paper we show that this extra factor can be greatly reduced if we allow the measurement process
to be adaptive. In the adaptive case, the measurements are chosen in rounds, and the choice of the mea-
surements in each round depends on the outcome of the measurements in the previous rounds. The adaptive
measurement model has received a fair amount of attention in the recent years [JXC08, CHNR08, HCN09,
HBCN09, MSW08, AWZ08], see also [Def10]. In particular [HBCN09] showed that adaptivity helps re-
ducing the approximation error in the presence of random noise. However, no asymptotic improvement to
the number of measurements needed for sparse recovery (as in Equation (1)) was previously known.
1In streaming applications, a data stream is modeled as a sequence of linear operations on an (implicit) vector x. Example
operations include increments or decrements of x's coordinates. Since such operations can be directly performed on the linear
sketch Ax, one can maintain the sketch using only O(m) words.
1
Results
In this paper we show that adaptivity can lead to very significant improvements in the number
of measurements over the bounds in Equations (2) and (3). We consider randomized sparse recovery with
ℓ2/ℓ2 guarantee, and show two results:
1. A scheme with m = O( 1
ǫ k log log(nǫ/k)) measurements and an approximation factor C = 1 + ǫ. For
low values of k this provides an exponential improvement over the best possible non-adaptive bound.
The scheme uses O(log∗ k · log log(nǫ/k)) rounds.
2. A scheme with m = O( 1
ǫ k log(k/ǫ) + k log(n/k)) and an approximation factor C = 1 + ǫ. For
low values of k and ǫ this offers a significant improvement over the best possible non-adaptive bound,
since the dependence on n and ǫ is "split" between two terms. The scheme uses only two rounds.
Implications Our new bounds lead to potentially significant improvements to efficiency of sparse recovery
schemes in a number of application domains. Naturally, not all applications support adaptive measurements.
For example, network monitoring requires the measurements to be performed simultaneously, since we
cannot ask the network to "re-run" the packets all over again. However, a surprising number of applications
are capable of supporting adaptivity. For example:
• Streaming algorithms for data analysis: since each measurement round can be implemented by one
pass over the data, adaptive schemes simply correspond to multiple-pass streaming algorithms (see
[McG09] for some examples of such algorithms).
• Compressed sensing of signals: several architectures for compressive sensing, e.g., the single-pixel
camera of [DDT+08], already perform the measurements in a sequential manner. In such cases the
measurements can be made adaptive2. Other architectures supporting adaptivity are under develop-
ment [Def10].
• Genetic data analysis and acqusition: as above.
Therefore, it seems likely that the results in this paper will be applicable in a wide variety of scenarios.
As an example application, we show how to solve the problem of finding a duplicate in a data stream of
n arbitrarily chosen items from the set {1, 2, . . . , n − 1} presented in an arbitrary order. Our algorithm uses
O(log n) bits of space and O(log log n) passes. It is known that for a single pass, Θ(log2 n) bits of space
is necessary and sufficient [JST11], and so our algorithm improves upon the best possible space complexity
using a single pass.
Techniques On a high-level, both of our schemes follow the same two-step process. First, we reduce the
problem of finding the best k-sparse approximation to the problem of finding the best 1-sparse approximation
(using relatively standard techniques). This is followed by solving the latter (simpler) problem.
The first scheme starts by "isolating" most of of the large coefficients by randomly sampling ≈ ǫ/k
fraction of the coordinates; this mostly follows the approach of [GLPS10] (cf. [GGI+02]). The crux of the
algorithm is in the identification of the isolated coefficients. Note that in order to accomplish this using
O(log log n) measurements (as opposed to O(log n) achieved by the "standard" binary search algorithm)
we need to "extract" significantly more than one bit of information per measurements. To achieve this, we
proceed as follows. First, observe that if the given vector (say, z) is exactly 1-sparse, then one can extract
the position of the non-zero entry (say zj) from two measurements a(z) = Pi zi, and b(z) = Pi izi,
2We note that, in any realistic sensing system, minimizing the number of measurements is only one of several considerations.
Other factors include: minimizing the computation time, minimizing the amount of communication needed to transfer the mea-
surement matrices to the sensor, satisfying constraints on the measurement matrix imposed by the hardware etc. A detailed cost
analysis covering all of these factors is architecture-specific, and beyond the scope of this paper.
2
since b(z)/a(z) = j. A similar algorithm works even if z contains some "very small" non-zero entries:
we just round b(z)/a(z) to the nearest integer. This algorithm is a special case of a general algorithm that
achieves O(log n/ log SN R) measurements to identify a single coordinate xj among n coordinates, where
SN R = x2
j /kx[n]\jk2 (SNR stands for signal-to-noise ratio). This is optimal as a function of n and the
SNR [DIPW10].
A natural approach would then be to partition [n] into two sets {1, . . . , n/2} and {n/2 + 1, . . . n}, find
the heavier of the two sets, and recurse. This would take O(log n) rounds. The key observation is that
not only do we recurse on a smaller-sized set of coordinates, but the SNR has also increased since x2
j has
remained the same but the squared norm of the tail has dropped by a constant factor. Therefore in the
next round we can afford to partition our set into more than two sets, since as long as we keep the ratio of
log(# of sets ) and log SN R constant, we only need O(1) measurements per round. This ultimately leads
to a scheme that finishes after O(log log n) rounds.
In the second scheme, we start by hashing the coordinates into a universe of size polynomial in k and
1/ǫ, in a way that approximately preserves the top coefficients without introducing spurious ones, and in
such a way that the mass of the tail of the vector does not increase significantly by hashing. This idea is
inspired by techniques in the data stream literature for estimating moments [KNPW10, TZ04] (cf. [CCF02,
CM06, GI10]). Here, though, we need stronger error bounds. This enables us to identify the positions of
those coefficients (in the hashed space) using only O( 1
ǫ k log(k/ǫ)) measurements. Once this is done, for
each large coefficient i in the hash space, we identify the actual large coefficient in the preimage of i. This
can be achieved using the number of measurements that does not depend on ǫ.
2 Preliminaries
We start from a few definitions. Let x be an n-dimensional vector.
Definition 2.1. Define
Hk(x) = arg max
S∈[n]
S=k
kxSk2
to be the largest k coefficients in x.
Definition 2.2. For any vector x, we define the "heavy hitters" to be those elements that are both (i) in the
top k and (ii) large relative to the mass outside the top k. We define
Definition 2.3. Define the error
Hk,ǫ(x) = {j ∈ Hk(x) x2
2
2}
xHk(x)(cid:13)(cid:13)(cid:13)
j ≥ ǫ(cid:13)(cid:13)(cid:13)
xHk(x)(cid:13)(cid:13)(cid:13)
2
2
Err2(x, k) =(cid:13)(cid:13)(cid:13)
For the sake of clarity, the analysis of the algorithm in section 4 assumes that the entries of x are sorted
by the absolute value (i.e., we have x1 ≥ x2 ≥ . . . ≥ xn). In this case, the set Hk(x) is equal to [k];
this allows us to simplify the notation and avoid double subscripts. The algorithms themselves are invariant
under the permutation of the coordinates of x.
Running times of the recovery algorithms
In the non-adaptive model, the running time of the recovery
algorithm is well-defined:
it is the number of operations performed by a procedure that takes Ax as its
input and produces an approximation x∗ to x. The time needed to generate the measurement vectors A, or
to encode the vector x using A, is not included. In the adaptive case, the distinction between the matrix
3
generation, encoding and recovery procedures does not exist, since new measurements are generated based
on the values of the prior ones. Moreover, the running time of the measurement generation procedure heavily
depends on the representation of the matrix. If we suppose that we may output the matrix in sparse form
and receive encodings in time bounded by the number of non-zero entries in the matrix, our algorithms run
in n logO(1) n time.
3 Full adaptivity
This section shows how to perform k-sparse recovery with O(k log log(n/k)) measurements. The core
of our algorithm is a method for performing 1-sparse recovery with O(log log n) measurements. We then
extend this to k-sparse recovery via repeated subsampling.
3.1 1-sparse recovery
This section discusses recovery of 1-sparse vectors with O(log log n) adaptive measurements. First, in
Lemma 3.1 we show that if the heavy hitter xj is Ω(n) times larger than the ℓ2 error (xj is "Ω(n)-heavy"),
we can find it with two non-adaptive measurements. This corresponds to non-adaptive 1-sparse recovery
with approximation factor C = Θ(n); achieving this with O(1) measurements is unsurprising, because the
lower bound [DIPW10] is Ω(log1+C n).
Lemma 3.1 is not directly very useful, since xj is unlikely to be that large. However, if xj is D times
larger than everything else, we can partition the coordinates of x into D random blocks of size N/D and
perform dimensionality reduction on each block. The result will in expectation be a vector of size D where
the block containing j is D times larger than anything else. The first lemma applies, so we can recover the
block containing j, which has a 1/√D fraction of the ℓ2 noise. Lemma 3.2 gives this result.
We then have that with two non-adaptive measurements of a D-heavy hitter we can restrict to a subset
where it is an Ω(D3/2)-heavy hitter. Iterating log log n times gives the result, as shown in Lemma 3.3.
Lemma 3.1. Suppose there exists a j with xj ≥ C n√δ (cid:13)(cid:13)x[n]\{j}(cid:13)(cid:13)2
adaptive measurements suffice to recover j with probability 1 − δ.
Proof. Let s : [n] → {±1} be chosen from a 2-wise independent hash family. Perform the measurements
a(x) =P s(i)xi and b(x) =P(n + i)s(i)xi. For recovery, output the closest integer to b/a − n.
Let z = x[n]\{j}. Then E[a(z)2] = kzk2
2. Hence with probability at least
1 − 2δ, we have both
for some constant C. Then two non-
Thus
2 and E[b(z)2] ≤ 4n2 kzk2
a(z) ≤p1/δ kzk2
b(z) ≤ 2np1/δ kzk2
b(x)
a(x)
=
s(j)(n + j)xj + b(z)
s(j)xj + a(z)
(cid:12)(cid:12)(cid:12)(cid:12)
b(x)
a(x) − (n + j)(cid:12)(cid:12)(cid:12)(cid:12)
s(j)xj + a(z)
(cid:12)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)(cid:12)
b(z) − (n + j)a(z)
≤b(z) + (n + j)a(z)
4np1/δ kzk2
xj − a(z)
xj − a(z)
≤
4
Suppose xj > (8n + 1)p1/δ kzk2. Then
(cid:12)(cid:12)(cid:12)(cid:12)
b(x)
a(x) − (n + j)(cid:12)(cid:12)(cid:12)(cid:12)
<
4np1/δ kzk2
8np1/δ kzk2
=1/2
so ı = j.
Lemma 3.2. Suppose there exists a j with xj ≥ C B2
δ2 (cid:13)(cid:13)x[n]\{j}(cid:13)(cid:13)2
for some constant C and parameters B
and δ. Then with two non-adaptive measurements, with probability 1 − δ we can find a set S ⊂ [n] such
that j ∈ S and(cid:13)(cid:13)xS\{j}(cid:13)(cid:13)2 ≤(cid:13)(cid:13)x[n]\{j}(cid:13)(cid:13)2 /B and S ≤ 1 + n/B2.
Proof. Let D = B2/δ, and let h : [n] → [D] and s : [n] → {±1} be chosen from pairwise independent
hash families. Then define Sp = {i ∈ [n] h(i) = p}. Define the matrix A ∈ RD×n by Ah(i),i = s(i) and
Ap,i = 0 elsewhere. Then
Let p∗ = h(j) and y = x[n]\{j}
(Az)p = Xi∈Sp
. We have that
s(i)zi.
E[Sp∗] =1 + (n − 1)/D
E[(Ay)2
2
p∗] = E[(cid:13)(cid:13)(cid:13)
ySp∗(cid:13)(cid:13)(cid:13)
E[kAyk2
2
] =kyk2
2] =kyk2
2
2 /D
Hence by Chebyshev's inequality, with probability at least 1 − 4δ all of the following hold:
Sp∗ ≤1 + (n − 1)/(Dδ) ≤ 1 + n/B2
ySp∗(cid:13)(cid:13)(cid:13)2 ≤kyk2 /√Dδ
(cid:13)(cid:13)(cid:13)
(Ay)p∗ ≤kyk2 /√Dδ
kAyk2 ≤kyk2 /√δ.
(4)
(5)
(6)
(7)
The combination of (6) and (7) imply
(Ax)p∗ ≥xj − (Ay)p∗ ≥ (CD/δ − 1/√Dδ)kyk2 ≥ (CD/δ − 1/√Dδ)√δ kAyk2 ≥
CD
2√δ kAyk2
and hence
(Ax)p∗ ≥
CD
2√δ (cid:13)(cid:13)(Ax)[D]\p∗(cid:13)(cid:13)2 .
As long as C/2 is larger than the constant in Lemma 3.1, this means two non-adaptive measurements suffice
to recover p∗ with probability 1 − δ. We then output the set Sp∗, which by (5) has
xSp∗\{j}(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)
ySp∗(cid:13)(cid:13)(cid:13)2 ≤ kyk2 /√Dδ =(cid:13)(cid:13)x[n]\{j}(cid:13)(cid:13)2 /√Dδ =(cid:13)(cid:13)x[n]\{j}(cid:13)(cid:13)2 /B
=(cid:13)(cid:13)(cid:13)
as desired. The overall failure probability is 1 − 5δ; rescaling δ and C gives the result.
Lemma 3.3. Suppose there exists a j with xj ≥ C(cid:13)(cid:13)x[n]\{j}(cid:13)(cid:13)2
adaptive measurements suffice to recover j with probability 1/2.
for some constant C. Then O(log log n)
5
procedure NONADAPTIVESHRINK(x, D)
For i ∈ [n], s1(i) ← {±1}, h(i) ← [D]
For i ∈ [D], s2(i) ← {±1}
a ←P s1(i)s2(h(i))xi
b ←P s1(i)s2(h(i))xi(D + h(i))
p∗ ← ROUND(b/a − D).
return {j∗ h(j∗) = p∗}.
end procedure
procedure ADAPTIVEONESPARSEREC(x)
⊲ Find smaller set S containing heavy coordinate xj
⊲ Observation
⊲ Observation
⊲ Recover heavy coordinate xj
S ← [n]
B ← 2, δ ← 1/4
while S > 1 do
end while
return S[0]
end procedure
S ← NONADAPTIVESHRINK(xS, 4B2/δ)
B ← B3/2, δ ← δ/2.
Algorithm 3.1: Adaptive 1-sparse recovery
Proof. Let C′ be the constant from Lemma 3.2. Define B0 = 2 and Bi = B3/2
i−1 for i ≥ 1. Define δi = 2−i/4
for i ≥ 0. Suppose C ≥ 16C′B2
Define r = O(log log n) so Br ≥ n. Starting with S0 = [n], our algorithm iteratively applies Lemma 3.2
with parameters B = 4Bi and δ = δi to xSi to identify a set Si+1 ⊂ Si with j ∈ Si+1, ending when i = r.
We prove by induction that Lemma 3.2 applies at the ith iteration. We chose C to match the base case.
0/δ2
0.
For the inductive step, suppose(cid:13)(cid:13)xSi\{j}(cid:13)(cid:13)2 ≤ xj /(C′16 B2
so the lemma applies in the next iteration as well, as desired.
(cid:13)(cid:13)xSi+1\{j}(cid:13)(cid:13)2 ≤ xj /(C′64
After r iterations, we have Sr ≤ 1 + n/B2
that any iteration fails is at mostP δi < 2δ0 = 1/2.
B3
i
δ2
i
3.2 k-sparse recovery
). Then by Lemma 3.2,
i
δ2
i
) = xj /(C′16
i+1
B2
δ2
i+1
)
r < 2, so we have uniquely identified j ∈ Sr. The probability
Given a 1-sparse recovery algorithm using m measurements, one can use subsampling to build a k-sparse
recovery algorithm using O(km) measurements and achieving constant success probability. Our method
for doing so is quite similar to one used in [GLPS10]. The main difference is that, in order to identify one
large coefficient among a subset of coordinates, we use the adaptive algorithm from the previous section as
opposed to error-correcting codes.
For intuition, straightforward subsampling at rate 1/k will, with constant probability, recover (say) 90%
of the heavy hitters using O(km) measurements. This reduces the problem to k/10-sparse recovery: we can
subsample at rate 10/k and recover 90% of the remainder with O(km/10) measurements, and repeat log k
times. The number of measurements decreases geometrically, for O(km) total measurements. Naively
doing this would multiply the failure probability and the approximation error by log k; however, we can
make the number of measurements decay less quickly than the sparsity. This allows the failure probability
and approximation ratios to also decay exponentially so their total remains constant.
6
To determine the number of rounds, note that the initial set of O(km) measurements can be done in
parallel for each subsampling, so only O(m) rounds are necessary to get the first 90% of heavy hitters.
Repeating log k times would require O(m log k) rounds. However, we can actually make the sparsity in
subsequent iterations decay super-exponentially, in fact as a power tower. This give O(m log∗ k) rounds.
Theorem 3.4. There exists an adaptive (1+ǫ)-approximate k-sparse recovery scheme with O( 1
measurements and success probability 1 − δ. It uses O(log∗ k log log(nǫ)) rounds.
ǫ k log 1
δ log log(nǫ/k))
To prove this, we start from the following lemma:
Lemma 3.5. We can perform O(log log(n/k)) adaptive measurements and recover an ı such that, for any
j ∈ Hk,1/k(x) we have Pr[ı = j] = Ω(1/k).
Proof. Let S = Hk(x). Let T ⊂ [n] contain each element independently with probability p = 1/(4C 2k),
where C is the constant in Lemma 3.3. Let j ∈ Hk,1/k(x). Then we have
so
2
2
2
E[(cid:13)(cid:13)xT\S(cid:13)(cid:13)
(cid:13)(cid:13)xT\S(cid:13)(cid:13)2 ≤p4p(cid:13)(cid:13)xS(cid:13)(cid:13)2 =
2] = p(cid:13)(cid:13)xS(cid:13)(cid:13)
C√k (cid:13)(cid:13)xS(cid:13)(cid:13)2 ≤ xj /C
1
with probability at least 3/4. Furthermore we have E[T \ S] < pn so T \ S < n/k with probability at
least 1 − 1/(4C 2) > 3/4. By the union bound, both these events occur with probability at least 1/2.
Independently of this, we have
Pr[T ∩ S = {j}] = p(1 − p)k−1 > p/e
so all these events hold with probability at least p/(2e). Assuming this,
and T ≤ 1 + n/k. But then Lemma 3.3 applies, and O(log log T) = O(log log(n/k)) measurements can
recover j from a sketch of xT with probability 1/2. This is independent of the previous probability, for a
total success chance of p/(4e) = Ω(1/k).
(cid:13)(cid:13)xT\{j}(cid:13)(cid:13)2 ≤ xj /C
Lemma 3.6. With O( 1
ǫ k log 1
f δ log log(nǫ/k)) adaptive measurements, we can recover T with T ≤ k and
Err2(xT , f k) ≤ (1 + ǫ) Err2(x, k)
with probability at least 1 − δ. The number of rounds required is O(log log(nǫ/k)).
Proof. Repeat Lemma 3.5 m = O( 1
f δ ) times in parallel with parameters n and k/ǫ to get coordinates
T ′ = {t1, t2, . . . , tm}. For each j ∈ Hk,ǫ/k(x) ⊆ Hk/ǫ,ǫ/k(x) and i ∈ [m], the lemma implies Pr[j = ti] ≥
ǫ/(Ck) for some constant C. Then Pr[j /∈ T ′] ≤ (1 − ǫ/(Ck))m ≤ e−ǫm/(Ck) ≤ f δ for appropriate m.
Thus
ǫ k log 1
E[(cid:12)(cid:12)Hk,ǫ/k(x) \ T ′(cid:12)(cid:12)] ≤ f δ(cid:12)(cid:12)Hk,ǫ/k(x)(cid:12)(cid:12) ≤f δk
Pr(cid:2)(cid:12)(cid:12)Hk,ǫ/k(x) \ T ′(cid:12)(cid:12) ≥ f k(cid:3) ≤δ.
Now, observe xT ′ directly and set T ⊆ T ′ to be the locations of the largest k values. Then, since
Hk,ǫ/k(x) ⊆ Hk(x),(cid:12)(cid:12)Hk,ǫ/k(x) \ T(cid:12)(cid:12) =(cid:12)(cid:12)Hk,ǫ/k(x) \ T ′(cid:12)(cid:12) ≤ f k with probability at least 1 − δ.
7
Suppose this occurs, and let y = xT . Then
Err2(y, f k) = min
2
2
2
2
2
2
S≤f k(cid:13)(cid:13)yS(cid:13)(cid:13)
yHk,ǫ/k(x)\T(cid:13)(cid:13)(cid:13)
≤(cid:13)(cid:13)(cid:13)
xHk,ǫ/k(x)(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)
xHk(x)\Hk,ǫ/k(x)(cid:13)(cid:13)(cid:13)
+(cid:13)(cid:13)(cid:13)
xHk(x)(cid:13)(cid:13)(cid:13)
xHk(x)\Hk,ǫ/k(x)(cid:13)(cid:13)(cid:13)
+ k(cid:13)(cid:13)(cid:13)
xHk(x)(cid:13)(cid:13)(cid:13)
≤(cid:13)(cid:13)(cid:13)
≤(1 + ǫ)(cid:13)(cid:13)(cid:13)
xHk(x)(cid:13)(cid:13)(cid:13)
2
=(1 + ǫ) Err2(x, k)
2
2
2
2
2
2
2
2
∞
as desired.
procedure ADAPTIVEKSPARSEREC(x, k, ǫ, δ)
⊲ Recover approximation x of x
R0 ← [n]
δ0 ← δ/2, ǫ0 ← ǫ/e, f0 ← 1/32, k0 ← k.
J ← {}
for i ← 0, . . . , O(log∗ k) do
for t ← 0, . . . , Θ( 1
ǫi
ki log 1
δi
) do
St ← SUBSAMPLE(Ri, Θ(ǫi/ki))
J.add(ADAPTIVEONESPARSEREC(xSt))
end for
Ri+1 ← [n] \ J
δi+1 ← δi/8
ǫi+1 ← ǫi/2
fi+1 ← 1/21/(4i+1fi)
ki+1 ← kifi
end for
x ← xJ
return x
⊲ While ki ≥ 1
⊲ Direct observation
end procedure
Algorithm 3.2: Adaptive k-sparse recovery
Theorem 3.7. We can perform O( 1
size at most 2k with
ǫ k log 1
δ log log(nǫ/k)) adaptive measurements and recover a set T of
(cid:13)(cid:13)xT(cid:13)(cid:13)2 ≤ (1 + ǫ)(cid:13)(cid:13)(cid:13)
with probability 1 − δ. The number of rounds required is O(log∗ k log log(nǫ)).
e·2i . Let f0 = 1/32 and fi = 2−1/(4ifi−1) for i > 0, and define
Proof. Define δi = δ
ki = kQj<i fj. Let R0 = [n].
Let r = O(log∗ k) such that fr−1 < 1/k. This is possible since αi = 1/(4i+1fi) satisfies the recurrence
α0 = 8 and αi = 2αi−1−2i−2 > 2αi−1/2. Thus αr−1 > k for r = O(log∗ k) and then fr−1 < 1/αr−1 <
1/k.
xHk(x)(cid:13)(cid:13)(cid:13)2
2·2i and ǫi = ǫ
.
8
For each round i = 0, . . . , r − 1, the algorithm runs Lemma 3.6 on xRi with parameters ǫi, ki, fi, and
The total number of measurements is
δi to get Ti. It sets Ri+1 = Ri \ Ti and repeats. At the end, it outputs T = ∪Ti.
O(X 1
log log(nǫi/ki)) ≤O(X 2i(ki/k) log(1/fi)
) log(log(k/ki) + log(nǫ/k)))
log log(nǫ/k)X 2i(ki/k) log(1/fi)(i + 1) log log(k/ki))
using the very crude bounds i + log(1/δ) ≤ (i + 1) log(1/δ) and log(a + b) ≤ 2 log a log b for a, b ≥ e.
But then
≤O(
k(i + log
1
fiδi
ki log
k log
1
δ
1
ǫ
ǫi
ǫ
1
δ
X 2i(ki/k) log(1/fi)(i + 1) log log(k/ki) ≤X 2i(i + 1)fi log(1/fi) log log(1/fi)
≤X 2i(i + 1)O(pfi)
=O(1)
since fi < O(1/16i), giving O( 1
ǫ k log 1
δ log log(nǫ/k) total measurements. The probability that any of
the iterations fail is at most P δi < δ. The result has size T ≤ P ki ≤ 2k. All that remains is the
approximation ratio(cid:13)(cid:13)xT(cid:13)(cid:13)2 = kxRrk2.
For each i, we have
Err2(xRi+1, ki+1) = Err2(xRi\Ti, fiki) ≤ (1 + ǫi) Err2(xRi, ki).
Furthermore, kr < kfr−1 < 1. Hence
kxRrk2
2 = Err2(xRr , kr) ≤ r−1
Yi=0
(1 + ǫi)! Err2(xR0, k0) = r−1
Yi=0
(1 + ǫi)! Err2(x, k)
i=0 (1 + ǫi) < eP ǫi < e, so
ButQr−1
and hence
as desired.
r−1
(1 + ǫi) < 1 +X eǫi ≤ 1 + 2ǫ
Yi=0
xHk(x)(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)xT(cid:13)(cid:13)2 = kxRrk2 ≤ (1 + ǫ)(cid:13)(cid:13)(cid:13)
Once we find the support T , we can observe xT directly with O(k) measurements to get a (1 + ǫ)-
approximate k-sparse recovery scheme, proving Theorem 3.4
4 Two-round adaptivity
The algorithms in this section are invariant under permutation. Therefore, for simplicity of notation, the
analysis assumes our vectors x is sorted: x1 ≥ . . . ≥ xn = 0.
We are given a 1-round k-sparse recovery algorithm for n-dimensional vectors x using m(k, ǫ, n, δ)
measurements with the guarantee that its output x satisfies kx − xkp ≤ (1 + ǫ) · kx[k]kp for a p ∈ {1, 2}
with probability at least 1 − δ. Moreover, suppose its output x has support on a set of size s(k, ǫ, n, δ). We
show the following black box two-round transformation.
9
Theorem 4.1. Assume s(k, ǫ, n, δ) = O(k). Then there is a 2-round sparse recovery algorithm for n-
dimensional vectors x, which, in the first round uses m(k, ǫ/5, poly(k/ǫ), 1/100) measurements and in the
second uses O(k · m(1, 1, n, Θ(1/k))) measurements. It succeeds with constant probability.
Corollary 4.2. For p = 2, there is a 2-round sparse recovery algorithm for n-dimensional vectors x such
that the total number of measurements is O( 1
ǫ k log(k/ǫ) + k log(n/k)).
Proof of Corollary 4.2. In the first round it suffices to use CountSketch with s(k, ǫ, n, 1/100) = 2k, which
holds for any ǫ > 0 [PW11]. We also have that m(k, ǫ/5, poly(k/ǫ), 1/100) = O( 1
ǫ k log(k/ǫ)). Us-
ing [CCF02, CM06, GI10], in the second round we can set m(1, 1, n, Θ(1/k)) = O(log n). The bound
follows by observing that 1
ǫ k log(k/ǫ) + k log(n) = O( 1
ǫ k log(k/ǫ) + k log(n/k)).
Proof of Theorem 4.1. In the first round we perform a dimensionality reduction of the n-dimensional input
vector x to a poly(k/ǫ)-dimensional input vector y. We then apply the black box sparse recovery algorithm
on the reduced vector y, obtaining a list of s(k, ǫ/5, poly(k/ǫ), 1/100) coordinates, and show for each
coordinate in the list, if we choose the largest preimage for it in x, then this list of coordinates can be used
to provide a 1 + ǫ approximation for x. In the second round we then identify which heavy coordinates in
x map to those found in the first round, for which it suffices to invoke the black box algorithm with only a
constant approximation. We place the estimated values of the heavy coordinates obtained in the first pass in
the locations of the heavy coordinates obtained in the second pass.
Let N = poly(k/ǫ) be determined below. Let h : [n] → [N ] and σ : [n] → {−1, 1} be Θ(log N )-wise
independent random functions. Define the vector y by yi = Pj h(j)=i σ(j)xj. Let Y (i) be the vector x
restricted to coordinates j ∈ [n] for which h(j) = i. Because the algorithm is invariant under permutation
of coordinates of y, we may assume for simplicity of notation that y is sorted: y1 ≥ . . . ≥ yN = 0.
We note that such a dimensionality reduction is often used in the streaming literature. For example, the
sketch of [TZ04] for ℓ2-norm estimation utilizes such a mapping. A "multishot" version (that uses several
functions h) has been used before in the context of sparse recovery [CCF02, CM06] (see [GI10] for an
overview). Here, however, we need to analyze a "single-shot" version.
Let p ∈ {1, 2}, and consider sparse recovery with the ℓp/ℓp guarantee. We can assume that kxkp = 1.
We need two facts concerning concentration of measure.
Fact 4.3. (see, e.g., Lemma 2 of [KNPW10]) Let X1, . . . , Xn be such that Xi has expectation µi and
variance v2
i , and Xi ≤ K almost surely. Then if the Xi are ℓ-wise independent for an even integer ℓ ≥ 2,
Pr"(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xi=1
where µ =Pi µi and v2 =Pi v2
Fact 4.4. (Khintchine inequality) ([Haa82]) For t ≥ 2, a vector z and a t-wise independent random sign
vector σ of the same number of dimensions, E[hz, σit] ≤ kzkt
≥ λ# ≤ 2O(ℓ)(cid:18)(cid:16)v√ℓ/λ(cid:17)ℓ
Xi − µ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
+ (Kℓ/λ)ℓ(cid:19)
n
i .
2(√t)t.
We start with a probabilistic lemma. Let Z(j) denote the vector Y (j) with the coordinate m(j) of largest
magnitude removed.
N 1/6(cid:17) and N be sufficiently large. Then with probability ≥ 99/100,
Lemma 4.5. Let r = O(cid:16)kx[k]kp · log N
1. ∀j ∈ [N ], kZ(j)kp ≤ r.
2. ∀i ∈ [N 1/3], σ(i) · yh(i) − xi ≤ r,
3. ky[k]kp ≤ (1 + O(1/√N )) · kx[k]kp + O(kr),
10
4. ∀j ∈ [N ], if h−1(j) ∩ [N 1/3] = ∅, then yj ≤ r,
5. ∀j ∈ [N ], kY (j)k0 = O(n/N + log N ).
Proof. We start by defining events E, F and G that will be helpful in the analysis, and showing that all of
them are satisfied simultaneously with constant probability.
Event E: Let E be the event that h(1), h(2), . . . , h(N 1/3) are distinct. Then Prh[E] ≥ 1 − 1/N 1/3.
Event F: Fix i ∈ [N ]. Let Z′ denote the vector Y (h(i)) with the coordinate i removed. Applying Fact 4.4
with t = Θ(log N ),
Pr
σ
[σ(i)yh(i) − xi ≥ 2√t · kZ(h(i))k2] ≤ Pr
≤ Pr
≤ Pr
σ
= Pr
σ
σ
σ
[σ(i)yh(i) − xi ≥ 2√t ·(cid:13)(cid:13)Z′(cid:13)(cid:13)2]
[σ(i)yh(i) − xit ≥ 2t(√t)t · kZ′kt
2]
t]]
[σ(i)yh(i) − xit ≥ 2t E[(cid:12)(cid:12)hσ, Z′i(cid:12)(cid:12)
[σ(i)yh(i) − xit ≥ 2t E[σ(i)yh(i) − xit] ≤ 1/N 4/3.
Let F be the event that for all i ∈ [N ], σ(i)yh(i) − xi ≤ 2√t · kZ(h(i))k2, so Prσ[F] ≥ 1 − 1/N.
Event G: Fix j ∈ [N ] and for each i ∈ {N 1/3 + 1, . . . , n}, let Xi = xip1h(i)=j (i.e., Xi = xip if
h(i) = j). We apply Lemma 4.3 to the Xi. In the notation of that lemma, µi = xip/N and v2
i ≤ xi2p/N,
2p/N. Also, K = xN 1/3+1p. Function h is Θ(log N )-wise
and so µ = kx
independent, so by Fact 4.3,
p/N and v2 ≤ kx
[N 1/3]k2p
[N 1/3]kp
kx
p
[N 1/3]kp
N
Xi −
Xi
≥ λ# ≤2O(ℓ)(cid:18)(cid:16)kx
[N 1/3]kp
2p
√ℓ/(λ√N )(cid:17)ℓ
+(cid:0)xN 1/3+1pℓ/λ(cid:1)ℓ(cid:19)
for any λ > 0 and an ℓ = Θ(log N ). For ℓ large enough, there is a
Pr"(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
[N 1/3]kp
λ = Θ(kx
2pp(log N )/N + xN 1/3+1p · log N )
for which this probability is ≤ N−2. Let G be the event that for all j ∈ [N ], kZ(j)kp
for some universal constant C > 0. Then Pr[G E] ≥ 1 − 1/N.
By a union bound, Pr[E ∧ F ∧ G] ≥ 999/1000 for N sufficiently large.
the event E ∧ F ∧ G holds (i.e., we condition on that event).
First Condition: This condition follows from the occurrence of G, and using that kx
[N 1/3]kp,
p. One just needs to make these
and kx
substitutions into the variable λ defining G and show the value r serves as an upper bound (in fact, there is
a lot of room to spare, e.g., r/ log N is also an upper bound).
[N 1/3]kp ≤ kx[k]kp, as well as (N 1/3 − k + 1)xN 1/3+1p ≤ kx[k]kp
We know proceed to proving the five conditions in the lemma statement. In the analysis we assume that
p ≤ C(cid:18)kx
[N 1/3]k2p ≤ kx
+ λ(cid:19)
[N1/3]kp
p
N
Second Condition: This condition follows from the joint occurrence of E, F, and G.
11
p + O(kr).
p ≤ ky′kp
p/(N 1/3 − k + 1) since m(j) /∈ [N 1/3].
2 ≤ (1 + O(1/√N ))kx[k]k2
2.
p + O(kr) ≤ (1 + O(1/√N ))kx[k]kp
2 ≤ (1 + O(1/√N ))kx′k2
p to ky[k]kp
We relate ky′kp
Third Condition:
For the third condition, let y′ denote the restriction of y to coordinates in the set
[N ] \ {h(1), h(2), ..., h(k)}. For p = 1 and for any choice of h and σ, ky′k1 ≤ kx[k]k1. For p = 2,
the vector y is the sketch of [TZ04] for ℓ2-estimation. By their analysis, with probability ≥ 999/1000,
ky′k2
2, where x′ is the vector whose support is [n] \ ∪k
i=1h−1(i) ⊆ [n] \ [k]. We
assume this occurs and add 1/1000 to our error probability. Hence, ky′k2
p. Consider any j = h(i) for an i ∈ [k] for which j is not among the top k
coordinates of y. Call such a j lost. By the first condition of the lemma, σ(i)yj − xi ≤ r. Since j is
not among the top k coordinates of y, there is a coordinate j′ among the top k coordinates of y for which
j′ /∈ h([k]) and yj ′ ≥ yj ≥ xi − r. We call such a j′ a substitute. We can bijectively map substitutes to
lost coordinates. It follows that ky[k]kp
Fourth Condition: This follows from the joint occurrence of E,F, and G, and using that xm(j)p ≤
kx[k]kp
Fifth Condition: For the fifth condition, fix j ∈ [N ]. We apply Fact 4.3 where the Xi are indicator
variables for the event h(i) = j. Then E[Xi] = 1/N and Var[Xi] < 1/N. In the notation of Fact 4.3,
µ = n/N, v2 < n/N, and K = 1. Setting ℓ = Θ(log N ) and λ = Θ(log N +p(n log N )/N ), we have
by a union bound that for all j ∈ [N ], kY (j)k0 ≤ n
N + Θ(log N +p(n log N )/N ) = O(n/N + log N ),
with probability at least 1 − 1/N.
By a union bound, all events jointly occur with probability at least 99/100, which completes the proof.
Event H: Let H be the event that the algorithm returns a vector y with ky − ykp ≤ (1 + ǫ/5)ky[k]kp. Then
Pr[H] ≥ 99/100. Let S be the support of y, so S = s(k, ǫ/5, N, 1/100). We condition on H.
In the second round we run the algorithm on Y (j) for each j ∈ S, each using m(1, 1,kY (j)k0, Θ(1/k)))-
measurements. Using the fifth condition of Lemma 4.5, we have that kY (j)k0 = O(ǫn/k + log(k)/ǫ) for
N = poly(k/ǫ) sufficiently large.
For each invocation on a vector Y (j) corresponding to a j ∈ S, the algorithm takes the largest (in
magnitude) coordinate HH(j) in the output vector, breaking ties arbitrarily. We output the vector x with
support equal to T = {HH(j) j ∈ S}. We assign the value σ(xj)yj to HH(j). We have
p + kx[n]\Tkp
p.
p + k(x − x)[n]\Tkp
p = k(x − x)Tkp
kx − xkp
p =k(x − x)Tkp
(8)
The rest of the analysis is devoted to bounding the RHS of equation 8.
Lemma 4.6. For N = poly(k/ǫ) sufficiently large, conditioned on the events of Lemma 4.5 and H,
kx[n]\Tkp
p ≤ (1 + ǫ/3)kx[k]kp
p.
Proof. If [k] \ T = ∅, the lemma follows by definition. Otherwise, if i ∈ ([k] \ T ), then i ∈ [k], and so
by the second condition of Lemma 4.5, xi ≤ yh(i) + r. We also use the third condition of Lemma 4.5 to
obtain ky[k]kp ≤ (1 + O(1/√N )) · kx[k]kp + O(kr). By the triangle inequality,
≤ k1/pr +
1/p
1/p
1/p
Xi∈[k]\T
xip
yh(i)p
≤ k1/pr +
≤k1/pr + (1 + ǫ/5) · ky[k]kp.
Xi ∈ [k]\T
Xi ∈ [N ]\S
yip
The lemma follows using that r = O(kx[k]k2 · (log N )/N 1/6) and N = poly(k/ǫ) is sufficiently large.
12
1/p
1/p
1/p
1/p
≤
Xj∈S
We bound k(x − x)Tkp
p using Lemma 4.5, S ≤ poly(k/ǫ), and that N = poly(k/ǫ) is sufficiently large.
(yj − yj + σ(HH(j)) · xHH(j) − yj)p
xHH(j) − σ(HH(j)) · yjp
k(x − x)Tkp ≤
Xj∈S
σ(HH(j)) · xHH(j) − yjp
yj − yjp
≤
Xj∈S
≤(1 + ǫ/5)ky[k]kp +
Xj∈S
≤(1 + ǫ/5)(1 + O(1/√N ))kx[k]kp + O(kr) +
Xj∈S
σ(HH(j)) · xHH(j) − yjp
≤(1 + ǫ/4)kx[k]kp +
Xj∈S
+
Xj∈S
σ(HH(j)) · xHH(j) − yjp
σ(HH((j)) · xHH(j) − yjp
1/p
1/p
1/p
p < 1
p ≤ 2 · kw[1]kp
p, we have w1p > 3
p ≥ w1 − w1p, so if w1p < 3
Event I: We condition on the event I that all second round invocations succeed. Note that Pr[I] ≥ 99/100.
We need the following lemma concerning 1-sparse recovery algorithms.
Lemma 4.7. Let w be a vector of real numbers. Suppose w1p > 9
which kw − wkp
Proof. kw − wkp
On the other hand, kw[1]kp
part, for j > 1 we have wjp < 1
10(cid:1)kwkp
then kw − wkp
p > (cid:0) 3
5 − 1
kw − wkp
p ≤ 2 · kw[1]kp
p.
It remains to boundPj∈S σ(HH(j))·xHH(j)− yjp. We show for every j ∈ S, σ(HH(j))·xHH(j)−
yjp is small.
p. Moreover, for all j > 1, wjp < 3
5(cid:1)kwkp
p, then kw − wkp
p > (cid:0) 9
p. This contradicts that kw − wkp
p ≤ 2 · kw[1]kp
10 · kwkp
2 · kwkp
p < 1
10 · kwkp
p.
p. For the second
5 · kwkp
p,
p, this contradicts that
p ≥ wj − wjp, so if wjp ≥ 3
10 · kwkp
10 · kwkp
p = 1
Recall that m(j) is the coordinate of Y (j) with the largest magnitude. There are two cases.
p. But since kw[1]kp
5 · kwkp
5 · kwkp
p. Now, kw − wkp
10 · kwkp
p. Then for any vector w for
5 · kwkp
p.
p = 3
10 − 3
Case 1: m(j) /∈ [N 1/3]. In this case observe that HH(j) /∈ [N 1/3] either, and h−1(j) ∩ [N 1/3] = ∅.
It follows by the fourth condition of Lemma 4.5 that yj ≤ r. Notice that xHH(j)p ≤ xm(j)p ≤ kx[k]kp
N 1/3−k .
Bounding σ(HH(j)) · xHH(j) − yj by xHH(j) + yj, it follows for N = poly(k/ǫ) large enough that
σ(HH(j)) · xHH(j) − yjp ≤ ǫ/4 · kx[k]kp/S).
Case 2: m(j) ∈ [N 1/3]. If HH(j) = m(j), then σ(HH(j)) · xHH(j) − yj ≤ r by the second con-
dition of Lemma 4.5, and therefore
p
for N = poly(k/ǫ) large enough.
σ(HH(j)) · xHH(j) − yjp ≤ rp ≤ ǫ/4 · kx[k]kp/S
13
Otherwise, HH(j) 6= m(j). From condition 2 of Lemma 4.5 and m(j) ∈ [N 1/3], it follows that
σ(HH(j)))xHH(j) − yj ≤σ(HH(j))xHH(j) − σ(m(j))xm(j) + σ(m(j))xm(j) − yj ≤ xHH(j) + xm(j) + r
Notice that xHH(j) + xm(j) ≤ 2xm(j) since m(j) is the coordinate of largest magnitude. Now, condi-
tioned on I, Lemma 4.7 implies that xm(j)p ≤ 9
p, or equivalently, xm(j) ≤ 101/p · kZ(j)kp.
Finally, by the first condition of Lemma 4.5, we have kZ(j)kp = O(r), and so σ(HH(j))xHH(j) − yjp =
O(rp), which as argued above, is small enough for N = poly(k/ǫ) sufficiently large.
10 · kY (j)kp
The proof of our theorem follows by a union bound over the events that we defined.
5 Adaptively Finding a Duplicate in a Data Stream
We consider the following FindDuplicate problem. We are given an adversarially ordered stream S of n
elements in {1, 2, . . . , n− 1} and the goal is to output an element that occurs at least twice, with probability
at least 1 − δ. We seek to minimize the space complexity of such an algorithm. We improve the space
complexity of [JST11] for FindDuplicate from O(log2 n) bits to O(log n) bits, though we use O(log log n)
passes instead of a single pass. Notice that [JST11] also proves a lower bound of Ω(log2 n) bits for a single
pass.
We use Lemma 3.3 of our multi-pass sparse recovery algorithm:
Fact 5.1. Suppose there exists an i with xi ≥ Ckx[n]\{i}k2 for some constant C. Then O(log log n)
adaptive measurements suffice to recover a set T of constant size so that i ∈ T with probability at least 1/2.
Further, all adaptive measurements are linear combinations with integer coefficients of magnitude bounded
by poly(n).
Our algorithm DuplicateFinder for this problem considers the equivalent formulation of FindDuplicate
in which we think of an underlying frequency vector x ∈ {−1, 0, 1, . . . , n − 1}n. We start by initializing
xi = −1 for all i. Each time item i occurs in the stream, we increment its frequency by 1. The task is
therefore to output an i for which xi > 0.
Theorem 5.2. There is an O(log log n)-pass, O(log n log 1/δ) bits of space per pass algorithm for solving
the FindDuplicate problem with probability at least 1 − δ.
Proof. We describe an algorithm DuplicateFinder which succeeds with probability at least 1/8. Since it
knows whether or not it succeeds, the probability can be amplified to 1 − δ by O(log 1/δ) independent par-
allel repetitions. It is easy to see that the pass and space complexity are as claimed, so we prove correctness.
14
DuplicateFinder(S)
1. Repeat the following procedure C = O(1) times independently.
(a) Select O(1)-wise independent uniform ti ∈ [0, 1] for i ∈ [n].
(b) Let ǫ > 0 be a sufficiently small constant. Let m = O(log 1/ǫ).
(c) Let z1, . . . , z4m be a pairwise-independent partition of the coordinates of z, where z = xi/ti for
all i.
(d) Run algorithm A independently on vectors z1, . . . , z4m.
(e) Let T1, . . . , T4m be the outputs of algorithm A on z1, . . . , z4m, respectively, as per Fact 5.1.
(f) Compute each xi for i ∈ ∪4m
j=1Tj in an extra pass. If there is an i for which xi > 0, then output
i.
2. If no coordinate i has been output, then output FAIL.
We use the following fact shown in the proof of Lemma 3 of [JST11].
Lemma 5.3. (see first paragraph of Lemma 3 of [JST11]) For a single index i ∈ [n] and t arbitrary, we
have
1
20
√mkxk1 ti = t] = O(ǫ),
Pr[kzHm(z)k2 >
where Hm(z) denotes the set of m largest (in magnitude) coordinates of z. Suppose zi > kxk1 for
. Conditioned on this
some value of i. This happens if ti < xi
kxk1
event, by Lemma 5.3 we have that with probability 1 − O(ǫ),
and occurs with probability equal to xi
kxk1
kzHm(z)k2 ≤
1
20
√mkxk1.
Suppose i occurs in zj for some value of j ∈ [4m]. Since the partition is pairwise-independent, the expected
number of ℓ ∈ Hm(z) \ {i} which occur in zj is at most m
4m , and so with probability at least 3/4 − O(ǫ),
the norm of zj with coordinate corresponding to coordinate i in z removed is at most
kzHm(z)k2 ≤
1
20
√mkxk1 ≤
1
20
√mzi.
Consider one of the C = O(1) independent repetitions of step 1.
Since m is a constant, by Fact 5.1, with probability at least 1/2, A outputs a set T which contains coordinate
i. Hence, with probability at least 3/8 − O(ǫ), if there is an i for which zi > kxk1, it is found by
DuplicateFinder.
Let pi = xi
. Then zi > kxk1 with probability pi. Since Pi xi > 0, we have Pi xi>0 pi > 1
2.
kxk1
For coordinates i for which xi > 0, let Wi = 1 if zi > kxk1, and let W =Pi Wi. Then E[W ] > 1/2
and by pairwise-independence, Var[W ] ≤ E[W ].
2 and Var[W ′] ≤ E[W ]
E[W ] > 1
have that with probability at least 1
coordinate i for which xi > 0 and zi > kxk1.
This completes the proof.
Let W ′ be the average of the random variable W over C independent repetitions. Then E[W ′] =
C , and so by Chebyshev's inequality for C = O(1) sufficiently large we
2, W ′ > 0, which means that in one of the C repetitions there is a
Hence, the overall probability of success is at least 1/2 · (3/8 − O(ǫ)) > 1/8, for ǫ sufficiently small.
15
Acknowledgements
This material is based upon work supported by the Space and Naval Warfare Systems Center Pacific under
Contract No. N66001-11-C-4092, David and Lucille Packard Fellowship, MADALGO (Center for Massive
Data Algorithmics, funded by the Danish National Research Association) and NSF grant CCF-1012042. E.
Price is supported in part by an NSF Graduate Research Fellowship.
References
[AWZ08] A. Aldroubi, H. Wang, and K. Zarringhalam. Sequential adaptive compressed sampling via
huffman codes. Preprint, 2008.
[BGK+10] A. Bruex, A. Gilbert, R. Kainkaryam, John Schiefelbein, and Peter Woolf. Poolmc: Smart
pooling of mRNA samples in microarray experiments. BMC Bioinformatics, 2010.
[CCF02] M. Charikar, K. Chen, and M. Farach-Colton. Finding frequent items in data streams. ICALP,
2002.
[CHNR08] R. Castro, J. Haupt, R. Nowak, and G. Raz. Finding needles in noisy haystacks. Proc. IEEE
Conf. Acoustics, Speech, and Signal Proc., page 51335136, 2008.
[CM04]
[CM06]
G. Cormode and S. Muthukrishnan. Improved data stream summaries: The count-min sketch
and its applications. LATIN, 2004.
G. Cormode and S. Muthukrishnan. Combinatorial algorithms for Compressed Sensing.
Proc. 40th Ann. Conf. Information Sciences and Systems, Princeton, Mar. 2006.
In
[CRT06]
E. J. Cand`es, J. Romberg, and T. Tao. Stable signal recovery from incomplete and inaccurate
measurements. Comm. Pure Appl. Math., 59(8):1208 -- 1223, 2006.
[DDT+08] M. Duarte, M. Davenport, D. Takhar, J. Laska, T. Sun, K. Kelly, and R. Baraniuk. Single-pixel
imaging via compressive sampling. IEEE Signal Processing Magazine, 2008.
[Def10]
Defense Sciences Office. Knowledge enhanced compressive measurement. Broad Agency
Announcement, DARPA-BAA-10-38, 2010.
[DIPW10] K. Do Ba, P. Indyk, E. Price, and D. Woodruff. Lower bounds for sparse recovery. SODA,
2010.
[Don06]
D. L. Donoho. Compressed Sensing. IEEE Trans. Info. Theory, 52(4):1289 -- 1306, Apr. 2006.
[FPRU10] S. Foucart, A. Pajor, H. Rauhut, and T. Ullrich. The gelfand widths of lp-balls for 0 < p ≤ 1.
J. Complexity, 2010.
[GG84]
A. Y. Garnaev and E. D. Gluskin. On widths of the euclidean ball. Sov. Math., Dokl., page
200204, 1984.
[GGI+02] A. C. Gilbert, S. Guha, P. Indyk, Y. Kotidis, S. Muthukrishnan, and M. J. Strauss. Fast, small-
space algorithms for approximate histogram maintenance. In ACM Symposium on Theoretical
Computer Science, 2002.
[GI10]
A. Gilbert and P. Indyk. Sparse recovery using sparse matrices. Proceedings of IEEE, 2010.
16
[GLPS10] Anna C. Gilbert, Yi Li, Ely Porat, and Martin J. Strauss. Approximate sparse recovery: opti-
mizing time and measurements. In STOC, pages 475 -- 484, 2010.
[Glu84]
[Haa82]
E. D. Gluskin. Norms of random matrices and widths of finite-dimensional sets. Math. USSR-
Sb., 48:173182, 1984.
Uffe Haagerup. The best constants in the Khintchine inequality. Studia Math., 70(3):231 -- 283,
1982.
[HBCN09] J. Haupt, R. Baraniuk, R. Castro, and R. Nowak. Compressive distilled sensing. Asilomar,
2009.
[HCN09]
J. Haupt, R. Castro, and R. Nowak. Adaptive sensing for sparse signal recovery. Proc. IEEE
13th Digital Sig. Proc./5th Sig. Proc. Education Workshop, page 702707, 2009.
[Ind07]
[JST11]
[JXC08]
[Kas77]
P. Indyk. Sketching, streaming and sublinear-space algorithms. Graduate course notes, avail-
able at http://stellar.mit.edu/S/course/6/fa07/6.895/, 2007.
Hossein Jowhari, Mert Saglam, and G´abor Tardos. Tight bounds for lp samplers, finding dupli-
cates in streams, and related problems. In PODS, pages 49 -- 58, 2011.
S. Ji, Y. Xue, and L. Carin. Bayesian compressive sensing. IEEE Trans. Signal Processing,
56(6):23462356, 2008.
B. S. Kashin. Diameters of some finite-dimensional sets and classes of smooth functions. Math.
USSR, Izv.,, 11:317333, 1977.
[KNPW10] Daniel M. Kane, Jelani Nelson, Ely Porat, and David P. Woodruff. Fast moment estimation in
data streams in optimal space. CoRR, abs/1007.4191, 2010.
[McG09] A. McGregor. Graph mining on streams. Encyclopedia of Database Systems, page 12711275,
2009.
[MSW08] D. M. Malioutov, S. Sanghavi, and A. S. Willsky. Compressed sensing with sequential obser-
vations. ICASSP, 2008.
[Mut05]
S. Muthukrishnan. Data streams: Algorithms and applications). Foundations and Trends in
Theoretical Computer Science, 2005.
[PW11]
E. Price and D. Woodruff. (1+eps)-approximate sparse recovery. FOCS, 2011.
[SAZ10]
N. Shental, A. Amir, and Or Zuk. Identification of rare alleles and their carriers using com-
pressed se(que)nsing. Nucleic Acids Research, 38(19):1 -- 22, 2010.
[TZ04]
Mikkel Thorup and Yin Zhang. Tabulation based 4-universal hashing with applications to
second moment estimation.
In Proceedings of the 15th Annual ACM-SIAM Symposium on
Discrete Algorithms (SODA), pages 615 -- 624, 2004.
17
|
1707.03478 | 2 | 1707 | 2018-02-01T18:13:10 | Round Compression for Parallel Matching Algorithms | [
"cs.DS",
"cs.DC"
] | For over a decade now we have been witnessing the success of {\em massive parallel computation} (MPC) frameworks, such as MapReduce, Hadoop, Dryad, or Spark. One of the reasons for their success is the fact that these frameworks are able to accurately capture the nature of large-scale computation. In particular, compared to the classic distributed algorithms or PRAM models, these frameworks allow for much more local computation. The fundamental question that arises in this context is though: can we leverage this additional power to obtain even faster parallel algorithms?
A prominent example here is the {\em maximum matching} problem---one of the most classic graph problems. It is well known that in the PRAM model one can compute a 2-approximate maximum matching in $O(\log{n})$ rounds. However, the exact complexity of this problem in the MPC framework is still far from understood. Lattanzi et al. showed that if each machine has $n^{1+\Omega(1)}$ memory, this problem can also be solved $2$-approximately in a constant number of rounds. These techniques, as well as the approaches developed in the follow up work, seem though to get stuck in a fundamental way at roughly $O(\log{n})$ rounds once we enter the near-linear memory regime. It is thus entirely possible that in this regime, which captures in particular the case of sparse graph computations, the best MPC round complexity matches what one can already get in the PRAM model, without the need to take advantage of the extra local computation power.
In this paper, we finally refute that perplexing possibility. That is, we break the above $O(\log n)$ round complexity bound even in the case of {\em slightly sublinear} memory per machine. In fact, our improvement here is {\em almost exponential}: we are able to deliver a $(2+\epsilon)$-approximation to maximum matching, for any fixed constant $\epsilon>0$, in $O((\log \log n)^2)$ rounds. | cs.DS | cs |
Round Compression for Parallel Matching Algorithms
Artur Czumaj
Jakub Ł acki
Aleksander M adry
University of Warwick
Google Research, New York
MIT
[email protected]
[email protected]
[email protected]
Slobodan Mitrovi´c
EPFL
Krzysztof Onak
IBM Research
[email protected]
[email protected]
Piotr Sankowski
University of Warsaw
[email protected]
Abstract
For over a decade now we have been witnessing the success of massive parallel computation (MPC)
frameworks, such as MapReduce, Hadoop, Dryad, or Spark. One of the reasons for their success is
the fact that these frameworks are able to accurately capture the nature of large-scale computation. In
particular, compared to the classic distributed algorithms or PRAM models, these frameworks allow for
much more local computation. The fundamental question that arises in this context is though: can we
leverage this additional power to obtain even faster parallel algorithms?
A prominent example here is the maximum matching problem-one of the most classic graph prob-
lems. It is well known that in the PRAM model one can compute a 2-approximate maximum matching in
O(log n) rounds. However, the exact complexity of this problem in the MPC framework is still far from
understood. Lattanzi et al. (SPAA 2011) showed that if each machine has n1+Ω(1) memory, this problem
can also be solved 2-approximately in a constant number of rounds. These techniques, as well as the
approaches developed in the follow up work, seem though to get stuck in a fundamental way at roughly
O(log n) rounds once we enter the (at most) near-linear memory regime. It is thus entirely possible that
in this regime, which captures in particular the case of sparse graph computations, the best MPC round
complexity matches what one can already get in the PRAM model, without the need to take advantage
of the extra local computation power.
In this paper, we finally refute that perplexing possibility. That is, we break the above O(log n) round
complexity bound even in the case of slightly sublinear memory per machine. In fact, our improvement
here is almost exponential: we are able to deliver a (2 + ǫ)-approximation to maximum matching, for
any fixed constant ǫ > 0, in O(cid:0)(log log n)2(cid:1) rounds.
To establish our result we need to deviate from the previous work in two important ways that are cru-
cial for exploiting the power of the MPC model, as compared to the PRAM model. Firstly, we use vertex–
based graph partitioning, instead of the edge–based approaches that were utilized so far. Secondly, we
develop a technique of round compression. This technique enables one to take a (distributed) algorithm
that computes an O(1)-approximation of maximum matching in O(log n) independent PRAM phases
and implement a super-constant number of these phases in only a constant number of MPC rounds.
1 Introduction
Over the last decade, massive parallelism became a major paradigm in computing, and we have witnessed
the deployment of a number of very successful massively parallel computation frameworks, such as MapRe-
duce [DG04, DG08], Hadoop [Whi12], Dryad [IBY+07], or Spark [ZCF+10]. This paradigm and the
corresponding models of computation are rather different from classical parallel algorithms models con-
sidered widely in literature, such as the PRAM model.
In particular, in this paper, we study the Mas-
sive Parallel Computation (MPC) model (also known as Massively Parallel Communication model) that
was abstracted out of capabilities of existing systems, starting with the work of Karloff, Suri, and Vas-
silvitskii [KSV10, GSZ11, BKS13, ANOY14, BKS14]. The main difference between this model and the
PRAM model is that the MPC model allows for much more (in principle, unbounded) local computation.
This enables it to capture a more "coarse–grained," and thus, potentially, more meaningful aspect of par-
allelism.
It is often possible to simulate one clock step of PRAM in a constant number of rounds on
MPC [KSV10, GSZ11]. This implies that algorithms for the PRAM model usually give rise to MPC al-
gorithms without incurring any asymptotic blow up in the number of parallel rounds. As a result, a vast
body of work on PRAM algorithms naturally translates to the new model.
It is thus natural to wonder: Are the MPC parallel round bounds "inherited" from the PRAM model
tight? In particular, which problems can be solved in significantly smaller number of MPC rounds than
what the lower bounds established for the PRAM model suggest?
It is not hard to come up with an example of a problem for which indeed the MPC parallel round
number is much smaller than its PRAM round complexity. For instance, computing the parity of n Boolean
values takes only O(1) parallel rounds in the MPC model when space per machine is nΩ(1), while on
PRAM it provably requires Ω(log n/ log log n) time [BH87] (as long as the total number of processors is
polynomial). However, the answer is typically less obvious for other problems. This is particularly the
case for graph problems, whose study in a variant of the MPC model was initiated already by Karloff et
al. [KSV10].
In this paper, we focus on one such problem, which is also one of the most central graph problems
both in sequential and parallel computations: maximum matching. Maximum matchings have been the
cornerstone of algorithmic research since 1950s and their study inspired many important ideas, including
the complexity class P [Edm65]. In the PRAM model we can compute (1 + ǫ)-approximate matching in
O(log n) rounds [LPP15] using randomization. Deterministically, a (2 + ǫ)-approximation can be computed
processors are located at graph nodes and can communicate only with neighbors.
in O(cid:0)log2 n(cid:1) rounds [FG17]. We note that these results hold in a distributed message passing setting, where
setting, Ω(cid:16)plog n/ log log n(cid:17) time lower bound is known for computing any constant approximation to
maximum matching [KMW06].
In such a distributed
So far, in the MPC setting, the prior results are due to Lattanzi, Moseley, Suri, and Vassilvitskii [LMSV11],
Ahn and Guha [AG15] and Assadi and Khanna [AK17]. Lattanzi et al. [LMSV11] put forth algorithms for
several graph problems, such as connected components, minimum spanning tree, and maximum match-
ing problem, that were based on a so-called filtering technique. In particular, using this technique, they
have obtained an algorithm that can compute a 2-approximation to maximum matching in O(1/δ) MPC
rounds, provided S, the space per machine, is significantly larger than the total number of vertices n, that
is S = Ω(cid:0)n1+δ(cid:1), for any constant δ ∈ (0, 1). Later on, Ahn and Guha [AG15] provided an improved algo-
rithm that computes a (1+ ǫ)-approximation in O(1/(δǫ)) rounds, provided S = Ω(cid:0)n1+δ(cid:1), for any constant
δ > 0. Both these results, however, crucially require that space per machine is significantly superlinear in
n, the number of vertices. In fact, if the space S is linear in n, which is a very natural setting for massively
1
parallel graph algorithms, the performance of both these algorithms degrades to O(log n) parallel rounds,
which matches what was known for the PRAM model. Recently, Assadi and Khanna [AK17] showed how
to construct randomized composable coresets of size O(n) that give an O(1)-approximation for maximum
matching. Their techniques apply to the MPC model only if the space per machine is O(n√n).
We also note that the known PRAM maximal independent set and maximal matching algorithms [Lub86,
ABI86, II86] can be used to find a maximal matching (i.e., 2-approximation to maximum matching) in
O(log n) MPC rounds as long as space per machine is at least nΩ(1) (i.e., S ≥ nc for some constant c > 0).
We omit further details here, except mentioning that a more or less direct simulation of those algorithms is
possible via an O(1)-round sorting subroutine [GSZ11].
The above results give rise to the following fundamental question: Can the maximum matching be
(approximately) solved in o(log n) parallel rounds in O(n) space per machine? The main result of this
paper is an affirmative answer to that question. We show that, for any S = Ω(n), one can obtain an O(1)-
approximation to maximum matching using O(cid:0)(log log n)2(cid:1) parallel MPC rounds. So, not only do we
break the existing Ω(log n) barrier, but also provide an exponential improvement over the previous work.
Our algorithm can also provide a (2 + ǫ), instead of O(1)-approximation, at the expense of the number of
parallel rounds increasing by a factor of O(log(1/ǫ)). Finally, our approach can also provide algorithms
that have o(log n) parallel round complexity also in the regime of S being (mildly) sublinear. For instance,
we obtain O(cid:0)(log log n)2(cid:1) MPC rounds even if space per machine is S = n/(log n)O(log log n). The exact
comparison of our bounds with previous results is given in Table 1.
1.1 The model
In this work, we adopt a version of the model introduced by Karloff, Suri, and Vassilvitskii [KSV10] and
refined in later works [GSZ11, BKS13, ANOY14]. We call it massive parallel computation (MPC), which
is a mutation of the name proposed by Beame et al. [BKS13].
In the MPC model, we have m machines at our disposal and each of them has S words of space. Initially,
each machine receives its share of the input. In our case, the input is a collection E of edges and each
machine receives approximately E/m of them.
The computation proceeds in rounds. During the round, each of the machines processes its local data
without communicating with other machines. At the end of each round, machines exchange messages. Each
message is sent only to a single machine specified by the machine that is sending the message. All messages
sent and received by each machine in each round have to fit into the machine's local memory. Hence, their
total length is bounded by S.1 This in particular implies that the total communication of the MPC model is
bounded by m · S in each round. The messages are processed by recipients in the next round.
machine has to fit in its local memory. Hence again, each machine can output at most S words.
At the end of the computation, machines collectively output the solution. The data output by each
If the input is of size N , one usually wants S sublinear in the N , and
The range of values for S and m.
the total space across all the machines to be at least N -so the input fits onto the machines-and ideally not
In this paper, the focus is on graph algorithms. If n is the number of vertices in the graph, the input size
much larger. Formally, one usually considers S ∈ Θ(cid:0)N 1−ǫ(cid:1), for some ǫ > 0.
can be as large as Θ(cid:0)n2(cid:1). Our parallel algorithm requires Θ(n) space per machine (or even slightly less),
which is polynomially less than the size of the input for dense graphs.
1This for instance allows a machine to send a single word to S/100 machines or S/100 words to one machine, but not S/100
words to S/100 machines if S = ω(1), even if the messages are identical.
2
Sparse graphs. Many practical large graphs are believed to have only O(n) edges. One natural example
is social networks, in which most participants are likely to have a bounded number of friends. The additional
advantage of our approach is that it allows for a small number of processing rounds even if a sparse input
graph does not fit onto a single machine. If a small number-say, f (n)-of machines is needed even to
store the graph, our algorithm still requires only O(cid:0)(log log n)2 + log f (n)(cid:1) rounds for O(n/f (n)) space
per machine.
Communication vs. computation complexity. The main focus of this work is the number of (communica-
tion) rounds required to finish computation. Also, even though we do not make an effort to explicitly bound
it, it is apparent from the design of our algorithms that every machines performs O(S polylog S) computa-
tion steps locally. This in particular implies that the overall work across all the machines is O(rN polylog S),
where r is the number of rounds and N is the input size (i.e., the number of edges).
and it is known that computing a (1 + ǫ)-approximate matching in the message passing model with Θ(n)
The total communication during the computation is O(rN ) words. This is at most O(cid:0)rn2(cid:1) words
edges per player may require Ω(cid:0)n2/(1 + ǫ)2(cid:1) bits of communication [HRVZ15]. Since our value of r is
O(cid:0)(log log n)2(cid:1) when Θ(n) edges are assigned to each player, we lose a factor of Θ(log n) compared to
this lower bound if words (and vertex identifiers) have Θ(log n) bits.
1.2 Our results
In our work, we focus on computing an O(1)-approximate maximum matching in the MPC model. We
collect our results and compare to the previous work in Table 1. The table presents two interesting regimes
Source
Approx.
[LMSV11]
2
[AG15]
1 + ǫ
2
O(1)
2 + ǫ
O(1)
2 + ǫ
here
Space
n1+Ω(1)
O(n)
O(cid:0)n1+1/p(cid:1)
nΩ(1)
O(n)
O(n)/f (n)
Remarks
Maximal matching
p > 1
Maximal matching
Simulate [Lub86, ABI86, II86]
ǫ ∈ (0, 1/2)
2 ≤ f (n) = O(cid:0)n1/2(cid:1)
Rounds
O(1)
O(log n)
O(p/ǫ)
O(log n)
O(cid:0)(log log n)2(cid:1)
O(cid:0)(log log n)2 · log(1/ǫ)(cid:1)
O(cid:0)(log log n)2 + log f (n)(cid:1)
O(cid:0)(log log n)2 + log f (n)(cid:1) · log(1/ǫ)
Table 1: Comparison of our results for computing approximate maximum size matchings to the previous
results for the MPC model.
for our algorithms. On the one hand, when the space per machine is S = O(n), we obtain an algorithm
that requires O((log log n)2) rounds. This is the first known algorithm that, with linear space per machine,
breaks the O(log n) round barrier. On the other hand, in the mildly sublinear regime of space per machine,
i.e., when S = O(n/f (n)), for some function f (n) that is no(1), we obtain an algorithm that still requires
o(log n) rounds. This, again is the first such result in this regime. In particular, we prove the following
result.
3
Theorem 1.1. There exists an MPC algorithm that constructs an O(1)-approximation to maximum matching
with constant probability in O(cid:0)(log log n)2 + max(cid:0)log n
space on each machine.
S , 0(cid:1)(cid:1) rounds, where S = nΩ(1) is the amount of
As a corollary, we obtain the following result that provides nearly 2-approximate maximum matching.
2 ), there exists an MPC algorithm that constructs a (2+ǫ)-approximation to
Assadi et al. [ABB+17] observe that one can use a technique of McGregor [McG05] to extend the
Corollary 1.2. For any ǫ ∈ (0, 1
maximum matching with 99/100 probability in O(cid:0)(log log n)2 + max(cid:0)log n
S = nΩ(1) is the amount of space on each machine.
algorithm to compute a (1 + ǫ)-approximation in O((log log n)2) · (1/ǫ)O(1/ǫ) rounds.
It should also be noted that (as pointed out to us by Seth Pettie) any O(1)-approximation algorithm for
unweighted matchings can be used to obtain a (2 + ǫ)-approximation algorithm for weighted matchings (see
Section 4 of his paper with Lotker and Patt-Shamir [LPP15] for details). In our setting this implies that
Theorem 1.1 yields an algorithm that computes a (2 + ǫ)-approximation to maximum weight matching in
O((log log n)2 · (1/ǫ)) rounds and O(n log n) space per machine.
S , 0(cid:1)(cid:1) · log(1/ǫ) rounds, where
1.3 Related work
We note that there were efforts at modeling MapReduce computation [FMS+10] before the work of Karloff
et al. Also a recent work [RVW16] investigates the complexity of the MPC model.
In the filtering technique, introduced by Lattanzi et al. [LMSV11], the input graph is iteratively sparsified
until it can be stored on a single machine. For the matching problem, the sparsification is achieved by first
obtaining a small sample of edges, then finding a maximal matching in the sample, and finally removing
all the matched vertices. Once a sufficiently small graph is obtained, a maximal matching is computed on
a single machine. In the S = Θ(n) regime, the authors show that their approach reduces the number of
edges by a constant factor in each iteration. Despite this guarantee, until the very last step, each iteration
may make little progress towards obtaining even an approximate maximal matching, resulting in a O(log n)
round complexity of the algorithm. Similarly, the results of Ahn and Guha [AG15] require n1+Ω(1) space
per machine to compute a O(1)-approximate maximum weight matching in a constant number of rounds
and do not imply a similar bound for the case of linear space.
We note that the algorithm of Lattanzi et al. [LMSV11] cannot be turned easily into a fast approximation
algorithm when space per machine is sublinear. Even with Θ(n) space, their method is able to remove only
a constant fraction of edges from the graph in each iteration, so Ω(log n) rounds are needed until only a
matching is left. When S = Θ(n), their algorithm works as follows: sample uniformly at random Θ(n)
edges of the graph, find maximal matching on the sampled set, remove the matched vertices, and repeat.
We do not provide a formal proof here, but on the following graph this algorithm requires Ω(log n) rounds,
even to discover a constant factor approximation. Consider a graph consisting of t separate regular graphs
of degree 2i, for 0 ≤ i ≤ t− 1, each on 2t vertices. This graph has t2t nodes and the algorithm requires Ω(t)
rounds even to find a constant approximate matching. The algorithm chooses edges uniformly at random,
and few edges are selected each round from all but the densest remaining subgraphs. Thus, it takes multiple
rounds until a matching of significant size is constructed for sparser subgraphs. This example emphasizes
the weakness of direct edge sampling and motivates our vertex sampling scheme that we introduce in this
paper.
Similarly, Ahn and Gupta [AG15] build on the filtering approach of Lattanzi et al. and design a primal-
dual method for computing a (1 + ǫ)-approximate weighted maximum matching. They show that each
4
iteration of their distributed algorithm either makes large progress in the dual, or they can construct a large
approximate matching. Regardless of their new insights, their approach is inherently edge-sampling based
and does not break the O(log n) round complexity barrier when S = O(n).
Despite the fact that MPC model is rather new, computing matching is an important problem in this
model, as the above mentioned two papers demonstrate. This is further witnessed by the fact that the
distributed and parallel complexity of maximal matching has been studied for many years already. The best
deterministic PRAM maximal matching algorithm, due to Israeli and Shiloach [IS86], runs in O(cid:0)log3 n(cid:1)
rounds. Israeli and Itai [II86] gave a randomized algorithm for this problem that runs in O(log n) rounds.
Their algorithm works as well in CONGEST, a distributed message-passing model with a processor assigned
to each vertex and a limit on the amount of information sent along each edge per round. A more recent paper
by Lotker, Patt-Shamir, and Pettie [LPP15] gives a (1+ǫ)-approximation to maximum matching in O(log n)
rounds also in the CONGEST model, for any constant ǫ > 0. On the deterministic front, in the LOCAL
model, which is a relaxation of CONGEST that allows for an arbitrary amount of data sent along each edge,
round algorithm by Fischer and Ghaffari [FG17].
On the negative side, Kuhn, Moscibroda, and Wattenhofer [KMW06] showed that any distributed algo-
a line of research initiated by Ha´n´ckowiak, Karo´nski, and Panconesi [HKP01, HKP99] led to an O(cid:0)log3 n(cid:1)-
rithm, randomized or deterministic, when communication is only between neighbors requires Ω(cid:16)plog n/ log log n(cid:17)
rounds to compute a constant approximation to maximum matching. This lower bound applies to all dis-
tributed algorithms that have been mentioned above. Our algorithm circumvents this lower bound by loosen-
ing the only possible assumption there is to be loosened: single-hop communication. In a sense, we assign
subgraphs to multiple machines and allow multi-hop communication between nodes in each subgraph.
Finally, the ideas behind the peeling algorithm that is a starting point for this paper can be traced back
to the papers of Israeli, Itai, and Shiloach [II86, IS86], which can be interpreted as matching high-degree
vertices first in order to reduce the maximum degree. A sample distributed algorithm given in a work
of Parnas and Ron [PR07] uses this idea to compute an O(log n) approximation for vertex cover. Their
algorithm was extended by Onak and Rubinfeld [OR10] in order to provide an O(1)-approximation for
vertex cover and maximum matching in a dynamic version of the problems. This was achieved by randomly
matching high-degree vertices to their neighbors in consecutive phases while reducing the maximum degree
in the remaining graph. This approach was further developed in the dynamic graph setting by a number of
papers [BHI15, BHN16, BHN17, BCH17]. Ideas similar to those in the paper of Parnas and Ron [PR07]
were also used to compute polylogarithmic approximation in the streaming model by Kapralov, Khanna,
and Sudan [KKS14]. Our version of the peeling algorithm was directly inspired by the work of Onak and
Rubinfeld [OR10] and features important modifications in order to make our analysis go through.
1.4 Future challenges
We show a parallel matching algorithm in the MPC model by taking an algorithm that can be seen as a
distributed algorithm in the so-called LOCAL model. This algorithm requires Θ(log n) rounds and can
be simulated in Θ(log n) MPC rounds relatively easily with nΩ(1) space per machine. We develop an ap-
proximate version of the algorithm that uses much fewer rounds by repeatedly compressing a superconstant
number of rounds of the original algorithm to O(1) rounds. It is a great question if this kind of speedup can
be obtained for other-either distributed or PRAM-algorithms.
As for the specific problem considered in this paper, an obvious question is whether our round complex-
ity is optimal. We conjecture that there is a better algorithm that requires O(log log n) rounds, the square
root of our complexity. Unfortunately, a factor of log n in one of our functions (see the logarithmic factor in
5
α, a parameter defined later in the paper) propagates to the round complexity, where it imposes a penalty of
log log n.
Note also that as opposed to the paper of Onak and Rubinfeld [OR10], we do not obtain an O(1)-
approximation to vertex cover. This stems from the fact that we discard so-called reference sets, which can
be much bigger than the minimum vertex cover. This is unfortunately necessary in our analysis. Is there a
way to fix this shortcoming of our approach?
Finally, we suspect that there is a simpler algorithm for the problem that avoids the intricacies of our
approach and proceeds by simply greedily matching high-degree vertices on induced subgraphs without
sophisticated sampling in every phase. Unfortunately, we do not know how to analyze this kind of approach.
1.5 Recent developments
Since an earlier version of this work was shared on arXiv, it has inspired two followup works. First, As-
sadi [Ass17] applied the round compression idea to the distributed O(log n)-approximation algorithm for
vertex cover of Parnas and Ron [PR07]. Using techniques from his recent work with Khanna [AK17], he
gave a simple MPC algorithm that in O(log log n) rounds and n/ polylog(n) space per machine computes
an O(log n)-approximation to minimum vertex cover.
Second, a new paper by Assadi et al. [ABB+17] addresses, among other things, several open questions
from this paper. They give an MPC algorithm that computes O(1)-approximation to both vertex cover
and maximum matching in O(log log n) rounds and O(n) space per machine (though the space is strictly
superlinear). Their result builds on techniques developed originally for dynamic matching algorithms [BS15,
BS16] and composable coresets [AK17]. It is worth to note that their construction critically relies on the
vertex sampling approach (i.e., random assignment of vertices to machines) introduced in our work.
1.6 Notation
For a graph G = (V, E) and V ′ ⊆ V , we write G[V ′] to denote the subgraph of G induced by V ′. Formally,
G[V ′] def= (V ′, E ∩ (V ′ × V ′)). We also write N (v) to denote the set of neighbors of a vertex v in G.
2 Overview
In this section we present the main ideas and techniques behind our result. Our paper contains two main
technical contributions.
First, our algorithm randomly partitions vertices across the machines, and on each machine considers
only the corresponding induced graph. We prove that it suffices to consider these induced subgraphs to
obtain an approximate maximum matching. Note that this approach greatly deviates from previous works,
that used edge based partitioning.
Second, we introduce a round compression technique. Namely, we start with an algorithm that executes
O(log n) phases and can be naturally implemented in O(log n) MPC rounds and then demonstrate how to
emulate this algorithm using only o(log n) MPC rounds. The underlying idea is quite simple: each machine
independently runs multiple phases of the initial algorithm. This approach, however, has obvious challenges
since the machines cannot communicate in a single round of the MPC algorithm. The rest of the section is
devoted to describing our approach and illustrating how to overcome these challenges.
6
2.1 Vertex based sampling
The algorithms for computing maximal matching in PRAM and their simulations in the MPC model [Lub86,
ABI86, IS86, II86] are designed to, roughly speaking, either halve the number of the edges or halve the
maximum degree in each round. Therefore, in the worst case those algorithms inherently require Ω(log n)
rounds to compute a maximal matching.
On the other hand, all the algorithm for the maximal matching problem in the MPC model prior to
ours ([LMSV11, AG15, AK17]) process the input graph by discarding edges, and eventually aggregate the
remaining edges on a single machine to decide which of them are part of the final matching. It is not known
how to design approaches similar to [LMSV11, AG15, AK17] while avoiding a step in which the maximal
matching computation is performed on a single machine. This seems to be a barrier for improving upon
O(log n) rounds, if the space available on each machine is O(n).
The starting point of our new approach is alleviating this issue by resorting to a more careful vertex based
sampling. Specifically, at each round, we randomly partition the vertex set into vertex sets V1, . . . , Vm and
consider induced graphs on those subsets independently. Such sampling scheme has the following handy
property: the union of matchings obtained across the machines is still a matching. Furthermore, we show
that for the appropriate setting of parameters this sampling scheme allows us to handle vertices of a wide
range of degrees in a single round, unlike handling only high-degree vertices (that is, vertices with degree
within a constant factor of the maximum degree) as guaranteed by [II86, IS86].
2.2 Global algorithm
To design an algorithm executed on machines locally, we start from a sequential peeling algorithm GlobalAlg
(see Algorithm 1), which is a modified version of an algorithm used by Onak and Rubinfeld [OR10]. The
algorithm had to be significantly adjusted in order to make our later analysis of a parallel version possible.
The execution of GlobalAlg is divided into Θ(log n) phases. In each phase, the algorithm first com-
putes a set H of high-degree vertices. Then it selects a set F of vertices, which we call friends. Next the
algorithm selects a matching fM between H and F , using a simple randomized strategy. F is carefully con-
structed so that both F andfM are likely to be of order Θ(H). Finally, the algorithm removes all vertices in
H∪F , hence reducing the maximum vertex degree in the graph by a constant factor, and proceeds to the next
phase. The central property of GlobalAlg is that it returns an O(1) approximation to maximum matching
with constant probability (Corollary 3.4). A detailed discussion of GlobalAlg is given in Section 3.
Global matching algorithm
Algorithm 1: GlobalAlg(G,e∆)
Input: Graph G = (V, E) of maximum degree at most e∆
1 ∆ ← e∆, M ← ∅, V ′ ← V
2 while ∆ ≥ 1 do
Output: A matching in G
3
4
5
6
/* Invariant: the maximum degree in G[V ′] is at most ∆ */
Let H ⊂ V ′ be a set of vertices of degree at least ∆/2 in G[V ′]. We call vertices in H heavy.
Create a set F of friends by selecting each vertex v ∈ V ′ independently with probability N (v) ∩ H/4∆.
Compute a matchingfM in G[H ∪ F ] using MatchHeavy(H, F ) and add it to M .
V ′ ← V ′ \ (H ∪ F ), ∆ ← ∆/2
7 return M
7
Algorithm 2: MatchHeavy(H, F )
Computing a matching in G[H ∪ F ]
Input: set H of heavy vertices and set F of friends
Output: a matching in G[H ∪ F ]
1 For every vertex v ∈ F pick uniformly at random a heavy neighbor v⋆ in N (v) ∩ H.
2 Independently at random color each vertex in H ∪ F either red or blue.
3 Select the following subset of edges: E⋆ ← {(v, v⋆) : v ∈ F ∧ v is red ∧ v⋆ ∈ H ∧ v⋆ is blue}.
4 For every blue vertex w incident to an edge in E⋆, select one such edge and add it to fM .
5 returnfM
2.3 Parallel emulation of the global algorithm (Section 4)
space.
The following two ways could be used to execute GlobalAlg in the MPC model: (1) place the whole
graph on one machine, and trivially execute all the phases of GlobalAlg in a single round; or (2) simulate
one phase of GlobalAlg in one MPC round while using O(n) space per machine, by distributing vertices
randomly onto machines (see Section 6.1 for details). However, each of these approaches has severe draw-
backs. The first approach requires Θ(E) space per machine, which is likely to be prohibitive for large
graphs. On the other hand, while the second approach uses O(n) space, it requires Θ(log n) rounds of MPC
computation. We achieve the best of both worlds by showing how to emulate the behavior of multiple phases
of GlobalAlg in a single MPC round with each machine using O(n) space, thus obtaining an MPC algo-
rithm requiring o(log n) rounds. More specifically, we show that it is possible to emulate the behavior of
GlobalAlg in O(cid:0)(log log n)2(cid:1) rounds with each machine using O(n) (or even only n/(log n)O(log log n))
Before we provide more details about our parallel multi-phase emulation of GlobalAlg, let us mention
the main obstacle such an emulation encounters. At the beginning of every phase, GlobalAlg has access
to the full graph. Therefore, it can easily compute the set of heavy vertices H. On the other hand, machines
in our MPC algorithm use O(n) space and thus have access only to a small subgraph of the input graph
(when E ≫ n). In the first phase this is not a big issue, as, thanks to randomness, each machine can
estimate the degrees of high-degree vertices. However, the degrees of vertices can significantly change from
phase to phase. Therefore, after each phase it is not clear how to select high-degree vertices in the next
phase without inspecting the entire graph again. Hence, one of the main challenges in designing a multi-
phase emulation of GlobalAlg is to ensure that machines at the beginning of every phase can estimate
global degrees of vertices well enough to identify the set of heavy vertices, while each machine still having
access only to its local subgraph. This property is achieved using a few modifications to the algorithm.
2.3.1 Preserving randomness
Our algorithm partitions the vertex set into m disjoint subsets Vi by assigning each vertex independently
and uniformly at random. Then the graph induced by each subset Vi is processed on a separate machine.
Each machine finds a set of heavy vertices, Hi, by estimating the global degree of each vertex of Vi. It is not
hard to argue (using a standard concentration bound) that there is enough randomness in the initial partition
so that local degrees in each induced subgraph roughly correspond to the global degrees. Hence, after the
described partitioning, sets H andSi∈[m] Hi have very similar properties. This observation crucially relies
on the fact that initially the vertices are distributed independently and uniformly at random.
However, if one attempts to execute the second phase of GlobalAlg without randomly reassigning
8
vertices to sets after the first phase, the remaining vertices are no longer distributed independently and
uniformly at random. In other words, after inspecting the neighborhood of every vertex locally and making
a decision based on it, the randomness of the initial random partition may significantly decrease.
Let us now make the following thought experiment. Imagine for a moment that there is an algorithm
that emulates multiples phases of GlobalAlg in parallel and in every phase inspects only the vertices that
end-up being matched. Then, from the point of view of the algorithm, the vertices that are not matched so
far are still distributed independently and uniformly at random across the machines. Or, saying in a different
way, if randomness of some vertices is not inspected while emulating a phase, then at the beginning of the
next phase those vertices still have the same distribution as in the beginning of that MPC round. But, how
does an algorithm learn about vertices that should be matched by inspecting no other vertex? How does the
algorithm learn even only about high-degree vertices without looking at their neighborhood?
In the sequel we show how to design an algorithm that looks only "slightly" at the vertices that do not end-up
being matched. As we prove, that is sufficient to design a multi-phase emulation of GlobalAlg.
We now discuss in more detail how to preserve two crucial properties of our vertex assignments through-
out the execution of multiple phases: independent and nearly-uniform distribution.
2.3.2
Independence (Lemma 4.3)
As noted above, it is not clear how to compute vertex degrees without inspecting their local neighborhood.
A key, and at first sight counter-intuitive, step in our approach is to estimate even local degrees of vertices
(in contrast to computing them exactly). To obtain the estimates, it suffices to examine only small neighbor-
hoods of vertices and in turn preserve the independent distribution of the intact ones. More precisely, we
sample a small set of vertices on each machine, called reference sets, and use the set to estimate the local
degrees of all vertices assigned to this machine. Furthermore, we show that with a proper adjustments of
GlobalAlg these estimates are sufficient for capturing high-degree vertices.
Very crucially, all the vertices that are used in computing a matching in one emulated phase (including
the reference sets) are discarded at the end of the phase, even if they do not participate in the obtained
matching. In this way we disregard the vertices which position is fixed and, intuitively, secure an indepen-
dent distribution of the vertices across the machines in the next phase.
We also note, without going into details, that obtaining full independence required modifying how the
set of friends is selected, compared to the original approach of Onak and Rubinfeld [OR10]. In their ap-
proach, each heavy vertex selected one friend at random. However, as before, in order to select exactly one
friend would require examining neighborhood of heavy vertices. This, however, introduces dependencies
between vertices that have not been selected. So instead, in our GlobalAlg, every vertex selects itself as
a friend independently and proportionally to the number of high-degree vertices (found using the reference
set), which again secures an independent distribution of the remaining vertices. The final properties of the
obtained sets in either approach are very similar.
2.3.3 Uniformity (Lemma 4.4)
A very convenient property in the task of emulating multiple phases of GlobalAlg is a uniform distribution
of vertices across all the machines at every phase – for such a distribution, we know the expected number of
neighbors of each desired type assigned to the same machine. Obtaining perfect uniformity seems difficult-
if not impossible in our setting-and we therefore settle for near uniformity of vertex assignments. The
probability of the assignment of each vertex to each machine is allowed to differ slightly from that in the
uniform distribution. Initially, the distribution of each vertex is uniform and with every phase it can deviate
9
more and more from the uniform distribution. We bound the rate of the decay with high probability and
execute multiple rounds as long as the deviation from the uniform distribution is negligible. More precisely,
in the execution of the entire parallel algorithm, the sufficiently uniform distribution is on average kept over
(log log n)2(cid:17) phases of the emulation of GlobalAlg.
Ω(cid:16) log n
µH(r)
1
1
2
1
2
1
r
Figure 1: An idealized version of µH : R → [0, 1], in which n was fixed to a small constant and the
multiplicative constant inside the exponentiation operator was lowered.
In order to achieve the near uniformity, we modify the procedure for selecting H, the set of high-degree
vertices. Instead of a hard threshold on the degrees of vertices that are included in H as in the sequential
algorithm, we randomize the selection by using a carefully crafted threshold function µH . This function
specifies the probability with which a vertex is included in H. It takes as input the ratio of the vertex's degree
to the current maximum degree (or, more precisely, the current upper bound on the maximum degree) and
it smoothly transitions from 0 to 1 in the neighborhood of the original hard threshold (see Figure 1). The
main intuition behind the introduction of this function is that we want to ensure that a vertex is not selected
for H with almost the same probability, independently of the machine on which it resides. Using a hard
threshold instead of µH could result in the following deficiency. Consider a vertex v that has slightly too
few neighbors to qualify as a heavy vertex. Still, it could happen, with a non-negligible probability, that the
reference set of some machine contains so many neighbors of v that v would be considered heavy on this
machine. However, if v is not included in the set of heavy vertices on that machine, it becomes clear after
even a single phase that the vertex is not on the given machine, i.e. the vertex is on the given machine with
probability zero. At this point the distribution is clearly no longer uniform.
Function µH has further useful properties that we extensively exploit in our analysis. We just note that
in order to ensure near uniformity with high probability, we also have to ensure that each vertex is selected
for F , the set of friends, with roughly the same probability on each machine.
2.4 Organization of the appendix
The appendix presents the details of our techniques. We start by analyzing GlobalAlg in Section 3. Then,
Section 4 describes how to emulate of a single phase of GlobalAlg in the MPC model. Section 5 gives
and analyzes our parallel algorithm by putting together components developed in the previous sections. The
resulting parallel algorithm can be implemented in the MPC model in a fairly straightforward way by using
the result of [GSZ11]. The details of the implementation are given in Section 6.
10
3 Global Algorithm
3.1 Overview
The starting point of our result is a peeling algorithm GlobalAlg that takes as input a graph G, and
removes from it vertices of lower and lower degree until no edge is left. See page 7 for its pseudocode. We
use the term phase to refer to an iteration of the main loop in Lines 2–6.
Each phase is associated with a threshold ∆. Initially, ∆ equals e∆, the upper bound on the maximum
vertex degree. In every phase, ∆ is divided by two until it becomes less than one and the algorithm stops.
Since during the execution of the algorithm we maintain the invariant that the maximum degree in the graph
is at most ∆, the graph has no edge left when the algorithm terminates.
In each phase the algorithm matches, in expectation, a constant fraction of the vertices it removes. We
use this fact to prove that, across all the phases, the algorithm computes a constant-factor approximate
matching.
We now describe in more detail the execution of each phase. First, the algorithm creates H, the set
of vertices that have degree at least ∆/2 (Line 3). We call these vertices heavy. Then, the algorithm uses
randomness to create F , a set of friends (Line 4). Each vertex v is independently included in F with
probability equal to the number of its heavy neighbors divided by 4∆. We show that E [F] = O(H)
and G[H ∪ F ] contains a matching of expected size Ω(H). This kind of matching is likely found by
MatchHeavy in Line 5.
Note that GlobalAlg could as well compute a maximal matching in G[H ∪ F ] instead of calling
MatchHeavy. However, for the purpose of the analysis, using MatchHeavy is simpler, as we can directly
relate the size of the obtained matching to the size of H. In addition, we later give a parallel version of
GlobalAlg, and MatchHeavy is easy to parallelize.
At the end of the phase, vertices in both H and F are removed from the graph, while the matching found
in G[H ∪ F ] is added to the global matching being constructed. It is easy to see, that by removing H, the
algorithm ensures that no vertex of degree larger than ∆/2 remains in the graph, and therefore the bound on
the maximum degree decreases by a factor of two.
3.2 Analysis
We start our analysis of the algorithm by showing that the execution of MatchHeavy in each phase of
GlobalAlg finds a relatively large matching in expectation.
Lemma 3.1. Consider one phase of GlobalAlg. Let H be the set of heavy vertices. MatchHeavy finds
40H.
Proof. Observe that the set E⋆ is a collection of vertex-disjoint stars: each edge connects a red vertex with
a blue vertex and the red vertices have degree 1. Thus, a subset of E⋆ forms a valid matching as long as no
a matching fM such that Eh(cid:12)(cid:12)(cid:12)fM(cid:12)(cid:12)(cid:12)i ≥ 1
blue vertex is incident to two matched edges. Note that this is guaranteed by how edges are added to fM in
The size of the computed matching is the number of blue vertices in H that have at least one incident
edge in E⋆. Let us now lower bound the number of such vertices. Consider an arbitrary u ∈ H. It has the
desired properties exactly when the following three independent events happen: some v ∈ F selects u in
Line 1, u is colored blue, and v is colored red. The joint probability of the two latter events is exactly 1
4 . The
Line 4.
11
probability that u is not selected by some v ∈ F is
1
≤(cid:18)1 −
1
4∆(cid:19)∆/2
This implies that u is selected by a neighbor v ∈ F with probability at least 1
at least 1
4∆(cid:19)N (u)∩V ′
2(cid:19) ≤ exp(cid:18)−
(cid:18)1 −
40 , u is blue and incident to an edge in E⋆. Hence, Eh(cid:12)(cid:12)(cid:12)fM(cid:12)(cid:12)(cid:12)i ≥ 1
Next we show an upper bound on the expected size of F , the set of friends.
≤ exp(cid:18)−
10 · 1
4 = 1
∆
1
4∆ ·
40 H.
1
8(cid:19) ≤
9
10
.
10 . Therefore, with probability
Lemma 3.2. Let H be the set of heavy vertices selected in a phase of GlobalAlg. The following bound
holds on the expected size of F , the set of friends, created in the same phase: E [F] ≤ 1
Proof. At the beginning of a phase, every vertex u ∈ V ′-including those in H-has its degree, N (u)∩V ′,
bounded by ∆. Reversing the order of the summation and applying this fact, we get:
4H.
E [F] = Xv∈V ′
N (v) ∩ H
4∆
= Xu∈H
N (u) ∩ V ′
4∆
≤ H · ∆
4∆
= H
4
.
We combine the last two bounds to lower bound the expected size of the matching computed by GlobalAlg.
Lemma 3.3. Consider an input graph G with an upper bounde∆ on the maximum vertex degree. GlobalAlg(G,e∆)
executes T def= ⌊loge∆⌋ + 1 phases. Let Hi, Fi, and fMi be the sets H, F , and fM constructed in phase i for
i ∈ [T ]. The following relationship holds on the expected sizes of these sets:
TXi=1
Eh(cid:12)(cid:12)(cid:12)fMi(cid:12)(cid:12)(cid:12)i ≥
1
50
TXi=1
E [Hi + Fi]
Proof. For each phase i ∈ [T ], by applying the expectation over all possible settings of the set Hi, we learn
from Lemmas 3.1 and 3.2 that
Eh(cid:12)(cid:12)(cid:12)fMi(cid:12)(cid:12)(cid:12)i ≥
It follows that
1
40
E [Hi]
and
E [Fi] ≤
1
4
E [Hi] .
1
50
E [Hi + Fi] ≤
1
50
E [Hi] +
1
200
E [Hi] =
1
40
and the statement of the lemma follows by summing over all phases.
E [Hi] ≤ Eh(cid:12)(cid:12)(cid:12)fMi(cid:12)(cid:12)(cid:12)i ,
We do not use this fact directly in our paper, but note that the last lemma can be used to show that
GlobalAlg can be used to find a large matching.
Corollary 3.4. GlobalAlg computes a constant factor approximation to the maximum matching with
Ω(1) probability.
Proof. First, note that GlobalAlg finds a correct matching, i.e., no two different edges in M share an
endpoint. This is implied by the fact that M is extended in every phase by a matching on a disjoint set of
vertices.
12
Let T and sets Hi, Fi, and fMi for i ∈ [T ] be defined as in the statement of Lemma 3.3. Let MOPT be a
maximum matching in the graph. Observe that at the end of the algorithm execution, the remaining graph is
empty. This implies that the size of the maximum matching can be bounded by the total number of removed
vertices, because each removed vertex decreases the maximum matching size by at most one:
Hi + Fi ≥ MOPT .
TXi=1
Eh(cid:12)(cid:12)(cid:12)fMi(cid:12)(cid:12)(cid:12)i ≥
Hence, using Lemma 3.3,
E [M] =
TXi=1
Since M ≤ MOPT, M ≥ 1
100 · MOPT + 1 · 1
strictly less than 1
1
50
TXi=1
100 MOPT = 1
E [Hi + Fi] ≥
1
50 MOPT .
100 MOPT with probability at least
100 . Otherwise, E [M] would be
1
50 MOPT, which is not possible.
4 Emulation of a Phase in a Randomly Partitioned Graph
In this section, we introduce a modified version of a single phase (one iteration of the main loop) of
GlobalAlg. Our modifications later allow for implementing the algorithm in the MPC model. The pseu-
docode of the new procedure, EmulatePhase, is presented as Algorithm 3. We partition the vertices of
the current graph into m sets Vi, 1 ≤ i ≤ m. Each vertex is assigned independently and almost uniformly at
random to one of the sets. For each set Vi, we run a subroutine LocalPhase (presented as Algorithm 4).
This subroutine runs a carefully crafted approximate version of one phase of GlobalAlg with an appropri-
ately rescaled threshold ∆. More precisely, the threshold passed to the subroutine is scaled down by a factor
of m, which corresponds to how approximately vertex degrees decrease in subgraphs induced by each of
the sets. The main intuition behind this modification is that we hope to break the problem up into smaller
subproblems on disjoint induced subgraph, and obtain similar global properties by solving the problem ap-
proximately on each smaller part. Later, in Section 5, we design an algorithm that assigns the subproblems
to different machines and solves them in parallel.
Algorithm 3: EmulatePhase(∆, G⋆, m,D)
Emulation of a single phase in a randomly partitioned graph
Input:
• threshold ∆
• induced subgraph G⋆ = (V⋆, E⋆) of maximum degree 3
• number m of subgraphs
• ǫ-near uniform and independent distribution D on assignments of V⋆ to [m]
2 ∆
Output: Remaining vertices and a matching
1 Pick a random assignment Φ : V⋆ → [m] from D
2 for i ∈ [m] do
Vi ← {v ∈ V⋆ : Φ(v) = i}
(V ′
i , Mi) ← LocalPhase(i, G⋆[Vi], ∆/m)
3
4
5 return (Sm
i ,Sm
i=1 V ′
i=1 Mi)
/* LocalPhase = Algorithm 4 */
13
A multiplicative constant used in the exponent of µH :
α def= 96 ln n.
The probability of the selection for a reference set:
µR
def=(cid:0)106 · log n(cid:1)−1
.
The probability of the selection for a heavy set (used with r equal to the ratio of the
estimated degree to the current threshold):
µH (r) def=( 1
2 exp(cid:0) α
2 (r − 1/2)(cid:1)
2 exp(cid:0)− α
1 − 1
2 (r − 1/2)(cid:1)
if r ≤ 1/2,
if r > 1/2.
The probability of the selection for the set of friends (used with r equal to the ratio of
the number of heavy neighbors to the current threshold):
µF (r) def=(max{r/4, 0} if r ≤ 4,
if r > 4.
1
Table 2: Global parameters α ∈ (1,∞) and µR ∈ (0, 1) and functions µH : R → [0, 1] and µF : R → [0, 1]
used in the parallel algorithm. α, µR, and µH depend on n, the total number of vertices in the graph.
We now discuss LocalPhase (i.e., Algorithm 4) in more detail. Table 2 introduces two parameters, α
and µR, and two functions, µH and µF , which are used in LocalPhase. Note first that α is a parameter
used in the definition of µH but it is not used in the pseudocode of LocalPhase (or EmulatePhase)
for anything else. It is, however, a convenient abbreviation in the analysis and the later parallel algorithm.
The other three mathematical objects specify probabilities with which vertices are included in sets that are
created in an execution of LocalPhase.
Apart from creating its own versions of H, the set of heavy vertices, and F , the set of friends, LocalPhase
constructs also a set Ri, which we refer to as a reference set. In Line 1, the algorithm puts each vertex in
Ri independently and with the same probability µR. The reference set is used to estimate the degrees of
other vertices in the same induced subgraph in Line 2. For each vertex vi, its estimate bdv is defined as
the number of v's neighbors in Ri multiplied by µ−1
R to compensate for sampling. Next, in Line 3, the
algorithm uses the estimates to create Hi, the set of heavy vertices. Recall that GlobalAlg uses a sharp
threshold for selecting heavy vertices: all vertices of degree at least ∆/2 are placed in Hi. LocalPhase
works differently. It divides the degree estimate by the current threshold ∆⋆ and uses function µH to decide
with what probability the corresponding vertex is included in Hi. A sketch of the function can be seen in
Figure 1. The function transitions from almost 0 to almost 1 in the neighborhood of 1
2 at a limited pace.
As a result vertices of degrees smaller than, say, 1
4 ∆ are very unlikely to be included in Hi and vertices of
degree greater than 3
4 ∆ are very likely to be included in Hi. GlobalAlg can be seen as an algorithm that
instead of µH , uses a step function that equals 0 for arguments less than 1
2 and abruptly jumps to 1 for larger
arguments. Observe that without µH , the vertices whose degrees barely qualify them as heavy could behave
very differently depending on which set they were assigned to. We use µH to guarantee a smooth behavior
in such cases. That is one of the key ingredients that we need for making sure that a set of vertices that
14
Algorithm 4: LocalPhase(i, Gi, ∆⋆)
Emulation of a single phase on an induced subgraph
Input:
• induced subgraph number i (useful only for the analysis)
• induced subgraph Gi = (Vi, Ei)
• threshold ∆⋆ ∈ R+
Output: Remaining vertices and a matching on Vi
1 Create a reference set Ri by independently selecting each vertex in Vi with probability µR.
2 For each v ∈ Vi, bdv ← N (v) ∩ Ri/µR.
3 Create a set Hi of heavy vertices by independently selecting each v ∈ Vi with probability µH(cid:16)bdv/∆⋆(cid:17).
4 Create a set Fi of friends by independently selecting each vertex in v ∈ Vi with probability
5 Compute a maximal matching Mi in G[Hi ∪ Fi].
6 return (Vi \ (Ri ∪ Hi ∪ Fi), Mi)
µF (N (v) ∩ Hi/∆⋆).
remains on one machine after a phase has almost the same statistical properties as a set of vertices obtained
by new random partitioning.
Finally, in Line 4, LocalPhase creates a set of friends. This step is almost identical to what happens
in the global algorithm. The only difference is that this time we have no upper bound on the number of
heavy neighbors of a vertex. As a result that number divided by 4∆⋆ can be greater than 1, in which case
we have to replace it with 1 in order to obtain a proper probability. This is taken care of by function µF .
Once Hi and Fi have been created, the algorithm finds a maximal matching Mi in the subgraph induced
by the union of these two sets. The algorithm discards from the further consideration not only Hi and Fi,
but also Ri. This eliminates dependencies in the possible distribution of assignments of vertices that have
not been removed yet if we condition this distribution on the configuration of sets that have been removed.
Intuitively, the probability of a vertex's inclusion in any of these sets depends only on Ri and Hi but not on
any other vertices. Hence, once we fix the sets of removed vertices, the assignment of the remaining vertices
to subgraphs is fully independent.2 The output of LocalPhase is a subset of Vi to be considered in later
phases and a matching Mi, which is used to expand the matching that we construct for the entire input
graph. We now introduce additional concepts and notation. They are useful for describing and analyzing
properties of the algorithm. A configuration describes sets Ri, Hi, and Fi, for 1 ≤ i ≤ m, constructed
in an execution of EmulatePhase. We use it for conditioning a distribution of vertex assignments as
described in the previous paragraph. We also formally define two important properties of distributions of
vertex assignments: independence and near uniformity.
Configurations. Let m and V⋆ be the parameters to EmulatePhase: the number of subgraphs and the
set of vertices in the graph to be partitioned, respectively. We say that
C =(cid:0){Ri}i∈[m],{Hi}i∈[m],{Fi}i∈[m](cid:1)
is an m-configuration if it represents a configuration of sets Ri, Hi, and Fi created by EmulatePhase in
the simulation of a phase. Recall that for any i ∈ [m], Ri, Hi, and Fi are the sets created (and removed) by
2By way of comparison, consider observing an experiment in which we toss the same coin twice. The bias of the coin is not
fixed but comes from a random distribution. If we do not know the bias, the outcomes of the coin tosses are not independent.
However, if we do know the bias, the outcomes are independent, even though they have the same bias.
15
the execution of LocalPhase for Vi, the i-th subset of vertices.
We say that a vertex v is fixed by C if it belongs to one of the sets in the configuration, i.e.,
v ∈ [i∈[m]
(Ri ∪ Hi ∪ Fi) .
Conditional distribution. Let D be a distribution on assignments ϕ : V⋆ → [m]. Suppose that we execute
EmulatePhase for D and let C be a non-zero probability m-configuration-composed of sets Ri, Hi, and
Fi for i ∈ [m]-that can be created in this setting. Let V ′
⋆ be the set of vertices in V⋆ that are not fixed by C.
We write D[C] to denote the conditional distribution of possible assignments of vertices in V ′
⋆ to [m], given
that for all i ∈ [m], Ri, Hi, and Fi in C were the sets constructed by LocalPhase for the i-th induced
subgraph.
Near uniformity and independence. Let D be a distribution on assignments ϕ : eV → [m] for some set
eV and m. For each vertex v ∈ eV , let pv : [m] → [0, 1] be the probability mass function of the marginal
distribution of v's assignment. For any ǫ ≥ 0, we say that D is ǫ-near uniform if for every vertex v and
every i ∈ [m], pv(i) ∈J(1 ± ǫ)/mK. We say that D is an independent distribution if the probability of every
assignment ϕ in D equals exactlyQv∈V ′ pv(ϕ(v)).
Concentration inequality. We use the following version of the Chernoff bound that depends on an upper
bound on the expectation of the underlying independent random variables. It can be shown by combining
two applications of the more standard version.
Lemma 4.1 (Chernoff bound). Let X1, . . . , Xk be independently distributed random variables taking values
in [0, 1]. Let X def= X1 +··· + Xk and let U ≥ 0 be an upper bound on the expectation of X, i.e., E[X] ≤ U .
For any δ ∈ [0, 1], Pr(X − E[X] > δU ) ≤ 2 exp(−δ2U/3).
Concise range notation. Multiple times throughout a paper, we want to denote a range around some
value. Instead of writing, say, [x − δ, x + δ], we introduce a more concise notation. In this specific case,
we would simply write Jx ± δK. More formally, let E be a numerical expression that apart from standard
operations also contains a single application of the binary or unary operator ±. We create two standard
numerical expressions from E: E− and E+ that replace ± with − and +, respectively. Now we define
JEK def= [min{E−, E+}, max{E−, E+}].
As another example, consider E = √101 ± 20. We have E− = √101 − 20 = 9 and E+ = √101 + 20 =
11. Henceq√101 ± 20y = [min{9, 11}, max{9, 11}] = [9, 11].
We now show the properties of EmulatePhase that we use to obtain our final parallel algorithm.
4.1 Outline of the section
We start by showing that EmulatePhase computes a large matching as follows. Each vertex belong-
ing to Hi or Fi that EmulatePhase removes in the calls to LocalPhase can decrease the maximum
matching size in the graph induced by the remaining vertices by one. We show that the matching that
EmulatePhase constructs in the process captures on average at least a constant fraction of that loss. We
also show that the effect of removing Ri is negligible. More precisely, in Section 4.2 we prove the following
lemma.
16
Lemma 4.2. Let ∆, G⋆ = (V⋆, E⋆), m, and D be parameters for EmulatePhase such that
• D is an independent and ǫ-near uniform distribution on assignments of vertices V⋆ to [m] for ǫ ∈
[0, 1/200],
m ≥ 4000µ−2
R ln2 n,
• ∆
• the maximum degree of a vertex in G⋆ is at most 3
2 ∆.
For each i ∈ [m], let Hi, Fi, and Mi be the sets constructed by LocalPhase for the i-th induced subgraph.
Then, the following relationship holds for their expected sizes:
Xi∈[m]
E [Hi ∪ Fi] ≤ n−9 + 1200 Xi∈[m]
E [Mi] .
Note that Lemma 4.2 requires that the vertices are distributed independently and near uniformly in the
m sets. This is trivially the case right after the vertices are partitioned independently at random. However,
in the final algorithm, after we partition the vertices, we run multiple phases on each machine. In the rest
of this section we show that running a single phase preserves independence of vertex distribution and only
slightly disturbs the uniformity (Lemma 4.3 and Lemma 4.4). As we have mentioned before, independence
stems from the fact that we use reference sets to estimate vertex degrees. We discard them at the end and
condition on them, which leads to the independence of the distribution of vertices that are not removed.
Lemma 4.3. Let D be an independent distribution of assignments of vertices in V⋆ to [m]. Let C be a non-
zero probability m-configuration that can be constructed by EmulatePhase for D. Let V ′
⋆ be the set of
vertices of V⋆ that are not fixed by C. Then D[C] is an independent distribution of vertices in V ′
⋆ on [m].
Independence of the vertex assignment is a very handy feature that allows us to use Chernoff-like con-
centration inequalities in the analysis of multiple phase emulation. However, although the vertex assignment
of non-removed vertices remains independent across machines from phase to phase, as stated by Lemma 4.3,
their distribution is not necessarily uniform. Fortunately, we can show it is near uniform.
The proof of near uniformity is the most involved proof in this paper. In a nutshell, the proof is structured as
follows. We pick an arbitrary vertex v that has not been removed and show that with high probability it has
the same number of neighbors in all sets Ri. The same property holds for v's neighbors in all sets Hi. We
use this to show that the probability of a fixed configuration of sets removed in a single phase is roughly the
same for all assignments of v to subgraphs. In other words, if v was distributed nearly uniformly before the
execution of EmulatePhase, it is distributed only slightly less uniformly after the execution.
Lemma 4.4. Let ∆, G⋆ = (V⋆, E⋆), m, and D be parameters for EmulatePhase such that
• D is an independent and ǫ-near uniform distribution on assignments of vertices V⋆ to [m] for ǫ ∈
[0, (200 ln n)−1],
m ≥ 4000µ−2
• ∆
R ln2 n.
Let C be an m-configuration constructed by EmulatePhase. With probability at least 1 − n−4 both the
following properties hold:
• The maximum degree in the graph induced by the vertices not fixed in C is bounded by 3
4 ∆.
• D[C] is 60α(cid:16)(cid:0) ∆
m(cid:1)−1/4
+ ǫ(cid:17)-near uniform.
17
4.2 Expected matching size
Now we prove Lemma 4.2, i.e. we show that EmulatePhase computes a large matching. In the proof
we argue that the expected total size of sets Hi and Fi is not significantly impacted by relatively low-degree
vertices classified as heavy or by an unlucky assignment of vertices to subgraphs resulting in local vertex
degrees not corresponding to global degrees. Namely, we show that the expected number of friends a heavy
vertex adds is O(1) and at the same time the probability that the vertex gets matched is Ω(1).
Lemma 4.2. Let ∆, G⋆ = (V⋆, E⋆), m, and D be parameters for EmulatePhase such that
• D is an independent and ǫ-near uniform distribution on assignments of vertices V⋆ to [m] for ǫ ∈
[0, 1/200],
m ≥ 4000µ−2
R ln2 n,
• ∆
• the maximum degree of a vertex in G⋆ is at most 3
2 ∆.
For each i ∈ [m], let Hi, Fi, and Mi be the sets constructed by LocalPhase for the i-th induced subgraph.
Then, the following relationship holds for their expected sizes:
Xi∈[m]
E [Hi ∪ Fi] ≤ n−9 + 1200 Xi∈[m]
E [Mi] .
Proof. We borrow more notation from EmulatePhase and the m executions of LocalPhase initiated
by it. For i ∈ [m], Vi is the set inducing the i-th subgraph. Value ∆⋆ = ∆
m is the rescaled threshold passed
to the executions of LocalPhase. Ri is the reference set created by LocalPhase for the i-th induced
subgraph.
For each induced subgraph, LocalPhase computes a maximal matching Mi in Line 5. While such a
matching is always large-its size is at least half the maximum matching size-it is hard to relate its size
directly to the sizes of Hi and Fi. Therefore, we first analyze the size of a matching that would be created
by MatchHeavy(G⋆[Hi ∪ Fi], Hi, Fi). We refer to this matching as fMi and we later use the inequality
(cid:12)(cid:12)(cid:12)fMi(cid:12)(cid:12)(cid:12) ≤ 2Mi.
We partition each Hi, i ∈ [m], into two sets: H ′
Si∈[m] Hi? Suppose that v ∈ Vk, where k ∈ [m]. The expected number of v's neighbors in Rk is at
most (1 + ǫ) · µR · 1
independence, Lemma 4.1, and the lower bound on ∆⋆, we obtain the following bound:
i is the subset of vertices in Hi of degree
i. We start by bounding the expected
8 ∆ is included in
16 µR∆⋆ due to the independence and ǫ-near uniformity of D[C]. Using the
i. What is the probability that a given vertex v of degree less than 1
less than 1
total size of sets H ′
i,t+1 is its complement, i.e., H ′′
i
i . H ′
def= Hi \ H ′
8 ∆/m ≤ 3
8 ∆ in G⋆. H ′′
i and H ′′
1
1
4
3 ·(cid:18) 1
3(cid:19)2
µR∆⋆(cid:21) ≤ 2 exp −
µR∆⋆! ≤ 2 exp (−27 ln n) = 2n−27.
Pr(cid:20)µRbdv >
4 ∆⋆, the probability that v is selected to Hk is at most µH(bdv/∆⋆) ≤ µH(1/4) ≤ 1
E [H ′
If bdv ≤ 1
thatPi∈[m]
We also partition the sets of friends, Fi for i ∈ [m], into two sets each: F ′
2 n−12. Hence
2 n−12 ≤ n−12. This implies
i and F ′′
i . This partition is
based on the execution of MatchHeavy for the i-th subgraph. In Line 1, this algorithm selects for every
v is selected to Hk-and therefore to H ′
i] ≤ n · n−12 = n−11.
k-with probability at most 2n−27 + 1
3
16
·
18
i . Obviously, a heavy vertex in H ′
vertex v ∈ Fi a random heavy neighbor v⋆ ∈ Hi. If v⋆ ∈ H ′
we assign v to F ′′
inequality and the upper bound onPi∈[m]
most n−11. Even if for all i ∈ [m], all vertices in Fi select a heavy neighbor in H ′
the total expected number of vertices in sets F ′
i, we assign v to F ′
i can be selected only if H ′
i . Analogously, if v⋆ ∈ H ′′
i ,
i is non-empty. By Markov's
i is non-empty is at
i whenever it is available,
Before we proceed to bounding sizes of the remaining sets, we prove that with high probability, all
vertices have a number of neighbors close to the expectation. Let ϕ : V⋆ → [m] be the assignment of
vertices to subgraphs. We define E as the event that for all v ∈ V⋆,
1
i], the probability that at least one set H ′
E [H ′
Eh(cid:12)(cid:12)(cid:12)F ′
i,t+1(cid:12)(cid:12)(cid:12)i ≤ n · n−11 = n−10.
i is at mostPi∈[m]
m N (v) ∩ V⋆ −(cid:12)(cid:12)N (v) ∩ Vϕ(v)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤
m N (v) ∩ V⋆ − E(cid:2)(cid:12)(cid:12)N (v) ∩ Vϕ(v)(cid:12)(cid:12)(cid:3)(cid:12)(cid:12)(cid:12)(cid:12) ≤ ǫ ·
∆
m ≤
3
400
2 + 3
1
16
(cid:12)(cid:12)(cid:12)(cid:12)
∆⋆.
∆⋆.
3
2
1
(cid:12)(cid:12)(cid:12)(cid:12)
Consider first one fixed v ∈ V⋆. The degree of v in G⋆ is N (v) ∩ V⋆ ≤ 3
2 ∆. Due to the near-uniformity
and independence,
Lemma 4.1, and the lower bound on ∆⋆ (i.e., ∆⋆ = ∆
This in particular implies that E(cid:2)(cid:12)(cid:12)N (v) ∩ Vϕ(v)(cid:12)(cid:12)(cid:3) ≤(cid:0) 3
Pr(cid:20)(cid:12)(cid:12)E(cid:2)(cid:12)(cid:12)N (v) ∩ Vϕ(v)(cid:12)(cid:12)(cid:3) −(cid:12)(cid:12)N (v) ∩ Vϕ(v)(cid:12)(cid:12)(cid:12)(cid:12) >
1
20
m ≥ 4000µ−2
400(cid:1) ∆⋆ ≤ 2∆⋆. Using the independence of D,
∆⋆(cid:21) ≤ 2 exp −
R ln2 n = 4 · 1015 · ln4 n),
3 ·(cid:18) 1
2(cid:19)2
≤ 2 exp(cid:0)−(1012 + 3) ln n(cid:1)
≤ n−(1012+2) ≤ n−12.
· 2∆⋆!
20 ·
1
1
As a result, with this probability, we have
1
∆⋆.
(cid:12)(cid:12)(cid:12)(cid:12)
1
20
1
16
∆⋆ +
3
400
∆⋆ ≤
m N (v) ∩ V⋆ −(cid:12)(cid:12)N (v) ∩ Vϕ(v)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤
By the union bound, this bound holds for all vertices in V⋆ simultaneously-and hence E occurs-with
probability at least 1 − n · n−12 = 1 − n−11.
i andPi∈[m] F ′′
If E does not occur, we can bound bothPi∈[m] H ′′
i by n. This contributes at most
n−11·n = n−10 to the expected size of each of these quantities. Suppose now that E occurs. Consider an arbi-
16 ∆⋆(cid:3) ⊆
i for some i. The number of neighbors of v in Vi lies in the range(cid:2) 1
trary v ∈ H ′′
16 ∆⋆, 2∆⋆(cid:3). Moreover, the expected number of vertices w ∈ F ′′
(cid:2) 1
that select v in w⋆ in Line 1 of
i ] ≤ 1
2 . It follows that E [F ′′
MatchHeavy is bounded by 2∆⋆ ·
i ], given E. We now
lower bound the expected size of fMi given E. What is the probability that some vertex w ∈ Fi selects v as
w⋆ in MatchHeavy and (v, w) is added tofMi?
This occurs if one of v's neighbors w is added to Fi and selects v as w⋆, and additionally, v and w are
colored blue and red, respectively. The number of v's neighbors is at least 1
16 ∆⋆. Since each vertex w in
Vi has at most 2∆⋆ neighbors, the number of heavy neighbors of w is bounded by the same number. This
implies that in the process of selecting Fi, only the first branch in the definition of µF is used and each
vertex w is included with probability exactly equal to the number of its neighbors in Hi divided by 4∆t+1.
Then each heavy neighbor of w is selected as w⋆ with probability one over the number of heavy neighbors
8 ∆⋆ − 1
E [H ′′
2 ∆⋆ + 1
16 ∆⋆, 3
= 1
4∆⋆
1
2
i
19
of w. What this implies is that each neighbor w of v is selected for Fi and selects v as w⋆ with probability
exactly (4∆⋆)−1. Hence the probability that v is not selected as w⋆ by any of its at least 1
16 ∆⋆ neighbors w
can be bounded by
16 ∆⋆
(cid:18)1 −
1
4∆⋆(cid:19) 1
≤ exp(cid:18)−
1
4∆⋆ ·
1
16
∆⋆(cid:19) = e−1/64.
Therefore the probability that v is selected by some vertex w ∈ Fi as w⋆ is at least 1 − e−1/64 ≥ 1/100.
Then with probability 1/4, these two vertices have appropriate colors and this or another edge incident to
v with the same properties is added to fMi. In summary, the probability that an edge (v, w) for some w as
described is added to fMi is at least 1/400. Since we do not count any edge in the matching twice for two
heavy vertices, by the linearity of expectation Eh(cid:12)(cid:12)(cid:12)fMi(cid:12)(cid:12)(cid:12)i ≥ 1
E(cid:2)(cid:12)(cid:12)H ′′
i ] given E. Overall, given E, we have
Xi∈[m]
E [H ′′
400
In general, without conditioning on E,
We now combine bounds on all terms to finish the proof of the lemma.
3
Eh(cid:12)(cid:12)(cid:12)fMi(cid:12)(cid:12)(cid:12)i .
2 Xi∈[m]
i(cid:12)(cid:12)(cid:3) ≤ 600 Xi∈[m]
E(cid:2)(cid:12)(cid:12)H ′′
i(cid:12)(cid:12) +(cid:12)(cid:12)F ′′
i(cid:12)(cid:12)(cid:3) ≤
Eh(cid:12)(cid:12)(cid:12)fMi(cid:12)(cid:12)(cid:12)i .
Xi∈[m]
i(cid:12)(cid:12)(cid:3) ≤ 2 · n−10 + 600 Xi∈[m]
E(cid:2)(cid:12)(cid:12)H ′′
i(cid:12)(cid:12) +(cid:12)(cid:12)F ′′
E [Hi ∪ Fi] ≤ Xi∈[m]
i(cid:12)(cid:12) +(cid:12)(cid:12)H ′′
i(cid:12)(cid:12) +(cid:12)(cid:12)F ′
E(cid:2)(cid:12)(cid:12)H ′
i(cid:12)(cid:12) +(cid:12)(cid:12)F ′′
i(cid:12)(cid:12)(cid:3)
Eh(cid:12)(cid:12)(cid:12)fMi(cid:12)(cid:12)(cid:12)i
≤ n−11 + n−10 + 2n−10 + 600 Xi∈[m]
≤ n−9 + 1200 Xi∈[m]
E [Mi] .
Xi∈[m]
4.3 Independence
Next we prove Lemma 4.3. We start with an auxiliary lemma that gives a simple criterion under which
an independent distribution remains independent after conditioning on a random event. Consider a random
vector with independently distributed coordinates. Suppose that for any value of the vector, a random event
E occurs when all coordinates "cooperate", where each coordinate cooperates independently with probabil-
ity that depends only on the value of that coordinate. We then show that the distribution of the vector's
coordinates given E remains independent.
Lemma 4.5. Let k be a positive integer and A an arbitrary finite set. Let X = (X1, . . . , Xk) be a random
vector in Ak with independently distributed coordinates. Let E be a random event of non-zero probability.
If there exist functions pi : A → [0, 1], for i ∈ [k], such that for any x = (x1, . . . , xk) ∈ Ak appearing with
non-zero probability,
then the conditional distribution of coordinates in X given E is independent as well.
Pr[EX = x] =
pi(xi),
kYi=1
20
Proof. Since the distribution of coordinates in X is independent, there are k probability mass functions
p′
i(xi). The
probability of E can be expressed as
i : A → [0, 1], i ∈ [k], such that for every x = (x1, . . . , xk) ∈ Ak, Pr[X = x] = Qk
i=1 p′
Pr[E] = Xx=(x1,...,xk)∈Ak
= Xx=(x1,...,xk)∈Ak
Pr[E ∧ X = x] = Xx=(x1,...,xk)∈Ak
kYi=1
kYi=1Xy∈A
pi(xi)p′
i(xi) =
Pr[X=x]>0
pi(y)p′
i(y).
Pr[EX = x] · Pr[X = x]
Note that since the probability of E is positive, each multiplicative termPy∈A pi(y)p′
i(y), i ∈ [k], in the
above expression is positive. We can express the probability of any vector x = (x1, . . . , xk) ∈ Ak given E
as follows:
Pr[X = xE] =
Pr[E ∧ X = x]
=
Pr[EX = x] · Pr[X = x]
Pr[E]
Pr[E]
i=1 pi(xi)p′
i(xi)
= Qk
Qk
i=1Py∈A pi(y)p′
kYi=1
i(xi)/Py∈A pi(y)p′
i(y)
=
pi(xi)p′
i(xi)
.
i(y)
Py∈A pi(y)p′
i(y) for each i ∈ [k]. Each p′′
We define p′′
i : A → [0, 1] as p′′
probability mass function on A. As a result we have Pr[X = xE] =Qk
i is a valid
i (xi), which proves that the
distribution of coordinates in X given E is still independent with each coordinate distributed according to
its probability mass function p′′
i .
i (x) def= pi(xi)p′
i=1 p′′
We now prove Lemma 4.3 by applying Lemma 4.5 thrice. We refer to functions pi, which describe the
probability of each coordinate cooperating, as cooperation probability functions.
Lemma 4.3. Let D be an independent distribution of assignments of vertices in V⋆ to [m]. Let C be a non-
zero probability m-configuration that can be constructed by EmulatePhase for D. Let V ′
⋆ be the set of
vertices of V⋆ that are not fixed by C. Then D[C] is an independent distribution of vertices in V ′
⋆ on [m].
Proof. C can be expressed as
C =(cid:0){R⋆
i }i∈[m],{H ⋆
i }i∈[m],{F ⋆
i }i∈[m](cid:1)
i , H ⋆
i , and F ⋆
for some subsets R⋆
i of V⋆, where i ∈ [m]. We write Φ to denote the random assignment of
vertices to sets selected in Line 1 of EmulatePhase. Φ is a random variable distributed according to D.
Let ER be the event that for all i ∈ [m], the reference set Ri generated for the i-th induced subgraph by
LocalPhase equals exactly R⋆
i . A vertex v that is assigned to a set Vi is included in Ri with probability
exactly µR, independently of other vertices. Hence once we fix an assignment ϕ : V⋆ → [m] of vertices to
sets Vi, we can express the probability of ER as a product of probabilities that each vertex cooperates. More
as follows.
formally, Pr[ERΦ = ϕ] =Qv∈V⋆ qv(ϕ(v)) for cooperation probability functions qv : [m] → [0, 1] defined
• If v ∈Si∈[m] R⋆
i , there is exactly one i ∈ [m] such that v ∈ R⋆
i . If v is not assigned to Vi, ER cannot
occur. If it is, v cooperates with ER with probability exactly µR, i.e., the probability of the selection
for Ri. For this kind of v, the cooperation probability function is
qv(i) def=(µR if v ∈ R⋆
i ,
if v 6∈ R⋆
i .
0
21
• If v 6∈Si∈[m] R⋆
i , v cooperates with ER if it is not selected for Rϕ(v), independently of its assignment
ϕ(v), which happens with probability exactly 1 − µR. Therefore, the cooperation probability can be
defined as qv(i) def= 1 − µR for all i ∈ [m].
We invoke Lemma 4.5 to conclude that the conditional distribution of values of Φ given ER is independent
as well.
We now define an event EH that both ER occurs and for all i ∈ [m], Hi, the set of heavy vertices
constructed for the i-th subgraph equals exactly H ⋆
i . We want to show that the conditional distribution of
values of Φ given EH is independent. Note that if Φ is selected from the conditional distribution given ER
(i.e., all sets Ri are as expected) and we fix the assignment φ : V⋆ → [m] of vertices to sets Vi, then each
vertex v ∈ V⋆ is assigned to Hφ(v)-this the only set Hi to which it can be assigned-independently of
other vertices. As a result, we can express the probability of EH given ER and ϕ being the assignment as a
q′
v(ϕ(v))
for cooperation probability functions q′
v : [m] → [0, 1] defined as follows, where ∆⋆ is the threshold used
in the m executions of LocalPhase.
product of cooperation probabilities for each vertex. More precisely, Pr[EHΦ = ϕ,ER] =Qv∈V⋆
• If v ∈ Si∈[m] H ⋆
i . EH can only occur if v is in-
cluded in the corresponding Hi. This cannot happen if v is not assigned to the corresponding Vi
by ϕ.
If v is assigned to this Vi, it has to be selected for Hi, which happens with probability
µH (N (v) ∩ R⋆
i /(µR∆⋆)). The cooperation probability function can be written in this case as
i , then there is exactly one i such that v ∈ H ⋆
v(i) def=(µH(N (v) ∩ R⋆
0
q′
i /(µR∆⋆))
if v ∈ H ⋆
i ,
if v 6∈ H ⋆
i .
• If v 6∈Si∈[m] H ⋆
EH to occur. This happens with probability 1 − µH(N (v) ∩ R⋆
v(i) def= 1 − µH(N (v) ∩ R⋆
q′
i , v cannot be included in Hi corresponding to the set Vi to which it is assigned for
i /(µR∆⋆)). Hence, we can define
i /(µR∆⋆)) for all i ∈ [m].
We can now invoke Lemma 4.5 to conclude that the distribution of values of Φ given EH is independent.
Finally, we define EF to be the event that both EH occurs and for each i ∈ [m], Fi, the set of friends
selected for the i-th induced subgraph, equals exactly F ⋆
i . We observe that once Φ is fixed to a specific
assignment ϕ : V⋆ → [m] and EH occurs (i.e., all sets Ri and Hi are as in C), then each vertex is indepen-
dently included in Fϕ(v) with some specific probability that depends only on Hϕ(v), which is already fixed.
In this setting, we can therefore express the probability of EF , which exactly specifies the composition of
sets Fi, as a product of values provided by some cooperation probability functions q′′
v : [m] → [0, 1]. More
v (ϕ(v)) for q′′
q′′
v that we define next.
i , then there is exactly one i such that v ∈ F ⋆
i . EF cannot occur if v is not assigned to
precisely, Pr[EFΦ = ϕ,EH ] =Qv∈V⋆
• If v ∈Si∈[m] F ⋆
Vi and selected for Fi. Hence, the cooperation probability function for v is
if v ∈ F ⋆
i ,
if v 6∈ F ⋆
i .
v (i) def=(µF (N (v) ∩ H ⋆
i /∆⋆)
q′′
0
• If v 6∈Si∈[m] F ⋆
EF to occur. Hence, q′′
v (i) def= 1 − µF (N (v) ∩ H ⋆
i⋆,t/∆t).
i , to whichever set Vi vertex v is assigned, it should not be included in Fi in order for
We invoke Lemma 4.5 to conclude that the distribution of values of Φ given EF is independent as well. This
is a distribution on assignments for the entire set V⋆. If we restrict it to assignments of V ′
⋆ ⊆ V⋆, we obtain
a distribution that first, is independent as well, and second, equals exactly D[C].
22
4.4 Near Uniformity
In this section we prove Lemma 4.4. We begin by showing a useful property of µH (see Table 2 for defini-
tion). Recall that GlobalAlg selects H, the set of heavy vertices, by taking all vertices of degree at least
∆/2. In LocalPhase the degree estimate of each vertex depends on the number of neighbors in the refer-
ence set in the vertex's induced subgraph. We want the decision taken for each vertex to be approximately
the same, independently of which subgraph it is assigned to. Therefore, we use µH-which specifies the
probability of the inclusion in the set of heavy vertices-which is relatively insensitive to small argument
changes. The next lemma proves that this is indeed the case. Small additive changes to the parameter x to
µH have small multiplicative impact on both µH(x) and 1 − µH(x).
Lemma 4.6 (Insensitivity of µH ). Let δ ∈ [0, (α/2)−1] = [0, (48 ln n)−1]. For any pair x and x′ of real
numbers such that x − x′ ≤ δ,
and
Proof. We define an auxiliary function f : R → [0, 1]:
µH(x′) ∈JµH(x)(1 ± αδ)K
1 − µH(x′) ∈J(1 − µH(x))(1 ± αδ)K .
f (r) def=( 1
2 exp(cid:0) α
2 r(cid:1)
if r ≤ 0,
2 exp(cid:0)− α
2 r(cid:1)
1 − 1
if r > 0.
It is easy to verify that for all r ∈ R, µH(r) = f (r − 1/2) and 1 − µH(r) = f (−(r − 1/2)). Therefore, in
order to prove the lemma, it suffices to prove that for any r and r′ such that r − r′ ≤ δ,
f (r)(1 − αδ) ≤ f (r′) ≤ f (r)(1 + αδ),
(1)
i.e., a small additive change to the argument of f has a limited multiplicative impact on the value of f .
Note that f is differentiable in both (−∞, 0) and (0,∞). Additionally, it is continuous in the entire
range-the left and right branch of the function meet at 0-and both the left and right derivatives at 0 are
equal. This implies that it is differentiable at 0 as well. Its derivative is
f ′(r) =( α
α
2 r(cid:1)
4 · exp(cid:0) α
4 · exp(cid:0)− α
2 r(cid:1)
if r ≤ 0,
if r > 0,
which is positive for all r, and therefore, f is strictly increasing. Note that f ′ is increasing in (−∞, 0] and
decreasing in [0,∞). Hence the global maximum of f ′ equals f ′(0) = α/4.
In order to prove Inequality 1 for all r and r′ such that r − r′ ≤ δ, we consider two cases. Suppose
first that r ≥ 0. By the upper bound on the derivative of f ,
f (r) −
α
4 · r − r′ ≤ f (r′) ≤ f (r) +
α
4 · r − r′.
Since r ≥ 0, f (r) ≥ 1/2. This leads to
α
2 · r − r′ ≤ f (r′) ≤ f (r) + f (r) ·
f (r) − f (r) ·
α
2 · r − r′.
By the bound on r − r′,
f (r)(1 − αδ) ≤ f (r′) ≤ f (r)(1 + αδ),
23
Suppose now that r < 0. Since f is increasing, it suffices to bound the value of f from below at r − δ
which finishes the proof in the first case.
and from above and at r + δ. For r − δ, we obtain
f (r − δ) =
1
2
exp(cid:16) α
≥ f (r)(cid:16)1 −
2
α
2
For r + δ, let us first define a function g : R → R as
1
2
g(y) def=
α
2
δ(cid:17)
(r − δ)(cid:17) = f (r) exp(cid:16)−
δ(cid:17) ≥ f (r)(1 − αδ).
exp(cid:16) α
(r + δ)(cid:17) = f (r) · exp(cid:16) α
y(cid:17) .
2
1
2
exp(cid:16) α
2
For y ≤ 0, f (y) = g(y). For y > 0, g′(y) ≥ f ′(y) and hence, for any y ∈ R, g(y) ≥ f (y). As a result, we
obtain
f (r + δ) ≤ g(r + δ) =
By the bound on δ in the lemma statement, α
2 δ ≤ 1. It follows from the convexity of the exponential function
that for any y ∈ [0, 1], exp(y) ≤ y · exp(1) + (1 − y) · exp(0) ≤ 3y + (1 − y) = 1 + 2y. Continuing the
reasoning,
2
δ(cid:17) .
which finishes the proof of Inequality (1).
f (r + δ) ≤ f (r) ·(cid:16)1 + 2 ·
α
2
δ(cid:17) = f (r)(1 + αδ),
The main result of this section is Lemma 4.4 that states that if a distribution D of vertex assignments
is near uniform, then EmulatePhase constructs a configuration C such that D[C] is near uniform as well,
and also, the maximum degree in the graph induced by the vertices not removed by EmulatePhase is
bounded.
Lemma 4.4. Let ∆, G⋆ = (V⋆, E⋆), m, and D be parameters for EmulatePhase such that
• D is an independent and ǫ-near uniform distribution on assignments of vertices V⋆ to [m] for ǫ ∈
[0, (200 ln n)−1],
m ≥ 4000µ−2
• ∆
R ln2 n.
Let C be an m-configuration constructed by EmulatePhase. With probability at least 1 − n−4 both the
following properties hold:
• The maximum degree in the graph induced by the vertices not fixed in C is bounded by 3
4 ∆.
• D[C] is 60α(cid:16)(cid:0) ∆
m(cid:1)−1/4
+ ǫ(cid:17)-near uniform.
Proof overview (of Lemma 4.4). This is the most intricate proof of the entire paper. We therefore provide
a short overview. First, we list again the variables in EmulatePhase and LocalPhase to which we
refer in the proof and define additional convenient symbols. Then we introduce five simple random events
(Events 1–5) that capture properties needed to prove Lemma 4.4. In Claim 4.7, we show that the proba-
bility of all these events occurring simultaneously is high. The proof of the claim follows mostly from a
repetitive application of the Chernoff bound. In the next claim, Claim 4.8, we show that the occurrence of
24
all the events has a few helpful consequences. First, high degree vertices get removed in the execution of
EmulatePhase (which is one of our final desired properties). Second, each vertex v that is not fixed in
C has a very similar number of neighbors in all sets Ri and it has a very similar number of neighbors in all
sets Hi. In the final proof of Lemma 4.4, we use the fact that this implies that to whichever set Vi vertex
v was assigned in EmulatePhase, the probability of its removal in EmulatePhase was more or less
the same. This leads to the conclusion that if v was distributed nearly uniformly in D, it is distributed only
slightly less uniformly in D[C].
Notation. To simplify the presentation, for the rest of Section 4.4, we assume that ∆, G⋆ = (V⋆, E⋆),
m, and D are the parameters to EmulatePhase as in the statement of Lemma 4.4. Additionally, for
each i ∈ [m], Ri, Hi, and Fi are the sets constructed by LocalPhase for the i-th subgraph in the ex-
ecution of EmulatePhase. We also write C to denote the corresponding m-configuration, i.e., C =
subgraph to which it was assigned. This estimate is computed in Line 2 of LocalPhase. We also use ∆⋆
to denote the rescaled threshold passed in all calls to LocalPhase, i.e., ∆⋆ = ∆
m .
(cid:0){Ri}i∈[m],{Hi}i∈[m],{Fi}i∈[m](cid:1). Furthermore, for each v ∈ V⋆, bdv is the estimate of v's degree in the
We also introduce additional notation, not present in EmulatePhase or LocalPhase. For each
def= N (v) ∩ V⋆, i.e., dv is the degree of v in G⋆. For each vertex v ∈ V⋆, we also introduce a
v ∈ V⋆, dv
def= µH(dv/∆), which can be seen as a very rough approximation of v's probability
notion of its weight: wv
of being selected for the set of heavy vertices. For any v ∈ V⋆ and U ⊆ V⋆, we also introduce notation for
the total weight of v's neighbors in U :
Wv(U ) def= Xu∈N (v)∩U
wu.
Finally, for all i ∈ [m] and v ∈ V⋆, we also introduce a slightly less intuitive notion of the expected number
of heavy neighbors of v in the i-th subgraph after the degree estimates are fixed in Line 2 of LocalPhase
and before vertices are assigned to the heavy set in Line 3:
hv,i
def= Xu∈N (v)∩Vi
Obviously, each hv,i is a random variable.
µH(cid:16)bdu/∆⋆(cid:17) .
Convenient random events. We now list five random events that we hope all to occur simultaneously with
high probability. The first event intuitively is the event that high-degree vertices are likely to be included in
the set of heavy vertices in Line 3 of LocalPhase.
Event 1
For each vertex v ∈ V⋆ such that dv ≥ 3
µH(cid:16)bdv/∆⋆(cid:17) ≥ 1 −
Another way to define this event would be to state that bdv for such vertices v is high, but this form is more
suitable for our applications later. The next event is the event that all such vertices are in fact classified as
4 ∆,
1
2
n−6.
25
heavy.
Each vertex v ∈ V⋆ such that dv ≥ 3
Event 2
4 ∆ belongs toSi∈[m] Hi.
The next event is the event that low-degree vertices have a number of neighbors in each set Ri close to the
mean. This implies that if we were able to move a low-degree vertex v to Vi, for any i ∈ [m], its estimated
degree bdv would not change significantly.
Event 3
For each vertex v ∈ V⋆ such that dv < 3
4 ∆ and each i ∈ [m],
1
µR N (v) ∩ Ri −
(cid:12)(cid:12)(cid:12)(cid:12)
⋆ +
ǫ∆⋆.
3
4
dv
m(cid:12)(cid:12)(cid:12)(cid:12) ≤ ∆3/4
As a reminder, we use Wv(U ) to denote the expected number of vertices in N (v) ∩ U that are selected
as heavy, where every vertex u is selected with respect to its global degree du. The next event shows that
Wv(Vi) does not deviate much from its mean.
Event 4
For each vertex v ∈ V⋆ such that dv < 3
4 ∆ and each i ∈ [m],
Wv(Vi) − Wv(V⋆)/m ≤ ∆3/4
⋆ +
3
4
ǫ∆⋆.
Recall that hv,i intuitively expresses the expected number of v's neighbors in the i-th induced subgraph at
some specific stage in the execution of LocalPhase for the i-th induced subgraph. The final event is the
event that for all bounded hv,i, the actual number of v's neighbors in Hi does not deviate significantly from
hv,i.
Event 5
For each vertex v ∈ V⋆ and each i ∈ [m], if hv,i ≤ 2∆⋆, then
N (v) ∩ Hi − hv,i ≤ ∆3/4
⋆
.
High probability of the random events. We now show that the probability of all the events occurring is
high. The proof follows mostly via elementary applications of the Chernoff bound.
Claim 4.7. If ǫ ∈ [0, 1/100] and ∆
ity at least 1 − n−4.
m ≥ 4000µ−2
R ln2 n, then Events 1–5 occur simultaneously with probabil-
R ln2 n = 4 · 1015 · ln4 n.
In the proof of the lemma, we extensively use the fact that ∆⋆ = ∆
Proof. We consider all events in order and later show by the union bound that all of them hold simulta-
m ≥
neously with high probability.
4000µ−2
First, we consider Event 1 and Event 2, which we handle together. Consider a vertex v such that dv ≥
4 ∆. Let i⋆ be the index of the set to which it is assigned. Since D is ǫ-near uniform, the expectation of
N (v) ∩ Ri⋆, the number of v's neighbors in Ri⋆, is at least (1 − ǫ) 3
400 µR∆⋆. Since vertices are
both assigned to machines independently and included in the reference set independently as well, we can
4 µR
∆
m ≥ 297
3
26
·
1
1
405
297
400
8 µR∆⋆ is at most
10 · 297
2 exp −
apply Lemma 4.1 to bound the deviation with high probability. The probability that the number of neighbors
is smaller than 9
400 µR∆⋆ ≥ 5
3 ·(cid:18) 1
10(cid:19)2
Hence with probability at least 1 − 1
v is not included in the set of heavy vertices in Line 3 of LocalPhase with probability at most 1
µR∆⋆(cid:19) ≤ 2n−9 ≤
8 ∆⋆ and µH(cid:16)bdv/∆⋆(cid:17) ≥ 1 − 1
µR∆⋆! ≤ 2 exp(cid:18)−
2 n−6, bdv ≥ 5
Therefore, v has the desired value of µH(cid:16)bdv/∆⋆(cid:17) and belongs to Hi⋆ with probability at least 1 − n−6. By
the union bound, this occurs for all high degree vertices with probability at least 1− n−5, in which case both
We now show that Event 3 occurs with high probability. Let v be an arbitrary vertex such that dv < 3
4 ∆
def= N (v) ∩ Ri. Xv,i is a random variable. Since D is ǫ-near uniform, E [Xv,i] ∈
and let i ∈ [m]. Let Xv,i
J(1 ± ǫ)µRdv/mK. In particular, due to the bounds on dv and ǫ, E[Xv,i] ≤ µR∆⋆. Due to the independence,
we can use Lemma 4.1 to bound the deviation of Xv,i from its expectation. We have
2 n−6. If this is the case,
2 n−6.
Event 1 and Event 2 occur.
n−6.
1
2
Pr(cid:16)Xv,i − E[Xv,i] > µR∆3/4
1
1
3
∆1/4
µR∆1/2
3 · 1
⋆ (cid:17) ≤ 2 exp−
· µR∆⋆
⋆ !2
⋆ (cid:19) ≤ 2n−21.
= 2 exp(cid:18)−
m(cid:12)(cid:12)(cid:12)(cid:12) ≤ µR∆3/4
m(cid:12)(cid:12)(cid:12)(cid:12) ≤ Xv,i − E[Xv,i] +(cid:12)(cid:12)(cid:12)(cid:12)E[Xv,i] − µR
m(cid:12)(cid:12)(cid:12)(cid:12) ≤ ∆3/4
1
µR N (v) ∩ Ri −
m(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)
≤ µR∆3/4
⋆ + ǫµR
ǫµR∆⋆.
⋆ +
⋆ +
ǫ∆⋆.
3
4
3
4
dv
dv
dv
dv
(cid:12)(cid:12)(cid:12)(cid:12)Xv,i − µR
(cid:12)(cid:12)(cid:12)(cid:12)
Xv,i
µR −
dv
m
We now move on to Event 4. Consider a vertex v such that dv < 3
By the union bound, this holds for all v and i of interest-and therefore, Event 3 occurs-with probability
at least 1 − V⋆ · m · 2n−21 ≥ 1 − n−5.
4 ∆ and i ∈ [m]. Note that since
the weight of every vertex is at most 1, Wv(V⋆)/m ≤ dv/m < 3
4 ∆⋆. Since D[C] is ǫ-near uniform,
E [Wv(Vi)] ∈ J(1 ± ǫ)Wv(V⋆)/mK. In particular, E [Wv(Vi)] ≤ 101
4 ∆⋆ ≤ ∆⋆. Since
vertices are assigned to machines independently, we can apply Lemma 4.1 to bound the deviation of Wv(Vi)
from the expectation:
⋆ (cid:17) ≤ 2 exp−
= 2 exp(cid:18)−
· ∆⋆
3 · 1
⋆ !2
⋆ (cid:19) ≤ 2n−21.
Pr(cid:16)Wv(Vi) − E [Wv(Vi)] > ∆3/4
100 Wv(V⋆)/m ≤ 101
100 · 3
∆1/4
∆1/2
1
3
1
27
Hence with probability 1 − 2n−21, we have
By dividing both sides by µR, we obtain the desired bound
As a result, with probability at least 1 − 2n−21,
Wv(Vi) − Wv(V⋆)/m ≤ Wv(Vi) − E [Wv(Vi)] + E [Wv(Vi)] − Wv(V⋆)/m
3
4
⋆ + ǫWv(V⋆)/m ≤ ∆3/4
⋆ + ǫdv/m ≤ ∆3/4
≤ ∆3/4
⋆ +
ǫ∆⋆.
By the union bound, this holds for all v and i of interest-and therefore, Event 4 occurs-with probability
at least 1 − V⋆ · m · 2n−21 ≥ 1 − n−5.
To show that Event 5 occurs with high probability, recall first that hv,i is the expected number of v's
neighbors to be added in Line 3 to Hi in the execution of LocalPhase for the i-th subgraph. Note that
the decision of adding a vertex to Hi is made independently for each neighbor of v. Fix a v ∈ V⋆ and
i ∈ [m] such that hv,i ≤ 2∆⋆. We apply Lemma 4.1 to bound the probability of a large deviation from the
expectation:
Pr(cid:16)N (v) ∩ Hi − hv,i > ∆3/4
⋆ (cid:17) ≤ 2 exp−
= 2 exp(cid:18)−
1
6
1
2∆1/4
· 2∆⋆
⋆ !2
3 · 1
⋆ (cid:19) ≤ 2n−10.
∆1/2
By the union bound the probability that this bound does not hold for some v and i such that hv,i ≤ 2∆⋆ is
by the union bound at most V⋆· m· 2n−10 ≤ n−5. Hence, Event 5 occurs with probability at least 1− n−5.
In summary, Events 1–5 occur simultaneously with probability at least 1− 4· n−5 ≥ 1− n−4 by another
application of the union bound.
Consequences of the random events. We now show that if all the random events occur, then a few helpful
properties hold for every vertex v that is not fixed by the constructed configuration C. Namely, v's degree is
at most 3
4 ∆, the number of v's neighbors is similar in all sets Ri is approximately the same, and the number
of v's neighbors is similar in all sets Hi.
Claim 4.8. If Events 1–5 occur for ǫ ∈ [0, (200 ln n)−1] and ∆
properties hold for every vertex v ∈ V⋆ that is not fixed by C:
R ln2 n, then the following
m ≥ 4000µ−2
1. dv < 3
4 ∆.
2. There exists χv such that for all i ∈ [m],
⋆ + ǫ∆⋆(cid:17)z .
m ≥ 4000µ−2
28
3. There exists ψv ∈(cid:2)0, 3
⋆ +
3
4
ǫ∆⋆(cid:19){ .
N (v) ∩ Ri /µR ∈sχv ±(cid:18)∆3/4
4 ∆⋆(cid:3) such that for all i ∈ [m],
N (v) ∩ Hi ∈rψv ± α(cid:16)∆3/4
R ln2 n = 4 · 1015 · ln4 n. To prove
Proof. We use in the proof of the claim the fact that ∆⋆ = ∆
the lemma, we fix a vertex v that is not fixed by C. The first property is directly implied by Event 2. Suppose
that dv ≥ 3
4 ∆. Then v is included in the Hi corresponding to the subgraph to which it has been assigned
and v is fixed by C. We obtain a contradiction that implies that dv < 3
4 ∆.
For the second property, we now know that dv < 3
4 ∆. The property follows then directly from Event 3
with χv
def= dv/m.
The last property requires a more complicated reasoning. We set ψv
any i ∈ [m]. By Event 4,
Wv(Vi) ∈sψv ±(cid:18)∆3/4
⋆ +
3
4
ǫ∆⋆(cid:19){ .
def= Wv(V⋆)/m < 3
4 ∆⋆. Consider
(2)
Consider now an arbitrary u ∈ V⋆. We bound the difference between wu = µH (du/∆), which can be seen
as the ideal probability of the inclusion in the set of heavy vertices, and µH(cid:16)bdu/∆⋆(cid:17), the actual probability
4 ǫ(cid:17). We
of this event in Line 3 of the appropriate execution of LocalPhase. Let δ⋆
consider two cases.
def= α(cid:16)∆−1/4
+ 3
⋆
• If du < 3
4 ∆, by Event 3, the monotonicity of µH , and Lemma 4.6,
µH(cid:16)bdu/∆⋆(cid:17) ∈sµH(cid:18)du
∆ ±(cid:18)∆−1/4
⋆
+
3
4
ǫ(cid:19)(cid:19){
⊆Jwu · (1 ± δ⋆)K .
Note that Lemma 4.6 is applied properly because ∆−1/4
(48 ln n)−1.
⋆
• If du ≥ 3
(cid:2)1 − 1
which is the same bound as in the previous case.
⋆
4 ∆, by Event 1, µH(cid:16)bdu/∆⋆(cid:17) ∈ (cid:2)1 − 1
2 n−12, 1(cid:3). Because ∆⋆ is relatively small, i.e., ∆⋆ ≤ n,
µH(cid:16)bdu/∆⋆(cid:17) ∈rwu(cid:16)1 ± ∆−1/4
µH(cid:16)bdu/∆⋆(cid:17) ∈u
hv,i = Xu∈N (v)∩Vi
=JWv (Vi) · (1 ± δ⋆)K .
+ 3
4 ǫ ≤ (200 ln n)−1 + (200 ln n)−1 ≤
2 n−6, 1(cid:3). Concurrently, wu ∈ [µH(3/4), 1] =
(cid:17)z ⊆Jwu · (1 ± δ⋆)K ,
v(1 ± δ⋆) · Xu∈N (v)∩Vi
wu}
~
(3)
It follows from the bound that we just obtained and the definitions of Wv and hv,i that
We now combine bounds (2) and (3):
hv,i ∈(cid:20)ψv (1 − δ⋆) −(cid:18)∆3/4
⊆sψv ±(cid:18)ψvδ⋆ +(cid:18)∆3/4
⋆ +
⋆ +
3
4
ǫ∆⋆(cid:19) (1 + δ⋆) , ψv (1 + δ⋆) +(cid:18)∆3/4
ǫ∆⋆(cid:19) (1 + δ⋆)(cid:19){ .
3
4
⋆ +
3
4
ǫ∆⋆(cid:19) (1 + δ⋆)(cid:21)
Due to the lower bound on ∆⋆, we obtain δ⋆ ≤ α(cid:0)(200 ln n)−1 + (200 ln n)−1(cid:1) ≤ 1. This enables us to
29
simplify and further transform the bound on hv,i:
⋆ +
hv,i ∈sψv ±(cid:18)ψvδ⋆ + 2(cid:18)∆3/4
⊆sψv ±(cid:18) 3
⊆sψv ± α(cid:18) 4
⋆ + ǫ∆⋆(cid:19){ .
α∆3/4
∆3/4
⋆ +
9
16
4
5
3
4
ǫ∆⋆(cid:19)(cid:19){
αǫ∆⋆ + 2∆3/4
⋆ +
3
2
ǫ∆⋆(cid:19){
By applying the bound on ∆⋆ again, we obtain a bound on the magnitude of the second term in the above
bound:
α(cid:18) 4
5
∆3/4
⋆ + ǫ∆⋆(cid:19) = α(cid:18) 4
5
∆−1/4
⋆
+ ǫ(cid:19) ∆⋆ ≤ 96 ln n(cid:18)
1
200 ln n
+
1
200 ln n(cid:19) ∆⋆ ≤ ∆⋆.
This implies that hv,i ≤ ψv+∆⋆ ≤ 2∆⋆. The condition in Event 5 holds, and therefore, N (v) ∩ Hi − hv,i ≤
∆3/4
. We combine this with the bound on hv,i to obtain
⋆
N (v) ∩ Hi ∈sψv ±(cid:18)α
4
5
∆3/4
⋆ + αǫ∆⋆ + ∆3/4
⋆ (cid:19){ ⊆rψv ± α(cid:16)∆3/4
⋆ + ǫ∆⋆(cid:17)z .
Wrapping up the proof of near uniformity. We now finally prove Lemma 4.4. Recall that it states that
an ǫ-near uniform D is very likely to result in a near uniform D[C] with a slightly worse parameter and
that all vertices not fixed by C have bounded degree. The proof combines the last two claims: Claim 4.7
and Claim 4.8. We learn that C, the m-configuration constructed in the process is very likely to have the
properties listed in Claim 4.8. One of those properties is exactly the property that all vertices not fixed by
C have bounded degree. Hence we have to prove only the near uniformity property. We accomplish this by
observing that the probability of C equal to a specific m-configuration C⋆ with good properties-those in
Claim 4.8-does not depend significantly on to which induced subgraph a given vertex v not fixed in C⋆ is
assigned. This can be used to show that the conditional distribution of v given that C = C⋆ is near uniform
as desired.
⋆
when EmulatePhase is run for D and has the properties specified by Claim 4.8. Consider an arbitrary
+ ǫ(cid:17)-near uniform for C with this set of properties.
Proof of Lemma 4.4. By combining Claim 4.7 and Claim 4.8, we learn that with probability at least 1− n−4,
all properties listed in the statement of Claim 4.8 hold for C, the configuration constructed by EmulatePhase.
Since one of the properties is exactly the same as in the statement of Lemma 4.4, it suffices to prove the
other one: that D[C] is 60α(cid:16)∆−1/4
Fix eC = (cid:16){eRi}i∈[m],{eHi}i∈[m],{eFi}i∈[m](cid:17) to be an m-configuration that has non-zero probability
vertex v ∈ V⋆. In order to prove the near uniformity of DheCi, we show that v is assigned by it almost
uniformly to [m]. Let E be the event that EmulatePhase constructs eC, i.e., C = eC. For each i ∈ [m], let
E→i be the event that v is assigned to the i-th induced subgraph. Let p : [m] → [0, 1] be the probability
mass function describing the probability of the assignment of v to each of the m subgraphs in D. Obviously,
p(i) = Pr[E→i] for all i ∈ [m]. Due to the ǫ-near uniformity of D, p(i) =q 1
def= Pr[EE→i]. In order to express all qi's in a suitable form, we conduct a
For each i ∈ [m], let qi
thought experiment. Suppose v were not present in the graph, but the distribution of all the other vertices in
m (1 ± ǫ)y.
30
does the probability of E change if we add v back and condition on its assignment to a machine i? Note first
that conditioning on E→i does not impact the distribution of the other vertices, because vertices are assigned
to machines independently in D. In order for E still to occur in this scenario, v cannot be assigned to any
behavior of other vertices, which only depends on the content of these sets and independent randomized
decisions to include vertices. As a result we can express qi as a product of q⋆ and three probabilities: of v
the modified D remained the same. Let q⋆ be the probability of E, i.e., C = eC, in this modified scenario. How
of eRi, eHi, or eFi, for which it is considered. Additionally, as long as this the case, v does not impact the
not being included in sets eRi, eHi, or eFi.
qi = q⋆ · (1 − µR) ·1 − µH
(cid:12)(cid:12)(cid:12)N (v) ∩ eHi(cid:12)(cid:12)(cid:12)
.
∆⋆
∆⋆
(4)
Using the properties listed in Claim 4.8, we have
(cid:12)(cid:12)(cid:12)N (v) ∩ eRi(cid:12)(cid:12)(cid:12) /µR
(cid:12)(cid:12)(cid:12)N (v) ∩ eRi(cid:12)(cid:12)(cid:12) /µR ∈sχv ±(cid:18)∆3/4
(cid:12)(cid:12)(cid:12)N (v) ∩ eHi(cid:12)(cid:12)(cid:12) ∈rψv ± α(cid:16)∆3/4
·1 − µF
ǫ∆⋆(cid:19){ ,
⋆ + ǫ∆⋆(cid:17)z ,
⋆ +
3
4
and
where χv and ψv are constants independent of machine i to which v has been assigned and ψ ≤ 3
4 ∆⋆. In the
next step, we use these bounds to derive bounds on the multiplicative terms in Equation (4) that may depend
on i. We also repeatedly use the bounds ∆⋆ = ∆
from the lemma statement. First, due to Lemma 4.6,
m ≥ 4000µ−2
1 − µH
(cid:12)(cid:12)(cid:12)N (v) ∩ eRi(cid:12)(cid:12)(cid:12) /µR
∆⋆
R ln2 n = 4 · 1015 · ln4 n and ǫ ≤ (200 ln n)−1
ǫ(cid:19)(cid:19){
∆⋆ ±(cid:18)∆−1/4
∆⋆(cid:19)(cid:19) ·(cid:18)1 ± α(cid:18)∆−1/4
ǫ(cid:19)(cid:19){ .
3
4
3
4
+
+
⋆
⋆
∈s1 − µH(cid:18) χv
⊆s(cid:18)1 − µH(cid:18) χv
(Note that the application of Lemma 4.6 was correct, because ∆−1/4
(96 ln n)−1.) Second,
⋆
+ 3
4 ǫ ≤ (200 ln n)−1 + (200 ln n)−1 <
Since ψv/∆⋆ ≤ 3
µF in the above bound is always less than 4, and therefore, only one branch of µF 's definitions gets applied.
Hence, we can eliminate µF :
⋆
1 − µF
4 and α(cid:16)∆−1/4
1 − µF
⋆
∆⋆
+ ǫ(cid:17)(cid:19){ .
∆⋆ ± α(cid:16)∆−1/4
∈s1 − µF(cid:18) ψv
(cid:12)(cid:12)(cid:12)N (v) ∩ eHi(cid:12)(cid:12)(cid:12)
+ ǫ(cid:17) ≤ (96 ln n) ·(cid:0)(200 ln n)−1 + (200 ln n)−1(cid:1) < 1, the argument to
(cid:12)(cid:12)(cid:12)N (v) ∩ eHi(cid:12)(cid:12)(cid:12)
4(cid:16)∆−1/4
+ ǫ(cid:17){ .
ψv
4∆⋆ ±
∆⋆
α
⋆
∈s1 −
31
Since 1 − ψv
4∆⋆ ≥ 3
4 , we can further transform the bound to
α
ψv
∆⋆
4∆⋆(cid:19)(cid:16)1 ±
(cid:12)(cid:12)(cid:12)N (v) ∩ eHi(cid:12)(cid:12)(cid:12)
1 − µF
∈s(cid:18)1 −
4 ǫ(cid:17) and δ2
+ ǫ(cid:17). As a result, every qi can be expressed as qi = ηvλiλ′
3(cid:16)∆−1/4
i ∈J1 ± δ2K. For every i, we can also write
Pr[E ∧ E→i] = Pr[EE→i] · Pr[E→i] = ηvλiλ′
3(cid:16)∆−1/4
+ ǫ(cid:17)(cid:17){ .
i · p(i) =
iλ′′
i ,
λiλ′
ηv
m
def= α
+ 3
⋆
⋆
i,
where ηv is a constant independent of i, λi ∈J1 ± δ1K, and λ′
Let δ1
def= α(cid:16)∆−1/4
⋆
where λ′′
D given E:
i ∈J1 ± ǫK. We now express the conditional probability of v being assigned to the i-th subgraph in
Note that for any i, this implies that
Pr[E→iE] =
Pr[E ∧ E→i]
j=1 Pr[E ∧ E→j]
=
Pm
λiλ′
iλ′′
i
j=1 λjλ′
.
jλ′′
j
Pm
1
m ·
(1 − δ1)(1 − δ2)(1 − ǫ)
(1 + δ1)(1 + δ2)(1 + ǫ) ≤ Pr[E→iE] ≤
1
m ·
(1 + δ1)(1 + δ2)(1 + ǫ)
(1 − δ1)(1 − δ2)(1 − ǫ)
.
(5)
Observe that
and
δ1 ≤ (96 ln n) ·(cid:0)(7000 ln n)−1 + (250 ln n)−1(cid:1) < 1/2,
3 · (96 ln n) ·(cid:0)(7000 ln n)−1 + (200 ln n)−1(cid:1) < 1/2.
1
δ2 ≤
Hence all of δ1, δ2, and ǫ are at most 1/2. We can therefore transform (5) to
1
m · (1 − δ1)2(1 − δ2)2(1 − ǫ)2 ≤ Pr[E→iE] ≤
1
m · (1 + δ1)(1 + δ2)(1 + ǫ)(1 + 2δ1)(1 + 2δ2)(1 + 2ǫ),
and then
Hence
1
m · (1 − 2δ1 − 2δ2 − 2ǫ) ≤ Pr[E→iE] ≤
1
m · (1 + 45δ1 + 45δ2 + 45ǫ).
Pr[E→iE] ∈s 1
m · (1 ± 45(δ1 + δ2 + ǫ)){ ⊆s 1
m ·(cid:16)1 ± 60α(cid:16)∆−1/4
⋆
+ ǫ(cid:17)(cid:17){ ,
which finishes the proof that DheCi is 60α(cid:16)∆−1/4
⋆
5 Parallel Algorithm
+ ǫ(cid:17)-near uniform.
In this section, we introduce our main parallel algorithm. It builds on the ideas introduced in EmulatePhase.
EmulatePhase randomly partitions the graph into m induced subgraphs and runs on each of them
LocalPhase, which resembles a phase of GlobalAlg. As we have seen, the algorithm performs well
even if vertices are assigned to subgraphs not exactly uniformly so long as the assignment is fully inde-
pendent. Additionally, with high probability, if we condition on the configuration of sets Ri, Hi, and Fi
32
that were removed, the distribution of assignments of the remaining vertices is still nearly uniform and also
independent.
These properties allow for the main idea behind the final parallel algorithm. We partition vertices ran-
domly into m induced subgraphs and then run LocalPhase multiple times on each of them with no repar-
titioning in the meantime. In each iteration, for a given subgraph, we halve the local threshold ∆⋆. This
corresponds to multiple phases of the original global algorithm. As long as we can show that this approach
leads to finding a large matching, the obvious gain is that multiple phases of the original algorithm translate
to O(1) parallel rounds. This approach enables our main result: the parallel round complexity reduction
from O(log n) to O((log log n)2).
Algorithm 5: ParallelAlg(G, S)
The final parallel matching algorithm
Input:
• graph G = (V, E) on n vertices
• parameter S ∈ Z+ such that S ≤ n and S = nΩ(1) (each machine uses O(S) space)
Output: matching in G
1 ∆ ← n, V ′ ← V , M ← ∅
2 while ∆ ≥ n
S (200 ln n)32 do
/* High-probability invariant: maximum degree in G[V ′] bounded by 3
2 ∆ */
3
4
5
6
7
8
9
10
11
/* number of phases to emulate */
Partition V ′ into m sets V1, . . . , Vm by assigning each vertex independently uniformly at random.
foreach i ∈ [m] do in parallel
/* number of machines used */
m ←jq n∆
S k
16 log120α (∆/m)(cid:7)
τ ←(cid:6) 1
for j ∈ [τ ] do (Vi, Mi,j) ← LocalPhase(cid:0)i, G[Vi], ∆/(cid:0)2j−1m(cid:1)(cid:1)
V ′ ←Sm
M ← M ∪Sm
If the number of edges in G[Vi] is greater than 8S, Vi ← ∅.
i=1Sτ
∆ ← ∆/2τ
i=1 Vi
j=1 Mi,j
12 Compute degrees of vertices V ′ in G[V ′] and remove from V ′ vertices of degree at least 2∆.
13 Directly simulate M ′ ← GlobalAlg(G[V ′], 2∆), using O(1) rounds per phase.
14 return M ∪ M ′
We present ParallelAlg, our parallel algorithm, as Algorithm 5. We write S to denote a parameter
specifying the amount of space per machine. After the initialization of variables, the algorithm enters
the main loop in Lines 2–11. The loop is executed as long as ∆, an approximate upper bound on the
maximum degree in the remaining graph, is large enough. The loop implements the idea of running multiple
iterations of LocalPhase on each induced subgraph in a random partition. At the beginning of the loop,
the algorithm decides on m, the number of machines, and τ , the number of phases to be emulated. Then it
creates a random partition of the current set of vertices that results in m induced subgraphs. Next for each
subgraph, the algorithm first runs a security check that the set of edges fits onto a single machine (see Line 7).
If it does not, which is highly unlikely, the entire subgraph is removed from the graph. Otherwise, the entire
subgraph is sent to a single machine that runs τ consecutive iterations of LocalPhase. Then the vertices
not removed in the executions of LocalPhase are collected for further computation and new matching
edges are added to the matching being constructed. During the execution of the loop, the maximum degree in
the graph induced by V ′, the set of vertices to be considered is bounded by 3
2 ∆ with high probability. Once
33
the loop finishes, we remove from the graph vertices of degree higher than 2∆-there should be none-and
we directly simulate GlobalAlg on the remaining graph, using O(1) rounds per phase.
5.1 Some Properties of Thresholds
Before we analyze the behavior of the algorithm, we observe that the value of ∆
m inside the main loop
is at least polylogarithmic and that the same property holds for the rescaled threshold that is passed to
LocalPhase.
Lemma 5.1. Consider a single iteration of the main loop of ParallelAlg (Lines 2–11). Let ∆ and m be
set as in this iteration. The following two properties hold:
• ∆/m ≥ (200 log n)16.
• The threshold ∆/(cid:0)2j−1m(cid:1) passed to LocalPhase in Line 8 is always at least (∆/m)15/16 ≥
R ln2 n.
4000µ−2
Proof. Let τ be also as in this iteration of the loop. The smallest threshold passed to LocalPhase is
∆/(2τ −1m). Let λ def= S∆/n, where S is the parameter to ParallelAlg. Due to the condition in
√λ. This implies that
Line 2, λ ≥ (200 ln n)32. Note that ∆ = λn/S. Hence m ≤ pn∆/S = n
∆/m ≥ √λ ≥ (200 ln n)16, which proves the first claim. Due to the definition of τ ,
S
This implies that
2τ −1 ≤ (120α)τ −1 ≤ (∆/m)1/16.
∆/(2τ −1m) ≥ (∆/m)15/16 ≥ (200 ln n)15 > 4 · 1015 · ln4 n = 4000µ−2
R ln2 n.
We also observe that the probability of any set of vertices deleted by the security check in Line 7 of
ParallelAlg is low as long as the maximum degree in the graph induced by the remaining vertices is
bounded.
Lemma 5.2. Consider a single iteration of the main loop of ParallelAlg and let ∆ and V ′ be as in that
iteration. If the maximum degree in G[V ′] is bounded by 3
2 ∆, then the probability of any subset of vertices
deleted in Line 7 is n−8.
Proof. Let m be as in the same iteration of the main loop of ParallelAlg. Consider a single vertex
v ∈ V ′. The expected number of v's neighbors assigned to the same subgraph is at most 3
2 ∆/m. Recall
that due to Lemma 5.1, ∆
m ≥ 200 ln n. Since the assignment of vertices to machines is fully independent, by
Lemma 4.1 (i.e., the Chernoff bound), the probability that v has more than 2∆/m neighbors is bounded by
2 exp −
1
3 ·(cid:18) 1
3(cid:19)2
3
2 ·
·
∆
m! ≤ 2 exp(cid:18)−
1
18 · 200 ln n(cid:19) ≤ n−10.
Therefore, by the union bound, with probability 1− n−9, no vertex has more than 2∆ neighbors in the same
induced subgraph. As V ′ ≤ n, the expected number of vertices in each set Vi constructed in the iteration
of the main loop is at most n/m ≥ ∆/m ≥ 200 ln n. What is the probability that Vi > 2n/m? By the
independence of vertex assignments and Lemma 4.1, the probability of such event is at most
2 exp(cid:18)−
1
3 ·
n
m(cid:19) ≤ 2 exp(cid:18)−
1
3 · 200 ln n(cid:19) ≤ n−10.
34
Again by the union bound, the event Vi ≤ 2n/m, for all i simultaneously, occurs with probability at least
1− n−9. Combining both bounds, with probability at least 1− 2n−9 ≥ 1− n−8, all induced subgraphs have
at most 2n/m vertices and the degree of every vertex is bounded by 2∆/m. Hence the number of edges in
2 · 2n
one induced subgraph is at most 1
m2 . By the definition of m and the setting of parameters
2q n∆
in the algorithm, m ≥ 1
S , where S is the parameter to ParallelAlg. This implies that the number
of edges is at most 2n∆/(cid:18) 1
S (cid:19)2
2q n∆
= 8S in every induced subgraph with probability 1 − n−8, in which
case no set Vi is deleted in Line 7 of ParallelAlg.
m = 2n∆
m · 2∆
5.2 Matching Size Analysis
The parallel algorithm runs multiple iterations of LocalPhase on each induced subgraph, without reparti-
tioning. A single iteration on all subgraphs corresponds to running EmulatePhase once. We now show
that in most cases, the global algorithm simulates EmulatePhase on a well behaved distribution with in-
dependently assigned vertices and all vertices distributed nearly uniformly conditioned on the configurations
of the previously removed sets Ri, Hi, and Fi. We also show that the maximum degree in the remaining
graph is likely to decrease gracefully during the process.
Lemma 5.3. With probability at least 1 − n−3:
• all parallel iterations of LocalPhase in ParallelAlg on each induced subgraph correspond
to running EmulatePhase on independent and (200 ln n)−1-near uniform distributions of assign-
ments,
• the maximum degree of the graph induced by the remaining vertices after the k-th simulation of
EmulatePhase is 3
2 ∆/2k.
def= ∆/(cid:0)2j−1m(cid:1) be the threshold passed to LocalPhase
Proof. We first consider a single iteration of the main loop in ParallelAlg. Let ∆, τ , and m be set as
in this iteration of the loop. For j ∈ [τ ], let ∆j
for the j-th iteration of LocalPhase on each of the induced subgraphs. The parallel algorithm assigns
vertices to subgraphs and then iteratively runs LocalPhase on each of them. In this analysis we ignore
the exact assignment of vertices to subgraphs until they get removed as a member of one of sets Ri, Hi, or
Fi. Instead we look at the conditional distribution on assignments given the configurations of sets Ri, Hi,
and Fi removed in the previous iterations corresponding to EmulatePhase. We write Dj, 1 ≤ j ≤ τ , to
denote this distribution of assignments before the execution of j-th iteration of LocalPhase on the induced
subgraphs, which corresponds to the j-th iteration of EmulatePhase for this iteration of the main loop of
ParallelAlg. Additionally, we write Dτ +1 to denote the same distribution after the τ -th iteration, i.e.,
at the end of the execution of the parallel block in Lines 6–8 of ParallelAlg. Due to Lemma 4.3, the
distributions of assignments are all independent. We define ǫj, j ∈ [τ + 1], to be the minimum positive value
such that Dj is ǫj-near uniform. Obviously, ǫ1 = 0, since the first distribution corresponds to a perfectly
uniform assignment. We want to apply Lemma 4.4 inductively to bound the value of ǫj+1 as a function of
ǫj with high probability. The lemma lists two conditions: ǫj must be at most (200 ln n)−1 and the threshold
passed to EmulatePhase has to be at least 4000µ−2
H ln2 n. The latter condition holds due to Lemma 5.1.
Hence as long as ǫj is sufficiently small, Lemma 4.4 implies that with probability at least 1 − n−4,
ǫj+1 ≤ 60α (cid:18) ∆
2τ −1m(cid:19)−1/4
+ ǫj! ≤ 60α (cid:18) ∆
m(cid:19)−15/64
+ ǫj! ,
35
and no high degree vertex survives in the remaining graph. One can easily show by induction that if this
for all j ∈ [τ + 1]. In particular,
by the definition of τ and Lemma 5.1, for any j ∈ [τ ],
recursion is satisfied for all 1 ≤ j ≤ τ , then ǫj ≤ (120α)j−1 ·(cid:0) ∆
·(cid:18) ∆
m(cid:19)−15/64
ǫj ≤ (120α)τ −1 ·(cid:18) ∆
m(cid:19)−15/64
m(cid:19)1/16
≤(cid:18) ∆
m(cid:1)−15/64
≤(cid:18) ∆
m(cid:19)−11/64
≤ (200 ln n)−1,
This implies that as long the unlikely events specified in Lemma 4.4 do not occur for any phase in any
iteration of the main loop of ParallelAlg, we obtain the desired properties: all intermediate distributions
of possible assignments are (200 ln n)−1-near uniform and the maximum degree in the graph decreases at
the expected rate. It remains to bound the probability of those unlikely events occurring for any phase. By
the union bound, their total probability is at most log n · n−4 ≤ n−3.
We now prove that the algorithm finds a large matching with constant probability.
Theorem 5.4. Let MOPT be an arbitrary maximum matching in a graph G. With Ω(1) probability, ParallelAlg
constructs a matching of size Ω(MOPT).
Proof. By combining Lemma 5.2 and Lemma 5.3, we learn that with probability at least 1−n·n−8−n−3 ≥
1 − 2n−3, we obtain a few useful properties. First, all relevant distributions corresponding to iterations of
EmulatePhase are independent and (200 ln n)−1-near uniform. Second, the maximum degree in the
graph induced by vertices still under consideration is bounded by 3
2 ∆ before and after every simulated
execution of EmulatePhase, where ∆ is the corresponding. As a result, no vertex is deleted in Lines 7
or 12 due to the security checks.
We now use Lemma 4.2 to lower bound the expected size of the matching created in every EmulatePhase
simulation. Let τ⋆ be the number of phases we simulate this way. We have τ⋆ ≤ log n. Let Hj , Fj, and
Mj be random variables equal to the total size of sets Hi, Fi, and Mi created in the j-th phase. If the corre-
sponding distribution in the j-th phase is near uniform and the maximum is bounded, Lemma 4.2 yields
i.e.,
1
E [Mj] ≥
E [Hj + Fj] ≤ n−9 + 1200 · E [Mj] ,
1200(cid:0)E [Hj + Fj] − n−9(cid:1) .
1200(cid:0)E [Hj + Fj] − n−9(cid:1) − 2n−3 ·
1
n
2
,
Xj∈[τ⋆]
E [Mj] ≥ Xj∈[τ⋆]
Overall, without the assumption that the conditions of Lemma 4.2 are always met, we obtain a lower bound
in which we consider the worst case scenario that we lose as much as n/2 edges in the size of the con-
structed matching when the unlikely negative events happen. ParallelAlg continues the construction of
a matching by directly simulating the global algorithm. Let τ ′
⋆ be the number of phases in that part of the
algorithm. We define H′
⋆], to be random variables equal to the size of sets H, F ,
j, and M′
j, F′
and fM in GlobalAlg in the j-th phase of the simulation. By Lemma 3.3, we have
j , for j ∈ [τ ′
E(cid:2)M′
j(cid:3) ≥ Xj∈[τ ′
⋆]
Xj∈[τ ′
⋆]
36
1
50(cid:0)E(cid:2)H′
j + F′
j(cid:3)(cid:1) .
By combining both bounds we obtain a lower bound on the size of the constructed matching. Let
M⋆
def= Xj∈[τ⋆]
E [Mj] + Xj∈[τ ′
⋆]
E(cid:2)M′
j(cid:3)
be the expected matching size, and let
V⋆
def= Xj∈[τ⋆]
E [Hj + Fj] + Xj∈[τ ′
⋆]
j + F′
E(cid:2)H′
j(cid:3) .
We have
M⋆ ≥
1
1200
V⋆ −
1
n2 .
Consider a maximum matching MOPT. At the end of the algorithm, the graph is empty. The expected
number of edges in MOPT incident to a vertex in one of the reference sets is bounded by MOPT · 2µR ·
log n ≤ 10−5MOPT. The expected number of edges removed by the security checks is bounded by n
2 ·n−3.
Hence the expected number of edges in MOPT deleted as incident to vertices that are heavy or are friends is
at least (1 − 10−5)MOPT − 1/(2n2). Since we can assume without the loss of generality that the graph is
non-empty, it is at least 1
n2 . For sufficiently
large n (say, n ≥ 50), M⋆ ≥ Ω (MOPT) and by an averaging argument, ParallelAlg has to output an
O(1)-multiplicative approximation to the maximum matching with Ω(1) probability. For smaller n, it is not
difficult to show that at least one edge is output by the algorithm with constant probability as long as it is
not empty.
2MOPT. Hence V⋆ ≥ 1
2MOPT, and M⋆ ≥ 1
2400MOPT − 1
Finally, we want to argue that the above procedure can be used to compute 2 + ǫ approximation to
maximum matching at the cost of increasing the running time by a factor of log(1/ǫ). The idea is to;
execute algorithm ParallelAlg to compute constant approximate matching; remove this matching from
the graph; and repeat.
Corollary 5.5. Let MOPT be an arbitrary maximum matching in a graph G. For any ǫ > 0, executing
ParallelAlg on G and removing a constructed matching repetitively, O(log(1/ǫ)) times, finds a multi-
plicative (2 + ǫ)-approximation to maximum matching, with Ω(1) probability.
Proof. Assume that the ParallelAlg succeeds with probability p and computes c-approximate matching.
Observe that each successful execution of ParallelAlg finds a matching Mc of size at least 1
cMOPT.
Removal of Mc from the graph decreases the size of optimal matching by at least 1
cMOPT and at most
by 2
cMOPT, because each edge of Mc can be incident to at most two edges of MOPT. Hence, when the
size of the remaining matching drops to at most ǫMOPT, we have an 2 + ǫ-multiplicative approximation
to maximum matching constructed. The number t of successful applications of ParallelAlg need to
satisfy.
(cid:18)1 −
1
c(cid:19)t
≤ ǫ.
This gives t = O(log(1/ǫ)). In ⌈t/p⌉ = O(log(1/ǫ)) executions, we have t successes with probability at
least 1/2 by the properties of the median of the binomial distribution.
37
6 MPC Implementation Details
In this section we present details of an MPC implementation of our algorithm. We also analyze its round and
space complexity. In the description we heavily use some of the subroutines described in [GSZ11]. While
the model used there is different, the properties of the distributed model used in [GSZ11] also hold in the
MPC model. Thus, the results carry over to the MPC model.
The results of [GSZ11] allow us to sort a set A of O(N ) key-value pairs of size O(1) and for every
element of a sorted list, compute its index. Moreover, we can also do a parallel search: given a collection A
of O(N ) key-value pairs and a collection of O(N ) queries, each containing a key of an element of A, we can
annotate each query with the corresponding key-value pair from A. Note that multiple queries may ask for
the same key, which is nontrivial to parallelize. If S = nΩ(1), all the above operations can be implemented
in O(1) rounds.
The search operation allows us to broadcast information from vertices to their incident edges. Namely,
we can build a collection of key-value pairs, where each key is a vertex and the value is the corresponding
information. Then, each edge {u, v} may issue two queries to obtain the information associated with u and
v.
6.1 GlobalAlg
We first show how to implement GlobalAlg, which is called in Line 13 of ParallelAlg.
Lemma 6.1. Let S = nΩ(1). There exists an implementation of GlobalAlg in the MPC model, which with
high probability executes O(lne∆) rounds and uses O(S) space per machine.
Proof. We first describe how to solve the following subproblem. Given a set X of marked vertices, for each
vertex v compute N (v) ∩ X. When all vertices are marked, this just computes the degree of every vertex.
The subproblem can be solved as follows. Create a set AX = {(u, v) u ∈ V, v ∈ X,{u, v} ∈
E} ∪ {(v,−∞), (v,∞) v ∈ V }, and sort its elements lexicographically. Denote the sorted sequence by
QX . Then, for each element of AX compute its index in QA.
Note that N (v) ∩ X is equal to the number of elements in QX between (v,−∞) and (v,∞). Thus,
having computed the indices of these two elements, we can compute N (v) ∩ X.
Let us now describe how to implement GlobalAlg. We can compute the degrees of all vertices, as
described above. Once we know the degrees, we can trivially mark the vertices in H. The next step is to
compute F and for that we need to obtain N (v) ∩ H, which can be done as described above.
After that, GlobalAlg computes a matching in G[H ∪ F ] by calling MatchHeavy (see Algorithm 2).
In the first step, MatchHeavy assigns to every v ∈ F a random neighbor v⋆ in H. This can again be
easily done by using the sequence QH (i.e. QX build for X = H). Note that for each v ∈ F we know the
number of neighbors of v that belong to H. Thus, each vertex v can pick an integer rv ∈ [1,N (v) ∩ H]
uniformly at random. Then, by adding rv and the index of (v,−∞) in QH , we obtain the index in QH , which
corresponds to an edge between v and its random neighbor in H. The remaining lines of MatchHeavy are
straightforward to implement. The vertices can trivially pick their colors. After that, the set E⋆ can be easily
computed by transmitting data from vertices to their adjacent edges. Implementing the following steps of
MatchHeavy is straightforward. Finally, picking the edges to be matched is analogous to the step, when
for each v ∈ F we picked a random neighbor in H.
Overall, each phase of GlobalAlg (that is, iteration of the main loop) is executed in O(1) rounds.
Thus, by Lemma 3.3, GlobalAlg can be simulated in O(lne∆) rounds as advertised.
38
6.2 Vertex and edge partitioning
We now show how to implement Line 5 and compute the set of edges that are used in each call to LocalPhase
in Line 8 of ParallelAlg. Our goal is to annotate each edge with the machine number it is supposed to
go to. To that end, once the vertices pick their machine numbers, we broadcast them to their adjacent edges.
Every edge that receives two equal numbers x is assigned to machine x.
In the implementation we do not check whether a machine is assigned too many edges (Line 7), but rather
show in Lemma 5.2 that not too many edges are assigned with high probability.
6.3 LocalPhase
We now discuss the implementation of LocalPhase. Observe that LocalPhase is executed locally.
Therefore, the for loop at Line 8 of ParallelAlg can also be executed locally on each machine. Thus,
we only explain how to process the output of LocalPhase.
Instead of returning the set of vertices and matched edges at Line 6 of LocalPhase, each vertex that
should be returned is marked as discarded, and each matched edge is marked as matched. After that,
we need to discard edges, whose at least one endpoint has been discarded. This can be done by broadcasting
information from vertices to adjacent edges. Note that some of the discarded edges might be also marked as
matched.
6.4 Putting all together
Lines 5, 7 and 8 can be implemented as described in sections 6.2 and 6.3. Lines 9 and 10 do not need an
actual implementation, as by that point all the vertices that are not marked as discarded constitute V ′,
and all the edges incident to V \ V ′ will be marked as discarded. Similarly, all the matched edges will be
marked as matched by the implementation of LocalPhase. All the edges and vertices that are marked
as discarded will be ignored in further processing. After all the rounds are over, the matching consists
of the edges marked as matched.
Let ∆⋆ be the value of ∆ at Line 12, and hence the value of ∆ at the end of the last while loop iteration.
Let ∆′ be the value of ∆ just before the last iteration, i.e. ∆⋆ = ∆′/2τ , for the corresponding τ . Now
consider the last call of LocalPhase at Line 8. The last invocation has ∆′/(2τ −1) as a parameter. On
the other hand, by Claim 4.7 and Claim 4.8 we know that after the last invocation of LocalPhase with
high probability there is no vertex that has degree greater then 3
4 ∆′/(2τ −1) < 2∆⋆. Therefore, with high
probability there is no vertex that should be removed at Line 12, and hence we do not implement that line
either.
An implementation of Line 13 is described in Section 6.1. Finally, we can state the following result.
Lemma 6.2. There exists an implementation of ParallelAlg in the MPC model that with high probabil-
ity executes O(cid:0)(log log n)2 + max(cid:0)log n
S , 0(cid:1)(cid:1) rounds.
Proof. In the proof we analyze the case S ≤ n. Otherwise, for the case S > n, we think of each machine
being split into ⌊S/n⌋ "smaller" machines, each of the smaller machines having space n.
We will analyze the number of iterations of the while loop ParallelAlg performs. Let ∆i and τi be
the value of ∆ and τ at the end of iteration i, respectively. Then, from Line 3 and Line 4 we have
τi =(cid:24) 1
16
log120α (∆i−1/m)(cid:25) ≥
1
16
log120α (∆i−1/m) ≥
log120αr S∆i−1
n
.
1
16
39
Define γ :=
1
32 log2 120α . By plugging in the above bound on τi, from Line 11, we derive
∆i = ∆i−1 · 2τi ≤ ∆i−1 · 2− 1
16 log120α q S∆i−1
n = ∆i−1 · 2−
S∆i−1
log2
32 log2 120α = ∆1−γ
n
S(cid:17)γ
i−1 (cid:16) n
(6)
To obtain the number of iterations the while loop of ParallelAlg performs, we derive for which
i ≥ 1 the condition at Line 2 does not hold.
Unraveling ∆i−1 further from (6) gives
∆i ≤ ∆(1−γ)i
0
S(cid:17)1−(1−γ)i
Observe that (c log log n)−1 ≤ γ ≤ (32 log log n)−1 < 1/2, for an absolute constant c and n ≥ 4.
1−(1−γ) = n(1−γ)i(cid:16) n
≤ n(1−γ)i(cid:16) n
S(cid:17)γ 1−(1−γ)i
(cid:16) n
S(cid:17)γ Pi−1
j=0(1−γ)j
For S ≤ n and as γ < 1/2 we have
On the other hand, for i⋆ = log log n
γ
.
n
S
S(cid:17)1−(1−γ)i
(cid:16) n
≤
≤ c(log log n)2 we have
n(1−γ)i⋆ < log n.
Now putting together (7), (8), and (9) we conclude
∆i⋆ <
n
S
ln n,
(7)
(8)
(9)
Total round complexity. Every iteration of the while loop can be executed in O(1) MPC rounds with
and hence the number of iteration the while loop of ParallelAlg performs is O(cid:0)(log log n)2(cid:1).
probability at least 1 − 1/n3. Since there are O(cid:0)(log log n)2(cid:1) iterations of the while loop, all the iterations
of the loop can be performed in O(cid:0)(log log n)2(cid:1) many rounds with probability at least 1 − 1/n2.
S (ln n)32(cid:1)(cid:1) rounds. Putting the both bounds
of Line 13 of ParallelAlg can be performed in O(cid:0)log(cid:0) n
S(cid:1) for the case
together we conclude that the round complexity of ParallelAlg is O(cid:0)(log log n)2 + log n
of space n) the round complexity is O(cid:0)(log log n)2(cid:1).
S ≤ n. For the case S > n (recall that in this regime we assume that each machine is divided into machines
On the other hand, by Lemma 6.1 and the condition at Line 2 of ParallelAlg, the computation
Acknowledgments
We thank Sepehr Assadi, Mohsen Ghaffari, Cameron Musco, Christopher Musco, Seth Pettie, and Govind
Ramnarayan for helpful discussions and comments. A. Czumaj was supported in part by the Centre for Dis-
crete Mathematics and its Applications (DIMAP), Royal Society International Exchanges Scheme 2013/R1,
IBM Faculty Award, and the EPSRC award EP/N011163/1. A. M adry was supported in part by an Alfred
P. Sloan Research Fellowship, Google Research Award, and the NSF grant CCF-1553428. S. Mitrovi´c was
supported in part by the Swiss NSF grant P1ELP2_161820. P. Sankowski was partially supported by grant
NCN2014/13/B/ST6/00770 of Polish National Science Center and ERC StG grant TOTAL no. 677651.
40
References
[ABB+17] Sepehr Assadi, MohammadHossein Bateni, Aaron Bernstein, Vahab S. Mirrokni, and Cliff
Stein. Coresets meet EDCS: algorithms for matching and vertex cover on massive graphs.
CoRR, abs/1711.03076, 2017.
[ABI86]
Noga Alon, László Babai, and Alon Itai. A fast and simple randomized parallel algorithm for
the maximal independent set problem. Journal of Algorithms, 7(4):567–583, 1986.
[AG15]
Kook Jin Ahn and Sudipto Guha. Access to data and number of iterations: Dual primal algo-
rithms for maximum matching under resource constraints. In Proceedings of the 27th ACM on
Symposium on Parallelism in Algorithms and Architectures, SPAA 2015, Portland, OR, USA,
June 13–15, 2015, pages 202–211, 2015.
[AK17]
Sepehr Assadi and Sanjeev Khanna. Randomized composable coresets for matching and vertex
cover. In Proceedings of the 29th ACM Symposium on Parallelism in Algorithms and Architec-
tures, SPAA 2017, Washington DC, USA, July 24–26, 2017, pages 3–12, 2017.
[ANOY14] Alexandr Andoni, Aleksandar Nikolov, Krzysztof Onak, and Grigory Yaroslavtsev. Parallel
algorithms for geometric graph problems.
In Proceedings of the 46th ACM Symposium on
Theory of Computing, STOC 2014, New York, NY, USA, May 31–June 3, 2014, pages 574–583,
2014.
[Ass17]
Sepehr Assadi. Simple round compression for parallel vertex cover. CoRR, abs/1709.04599,
September 2017.
[BCH17]
Sayan Bhattacharya, Deeparnab Chakrabarty, and Monika Henzinger. Deterministic fully dy-
namic approximate vertex cover and fractional matching in O(1) amortized update time.
In
Proceedings of the 19th International Conference on Integer Programming and Combinatorial
Optimization, IPCO 2017, Waterloo, ON, Canada, June 26–28, 2017, pages 86–98, 2017.
[BH87]
[BHI15]
[BHN16]
[BHN17]
Paul Beame and Johan Hastad. Optimal bounds for decision problems on the CRCW PRAM. In
Proceedings of the 19th Annual ACM Symposium on Theory of Computing, STOC 1987, pages
83–93, New York, NY, USA, 1987.
Sayan Bhattacharya, Monika Henzinger, and Giuseppe F. Italiano. Deterministic fully dynamic
data structures for vertex cover and matching. In Proceedings of the 26th Annual ACM-SIAM
Symposium on Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4–6, 2015,
pages 785–804, 2015.
Sayan Bhattacharya, Monika Henzinger, and Danupon Nanongkai. New deterministic approx-
imation algorithms for fully dynamic matching.
In Proceedings of the 48th Annual ACM
SIGACT Symposium on Theory of Computing, STOC 2016, Cambridge, MA, USA, June 18–
21, 2016, pages 398–411, 2016.
Sayan Bhattacharya, Monika Henzinger, and Danupon Nanongkai. Fully dynamic approxi-
mate maximum matching and minimum vertex cover in O(log3 n) worst case update time. In
Proceedings of the 28th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2017,
Barcelona, Spain, Hotel Porta Fira, January 16–19, pages 470–489, 2017.
41
[BKS13]
Paul Beame, Paraschos Koutris, and Dan Suciu. Communication steps for parallel query pro-
cessing. In Proceedings of the 32nd ACM SIGMOD-SIGACT-SIGART Symposium on Princi-
ples of Database Systems, PODS 2013, New York, NY, USA, June 22–27, 2013, pages 273–284,
2013.
[BKS14]
Paul Beame, Paraschos Koutris, and Dan Suciu. Skew in parallel query processing.
In Pro-
ceedings of the 33rd ACM SIGMOD-SIGACT-SIGART Symposium on Principles of Database
Systems, PODS'14, Snowbird, UT, USA, June 22–27, 2014, pages 212–223, 2014.
[BS15]
[BS16]
[DG04]
Aaron Bernstein and Cliff Stein. Fully dynamic matching in bipartite graphs. In Proceedings of
the 42nd International Colloquium on Automata, Languages, and Programming, ICALP 2015,
Kyoto, Japan, July 6–10, 2015, Proceedings, Part I, pages 167–179, 2015.
Aaron Bernstein and Cliff Stein. Faster fully dynamic matchings with small approximation ra-
tios. In Proceedings of the 27th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2016, Arlington, VA, USA, January 10–12, 2016, pages 692–711, 2016.
Jeffrey Dean and Sanjay Ghemawat. MapReduce: Simplified data processing on large clusters.
In Proceedings of the 6th Conference on Symposium on Opearting Systems Design & Imple-
mentation, Volume 6, OSDI'04, pages 10–10, Berkeley, CA, USA, 2004. USENIX Association.
[DG08]
Jeffrey Dean and Sanjay Ghemawat. MapReduce: Simplified data processing on large clusters.
Commununication of the ACM, 51(1):107–113, January 2008.
[Edm65]
Jack Edmonds. Paths, trees and flowers. Canadian Journal of Mathematics, pages 449–467,
1965.
[FG17]
Manuela Fischer and Mohsen Ghaffari. Deterministic distributed matching: Simpler, faster,
better. CoRR, abs/1703.00900, 2017.
[FMS+10] Jon Feldman, S. Muthukrishnan, Anastasios Sidiropoulos, Clifford Stein, and Zoya Svitk-
ina. On distributing symmetric streaming computations. ACM Transactions on Algorithms,
6(4):66:1–66:19, 2010.
[GSZ11] Michael T. Goodrich, Nodari Sitchinava, and Qin Zhang. Sorting, searching, and simulation
in the MapReduce framework. In International Symposium on Algorithms and Computation,
pages 374–383. Springer, 2011.
[HKP99] Michał Ha´n´ckowiak, Michał Karo´nski, and Alessandro Panconesi. A faster distributed algo-
rithm for computing maximal matchings deterministically.
In Proceedings of the Eighteenth
Annual ACM Symposium on Principles of Distributed Computing, PODC '99, pages 219–228,
New York, NY, USA, 1999. ACM.
[HKP01] Michał Ha´n´ckowiak, Michał Karo´nski, and Alessandro Panconesi. On the distributed complex-
ity of computing maximal matchings. SIAM Journal on Discrete Mathematics, 15(1):41–57,
2001.
[HRVZ15] Zengfeng Huang, Božidar Radunovi´c, Milan Vojnovi´c, and Qin Zhang. Communication com-
plexity of approximate matching in distributed graphs.
In Proceedings of the 32nd Interna-
tional Symposium on Theoretical Aspects of Computer Science, STACS 2015, March 4–7, 2015,
Garching, Germany, pages 460–473, 2015.
42
[IBY+07] Michael Isard, Mihai Budiu, Yuan Yu, Andrew Birrell, and Dennis Fetterly. Dryad: Distributed
data-parallel programs from sequential building blocks. SIGOPS Operating Systems Review,
41(3):59–72, March 2007.
[II86]
[IS86]
Amos Israeli and Alon Itai. A fast and simple randomized parallel algorithm for maximal
matching. Information Processing Letters, 22(2):77–80, 1986.
Amos Israeli and Yossi Shiloach. An improved parallel algorithm for maximal matching. Infor-
mation Processing Letters, 22(2):57–60, 1986.
[KKS14] Michael Kapralov, Sanjeev Khanna, and Madhu Sudan. Approximating matching size from
random streams. In Proceedings of the 25th Annual ACM-SIAM Symposium on Discrete Algo-
rithms, SODA 2014, Portland, Oregon, USA, January 5–7, 2014, pages 734–751, 2014.
[KMW06] Fabian Kuhn, Thomas Moscibroda, and Roger Wattenhofer. The price of being near-sighted.
In Proceedings of the 17th Annual ACM-SIAM Symposium on Discrete Algorithm, SODA 2006,
pages 980–989, Philadelphia, PA, USA, 2006. Society for Industrial and Applied Mathematics.
[KSV10] Howard J. Karloff, Siddharth Suri, and Sergei Vassilvitskii. A model of computation for MapRe-
duce. In Proceedings of the 21st Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2010, Austin, Texas, USA, January 17–19, 2010, pages 938–948, 2010.
[LMSV11] Silvio Lattanzi, Benjamin Moseley, Siddharth Suri, and Sergei Vassilvitskii. Filtering: a method
for solving graph problems in MapReduce. In Proceedings of the 23rd Annual ACM Symposium
on Parallelism in Algorithms and Architectures, SPAA 2011, San Jose, CA, USA, June 4–6, 2011,
pages 85–94, 2011.
[LPP15]
Zvi Lotker, Boaz Patt-Shamir, and Seth Pettie.
Journal of the ACM, 62(5):38:1–38:17, November 2015.
Improved distributed approximate matching.
[Lub86] Michael Luby. A simple parallel algorithm for the maximal independent set problem. SIAM
Journal on Computing, 15(4):1036–1053, 1986.
[McG05] Andrew McGregor. Finding graph matchings in data streams. In Approximation, Randomization
and Combinatorial Optimization, Algorithms and Techniques, 8th International Workshop on
Approximation Algorithms for Combinatorial Optimization Problems, APPROX 2005 and 9th
InternationalWorkshop on Randomization and Computation, RANDOM 2005, Berkeley, CA,
USA, August 22-24, 2005, Proceedings, pages 170–181, 2005.
[OR10]
Krzysztof Onak and Ronitt Rubinfeld. Maintaining a large matching and a small vertex cover.
In Proceedings of the 42nd ACM Symposium on Theory of Computing, STOC 2010, Cambridge,
Massachusetts, USA, 5–8 June 2010, pages 457–464, 2010.
[PR07]
Michal Parnas and Dana Ron. Approximating the minimum vertex cover in sublinear time and a
connection to distributed algorithms. Theoretical Computer Science, 381(1-3):183–196, 2007.
[RVW16] Tim Roughgarden, Sergei Vassilvitskii, and Joshua R. Wang. Shuffles and circuits: On lower
bounds for modern parallel computation. In Proceedings of the 28th ACM Symposium on Par-
allelism in Algorithms and Architectures, SPAA 2016, Asilomar State Beach/Pacific Grove, CA,
USA, July 11–13, 2016, pages 1–12, 2016.
43
[Whi12]
Tom White. Hadoop: The Definitive Guide. O'Reilly Media, Inc., 2012.
[ZCF+10] Matei Zaharia, Mosharaf Chowdhury, Michael J. Franklin, Scott Shenker, and Ion Stoica. Spark:
Cluster computing with working sets. In 2nd USENIX Workshop on Hot Topics in Cloud Com-
puting, HotCloud'10, Boston, MA, USA, June 22, 2010.
44
|
1207.5215 | 2 | 1207 | 2012-07-30T07:37:53 | Density Functions subject to a Co-Matroid Constraint | [
"cs.DS",
"cs.DM"
] | In this paper we consider the problem of finding the {\em densest} subset subject to {\em co-matroid constraints}. We are given a {\em monotone supermodular} set function $f$ defined over a universe $U$, and the density of a subset $S$ is defined to be $f(S)/\crd{S}$. This generalizes the concept of graph density. Co-matroid constraints are the following: given matroid $\calM$ a set $S$ is feasible, iff the complement of $S$ is {\em independent} in the matroid. Under such constraints, the problem becomes $\np$-hard. The specific case of graph density has been considered in literature under specific co-matroid constraints, for example, the cardinality matroid and the partition matroid. We show a 2-approximation for finding the densest subset subject to co-matroid constraints. Thus, for instance, we improve the approximation guarantees for the result for partition matroids in the literature. | cs.DS | cs | Density Functions subject to a Co-Matroid
Constraint∗
Venkatesan T. Chakaravarthy1, Natwar Modani1,
Sivaramakrishnan R. Natarajan2, Sambuddha Roy1, and Yogish
Sabharwal1
1
2
IBM Research, New Delhi, India
{vechakra,namodani,sambuddha,ysabharwal}@in.ibm.com
IIT Chennai, India
[email protected]
Abstract
In this paper we consider the problem of finding the densest subset subject to co-matroid con-
straints. We are given a monotone supermodular set function f defined over a universe U , and the
density of a subset S is defined to be f (S)/S. This generalizes the concept of graph density. Co-
matroid constraints are the following: given matroid M a set S is feasible, iff the complement of S
is independent in the matroid. Under such constraints, the problem becomes NP-hard. The spe-
cific case of graph density has been considered in literature under specific co-matroid constraints,
for example, the cardinality matroid and the partition matroid. We show a 2-approximation for
finding the densest subset subject to co-matroid constraints. Thus, for instance, we improve the
approximation guarantees for the result for partition matroids in the literature.
1
Introduction
In this paper, we consider the problem of computing the densest subset with respect to a
monotone supermodular function subject to co-matroid constraints. Given a universe U of
n elements, a function f : 2U → R+ is supermodular iff
f (A) + f (B) 6 f (A ∪ B) + f (A ∩ B)
for all A, B ⊆ U . If the sign of the inequality is reversed for all A, B, then we call the function
submodular. The function f is said to be monotone if f (A) 6 f (B) whenever A ⊆ B; we
assume f (∅) = 0. We define a density function d : 2U → R+ as d(S) , f (S)/S. Consider
the problem of maximizing the density function d(S) given oracle access to the function f .
We observe that the above problem can be solved in polynomial time (see Theorem 6).
The main problem considered in this paper is to maximize d(S) subject to certain
constraints that we call co-matroid constraints.
In this scenario, we are given a matroid
M = (U, I) where I ⊆ 2U is the family of independent sets (we give the formal definition of
a matroid in Section 2). A set S is considered feasible iff the complement of S is independent
i.e. S ∈ I. The problem is to find the densest feasible subset S given oracle access to f and
M. We denote this problem as DEN-M.
We note that even special cases of the DEN-M problem are NP-hard [14]. The main result
in this paper is the following:
◮ Theorem 1. Given a monotone supermodular function f over a universe U , and a matroid
M defined over the same universe, there is a 2-approximation algorithm for the DEN-M
problem.
∗ Work done by the third author while he was interning at IBM Research
2
1
0
2
l
u
J
0
3
]
S
D
.
s
c
[
2
v
5
1
2
5
.
7
0
2
1
:
v
i
X
r
a
2
Density Functions
Alternatively one could have considered the same problem under matroid constraints
(instead of co-matroid constraints). We note that this problem is significantly harder, since
the Densest Subgraph problem can be reduced to special cases of this problem (see [2, 14]).
The Densest Subgraph problem is notoriously hard: the best factor approximation known
to date is O(n1/4+ǫ) for any ǫ > 0 [3].
Special cases of the DEN-M problem have been extensively studied in the context of graph
density, and we discuss this next.
1.1 Comparison to Graph Density
Given an undirected graph G = (V, E), the density d(S) of a subgraph on vertex set S is
defined as the quantity E(S)
, where E(S) is the set of edges in the subgraph induced by the
S
vertex set S. The densest subgraph problem is to find the subgraph S of G that maximizes
the density.
The concept of graph density is ubiquitous, more so in the context of social networks.
In the context of social networks, the problem is to detect communities: collections of
individuals who are relatively well connected as compared to other parts of the social network
graph.
The results relating to graph density have been fruitfully applied to finding communities
in the social network graph (or even web graphs, gene annotation graphs [15], problems
related to the formation of most effective teams [9], etc.). Also, note that graph density
appears naturally in the study of threshold phenomena in random graphs, see [1].
Motivated by applications in social networks, the graph density problem and its variants
have been well studied. Goldberg [11] proved that the densest subgraph problem can be
solved optimally in polynomial time: he showed this via a reduction to a series of max-flow
computations. Later, others [6, 14] have given new proofs for the above result, motivated
by considerations to extend the result to some generalizations and variants.
Andersen and Chellapilla [2] studied the following generalization of the above problem.
Here, the input also includes an integer k, and the goal is to find the densest subgraph
S subject to the constraint S > k. This corresponds to finding sufficiently large dense
subgraphs in social networks. This problem is NP-hard [14]. Andersen and Chellapilla [2]
gave a 2-approximation algorithm. Khuller and Saha [14] give two alternative algorithms:
one of them is a greedy procedure, while the other is LP-based. Both the algorithms have
2-factor guarantees.
Gajewar and Sarma [9] consider a further generalization. The input also includes a
partition of the vertex set into U1, U2, · · · , Ut, and non-negative integers r1, r2, · · · , rt. The
goal is to find the densest subgraph S subject to the constraint that for all 1 6 i 6 t,
S ∩ Ui > ri. They gave a 3-approximation algorithm by extending the greedy procedure of
Khuller and Saha [14].
We make the following observations: (i) The objective function E(S) is monotone and
supermodular. (ii) The constraint S > k (considered by [2]) is a co-matroid constraint;
this corresponds to the cardinality matroid. (iii) The constraint considered by Gajewar and
Sarma [9] is also a co-matroid constraint; this corresponds to the partition matroid (formal
definitions are provided in Section 2). Consequently, our main result Theorem 1 improves
upon the above results in three directions:
Objective function: Our results apply to general monotone supermodular functions f
instead of the specific set function E(S) in graphs.
Constraints: We allow co-matroid constraints corresponding to arbitrary matroids.
V. Chakaravarthy et. al
3
Approximation Factor: For the problem considered by Gajewar and Sarma [9], we im-
prove the approximation guarantee from 3 to 2. We match the best factor known for the
at-least-k densest subgraph problem considered in [2, 14].
1.2 Other Results
Knapsack Covering Constraints:
We also consider the following variant of the DEN-M problem. In this variant, we will
have a weight wi (for i = 1, · · · , U ) for every element i ∈ U , and a number k ∈ N. A set S
of elements is feasible if and only if the following condition holds:
X
i∈S
wi > k
We call this a knapsack covering constraint. We extend the proof of Theorem 1 to show the
following:
◮ Theorem 2. Suppose we are given a monotone supermodular function f over a universe U ,
weights wi for every element i ∈ U , and a number k ∈ N. Then there is a 3-approximation
algorithm for maximizing the density function d(S) subject to knapsack covering constraints
corresponding to the weights wi and the number k.
Dependency Constraints:
Saha et. al[15] consider a variant of the graph density problem. In this version, we are
given a specific collection of vertices A ⊆ V ; a subset S of vertices is feasible iff A ⊆ S.
We call this restriction the subset constraint. The objective is to find the densest subgraph
among subsets satisfying a subset constraint. Saha et. al[15] prove that this problem is
solvable in polynomial time by reducing this problem to a series of max-flow computations.
We study a generalization of the subset constraint problem. Here, we are given a mono-
tone supermodular function f defined over universe U . Additionally, we are given a directed
graph D = (U, ~A) over the universe U . A feasible solution S has to satisfy the following
if a ∈ S, then every vertex of the digraph D reachable from a also has to be-
property:
long to S. Alternatively, a ∈ S and (a, b) ∈ ~A implies that b ∈ S. We call the digraph
D as the dependency graph and such constraints as dependency constraints. The goal is to
find the densest subset S subject to the dependency constraints. We call this the DENdep
problem. We note that the concept of dependency constraints generalizes that of the subset
constraints: construct a digraph D by drawing directed arcs from every vertex in U to every
vertex in A. The motivation for this problem comes from certain considerations in social
networks, where we are to find the densest subgraph but with the restriction that in the
solution subgraph all the members of a sub-community (say, a family) are present or absent
simultaneously. In literature, such a solution S that satisfies the dependency constraints is
also called a closure (see [18], Section 3.7.2). Thus our problem can be rephrased as that of
finding the densest subset over all closures.
We note that dependency constraints are incomparable with co-matroid constraints. In
it is not true that if S is a
fact dependency constraints are not even upward monotone:
feasible subset, any superset of S is feasible.
Our result is as follows:
◮ Theorem 3. The DENdep problem is solvable in polynomial time.
The salient features of the above result are as follows:
4
Density Functions
While the result in [15] is specific to graph density, our result holds for density functions
arising from arbitrary monotone supermodular functions.
Our proof of this result is LP-based. The work of [15] is based on max-flow computations.
We can extend our LP-based approach (via convex programs) to the case for density
functions arising from arbitrary monotone supermodular f , while we are not aware as to
how to extend the max-flow based computation.
The proof technique, inspired by Iwata and Nagano [13] also extends to show "small
support" results: thus, for instance, we can show that for the LP considered by [14] for
the at-least-k-densest subgraph problem, every non-zero component of any basic feasible
solution is one of two values.
Combination of Constraints:
We also explore the problem of finding the densest subset subject to a combination of the
constraints considered. We are able to prove results for the problem of maximizing a density
function subject to (a) co-matroid constraints and (b) subset constraints. Suppose we are
given a monotone supermodular function f over a universe U , a matroid M = (U, I), and
a subset of elements A ⊆ U . A subset S is called feasible iff (1) S satisfies the co-matroid
constraints wrt M (i.e. S ∈ I) and (2) S satisfies the subset constraint wrt A (i.e. A ⊆ S).
We show the following:
◮ Theorem 4. There is a 2-approximation algorithm for the problem of maximizing the
density function d(S) corresponding to a monotone supermodular function f , subject to the
co-matroid and subset constraints.
1.3 Related Work
Recently, there has been a considerable interest in the problems of optimizing submodular
functions under various types of constraints. The most common constraints that are con-
sidered are matroid constraints, knapsack constraints or combinations of the two varieties.
Thus for instance, Calinescu et. al [5] considered the problem of maximizing a monotone
submodular function subject to a matroid constraint. They provide an algorithm and show
that it yields a (1 − 1/e)-approximation: this result is essentially optimal (also see the re-
cent paper [8] for a combinatorial algorithm for the same). Goemans and Soto [10] consider
the problem of minimizing a symmetric submodular function subject to arbitrary matroid
constraints. They prove the surprising result that this problem can be solved in polyno-
mial time. In fact, their result extends to the significantly more general case of hereditary
constraints; the problem of extending our results to arbitrary hereditary functions is left
open.
The density functions that we consider may be considered as "close" to the notion of
supermodular functions. To the best of our knowledge, the general question of maximizing
density functions subject to a (co-)matroid constraint has never been considered before.
1.4 Proof Techniques
We employ a greedy algorithm to prove Theorems 1 and 2. Khuller and Saha [14] and
Gajewar and Sarma [9] had considered a natural greedy algorithm for the problem of max-
imizing graph density subject to co-matroid constraints corresponding to the cardinality
matroid and partition matroid respectively. Our greedy algorithm can be viewed as an ab-
straction of the natural greedy algorithm to the generalized scenario of arbitrary monotone
supermodular functions. However, our analysis is different from that in [14, 9]: the efficacy
V. Chakaravarthy et. al
5
of our analysis is reflected in the fact that we improve on the guarantees provided by [9].
While they provide a 3-approximation algorithm for the graph density problem with par-
tition matroid constraints, we use the modified analysis to obtain a 2-factor guarantee. In
both of the earlier papers [14, 9], a particular stopping condition is employed to define a set
Dℓ useful in the analysis. For instance, in Section 4.1 of [9] they define Dℓ using the optimal
set H ∗ directly. We choose a different stopping condition to define the set Dℓ; it turns out
that this choice is crucial for achieving a 2-factor guarantee.
We prove Theorem 3 using LP-based techniques. In fact, we provide two proofs for the
same. Both our techniques also provide alternate new proofs of the basic result that graph
density is computable in polynomial time. The first proof method is inspired by Iwata and
Nagano [13]. The second proof method invokes Cramer's rule to derive the conclusion.
1.5 Organization
We present the relevant definitions in Section 2. We proceed to give the proof of Theorem 1
in Section 3, while the proof of Theorem 2 is presented in Section 4. The proof of Theorem 3
is presented in Section 5. and the proof of Theorem 4 is in Section 6.
2
Preliminaries
In this paper, we will use the following notation: given disjoint sets A and B we will use
A+B to serve as shorthand for A∪B. Vice versa, when we write A+B it will hold implicitly
that the sets A and B are disjoint.
Monotone: A set function f is called monotone if f (S) 6 f (T ) whenever S ⊆ T .
Supermodular: A set function f : 2U → R+ over a universe U is called supermodular if
the following holds for any two sets A, B ⊆ U :
f (A) + f (B) 6 f (A ∪ B) + f (A ∩ B)
If the inequality holds (for every A, B) with the sign reversed, then the function f is called
submodular. In this paper, we will use the following equivalent definition of supermodularity:
given disjoint sets A, B and C,
f (A + C) − f (A) 6 f (A + B + C) − f (A + B)
We can think of this as follows: the marginal utility of the set of elements C to the set A
increases as the set becomes "larger" (A + B instead of A).
It is well known (see [12, 17]) that supermodular functions can be maximized in poly-
nomial time (whereas submodular functions can be minimized in polynomial time). Let us
record this as:
◮ Theorem 5. Any supermodular function f : 2U → R+ can be maximized in polynomial
time.
We also state the following folklore corollary:
◮ Corollary 6. Given any supermodular function f : 2U → R+, we can find maxS
polynomial time.
f (S)
S
in
For completeness, a proof of this Corollary is included in Section 5.2.
Density Function: Given a function f over U , the density of a set S is defined to be
d(S) = f (S)
S .
Matroid: A matroid is a pair M = (U, I) where I ⊆ 2U , and
6
Density Functions
i ← 1
Hi ← arg maxX
Di ← Hi
while Di infeasible do
f (X)
X
Hi+1 ← arg maxX:X∩Di=∅
Di+1 ← Di + Hi+1
i ← i + 1
end while
L ← i
for i = 1 → L do
f (Di+X)−f (Di)
X
Add arbitrary vertices to Di to make it minimal feasible
Call the result D′
i
end for
Output the subset among the D′
i's with the highest density
Figure 1 Main Algorithm
1. (Hereditary Property) ∀B ∈ I, A ⊂ B =⇒ A ∈ I.
2. (Extension Property) ∀A, B ∈ I : A < B =⇒ ∃x ∈ B \ A : A + x ∈ I
Matroids are generalizations of vector spaces in linear algebra and are ubiquitous in combin-
atorial optimization because of their connection with greedy algorithms. Typically the sets
in I are called independent sets, this being an abstraction of linear independence in linear
algebra. The maximal independent sets in a matroid are called the bases (again preserving
the terminology from linear algebra). An important fact for matroids is that all bases have
equal cardinality -- this is an outcome of the Extension Property of matroids.
Any matroid is equipped with a rank function r : 2U → R+. The rank of a subset S
is defined to be the size of the largest independent set contained in the subset S. By the
Extension Property, this is well-defined. See the excellent text by Schrijver [16] for details.
Two commonly encountered matroids are the (i) Cardinality Matroid: Given a universe U
and r ∈ N, the cardinality matroid is the matroid M = (U, I), where a set A is independent
(i.e. belongs to I) iff A 6 r. (ii) Partition Matroid: Given a universe U and a partition of
U as U1, · · · , Ur and non-negative integers r1, · · · , rt, the partition matroid is M = (U, I),
where a set A belongs to I iff A ∩ Ui 6 ri for all i = 1, 2, · · · , t.
Convex Programs: We will need the definition of a convex program, and that they can
be solved to arbitrary precision in polynomial time, via the ellipsoid method(see [12]). We
refer the reader to the excellent text [4].
3
Proof of Theorem 1
We first present the algorithm and then its analysis. To get started, we describe the intuition
behind the algorithm.
Note that co-matroid constraints are upward monotone:
if a set S is feasible for such
constraints, then any superset of S is also feasible. Thus, it makes sense to find a maximal
subset of U with the maximum density. In the following description of the algorithm, one
may note that the sets D1, D2, · · · , Di are an attempt to find the maximal subset with the
largest density. Given this rough outline, the algorithm is presented in Figure 1.
We note that we can find the maximum maxX:X∩Di=∅
f (Di+X)−f (Di)
X
in polynomial time.
V. Chakaravarthy et. al
7
This is because the function f (Di + X) for a fixed Di is supermodular (and we appeal to
Corollary 6).
Let H ∗ denote the optimal solution, i.e. the subset that maximizes the density d(S)
subject to the co-matroid constraints. Let d∗ denote the optimal density, so that f (H ∗) =
d∗·H ∗.
We can make the following easy claim:
◮ Claim 1. The subset D1 obeys the inequality d(D1) > d∗.
This is because D1 is the densest subset in the universe U , while d∗ is the density of a
specific subset H ∗.
In the following, we will have occasion to apply the following lemmas.
◮ Lemma 7. Let a, b, c, d, θ ∈ R+ be such that the inequalities a
b
it is true that a+c
b+d
> θ. Thus, if a
b
d , then a+c
d (by setting θ = c
> θ and c
d
> c
> c
b+d
d ).
> θ hold. Then
Also,
◮ Lemma 8. Let a, b, c, d ∈ R+ be real numbers such that a
b
> a
b holds.
> c−a
d−b holds.
Suppose a > c, b > d. Then the inequality a−c
b−d
Suppose c > a, d > b. Then the inequality c
d
> c
d holds.
We make the following claim:
◮ Claim 2. The sequence of subsets D1, D2, · · · , DL obeys the following ordering:
f (D1)
D1
>
f (D2) − f (D1)
D2 − D1
> · · · >
f (Di+1) − f (Di)
Di+1 − Di
> · · · >
f (DL) − f (DL−1)
DL − DL−1
Proof. Consider any term in this sequence, say f (Di+1)−f (Di)
Di+1−Di
as arg max of f (Di+X)−f (Di)
f (Di+X)−f (Di)
quantity is larger than f (Di+2)−f (Di)
Di+2−Di
part of Lemma 8, we get that
. Therefore, maxX
X
X
. Note that Hi+1 was chosen
= f (Di+1)−f (Di)
Di+1−Di
. Hence this
(as long as Di+2 is well defined). Now from the second
f (Di+1) − f (Di)
Di+1 − Di
>
f (Di+2) − f (Di)
Di+2 − Di
>
f (Di+2) − f (Di+1)
Di+2 − Di+1
◭
Via an application of Lemma 7, we then have:
◮ Claim 3. Given any i (1 6 i 6 L), the following holds:
f (Di)
Di
>
f (Di) − f (Di−1)
Di − Di−1
Proof. We will the prove the statement by induction.
Base Case: We implicitly assume that D0 = ∅, and hence the case for i = 1 holds.
Induction Step: Assume the statement by induction for i = k, and we prove it for i = k+1.
Thus, by hypothesis we have
f (Dk)
Dk
>
f (Dk) − f (Dk−1)
Dk − Dk−1
Now by Claim 2 we have that
f (Dk) − f (Dk−1)
Dk − Dk−1
>
f (Dk+1) − f (Dk)
Dk+1 − Dk
8
Density Functions
Thus,
f (Dk)
Dk
>
f (Dk+1) − f (Dk)
Dk+1 − Dk
Applying Lemma 7, we get:
f (Dk+1)
Dk+1
>
f (Dk+1) − f (Dk)
Dk+1 − Dk
Thus we have proven the Claim by induction.
◭
The analysis will be broken up into two parts. We will consider the set Dℓ in the sequence
D1, D2, · · · , DL such that the following hold:
f (Dℓ) − f (Dℓ−1)
Dℓ − Dℓ−1
>
d∗
2
but
f (Dℓ+1) − f (Dℓ)
Dℓ+1 − Dℓ
<
d∗
2
Since d(D1) ≥ d∗ by Claim 1, such an ℓ will exist or ℓ = L. If ℓ = L, then we have a feasible
solution DL with the property that f (DL)−f (DL−1)
2 . Therefore, by Claim 3 we have
DL−DL−1
that d(DL) > d∗
2 and we are done in this case.
> d∗
D′
So we may assume that ℓ < L so that Dℓ is not feasible. In this case, we will prove that
ℓ has the correct density, i.e. that d(D′
To this end, we will prove two facts about Dℓ and that will yield the desired result:
ℓ) > d∗
2 .
◮ Claim 4.
f (Dℓ) − f (Dℓ ∩ H ∗) >
d∗
2
(Dℓ − Dℓ ∩ H ∗)
Proof. Note that Dℓ = H1 + H2 + · · · + Hℓ. For brevity, for 1 6 i 6 ℓ, denote Hi ∩ H ∗ as
Ai (thus, Ai ⊆ Hi for every i). Thus, Dℓ ∩ H ∗ = A1 + A2 + · · · + Aℓ.
We will prove the following statement by induction on i (for 1 6 i 6 ℓ):
f (H1 +H2 +· · ·+Hi)−f (A1 +A2 +· · ·+Ai) >
d∗
2
(H1 +H2 +· · ·+Hi−A1 +A2 +· · ·+Ai)
Base Case: For i = 1, we have to prove that:
f (H1) − f (A1)
H1 − A1
>
d∗
2
Since H1 is the densest subset, we have
f (H1)
H1
>
f (A1)
A1
and we may apply (the first part of) Lemma 8 to obtain the desired.
Induction Step: Assume the statement to be true for i, and we will prove it for i + 1.
V. Chakaravarthy et. al
9
Consider the following chain:
f (H1 + · · · + Hi + Hi+1) − f (H1 + · · · + Hi)
Hi+1
f (H1 + · · · + Hi + Ai+1) − f (H1 + · · · + Hi)
Ai+1
f (A1 + · · · + Ai + Ai+1) − f (A1 + · · · + Ai)
Ai+1
Hi+1arg max
>
supermodular
>
We would now like to apply Lemma 8 to the first and last terms in the above chain. To
this end, let us check the preconditions:
f (H1 + · · · + Hi + Hi+1) − f (H1 + · · · + Hi)
supermodular
monotone
>
f (H1 + · · · + Hi + Ai+1) − f (H1 + · · · + Hi)
>
f (A1 + · · · + Ai + Ai+1) − f (A1 + · · · + Ai)
Also, clearly, Hi+1 > Ai+1.
Thus, the preconditions in Lemma 8 hold, and we have that
f (H1 + · · · + Hi+1) − f (A1 + · · · + Ai+1) − f (H1 + · · · + Hi) + f (A1 + · · · + Ai)
f (H1 + · · · + Hi + Hi+1) − f (H1 + · · · + Hi)
Hi+1
>
d∗
2
Hi+1 − Ai+1
>
Applying Lemma 7 to the first term in the above chain and the induction statement for
◭
i, we obtain the desired result for i + 1. Hence done.
The next claim lower bounds the value f (Dℓ ∩ H ∗).
Building up to the Claim, let us note that Dℓ ∩ H ∗ 6= ∅. If the intersection were empty,
then H ∗ is a subgraph of density d∗, and so Hℓ+1 would be a subgraph of density at least
d∗. But then,
f (Dℓ + Hℓ+1) − f (Dℓ)
Hℓ+1
supermodular
>
f (Hℓ+1)
Hℓ+1
> d∗
But this contradicts the choice of Dℓ.
◮ Claim 5.
f (Dℓ ∩ H ∗) >
d∗
2
Dℓ ∩ H ∗ +
d∗
2
H ∗
Proof. Let X = H ∗ −Dℓ ∩H ∗. Then, X ∩Dℓ = ∅ and Dℓ +X = Dℓ ∪H ∗. Then by definition
of Dℓ, we know that f (Dℓ+X)−f (Dℓ)
Dℓ+1−Dℓ < d∗/2. Thus, f (Dℓ ∪ H ∗) − f (Dℓ) 6
d∗
2 (H ∗ − Dℓ ∩ H ∗).
6 f (Dℓ+1)−f (Dℓ)
X
Therefore, f (Dℓ ∪ H ∗) + f (Dℓ ∩ H ∗) 6 f (Dℓ) + f (Dℓ ∩ H ∗) + d∗
Applying supermodularity we have that f (Dℓ ∪ H ∗) + f (Dℓ ∩ H ∗) > f (Dℓ) + f (H ∗).
2 (H ∗ − Dℓ ∩ H ∗) > f (H ∗). The claim
◭
Thus, cancelling f (Dℓ) gives us that f (Dℓ ∩ H ∗) + d∗
follows by observing that d∗ = f (H ∗)
H ∗ .
2 (H ∗ − Dℓ ∩ H ∗).
Note that this claim also implies that the density of the set Dℓ ∩ H ∗ is at least d∗.
Intuitively, Dℓ ∩ H ∗ is a subset that has "enough f -value" as well as a "good" density.
10
Density Functions
We may now combine the statements of Claim 4 and Claim 5 to get the following chain
of inequalities:
f (Dℓ)
Claim 4
> f (Dℓ ∩ H ∗) +
d∗
2
Dℓ −
d∗
2
Dℓ ∩ H ∗
Claim 5
>
d∗
2
Dℓ +
d∗
2
H ∗
Consider D′
ℓ: this is obtained from Dℓ by adding suitably many elements to make Dℓ
feasible. Let r be the minimum number of elements to be added to Dℓ so as to make it
feasible. Since H ∗ is a feasible solution too, clearly, r 6 H ∗. With this motivation, we
define the Extension Problem for a matroid M. The input is a matroid M = (U, I) and a
subset A ⊆ U . The goal is to find a subset T of minimum cardinality such that A ∪ T ∈ I.
Lemma 9 shows that we can find such a subset T in polynomial time. Thus, we would have
that:
d(D′
ℓ) =
f (D′
ℓ)
Dℓ + r
>
f (Dℓ)
Dℓ + r
>
f (Dℓ)
Dℓ + H ∗
> d∗/2
and we are done with the proof of Theorem 1, modulo the proof of Lemma 9.
We proceed to present the lemma and its proof:
◮ Lemma 9. The Extension Problem for matroid M and subset A can be solved in polyno-
mial time.
Proof. The proof considers the base polyhedron of the matroid (see the text by Schrijver [16]).
We will have a variable xi for each element i ∈ U \ A, where xi = 1 would indicate that we
pick the element i in our solution T . For brevity, we will also maintain a variable yi that
indicates whether i is absent from the solution T . Thus for every i, we will maintain that
xi + yi = 1. Given an arbitrary set S, we will let r(S) denote the rank of the subset S in
the matroid M.
The following is a valid integer program for the Extension Problem (where y(S) is short-
hand for Pi∈S yi). The linear program to the right is the relaxation of the integer program,
and with variables xi eliminated.
min X
xi
s.t.
IP1 :
i∈U
xi + yi = 1
y(S) 6 r(S)
xi = 1
xi, yi ∈ {0, 1} for all i ∈ U .
for all S ⊆ U
for all i ∈ U
for all i ∈ A
min X
(1 − yi)
LP1 :
s.t.
i∈U
y(S) 6 r(S)
yi = 0
yi > 0
for all S ⊆ U
for all i ∈ A
for all i ∈ U .
The linear program LP1 can also be formulated as a maximization question. To be
precise, let VAL(LP1) denote the value of the program LP1. Then VAL(LP1) = U −
VAL(LP2), where LP2 is as follows:
max X
yi
LP2 :
s.t.
i∈U
y(S) 6 r(S)
yi = 0
yi > 0
for all S ⊆ U
for all i ∈ A
for all i ∈ U .
Now, by folklore results in matroid theory (cf. [16]), we have that solutions to LP2 are
integral and can be found by a greedy algorithm. Thus, we can solve IP1 in polynomial
◭
time, and this proves the statement of the Lemma.
V. Chakaravarthy et. al
11
i ← 1
Hi ← arg maxX
Di ← Hi
while Di infeasible do
f (X)
X
Hi+1 ← arg maxX:X∩Di=∅
Di+1 ← Di + Hi+1
i ← i + 1
end while
L ← i
for i = 1 → L do
f (Di+X)−f (Di)
X
Order the vertices i in U \ Di by non-increasing order of weights wi
Add vertices from U \ Di in this order to Di until feasibility is attained
Let the result be D′
i
end for
Output the subset among the D′
i's with the highest density
Figure 2 Algorithm for Knapsack constraints
4
Proof of Theorem 2
In order to prove this result, we will have to modify the algorithm presented in Section 3.
In the analysis, we will correspondingly modify the definition of the set Dℓ. Then we will
apply (modified versions of) Claim 4 and Claim 5 to derive the result.
The modified algorithm is as shown in Figure 2.
Consider the set Dℓ in the sequence D1, D2, · · · , DL such that the following hold:
f (Dℓ) − f (Dℓ−1)
Dℓ − Dℓ−1
>
d∗
3
but
f (Dℓ+1) − f (Dℓ)
Dℓ+1 − Dℓ
<
d∗
3
As earlier, if there is no such ℓ < L for which this holds, this implies that DL satisfies
f (DL)−f (DL−1)
> d∗/3. But this gives a 3-approximation in this case since DL is feasible
DL−DL−1
and
d(DL) >
f (DL) − f (DL−1)
DL − DL−1
Let us consider the other case where ℓ < L and Dℓ is infeasible. Let H ∗ denote the
optimal solution.
We state modified versions of Claim 4 and Claim 5.
◮ Claim 6. (modified Claim 4)
f (Dℓ) − f (Dℓ ∩ H ∗) >
d∗
3
(Dℓ − Dℓ ∩ H ∗)
◮ Claim 7. (modified Claim 5)
f (Dℓ ∩ H ∗) >
d∗
3
Dℓ ∩ H ∗ +
2d∗
3
H ∗
12
Density Functions
These modified claims may be proven analogously to the original claims, taking into
account the new definition for Dℓ.
Now note that given a set Dℓ, in order to make the set feasible for the knapsack cover
constraint, we pick the elements with the largest weights wi so that feasibility is attained.
The usual knapsack greedy algorithm shows that this is a 2-approximation to the optimal
knapsack cover. Thus, if we add r elements, then r ≤ 2H ∗. Thus we have that,
f (D′
ℓ)
D′
ℓ
>
f (Dℓ)
Dℓ + r
>
f (Dℓ)
Dℓ + 2H ∗
>
d∗
3
Thus we have proven a 3-approximation.
5
Proof of Theorem 3
We will present the proof for the case of the graph density function, i.e. where f (S) = E(S).
The proof for arbitrary f will require a passage to the Lovász Extension Lf (x) of a set
function f (S). In fact we will present two proofs of this fact for the special case of the graph
density function. To the best of our knowledge, both the proofs are new, and seems simpler
than existing proofs. For both the proofs, we will use the same LP.
First Proof:
We will augment the LP that Charikar [6] uses to prove that graph density is computable
in polynomial time. Given a graph G = (V, E), there are edge variables ye and vertex
variables xi in the LP. We are also given an auxiliary dependency digraph D = (V, ~A) on
In the augmented LP, we also have constraints xi 6 xj if there is an
the vertex set V .
arc from i to j in the digraph D = (V, ~A). The DENdep problem is modelled by the linear
program LP3.
max X
s.t. X
e∈E
ye
xi = 1
LP3 :
i
ye 6 xi
xi 6 xj
xi > 0
for all e ∼ i, e ∈ E
for all (i, j) ∈ ~A
for all i ∈ V (G) .
CP1 :
max X
s.t. X
e=(i,j)∈E
min{xi, xj}
xi = 1
i
xi 6 xj
xi > 0
for all (i, j) ∈ ~A
for all i ∈ V (G) .
Suppose we are given an optimal solution H ∗ to the DENdep problem. Let VAL(LP3)
denote the feasible value of this LP: we will prove that VAL(LP3) = d(H ∗).
VAL(LP3) > d(H ∗):
We let H ∗ = ℓ, and xi = 1/ℓ for i ∈ H ∗, and 0 otherwise. Likewise, we set ye = 1/ℓ for
e ∈ E(H ∗), and 0 otherwise. Note that H ∗ is feasible, so if a ∈ H ∗ and (a, b) ∈ ~A, then it
also holds that b ∈ H ∗. We may check that the assignment x and y is feasible for the LP.
So, d(H ∗) = E(H ∗)
VAL(LP3) 6 d(H ∗):
is achieved as the value of a feasible assignment to the LP.
ℓ
In the rest of the proof, we will prove that there exists a subgraph H such that VAL 6
d(H). First, it is easy to observe that in any optimal solution of the above LP, the variables
ye will take the values min{xi, xj} where e = (i, j). Thus, we may eliminate the variables ye
from the program LP3 to obtain the program CP1. We claim that CP1 is a convex program.
Given two concave functions, the min operator preserves concavity. Thus, the objective
function of the above modified program is concave. Hence we have a convex program: here,
the objective to be maximized is concave, subject to linear constraints. We may solve the
V. Chakaravarthy et. al
13
program CP1 and get an output optimal solution x∗. Relabel the vertices of V such that the
following holds: x∗
n. If there are two vertices with (modified) indices a and b
where a < b and there is an arc (a, b) ∈ ~A, then we have the equalities x∗
a+1 = · · · = x∗
b .
We will replace the inequalities in the program CP1 as follows:
2 > · · · > x∗
a = x∗
1 > x∗
max X
s.t. X
xi = 1
e=(i,j)∈E:i<j
xj
LP4 :
i
xi > xi+1
xn > 0
for all i ∈ {1, 2, · · · , (n − 1)}
.
where some of the inequalities xi > xi+1 may be equalities if there is an index a with a 6 i
and an index b with b > (i+1) such that (a, b) ∈ ~A. Note also that because of the ordering of
the variables of this LP, the objective function also simplifies and becomes a linear function.
Clearly x∗ is a feasible solution to this LP. Thus the value of this LP is no less than the
value of CP1. Consider a BFS x to LP4. The program LP4 has (n + 1) constraints, and
n variables. Given the BFS x, call a constraint non-tight if it does not hold with equality
under the solution x. Thus, there may be at most one non-tight constraint in LP4. In other
words, there is at most one constraint xi > xi+1 that is a strict inequality. This, in turn,
implies that all the non-zero values in x are equal. Let there be ℓ such non-zero values.
From the equality Pi xi = 1, we get that each non-zero xi = 1/ℓ. Let H denote the set of
indices i which have non-zero xi values. Then the objective value corresponding to this BFS
x is E(H)/ℓ = d(H).
Thus we have proven that d(H) > VAL(LP4) > VAL(CP1) = VAL(LP3), as required.
This completes the proof of Theorem 3.
◭
Remarks about the proof:
We remark that the objective in the convex program CP1 is precisely the Lovász Ex-
tension Lf (x) for the specific function f = E(S). Thus our proof shows that the LP
provided by Charikar [6] is precisely the Lovász Extension for the specific supermodular
function E(S).
Note that there are other proofs possible for this result. For instance, one can follow the
basic argument of Charikar to show that LP3 satisfies d(H ∗) = VAL(LP3). The proof
we provide above is new, and is inspired by the work of Iwata and Nagano [13].
Via our proof, we also prove that any BFS for the basic graph density LP has the property
that all the non-zero values are equal. This fact is not new: it was proven by Khuller
and Saha [14] but we believe our proof of this fact is more transparent.
Second Proof:
Again, let us consider the program LP3.
Similar to the above, it will suffice to prove that the BFS solutions to this LP have the
property that all non-zero components are equal.
Consider the constraint matrix B that consists of the LHS of the non-trivial constraints in
the above LP, without the constraint Pi xi = 1. Thus B consists of rows for the constraints
ye 6 xi (for e ∈ E : e ∼ i) and the constraints xi 6 xj for (i, j) ∈ ~A. The matrix B is TUM:
this is because it can easily be realised as the incidence matrix of a digraph.
Thus the original constraint matrix consists of the matrix B augmented by a single (non-
trivial) constraint, consisting of the sum of the xi's being equal to 1; and also the (trivial)
14
Density Functions
nonnegativity constraints xi ≥ 0 and ye ≥ 0. Let B′ denote this augmented matrix. Note
that B′ need not be TUM.
Consider a basic feasible solution (BFS) ev = (y1, · · · , ye, · · · , ym, x1, · · · , xi, · · · , xn).
Since ev consists of (m + n) variables, there are (m + n) constraints in the constraint matrix
B′ that are tight. Consider the submatrix T formed by the tight constraints in the matrix
B′. Since the constraint Pi xi = 1 is always tight, this will be included as a row in the
matrix T . Without loss of generality, let this row be the last row r of T . Thus, ev is the
unique solution to the linear system T v = b, where bT = (0, 0, · · · , 0, 1).
Note that, by previous considerations, the submatrix T ′ of the matrix T consisting of all
the rows of T but the last one, is TUM (since T ′ is then a submatrix of the matrix B). The
sth component (for 1 ≤ s ≤ (m + n)) of ev may be found by Cramer's rule as evs = det(Ts)
det(T ) ,
where Ts is the matrix T with the sth column replaced by the vector b.
Note that det(T ) is at most V (G) = n. This is because the row r has at exactly n 1's,
so we may expand the determinant by row r. Any sub-determinant to be computed in this
row-wise expansion of the determinant is a submatrix of T , thus is TUM. Therefore, det(T )
is a sum of at most n +1's and −1's thus, is (say) k where k ≤ n.
Consider the computation of det(Ts). The matrix Ts has its sth column replaced by the
vector b, which has precisely one 1. So we may expand the determinant of Ts by its sth
column, and thereby, the determinant is that of a square submatrix of the matrix T . This
means that det(Ts) is 0, 1 or −1.
Thus, every component of ev is precisely 0 or 1
LP is ≥ 0, thus the third possibility is excluded. This completes the proof.
k , or − 1
k . However every component in the
◭
5.1 Arbitrary monotone supermodular functions:
We now proceed to consider the case where we are given an arbitrary monotone supermodular
function f over the universe U and a directed graph D = (U, ~A), where the arcs in ~A specify
the dependencies.
To extend our results to this, we will need the concept of the Lovász Extension.
The Lovász Extension Lf : [0, 1]U → R, first defined by Lovász, is an extension of an
arbitrary set function f : 2U → R. We proceed with the formal definition:
◮ Definition 10. (Lovász Extension) Fix x ∈ [0, 1]U , and let U = {v1, v2, · · · , vn} be such
that x(v1) ≥ x(v2) ≥ · · · ≥ x(vn). For 0 ≤ i ≤ n, let Si = {v1, v2, · · · , vi}. Let {λi}n
i=0 be
the unique coefficients with λi ≥ 0, and Pi λi = 1 such that:
x =
nX
i=0
λi1Si
It is easy to see that λn = x(vn), and for 0 ≤ i < n, we have λi = x(vi) − x(vi+1) and
λ0 = 1 − x(v1). The value of the Lovász Extension of f at x is defined as
Lf (x) = X
λif (Si)
i
For motivation behind the definition, refer to the excellent survey on submodular func-
tions by Dughmi [7].
The Lovász Extension enjoys the following properties:
Lf is concave iff f is supermodular.
If f is supermodular, the maximum value of f (S) is the same as the maximum value of
Lf (x).
V. Chakaravarthy et. al
15
Restricted to the subspace x1 ≥ x2 ≥ · · · ≥ xn, the function Lf is linear.
We are now ready to describe our convex program CP for computing the densest subset
of the universe U subject to dependency constraints. For details on convex programming,
one may consult the text [4].
The program has variables x1, x2, · · · , xn corresponding to the elements i ∈ U . Since
f (S) is supermodular, the corresponding Lovász Extension LE(x) is concave.
max L(x)
CP :
s.t.
hx,~1i = 1
xi 6 xj
xi > 0
for all (i, j) ∈ ~A
for all i ∈ V (G) .
This convex programming problem can be solved to arbitrary precision in polynomial
time by the ellipsoid method (see [12]).
1 > x∗
As in the first proof above, we will relabel the elements of the universe so that x∗
2 >
n. But now, by the property of the Lovász Extension, we see that L(x) is a linear
· · · > x∗
function in this subspace.
Now, the rest of the first proof carries over and gives us the result for arbitrary monotone
supermodular f .
5.2 Proof of Corollary 6
There are many ways to see this. One way is to consider the convex program above for the
Lovász Extension of the monotone supermodular function f .
max L(x)
CP2 :
s.t.
hx,~1i = 1
xi > xi+1
xn > 0
for all i ∈ {1, 2, · · · , (n − 1)}
.
As in the proof above, we can see that this has solutions x∗ where all the nonzero xi's are
equal, and that this corresponds to choosing a subset S so that L(x∗) = f (S)/S. Thus,
we see that maxS f (S)/S is computable in polynomial time.
Yet another way of verifying Corollary 6 is to consider the sequence of functions g(α, S) ,
f (S) − αS (for fixed α > 0). Note that each g(α, S) is supermodular for any fixed α, and
so can be maximized in polynomial time. Also observe that if maxS f (S)/S > α for some
α, then maxS g(α, S) > 0. Conversely, if maxS f (S)/S 6 α then maxS g(α, S) 6 0. Thus,
we can find maxS f (S)/S by a binary search over α and maximizing the corresponding
functions g(α, S).
6
Proof of Theorem 4
To fix the notation, in this problem, we are given a monotone supermodular function f over
a universe U , a matroid M = (U, I), and a set A ⊆ U .
16
Density Functions
The only modification that we have to make to Algorithm 1 is that we will choose the first
set H1 such that H1 maximizes the density of all subsets that contain the set A. Note that we
can do this in polynomial time. Apart from this, the construction of the sets H2, H3, · · · , HL
and the sets D1, D2, · · · , DL are the same as in Algorithm 1. So, each Di contains A, and
the candidate feasible solutions D′
i also contain A. The analysis of the modified algorithm
is the same as in Section 3. Thus we obtain a 2-approximation algorithm as promised in
Theorem 4.
7 Open Problems
One interesting open direction is to investigate the maximization of density functions subject
to combinations of constraints. In this paper, we consider the combination of a single matroid
and a subset constraint. In general, one could ask similar questions about combinations of
multiple matroid constraints or a matroid and a dependency constraint for instance. Another
open question is to derive a LP-based technique to prove the result in Theorem 1.
References
1 Noga Alon and Joel H. Spencer. The Probabilistic Method. Wiley, New York, 1992.
2 Reid Andersen and Kumar Chellapilla. Finding dense subgraphs with size bounds.
In
WAW, pages 25 -- 37, 2009.
3 Aditya Bhaskara, Moses Charikar, Eden Chlamtac, Uriel Feige, and Aravindan Vi-
jayaraghavan. Detecting high log-densities: an o(n1/4) approximation for densest k-
subgraph. In STOC, pages 201 -- 210, 2010.
Stephen Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University
Press, 2004.
4
5 Gruia Călinescu, Chandra Chekuri, Martin Pál, and Jan Vondrák. Maximizing a monotone
submodular function subject to a matroid constraint. SIAM J. Comput., 40(6):1740 -- 1766,
2011.
6 Moses Charikar. Greedy approximation algorithms for finding dense components in a graph.
7
In APPROX, pages 84 -- 95, 2000.
Shaddin Dughmi. Submodular functions: Extensions, distributions, and algorithms. a
survey. CoRR, abs/0912.0322, 2009.
8 Yuval Filmus and Justin Ward. A tight combinatorial algorithm for submodular maximiz-
ation subject to a matroid constraint. In FOCS, 2012.
9 Amita Gajewar and Atish Das Sarma. Multi-skill collaborative teams based on densest
subgraphs. In SDM, pages 165 -- 176, 2012.
10 Michel X. Goemans and José A. Soto. Symmetric submodular function minimization under
hereditary family constraints. CoRR, abs/1007.2140, 2010.
11 Andrew V. Goldberg. Finding a maximum density subgraph. Technical report, UC Berke-
ley, 1984.
12 Martin Grötschel, László Lovász, and Alexander Schrijver. The ellipsoid method and its
13
consequences in combinatorial optimization. Combinatorica, 1(2):169 -- 197, 1981.
Satoru Iwata and Kiyohito Nagano. Submodular function minimization under covering
constraints. In FOCS, pages 671 -- 680, 2009.
Samir Khuller and Barna Saha. On finding dense subgraphs. In ICALP, 2009.
14
15 Barna Saha, Allison Hoch, Samir Khuller, Louiqa Raschid, and Xiao-Ning Zhang. Dense
In RECOMB,
subgraphs with restrictions and applications to gene annotation graphs.
pages 456 -- 472, 2010.
16 A. Schrijver. Combinatorial Optimization - Polyhedra and Efficiency. Springer, 2003.
V. Chakaravarthy et. al
17
17 Alexander Schrijver. A combinatorial algorithm minimizing submodular functions in
strongly polynomial time. J. Comb. Theory, Ser. B, 80(2):346 -- 355, 2000.
18 D.M. Topkis. Supermodularity and Complementarity. Princeton University Press, 1998.
|
1704.04538 | 2 | 1704 | 2017-04-21T13:01:58 | A Simple Randomized Algorithm to Compute Harmonic Numbers and Logarithms | [
"cs.DS"
] | Given a list of N numbers, the maximum can be computed in N iterations. During these N iterations, the maximum gets updated on average as many times as the Nth harmonic number. We first use this fact to approximate the Nth harmonic number as a side effect. Further, using the fact the Nth harmonic number is equal to the natural logarithm of N plus a constant that goes to zero with N, we approximate the natural logarithm from the harmonic number. To improve accuracy, we repeat the computation over many lists of uniformly generated random numbers. The algorithm is easily extended to approximate logarithms with integer bases or rational arguments. | cs.DS | cs |
A Simple Randomized Algorithm to Compute
Harmonic Numbers and Logarithms
Ali Dasdan
KD Consulting
Saratoga, CA, USA
[email protected]
October 19, 2018
Abstract
Given a list of N numbers, the maximum can be computed in N
iterations. During these N iterations, the maximum gets updated on
average as many times as the Nth harmonic number. We first use
this fact to approximate the Nth harmonic number as a side effect.
Further, using the fact the Nth harmonic number is equal to the nat-
ural logarithm of N plus a constant that goes to zero with N, we
approximate the natural logarithm from the harmonic number. To
improve accuracy, we repeat the computation over many lists of uni-
formly generated random numbers. The algorithm is easily extended
to approximate logarithms with integer bases or rational arguments.
1
Introduction
We approximately compute the harmonic number and the natural logarithm
of an integer as a side effect of computing the maximum of a list of x numbers
randomly drawn from a uniform distribution. The key point of this compu-
tation is that it basically uses only counting. To improve accuracy, we repeat
the computation multiple times and take the average. Using the basic prop-
erties of the natural logarithm function, it is simple to extend the algorithm
to approximate the logarithms with integer bases or rational arguments. The
details follow.
1
2 Computing the Maximum
Figure 1: A Python function to compute the maximum over a list of x
numbers in the interval [0.0, 1.0), where x > 0.
Given a list of x numbers in the interval [0.0, 1.0), where x > 0, the algo-
rithm (written in the Python programming language) in Figure 1 computes
the maximum in x iterations. It is well known that during these iterations,
the maximum gets updated Hx times on average, where Hx is the xth har-
monic number [1]. The reason for this fact is that in a list of x numbers
randomly drawn from a uniform distribution, each number has a probability
of 1/x of being the maximum.
3 Computing the Harmonic Number and the
Natural Logarithm
The xth harmonic number Hx is defined as the series
x(cid:88)
i=1
Hx =
1
i
= ln(x) + γ + x
(1)
where γ is the Euler-Mascheroni constant (roughly equal to 0.57721) and x,
which is in the interval (
2x), approaches 0 as x goes to infinity [2].
1
2(x+1) , 1
We can rewrite this equation to compute ln(x) as
ln(x) = Hx − γ − x.
(2)
This means an approximation to Hx can be converted to an approximation
to ln(x).
The algorithm (written in the Python programming language) is given in
Figure 2. The inner loop computes the maximum over x uniformly random
2
Figure 2: A Python function to approximate the natural logarithm as an
average over n iterations of the maximum computation. Each maximum
computation goes over x uniformly random numbers and produces a new
approximation to the xth harmonic number Hx.
numbers. The outer loop with n iterations is for accuracy improvement; it
computes an approximation to Hx every iteration as a side effect of the max-
imum computation rather than the result of the summation in Equation 1.
This Hx computation is an approximation due to two reasons: 1) Hx is never
an integer except for x = 1 [2], and 2) it is a probabilistic estimate. After
these loops exit, the final Hx is set to the average over all these approx-
imations. The natural logarithm is then approximated at the end of this
algorithm using Equation 2, where we set x to its upper bound of 1/(2x).
4 Results
Some results from limited experiments as shown in Figure 3 indicate that the
approximation quality is relatively good especially with larger arguments. In
this figure, the approximate ln(x) is the value computed by the algorithm in
the previous section and the library ln(x) is the log function from the Python
Math library.
3
Though we used 1000 repetitions for the results shown, separate limited
experiments show that the approximation quality gets better after as low as
10 repetitions.
Figure 3: The approximation results on powers of 4 from 1 to 8 with 1000
iterations. The library ln(x) is math.log() from Python's math library.
Both relative and absolute errors are small. Relative error decreases with
larger inputs. Note that we multiplied the absolute error by 10 to make it
visible in this plot.
5 Pros and Cons
This algorithm to approximate the harmonic number and the natural loga-
rithm takes time proportional to the product of x and n. Even for n = 1,
the time is linear in x. As such, this is probably not an efficient way of
computing the harmonic number or the natural logarithm. What is the use
of this algorithm then?
One reason, possibly the main reason, why this algorithm may be inter-
esting is that it approximately computes a function that occurs in its own
time complexity analysis. Here the functions are the harmonic number as
well as the natural logarithm. Another reason is that this algorithm uses
integer arithmetic only except for the final averaging and error computation.
Finally, this algorithm is easily parallelizable since the maximum of a list is
equal to the maximum over the maximums of parts of the list.
4
In the technical literature, there are of course many formulas and algo-
rithms for computing both functions [2]. This is expected as the harmonic
number and the natural logarithm are so fundamental. This paper is not
meant to provide any comparisons with those algorithms or to claim that it
is better; it is mainly a fun application on the use of the side effect of a well
known and simple algorithm, namely, the maximum computation.
6 Conclusions
We provide a simple algorithm that exploits the time complexity expression of
the maximum computation of a list of numbers to approximate the harmonic
number and then using it to approximate the natural logarithm. Limited
experiments show that the approximations are good with small relative and
absolute errors. We hope others may find this algorithm interesting enough
to study and potentially improve. At a minimum this paper might hopefully
inspire some exercises for students of a basic algorithms book like [1].
References
[1] T. Cormen, C. S. R. Rivest, and C. Leiserson. Introduction to Algorithms.
McGraw-Hill Higher Education, 2nd edition, 2001.
[2] J. Sondow and E. Weisstein.
Harmonic number.
MathWorld -- A Wolfram Web Resource. Wolfram, Apr
http://mathworld.wolfram.com/HarmonicNumber.html.
In From
2007.
5
|
1208.5247 | 1 | 1208 | 2012-08-26T20:03:04 | Faster Clustering via Preprocessing | [
"cs.DS",
"cs.CG"
] | We examine the efficiency of clustering a set of points, when the encompassing metric space may be preprocessed in advance. In computational problems of this genre, there is a first stage of preprocessing, whose input is a collection of points $M$; the next stage receives as input a query set $Q\subset M$, and should report a clustering of $Q$ according to some objective, such as 1-median, in which case the answer is a point $a\in M$ minimizing $\sum_{q\in Q} d_M(a,q)$.
We design fast algorithms that approximately solve such problems under standard clustering objectives like $p$-center and $p$-median, when the metric $M$ has low doubling dimension. By leveraging the preprocessing stage, our algorithms achieve query time that is near-linear in the query size $n=|Q|$, and is (almost) independent of the total number of points $m=|M|$. | cs.DS | cs | Faster Clustering via Preprocessing∗
Tsvi Kopelowitz
Robert Krauthgamer
Weizmann Institute of Science
{tsvi.kopelowitz,robert.krauthgamer}@weizmann.ac.il
October 26, 2018
Abstract
We examine the efficiency of clustering a set of points, when the encompassing metric space
may be preprocessed in advance. In computational problems of this genre, there is a first stage
of preprocessing, whose input is a collection of points M ; the next stage receives as input a
query set Q ⊂ M , and should report a clustering of Q according to some objective, such as
1-median, in which case the answer is a point a ∈ M minimizing Pq∈Q dM (a, q).
We design fast algorithms that approximately solve such problems under standard clustering
objectives like p-center and p-median, when the metric M has low doubling dimension. By
leveraging the preprocessing stage, our algorithms achieve query time that is near-linear in the
query size n = Q, and is (almost) independent of the total number of points m = M .
2
1
0
2
g
u
A
6
2
]
S
D
.
s
c
[
1
v
7
4
2
5
.
8
0
2
1
:
v
i
X
r
a
∗This work was supported in part by The Israel Science Foundation (grant #452/08), by a US-Israel BSF grant
#2010418, and by the Citi Foundation.
1
Introduction
Clustering is a ubiquitous computational task of prime importance in numerous applications and
domains, including machine learning, image processing, and bioinformatics. While the clustering
problem has several variations, it often falls within the following framework of metric clustering:
given a set of points Q in a metric space (M, d), choose a set of centers C (in that same metric
space) so as to minimize some objective function of Q and the centers C. For example, in the
p-median problem, the goal is to find a set of p centers C ⊆ M that minimizes the objective
med(Q, C) := Pq∈Q d(q, C),
where we define d(q, C) := minc∈C d(q, c).
Our focus here is on understanding whether an initial preprocessing stage can speed up the
process of (metric) clustering. Concretely, we are interested in algorithms for efficient clustering
of Q when the metric M can be preprocessed in advance. Throughout, we denote the number of
center candidates by m = M , and the number of query points by n = Q. The goal is to answer
queries with time close to linear in n and (almost) independent of m. To our knowledge, no previous
research on (metric) clustering problems has addressed the issue of preprocessing. Past work has
largely focused on the offline problem, where the entire input is given at once, either because M is
implicit (e.g., a Euclidean space) or because M is given together with Q (called discrete centers).
Other past work studied the online version, where points arrive one by one (the data-stream model).
Clustering with preprocessing can model, for example, the following scenario. Consider a huge
corpus of documents (M ) with distances between the documents (d) defining a metric space. Given
a relatively small subset of the documents (Q), we may wish to quickly cluster them using centers
from the corpus. Since preprocessing needs to be done only once, it has the benefit that even a
huge corpus can be processed, by pooling together many machines or by running it for several days.
The first problem we consider is the p-median problem defined above. A second problem of
interest, called p-center, is to find a set of p centers C ⊆ M that minimizes the objective
cntr(Q, C) := maxq∈Q{d(q, C)}.
Observe that when n = 1 and p = 1, both the p-median and p-center problems receive a sin-
gle input point q and seek the point of M that is closest to q, which is precisely the famous
nearest neighbor search (NNS) problem. Even for this special case of NNS (i.e., n = p = 1),
Krauthgamer and Lee [KL04a] have shown that achieving approximation factor better than 7
5 in a
general metric requires query time that depends on the doubling dimension of the metric, regard-
less of the preprocessing. (Throughout, we denote the doubling dimension by ddim = ddim(M );
see Section 1.4 for a formal definition.) It thus follows that for general n and p, one must con-
sider metrics M whose doubling dimension is bounded, and we indeed assume as such. We also
assume that computing the distance between two points takes O(1) time. This whole approach
follows an established line of research that covers a host of problems including nearest neighbor
search [KL04b, HM06, BKL06, CG06] as well as routing, distance estimation, the traveling sales-
man problem and classification see e.g. [AGGM06, KRX08], [KSW09, Sli07], [Tal04, BGK12], and
[BLL09, GKK10].
1
1.1 Results
We provide the first clustering algorithms that leverage a preprocessing stage to obtain improved
query time. Specifically, we design algorithms that compute (1+ε) -- approximation for the p-median
and p-center problems; the precise time and space bounds are presented in Table 1. Observe that
the query time is near-linear in n and is (almost) independent of M , assuming the other parameters
(ε−1, p and ddim) are small. For sake of simplicity, we let our results depend on the aspect ratio
of M , denoted ∆ = ∆(M ). Such bounds can usually be refined, replacing e.g. log ∆ terms with
log n, by adapting our algorithms using known techniques and data structures, but it would clutter
the presentation of our main ideas. Interestingly, we use essentially the same data structure for all
problems solved. All our space bounds are expressed in terms of machine words, which as usual
can accommodate a pointer to a data point or a single distance value.
Problem
1-median
Theorem 4.1
p-median
Theorem 7.1
p-center
Theorem 6.1
Preprocessing time
Space
Query Time
2O(ddim)m log ∆ log log ∆
2O(ddim)m
O(n log n + 2O(ddim) log ∆ + ε−O(ddim))
2O(ddim)m log ∆ log log ∆
2O(ddim)m
2O(ddim)m log ∆ log log ∆
2O(ddim)m
O(n log n + ε−O(p·ddim)(p log n)O(p)
+ ε−O(ddim)(p log n)O(1) · log log log ∆)
O(n log n + p log log log ∆
+ pp+1 · ε−O(p·ddim))
Table 1: Our algorithms for (1 + ε) -- approximation of clustering problems.
We point two possible extensions of our results. First, one may ask about updates to M , i.e.,
inserting and deleting points. Our data structure is similar to previous work on NNS, and thus we
expect the methods known there (see e.g. [KL04b]) to apply also in our case, although we did not
check all the details. Second, we assume throughout that Q ⊂ M . One may remove this restriction,
possibly adapting the definition of ddim and ∆ to refer to M ∪ Q. Again, we have not checked
the details, but we expect this is possible by roughly applying the procedure of inserting Q to M
before executing the query Q, except that now we cannot use points of Q \ M as centers.
Our bounds for clustering with preprocessing are in a new model that was not studied before, and
thus cannot be compared directly with previous work. But of course, all of our results immediately
imply also algorithms for the respective offline problems, where the input includes both Q and
M . Since our preprocessing time is near linear in m, these are pretty efficient as well. Even
for the offline problems, our results are new, as we are not aware of previous work on clustering
(p-median and p-center) in metric spaces of bounded doubling dimension. Notice that a naive
algorithm, which exhaustively tries all possible sets of centers (with no preprocessing) finds an
optimal solution but takes runtime (cid:0)m
p(cid:1)np, which is significantly higher even for p = 1. Another
possible comparison is with the respective Euclidean problems; this is only for the sake of analogy
and is discussed in Section 1.3. We also point out that our 1-median algorithm is deterministic,
while previous algorithms achieving (1+ ε) -- approximation for 1-median, even in Euclidean metrics,
are randomized [Ind99, BHI02, KSS10].
2
1.2 Techniques
Our algorithms build on several techniques from prior work. The common algorithmic paradigm for
the NNS problem in metrics with low doubling dimension [KL04b, HM06, BKL06, CG06] (which in
our context is just the special case p = n = 1), is to look for an answer (center point) by restricting
attention to a sequence of search balls, whose radii are decreasing, usually by a constant factor.
When the ball's radius becomes small enough, the algorithms revert to exhaustively trying a small
set of candidates inside the search ball, with the property that at least one candidate in the set
must be a good enough approximation to an optimal answer. Such a set of candidates is sometimes
called a centroid set [HM04]. We follow this paradigm, but extend and modify it for our needs.
We further borrow a technique of constructing a coreset [BHI02], which essentially assigns points
in Q to a small set of "representatives" R, so that solving the clustering problem on the weighted
set R provides a good approximation for clustering Q. The weight of a representative r ∈ R is
simply the number of query points q ∈ Q assigned to it. In contrast to previous work on coresets
and on centroid sets, we have the leverage of preprocessing M , and our challenge is to quickly
construct such sets for Q during query time.
We also devise a new technique (new at least in the context of our clustering problems) of
"projecting" the data structure constructed for M (during preprocessing) onto the query set Q ⊂ M .
While a data structure for Q can be constructed from scratch in time 2O(ddim)n log n, the projection
can be constructed even faster, in time O(n log n). But even more importantly, the projected data
structure inherently provides hooks into the larger set M , and these hooks are crucial for our goal
of locating centers in M , which is (generally) a much richer point set than Q.
1.3 Related Work
Metrics with bounded doubling dimension are known to generalize Euclidean metrics of fixed-
dimension. Below we briefly mention known algorithms achieving (1 + ε) -- approximation for the
p-median and p-center problems in Euclidean spaces of fixed dimension D. These are only intended
to be a crude analogy to our results, possibly providing yet another perspective. Often, different
tradeoffs are possible between the number of centers p and the dimension D. We do not discuss
approximation algorithms for general metrics, as these do not achieve (1 + ε) -- approximation.
We start with the p-median problem. Arora, Raghavan and Rao [ARR98] were the first to
obtain (1 + ε) -- approximation, via a divide-and-conquer approach based on quadtrees and dynamic
programming. This approach was later improved by Kolliopoulos and Rao [KR07] and by Badoiu,
Har-Peled, and Indyk [BHI02]. Har-Peled and Mazumdar [HM04] added another technique of find-
ing coresets, and obtained running time O(n + pp+2ε−(2D+1)p logp+1 n logp 1
ε ). Kumar, Sabharwal,
and Sen [KSS10] showed a different approach, based on finding centroid sets, that runs in time
2(p/ε)O(1)
n. These approaches were later combined by Chen [Che06], who obtains improved runtime
when the dimension D is large.
For the p-center problem, Agarwal and Procopiuc [AP02] obtain (1 + ε) -- approximation in time
O(n log p) + (p/ε)O(Dp1−1/D ). Badoiu, Har-Peled, and Indyk [BHI02] show an algorithm that runs
in time pO(p/ε2)Dn.
3
1.4 Preliminaries
Let (M, d) be a finite metric space. The doubling dimension of M , denoted ddim = ddim(M ), is
the smallest k > 0 such that every ball (in M ) can be covered by 2k balls of half the radius. We
denote the diameter of the metric by diam(M ) := maxx,y∈M d(x, y), and its aspect ratio (or spread)
by ∆(M ) := max{d(x,y): x,y∈M }
min{d(x,y): x6=y∈M } .
Let r > 0. An r-net of a point set S ⊂ M is a subset N ⊆ S satisfying: (a) packing property:
for all x, y ∈ N we have d(x, y) ≥ r; and (b) covering property: for all x ∈ S we have d(x, N ) < r.
1 Such a net always exists, and can be constructed greedily by considering the points one by one
in an arbitrary order.
2 Our Data Structure
The Net Hierarchy. Our data structure is based on a lot of previous work on algorithms and
data structures for doubling metrics, in particular for nearest neighbor search [KL04b, HM06,
BKL06, CG06]. But despite the overall similarity, some technical details differ slightly from each of
those papers. Let itop := ⌈log2 diam(M )⌉, and assume for simplicity that the minimum interpoint
distance in M is minx6=y∈M d(x, y) = 1 (otherwise we need to introduce ibot as its logarithm).
Let Y0 = M , and for i = 1, . . . , itop let Yi be a 2i-net of Yi−1. Note that it is not necessarily a 2i-
net of M , but it does cover M indirectly via the nets at lower levels. We sometimes refer to Yi as the
level i net. By definition, Yi ⊆ Yi−1, so when we refer to y ∈ Yi we mean the copy of y which is in Yi.
These nets form a natural hierarchy, with Y0 being on the bottom, and a singleton Yitop = {ytop} at
the top of the hierarchy. This hierarchy may be represented by a directed acyclic graph GM , whose
vertex set is the union of all the nets Yi (so a point y ∈ M may have multiple copies in this graph),
and with an arc from every yi ∈ Yi to every yi−1 ∈ Yi−1 for which d(yi, yi−1) ≤ 2i. We prefer not
to maintain the graph GM explicitly; instead, our data structure has two main components, a tree
T and a collection of c-lists, which are defined below.
The Tree T . The hierarchy is represented by a tree that is defined as follows. First construct
GM as explained above. Next, every node in GM keeps only one of its incoming arcs that is chosen
arbitrarily except for giving higher priority to the arc coming from another copy of the same point
of M (if it exists). The surviving arcs define (when ignoring the edge orientations) a tree, denoted
T = TM , which is rooted at ytop. Because of the prioritization rule, whenever a point y ∈ Yi has
only one child in the tree T , this child must correspond to the same point y but in Yi−1. Thus,
every non-branching path in T consists of copies of the same point in M in consecutive nets. By
contracting each such path while recording the range of nets in which it participates, we can store
the tree T more compactly, using only O(m) space (recall the tree has m leaves). However, as
explained a bit later, we actually employ a more limited compaction, that results with a weaker
space bound 2O(ddim)m.
We supplement T with a data structure that supports constant-time lowest common ancestor
(LCA) queries using an additional 2O(ddim)m words [HT84] (see also [BF00] for a simplified ver-
sion). For the 1-center and p-center algorithms, we supplement T also with a data structure for
weighted level ancestor queries [FM96, KL07], which locate an ancestor of q ∈ M at level i (i.e.,
1Another common definition has a strict inequality in condition (a) rather than in (b). Our analysis can be
adapted to this definition by changing constants.
4
in Yi) in O(log log log ∆) time. The preprocessing for the weighted level ancestor queries requires
2O(ddim)m log log log ∆ time.
The c-Lists. For some constant c ≥ 1 that will be determined later, we maintain for every net
point y ∈ Yi a so-called c-list
Ly,i,c := {z ∈ Yi−1 : d(y, z) ≤ c · 2i}.
The c-lists allow us to traverse the ball of radius c2i in the next level of the hierarchy. If c = 1,
this list can be viewed as the set of arcs leaving y ∈ Yi in GM . When c ≥ 1, these lists can be used
(via straightforward filtering) to recover the arcs of GM . Since c is an absolute constant, the size
of each c-list is at most cO(ddim) ≤ 2O(ddim) (see e.g. [GKL03, KL04b]). We do not store the c-list
explicitly for every point in every net, as this might require too much space. We say that a c-list
of a point y ∈ Yi is trivial if it has size 1, in which case the only point in this list must be the copy
of y in Yi−1. We store only nontrivial c-lists, the number of which is at most 2O(ddim)m [KL04b,
Theorem 2.1]. It follows that the total space usage for the c-lists is 2O(ddim)m.
The nontrivial c-lists also limit the compaction of the tree T as follows. We compact T only
along paths whose nodes are both non-branching and have trivial c-lists. By the above bound on
the number of nontrivial c-lists, our limited compaction of T uses at most 2O(ddim)m space.
Preprocessing time. The preprocessing stage first employs the data structure of [KL04b] to
construct the c-lists in 2O(ddim)m log ∆ log log ∆ time, by simply inserting the data points one after
the other. We then scan this structure, from top to bottom, to introduce direct pointers as dictated
by the c-lists (i.e., from Ly,i,c to relevant Lz,i−1,c), and also construct the tree T in its compacted
version. This entire process takes 2O(ddim)m log ∆ log log ∆ time.
Projected Tree. A key tool in getting faster runtime is a projection of the tree T = TM onto a
subset of points Q ⊆ M . The idea is to consider the subtree of T induced by the leaves that are
points in Q. We will denote this projected tree T Q. Observe that this projected tree might be very
different from the tree TQ that would be constructed for Q independently of M ; in particular, the
latter cannot contain points from M \ Q. In the interest of runtime, we maintain the projected tree
somewhat implicitly; what the data structure stores explicitly is a compacted version, in which all
non-branching paths of T Q are contracted, and this clearly uses only O(n) space. Notice that such
a contracted path of T Q might contain nodes that are branching in T , which possibly correspond
to distinct data points in M . Although we have only the compacted version of T Q at hand, we
can implement a traversal down the un-compacted tree T Q, as described in Lemma 2.2.
To aid in the construction of the projected tree, we number the leaves of T in depth-first search
(DFS) order. In addition, for every node u ∈ T Q we denote by wt(u) the number of leaves in its
subtree. We can compute the weight of all the nodes in T Q in time O(n) by a simple scan.
Lemma 2.1. When a query Q is given, the compacted version of T Q can be computed in time
O(n log n).
Proof (Sketch). To create T Q, first sort Q according to the DFS numbering. Notice that the order
in which points from Q are encountered when performing a DFS on T is exactly the order in which
they would be encountered had we performed a DFS on T Q. Hence the sorted Q gives us this
5
order. We now use LCA queries to simulate the DFS on T Q, in order to construct T Q. This is
done as follows. Denote by Qi ⊆ Q the first i − 1 points in the ordered Q. We scan Q by the DFS
order, and when we reach the ith point, say qi ∈ Q, we assume we have already constructed T Qi−1
on the first i − 1 points of Q. We now wish to insert qi to this tree to obtain T Qi. To do this,
compute u = LCA(qi−1, qi). This node u, which has to be part of T Qi, is either on the path from
the leaf corresponding to qi to the root of T Qi−1, or is an ancestor of the root of T Qi−1. To locate
its exact position, we traverse T Qi−1 from the leaf corresponding to qi−1 upwards towards the root,
testing at each node v if this is the location into which u should be inserted. The testing at v is
performed via an LCA query between v and u. If the LCA query returns u, then the traversal
needs to continue. If not, then u is inserted as a child of v, either breaking an edge or inserting a
new leaf. The entire process simulates the DFS search on T Q and hence takes O(n) time.
The next lemma is used to traverse the un-compacted tree T Q while using the data structure
of its compacted version.
Lemma 2.2. Given the compacted T Q, a node v in the un-compacted version of T Q together with
its weight wt(v), and node w which is the closest descendant of v in the compacted T Q (and could
possibly be v itself ), it is possible to locate the children of v in the un-compacted T Q, together with
their weights, in time 2O(ddim).
Proof. Suppose first that v is a branching node in T Q. For each child u of v in the compacted
tree T Q, we find the respective child of v in the un-compacted T Q as follows: Run an LCA query
between u and every child of v in T . All of those queries will return v, except for one query that
will return the required child of v in the un-compacted tree (the one that is also an ancestor of u).
The total time for all such queries is 2O(ddim).
Suppose next that v is a non-branching node, and hence is not a part of the compacted version
of T Q. We find the child of v that is an ancestor of w in the un-compacted tree as follows: Perform
an LCA query between w and each of v's children in T . All of those queries will return v, except
for one query that will not return v, but rather the child of v that is also an ancestor of w, denoted
by w′. Notice that in this case, w is also the closest descendant of w′ in the compacted T Q, which
is needed to continue our traversal and proceed to w′.
Standard Operations on the Net Hierarchy. A basic operation in a net hierarchy is a recur-
sive scan, where given a point yi ∈ Yi, we scan its c-lists and apply the same procedure recursively
on these points. During this process, we discard any duplicates we find (e.g., if we reach the same
point in Yi−2 via different points in Yi−1).
Definition 2.1. Let y ∈ Yi. A point x ∈ M is called a c-list-descendant of y if it can be reached
from y using a recursive scan of the c-lists. We then also say that y is a c-list-ancestor of x.
Lemma 2.3. Let y ∈ Yi, and let x ∈ Yj be a c-list-descendant of y. Then d(x, y) ≤ c2i+1 − c2j+1.
Proof. The proof is by induction on i − j. The base case is trivial. For the inductive step, for every
y ∈ Yi, the distance between y and any of the points in its c-list is at most c2i. For every x ∈ Yj
which is a c-list-descendant of y, there exists a yi−1 ∈ Ly,i,c such that x is a c-list-descendant of
yi−1. Therefore, d(x, y) ≤ d(x, yi−1) + d(yi−1, y) ≤ c2i − c2j+1 + c2i = c2i+1 − c2j+1.
Notice that a point x can be a c-list-descendant of y even if in the tree T it is not a descendant
of y. However, ancestors and descendants in T also have bounds on the distance between them.
6
Lemma 2.4. Let y be an ancestor of x in T , such that y ∈ Yi and x ∈ Yj, where i > j. Then
d(x, y) ≤ 2i+1 − 2j+1 < 2i+1.
Proof. The distance between a parent from Yi and its child in T is at most 2i. Therefore, by
summation on the path from y to x in T , and the triangle inequality d(x, y) ≤ Pi
k=j+1 2k =
2i+1 − 2j+1.
The following lemma is crucial to searching the vicinity of a given point with some refinement
factor ε > 0, by executing a recursive scan with limited depth. This process will be used several
times in our various algorithms.
Lemma 2.5 (Descendents Search with Refinement ε). Let y ∈ Yi and x ∈ M be such that d(x, y) ≤
2i, and suppose c ≥ 3. Then for every refinement constant 0 < ε ≤ 1/2, a recursive scan of c-lists
that stops at level i − log(1/ε) will traverse a point x′ ∈ Yi−log(1/ε) for which d(x, x′) ≤ ε2i. In
addition, the number of points traversed in such a scan is at most ε−O(ddim).
Proof. Let xε be the ancestor of x in T who is in the Yi−log(1/ε) net, and so by Lemma 2.4 we
have that d(x, xε) ≤ 2i−log 1
ε + 1 ≤ j ≤ i and
every yj ∈ Yj such that d(x, yj) ≤ 2j+1, the recursive scan of c-lists from yj will reach xε. This
will suffice as for every x such that d(x, y) ≤ 2i, we also have d(x, y) ≤ 2i+1. For the base case,
j = i − log 1
ε + 1, and so d(yj, xε) ≤ d(yj, x) + d(x, xε) ≤ 2j+1 + 2j ≤ c2j, and so xε is in the c-list
for yj.
ε +1. We prove by induction that for every i − log 1
For the induction step, assume that the claim is correct for j − 1. Consider xj−1 ∈ Yj−1 which is
the ancestor of xε in T , and therefore is also an ancestor of x. Then by Lemma 2.4, d(x, xj−1) ≤ 2j
and by the induction hypothesis, a recursive scan on the c-lists starting from xj−1 will reach xε.
Then d(xj−1, yj) ≤ d(xj−1, x) + d(x, yj) ≤ 2j + 2j+1 ≤ c2j, and so a recursive scan on the c-lists
starting from yj must go through xj−1 and eventually reach xε.
The number of points traversed can be bounded as follows. Each point x not in Yi−log(1/ε) that
is encountered needs to scan its c-list which is of size 2O(ddim). So at k levels beneath y we scan
at most 2O(ddim ·k) points. The last level scanned is when k = log 1
ε , so using a geometric series we
obtain that the total number of points scanned is 2O(ddim · log(1/ε)) = ε−O(ddim).
3 A Simple Algorithm for 1-median
In this section we provide a simple iterative algorithm for 1-median, for the purpose of explaining
the basic approach used by our main result for 1-median in Section 4. This basic approach is similar,
but not identical, to the known algorithms for NNS [KL04b, HM06, BKL06, CG06]. We remark
that there is a well-known randomized algorithm that achieves an (expected) 2 -- approximation for
1-median by picking a random point from Q to be the center. Below, we present a deterministic
6 -- approximation algorithm, which has the advantage that it is then easily refined to achieve (1+ε) --
approximation. Unlike that randomized algorithm, ours can probably be adapted to the case where
Q need not be a subset of M (or alternatively, when the center must come from M \ Q).
Theorem 3.1. There is an algorithm that preprocesses a finite metric M in 2O(ddim)m log ∆(M ) log log ∆(M )
time using 2O(ddim)m space, so that subsequent 1-median queries on a set Q ⊆ M , can be answered
within (1 + ε) -- approximation (for any desired 0 < ε ≤ 1
2 ) in time n(2O(ddim) log ∆ + ε−O(ddim)).
7
The preprocessing algorithm simply builds the net hierarchy for the metric M (see Section 2).
The query algorithm is described in Figure 1. For convenience, we use the shorthand med(y, Q)
for med({y}, Q). By convention, for all i < 0 we define Yi := M (similarly to Y0), and note that
the corresponding c-lists can be computed on the fly by a direct filtering of the respective c-list at
level 0.
1. let y ← ytop
2. for each i from itop down to −3
let y ← argminz∈Ly,i,7 med(Q, z)
3.
if med(Q, y) > 3n · 2i−1 then return y.
4.
else y ← y.
5.
6. return y.
Figure 1: Simple algorithm for 1-median query on a set Q ⊂ M
Correctness Analysis. Assume for now that the algorithm returns from line 4. (We discuss
later the more special case where the algorithm reaches line 6.) For the following, let i∗ be the final
value of i (i.e., at line 4), and let y and y refer to their values in the algorithm at the end of the
execution. It can be verified that the condition in line 4 must fail at least once, by considering ytop
as a potential center y, and bounding the distance between every point in Q ⊂ M to ytop using
Lemma 2.4. Therefore,
Pq∈Q d(q, y) > 3n · 2i∗−1,
and Pq∈Q d(q, y) ≤ 3n · 2i∗
.
(1)
Let a ∈ M be an optimal solution to the 1-median problem on Q. Let ai∗−1 ∈ Yi∗−1 be an ancestor
of a in T . Then d(ai∗−1, a) ≤ 2i∗
Lemma 3.2. d(ai∗−1, y) ≤ 7 · 2i∗
, and thus ai∗−1 ∈ Ly,i∗,7.
by Lemma 2.4.
Proof. Using the triangle inequality, the optimality of a ∈ M , and then (1),
n · d(ai∗−1, y) ≤ X
d(ai∗−1, a) + 2 X
[d(ai∗−1, a) + d(a, q) + d(q, y)] ≤ X
q∈Q
q∈Q
q∈Q
d(q, y) ≤ 7 · n2i∗
.
Lemma 3.3. Pq∈Q d(q, a) > n · 2i∗−1.
Proof. By Lemma 3.2, when the algorithm computes y in the final iteration, one of the options it
considers is ai∗−1, and so
d(q, y) ≤ X
[d(q, a) + d(a, ai∗−1)] ≤ X
d(q, ai∗−1) ≤ X
d(q, a) + n · 2i∗
.
q∈Q
q∈Q
q∈Q
X
q∈Q
Combining this with (1) and rearranging, the lemma follows.
Thus, if we returned from line 4, then using (1), the approximation factor achieved is
Pq∈Q d(q,y)
Pq∈Q d(q,a) <
3·n·2i∗
n·2i∗−1 = 6. If we returned from line 6, Lemma 3.2 holds also for i = −3, and thus at the last
execution of line 3, we have d(a, y) ≤ 7
23 . But since there cannot be two points with distance less
than 7
23 < 1, we see that y = a, and the returned point is an optimal solution a. We remark that
a similar effect can be achieved by stopping at i = 0, possibly increasing the value of c.
8
3.1 Refinement to (1 + ε) -- approximation
We now improve the approximation factor to 1 + ε for an arbitrary ε ∈ (0, 1
2 ]. We can utilize the
fact that a is a descendant of ai∗−1 in T , so d(a, ai∗−1) ≤ 2i∗
, and that ai∗−1 ∈ Ly,i∗,c. As such, we
perform a descendant search with refinement constant ε/2, starting from each member of Ly,i∗,c.
. For each point
By Lemma 2.5, we are guaranteed to traverse a point a ε
2
x traversed in this process, we compute med(Q, x), and eventually report a center candidate x with
minimal objective value med(Q, x). Using (1) again, this objective value is
such that d(a, a ε
2
) ≤ ε
2 2i∗
med(Q, x) ≤ med(Q, a ε
2
) ≤ X
q∈Q
hd(q, a) + d(q, a ε
2
)i ≤ med(Q, a) + ε
2 n2i∗
≤ (1 + ε) med(Q, a).
3.2 Runtime Analysis
The running time of the first part of the algorithm is 2O(ddim)n log ∆, as there are at most O(log ∆)
levels, and at each level we compute the distance from every point in Q to every point z ∈ Ly,i,7.
In the second part of the algorithm (the descendants search) we compute the cost of each of the
ε−O(ddim) center candidates in O(n) time. The total runtime is n(2O(ddim) log ∆ + ε−O(ddim)), and
the space usage is just that of the hierarchy, which is 2O(ddim)m.
4 An Efficient Algorithm for 1-median
Theorem 4.1. There is an algorithm that preprocesses a finite metric M of size m in time
2O(ddim)m log ∆(M ) log log ∆(M ) using 2O(ddim)m memory words, so that subsequent 1-median
queries on a set Q ⊆ M of size n can be answered within approximation factor 1 + ε (for any
desired 0 < ε ≤ 1/2) in time O(n log n) + 2O(ddim) log ∆(M ) + ε−O(ddim).
This theorem builds on the simple algorithm from Section 3, refining the approach therein using
two main ideas. First, as we iterate down the levels i, some query points q ∈ Q might get further
away from the current center yi ∈ Yi. But then, picking any c-list-descendant of yi as the final
center will give approximately the same contribution from those far query points. This speeds up
the traversal down the hierarchy as query points need not be considered once they get far enough
from yi. The second idea is to cluster query points that are close to each other, relative to the
current level i, into one (weighted) representative point. This (crude) clustering must be computed
quickly, and indeed it is achieved using the projection tree T Q. Once we bound the number of
weighted representatives under consideration in each iteration, we obtain a significant speedup.
Algorithm Description. We first describe a constant factor approximation algorithm, which is
detailed in Figure 2 using α, c′ > 0 to denote sufficiently large constants. Similarly to the simple
algorithm in Section 3, the algorithm iterates (in lines 3 -- 12) down the levels i, while maintaining a
candidate center yi ∈ Yi. However, the iterations here start at the root of T Q (instead of at ytop).
Observe that the next candidate yi−1 is always chosen from the c-list of yi (lines 9,12).
During the iterations, the algorithm maintains also a set Ri of representatives to some points of
Q, those points that are not too far, as explained next. The level i representative of a point q ∈ Q,
denoted ri(q), is the (unique) ancestor r ∈ Yi of q in T Q. Notice that this is the same ancestor as in
the tree T . The algorithm also uses, for each representative r ∈ Ri, a weight denoted wt(r), which
9
1 . compute T Q
2 . let iroot ← level(root(T Q)); yiroot−1 ← root(T Q); Riroot ← {root(T Q)}; sum ← 0
3 . foreach i from iroot − 1 down to 0
4 .
5 .
6 .
7 .
8.
9.
10.
11.
12.
13. return y−1
if d(r, yi) > c′ · 2i
then let sum ← sum + wt(r) · d(r, yi)
else let Ri ← Ri ∪ {children of r in non-compacted T Q}.
let y ← argminx∈Lyi,i,c{Pr∈Ri
if sum + Pr∈Ri
then return yi
else yi−1 ← y
let Ri ← ∅
foreach r ∈ Ri+1
d(r, x) · wt(r)}
d(r, y) · wt(r) > α · n · 2i−1
Figure 2: Efficient Algorithm for 1-median query on a set Q ⊂ M
is the number of points in Q that have r as an ancestor in T Q. This weight is calculated for each
node in T Q during the tree's construction in line 1. The set of representatives Ri is constructed
(in lines 4,8) from children of Ri+1 in T Q, which clearly maintains the invariant Ri ⊂ Yi. In this
process, we skip (via the condition in line 6) representatives r ∈ Ri+1 that are far enough from yi,
in which case we add their weighted distance wt(r) · d(r, yi) to a variable called sum. The purpose
of this variable is to accumulate all those weighted distances, but note that each weighted distance
is taken relative to yi at the iteration in which the representative r fails the condition in line 6.
Denote by sumi the value of variable sum at the end of iteration i. For representatives r ∈ Ri+1
that are close enough to yi, we need to compute their children in the un-compacted T Q (in line
8). For simplicity sake, the algorithm's description assumes that the tree T Q is available in its
un-compacted version. The necessary operations can be implemented using the data structure for
the compacted version by Lemma 2.2.
4.1 Correctness Analysis
We say a point q ∈ Q is far at level i if it has no representative in Ri, which means that during
some iteration i′ > i its representative was skipped. A point q ∈ Q is near if it is not far. Let Fi
denote the points of Q that are far at level i, and similarly Ni = Q \ Fi for the points that are near.
Notice that Fi ⊇ Fi+1 and Ni ⊆ Ni+1.
Let i∗ be the value of i at the end of the execution. This is the "last" level (time-wise) considered
by the algorithm, and the analysis shall rely on the corresponding partition Q = Ni∗ ∪ Fi∗ . For
q ∈ Q, we denote its representative in Ri by ri(q). We let rq be the "last" representative of q,
formally defined as follows. If q ∈ Fi∗, define iq as the smallest i such that q ∈ Ni. Intuitively,
this is the "last level" in which q has a representative, and also the (unique) value of i such that
q ∈ Ni \ Ni−1 = Fi−1 \ Fi (assuming by convention Ni∗−1 = ∅ and Fi∗−1 = Q). Otherwise (i.e.,
q ∈ Ni∗), define iq := i∗. In both cases, let rq := riq (q). Notice that rq ∈ Yiq .
At iteration i, the variable called sum receives (in line 7) a contribution for every point q ∈
Fi \Fi+1. Observe that this contribution is proportional to d(ri+1(q), yi), and the last representative
10
of q is at level iq = i + 1. Hence, rq = riq (q) = ri+1(q), and by the condition in line 6,
d(rq, yiq−1) = d(ri+1(q), yi) > c′2i = c′2iq−1.
Summing the aforementioned contributions over all iterations up to i, we see that
sumi = X
i′≥i
X
q∈Fi′ \Fi′+1
d(ri′+1(q), yi′) = X
q∈Fi
d(rq, yiq−1).
In addition, rq ∈ Yiq and is an ancestor of q in T .Thus, by Lemma 2.4,
∀q ∈ Q,
∀q ∈ Ni∗,
d(q, rq) ≤ 2iq+1
d(q, rq) ≤ 2i∗+1.
(2)
(3)
(4)
(5)
Below, y refers to its value at the end of the execution.
We assume from now on that the algorithm halts during some iteration and returns the value
from line 11. Similarly to Section 3, the special case where the algorithm returns from line 13 is
proved by replacing Eqn. (6) and its consequences with the fact that we reached iteration i = 0.
Thus, at the last iteration, i = i∗, the algorithm halts, and
sumi∗ + X
r∈Ri∗
d(r, y) · wt(r) = sumi∗ + X
q∈Ni∗
d(rq, y) > α · n · 2i∗−1.
(6)
Similarly, at the previous to last iteration i = i∗ + 1 and y is assigned yi∗, hence
sumi∗+1 + X
r∈Ri∗+1
d(r, yi∗ ) · wt(t) = sumi∗+1 + X
q∈Ni∗+1
d(ri∗+1(q), yi∗ ) ≤ α · n · 2i∗
.
(7)
This inequality holds even in the special case where i∗ = iroot − 1 and there was no previous to last
iteration. Indeed, we have that Fi∗+1 = ∅, sumi∗+1 = 0, Ri∗+1 = {root(T Q)} and yi∗ = root(T Q),
and therefore, sumi∗+1 + Pr∈Ri∗+1
Lemma 4.2. sumi∗ + Pq∈Ni∗ d(rq, yi∗ ) ≤ (α + 2) · n · 2i∗
Proof. We write the lefthand-side as
d(r, yi∗ ) = 0 ≤ α · n · 2i∗
.
.
sumi∗ + X
q∈Ni∗
d(rq, yi∗) =
= sumi∗+1 + X
q∈Ni∗ +1−Ni∗
≤ sumi∗+1 + X
q∈Ni∗ +1−Ni∗
d(ri∗+1(q), yi∗ ) + X
q∈Ni∗
d(ri∗+1(q), yi∗ ) + X
q∈Ni∗
≤ sumi∗+1 + X
q∈Ni∗ +1
d(ri∗+1(q), yi∗ ) + n2i∗+1,
d(ri∗ (q), yi∗)
[d(ri∗ (q), ri∗+1(q)) + d(ri∗+1(q), yi∗ )]
where the last inequality follows from ri∗(q) being a child of ri∗+1(q) in T Q. The lemma then
follows by plugging in Eqn. (7).
11
For every q ∈ Fi∗ , we have by Eqn. (2) that d(q, rq) ≤ 2iq+1 ≤ 4
In addition,
d(yiq−1, yi∗) ≤ c2iq ≤ 2c
c′ d(rq, yiq−1), because yi∗ is a c-list-descendant of yiq−1 and thus Lemma 2.3
applies. To simplify notation, define β := 4
c′ and notice it can be made an arbitrarily small
positive constant by controlling c′. For example, it is always possible to make β = 1
2 . We can now
show that with respect to the query points Fi∗ , our estimate sumi∗ is a good approximation for the
cost of picking yi∗ as the center.
c′ d(rq, yiq−1).
c′ + 2c
X
q∈Fi∗
d(q, yi∗ ) ≤ X
q∈Fi∗
(cid:2)d(q, rq) + d(rq, yiq−1) + d(yiq−1, yi∗)(cid:3)
≤ (1 + 4
c′ + 2c
c′ ) X
q∈Fi∗
d(rq, yiq−1) = (1 + β)sumi∗.
(8)
In addition, we show that with respect to the query points Ni∗, the representatives give a good
approximation as well.
X
q∈Ni∗
d(q, yi∗ ) ≤ X
q∈Ni∗
[d(q, rq) + d(rq, yi∗)] ≤ n2i∗+1 + X
q∈Ni∗
d(rq, yi∗).
(9)
Let a ∈ M be an optimal solution to the 1-median problem Q, and let ai∗−1 ∈ Yi∗−1 be an
. We next prove that yi∗ is near ai∗−1, and thus also
ancestor of a in T . Thus, d(ai∗−1, a) ≤ 2i∗
near a itself.
Lemma 4.3. d(ai∗−1, yi∗) ≤ c2i∗
and therefore ai∗−1 ∈ Lyi∗ ,i∗,c.
Proof. We start with the lefthand-side multiplied by n
n · d(ai∗−1, yi∗ ) = X
q∈Q
d(ai∗−1, yi∗)
[d(ai∗−1, a) + d(a, q) + d(q, yi∗ )]
by triangle inequality
≤ X
q∈Q
≤ n2i∗
= n2i∗
≤ n2i∗
≤ 5n2i∗
≤ cn2i∗
.
+ 2 X
q∈Q
d(q, yi∗)
+ 2h X
q∈Fi∗
d(q, yi∗ ) + X
q∈Ni∗
d(q, yi∗)i
by optimality of a
+ 2(cid:2)(1 + β)sumi∗ + n2i∗+1 + X
+ 2(1 + β)(α + 2) · n · 2i∗
q∈Ni∗
d(rq, yi∗)(cid:3)
by Eqns. (8),(9)
by Lemma 4.2
For the last inequality we need to pick a large enough constant c > 0. (Recall that β can be made
to be 1
2 increasing c′ as needed.) Dividing all by n completes the proof.
We now want to prove the guarantee of our approximation. To this end we need an upper bound
on the cost of the algorithm's solution, which we establish by analyzing the stopping condition
iteration.
Lemma 4.4. Pq∈Q d(q, yi∗) ≤ [2 + (1 + β)(α + 2)]n2i∗
.
12
Proof. First, each q is close to rq and thus
X
q∈Ni∗
d(q, yi∗ ) ≤ X
q∈Ni∗
[d(rq, q) + d(rq, yi∗)] ≤ n2i∗+1 + X
q∈Ni∗
d(rq, yi∗).
Thus,
X
q∈Q
d(q, yi∗ ) = X
q∈Ni∗
d(q, yi∗) + X
q∈Fi∗
d(q, yi∗)
≤ n2i∗+1 + X
q∈Ni∗
d(rq, yi∗) + (1 + β)sumi∗
≤ n2i∗+1 + (1 + β)(sumi∗ + X
q∈Ni∗
≤ [2 + (1 + β)(α + 2)] n2i∗
.
d(rq, yi∗))
Lemma 4.5. sumi∗ ≤ 1
1−β Pq∈Fi∗ d(q, a).
Proof. First notice that
d(a, yiq −1) ≤ d(a, ai∗−1) + d(ai∗−1, yiq−1)
+ c2iq − c2i∗
≤ 2i∗
≤ c2iq
2c
c′ d(rq, yiq−1),
≤
where the bound on d(ai∗−1, yiq−1) follows from Lemma 2.3. Therefore,
X
q∈Fi∗
d(q, a) ≥ X
q∈Fi∗
[d(rq, yiq−1) − d(q, rq) − d(a, yiq −1)] ≥ (1 − β)sumi∗.
We are now ready to provide a lower bound on the optimal solution. Recall that y refers to its
value at the end of the algorithm.
Lemma 4.6. Pq∈Q d(q, a) > (α/2 − 3)(1 − β)n2i∗
(assuming the algorithm returns from line 11).
13
Proof.
α · n · 2i∗−1 < sumi∗ + X
q∈Ni∗
d(rq, y)
(6)
≤ sumi∗ + X
q∈Ni∗
≤ sumi∗ + X
q∈Ni∗
d(rq, ai∗−1)
by Lemma 4.3 and choice of y
[d(rq, q) + d(q, a) + d(a, ai∗−1)]
≤ sumi∗ + n2i∗+1 + X
q∈Ni∗
d(q, a) + n2i∗
≤ 3n2i∗
≤ 3n2i∗
+ 1
1−β X
q∈Fi∗
d(q, a) + X
q∈Ni∗
d(q, a)
by Lemma 4.5
+ 1
1−β X
q∈Q
d(q, a).
We conclude that the algorithm achieves approximation factor
Pq∈Q d(q, yi∗ )
Pq∈Q d(q, a)
≤
[2 + (1 + β)(α + 2)]n2i∗
(1 − β)(α/2 − 3)n2i∗ =
[2 + (1 + β)(α + 2)]
(1 − β)(α/2 − 3)
.
4.2 Refinement to (1 + ε) -- approximation
Our goal now is to improve the approximation factor to 1 + ε for arbitrary ε > 0. We can utilize the
fact that a is a descendant of ai∗−1 in T , so d(a, ai∗−1) ≤ 2i∗
, and that ai∗−1 ∈ Lyi∗ ,i∗,c. As such,
we can perform a descendant search, as in Lemma 2.5, starting from each member of Lyi∗ ,i∗,c, with
refinement constant ε′ = Θ(ε). By Lemma 2.5 we are guaranteed to traverse a point aε′ such that
d(a, aε′) ≤ ε′2i∗
. However, we wish to avoid the high runtime of computing the cost of each center
candidate by summing the distances from all of Q to that point. Instead, we once again speed up
the process by removing far points, and using weighted representatives for the rest.
Speeding up the descendants search. Define the set of far points F = {q ∈ Q : d(q, yi∗ ) >
2ε′ 2i∗} for some ε′ > 0 to be determined later. The set of near points is N := Q \ F . The
3c
points in F are ignored in this phase of the algorithm. For the points in N we wish to find good
representatives so that the number of representatives is few, and the additive distortion caused
by replacing the query points in N with their representative is very small. To this end, consider
the set of representatives obtained as follows. Each point q ∈ N is mapped to its ancestor in the
compacted T Q which is in Yk for the largest k ≤ i∗ − log(1/ε′′) for ε′′ > 0 to be determined later.
Call this set of representatives Rε′′, and give each r ∈ Rε′′ a weight wt(r) which is the number of
points in N that were mapped to r. Notice that the process of this mapping and weighting can
be done efficiently by scanning the compacted T Q in O(n) time. Now, for each center candidate
x obtained by a descendants search from each of the points in Lyi∗ ,i∗,c by using Lemma 2.5 with
refinement constant ε′′, we compute Pr∈Rε′′ d(r, x) · wt(r), and take the candidate which minimizes
this cost.
14
We want to argue that the candidate returned is a 1 + ε approximation from the optimum.
Denote this candidate by x. Notice that one of the candidates must be a point aε′′ which is an
ancestor of a in T and is in Yk for some k ≤ i∗ − log(1/ε′′). Therefore, Pr∈Rε′′ d(r, x) · wt(r) ≤
Pr∈Rε′′ d(r, aε′′ ) · wt(r).
Lemma 4.7. Pq∈Q d(q, x) ≤ (1 + ε)Pq∈Q d(q, a).
Proof. Denote by xi∗−1 the c-list-ancestor of x in Lyi∗ ,i∗,c ⊆ Yi∗−1. First, for every q ∈ F ,
d(q, x) ≤ d(q, yi∗ ) + d(yi∗, xi∗−1) + d(xi∗−1, x)
≤ d(q, yi∗ ) + c2i∗−1 + c2i∗
≤ d(q, yi∗ ) + ε′d(q, yi∗ )
= (1 + ε′)d(q, yi∗ ),
and similarly,
by xi∗−1 ∈ Lyi∗ ,i∗,c and Lemma 2.3
since q ∈ F
d(q, yi∗) ≤ d(q, a) + d(a, ai∗−1) + d(ai∗−1, yi∗)
≤ d(q, a) + 2i∗
+ c2i∗
≤ d(q, a) + ε′d(q, yi∗ ).
by Lemma 4.3
since q ∈ F
Therefore, d(q, yi∗ ) ≤ d(q,a)
1−ε′ . Combining this with our earlier inequality, we get
d(q, x) ≤
X
q∈F
1 + ε′
1 − ε′ X
q∈F
d(q, a).
For the near points, we have
X
q∈N
d(q, x) ≤ X
q∈N
d(q, aε′′) ≤ X
q∈N
(cid:2)d(q, a) + d(a, aε′′)(cid:3) ≤ X
q∈N
(cid:2)d(q, a) + ε′′2i∗+1(cid:3).
Altogether, the cost of the reported center candidate x is
(10)
(11)
d(q, x) ≤
X
q∈Q
≤
≤
q∈F
1 + ε′
1 − ε′ X
1 + ε′
1 − ε′ X
1 + ε′
1 − ε′ X
q∈Q
q∈Q
d(q, a) + X
q∈N
[d(q, a) + ε′′2i∗+1]
by Eqns. (10),(11)
d(q, a) + 2ε′′n2i∗
d(q, a) +
2ε′′
(α/2 − 3)(1 − β) X
q∈Q
d(q, a)
by Lemma 4.6
= (cid:16) 1 + ε′
1 − ε′ +
2ε′′
(α/2 − 3)(1 − β)(cid:17) X
q∈Q
d(q, a).
Setting ε′′ := (α/2−3)(1−β)ε
2
and ε′ := ε
ε+2 , we get that Pq∈Q d(q, x) ≤ (1 + ε)Pq∈Q d(q, a).
15
4.3 Runtime Analysis
In the first part, the compacted version of T Q is constructed (in line 1) in time O(n log n) using
Lemma 2.1. At each iteration i we locate y (in line 9), which becomes yi∗. The runtime of this step
is proportional to the number of candidates in the c-list multiplied by the size Ri. The number of
candidates is Ly,i,c ≤ 2O(ddim). The size of Ri is at most the number of points in Yi which are at
most c′2i away from yi. We conclude that Ri ≤ 2ddim log(c′2i+1/2i) ≤ 2O(ddim).
Computing Ri from Ri+1 takes 2O(ddim) time per member of Ri+1, for a total of 2O(ddim)
per iteration i. Thus, the total time spent on finding yi∗ is O(n log n + 2O(ddim) log ∆). For the
descendants search used in the refinement to (1 + ε) -- approximation, we can bound the number of
representatives as follows.
Lemma 4.8. Rε′′ ≤ ε−O(ddim).
Proof. If all of the representatives are in Yi∗−log(1/ε′′) then the size of Rε′′ is at most the number
of points in Yi∗−log(1/ε′′) which are at most 3c
away from yi∗. The number of such points is
bounded above by
2ε′ 2i∗
2O(ddim log( 3c
2ε′ 2i∗
/2i∗
−log(1/ε′′
)) = 2O(ddim log 3c
2ε′ε′′ ) ≤ ε−O(ddim).
However, the representatives do not all have to be in Yi∗−log(1/ε′′). To overcome this, we charge
each representative in Rε′′ to a different point in Yi∗−log(1/ε′′). This mapping is done by assigning
to each point in Rε′′ its ancestor in the un-compacted T Q which is in Yi∗−log(1/ε′′). Notice that
no two points in Rε′′ can be assigned to the same point in Yi∗−log(1/ε′′), as otherwise there would
be another node in the compacted T Q which is an ancestor of those two points and in Yk for
k ≤ i∗ − log(1/ε′′), which contradicts the method in which the representatives were picked.
The total cost of the descendants search is the number of representatives multiplied by the
number of center candidates. From Lemma 2.5 we know that the number of candidates is at most
ε′′−O(ddim) ≤ ε−O(ddim), and therefore, the runtime of this refinement stage is bounded by ε−O(ddim).
Overall, the runtime of computing a (1 + ε) -- approximation for the 1-median is O(n log n) +
2O(ddim) log ∆(M ) + ε−O(ddim), and this completes the proof of Theorem 4.1.
5 Algorithm for 1-Center
It is helpful to see the solution for the 1-center problem prior to seeing the solution for the p-center
problem, as many of the ideas are similar, and the exposition with only one center is simpler.
Therefore, we first prove the following.
Theorem 5.1. There is an algorithm that preprocesses a finite metric M in time 2O(ddim)m log ∆ log log ∆(M )
using 2O(ddim)m memory words, where m = M , ddim = ddim(M ) and ∆ = ∆(M ), so that subse-
quent 1-center queries on a set Q ⊆ M , can be answered with approximation factor 1 + ε, for any
desired 0 < ε ≤ 1/2, in time O(n log n + log log log ∆ + ε−O(ddim)) , where n = Q.
The preprocessing algorithm simply builds the net hierarchy for the metric M , and prepares it
for weighted level ancestor queries (see Section 2). For the query, we first recall a trivial algorithm
that provides a 2 -- approximation for the 1-center problem on a query set Q ⊂ M , and then refine
it to provide a (1 + ε) -- approximation.
16
Let a ∈ M be an optimal center, and denote its value by OPT := maxq∈Q d(a, q). Notice that
every point y ∈ Q gives a 2 -- approximation, because its objective value is
ALG0 := max
q∈Q
d(y, q) ≤ max
q∈Q
{d(y, a) + d(a, y)} ≤ 2 · OPT.
We thus pick any point y ∈ Q as our first approximation, and proceed to the refinement stage.
Refinement to (1 + ε) -- approximation. Let i be an integer such that 2i−1 < ALG0 ≤ 2i, and
notice that OP T ≤ ALG0 ≤ 2i. We begin by locating the ancestor yi ∈ Yi of y in T . This can be
done using a weighted level ancestor query [FM96, KL07]. We next show that a is fairly close to
yi.
Lemma 5.2. Let ai−1 ∈ Yi−1 be an ancestor of a in T . Then ai−1 ∈ Lyi,i,6.
Proof. For every point q ∈ Q,
d(ai−1, yi) ≤ d(ai−1, a) + d(a, q) + d(q, y) + d(y, yi) ≤ 2i + OPT + 2 · OPT + 2i+1 ≤ 6 · 2i.
This lemma implies that the optimal center a is a descendant in T of some point in Lyi,i,6.
Executing a descendants search from all the points in Lyi,i,6 by using Lemma 2.5 with refinement
constant ε will guarantee that we traverse a point aε such that d(a, aε) ≤ ε2i. Denote the set of the
points seen in such a descendants search by D. Unfortunately, this process computes (separately)
the cost of each candidate traversed by taking the maximum distances from all of Q to that
candidate, which would take time ε−O(ddim)n. We can speed up this process by using (a few)
representatives of Q, as is explained next.
Speeding up the descendants search. We wish to find a bounded-size set of representatives
for the points in Q, such that the distortion caused by considering them (instead of Q) is small. To
this end, consider the set of representatives obtained as follows. Each point q ∈ Q is mapped to its
ancestor in the compacted T Q which is in Yk for the largest k ≤ i − log(1/ε′), for some refinement
constant ε′ = Θ(ε) to be determined later. Call this set of representatives Rε′. Notice that Rε′
is a subset of the compacted T Q and thus the process of this mapping can be done efficiently by
scanning the compacted T Q in linear time. Now, for each center candidate x ∈ D we compute
maxr∈Rε′ d(r, x), and return the candidate x that minimizes this cost.
The next lemma shows that this algorithm achieves (1 + ε) -- approximation.
Lemma 5.3. cntr(Q, {x}) = maxq∈Q d(x, q) ≤ (1 + ε)OPT.
Proof. Every q ∈ Q has a representative in Rε′, for which we can apply Lemma 2.4 and the triangle
inequality, and thus
max
q∈Q
d(x, q) < max
r∈Rε′
d(x, r) + ε′2i+1.
Recall that one of the center candidates is some aε′ ∈ Yi−log(1/ε′) that is an ancestor of a in T .
Therefore, the returned center x satisfies
max
r∈Rε′
d(x, r) ≤ max
r∈Rε′
d(aε′, r).
17
Let r∗ ∈ Rε′ be a maximizer for the righthand side, and let q∗ ∈ Q be such that r∗ is a representative
of q∗. Using the triangle inequality and Lemma 2.4 again,
d(aε′, r∗) ≤ d(a, aε′) + d(a, q∗) + d(q∗, r∗) ≤ OPT + 2 · ε′2i+1.
Recalling from earlier that 2i < 2ALG0 ≤ 4OPT, we finally combine the inequalities above and
conclude that maxq∈Q d(x, q) ≤ OPT + 3 · ε2i+1 ≤ (1 + 24ε)OPT. To complete the proof, set ε′ to
be a power of 2 in the range [ ε
48 , ε
24 ].
Runtime. The running time of the above query process is as follows. Locating yi using a weighted
level ancestor query takes O(log log log ∆) time as there are only log ∆ possible nets. After con-
structing T Q in O(n log n) time, the mapping of every q ∈ Q to its representative takes, alto-
gether, O(n) time. The descendants search for each of the O(2ddim) points in Lyi,i,6 takes time
ε′−O(ddim) ≤ ε−O(ddim) time, which also bounds the number of candidates. The number of repre-
sentatives for Q is bounded by the following lemma.
Lemma 5.4. Rε′ ≤ ε−O(ddim).
Proof. If all of the representatives are in Yi−log 1
points in Yi−log 1
by
then the size of Rε′ is at most the number of
which are at most 2i away from yi. The number of such points is bounded above
ε′
ε′
2ddim · log(2i/2i−log(1/ε′)) = (1/ε′)O(ddim) = ε−O(ddim).
However, the representatives do not all have to be in Yi∗−log(1/ε′). To overcome this, we charge each
representative in Rε′ to a different point in Yi∗−log(1/ε′). This mapping is done by assigning to each
point in Rε′ its ancestor in the un-compacted T Q which is in Yi∗−log(1/ε′). Notice that no two points
in Rε′ can be assigned to the same point in Yi∗−log(1/ε′), as otherwise there would be another node
in the compacted T Q which is an ancestor of those two points and in Yk for k ≤ i∗ − log(1/ε′),
which contradicts the method in which the representatives were picked.
It follows that the time it takes to evaluate the cost of all center candidates is (altogether)
ε−O(ddim), and thus the algorithm's total runtime of is O(n log n + log log log ∆ + ε−O(ddim)).
6 Algorithm for p-center
Theorem 6.1. There is an algorithm that preprocesses a finite metric M in time 2O(ddim)m log ∆ log log ∆
using 2O(ddim)m memory words, where m = M , ddim = ddim(M ) and ∆ = ∆(M ), so that subse-
quent p-center queries on a set Q ⊆ M , can be answered with approximation factor 1 + ε, for any
desired 0 < ε ≤ 1/2, in time O(n log n + p log log log ∆ + pp+1ε−O(p·ddim)) , where n = Q.
The preprocessing algorithm simply builds the net hierarchy for the metric M , and prepares
it for weighted level ancestor queries (see Section 2). For the query, we first use the algorithm
of Gonzalez from [Gon85] on Q, which obtains a 2-approximation for the p-center in O(p · n)
In other words, the algorithm locates a set B ⊂ Q of size p such that if we denote its
time.
objective value as ALG0 := maxq∈Q d(q, B), and if A ⊂ M is an optimal p-center set with value
OPT := maxq∈Q d(q, A), then ALG0 ≤ 2 · OPT.
18
6.1 Refinement to (1 + ε) -- approximation.
Let i be an integer such that 2i−1 < ALG0 ≤ 2i. For each b ∈ B locate the ancestor bi ∈ Yi of b in
T . This can be done using a weighted level ancestor query [FM96, KL07]. For a center a ∈ OPT,
let ai−1 ∈ Yi−1 be an ancestor of a in T .
Lemma 6.2. For every a ∈ OPT there exists a point b ∈ B such that ai−1 ∈ Lbi,i,6.
For every point q ∈ Q which is assigned to a in OPT, let b be the center of the cluster of q in
B. Then d(ai−1, bi) ≤ d(ai, a) + d(a, q) + d(q, b) + d(b, bi) ≤ 2i + OPT + 2 · OPT + 2i+1 ≤ 6 · 2i.
This implies that a center a ∈ A is a descendant in T of some point in Sb∈B Lbi,i,6. Performing
a descendants search from each of the points in Lbi,i,6 by using Lemma 2.5 for some refinement
constant ε′ = θ(ε) to be determined later, will guarantee that for each a ∈ OPT we traverse a point
a such that d(a, a) ≤ ε′2i. Denote the union of the points seen in such a descendants search by D.
Unfortunately, this process computes (separately) the cost of each subset of size p of candidates
traversed by taking the maximum distances from all of Q to that subset, which would take time
ε−O(ddim)np. We can speed up this process by using (a few) representatives of Q.
Speeding up the descendants search. We wish to find a bounded-size set of representatives
for the points in Q, such that the distortion caused by considering them (instead of Q) is small. To
this end, consider the set of representatives obtained as follows. Each point q ∈ Q is mapped to its
ancestor in the compacted T Q which is in Yk for the largest k ≤ i − log(1/ε′), for some refinement
constant ε′ = Θ(ε) to be determined later. Call this set of representatives Rε′. Notice that Rε′
is a subset of the compacted T Q and thus the process of this mapping can be done efficiently by
scanning the compacted T Q in linear time. Now, for each set of p center candidates X ⊂ D we
compute maxr∈Rε′ d(r, X), and take the set of candidates X that minimizes this cost.
The next lemma shows that this algorithm achieves (1 + ε) -- approximation.
Lemma 6.3. cntr(Q, { X}) = maxq∈Q d( X, q) ≤ (1 + ε)OPT.
Proof. Every q ∈ Q has a representative in Rε′, for which we can apply Lemma 2.4 and the triangle
inequality, and thus
max
q∈Q
d( X, q) < max
r∈Rε′
d( X, r) + ε′2i+1.
Recall that one of the sets of center candidates is some Aε ⊂ Yi−log(1/ε′) that is the set of ancestors
of every a ∈ A in T , where A is an optimal solution. Therefore, the returned center set X satisfies
max
r∈Rε′
d( X, r) ≤ max
r∈Rε′
d(Aε′, r).
Let r∗ ∈ Rε′ be a maximizer for the righthand side, and let q∗ ∈ Q be such that r∗ is a representative
of q∗. Let a ∈ A be the center in A which is closest to q∗, and let aε′ be the ancestor of a in Aε′.
Using the triangle inequality and Lemma 2.4 again,
d(Aε′, r∗) ≤ d(aε′, a) + d(a, q) + d(q, r∗) ≤ OPT + 2 · ε′2i+1.
Recalling from earlier that 2i < 2ALG0 ≤ 4OPT, we finally combine the inequalities above and
conclude that maxq∈Q d( X, q) ≤ OPT + 3 · ε2i+1 ≤ (1 + 24ε)OPT. To complete the proof, set ε′ to
be a power of 2 in the range [ ε
48 , ε
24 ].
19
6.2 Runtime
The running time of the above process is as follows. Locating bi for all b ∈ B using a weighted level
ancestor queries takes O(p log log log ∆) as there are only log ∆ possible nets. After constructing T Q
in O(n log n) time, the mapping of each q ∈ Q to its representative takes another O(n) time. The
descendants search from all of the O(2ddim) points in Lbi,i,6 takes O(pε′−O(ddim)) = O(pε−O(ddim)),
which also bounds the number of candidates. The number of representatives can be bounded by
the following lemma.
Lemma 6.4. Rε′ ≤ pε−O(ddim).
Proof. If all of the representatives are in Yi∗−log(1/ε′) then the size of Rε′ is at most the number of
points in Yi−log(1/ε′) which are at most 2i away from each of the p points in B. The number of such
points is bounded above by
pε′−O(ddim) ≤ pε−O(ddim).
However, the representatives do not all have to be in Yi∗−log(1/ε′). To overcome this, we charge each
representative in Rε′ to a different point in Yi∗−log(1/ε′). This mapping is done by assigning to each
point in Rε′ its ancestor in the un-compacted T Q which is in Yi∗−log(1/ε′). Notice that no two points
in Rε′ can be assigned to the same point in Yi∗−log(1/ε′), as otherwise there would be another node
in the compacted T Q which is an ancestor of those two points and in Yk for k ≤ i∗ − log(1/ε′),
which contradicts the method in which the representatives were picked.
Thus, the time it takes to test each of the (cid:0)pε−O(ddim)
(cid:1) candidates is at most O(pε−O(ddim)), and
the total runtime of the algorithm is O(n log n + p log log log ∆ + pp+1ε−p·ddim). Notice that the
runtime of the algorithm of Gonzalez is O(np), which is always bounded from above by O(n log n +
pp−1), and can thus be absorbed by the other terms.
p
7 Algorithm for p-median
Theorem 7.1. There is an algorithm that preprocesses a finite metric M in time 2O(ddim)m log ∆ log log ∆
using 2O(ddim)m memory words, where m = M , ddim = ddim(M ), and ∆ = ∆(M ), so that subse-
quent p-median queries on a set Q ⊆ M of size n, can be answered within approximation factor 1+ε,
for any desired ε ∈ (0, 1
2 ], in time O(n log n) + ε−O(ddim)(p · log n)O(1) · log log log ∆ + ε−O(p·ddim)(p ·
log n)O(p).
To a large extent, we follow an algorithm of Har-Peled and Mazumdar [HM04] for approximating
p-median clustering in Euclidean space. Their algorithm runs in time roughly O(n + exp(ε−d)(p ·
log n)O(1)), where d is the dimension in Euclidean space (in a scenario without preprocessing). In
order to give a flavor of our preprocessing model, we focus on the case of small p and employ
an abridged version of their algorithm, with runtime that grows exponentially with p. We note
that following their techniques more closely may possibly reduce the runtime, like eliminating the
exponential dependence on p.
Proof (Sketch). At a high level, the algorithm of Har-Peled and Mazumdar [HM04] works as follows.
First, construct a set A of p := p · logO(1) n centers that provides a constant factor approximation
of the p-median (formally, it is a bicriteria approximation, since p > p). Next, construct a core-set
20
S by building an exponential grid (as defined below) around each of the centers in A, and mapping
each point in Q to its (approximate) closest grid point, using near neighbor search. The size of the
core-set is roughly S ≤ ε−dA log n (but of course these points have weights that add up to n).
This means that every solution to the p-median problem on S is a good approximation for the p-
median problem on Q. Finally, construct another set of exponential grids around each of the points
in S to obtain a centroid set D, i.e., set of potential centers in the ambient (Euclidean) space, of size
roughly D ≤ ε−dSO(1) ≤ ε−2d(p · log n)O(1). Finally, use a variant of the dynamic programming
algorithm of Kolliopoulos and Rao [KR07] to quickly compute a near-optimal p-median of S among
the potential centers D.
This algorithm of Har-Peled and Mazumdar [HM04] carries over to our scenario (possibly using
some different black-box data structures, for example to solve nearest neighbor search), except for
the following two main ingredients. The first is the construction of the exponential grid, which
strongly relies on being in Euclidean space, and the ability to define points in ambient space, which
we do not enjoy in doubling dimension metrics. The second is the dynamic programming solution
of Kolliopoulos and Rao [KR07], which also exploits the Euclidean space structure. We solve the
exponential grid using T , as shown below, and skip the use of dynamic programming by performing
a brute-force search over all size p subsets (of the centroid set). It is plausible that our runtime can
be improved by adapting the solution of Kolliopoulos and Rao [KR07] to work in our case as well,
and we leave this for future work.
Exponential grid. The exponential grid of Har-Peled and Mazumdar [HM04] around a point r
with length parameter R > 0 roughly works as follows. They build O(log n) axis-parallel squares,
where the jth square has side length of R2j and is partitioned into sub-squares (i.e., a grid) of side
length O(εR2j /d), the idea being that areas closer to the point r have smaller cell size, while areas
further away have larger cell size. This construction does not carry over to doubling dimension
metrics as we cannot define grid points in ambient space. However, we make use of T to provide a
set with similar properties.
We provide a sketch of the idea in order to ease presentation, but point out that some of our
constants can to be refined. Given r, we use a weighted level ancestor query [FM96, KL07] to locate
its ancestor rlog R ∈ Ylog R in T . A descendants search, using Lemma 2.5 starting from rlog R with
refinement constant ǫ will give us a good resolution for points that are roughly at most distance R
away from r. Let rj+log R ∈ Yj+log R be an ancestor of r in T , for 0 ≤ j ≤ O(log n). We perform a
descendants search using Lemma 2.5 starting from each such rj+log R, with refinement constant ε.
Notice that for a specific j, for points within distance 2j+log R = 2jR from r, the descendant search
starting from rj+log R reaches a set of points which are in Yj+log R−log(1/ε) which is similar to the
resolution obtained from the exponential grid in Euclidean space.
The union of all of the points seen during all of the descendants searches on all O(log n) levels
provides a set of size ε−O(ddim) log n, which gives us (i.e., in doubling dimension metrics) the same
properties as the exponential grid does in a Euclidean space. Thus we obtain a core-set S and
centroid set D both of size at most ε−O(ddim)(p · log n)O(1). Finally, perform an exhaustive search
through all subsets of size p of the centroid set D and compute the cost of each such set, which takes
total time (cid:0)D
p (cid:1) · O(Sp) ≤ ε−O(p·ddim)(p · log n)O(p). Notice that a weighted level ancestor query
is performed for each of the points in the core-set S, which increases the runtime by ε−O(ddim)(p ·
log n)O(1) · log log log ∆.
21
References
[AGGM06] I. Abraham, C. Gavoille, A. V. Goldberg, and D. Malkhi. Routing in networks with low doubling
dimension. In 26th IEEE International Conference on Distributed Computing Systems, page 75.
IEEE, 2006.
[AP02]
P. K. Agarwal and C. M. Procopiuc. Exact and approximation algorithms for clustering. Algo-
rithmica, 33:201 -- 226, 2002.
[ARR98]
S. Arora, P. Raghavan, and S. Rao. Approximation schemes for euclidean k-medians and related
problems. In 13th Annual ACM Symposium on the Theory of Computing, pages 106 -- 113, 1998.
[BF00]
M. A. Bender and M. Farach-Colton. The LCA problem revisited. In LATIN 2000: Theoretical
Informatics, pages 88 -- 94, 2000.
[BGK12] Y. Bartal, L.-A. Gottlieb, and R. Krauthgamer. The traveling salesman problem: Low-
dimensionality implies a polynomial time approximation scheme. In 44th symposium on Theory
of Computing, pages 663 -- 672. ACM, 2012.
[BHI02]
[BKL06]
[BLL09]
[CG06]
[Che06]
[FM96]
M. Badoiu, S. Har-Peled, and P. Indyk. Approximate clustering via core-sets. In 34th Annual
ACM Symposium on Theory of Computing, pages 250 -- 257, 2002.
A. Beygelzimer, S. Kakade, and J. Langford. Cover trees for nearest neighbor. In 23rd interna-
tional conference on Machine learning, pages 97 -- 104. ACM, 2006.
N. H. Bshouty, Y. Li, and P. M. Long. Using the doubling dimension to analyze the generalization
of learning algorithms. Journal of Computer and System Sciences, 75(6):323 -- 335, 2009.
R. Cole and L.-A. Gottlieb. Searching dynamic point sets in spaces with bounded doubling
dimension.
In 38th annual ACM symposium on Theory of computing, pages 574 -- 583. ACM,
2006.
K. Chen. On k-median clustering in high dimensions. In 17th Annual ACM-SIAM Symposium
on Discrete algorithm, pages 1177 -- 1185. ACM, 2006.
M. Farach and S. Muthukrishnan. Perfect hashing for strings: Formalization and algorithms. In
7th Annual Symposium on Combinatorial Pattern Matching, pages 130 -- 140, 1996.
[GKK10]
L.-A. Gottlieb, L. Kontorovich, and R. Krauthgamer. Efficient classification for metric data. In
23rd Conference on Learning Theory, pages 433 -- 440. Omnipress, 2010.
[GKL03]
A. Gupta, R. Krauthgamer, and J. R. Lee. Bounded geometries, fractals, and low-distortion
embeddings.
In 44th Annual IEEE Symposium on Foundations of Computer Science, pages
534 -- 543, October 2003.
[Gon85]
T. F. Gonzalez. Clustering to minimize the maximum intercluster distance. Theoret. Comput.
Sci., 38(2-3):293 -- 306, 1985.
[HM04]
[HM06]
[HT84]
[Ind99]
S. Har-Peled and S. Mazumdar. On coresets for k-means and k-median clustering.
Annual ACM Symposium on Theory of Computing,, pages 291 -- 300, 2004.
In 36th
S. Har-Peled and M. Mendel. Fast construction of nets in low-dimensional metrics and their
applications. SIAM Journal on Computing, 35(5):1148 -- 1184, 2006.
D. Harel and R. E. Tarjan. Fast algorithms for finding nearest common ancestors. SIAM J.
Comput., 13(2):338 -- 355, 1984.
P. Indyk. Sublinear time algorithms for metric space problems. In Proceedings of the 31st Annual
ACM Symposium on Theory of Computing, pages 428 -- 434. ACM, 1999.
22
[KL04a]
In
R. Krauthgamer and J. R. Lee. The black-box complexity of nearest neighbor search.
31st International Colloquium on Automata, Languages and Programming, Lecture Notes in
Computer Science, pages 858 -- 869. Springer, July 2004.
[KL04b]
R. Krauthgamer and J. R. Lee. Navigating nets: Simple algorithms for proximity search. In
15th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 791 -- 801, January 2004.
[KL07]
[KR07]
T. Kopelowitz and M. Lewenstein. Dynamic weighted ancestors. In 18th Annual ACM-SIAM
Symposium on Discrete Algorithms,, pages 565 -- 574, 2007.
S. G. Kolliopoulos and S. Rao. A nearly linear-time approximation scheme for the euclidean
k-median problem. SIAM J. Comput., 37(3):757 -- 782, 2007.
[KRX08] G. Konjevod, A. W. Richa, and D. Xia. Dynamic routing and location services in metrics of low
doubling dimension. In 22nd International Symposium on Distributed Computing, volume 5218
of Lecture Notes in Computer Science, pages 379 -- 393. Springer, 2008.
[KSS10]
A. Kumar, Y. Sabharwal, and S. Sen. Linear-time approximation schemes for clustering problems
in any dimensions. J. ACM, 57(2), 2010.
[KSW09]
J. Kleinberg, A. Slivkins, and T. Wexler. Triangulation and embedding using small sets of
beacons. J. ACM, 56:32:1 -- 32:37, September 2009.
[Sli07]
[Tal04]
A. Slivkins. Distance estimation and object location via rings of neighbors. Distributed Com-
puting, 19:313 -- 333, 2007.
K. Talwar. Bypassing the embedding: Algorithms for low dimensional metrics. In Proceedings
of the 36th Annual ACM Symposium on Theory of Computing, pages 281 -- 290, 2004.
23
|
1703.06053 | 2 | 1703 | 2017-03-21T12:29:29 | Fast Non-Monotone Submodular Maximisation Subject to a Matroid Constraint | [
"cs.DS"
] | In this work we present the first practical $\left(\frac{1}{e}-\epsilon\right)$-approximation algorithm to maximise a general non-negative submodular function subject to a matroid constraint. Our algorithm is based on combining the decreasing-threshold procedure of Badanidiyuru and Vondrak (SODA 2014) with a smoother version of the measured continuous greedy algorithm of Feldman et al. (FOCS 2011). This enables us to obtain an algorithm that requires $O(\frac{nr^2}{\epsilon^4} \big(\frac{a+b}{a}\big)^2 \log^2({\frac{n}{\epsilon}}))$ value oracle calls, where $n$ is the cardinality of the ground set, $r$ is the matroid rank, and $ b, a \in \mathbb{R}^+$ are the absolute values of the minimum and maximum marginal values that the function $f$ can take i.e.: $ -b \leq f_S(i) \leq a$, for all $i\in E$ and $S\subseteq E$, (here, $E$ is the ground set). The additional value oracle calls with respect to the work of Badanidiyuru and Vondrak come from the greater spread in the sampling of the multilinear extension that the possibility of negative marginal values introduce. | cs.DS | cs |
Fast Non-Monotone Submodular Maximisation
Subject to a Matroid Constraint
Pau Segui-Gasco ∗†
Hyo-Sang Shin †
Abstract
In this work we present the first practical (cid:0) 1
e − ǫ(cid:1)-approximation algorithm to maximise a general
non-negative submodular function subject to a matroid constraint. Our algorithm is based on
combining the decreasing-threshold procedure of Badanidiyuru and Vondrak (SODA 2014) with a
smoother version of the measured continuous greedy algorithm of Feldman et al. (FOCS 2011).
This enables us to obtain an algorithm that requires O( nr2
log2( n
ǫ )) value oracle calls, where
d, ¯d ∈ R+ are the absolute values
n is the cardinality of the ground set, r is the matroid rank, and
¯
d ≤ fS(i) ≤ ¯d,
of the minimum and maximum marginal values that the function f can take i.e.: −
¯
for all i ∈ E and S ⊆ E, where E is the ground set. The additional value oracle calls with respect
to the work of Badanidiyuru and Vondrak come from the greater spread in the sampling of the
multilinear extension that the possibility of negative marginal values introduce.
ǫ4 (cid:0)
¯d+
d
¯
¯d (cid:1)2
1 Introduction
Submodular maximisation problems have drawn a lot of attention recently [1]. This interest is due
to a good confluence of theoretical results and practical applicability. Intuitively, a submodular
set function is said to be so because it exhibits diminishing marginal returns, i.e.: the marginal
value that an element adds to a set decreases as the size of the set increases. This simple property
arises naturally in many applications and is what enables good theoretical tractability. It has been
used in a variety of application domains, to name but a few: markets [2, 3], influence in networks
[4], document summarisation [5], and sensor placement [6, 7]. In particular, this paper focuses on
matroid-constrained submodular maximization of a general non-negative submodular function.
Matroids are an incredibly powerful abstraction of independence. They capture seemingly
disconnected notions such as linear independence, forests in graphs, traversals, among many others.
Interestingly, with linear sum objectives (modular function), they are inextricably linked with the
greedy algorithm. If the greedy algorithm is optimal, then there is an implicit matroid; if there is a
matroid then the greedy algorithm is optimal [8]. They provide a flexible framework to characterise
a variety of relevant constraints such as: partitions, schedules, cardinality, or even rigidity.
∗[email protected]
†Centre for Autonomous and Cyber-Physical Systems, Institute of Aerospace Sciences, School of
Aerospace Transport and Manufacturing, Cranfield University, Bedfordshire, UK.
1
1 Introduction
2
Importantly, the combination of a submodular function and a matroid constraint plays a uni-
fying role for many well-known combinatorial optimisation problems such as: Submodular Welfare
(also known as Submodular Task Allocation), Max k-Cover, Max Generalised Assignment, Max
Facility Location, and Constrained Max Cut (e.g. Max Bisection) among others.
However, maximising a submodular function subject to a matroid constraint is NP-Hard. Hence
much of the research effort has focused on obtaining good approximation algorithms. A classic re-
sult by Nemhauser, Wolsey, and Fisher [9] shows that the Greedy Algorithm is a 1
2 -approximation
for non-negative monotone submodular functions. More recently, a fruitful avenue of research was
spurred by Vondrak [10, 11] by showing that solving a relaxation of the problem based on the
multilinear extension using the Continuous Greedy Algorithm and then rounding the result yields
good approximation algorithms. They present their result in the context of non-negative monotone
submodular functions to yield a (1 − 1
e )-approximation. Shortly after, Feldman et al. [12] modified
Vondrak's algorithm to develop the Measured Continuous Greedy Algorithm which supported both
non-negative non-monotone and monotone submodular functions. Their algorithm was the first
to achieve a 1
e -approximation for general non-negative submodular functions subject to a matroid
constraint. Both continuous greedy algorithms can find the aforementioned approximation bounds
for solving relaxation problems subject to the more general constraint class of down-closed solvable
polytopes. An important breakthrough in this field are Contention Resolution Schemes, a rounding
framework proposed in [13], because they provide a paradigm for developing rounding schemes for a
combination of useful constraints including matroids and knapsacks. Thus the combination of con-
tinuous greedy relaxations and Contention Resolution schemes enabled approximation algorithms
for many important submodular maximization problems.
However, continuous greedy algorithms are impractically slow. Vondrak's algorithm required
Θ(n8) value oracle calls, while Feldman's Measured Continuous Greedy requires O(n6) assuming
oracle access to the multilinear extension, which needs to be sampled in general, creating additional
overhead, possibly around O(n3) or O(n2).
algorithm that uses O(cid:0) nr
To remedy this, Badanidiyuru and Vondrak [14] proposed an efficient(cid:0)1 − 1
e − ǫ(cid:1)-approximation
ǫ(cid:1) value oracle calls, for non-negative monotone submodular function
maximisation subject to a matroid constraint. Where n is the cardinality of the ground set, and
r is the rank of the matroid. They achieve this by using a Decreasing-Threshold procedure that
enables a reduction of both the number of steps and the number of samples needed at each step of
the continuous greedy process.
ǫ4 log2 n
In the inapproximability front, for matroid constrained non-negative monotone submodular
maximisation Feige [15] showed that improving over the 1 − 1
e threshold is NP-Hard, and so the
continuous greedy guarantees are tight. In the non-monotone case, [16] showed that no polynomial
time algorithm can achieve an approximation better than 0.478. Closing the gap between 1
e and
0.478 remains an important open problem where recent advances have been made: Ene and Nguyen
in [17] give a 0.372-approximation, while Feldman et al in [18] improve it to a 0.385-approximation.
This improvements are relatively small, recall that 1
e ≈ 0.368, but show that there is room for future
improvement. Both algorithms are based on the measured continuous greedy, and would therefore
benefit from the techniques presented here.
Our contribution in this work is the first(cid:0) 1
non-negative submodular function subject to a matroid constraint with a practical time complexity.
Our algorithm is based on combining the Decreasing-Threshold procedure of [14] with a smoother
version of the measured continuous greedy algorithm of [12]. This allows us to obtain an algorithm
that requires O( nr2
d, ¯d ∈ R+ are the absolute values of
¯
ǫ )) value oracle calls, where
e − ǫ(cid:1)-approximation algorithm to maximise a general
log2( n
ǫ4 (cid:0)
¯d+
d
¯
¯d (cid:1)2
1 Introduction
3
d ≤ fS(i) ≤ ¯d,
the minimum and maximum marginal values that the function f can take, i.e. −
¯
for all i ∈ E and S ⊆ E. The additional oracle calls with respect to [14] come from the additional
spread in the sampling of the multilinear extension that the possibility of negative marginal values
introduces.
Definitions Let us now introduce the definitions of a matroid and a submodular function. A
function f : 2E → R+ on a set E is said to be submodular if:
f (A) + f (B) ≥ f (A ∪ B) + f (A ∩ B), for all A, B ⊆ E.
(1)
A more intuitive, but equivalent, definition can be formulated in terms of the marginal value added
by an element: given Y, X ⊆ E satisfying X ⊆ Y and x ∈ E \ Y , then f (X + x) − f (X) ≥
f (Y + x) − f (Y ). Herein we overload the symbol + (−) to use it as shorthand notation for the
addition (subtraction) of an element to a set, i.e. S + i = S ∪ {i} (S − i = S \ {i}). A matroid
can be defined as follows [8]: a pair M = (E, I) is called a matroid if E is a finite set and I is a
nonempty collection of subsets of E satisfying: ∅ ∈ I; if A ∈ I and B ⊆ A, then B ∈ I; and if
A, B ∈ I and A < B, then A + z ∈ I for some z ∈ B \ A.
A matroid base B ⊆ E is a maximally independent set B ∈ I, that becomes dependent by
adding an additional element e ∈ E \ B, i.e. B + e /∈ I. A key property of matroids is that we
can exchange elements between bases, let B denote the set of all bases of a matroid M, then the
exchange property can be is formalised as follows:
Lemma 1.1. (Corollary 39.12A in [8]) Let M = (N, I) be a matroid, and B1, B2 ∈ B be two bases.
Then there is a bijection φ : B1 → B2 such that for every b ∈ B1 we have B1 − b + φ(b) ∈ B.
Now we can formally define our target problem:
Given a ground set E, a matroid defined upon it M = (E, I), and a general non-negative submodular
function f : 2E → R+, find:
max
S∈I
f (S).
(2)
A common approach to solve combinatorial optimisation problems is by using relaxation and
rounding. This involves solving a continuous problem, in which we allow fractional solutions, and
a rounding procedure which transforms the fractional output back to a discrete solution. Now let
us explain the relaxation that is used in continuous greedy algorithms. There are two key elements
needed to define the relaxation: the domain that contains fractional solutions, and a function to
evaluate fractional solutions. In our problem, the most natural representation for the domain is the
matroid polytope P (M), which is the convex hull of the incidence vectors of the independent sets in
the matroid, i.e.: P (M) = Conv({1S : ∀S ∈ I}) [8]. Here 1S denotes the incidence vector of a set
S ⊆ E, which contains a 1 for each element in S and 0 elsewhere, thus 1S ∈ [0, 1]E. The function
for evaluation of fractional solutions is the multilinear extension: given a set E and a submodular
set function f : 2E → R+, its multilinear extension F : [0, 1]E → R+ is defined as:
F (y) , E[f (R(y))] = XS⊆E
f (S)Yi∈S
yiYj /∈S
(1 − yj),
(3)
where R(y) is a random set containing each element i ∈ E with probability yi. Additionally, F can
be lower bounded in terms of f :
2 Algorithm
4
Lemma 1.2. (from [12]) Let f : 2E → R+ be a submodular function; let F : [0, 1]E → R+ be the
multilinear extension of f ; and let y ∈ [0, 1]E be a vector whose components are bounded by a, i.e.
yi ≤ a, ∀i ∈ E. Then, for every S ⊆ E, we have that
F (y ∨ 1S) ≥ (1 − a)f (S).
(4)
A key magnitude used in the Measured Continuous Greedy Algorithm is the marginal value
of an element e ∈ E given a fractional solution y ∈ P (M), we refer to it by ∆Fe(y), defined as
follows:
∆Fe(y) , E[fR(y)(e)] = F (y ∨ 1e) − F (y).
(5)
In other words, it is the value that would be created by adding the full element i to the fractional
solution y. Here we have used the common shorthand notation for the marginal value of an element
fS(e) = f (S + e) − f (S).
In general, given an arbitrary non-negative submodular function there is no closed form of
multilinear extension that enables us to evaluate it efficiently. The usual way to deal with this is
to sample it, and so we need the following concentration inequality that enables us to bound the
sampling error:
Lemma 1.3. (Hoeffding Bound, Theorem 2 in [19]) Let X1, ..., Xm be independent random variables
such that for each i, a ≤ Xi ≤ b, with a, b ∈ R. Let X = 1
i=1 Xi. Then
Pr[ X − E(X) > t] ≤ 2e
.
mPm
− 2t2
(b−a)2 m
In this work we shall not dwell on the rounding step. Suffice it to say that there exist algorithms,
such as Contention Resolution Schemes [20], Pipage-Rounding [21, 11, 22], or Swap-Rounding [13],
that given a general non-negative submodular function f , its multilinear extension F , and a point
y ∈ P (M), find a set S ∈ I, such that f (S) ≥ F (y). The most efficient technique is Swap-Rounding,
and its results are used in [14]. However, the swap-rounding results in [13] hinge around Chernoff-like
concentration bounds for monotone submodular functions, but in light of the concentration bounds
for non-monotone submodular functions in [23], we believe that this technique can be adapted to
work with non-monotone submodular functions.
2 Algorithm
The algorithm we present here is based on the Decreasing-Threshold procedure that enabled Badani-
diyuru and Vondrak [14] to achieve a very efficient algorithm, O( nr
ǫ ), for maximising a non-
negative monotone submodular function subject to a matroid constraint. The continuous greedy
algorithm finds an approximate solution to the relaxation, y ∈ P (M), by integrating between t = 0
to t = 1 the differential equation dy
basically the standard continuous greedy algorithm, but uses a much more efficient Decreasing-
Threshold procedure to find the maximum marginal improvement direction, i.e. vi ∀i ∈ E in the
equation above, in each iteration. To enable an approximation algorithm for the non-monotone
case Feldman [12] proposed the measured continuous greedy algorithm, which similarly integrates
dt = argmaxv∈P (M)Pi∈E ∆Fi(y)vi. The algorithm in [14] is
ǫ4 log2 n
2 Algorithm
5
Algorithm 1: Accelerated Measured Continuous Greedy
Input : f : 2E → R+, ǫ ∈ [0, 1], I ⊆ 2E.
Output: A point y ∈ P (M), such that F (y) ≥ ( 1
e − 2ǫ)f (OP T ).
// Initialisation
y(0) ← 0
δ ← ǫ
// Main Loop
for t = {0, δ, 2δ, 3δ, . . . , 1 − δ} do
B(t) ← Decreasing-Threshold(f, y(t), ǫ, δ, I )
for i ∈ E do
if i ∈ B(t) then
yi(t + δ) ← 1 + e−δ(yi(t) − 1)
else
yi(t + δ) ← yi(t)
Return: y(1)
Algorithm 2: Decreasing-Threshold
Input : f : 2E → R+, y ∈ [0, 1]E , ǫ ∈ [0, 1], δ ∈ [0, 1], I ⊆ 2E .
Output: A set B ⊆ E, such that B ∈ I.
// Initialisation
B ← ∅;
¯d ← maxi∈E f (i);
¯y′ ← maxi∈E(1 + e−δ(yi − 1));
// Main Loop
for (w = d; w ≥ ǫ
for e ∈ E do
¯d
r (1 − ¯y′); w ← w(1 − ǫ)) do
we(B, y) ← ∆Fe(y(B, δ)),
// averaging O(cid:18) r2
ǫ2(cid:0)
if we(B, y) ≥ w and B + e ∈ I then
¯d+
d
¯
¯d (cid:1)2
log(E)(cid:19) iid random samples.
B ← B + e
Return: B
*Note that the notation y(B, δ) means yi(B, δ) = yi(t) for i /∈ B, and
yi(B, δ) = 1 + e−δ(yi − 1) for i ∈ B.
3 Algorithm Analysis
6
dy(t)
here).
dt = (1 − y(t)) ⊙ argmaxv∈P (M)Pi∈E ∆Fi(y)vi (⊙ represents element by element multiplication
The advantage of the Decreasing-Threshold procedure is that it requires a smaller number
of integration steps and a smaller number of samples per step. Here we apply its key idea to
the measured continuous greedy. However, we need to perform two crucial modifications. The
first modification is a smoother version of the integration rule in the measured continuous greedy
algorithm of Feldman et al. [12]. We use an update step that makes the algorithm more continuous-
like. Instead of using the integration step y(t + δ) = y(t) + δ(1 − y(t)) as proposed by [12], we
increment the solution at a rate of dy(t)
dt = 1 − y(t) by using the integration rule: y(t + δ) =
1 + e−δ(y(t) − 1), see Algorithm 1. This allows us to smooth the tradeoff between error and running
time. The second modification is in the sampling error up to which we estimate the marginal values
of the multilinear extension in the Decreasing-Threshold procedure. Badanidiyuru and Vondrak
[14] used an additive and multiplicative error bound which enabled them to use only O( 1
ǫ2 r log(n))
samples. However, since our algorithm is meant to work with non-monotone functions, we may
have negative marginal values, this increases the number of samples required for two reasons: first,
it prevents us from using a multiplicative-additive error bound; and second the marginal values,
when sampled, have inherently more spread. To quantify our error we use Hoeffding's concentration
d, ¯d ∈ R+ are the absolute values of
¯
d ≤ fS(i) ≤ ¯d, for
the minimum and maximum marginal values that the function f can take, i.e. −
¯
all i ∈ E and S ⊆ E. (Recall that if f were to be monotone we would have
d = 0.) The resulting
¯
)
value oracle calls. Using only an additive bound instead of an additive and multiplicative introduces
an extra O(r) factor in the number of samples required per evaluation of the multilinear extension.
inequality, which forces us to use O(cid:0) r2
ǫ2(cid:0)
Decreasing-Threshold procedure is presented in Algorithm 2, and it takes an additional O(r(cid:0)
While we also incur the additional O(cid:0)(cid:0)
¯d (cid:1)2
¯d (cid:1)2(cid:1) term to cope with the larger spread that is introduced
log(n)(cid:1), where
¯d (cid:1)2
by the possibility of negative marginal values. This results in an algorithm that finds an 1
e − ǫ
approximation with a number of value oracle calls of O( nr2
ǫ )). It is not as efficient
as the algorithm for the monotone case, but to the best of our knowledge, this constitutes the first
practical 1
e − ǫ approximation algorithm for general non-negative submodular function subject to a
matroid constraint.
¯d (cid:1)2
ǫ4 (cid:0)
log2( n
¯d+
d
¯
¯d+
d
¯
¯d+
d
¯
¯d+
d
¯
3 Algorithm Analysis
We split the analysis of the accelerated measured continuous greedy algorithm in three parts:
first we show that the solution produced is feasible, i.e. y(1) ∈ P (M); then we prove the 1
e − ǫ
approximation; and finally we study its running time.
Feasibility
Theorem 3.1. The accelerated measured continuous greedy algorithm produces a feasible fractional
solution, i.e., y(1) ∈ P (M).
Proof. We follow the approach used by [12]. We first define a vector x that coordinate wise upper-
bounds y(1). Then, given that P (M) is down-monotone, we only need to show that x is in P (M)
1
δ −1
l=0 1B(y(lδ)). This is a coordinate-wise
upper bound of y(1) because when i ∈ B, we have that yi(t + δ) − yi(t) = 1 + e−δ(yi(t) − 1) − yi(t) =
to show that y(1) ∈ P (M). Consider the vector x = δP
3 Algorithm Analysis
7
(1 − e−δ)(1 − yi(t)) ≤ 1 − e−δ ≤ δ, for all δ ∈ [0, 1]; and when i /∈ B yi(t + δ) − yi(t) = 0. We now
show that x is in P (M). First, note that by definition 1B ∈ P (M). Then, observe that x/δ is the
sum of 1
δ points in P (M), thus (x/δ)/(1/δ) = x is a convex combination of points in P (M), hence
x ∈ P (M), and consequently y(1) ∈ P (M).
In fact, it is possible to find a point that still lies in P (M) even with a t > 1, specifically
stopping at a value depending on a magnitude called the density of the matroid. This yields tighter
approximation bounds that match those found for particular matroids, such as partition matroids.
Asymptotically, however, these bounds are the same as those presented here. Therefore, in the aim
of simplicity, the analysis is carried out with a stopping time of 1. The interested reader is referred
to [12].
Now let us establish a bound to the coordinates of y(t) that will become useful later in the
analysis of the approximation ratio.
Lemma 3.2. At time 0 ≤ t ≤ 1, we have that:
yi(t) ≤ 1 − e−t,
∀i ∈ E.
(6)
Proof. Consider the recurrence g(n + 1) = 1 + e−δ(1 − g(n)), with g(0) = 0. This recurrence has the
following solution: g(n) = 1 − e−nδ. Now, in our algorithm t is incremented linearly, so the number
of iteration, n, and t are related by t = nδ. At t, all the coordinates of y either, they stay constant,
i.e. yi(t + δ) = yi(t), when i /∈ B, or increase, i.e. yi(t + δ) = 1 + e−δ(yi(t) − 1), when i ∈ B. Hence,
given that g(n) is non-decreasing, the recurrence g is an upper bound because it corresponds to
incrementing in each and every iteration. Consequently, at t, y(t) ≤ g( t
δ ) = 1 − e−t.
Approximation Ratio First we show the gain that the algorithm makes in a single step, and then
use this to build a recurrence relation that yields the approximation ratio. But before we do that,
we bound the error introduced by sampling:
Corollary 3.3. Given a non-negative submodular function f : 2E → R+, and a point y ∈ [0, 1]E;
d ≤ fS(j) ≤ ¯d for all
d, ¯d ∈ R+ be the minimum and maximum marginal values of f , such that −
let
¯
¯
S ⊆ E and j ∈ E; let R1, R2, ..., Rm be iid samples drawn from R(y), let wj(y) = 1
i=1 fRi(j);
and let f (OP T ) = maxS∈I f (S). Then,
mPm
Pr(wj (y) − ∆Fj (y) ≥ βf (OP T )) ≤ 2e
d(cid:1)2
−2mβ2(cid:0) ¯d
¯d+
¯
.
Proof. Immediate application from the Hoeffding bound in Lemma 1.3, noting that f (OP T ) =
maxS∈I f (S) ≥ maxe∈E f ({e}) = ¯d.
Now we can present the improvement made by the algorithm in a single step:
Lemma 3.4. Let OP T be an optimal solution. Given a fractional solution y, the Decreasing-
Threshold produces a set B such that, with y′ = 1 + e−1Bδ ⊙ (y − 1), we have:
F (y′) − F (y) ≥ (1 − e−δ)(cid:18)(1 − 4ǫ)(1 − ¯y′)f (OP T ) − F (y′)(cid:19)
(7)
3 Algorithm Analysis
8
Proof. This proof follows closely the proof of Claim 4.1 in [14] but with several modifications to
avoid assuming monotonicity, namely: the stopping threshold, the sampling error, and the increment
bound. Assume that the Decreasing-Threshold procedure returns a sequence of r elements B =
{b1, b2, . . . , br}, indexed in the order in which they were chosen. Let O = {o1, o2, . . . , or} be an
optimal solution indexed as per the exchange property of the matroids in Lemma 1.1 such that
φ(bi) = oi. Additionally, let Bi and Oi denote the first i elements of B and O respectively, i.e. Bi
is the sequence in which the elements have been added to B in algorithm 2 up until the ith element
was added. If the procedure returns fewer than r elements or the optimal solution contained fewer
than r elements, formally we just add dummy elements with value 0, so that B = r and O = r.
Now let us bound the marginal values of the elements selected by the Decreasing-Threshold
procedure with respect to those in the optimal solution. Recall that y(S, δ) is the notation that we
use to refer to the point such that yk(S, δ) = yk if k /∈ S and yk(S, δ) = 1 + e−δ(yk − 1) if k ∈ S.
When bi is selected, let w be the current threshold, hence wbi (y(Bi−1, δ)) ≥ w. At this point oi is
a candidate element, thus we have one of two situations depending on whether the procedure has
finished: if the threshold has not dropped below ǫ
r d(1 − ¯y′), the value of woi (y(Bi−1, δ)) must be
below the threshold in the previous iteration, i.e. woi (y(Bi−1, δ)) ≤ w
(1−ǫ) (otherwise it would have
been chosen already); conversely, if the procedure has terminated, bi is a dummy element with value
¯d
r (1− ¯y′).
0, and the value of woi (y(Bi−1, δ)) is below the stopping threshold, i.e. woi(y(Bi−1, δ)) ≤ ǫ
Consequently, we can relate the marginal value estimate of bi to that of oi:
wbi (y(Bi−1, δ)) ≥ (1 − ǫ)woi (y(Bi−1, δ)) − ǫ
¯d
r
(1 − ¯y′).
Note that when bi is selected we have that Bi−1 + bi ∈ I, and by the definition of oi (i.e. the
matroid exchange property) Bi−1 + oi ∈ I.
Now we need to bound the error incurred by sampling. From Lemma 3.3 we can sample
r f (OP T )(1 − ¯y′) by taking the average of
the marginal values up to an additive error of β = ǫ
log(E)(cid:19) samples with high probability (i.e. with a bad estimate probability decreas-
E ). (We do not have a term O( 1
ing with 1
time, t = 1, and the update rule enforce (see Lemma 3.2) that (1 − ¯y′) ≥ 1
1−¯y ′ ) in the number of samples because our stopping
e ). Thus, we can write:
O(cid:18) r2
ǫ2(cid:0)
¯d+
d
¯
¯d (cid:1)2
wbi (y(Bi−1, δ)) ≤ ∆Fbi (y(Bi−1, δ)) +
woi(y(Bi−1, δ)) ≥ ∆Foi(y(Bi−1, δ)) −
f (OP T )(1 − ¯y′)
f (OP T )(1 − ¯y′)
ǫ
r
ǫ
r
which, with ¯d ≤ f (OP T ), can be combined with the prior bound to yield:
∆Fbi (y(Bi−1, δ)) ≥ (1 − ǫ)∆Foi (y(Bi−1, δ)) − 3
ǫ
r
f (OP T )(1 − ¯y′).
(8)
Then, we can bound the improvement that the Decreasing-Threshold procedure obtains:
3 Algorithm Analysis
9
F (y′) − F (y)
r
(F (y(Bi, δ)) − F (y(Bi−1, δ))
(y′
bi − ybi)
∂F
∂ybi(cid:12)(cid:12)(cid:12)(cid:12)y=y(Bi−1,δ)
∂F
(1 − e−δ)(1 − ybi)
=
=
=
r
Xi=1
Xi=1
Xi=1
r
∂ybi(cid:12)(cid:12)(cid:12)(cid:12)y=y(Bi−1,δ)
r
r
ǫ
r
= (1 − e−δ)
∆Fbi (y(Bi−1, δ))
Xi=1
Xi=1(cid:18)(1 − ǫ)∆Foi (y(Bi−1, δ)) − 3
≥ (1 − e−δ)
f (OP T )(1 − ¯y′)(cid:19)
Xi=1(cid:18)∆Foi(y(Bi−1, δ))(cid:19) − 3ǫf (OP T )(1 − ¯y′)(cid:19)
= (1 − e−δ)(cid:18)(1 − ǫ)
≥ (1 − e−δ)(cid:18)(1 − ǫ)(cid:0)F (y′ ∨ 1OPT) − F (y′)(cid:1) − 3ǫf (OP T )(1 − ¯y′)(cid:19)
≥ (1 − e−δ)(cid:18)(1 − 4ǫ)(1 − ¯y′)f (OP T ) − F (y′)(cid:19).
r
The second equality comes from the the multilinearity of F . With the update step y′ =
1 + e−δ(y − 1), we have that the increment is y′ − y = (1 − e−δ)(1 − y), which gives the third
inequality. The fourth equality is by definition of ∆Fe. The fifth inequality is by the bound in
equation 8, the sixth by submodularity, and the last one by Lemma 1.2.
Finally, we can use the above result to build a recurrence relation that yields the 1
e − ǫ approx-
imation ratio.
Theorem 3.5. The accelerated measured continuous greedy algorithm returns a point y∗ ∈ P (M),
such that:
F (y∗) ≥(cid:18) 1
e
− 2ǫ(cid:19)f (OP T ).
(9)
Proof. From Lemma 3.4 we have that:
F (y(t + δ)) − F (y(t)) ≥ (1 − e−δ)(cid:18)(1 − 4ǫ)(1 − ¯y(t + δ))f (OP T ) − F (y(t + δ))(cid:19)
We can now use the bound on the value of the coordinates of y in Lemma 3.2, we have that
yi(t) ≤ 1 − e−t ∀i ∈ E, hence we can write:
F (y(t + δ)) − F (y(t)) ≥ (1 − e−δ)(cid:18)(1 − 4ǫ)e−(δ+t)f (OP T ) − F (y(t + δ))(cid:19).
(10)
3 Algorithm Analysis
10
Consider the recurrence relation a(n + 1) − a(n) = k1(k2 exp(−(n + 1)ǫ) − a(n + 1)), which,
with a(0) = a0, has the following solution
a(n) =
(
1
k1+1 )n(a0(−k1 + eǫ − 1) + k1k2) − k1k2e−nǫ
−k1 + eǫ − 1
.
(11)
This recurrence is equivalent to equation 3 if we set k1 = (1 − e−δ), k2 = (1 − 4ǫ), n = t
δ , and δ = ǫ.
Thus, substituting and simplifying assuming that F (0) ≥ 0 (due to the non-negativity of f ), we
have the following lower bound on the value of the solution y(t) for any time t ∈ [0, 1]:
F (y(t)) ≥
(1 − e−ǫ)(1 − 4ǫ)eǫ(−( t
ǫ +1))(eǫ( t
e−ǫ + eǫ − 2
ǫ +1)(
So, when the algorithm ends at t = 1, we have:
F (y(1)) ≥
(1 − e−ǫ)(1 − 4ǫ)eǫ(−( 1
ǫ +1))(eǫ( 1
e−ǫ + eǫ − 2
ǫ +1)(
1
2−e−ǫ )t/ǫ − eǫ)
f (OP T ).
1
2−e−ǫ )1/ǫ − eǫ)
f (OP T ).
We can find a more intuitive version of this bound by observing that, clearly, for 0 ≤ ǫ ≤ 1:
1
e
−
Hence:
(1 − e−ǫ)(1 − 4ǫ)eǫ(−( 1
ǫ +1))(eǫ( 1
e−ǫ + eǫ − 2
ǫ +1)(
1
2−e−ǫ )1/ǫ − eǫ)
≤
5
e
ǫ ≤ 2ǫ
F (y(1)) ≥(cid:18) 1
e
− 2ǫ(cid:19)f (OP T ).
(12)
(13)
Finally, from Lemma 3.2 we have that y(1) ∈ P (M).
Running Time As it is common in the submodular maximisation literature, we quantify the
running time in terms of value oracle calls to the submodular function and matroid independence
oracle calls. We analyse the running time of the algorithm in two steps: first we study the running
time of the Decreasing-Threshold procedure, and then that of the the continuous greedy.
Lemma 3.6. The Decreasing-Threshold procedure makes
O(cid:18) Er2
ǫ3
log(E) log( r
ǫ )(cid:16) ¯d+
d
¯
¯d (cid:17)2(cid:19) value oracle calls, and O(cid:0) E
log r
ǫ
ǫ(cid:1) independence oracle calls.
log ǫ
Proof. First, the number of values that the threshold takes in the Decreasing-Threshold procedure
to reach the stopping threshold is, considering that the term (1 − ¯y) ≥ 1
e due to Lemma 3.2,
log (1−ǫ) ). Second, for each threshold value, the algorithm performs O(E) estimates of ∆Fe,
O(
and O(E) calls to the independence oracle. Therefore, the number of independence oracle calls is
the number of calls per threshold step multiplied by the number of threshold steps, i.e.:
r
O(cid:18) E
ǫ
log
r
ǫ(cid:19),
(14)
4 Discussion and Future Work
11
which has been simplified noting that
log ǫ
log 1−ǫ ≤ 1
r
ǫ log( r
ǫ ).
the number of value oracle calls is
Now, each estimate of ∆Fe requires O(cid:18) r2
ǫ2(cid:0)
O Er2
log(E) log(cid:16) r
log(E)(cid:19) samples. Hence, we can conclude that
ǫ(cid:17)(cid:18) ¯d +
¯d (cid:19)2! .
¯d (cid:1)2
¯d+
d
¯
d
¯
ǫ3
We can now quantify the running time of the whole algorithm.
Theorem 3.7. The accelerated measured continuous greedy algorithm makes
O(cid:18) Er2
ǫ4
log(E) log( r
ǫ )(cid:16) ¯d+
d
¯
¯d (cid:17)2(cid:19) value oracle calls, and O(cid:0) E
ǫ2 log r
ǫ(cid:1) independence oracle calls.
Proof. Considering that the number of steps of the procedure, with δ = ǫ, is 1
ǫ . The number of calls
to boths oracles is simply the number of calls by the Decreasing-Threshold procedure (Algorithm
2) in Lemma 3.6 multiplied by the number of steps in the continuous greedy algorithm (Algorithm
1).
4 Discussion and Future Work
In this paper we have presented a 1
e −ǫ-approximation algorithm for general non-negative submodu-
lar function maximisation that requires O( nr2
ǫ )) value oracle calls. This is the fastest
1
e -approximation algorithm currently available, which enables the use of general (non-monotone)
matroid-constrained submodular maximisation for many applications for which existing algorithms
were implausibly slow. We think this is of great significance because there has been a recent
surge of interest for applying submodular maximisation in fields where large problem instances are
paramount, such as Machine Learning [24, 25, 5], particularly in the field of summarisation where
non-monotone submodular functions are natural [26, 27]. Our algorithm is slower than the one
¯d (cid:1)2
ǫ4 (cid:0)
log2( n
¯d+
d
¯
presented for the monotone case in [14] by O(cid:0)r(cid:0)
values of the multilinear extension up to an additive and multiplicative bound. If we could, then
we would reduce the additional value oracle calls required to achieve an algorithm with the same
running time, we believe this might be possible.
¯d (cid:1)2(cid:1) due to the inability to sample the marginal
¯d+
d
¯
A future avenue of research would be to combine our work with the very interesting results in
[28], where an efficient algorithm is proposed to allow the trade-off of value oracle calls and
matroid independence calls for non-negative monotone submodular functions, to enable query
trade-off for general non-negative submodular functions. Another interesting path is to combine
the more continuous-like measured continuous greedy update step that we present here with the
acceleration techniques for strong submodular functions presented in [29] to produce an adaptive
step algorithm. This way, in each step we could use a large δ that extended to the boundary of
the region of validity of the set B, instead of taking a δ that is small enough to satisfy the worst
case. Another obvious improvement on the algorithms presented here would be to combine the
ideas from the Lazy Greedy Algorithm [30] to adaptively change the decrement of the threshold in
the Decreasing-Threshold procedure.
4 Discussion and Future Work
12
References
[1] A. Krause and D. Golovin, "Submodular function maximization," Tractability: Practical
Approaches to Hard Problems 3 no. 19, (2012) 8.
[2] B. Lehmann, D. Lehmann, and N. Nisan, "Combinatorial auctions with decreasing marginal
utilities," in Proceedings of the 3rd ACM conference on Electronic Commerce, pp. 18 -- 28,
ACM. 2001.
[3] S. Dughmi, T. Roughgarden, and M. Sundararajan, "Revenue submodularity," in Proceedings
of the 10th ACM conference on Electronic commerce, pp. 243 -- 252, ACM. 2009.
[4] D. Kempe, J. Kleinberg, and ´E. Tardos, "Maximizing the spread of influence through a social
network," in Proceedings of the ninth ACM SIGKDD international conference on Knowledge
discovery and data mining, pp. 137 -- 146, ACM. 2003.
[5] H. Lin and J. Bilmes, "Multi-document summarization via budgeted maximization of
submodular functions," in Human Language Technologies: The 2010 Annual Conference of
the North American Chapter of the Association for Computational Linguistics, pp. 912 -- 920,
Association for Computational Linguistics. 2010.
[6] A. Krause, J. Leskovec, C. Guestrin, J. VanBriesen, and C. Faloutsos, "Efficient sensor
placement optimization for securing large water distribution networks," Journal of Water
Resources Planning and Management 134 no. 6, (2008) 516 -- 526.
[7] J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J. VanBriesen, and N. Glance,
"Cost-effective outbreak detection in networks," in Proceedings of the 13th ACM SIGKDD
international conference on Knowledge discovery and data mining, pp. 420 -- 429, ACM. 2007.
[8] A. Schrijver, Combinatorial Optimization: Polyhedra and Efficiency. Algorithms and
Combinatorics. Springer Berlin Heidelberg, 2003.
[9] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher, "An analysis of approximations for
maximizing submodular set functionsI,"
Mathematical Programming 14 no. 1, (Dec, 1978) 265 -- 294.
[10] J. Vondrak,
"Optimal approximation for the submodular welfare problem in the value oracle model," in
Proceedings of the fourtieth annual ACM symposium on Theory of computing - STOC 08,
p. 67. ACM Press, New York, New York, USA, May, 2008.
[11] G. Calinescu, C. Chekuri, M. P´al, and J. Vondr´ak, "Maximizing a Monotone Submodular
Function Subject to a Matroid Constraint,"
SIAM Journal on Computing 40 no. 6, (Jan, 2011) 1740 -- 1766.
[12] M. Feldman, J. Naor, and R. Schwartz,
"A Unified Continuous Greedy Algorithm for Submodular Maximization," in 2011 IEEE
52nd Annual Symposium on Foundations of Computer Science, pp. 570 -- 579. IEEE, Oct,
2011.
[13] J. Vondr´ak, C. Chekuri, and R. Zenklusen, "Submodular function maximization via the
multilinear relaxation and contention resolution schemes," in Proceedings of the forty-third
annual ACM symposium on Theory of computing, pp. 783 -- 792, ACM. 2011.
4 Discussion and Future Work
13
[14] A. Badanidiyuru and J. Vondr´ak, "Fast algorithms for maximizing submodular functions," in
Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms,
pp. 1497 -- 1514, Society for Industrial and Applied Mathematics. 2014.
[15] U. Feige, "A threshold of ln n for approximating set cover," Journal of the ACM (JACM) 45
no. 4, (1998) 634 -- 652.
[16] S. O. Gharan and J. Vondr´ak, "Submodular maximization by simulated annealing," in
Proceedings of the twenty-second annual ACM-SIAM symposium on Discrete Algorithms,
pp. 1098 -- 1116, Society for Industrial and Applied Mathematics. 2011.
[17] A. Ene and H. L. Nguyen, "Constrained submodular maximization: Beyond 1/e," in
Foundations of Computer Science (FOCS), 2016 IEEE 57th Annual Symposium on,
pp. 248 -- 257, IEEE. 2016.
[18] N. Buchbinder and M. Feldman, "Constrained Submodular Maximization via a
Non-symmetric Technique," arXiv:1611.03253.
[19] W. Hoeffding, "Probability inequalities for sums of bounded random variables," Journal of
the American statistical association 58 no. 301, (1963) 13 -- 30.
[20] C. Chekuri, J. Vondr´ak, and R. Zenklusen, "Submodular Function Maximization via the
Multilinear Relaxation and Contention Resolution Schemes,"
SIAM Journal on Computing 43 no. 6, (Nov, 2014) 1831 -- 1879.
[21] A. Ageev and M. Sviridenko, "Pipage Rounding: A New Method of Constructing Algorithms
with Proven Performance Guarantee,"
Journal of Combinatorial Optimization 8 no. 3, (Sep, 2004) 307 -- 328.
[22] J. Vondr´ak, "Symmetry and approximability of submodular maximization problems," SIAM
Journal on Computing 42 no. 1, (2013) 265 -- 304.
[23] J. Vondrak, "A note on concentration of submodular functions," arXiv:1005.2791.
[24] B. Mirzasoleiman, A. Badanidiyuru, and A. Karbasi, "Fast constrained submodular
maximization: Personalized data summarization," in ICLM'16: Proceedings of the 33rd
International Conference on Machine Learning (ICML). 2016.
[25] J. Bilmes and W. Bai, "Deep Submodular Functions," arXiv:1701.08939.
[26] S. Tschiatschek, R. K. Iyer, H. Wei, and J. A. Bilmes, "Learning mixtures of submodular
functions for image collection summarization," in Advances in neural information processing
systems, pp. 1413 -- 1421. 2014.
[27] A. Dasgupta, R. Kumar, and S. Ravi, "Summarization Through Submodularity and
Dispersion.," in ACL (1), pp. 1014 -- 1022. 2013.
[28] N. Buchbinder, M. Feldman, and R. Schwartz, "Comparing Apples and Oranges: Query
Tradeoff in Submodular Maximization," arXiv:1410.0773.
[29] Z. Wang, B. Moran, X. Wang, and Q. Pan, "An accelerated continuous greedy algorithm for
maximizing strong submodular functions,"
Journal of Combinatorial Optimization 30 no. 4, (2015) 1107 -- 1124.
[30] M. Minoux, "Accelerated greedy algorithms for maximizing submodular set functions," in
Optimization Techniques, pp. 234 -- 243. Springer-Verlag, Berlin/Heidelberg, 1978.
|
1505.01140 | 1 | 1505 | 2015-05-04T18:47:58 | Depth-first search in split-by-edges trees | [
"cs.DS"
] | A layerwise search in a split-by-edges tree (as defined by Br{\ae}ndeland, 2015) of agiven graph produces a maximum independent set in exponential time. A depth-first search produces an independent set, which may or may not be a maximum, in linear time, but the worst case success rate is maybe not high enough to make it really interesting. What may make depth-first searching in split-by-edges trees interesting, though, is the pronounced oscillation of its success rate along the graph size axis. | cs.DS | cs | Depth-first search in split-by-edges trees
Asbjørn Braendeland
Abstract
A split-by-edges tree of a graph G is a binary tree T where the root = V(G), every leaf is an independent
set in G, and for every other node N in T with children L and R there is an edge uv in G such that u and v
are in N, L = N – v and R = N – u. Every maximum independent set in G is in the same layer M, which is
guarantied to be found in a layer by layer search of T. Every depth-first search in T will reach M, and may
stop there, if a leaf was hit, or will pass through, to another, smaller, leaf below. For random graphs of a
given order and increasing size the succes rates of depth-first searches in split-by-edges trees oscillate
between local minima and maxima.
The split by edges tree and the uniquified split by edges tree are defined in [1] as follows.
Definition 1: Let G be graph and let T be a binary tree of subsets of V(G). Then T is a split-by-edges
tree, or SBE-tree, of G if and only if the root of T = V(G), every leaf in T is an independent set of G,
and for every other node N in T with children L and R there is a pair of vertices {u, v} N such that
L = N – u, R = N – v, and u and v are adjacent in G.
Definition 2: An SBE-tree minus its duplicate nodes is a uniquified SBE-tree, or a USBE-tree.
By Corollary 1.1. in [1], which says that every maximal independent set of G is a leaf in every SBE-tre
of G, there is a layer MI in T that contains every maximum independent set in G[1], and the distance
from the root of T to MI is n – (G), when n = V(G), and (G) is the independence number of G; and
it follows from definition 2 that there is a corresponding layer M’I in the uniquified version T’ of T.
For random graphs a layer-by-layer search, which is akin to a breadth-first search, through T’ produ-
ces M’I in maximum 2O(0.37n) time [1]. A depth-first search through T reaches MI in n – (G) time, but
since MI usually also contains non-independent sets, it is far from guarantied that an independent set
will be hit, and if not, the search continues until a smaller independent set is found further down.
Whether the SBE-tree being searched is uniquified or not, depends on the search method. In a layer-by-
layer search the layers are created by the search procedure, which can keep duplicates out of the way.
But since the cardinalities of the nodes are unique to each layer, duplicates can only occur layerwise,
thus in a depth-first search, where only one node or two sibling nodes are examined per layer, the
existence of duplicates is not relevant—which is to say that the tree being searched is not uniquified.
The vertices of a graph G can be ordered by degree[1]. A layer-by-layer search for a maximum inde-
pendent set in a USBE-tree of G is more efficient if the order is descending and less efficient if the
order is ascending than if the order is arbitrary, due to the fact that a descending order gives a slimmer
USBE-tree. Somewhat surprisingly, the effect is the opposite for depth-first search. The real surprise is
that there is an effect at all, since in an SBE-tree the width of layer Ll = 2l all the way down to the layer
that contains the maximum independent sets, regardless of how the graph is arranged.
Independent of vertex ordering, the DFS success rate increases if we, instead of always choosing the
left branch, choose the one with the fewest edges, or even better, the most stable branch, when the
stability of a node N on n vertices is given by the formula
∑
𝑣 𝑁
𝑛
deg𝐺(𝑁)(𝑣) + 1
1
(1)
Figure 1. The curves show the number of successful MIS searches for six variations of depth-first search in the
SBE trees of random graphs on 24 vertices. 10 000 graphs where searched for each size ranging from 24 to 276.
(That there for some m are less than 10 000 different graphs, has no bearing on the shapes of the curves.)
We will return to the peculiar oscillations shown in Figure 1, presently.
The curves keep their shapes as n grows, but the success rates drop rather quickly, as shown in Figure 2.
Figure 2. The histograms show the number of successful depth-first, most-stable-branch MIS searches
in the SBE trees of six random graphs of order 24 to 64. 100 graphs where searched for each size, ranging
from n to n(n – 1)/2.The minimum success ratio drops from about 0.71, for n = 24, to about 0.25, for n = 64.
The accuracy of an MIS algorithm A for a graph G is the ratio A(G) /(G), when A(G) is the result of
applying A to G. Accuracy is subject to the same size related oscillations as success. For n = 24 to 64
the lowest accuracy of depth-first, most-stable-branch searches in SBE-trees falls from 0.917 to 0.841.
(A different, but nearly equivalent, type of measure is defined in [2].)
2
In Figure 1 and Figure 2, the numbers of successful depth-first searches for random graphs of a fixed
order and increasing size oscillate between local minima and maxima. This oscillation shows up con-
sistently in every comparable test, including tests where different random graph generators were used.
In the grids on top of Figure 3 the red lines and the blue dottet lines indicate local maxima and minima,
respectively, for depth-first searches in graphs of orders 15 to 40. The grid points can be seen as co-
ordinates (n, m) of the graph orders and sizes that give the local extrema. The grids are based on 100
runs for each (n, m)1, and the vertical lines have been slightly smoothed.
Figure 3.
For comparison Figure 3 also contains a grid of (𝑛
2) values, where each dotted line indicates one value.
Each vertical line contains a segment of the partial sums of N = (1, …, 40). Conversely, the difference
sequences of each vertical line is a segment of N. The corresponding difference sequences for the two
upper grids are close to some, but not quite equal to any segments of N.
1 The fact that there is only one graph with n vertices and n(n–1)/2 edges is immaterial, since no algorithm ever fails for this
value of m.
3
Reference
[1]
[2]
[3]
Braendeland, Asbjørn. Split-by-edges trees. 2015. arXiv:1504.07626
Halldórssonz, Magnús M. Radhakrishnan, Jaikumar. Greed is good: Approximating independent sets in
sparse and bounded-degree graphs. 1997
Knuth, Donald E. (1997), The Art Of Computer Programming Vol 1. 3rd ed, Boston: Addison-Wesley,
ISBN 0-201-89683-4.
4
|
1304.7055 | 1 | 1304 | 2013-04-26T02:28:22 | An LP-based 3/2-approximation algorithm for the graphic s-t path TSP | [
"cs.DS"
] | We design a new LP-based algorithm for the graphic $s$-$t$ path Traveling Salesman Problem (TSP), which achieves the best approximation factor of 1.5. The algorithm is based on the idea of narrow cuts due to An, Kleinberg, and Shmoys. It partly answers an open question of Seb\H{o}. | cs.DS | cs |
An LP-based 3
2-approximation algorithm for the graphic s-t path
TSP
Zhihan Gao∗
Abstract
We design a new LP-based algorithm for the graphic s-t path Traveling Salesman Problem
(TSP), which achieves the best approximation factor of 1.5. The algorithm is based on the idea
of narrow cuts due to An, Kleinberg, and Shmoys. It partly answers an open question of Sebo.
Keywords:
approximation algorithms, linear programming, s-t path TSP.
1
Introduction
The metric Traveling Salesman Problem (TSP) is one of the most well-known problems in the area
of combinatorial optimization. For the metric TSP, Christofides [4] presented an algorithm that
achieves an approximation guarantee of 3
2 . Hoogeveen [8] extended the algorithm to the metric
s-t path TSP, and proved an approximation guarantee of 5
3 . This had been the best approxima-
tion factor for decades until the recent paper [1, An, Kleinberg, and Shmoys] improved on the 5
3
approximation guarantee and presented an algorithm that achieves an approximation guarantee of
1+√5
2 ≈ 1.61803. Most recently, [11, Sebo] further improved the approximation factor to 1.6.
For the graphic s-t path TSP, a special case of the metric s-t path TSP, [2, An and Shmoys]
provided a sightly improved performance guarantee of ( 5
3 − ǫ). The paper [9, Momke and Svensson]
gave a 1.586-approximation algorithm for the graphic s-t path TSP. [10, Mucha] improved the
analysis of [9] and obtained a 19
12 + ǫ ≈ 1.58333 + ǫ approximation guarantee for any ǫ > 0 for the
graphic s-t path TSP. Recently, [12, Sebo and Vygen] gave the first 1.5-approximation algorithm
for the graphic s-t path TSP. Their algorithm and its analysis are sophisticated, and are based on
ear decomposition. The algorithm applies both local and global optimization to the ears.
In this paper, we present a new 1.5-approximation algorithm for the graphic s-t path TSP.
Compared with the algorithm from [12, Sebo and Vygen], our algorithm and its analysis are much
simpler. The notion of narrow cuts for s-t path TSP was introduced by [1, An, Kleinberg, and
Shmoys]. Our algorithm is based on this idea. In [11], Sebo posed an open question on applying the
"Best of Many Christofides" algorithm in [1] to achieve the best approximation guarantees known
for the graphic special cases of TSP and its variants. Although our algorithm seems different from
the "Best of Many Christofides" algorithm, they share the idea of narrow cuts. From this point of
view, our algorithm answers this question partly for the graphic s-t path TSP. The key point of our
algorithm is to find a minimal spanning tree that intersects every narrow cut in an odd number of
edges. Such a tree guarantees that the minimum size of the edges fixing the wrong degree vertices
∗([email protected]) Dept. of Comb. & Opt., University of Waterloo, Waterloo, Ontario N2L3G1, Canada.
1
is at most half of the optimal value of the linear programming relaxation. Finally, the union of the
spanning tree and the fixing edges gives us the 1.5-approximation guarantee.
2 Preliminaries
Let G = (V, E) be a connected graph with unit cost on each edge. Let s, t be two given vertices in
G. Consider the metric completion (G′, c′) of G where c′ is the cost function on each edge e = (u, v)
in G′ such that c′e is the minimal cost of any u-v path in G. The graphic s-t path TSP is to find a
minimum cost Hamiltonian path from s to t in G′ with edge costs c′. Denote the cost of this path
by OP T (G).
For any vertex subset φ ( S ( V , we define δG(S) = {(u, v) ∈ E : u ∈ S, v /∈ S}. If there is
no ambiguity, we use δ(S) for short. In particular, if S = {v}, we use δ(v) instead of δ({v}). Let
W = {W1, W2, . . . , Wl} be a partition of vertex set V . Define δ(W) = ∪1≤i≤lδ(Wi). Let x ∈ RE.
For any F ⊆ E, we define x(F ) = Pe∈F xe. Let 2G be the graph obtained from G by doubling
every edge of G. The graphic s-t path TSP of G is equivalent to finding a minimum-size trail in
2G from s to t visiting every vertex at least once (multiple visits are allowed). That is to find a
minimum-size connected spanning subgraph of 2G with {s, t} as the odd-degree vertex set. The
following linear program (LP) on the original graph G is a relaxation of the graphic s-t path TSP:
(L.P.1) minimize : Pe∈E xe
subject to : x(δ(W)) ≥ W − 1 ∀ partition W of V
x(δ(S)) ≥ 2
2 ≥ xe ≥ 0
∀∅ ( S ( V, S ∩ {s, t} even
∀e ∈ E
Let x∗ be an optimal solution of (L.P.1). Note that (L.P.1) can be solved in polynomial time
via the ellipsoid method [7]. We know that Pe∈E x∗e ≤ OP T (G). Let S ⊆ V . If S ∩ {s, t} = 1,
we call S an s-t cut. Furthermore, if x∗(δ(S)) < 2, we call S a narrow cut.
Lemma 2.1 [1, Lemma 1] Let S1, S2 ⊆ V be two distinct narrow cuts such that s ∈ S1 and s ∈ S2.
Then S1 ( S2 or S2 ( S1.
Hence, we know that the set of narrow cuts containing s forms a nested family. Let S1, S2, . . . , Sk
be all the narrow cuts containing s such that s ∈ S1 ( S2 ( S3 · · · ( Sk ( V . Define Li = Si\Si−1
for i = 1, 2, . . . , k, k + 1 where S0 = φ and Sk+1 = V . Note that each Li is nonempty and
∪1≤i≤k+1Li = V .
Let T be a nonempty subset of V with T even. For F ⊆ E, if the set of odd-degree vertices
of graph (V, F ) is T , then we call F a T -join. Note that if G is connected, then a T -join always
exists. For any S ⊆ V , if S ∩ T is odd, we call it T -odd cut. The following LP formulates the
problem of finding a T -join of minimum size:
(L.P.2) minimize : Pe∈E xe
subject to :
x(δ(S)) ≥ 1 ∀ T -odd S
xe ≥ 0 ∀e ∈ E
Lemma 2.2 [6] The optimal value of (L.P.2) is the same as the minimum size of a T -join.
Let F ⊆ E. For any v ∈ V , we call v a wrong degree vertex with respect to F if
δ(v) ∩ F is (even if v ∈ {s, t}
if v /∈ {s, t}.
odd
(1)
2
We use the next lemma through the rest of the paper.
Lemma 2.3 [3, Lemma 1] Let G = (V, E) be a graph, let s, t be two vertices of G, let F be a set
of edges of G, and let T be the set of wrong-degree vertices with respect to F . Then, for any S ⊆ V ,
if S is T -odd and also satisfies S ∩ {s, t} = 1, then δ(S) ∩ F is even.
3 LP-based 3
2-approximation algorithm
In this section, we give an LP-based 3
the algorithm, we need some lemmas.
2 -approximation algorithm for s-t path TSP. Before stating
Lemma 3.1 There is a polynomial-time combinatorial algorithm to find all narrow cuts S1, S2, . . . , Sk.
Proof. Compute the Gomory-Hu tree for the terminal vertex set V with respect to the capacity
x∗ (See [5, Section 3.5.2]). After that, for each edge of the s-t path in the Gomory-Hu tree, check
the corresponding cut. We claim that each such cut with x∗ capacity less than 2 is a narrow
cut, and there are no other narrow cuts. The correctness of this claim follows from the following
observation: For any u ∈ Li, v ∈ Li+1, the narrow cut Si is the unique minimum u-v cut, and
furthermore, Si is also a s-t cut.
(cid:3)
Let H be the support graph of x∗. For any L ⊆ V (H), the subgraph of H induced by L is
denoted by H(L).
Lemma 3.2 For 1 ≤ p ≤ q ≤ k + 1, H(∪p≤i≤qLi) is connected.
Consider the graph H which is the support graph of x∗. Note that x∗(δH (S)) =
Proof.
x∗(δG(S)) for any φ ⊂ S ⊂ V . In this proof, the notation refers to H, e.g., δ(S) means δH(S). Let
L = ∪p≤i≤qLi. We divide the proof into several cases:
Case 1: p = 1 and q = k + 1, i.e., H = H(L). The partition cut constraint in (L.P.1) implies that
H is connected.
Case 2: p = 1 and q < k + 1. Suppose H(L) is not connected. Then, there exist two nonempty
vertex sets U1 and U2 such that U1, U2 is a partition of L and there exists no edge between U1 and U2
in H. Without loss of generality, we can assume that s ∈ U1. By the constraints of (L.P.1), we have
x∗(δ(U1)) ≥ 1 and x∗(δ(U2)) ≥ 2. However, L = Sq is a narrow cut, which implies x∗(δ(L)) < 2.
Note that δ(U1) ∩ δ(U2) = φ and δ(L) = δ(U1) ∪ δ(U2). So, 2 > x∗(δ(L)) = x∗(δ(U1)) + x∗(δ(U2)) ≥
1 + 2 = 3. This is a contradiction.
Case 3: p > 1 and q = k + 1. By the symmetry of s and t, it is the same as Case 2.
Case 4: p > 1 and q < k + 1. Suppose H(L) is not connected. Then, similarly there exist two
nonempty vertex sets U1 and U2 such that U1, U2 is a partition of L and there exists no edge
between U1 and U2 in H. In this case, by the constraints of (L.P.1), we have x∗(δ(U1)) ≥ 2 and
x∗(δ(U2)) ≥ 2. Let Y1 = ∪1≤i≤qLi and Y2 = ∪p≤i≤k+1Li. Note that Y1 and Y2 are two narrow cuts.
Also, δ(U1) ∪ δ(U2) ⊆ δ(Y1) ∪ δ(Y2). Note that δ(U1) ∩ δ(U2) = φ by the definition of U1 and U2.
Thus, 4 > x∗(δ(Y1)) + x∗(δ(Y2)) ≥ x∗(δ(Y1) ∪ δ(Y2)) ≥ x∗(δ(U1) ∪ δ(U2)) = x∗(δ(U1)) + x∗(δ(U2)) ≥
2 + 2 = 4. This is a contradiction.
(cid:3)
Corollary 3.3 For every 1 ≤ i ≤ k + 1, H(Li) is connected, and moreover there exists an edge
connecting Li and Li+1 in H.
3
Algorithm 1 LP-based approximation for the graphic s-t path TSP
Step 1. Find an optimal solution x∗ of (L.P.1) and construct the support graph H of x∗.
Step 2. Find the narrow cuts S1, S2, . . . , Sk containing s, and get the corresponding sets
L1, L2, . . . , Lk+1 (recall: Li = Si\Si−1 where S0 = φ and Sk+1 = V ). If no narrow cuts exist, take
J as a spanning tree in G and go to Step 6.
Step 3. For 1 ≤ i ≤ k + 1, find a spanning tree Ji on H(Li).
Step 4. Take an edge ei from H connecting Li to Li+1 for 1 ≤ i ≤ k. Let Eb = ∪1≤i≤k{ei}.
Step 5. Construct a spanning tree J = (∪1≤i≤k+1Ji) ∪ Eb.
Step 6. Let T be the wrong degree vertex set of J. Find the minimum size T -join F in G.
Step 7. Output J ∪F (disjoint union of edge sets in 2G).
Lemma 3.1 provides a polynomial algorithm for Step 2, and Corollary 3.3 guarantees that Step
3 and Step 4 are feasible. Thus, the LP-based algorithm runs in polynomial time.
Lemma 3.4 For F in the LP-based algorithm, we have
F ≤
x∗e.
1
2Xe∈E
Proof. Firstly, we claim x∗(δ(S)) ≥ 2 for every T -odd cut where T is the wrong degree vertex
set of J in the LP-based algorithm. Let S be a T -odd cut. There are two cases to be considered.
Case 1: S is not an s-t cut. Then, by the constraint of (L.P.1), we have x∗(δ(S)) ≥ 2
Case 2: S is an s-t cut. If there exist no narrow cuts, then clearly x∗(δ(S)) ≥ 2. Otherwise, for
any narrow cut S′, we have J ∩ δ(S′) = 1 by Step 4 of the algorithm. However, by Lemma 2.3,
we have J ∩ δ(S) is even. This means S is not a narrow cut. Thus, x∗(δ(S)) ≥ 2.
By the claim, we know 1
of (L.P.2). By Lemma 2.2, we have F ≤ 1
2 x∗(δ(S)) ≥ 1 for every T -odd cut S. This implies 1
2 x∗ is a feasible solution
(cid:3)
2Pe∈E x∗e. This completes the proof.
Remark 3.5 In fact, if we can find a spanning tree J such that J ∩ δ(S) is odd for each narrow
cut S, then we can find an edge set F to correct the wrong degree vertices in J such that F ≤
1
2Pe∈E x∗e. This also holds for the (general) metric s-t path TSP, i.e., for metric costs, if we can
find such a spanning tree J, then the minimum cost of the edges fixing the wrong degree vertices in
J is at most half of the cost of the optimal solution of LP.
Theorem 3.6 The LP-based algorithm is a 3
2 -approximation for the graphic s-t path TSP.
Pe∈E x∗e ≤ OP T (G). Also note that F ≤ 1
Proof. Note that J is a spanning tree of G. We consider J as an edge set. So, J = V − 1 ≤
2 OP T (G). Since J ∪F is a connected
spanning subgraph of 2G with {s, t} as the odd-degree vertex set, this gives a s-t Hamiltonian path
on the metric completion of G with cost at most J + F . Therefore, the LP-based algorithm is a
3
2 -approximation algorithm.
(cid:3)
2Pe∈E x∗e ≤ 1
4
Remark 3.7 By the proof of Theorem 3.6, we can obtain an upper bound 3
2 for the integrality
ratio of the (L.P.1). Furthermore, this also implies that the integrality ratio of the path-variant
Held-Karp relaxation (See [1]) is at most 3
2 when restricted to graphic metric. Note that [1, Figure
1(b)] presented an example with graphic metric to show the lower bound 3
2 for the integrality ratio
of the path-variant Held-Karp relaxation. Hence, from this point of view, our algorithm achieves
the best possible approximation guarantee that the LP-based algorithms can get for the graphic s-t
path TSP.
Acknowledgements. The author is grateful to Joseph Cheriyan for stimulating discussions and
indispensable help, and to Zachary Friggstad, Laura Sanit`a and Chaitanya Swamy for their useful
comments.
References
[1] H.-C. An, R. Kleinberg, and D. B. Shmoys. Improving Christofides' algorithm for the s-t path
TSP. In STOC, pages 875 -- 886, 2012.
[2] H.-C. An and D. B. Shmoys. LP-based approximation algorithms for Traveling Salesman Path
Problems. CoRR, abs/1105.2391, 2011.
[3] J. Cheriyan, Z. Friggstad, and Z. Gao. Approximating minimum-cost connected T-joins. In
APPROX-RANDOM, pages 110 -- 121, 2012.
[4] N. Christofides. Worst-case analysis of a new heuristic for the travelling salesman problem.
Technical Report 388, Graduate School of Industrial Administration, CMU, 1976.
[5] W. J. Cook, W. H. Cunningham, W. R. Pulleyblank, and A. Schrijver. Combinatorial Opti-
mization. Wiley-Interscience, 1998.
[6] J. Edmonds and E. L. Johnson. Matching: A well-solved class of integer linear programs. In
Combinatorial Optimization, pages 27 -- 30, 2001.
[7] M. Grotschel, L. Lov´asz, and A. Schrijver. The ellipsoid method and its consequences in
combinatorial optimization. Combinatorica, 1(2):169 -- 197, 1981.
[8] J. Hoogeveen. Analysis of Christofides' heuristic: Some paths are more difficult than cycles.
Operations Research Letters, 10:291 -- 295, 1991.
[9] T. Momke and O. Svensson. Approximating graphic TSP by matchings.
In FOCS, pages
560 -- 569, 2011.
[10] M. Mucha. 13/9-approximation for graphic TSP. In STACS, pages 30 -- 41, 2012.
[11] A. Sebo. Eight-fifth approximation for TSP paths. CoRR, abs/1209.3523, 2012. To appear in
IPCO 2013.
[12] A. Sebo and J. Vygen. Shorter tours by nicer ears: 7/5-approximation for graphic TSP, 3/2
for the path version, and 4/3 for two-edge-connected subgraphs. CoRR, abs/1201.1870, 2012.
5
|
1602.08820 | 1 | 1602 | 2016-02-29T04:38:50 | Compressing Graphs and Indexes with Recursive Graph Bisection | [
"cs.DS",
"cs.SI"
] | Graph reordering is a powerful technique to increase the locality of the representations of graphs, which can be helpful in several applications. We study how the technique can be used to improve compression of graphs and inverted indexes.
We extend the recent theoretical model of Chierichetti et al. (KDD 2009) for graph compression, and show how it can be employed for compression-friendly reordering of social networks and web graphs and for assigning document identifiers in inverted indexes. We design and implement a novel theoretically sound reordering algorithm that is based on recursive graph bisection.
Our experiments show a significant improvement of the compression rate of graph and indexes over existing heuristics. The new method is relatively simple and allows efficient parallel and distributed implementations, which is demonstrated on graphs with billions of vertices and hundreds of billions of edges. | cs.DS | cs |
Compressing Graphs and Indexes with Recursive Graph Bisection
Laxman Dhulipala1, Igor Kabiljo2, Brian Karrer2, Giuseppe Ottaviano2, Sergey Pupyrev2,
and Alon Shalita2
1Carnegie Mellon University
2Facebook
Abstract
Graph reordering is a powerful technique to increase the locality of the representations of
graphs, which can be helpful in several applications. We study how the technique can be used
to improve compression of graphs and inverted indexes.
We extend the recent theoretical model of Chierichetti et al. (KDD 2009) for graph compres-
sion, and show how it can be employed for compression-friendly reordering of social networks and
web graphs and for assigning document identifiers in inverted indexes. We design and implement
a novel theoretically sound reordering algorithm that is based on recursive graph bisection.
Our experiments show a significant improvement of the compression rate of graph and indexes
over existing heuristics. The new method is relatively simple and allows efficient parallel and
distributed implementations, which is demonstrated on graphs with billions of vertices and
hundreds of billions of edges.
1
Introduction
Many real-world systems and applications use in-memory representation of indexes for serving ad-
jacency information in a graph. A popular example is social networks in which the list of friends is
stored for every user. Another example is an inverted index for a collection of documents that stores,
for every term, the list of documents where the term occurs. Maintaining these indexes requires a
compact, yet efficient, representation of graphs.
How to represent and compress such information? Many techniques for graph and index com-
pression have been studied in the literature [24, 37]. Most techniques first sort vertex identifiers in
an adjacency list, and then replace the identifiers (except the first) with differences between con-
secutive ones. The resulting gaps are encoded using some integer compression algorithm. Note that
using gaps instead of original identifiers decreases the values needed to be compressed and results in
a higher compression ratio. We stress that the success of applying a particular encoding algorithm
strongly depends on the distribution of gaps in an adjacency list: a sequence of small and regular
gaps is more compressible than a sequence of large and random ones.
This observation has motivated the approach of assigning identifiers in a way that optimizes
compression. Graph reordering has been successfully applied for social networks [8, 13]. In that
scenario, placing similar social actors nearby in the resulting order yields a significant compression
improvement. Similarly, lexicographic locality is utilized for compressing the Web graph: when pages
are ordered by URL, proximal pages have similar sets of neighbors, which results in an increased
1
compression ratio of the graph, when compared with the compression obtained using the original
graph [9,29]. In the context of index compression, the corresponding approach is called the document
identifier assignment problem. Prior work shows that for many collections, index compression can
be significantly improved by assigning close identifiers to similar documents [6, 7, 15, 32, 34].
In this paper, we study the problem of finding the best "compression-friendly" order for a graph
or an inverted index. While graph reordering and document identifier assignment are often studied
independently, we propose a unified model that generalizes both of the problems. Although a number
of heuristics for the problems exists, none of them provides any guarantees on the resulting quality.
In contrast, our algorithm is inspired by a theoretical approach with provable guarantees on the
final quality, and it is designed to directly optimize the resulting compression ratio. Our main
contributions are the following.
• We analyze and extend the formal model of graph compression suggested in [13]; the new
model is suitable for both graph reordering and document identifier assignment problems. We
show that the underlying optimization problem is NP-hard (thus, resolving an open question
stated in [13]), and suggest an efficient approach for solving the problem, which is based on
approximation algorithms for graph reordering.
• Based on the theoretical result, we develop a practical algorithm for constructing compression-
friendly vertex orders. The algorithm uses recursive graph bisection as a subroutine and tries
to optimize a desired objective at every recursion step. Our objective corresponds to the size
of the graph compressed using delta-encoding. The algorithm is surprisingly simple, which
allows for efficient parallel and distributed implementations.
• Finally, we perform an extensive set of experiments on a collection of large real-world graphs,
including social networks, web graphs, and search indexes. The experiments indicate that
our new method outperforms the state-of-the-art graph reordering techniques, improving the
resulting compression ratio. Our implementation is highly scalable and is able to process a
billion-vertex graph in a few hours.
The paper is organized as follows. We first discuss existing approaches for graph reordering,
assigning document identifiers, and the most popular encoding schemes for graph and index repre-
sentation (Section 2). Then we consider algorithmic aspects of the underlying optimization problem.
We analyze the models for graph compression suggested by Chierichetti et al. [13] and suggest our
generalization in Section 3.1. Next, in Section 3.2, we examine existing theoretical techniques for the
graph reordering problem and use the ideas to design a practical algorithm. A detailed description
of the algorithm along with the implementation details is presented in Section 4, which is followed
by experimental Section 5. We conclude the paper with the most promising future directions in
Section 6.
2 Related Work
There exists a rich literature on graph and index compression, that can be roughly divided into three
categories: (1) structural approaches that find and merge repeating graph patterns (e.g., cliques),
(2) encoding adjacency data represented by a list of integers given some vertex/document order, and
(3) finding a suitable order of graph vertices. Our focus is on the ordering techniques. We discuss
the existing approaches for graph reordering, followed by an overview of techniques for document
identifier assignment. Since many integer encoding algorithms can benefit from such a reordering,
we also outline the most popular encoding schemes.
2
Graph Reordering Among the first approaches for compressing large-scale graphs is a work by
Boldi and Vigna [9], who compress web graphs using a lexicographical order of the URLs. Their
compression method relies on two properties: locality (most links lead to pages within the same
host) and similarity (pages on the same host often share the same links). Later Apostolico and
Drovandi [2] suggest one of the first ways to compress a graph assuming no a priori knowledge of
the graph. The technique is based on a breadth-first traversal of the graph vertices and achieves a
better compression rate using an entropy-based encoding.
Chierichetti et al. [13] consider the theoretical aspect of the reordering problem motivated by
compressing social networks. They develop a simple but practical heuristic for the problem, called
shingle ordering. The heuristic is based on obtaining a fingerprint of the neighbors of a vertex and
positioning vertices with identical fingerprints close to each other. If the fingerprint can capture
locality and similarity of the vertices, then it can be effective for compression. This approach is also
called minwise hashing and was originally applied by Broder [10] for finding duplicate web pages.
Boldi et al. [8] suggest a reordering algorithm, called Layered Label Propagation, to compress
social networks. The algorithm is built on a scalable graph clustering technique by label propa-
gation [28]. The idea is to assign a label for every vertex of a graph based on the labels of its
neighbors. The process is executed in rounds until no more updates take place. Since the standard
label propagation described in [28] tends to produce a giant cluster, the authors of [8] construct a
hierarchy of clusters. The vertices of the same cluster are then placed together in the final order.
The three-step multiscale paradigm is often employed for the graph ordering problems. First,
a sequence of coarsened graphs, each approximating the original graph but having a smaller size,
is created. Then the problem is solved on the coarsest level by an exhaustive search. Finally, the
process is reverted by an uncoarsening procedure so that a solution for every graph in the sequence
is based on the solution for a previous smaller graph. Safro and Temkin [31] employ the algebraic
multigrid methodology in which the sequence of coarsened graphs is constructed using a projection
of graph Laplacians into a lower-dimensional space.
Spectral methods have also been successfully applied to graph ordering problems [20]. Sequencing
the vertices is done by sorting them according to corresponding elements of the second smallest
eigenvector of graph Laplacian (also called the Fiedler vector). It is known that the order yields the
best non-trivial solution to a relaxation of the quadratic graph ordering problem, and hence, it is a
good heuristic for computing linear arrangements.
Recently Kang and Faloutsos [23] present another technique, called SlashBurn. Their method
constructs a permutation of graph vertices so that its adjacency matrix is comprised of a few nonzero
blocks. Such dense blocks are easier to encode, which is beneficial for compression.
In our experiments, we compare our new algorithm with all of the methods, which are either
easy to implement, or come with the source code provided by the authors.
Document Identifier Assignment Several papers study how to assign document identifiers in
a document collection for better compression of an inverted index. A popular idea is to perform
a clustering on the collection and assign close identifiers to similar documents. Shieh et al. [32]
propose a reassignment heuristic motivated by the maximum travelling salesman problem (TSP).
The heuristic computes a pairwise similarity between every pairs of documents (proportional to
the number of shared terms), and then finds the longest path traversing the documents in the
graph. An alternative algorithm calculating cosine similarities between documents is suggested by
Blandford and Blelloch [7]. Both methods are computationally expensive and are limited to fairly
3
small datasets. The similarity-based approach is later improved by Blanco and Barreiro [6] and
by Ding et al. [15], who make it scalable by reducing the size of the similarity graph, respectively
through dimensionality reduction and locality sensitive hashing.
The approach by Silvestri [34] simply sorts the collection of web pages by their URLs and then
assigns document identifiers according to the order. The method performs very well in practice and
is highly scalable. This technique however does not generalize to document collections that do not
have URL-like identifiers.
Encoding Schemes Our algorithm is not tailored specifically for an encoding scheme; any method
that can take advantage of lists with higher local density (clustering) should benefit from the
reordering. For our experiment we choose a few encoding schemes that should be representative
of the state-of-the-art.
Most graph compression schemes build on delta-encoding, that is, sorting the adjacency lists
(posting lists in the inverted indexes case) so that the gaps between consecutive elements are positive,
and then encoding these gaps using a variable-length integer code. The WebGraph framework adds
the ability to copy portions of the adjacency lists from other vertices, and has special cases for runs
of consecutive integers. Introduced in 2004 by Boldi and Vigna [9], it is still widely used to compress
web graphs and social networks.
Inverted indexes are usually compressed with more specialized techniques in order to enable
fast skipping, which enables efficient list intersection. We perform our experiments with Parti-
tioned Elias-Fano and Binary Interpolative Coding. The former, introduced by Ottaviano and Ven-
turini [26], provides one of the best compromise between decoding speed and compression ratio. The
latter, introduced by Moffat and Stuiver [25], has the highest compression ratio in the literature,
with the trade-off of slower decoding by several times. Both techniques directly encode monotone
lists, without going through delta-encoding.
3 Algorithmic Aspects
Graph reordering is a combinatorial optimization problem with a goal to find a linear layout of
an input graph so that a certain objective function (referred to as a cost function or just cost) is
optimized. A linear layout of a graph G = (V, E) with n = V vertices is a bijection π : V →
{1, . . . , n}. A layout is also called an order, an arrangement, or a numbering of the vertices. In
practice it is desirable that "similar" vertices of the graph are "close" in π. This leads to a number
of problems that we define next.
The minimum linear arrangement (MLA) problem is to find a layout π so that
(cid:88)
π(u) − π(v)
(u,v)∈E
is minimized. This is a classical NP-hard problem [18], even when restricted to certain graph classes.
The problem is APX-hard under Unique Games Conjecture [14], that is, it is unlikely that an efficient
approximation algorithm exists. Charikar et al. [11] suggested the best currently known algorithm
with approximation factor O(√log n log log n); see [27] for a survey of results on MLA.
A closely related problem is minimum logarithmic arrangement (MLogA) in which the goal is
4
to minimize
(cid:88)
(u,v)∈E
log π(u) − π(v).
Here and in the following we denote log(x) = 1 + (cid:98)log2(x)(cid:99), that is, the number of bits needed to
represent an integer x. The problem is also NP-hard, and one can show that the optimal solutions
of MLA and MLogA are quite different on some graphs [13]. In practice a graph is represented
in memory as an adjacency list using an encoding scheme; hence, the gaps induced by consecutive
neighbors of a vertex are important for compression. For this reason, the minimum logarithmic
cost compressing the list out(v) under π is related to fπ(v, out(v)) = (cid:80)k−1
gap arrangement (MLogGapA) problem is introduced. For a vertex v ∈ V of degree k and an
order π, consider the neighbors out(v) = (v1, . . . , vk) of v such that π(v1) < ··· < π(vk). Then the
i=1 log π(vi+1) − π(vi).
MLogGapA consists in finding an order π, which minimizes
(cid:88)
v∈V
fπ(v, out(v)).
To the best of our knowledge, MLogA and MLogGapA are introduced quite recently by Chierichetti
et al. [13]. They show that MLogA is NP-hard but left the computational complexity of MLogGapA
open. Since the latter problem is arguably more important for applications, we address the open
question of complexity of the problem.
Theorem 1. MLogGapA is NP-hard.
Proof. We prove the theorem by using the hardness of MLogA, which is known to be NP-hard [13].
Let G = (V, E) be an instance of MLogA. We build a bipartite graph G(cid:48) = (V (cid:48), E(cid:48)) by splitting
every edge of E by a degree-2 vertex. Formally, we add E new vertices so that V (cid:48) = V ∪ U, where
V = {v1, . . . , vn} and U = {u1, . . . , um}. For every edge (a, b) ∈ E, we have two edges in E(cid:48), that
is, (a, ui) and (b, ui) for some 1 ≤ i ≤ m. Next we show that an optimal solution for MLogGapA
on G(cid:48) yields an optimal solution for MLogA on G, which proves the claim of the theorem.
Let R be an optimal order of V (cid:48) for MLogGapA. Observe that without loss of generality, the
vertices of V and U are separated in R, that is, R = (vi1, . . . , vin, uj1, . . . , ujm). Otherwise, the
vertices can be reordered so that the total objective is not increased. To this end, we "move" all
the vertices of V to the left of R by preserving their relative order. It is easy to see that the gaps
between vertices of V and the gaps between vertices of U can only decrease.
Now the cost of MLogGapA is(cid:88)
v∈V
(cid:88)
u∈U
fπu(v, out(v)) +
fπv (u, out(u)),
where πu = (uj1, . . . , ujm) and πv = (vi1, . . . , vin). Notice that the second term of the sum depends
only on the order πv of the vertices in V , and it equals to the cost of MLogA for graph G. Since
R is optimal for MLogGapA, the order πv = (vi1, . . . , vin) is also optimal for MLogA.
Most of the previous works consider the MLogA problem for graph compression, and the
algorithms are not directly suitable for index compression. Contrarily, an inverted index is generally
represented by a directed graph (with edges from terms to documents), which is not captured by
the MLogGapA problem. In the following, we suggest a model, which generalizes both MLogA
and MLogGapA and better expresses graph and index compression.
5
3.1 Model for Graph and Index Compression
Intuitively, our new model is a bipartite graph comprising of query and data vertices. A query vertex
might correspond to an actor in a social network or to a term in an inverted index. Data vertices
are an actor's friends or documents containing the term, respectively. The goal is to find a layout
of data vertices.
Formally, let G = (Q ∪ D, E) be an undirected unweighted bipartite graph with disjoint sets of
vertices Q and D. We denote D = n and E = m. The goal is to find a permutation, π, of data
vertices, D, so that the following objective is minimized:
(cid:88)
degq −1(cid:88)
q∈Q
i=1
log(π(ui+1) − π(ui)),
where degq is the degree of query vertex q ∈ Q, and q's neighbors are {u1, . . . , udegq} with π(u1) <
··· < π(udegq ). Note that the objective is closely related to minimizing the number of bits needed
to store a graph or an index represented using the delta-encoding scheme. We call the optimiza-
tion problem bipartite minimum logarithmic arrangement (BiMLogA), and the corresponding cost
averaged over the number of gaps LogGap.
Note that BiMLogA is different from MLogGapA in that the latter does not differentiate
between data and query vertices (that is, every vertex is query and data in MLogGapA), which is
unrealistic in some applications. It is easy to see that the new problem generalizes both MLogA
and MLogGapA: to model MLogA, we add a query vertex for every edge of the input graph, as
in the proof of Theorem 1; to model MLogGapA, we add a query for every vertex of the input
graph; see Figure 1. Moreover, the new approach can be naturally applied for compressing directed
graphs; to this end, we only consider gaps induced by outgoing edges of a vertex. Clearly, given an
algorithm for BiMLogA, we can easily solve both MLogA and MLogGapA. Therefore, we focus
on this new problem in the next sections.
(a) Original graph
(b) MLogA
(c) MLogGapA
Figure 1: Modeling of MLogA and MLogGapA with a bipartite graph with query (red) and
data (blue) vertices.
How can one solve the above ordering problems? Next we discuss the existing theoretical ap-
proaches for solving graph ordering problems. We focus on approximation algorithms, that is, effi-
cient algorithms for NP-hard problems that produce sub-optimal solutions with provable quality.
3.2 Approximation Algorithms
To the best of our knowledge, no approximation algorithms exist for the new variants of the graph
ordering problem. However, a simple observation shows that every algorithm has approximation
factor O(log n). Note that the cost of a gap between u ∈ D and v ∈ D in BiMLogA cannot exceed
6
2134213412132334213410203040 log n, as π(v) − π(u) ≤ n for every permutation π. On the other hand, the cost of a gap is at least
1. Therefore, an arbitrary order of vertices yields a solution with the cost, which is at most log n
times greater than the optimum.
In contrast, the well-studied MLA does not admit such a simple approximation and requires more
involved algorithms. We observe that most of the existing algorithms for MLA and related ordering
problems employ the divide-and-conquer approach; see Algorithm 1. Such algorithms partition the
vertex set into two sets of roughly equal size, compute recursively an order of each part, and "glue"
the orderings of the parts together. The crucial part is graph bisection or more generally balanced
graph partitioning, if the graph is split into more than two parts.
Input: graph G
1. Find a bisection (G1, G2) of G;
2. Recursively find linear arrangements for G1 and G2;
3. Concatenate the resulting orderings;
Algorithm 1: Graph Reordering using Graph Bisection
The first non-trivial approximation algorithm for MLA follows the above approach. Hancen [19]
proves that Algorithm 1 yields an O(α log n)-approximation for MLA, where α indicates how close
is the solution of the first step (bisection of G) to the optimum. Later, Charikar et al. [12] shows that
a tighter analysis is possible, and the algorithm is in fact O(α)-approximation for d-dimensional
MLA. Currently, α = O(√log n) is the best known bound [3]. Subsequently, the idea of Algorithm 1
was employed by Even et al. [16], Rao and Richa [30], and Charikar et al. [11] for composing
approximation algorithms for MLA. The techniques use the recursive divide-and-conquer approach
and utilize a spreading metric by solving a linear program with an exponential number of constraints.
Inspired by the algorithms, we design a practical approach for the BiMLogA problem. While
solving a linear program is not feasible for large graphs, we utilize recursive graph partitioning in
designing the algorithm. Next we describe all the steps and provide implementation-specific details.
4 Compression-Friendly Graph Reordering
Assume that the input is an undirected bipartite graph G = (Q∪D, E), and the goal is to compute
an order of D. On a high level, our algorithm is quite simple; see Algorithm 1.
The reordering method is based on the graph bisection problem, which asks for a partition of
graph vertices into two sets of equal cardinality so as to minimize an objective function. Given an
input graph G with D = n, we apply the bisection algorithm to obtain two disjoint sets V1, V2 ⊆ D
with V1 = (cid:98)n/2(cid:99) and V2 = (cid:100)n/2(cid:101). We shall lay out V1 on the set {1, . . . ,(cid:98)n/2(cid:99)} and lay out V2
on the set {(cid:100)n/2(cid:101), . . . , n}. Thus, we have divided the problem into two problems of half the size,
and we recursively compute good layouts for the graphs induced by V1 and V2, which we call G1
and G2, respectively. Of course, when there is only one vertex in G, the order is trivial.
How to bisect the vertices of the graph? We use a graph bisection method, similar to the popular
Kernighan-Lin heuristic [21]; see Algorithm 2. Initially we split D into two sets, V1 and V2, and define
a computational cost of the partition, which indicates how "compression-friendly" the partition is.
Next we exchange pairs of vertices in V1 and V2 trying to improve the cost. To this end we compute,
for every vertex v ∈ D, the move gain, that is, the difference of the cost after moving v from its
7
current set to another one. Then the vertices of V1 (V2) are sorted in the decreasing order of the
gains to produce list S1 (S2). Finally, we traverse the lists S1 and S2 in the order and exchange the
pairs of vertices, if the sum of their move gains is positive. Note that unlike classical graph bisection
heuristics [17,21], we do not update move gains after every swap. The process is repeated until the
convergence criterion is met (no swapped vertices) or the maximum number of iterations is reached.
Input : graph G = (Q ∪ D, E)
Output: graphs G1 = (Q ∪ V1, E1), G2 = (Q ∪ V2, E2)
determine an initial partition of D into V1 and V2;
repeat
gains[v] ← ComputeM oveGain(v)
for v ∈ D do
S1 ← sorted V1 in descending order of gains;
S2 ← sorted V2 in descending order of gains;
for v ∈ S1, u ∈ S2 do
if gains[v] + gains[u] > 0 then
exchange v and u in the sets;
else break
until converged or iteration limit exceeded;
return graphs induced by Q ∪ V1 and Q ∪ V2
Algorithm 2: Graph Bisection
To initialize the bisection, we consider the following two alternatives. A simpler one is to arbi-
trarily split D into two equal-sized sets. Another approach is based on shingle ordering (minwise
hashing) suggested in [13]. To this end, we order the vertices as described in [13] and assign the first
(cid:98)n/2(cid:99) vertices to V1 and the last (cid:100)n/2(cid:101) to V2.
Algorithm 2 tries to minimize the following objective function of the sets V1 and V2, which is
motivated by BiMLogA. For every vertex q ∈ Q, let deg1(q) = {(q, v) : v ∈ V1}, that is, the
number of adjacent vertices in set V1; define deg2(q) similarly. Then the cost of the partition is
(cid:18)
(cid:88)
q∈Q
deg1(q) log(
n1
deg1(q) + 1
) + deg2(q) log(
n2
deg2(q) + 1
)
,
(cid:19)
where n1 = V1 and n2 = V2. The cost estimates the required number of bits needed to represent
G using delta-encoding. If the neighbors of q ∈ Q are uniformly distributed in the final arrangement
of V1 and V2, then the the average gap between consecutive numbers in the q's adjacency list is
gap1 := n1/(deg1(q)+1) and gap2 := n2/(deg2(q)+1) for V1 and V2, respectively; see Figure 2. There
are (deg1(q) − 1) gaps between vertices in V1 and (deg2(q) − 1) gaps between vertices in V2. Hence,
we need approximately (deg1(q)− 1) log(gap1) + (deg2(q)− 1) log(gap2) bits to compress the within-
group gaps. In addition, we have to account for the average gap between the last vertex of V1 and
the first vertex of V2, which is (gap1 + gap2). Assuming that n1 = n2, we have log(gap1 + gap2) =
log(gap1) + log(gap2) + C, where C is a constant with respect to the data vertex assignment,
and hence, it can be ignored in the optimization. Adding this between-group contribution to the
within-group contributions gives the above expression.
8
Figure 2: Partitioning D into V1 and V2 for a query q ∈ Q with deg1(q) = 3 and deg2(q) = 2.
Note that using the cost function, it is straightforward to implement ComputeM oveGain(v)
function from Algorithm 2 by traversing all the edges (q, v) ∈ E for v ∈ D and summing up the
cost differences of moving v to another set.
Combining all the steps of Algorithm 1 and Algorithm 2, we have the following claim.
Implementation
Theorem 2. The algorithm produces a vertex order in O(m log n + n log2 n) time.
Proof. There are (cid:100)log n(cid:101) levels of recursion. Each call of graph bisection requires computing move
gains and sorting of n elements. The former can be done in O(m) steps, while the latter requires
O(n log n) steps. Summing over all subproblems, we get the claim of the theorem.
4.1
Due to the simplicity of the algorithm, it can be efficiently implemented in parallel or distributed
manner. For the former, we notice that two different recursive calls of Algorithm 1 are indepen-
dent, and thus, can be executed in parallel. Analogously, a single bisection procedure can easily
be parallelized, as each of its steps computes independent values for every vertex, and a parallel
implementation of sorting can be used. In our implementation, we employ the fork-join computation
model in which small enough graphs are processed sequentially, while larger graphs which occur on
the first few levels of recursion are solved in parallel manner.
Our distributed implementation relies on the vertex-centric programming model and runs in the
Giraph framework1. In Giraph, a computation is split into supersteps that consists of processing
steps: (i) a vertex executes a user-defined function based on local vertex data and on data from
adjacent vertices, (ii) the resulting output is sent along outgoing edges. Supersteps end with a
synchronization barrier, which guarantees that messages sent in a given superstep are received at
the beginning of the next superstep. The whole computation is executed iteratively for a certain
number of rounds, or until a convergence property is met.
Algorithm 2 is implemented in the vertex-centric model with a simple modification. The first
two supersteps compute move gains for all data vertices. To this end, every query vertex calculates
the differences of the cost function when its neighbor moves from a set to another one. Then, every
data vertex sums up the differences over its query neighbors. Given the move gains, we exchange the
vertices as follows. Instead of sorting the move gains, we construct, for both sets, an approximate
histogram (e.g., as described in [5]) of the gain values. Since the size of the histograms is small
enough, we collect the data on a dedicated host, and decide how many vertices from each bin should
exchange its set. On the last superstep, this information is propagated over all data vertices and
the corresponding swaps take effect.
1http://giraph.apache.org
9
qV1V2{gap1(q){gap2(q) 5 Experiments
We design our experiments to answer two primary questions: (i) How well does our algorithm
compress graphs and indexes in comparison with existing techniques? (ii) How do various parameters
of the algorithm contribute to the solution, and what are the best parameters?
5.1 Dataset
For our experiments, we use several publicly available web graphs, social networks, and inverted
document indexes; see Table 1. In addition, we run evaluation on two large subgraphs of the Face-
book friendship graph and a sample of the Facebook search index. These private datasets serve to
demonstrate scalability of our approach. We do not release the datasets and our source code due to
corporate restrictions. Before running the tests, all the graphs are made unweighted and converted
to bipartite graphs as described in Section 3.1. Our dataset is as follows.
• Enron represents an email communication network; data is available at https://snap.stanford.
edu/data.
• AS-Oregon is an Autonomous Systems peering information inferred from Oregon route-views
in 2001; data is available at https://snap.stanford.edu/data.
• FB-NewOrlean contains a list of all of the user-to-user links from the Facebook New Orleans
network; the data was crawled and anonymized in 2009 [36].
• web-Google represents web pages with hyperlinks between them. The data was released in
2002 by Google; data is available at https://snap.stanford.edu/data.
• LiveJournal is an undirected version of the public social graph (snapshot from 2006) con-
taining 4.8 million vertices and 42.9 million edges [35].
• Twitter is a public graph of tweets, with about 41 million vertices (twitter accounts) and 2.4
billion edges (denoting followership) [22].
• Gov2 is an inverted index built on the TREC 2004 Terabyte Track test collection, consisting
of 25 million .gov sites crawled in early 2004.
• ClueWeb09 is an inverted index built on the ClueWeb 2009 TREC Category B test collection,
consisting of 50 million English web pages crawled in 2009.
• FB-Posts-1B is an inverted index built on a sample of one billion Facebook posts, containing
the longest posting lists. Since the posts have no hierarchical URLs, the Natural order for
this index is random.
• FB-300M and FB-1B are two subgraphs of the Facebook friendship graph; the data was anonymized
before processing.
To build the inverted indexes for Gov2 and ClueWeb09 the body text was extracted using Apache
Tika2 and the words were lowercased and stemmed using the Porter2 stemmer; no stopwords were
removed. We consider only long posting lists containing more than 4096 elements.
5.2 Techniques
We compare our new algorithm (referred to as BP) with the following competitors.
• Natural is the most basic order defined for a graph. For web graphs and document indexes,
the order is the URL lexicographic ordering used in [9, 29]. For social networks, the order is
induced by the original adjacency matrix.
2http://tika.apache.org
10
Graph
Enron
AS-Oregon
FB-NewOrlean
web-Google
LiveJournal
Twitter
Gov2
ClueWeb09
FB-Posts-1B
FB-300M
FB-1B
Q
9,660
13,579
63,392
356,648
4,847,571
41,652,230
39,187
96,741
60 × 103
300 × 106
1 × 109
D
9,660
13,579
63,392
356,648
4,847,571
41,652,230
24,618,755
50,130,884
1 × 109
300 × 106
1 × 109
E
224,896
74,896
1,633,662
5,186,648
85,702,474
2,405,026,092
5,322,924,226
14,858,911,083
20 × 109
90 × 109
300 × 109
Table 1: Basic properties of our dataset.
• BFS is given by the bread-first search graph traversal algorithm as utilized in [2].
• Minhash is the lexicographic order of 10 minwise hashes of the adjacency sets. The same
approach with only 2 hashes is called double shingle in [13].
• TSP is a heuristic for document reordering suggested by Shieh et al. [32], which is based on solv-
ing the maximum travelling salesman problem. We implemented the variant of the algorithm
that performs best in the authors' experiments. Since the algorithm is computationally expen-
sive, we run it on small instances only. The sparsification techniques presented in [6,15] would
allow us to scale to the larger graphs, but they are too complex to re-implement faithfully.
• LLP represents an order computed by the Layered Label Propagation algorithm [8].
• Spectral order is given by the second smallest eigenvector of the Laplacian matrix of the
graph [20].
• Multiscale is an algorithm based on the multi-level algebraic methodology suggested for
solving MLogA [31].
• SlashBurn is a method for matrix reordering [23].
5.3 Effect of BP parameters
BP has a number of parameters that can affect its quality and performance. In the following we
discuss some of the parameters and explain our choice of their default values.
An important aspect of BP is how two sets, V1 and V2, are initialized in Algorithm 2. Arguably the
initialization procedure might affect the quality of the final vertex order. To verify the hypothesis,
we implemented four initialization techniques that bisect a given graph: Random, Natural, BFS,
and Minhash. The techniques order the data vertices, D, using the corresponding algorithm, and
then split the order into two sets of equal size. In the experiment, we intentionally consider only the
simplest and most efficient bisection techniques so as to keep the complexity of the BP algorithm low.
Figure 3 illustrates the effect of the initialization methods for graph bisection. Note that initialization
plays a role to some extent, and there is no consistent winner. BFS is the best initialization for three
of the graphs but does not produce an improvement on the indexes. One explanation is that the
indexes contain high-degree query vertices, that make the BFS order essentially random. Overall, the
difference between the final results is not substantial, and even the worst initialization yields better
orders than the alternative algorithms do. Therefore, we utilize the simplest approach, Random, for
graphs and Minhash for indexes as the default technique for bisection initialization.
11
Figure 3: LogGap cost of the resulting order pro-
duced with different initialization approaches
for graph bisection.
Figure 4: LogGap cost of the resulting order pro-
duced with a fixed depth of recursion. Note that
the last few splits make insignificant contribu-
tions to the final quality.
Figure 5: The average percentage of moved vertices on an iteration of Algorithm 2 for various levels
of recursion. The data is computed for LiveJournal.
Is it always necessary to perform log n levels of recursion to get a reasonable solution? Figure 4
shows the quality of the resulting vertex order after a fixed number, i, of recursion splits. For every
i (that is, when there are 2i disjoint sets), we stop the algorithm and measure the quality of the
order induced by the random assignment of vertices respecting the partition. It turns out that graph
bisection is beneficial only when D contains more than a few tens of vertices. In our implementation,
we set (log n−5) for the depth of recursion, which slightly reduces the overall running time. It might
be possible to improve the final quality by finding an optimal solution (e.g., using an exhaustive
search or a linear program) for small subgraphs on the lowest levels of the recursion. We leave the
investigation for future research.
Figure 5 illustrates the speed of convergence of our optimization procedure utilized for improv-
ing graph partitioning in Algorithm 2. The two sets approach a locally optimal state within a few
iterations. The number of required iterations increases, as the depth of recursion gets larger. Gen-
erally, the number of moved vertices per iteration does not exceed 1% after 20 iterations, even for
the deepest recursion levels. Therefore, we use 20 as the default number of iterations in all our
experiments.
12
6420LogGap costAS-OregonFB-NewOrleanweb-GoogleLiveJournalGov2 Random Natural BFS Minhash1614121086420LogGap cost20151050depth of recursion AS-Oregon FB-NewOrlean web-Google LiveJournal302520151050moved vertices per iteration, %20151050iterationRecursion depth: 0 3 9 12 15 (a) Natural
(b) BFS
(c) Minhash
(d) TSP
(e) LLP
(f) Spectral
(g) Multiscale
(h) SlashBurn
(i) BP
Figure 6: Adjacency matrices of FB-NewOrlean after applying various reordering algorithms;
nonzero elements are blue.
5.4 Compression ratio
Table 2 presents a comparison of various reordering methods on social networks and web graphs.
We evaluate the following measures: (i) the cost of the BiMLogA problem (LogGap), (ii) the
cost of the MLogA problem (the logarithmic difference averaged over the edges, Log), (iii) the
average number of bits per edge needed to encode the graph with WebGraph [9] (referred to as
BV). The results suggest that BP yields the best compression on all but one instance, providing an
5−20% improvement over the best alternative. An average gain over a non-reordered solution reaches
impressive 50%. The runner-up approaches, TSP, LLP, and Multiscale, also significantly outperform
the natural order. However, their straightforward implementations are not scalable for large graphs
(none of them is able to process Twitter within a few hours), while efficient implementations are
arguably more complicated than BP.
The computed results for FB-300M and FB-1B demonstrate that the new reordering technique is
beneficial for very large graphs, too. Unfortunately, we were not able to calculate the compression
rate for the graphs, as WebGraph [9] does not provide distributed implementation. However, the
experiment indicates that BP outperforms Natural by around 50% and outperforms Minhash by
around 30%.
The compression ratio of inverted indexes is illustrated in Table 3, where we evaluate the Parti-
tioned Elias-Fano [26] encoding and Binary Interpolative Coding [25] (respectively PEF and BIC).
Here the results are reported in average bits per edge. Again, our new algorithm largely outperforms
existing approaches in terms of both LogGap cost and compression rate. BP has a large impact on
the indexes, achieving a 22% and a 15% compression improvement over alternatives; these gains are
almost identical for PEF and BIC.
An interesting question is why does the new algorithm perform best on most of the tested graphs.
In Figure 7 we analyze the number of gaps between consecutive numbers of graph adjacency lists.
It turns out that BP and LLP have quite similar gap distributions, having notably more shorter gaps
than the alternative methods. Note that the number of edges that the BV encoding is able to copy
is related to the number of consecutive integers in the adjacency lists; hence short gaps strongly
influences its performance. At the same time, BP is slightly better at longer gaps, which is a reason
13
Graph
Enron
AS-Oregon
FB-NewOrlean
web-Google
LiveJournal
Twitter
FB-300M
FB-1B
5.01
4.86
4.91
3.95
3.96
5.43
4.23
5.11
3.69
7.88
4.71
4.47
3.59
4.42
5.64
4.53
4.50
3.15
9.74
7.16
7.06
5.62
5.37
7.64
5.90
8.37
4.99
Algorithm LogGap
Natural
BFS
Minhash
TSP
LLP
Spectral
Multiscale
SlashBurn
BP
Natural
BFS
Minhash
TSP
LLP
Spectral
Multiscale
SlashBurn
BP
Natural
BFS
Minhash
TSP
LLP
Spectral
Multiscale
SlashBurn
BP
Natural
BFS
Minhash
TSP
LLP
Spectral
Multiscale
SlashBurn
BP
Natural
BFS
Minhash
LLP
BP
Natural
BFS
Minhash
BP
Natural
Minhash
BP
Natural
Minhash
BP
13.39
5.57
5.65
3.28
3.75
6.68
2.72
8.02
3.17
10.43
10.52
10.79
7.46
7.03
15.23
12.87
10.43
7.91
17.65
13.06
8.39
19.63
14.60
8.66
14
BV
7.80
7.70
7.68
6.58
6.51
8.60
6.90
8.05
6.24
13.34
7.97
7.56
6.66
7.47
8.76
7.31
8.74
6.25
14.64
10.79
10.62
8.96
8.54
11.79
9.25
12.65
8.16
20.08
7.69
6.87
4.77
5.13
9.16
4.10
10.29
4.68
14.61
14.69
15.07
11.12
10.73
21.56
17.99
14.76
11.62
Log
9.82
9.97
10.12
9.46
8.55
9.41
8.00
10.18
8.26
12.06
11.06
11.17
10.39
8.32
9.53
7.23
10.66
9.21
14.29
12.63
12.57
11.61
9.41
11.49
9.58
13.06
9.45
16.74
11.21
13.14
7.99
6.70
10.25
4.82
14.46
7.74
17.44
17.59
17.76
12.25
12.79
23.65
22.69
21.98
20.50
25.34
24.9
18.13
27.22
26.89
18.36
Table 2: Reordering results of various algorithms on graphs: the costs of MLogA, BiMLogA, and
the number of bits per edge required by BV. The best results in every column are highlighted. We
present the results that completed the computation within a few hours.
Index
Gov2
ClueWeb09
FB-Posts-1B
2.52
2.44
2.52
1.95
3.12
3.00
3.12
2.44
Algorithm LogGap PEF BIC
Natural
BFS
Minhash
BP
Natural
BFS
Minhash
BP
Natural
Minhash
BP
4.99
4.99
4.99
4.34
4.05
4.06
4.05
3.50
2.12
2.07
2.12
1.81
2.91
2.91
2.91
2.55
8.03
3.41
2.95
10.19
4.96
4.18
9.95
4.24
3.61
Table 3: Reordering results of various algorithms on inverted indexes with highlighted best results.
(a) FB-NewOrlean
(b) LiveJournal
Figure 7: Distribution of gaps between consecutive elements of graph adjacency lists induced by
various algorithms.
why the new algorithm yields a higher compression ratio.
We point out that the cost of BiMLogA, LogGap, is more relevant for the compression rate
than the cost of MLogA; see Figure 8. The observation agrees with the previous evaluation of Boldi
et al. [8] and motivates our research on the former problem. The Pearson correlation coefficients
between the LogGap cost and the average number of bit per edge using BV, PEF, and BIC
encoding schemes are 0.9853, 0.8487, and 0.8436, respectively. While the high correlation between
LogGap and BV is observed earlier [9,13], the relation between LogGap and PEF or BIC is a new
phenomenon. A possible explanation is that the schemes encode a sequence of k integers in the
range [1..n] using close to the information-theoretic minimum of k(1 +(cid:98)log2(n/k)(cid:99)) bits [26], which
is equivalent to our optimization function utilized in Algorithm 2. It might be possible to construct
a better model for the two encoding schemes, where the cost of the optimization problem has a
higher correlation with the final compression ratio. For example, this can be achieved by increasing
the weights of "short" gaps that are generally require more than log(gap) bits. We leave the question
for future investigation.
Figure 6 presents an alternative comparison of the impact of the reordering algorithms on the
FB-NewOrlean graph. Note that only BP and LLP are able to find communities in the graph (dense
subgraphs), that can be compressed efficiently. The recursive nature of BP is also clearly visible.
15
500x1034003002001000gap count161412108642log2(gap length) Natural BFS Minhash TSP LLP Spectral Multiscale SlashBurn BP20x106151050gap count222018161412108642log2(gap length) Natural BFS Minhash LLP BP Figure 8: LogGap cost against the average number of bits per edge using various encoding schemes.
5.5 Running time
We created and tested two implementations of our algorithm, parallel and distributed. The parallel
version is implemented in C++11 and compiled with the highest optimization settings. The tests
are performed on a machine with Intel(R) Xeon(R) CPU E5-2660 @ 2.20GHz (32 cores) with 128GB
RAM. Our algorithm is highly scalable; the largest instances of our dataset, Gov2, ClueWeb09 and
FB-Posts-1B, are processed with BP within 29, 129, and 163 minutes, respectively. In contrast, even
the simplest Minhash takes 14, 42, and 70 minutes for the indexes. Natural and BFS also have
comparable running times on the graphs. Our largest graphs, Twitter and LiveJournal, require
149 and 3 minutes; all the smaller graphs are processed within a few seconds. In comparison, the
author's implementation of LLP with the default settings takes 23 minutes on LiveJournal and
is not able to process Twitter within a reasonable time. The other alternative methods are less
efficient; for instance, Multiscale runs 12 minutes and TSP runs 3 minutes on web-Google. The
single-machine implementation of BP is also memory-efficient, utilizing less than twice the space
required to store the graph edges.
The distributed version of BP is implemented in Java. We run experiments using the distributed
implementation only on FB-300M and FB-1B graphs, using a cluster of a few tens of machines.
FB-300M is processed within 350 machine-hours, while the computation on FB-1B takes around 2800
machine-hours. In comparison, the running time of the Minhash algorithm is 20 and 60 machine-
hours on the same cluster configuration, respectively. Despite the fact that our implementation is
a part of a general graph partitioning framework [1], which is not specifically optimized for the
problem, BP scales almost linearly with the size of the utilized cluster and processes huge graphs
within a few hours.
6 Conclusions and Future Work
We presented a new theoretically sound algorithm for graph reordering problem and experimentally
proved that the resulting vertex orders allow to compress graphs and indexes more efficiently than
the existing approaches. The method is highly scalable, which is demonstrated via evaluation on
several graphs with billions of vertices and edges. While we see impressive gains in the compression
ratio, we believe there is still much room for further improvement. In particular, our graph bisection
technique ignore the freedom of orienting the decomposition tree. An interesting question is whether
a postprocessing step that "flips" left and right children of tree nodes can be helpful. It is shown
16
20151050LogGap cost1614121086420number of bits per edge PEF BIC BV in [4] that there is an O(n2.2)-time algorithm that computes an optimal tree orientation for the
MLA problem. Whether there exists a similar algorithm for MLogA or BiMLogA, is open.
While our primary motivation is compression, graph reordering plays an important role in a
number of applications. In particular, various graph traversal algorithms can be accelerated if the in-
memory graph layout takes advantage of the cache architecture. Improving vertex and edge locality
is important for fast node/link access operations, and thus can be beneficial for generic graph
algorithms and applications [33]. We are currently working on exploring this area and investigating
how reordering of graph vertices can improve cache and memory utilization.
From the theoretical point of view, it is interesting to devise better approximation algorithms
for the MLogA and BiMLogA problems.
7 Acknowledgments
We thank Yaroslav Akhremtsev, Mayank Pundir, and Arun Sharma for fruitful discussions of the
problem. We thank Ilya Safro for the help with running experiments with the Multiscale method.
References
[1] A, B. Shalita, I. K. Karrer, A. Sharma, A. Presta, A. Adcock, H. Kllapi, and M. Stumm.
Social hash: An assignment framework for optimizing distributed systems operations on social
networks. In Networked Systems Design and Implementation, 2016.
[2] A. Apostolico and G. Drovandi. Graph compression by BFS. Algorithms, 2(3):1031–1044, 2009.
[3] S. Arora, S. Rao, and U. Vazirani. Expander flows, geometric embeddings and graph partition-
ing. Journal of the ACM, 56(2):5, 2009.
[4] R. Bar-Yehuda, G. Even, J. Feldman, and J. Naor. Computing an optimal orientation of a
balanced decomposition tree for linear arrangement problems. JGAA, 5(4):1–27, 2001.
[5] Y. Ben-Haim and E. Tom-Tov. A streaming parallel decision tree algorithm. The Journal of
Machine Learning Research, 11:849–872, 2010.
[6] R. Blanco and Á. Barreiro. Document identifier reassignment through dimensionality reduction.
In Adv. Inf. Retr., pages 375–387. 2005.
[7] D. Blandford and G. Blelloch.
Index compression through document reordering.
In Data
Compression Conference, pages 342–351, 2002.
[8] P. Boldi, M. Rosa, M. Santini, and S. Vigna. Layered label propagation: A multiresolution
coordinate-free ordering for compressing social networks. In World Wide Web, pages 587–596,
2011.
[9] P. Boldi and S. Vigna. The WebGraph framework I: Compression techniques. In World Wide
Web, pages 595–602, 2004.
[10] A. Z. Broder. On the resemblance and containment of documents. In Compression and Com-
plexity of Sequences, pages 21–29, 1997.
17
[11] M. Charikar, M. T. Hajiaghayi, H. Karloff, and S. Rao. l2
2 spreading metrics for vertex ordering
problems. Algorithmica, 56(4):577–604, 2010.
[12] M. Charikar, K. Makarychev, and Y. Makarychev. A divide and conquer algorithm for d-
dimensional arrangement. In Symposium on Discrete Algorithms, pages 541–546, 2007.
[13] F. Chierichetti, R. Kumar, S. Lattanzi, M. Mitzenmacher, A. Panconesi, and P. Raghavan. On
compressing social networks. In Knowledge Discovery and Data Mining, pages 219–228, 2009.
[14] N. R. Devanur, S. A. Khot, R. Saket, and N. K. Vishnoi.
Integrality gaps for sparsest cut
and minimum linear arrangement problems. In Symposium on Theory of Computing, pages
537–546, 2006.
[15] S. Ding, J. Attenberg, and T. Suel. Scalable techniques for document identifier assignment in
inverted indexes. In World Wide Web, pages 311–320, 2010.
[16] G. Even, J. S. Naor, S. Rao, and B. Schieber. Divide-and-conquer approximation algorithms
via spreading metrics. Journal of the ACM, 47(4):585–616, 2000.
[17] C. M. Fiduccia and R. M. Mattheyses. A linear-time heuristic for improving network partitions.
In Design Automation, pages 175–181, 1982.
[18] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of
NP-Completeness. W. H. Freeman & Co., 1979.
[19] M. D. Hansen. Approximation algorithms for geometric embeddings in the plane with appli-
cations to parallel processing problems. In Foundations of Computer Science, pages 604–609,
1989.
[20] M. Juvan and B. Mohar. Optimal linear labelings and eigenvalues of graphs. Discrete Applied
Mathematics, 36(2):153–168, 1992.
[21] B. W. Kernighan and S. Lin. An efficient heuristic procedure for partitioning graphs. Bell
System Technical Journal, 49(2):291–307, 1970.
[22] H. Kwak, C. Lee, H. Park, and S. Moon. What is Twitter, a social network or a news media?
In World Wide Web, pages 591–600, 2010.
[23] Y. Lim, U. Kang, and C. Faloutsos. SlashBurn: Graph compression and mining beyond caveman
communities. IEEE Transactions on Knowledge and Data Engineering, 26(12):3077–3089, 2014.
[24] S. Maneth and F. Peternek. A survey on methods and systems for graph compression. arXiv
preprint arXiv:1504.00616, 2015.
[25] A. Moffat and L. Stuiver. Binary interpolative coding for effective index compression. Infor-
mation Retrieval, 3(1), 2000.
[26] G. Ottaviano and R. Venturini. Partitioned Elias-Fano indexes.
2014.
In SIGIR, pages 273–282,
[27] J. Petit. Addenda to the survey of layout problems. Bulletin of EATCS, 3(105), 2013.
18
[28] U. N. Raghavan, R. Albert, and S. Kumara. Near linear time algorithm to detect community
structures in large-scale networks. Physical Review E, 76(3):036106, 2007.
[29] K. H. Randall, R. Stata, R. G. Wickremesinghe, and J. L. Wiener. The link database: Fast
access to graphs of the web. In Data Compression Conference, pages 122–131, 2002.
[30] S. Rao and A. W. Richa. New approximation techniques for some ordering problems.
Symposium on Discrete Algorithms, pages 211–219, 1998.
In
[31] I. Safro and B. Temkin. Multiscale approach for the network compression-friendly ordering.
Journal of Discrete Algorithms, 9(2):190–202, 2011.
[32] W.-Y. Shieh, T.-F. Chen, J. J.-J. Shann, and C.-P. Chung. Inverted file compression through
document identifier reassignment. Information Processing & Management, 39(1):117–131, 2003.
[33] J. Shun, L. Dhulipala, and G. E. Blelloch. Smaller and faster: Parallel processing of compressed
graphs with Ligra+. In Data Compression Conference, pages 403–412, 2015.
[34] F. Silvestri. Sorting out the document identifier assignment problem. In European Conference
on IR Research, pages 101–112. Springer, 2007.
[35] J. Ugander and L. Backstrom. Balanced label propagation for partitioning massive graphs. In
Web Search and Data Mining, pages 507–516, 2013.
[36] B. Viswanath, A. Mislove, M. Cha, and K. P. Gummadi. On the evolution of user interaction
in Facebook. In Workshop on Social Networks, 2009.
[37] I. H. Witten, A. Moffat, and T. C. Bell. Managing gigabytes: compressing and indexing docu-
ments and images. Morgan Kaufmann, 1999.
19
|
1505.01962 | 2 | 1505 | 2015-10-16T14:33:46 | Applying Sorting Networks to Synthesize Optimized Sorting Libraries | [
"cs.DS",
"cs.MS"
] | This paper shows an application of the theory of sorting networks to facilitate the synthesis of optimized general purpose sorting libraries. Standard sorting libraries are often based on combinations of the classic Quicksort algorithm with insertion sort applied as the base case for small fixed numbers of inputs. Unrolling the code for the base case by ignoring loop conditions eliminates branching and results in code which is equivalent to a sorting network. This enables the application of further program transformations based on sorting network optimizations, and eventually the synthesis of code from sorting networks. We show that if considering the number of comparisons and swaps then theory predicts no real advantage of this approach. However, significant speed-ups are obtained when taking advantage of instruction level parallelism and non-branching conditional assignment instructions, both of which are common in modern CPU architectures. We provide empirical evidence that using code synthesized from efficient sorting networks as the base case for Quicksort libraries results in significant real-world speed-ups. | cs.DS | cs |
Applying Sorting Networks to Synthesize
Optimized Sorting Libraries(cid:63)
Michael Codish1, Lu´ıs Cruz-Filipe2,
Markus Nebel2, and Peter Schneider-Kamp2
1 Department of Computer Science, Ben-Gurion University of the Negev, Israel
2 Dept. Mathematics and Computer Science, Univ. of Southern Denmark,Denmark
Abstract. This paper presents an application of the theory of sorting
networks to facilitate the synthesis of optimized general-purpose sorting
libraries. Standard sorting libraries are often based on combinations of
the classic Quicksort algorithm with insertion sort applied as base case
for small, fixed, numbers of inputs. Unrolling the code for the base case
by ignoring loop conditions eliminates branching, resulting in code equiv-
alent to a sorting network. This enables further program transformations
based on sorting network optimizations, and eventually the synthesis of
code from sorting networks. We show that, if considering the number
of comparisons and swaps, the theory predicts no real advantage of this
approach. However, significant speed-ups are obtained when taking ad-
vantage of instruction level parallelism and non-branching conditional
assignment instructions, both of which are common in modern CPU ar-
chitectures. We provide empirical evidence that using code synthesized
from efficient sorting networks as the base case for Quicksort libraries
results in significant real-world speed-ups.
1
Introduction
General-purpose sorting algorithms are based on comparing, and possibly ex-
changing, pairs of inputs. If the order of these comparisons is predetermined by
the number of inputs to sort and does not depend on their concrete values, then
the algorithm is said to be data-oblivious. Such algorithms are well suited for
e.g. parallel sorting or secure multi-party computations.
Sorting functions in state-of-the-art programming language libraries (such
as the GNU C Library) are typically based on a variant of Quicksort, where
the base cases of the recursion apply insertion sort: once the subsequence to
sort considered by Quicksort falls under a certain length M , it is sorted using
insertion sort. The reasons for using such base cases is that, both theoretically
and empirically, insertion sort is faster than Quicksort for sorting small numbers
of elements. Typical values of M are 4 (e.g. in the GNU C library) or 8.
Generalizing this construction, we can take any sorting algorithm based on
the divide-and-conquer approach (e.g. Quicksort, merge sort), and use another
(cid:63) Supported by the Israel Science Foundation, grant 182/13 and by the Danish Council
for Independent Research, Natural Sciences.
sorting method once the number of elements to sort in one partition does not
exceed a pre-defined limit M . The guiding idea here is that, by supplying opti-
mized code for sorting up to M inputs, the overall performance of the sorting
algorithm can be improved. One obvious way to supply optimized code for sort-
ing up to M inputs is to provide a unique optimized implementaton of sorting
m elements, for each m ≤ M .
This approach leads directly to the following problem: For a given fixed num-
ber M , how can we obtain an efficient way to sort M elements on a modern
CPU? Similar questions have been asked since the 1950s, though obviously with
a different notion of what constitutes a modern CPU.
Sorting networks are a classical model of comparison-based sorting that pro-
vides a framework for addressing such questions. In a sorting network, n inputs
are fed into n channels, connected pairwise by comparators. Each comparator
compares the two inputs from its two channels, and outputs them sorted back
to the same two channels. Consecutive comparators can be viewed as a "parallel
layer" if no two touch the same channel. Sorting networks are data-oblivious al-
gorithms, as the sequence of comparisons performed is independent of the actual
input. For this reason, they are typically viewed as hardware-oriented algorithms,
where data-obliviousness is a requirement and a fixed number of inputs is given.
In this paper, we examine how the theory of sorting networks can improve
the performance of general-purpose software sorting algorithms. We show that
replacing the insertion sort base case of a Quicksort implementation as found in
standard C libraries by optimized code synthesized from logical descriptions of
sorting networks leads to significant improvements in execution times.
The idea of using sorting networks to guide the synthesis of optimized code
for base cases of sorting algorithms may seem rather obvious, and, indeed, has
been pursued earlier. A straightforward attempt, described in [10], has not re-
sulted in significant improvements, though. In this paper we show that this
is not unexpected, providing theoretical and empirical insight into the reasons
for these rather discouraging results. In a nutshell, we provide an average case
analysis of the complexity w.r.t. measures such as number of comparisons and
number of swaps. From the complexity point of view, code synthesized from sort-
ing networks can be expected to perform slightly worse than unrolled insertion
sort. Fortunately, for small numbers (asymptotic) complexity arguments are not
always a good predictor of real-world performance.
The approach taken in [7] matches the advantages of sorting networks with
the vectorization instruction sets available in some modern CPU architectures.
The authors obtain significant speedups by implementing parallel comparators
as vector operations, but they require a complex heuristic algorithm to generate
sequences of bit shuffling code that needs to be executed between comparators.
Their approach is also not fully general, as they target a particular architecture.
In this paper, we combine the best of both these attempts by providing a
straightforward implementation of sorting networks that still takes advantage of
the features of modern CPU architectures, while keeping generality. We obtain
speedups comparable to [7], but our requirements to the instruction set are sat-
isfied by virtually all modern CPUs, including those without vector operations.
The success of our approach is based on two observations.
-- Sorting networks are data-oblivious and the order of comparisons is fully
determined at compile time, i.e., they are free of any control-flow branching.
Comparators can also be implemented without branching, and on modern
CPU architectures even efficiently so.
-- Sorting networks are inherently parallel, i.e., comparators at the same level
can be performed in parallel. Conveniently, this maps directly to implicit in-
struction level parallelism (ILP) common in modern CPU architectures. This
feature allows parallel execution of several instructions on a single thread of
a single core, as long as they are working on disjoint sets of registers.
Avoiding branching and exploiting ILP are tasks also performed through
program transformations by the optimization stages of modern C compilers,
e.g., by unrolling loops and reordering instructions to minimize data-dependence
between neighbouring instructions. They are though both restricted by the data-
dependencies of the algorithms being compiled and, consequently, of only limited
use for data-dependent sorting algorithms, like insertion sort.
Throughout this paper, for empirical evaluations we run all code on an Intel
Core i7, measuring runtime in CPU cycles using the time stamp counter register
using the RDTSC instruction. As a compiler for all benchmarks, we used LLVM
6.1.0 with clang-602.0.49 as frontend on Max OS X 10.10.2. We also tried GCC
4.8.2 on Ubuntu with Linux kernel 3.13.0-36, yielding comparable results.
The remainder of the paper is organized as follows. Section 2 provides back-
ground information and formal definitions for both sorting algorithms and hard-
ware features. In Section 3, we theoretically compare Quicksort and the best
known sorting networks w.r.t. numbers of comparisons and swaps. We aggres-
sively unroll insertion sort until we obtain a sorting network in Section 4, and
in Section 5 we show how to implement individual comparators efficiently. We
empirically evaluate our contribution as a base case of Quicksort in Section 6,
before concluding and giving an outlook on future work in Section 7.
2 Background
2.1 Quicksort with Insertion Sort for Base Case
For decades, Quicksort has been used in practice, due to its efficiency in the
average case. Since its first publication by Hoare [8], several modifications were
suggested to improve it further. Examples are the clever choice of the pivot, or
the use of a different sorting algorithm, e.g., insertion sort, for small subprob-
lem sizes. Most such suggestions have in common that the empirically observed
efficiency can be explained on theoretical grounds by analyzing the expected
number of comparisons, swaps, and partitioning stages (see [13] for details).
Figure 1 presents a comparison of the common spectrum of data-dependent
sorting algorithms for small numbers of inputs, depicting the number of inputs
1600
1400
1200
1000
800
600
400
200
0
Quicksort (C library, with compare function)
selection sort (Sedgewick)
Quicksort (Sedgewick)
shell sort (Sedgewick)
insertion sort (Sedgewick)
+
×
∗
(cid:3)
(cid:4)
∗
+
×
(cid:4)
(cid:3)
∗
×
+
(cid:4)
(cid:3)
2
+
∗
×
(cid:4)
(cid:3)
4
+
∗
×
(cid:4)
(cid:3)
+
∗
(cid:3)
×
(cid:4)
6
+
∗
×
(cid:3)
(cid:4)
+
∗
×
(cid:3)
(cid:4)
8
+
×
∗
(cid:3)
(cid:4)
+
×
∗
(cid:3)
(cid:4)
+
×
∗
(cid:3)
(cid:4)
+
×
∗
(cid:3)
(cid:4)
+
×
∗
(cid:3)
(cid:4)
+
×
∗
(cid:3)
(cid:4)
10
12
14
Fig. 1. Comparison of different sorting algorithms for small numbers of inputs.
(x-axis) together with the number of cycles required to sort them (y-axis), av-
eraged over 100 million random executions. The upper curve in the figure is
obtained from the standard Quicksort implementation in the C library (which is
at some disadvantage, as it requires a general compare function as an argument).
The remaining curves are derived from applying standard sorting algorithms, as
detailed by Sedgewick [14]; the code was taken directly from the book's web page,
http://algs4.cs.princeton.edu/home/. Insertion sort is the clear winner.
2.2 Sorting Networks
A comparator network on n channels is a finite sequence C = c1, . . . , ck of
comparators, where each comparator c(cid:96) is a pair (i(cid:96), j(cid:96)) with 1 ≤ i(cid:96) < j(cid:96) ≤ n. The
size of C is the number k of comparators it contains. Given an input x ∈ Dn,
where D is any totally ordered domain, the output of C on x is the sequence
C(x) = xn, where x(cid:96) is defined inductively as follows: x0 = x, and x(cid:96) is obtained
from x(cid:96)−1 by swapping the elements in positions i(cid:96) and j(cid:96), in case xi(cid:96) < xj(cid:96).
C is a sorting network if C(x) is sorted for all C ∈ Dn. It is well known (see
e.g. [9]) that this property is independent of the concrete domain D.
Comparators may act in parallel. A comparator network C has depth d if C is
the concatenation of L1, . . . , Ld, where each Li is a layer : a comparator network
with the property that no two of its comparators act on a common channel.
Figure 2 depicts a sorting network on 5 channels in the graphical notation
we will use throughout this paper. Comparators are depicted as vertical lines,
and layers are separated by a dashed line. The numbers illustrate how the input
10101 ∈ {0, 1}5 propagates through the network. This network has 6 layers and
9 comparators.
There are two main notions of optimality of sorting networks in common use:
size optimality, where one minimizes the number of comparators used in the
network; and depth optimality, where one minimizes the number of execution
steps, taking into account that some comparators can be executed in parallel.
1
0
1
0
1
0
1
1
0
1
0
1
1
0
1
0
1
0
1
1
0
1
0
1
1
0
1
0
1
1
0
0
1
1
1
Fig. 2. A sorting network on 5 channels operating on the input 10101.
Given n inputs, finding the minimal size sn and depth tn of a sorting net-
work is an extremely hard problem that has seen significant progress in recent
years. The table below details the best currently known bounds. The values for
n ≤ 8 are already listed in [9]; the values of t9 and t10 were proven exact by
Parberry [11], those of t11 -- t16 by Bundala and Z´avodn´y [1], and t17 was recently
computed by Ehlers and Muller [5] using results from [3,4]. Finally, the values
of s9 and s10 were first given in [2].
n 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
sn 0 1 3 5 9 12 16 19 25 29
35 39 45 51 56 60 73
33 37 41 45 49 53 58
tn 0 1 3 3 5 5 6 6 7
7
8
8
9
9
9
9 10
Oblivious versions of classic sorting algorithms can also be implemented as
sorting networks, as described in [9]. Figure 3 (a) shows an oblivious version of
insertion-sort. The vertical dashed lines highlight the 4 iterations of "insertion"
required to sort 5 elements. Figure 3 (b) shows the same network, with compara-
tors arranged in parallel layers. Bubble-sort can also be implemented as a sorting
network as illustrated in Figure 3 (c), where the vertical dashed lines illustrate
the 4 iterations of the classic bubble-sort algorithm. When ordered according to
layers, this network becomes identical to the one in Figure 3 (b).
(a)
(b)
(c)
Fig. 3. Sorting networks for insertion sort (a) and bubble-sort (c) on 5 inputs, dashed
lines separating iterations. When parallelized, both networks become the same (b).
2.3 Modern CPU Architectures
Modern CPU architectures allow multiple instructions to be performed in par-
allel on a single thread. This ability is called instruction-level parallelism (ILP),
and is built on three modern micro-architectural techniques1:
-- superscalar instruction pipelines, i.e., pipelines with the ability to hold and
execute multiple instructions at the same time
-- dynamic out-of-order execution, i.e., dynamic reordering of instructions re-
specting data dependencies
-- redundant execution units, i.e., multiple Arithmetic Logic Units per core
Together, these features allow execution of instructions in an order that mini-
mizes data dependencies, so that multiple redundant execution units can be used
at the same time. This is often termed implicit ILP, in contrast to the explicit
ILP found in vector operations.
Example 1. Consider the C expression (x+y)*(z+u). Assume the variables x, y,
z, and u are loaded in registers eax, ebx, ecx, and edx. Then the evaluation of the
above expression is compiled to three machine instructions: ADD eax,ebx; ADD
ecx,edx; MUL eax,ecx, with the result in ecx. Here, the first two instructions
are data-independent and can be executed in parallel, while the last one depends
on the results of those, and is executed in another CPU cycle.
Conditional branching instructions are the most expensive instructions on
pipelined CPUs, as they require flushing and refilling the pipeline. In order to
minimize their cost, modern CPU architectures employ dynamic branch predic-
tion. By keeping the pipeline filled with the instructions of the predicted branch,
the cost of branching is severely alleviated. Unfortunately, branch prediction can-
not be perfect, and when the wrong branch is predicted, the pipeline needs to
be flushed and refilled -- an operation taking many CPU cycles.
In order to avoid branching instructions for "small" decisions, e.g., decid-
ing whether to assign a value or not, modern CPU architectures also feature
conditional instructions. Depending on flags set by e.g. a comparison, either an
assignment of a value of a register will be performed, or the instruction will be
ignored. In both cases, the pipeline is filled with the subsequent instructions, and
the cost of the operation is smaller than a possible branch prediction failure.
Example 2. Consider the C statement if (x == 42) x = 23; with variable x
loaded in eax. Without conditional move instructions, this is compiled to code
with a conditional branching instruction, i.e. CMP eax,42; JNZ after; MOV
eax, 23, where after is the address of the instruction following the MOV in-
struction. Alternatively, using conditional instructions, we obtain CMP eax, 42;
CMOVZ eax, 23. This code not only saves one machine code instruction, but most
importantly avoids the huge performance impact of a mispredicted branch.
1 For details on these features of modern microarchitectures see e.g. [6,15].
Table 1. Average number of comparisons and swaps when executing optimal sorting
networks with at most M = 14 inputs.
n
1
comparisons 0
2
1
3
3
4
5
5
9
6
7
8
12 16
19
9
25
10
29
11
35
12
39
13
45
14
51
swaps
0.0 0.5 1.5 2.7 4.8 6.6 8.6 10.6 13.0 11.1 19.4 22.4 20.0 26.5
3 Quicksort with Sorting Networks for Base Case
The general theme of this paper is to derive, from sorting networks, optimized
code to sort small numbers of inputs, and then to apply this code as the base
case in a Quicksort algorithm. In this section, we compare precise average case
results for the number of comparisons and swaps performed by a classic Quicksort
algorithm and by a modification that uses sorting networks on subproblems of
size at most 14. We choose 14 for this analysis, as it is the largest value n for
which we could conveniently measure the number of comparisons and swaps
for all n! permutations. We used the best-known (w.r.t. size) sorting networks
(optimal for up to 10 inputs) in order to obtain the most favorable comparison
numbers for sorting networks. To this end, we assume the algorithm to act on
random permutations of size n, each being the input with equal probability.
Let Cn (resp. Sn) denote the expected number of comparisons (resp. swaps)
performed by classic Quicksort on (random) inputs of size n. Let furthermore Cn
and Sn denote the corresponding quantities for Quicksort using sorting networks
for inputs smaller than 15. It is standard to set up recurrence relations for those
quantities which typically obey a pattern such as:
Tn(a, b) =
a · n + b + 1
g(n)
n
1≤j≤n Tj−1(a, b) + Tn−j(a, b)
if n > M ,
otherwise.
(cid:40)
(cid:80)
Here, a and b have to be chosen properly to reflect the parameter's (compar-
isons, swaps) behavior, M determines the maximum subproblem size for which
a different algorithm (insertion sort, sorting networks) is used, and g accounts
for the costs of that algorithm. In order to analyze classic Quicksort as proposed
by Hoare, we have to choose a = 1, b = −1 (resp. a = 1
3 ) for compar-
isons (resp. swaps), together with M = 0 and g(0) = 0. For the analysis of our
proposed modification using sorting networks for subproblems of small sizes, we
set M = 14 together with the values for g as given in Table 1. Using standard
algebraic manipulations, it is possible to solve this recurrence explicitly to obtain
a formula for Tn(a, b) in terms of n, M , a and b. Defining tn = a · n + b and
∇tn = tn − tn−1, one finds (see [12] for details) that, for n > M ,
6 , b = 2
Tn(a, b) = 2(n + 1)
∇tk
k + 1
+
n + 1
M + 2
(tM +1 + TM +1(a, b)) − tn .
(cid:88)
M +2≤k≤n
Computing the closed form expressions for(cid:80)
choices of tn, we finally get
Cn = 2n ln(n + 1) − 2.84557n + o(n)
Cn = 2n ln(n + 1) − 2.44869n + o(n)
M +2≤k≤n
∇tk
k+1 for the different
Sn =
Sn =
1
3
1
3
n ln(n + 1) + 0.359072n + o(n)
n ln(n + 1) + 0.524887n + o(n)
We see that, when increasing n, both parameters get worse by our modification
of classic Quicksort. Even for small n and optimal size sorting networks, there
is no advantage w.r.t. the numbers of comparisons or swaps. In conclusion, we
cannot hope to get a faster sorting algorithm simply by switching to sorting
networks for small subproblems -- at least not on grounds of our theoretical
investigations. And, by transitivity, replacing insertion sort by sorting networks
in the base case should result in an even worse behavior w.r.t. both parameters.
4 Unrolling the Base Case
In this section, we show how to unroll an implementation of insertion sort, step
by step, until we finally obtain code equivalent to a sorting network. We take
the basic insertion sort code from Sedgewick [14], and, for illustration, assume
that the fixed number of inputs is n = 5. We experimented also with optimized
variants (e.g. making use of sentinels to avoid the j>0 check), but did not find
any of them to be faster for small inputs given a modern C compiler.
# define SWAP (x , y ) { int tmp = a [ x ]; a [ x ] = a [ y ]; a [ y ] = tmp ;}
static inline void sort5 ( int *a , int n ) {
n =5
for ( int i = 1; i < n ; i ++)
for ( int j = i ; j > 0 && a [ j ] < a [j -1]; j - -)
SWAP (j -1 , j )
}
Applying partial evaluation and (outer) loop unrolling results in:
static inline void s o r t 5 _ u n r o l l e d ( int * a ) {
for ( int j = 1; j > 0 && a [ j ] < a [j -1]; j - -)
SWAP (j -1 , j )
for ( int j = 2; j > 0 && a [ j ] < a [j -1]; j - -)
SWAP (j -1 , j )
for ( int j = 3; j > 0 && a [ j ] < a [j -1]; j - -)
SWAP (j -1 , j )
for ( int j = 4; j > 0 && a [ j ] < a [j -1]; j - -)
SWAP (j -1 , j )
}
The condition in the inner loop is data-dependent, hence no sound and complete
program transformation can be applied to unroll them. To address this, we move
the data-dependent part of the loop condition to the statement in the body of
the loop, while always iterating the variable j down to 1.
static inline void s o r t 5 _ o b l i v i o u s ( int * a ) {
for ( int j = 1; j > 0; j - -)
if ( a [ j ] < a [j -1]) SWAP (j -1 , j )
for ( int j = 2; j > 0; j - -)
if ( a [ j ] < a [j -1]) SWAP (j -1 , j )
for ( int j = 3; j > 0; j - -)
if ( a [ j ] < a [j -1]) SWAP (j -1 , j )
for ( int j = 4; j > 0; j - -)
if ( a [ j ] < a [j -1]) SWAP (j -1 , j )
}
Now we can now apply (inner) loop unrolling and obtain:
static inline void s o r t 5 _ o b l i v o u s _ u n r o l l e d ( int * a ) {
if ( a [1] < a [0]) SWAP (0 , 1)
if ( a [2] < a [1]) SWAP (1 , 2)
if ( a [1] < a [0]) SWAP (0 , 1)
if ( a [3] < a [2]) SWAP (2 , 3)
if ( a [2] < a [1]) SWAP (1 , 2)
if ( a [1] < a [0]) SWAP (0 , 1)
if ( a [4] < a [3]) SWAP (3 , 4)
if ( a [3] < a [2]) SWAP (2 , 3)
if ( a [2] < a [1]) SWAP (1 , 2)
if ( a [1] < a [0]) SWAP (0 , 1)
}
All the statements in the body of sort5 oblivous unrolled are now conditional
swaps. For readability, we move the condition into the macro. COMPs on the same
line indicate that they originate from the same iteration of insertion sort:
# define COMP (x , y ) { if ( a [ y ] < a [ x ]) SWAP (x , y ) }
static inline void s o r t 5 _ f i g 3 a ( int * a ) {
COMP (0 , 1)
COMP (1 , 2)
COMP (2 , 3)
COMP (3 , 4)
COMP (0 , 1)
COMP (1 , 2)
COMP (2 , 3)
}
COMP (0 , 1)
COMP (1 , 2)
COMP (0 , 1)
This sequence is equivalent to the sorting network in Figure 3 (a). Thus, we
can apply the reordering of comparators that resulted in Figure 3 (b) to obtain
the following implementation, where we reduce the number of layers to 7 (here,
COMPs on the same line indicate a layer in the sorting network):
static inline void s o r t 5 _ f i g 3 b ( int * a ) {
COMP (0 , 1)
COMP (1 , 2)
COMP (0 , 1)
COMP (1 , 2)
COMP (0 , 1)
COMP (1 , 2)
COMP (0 , 1)
}
COMP (2 , 3)
COMP (3 , 4)
COMP (2 , 3)
700
600
500
400
300
200
100
0
insertion sort
unrolled insertion sort
insertion sorting network
compressed insertion sorting network
optimal sorting network
+
×
∗
(cid:3)
(cid:4)
×
∗
(cid:4)
(cid:3)
+
∗
×
(cid:4)
(cid:3)
+
2
×
∗
(cid:3)
(cid:4)
+
4
×
∗
(cid:3)
(cid:4)
+
×
∗
(cid:3)
(cid:4)
+
6
×
∗
(cid:4)
(cid:3)
+
∗
+
×
(cid:3)
(cid:4)
8
∗
(cid:3)
+
×
(cid:4)
∗
(cid:3)
+
×
(cid:4)
∗
+
(cid:3)
×
(cid:4)
∗
(cid:3)
+
×
(cid:4)
∗
(cid:3)
+
×
(cid:4)
+
∗
(cid:3)
×
(cid:4)
10
12
14
Fig. 4. Comparison of insertion sort with (unrolled) comparator based code for small
numbers of inputs.
Figure 4 presents a comparison of a standard insertion sort (code from [14])
with the several optimized versions, depicting the number of inputs (x-axis) to-
gether with the number of cycles required to sort them (y-axis), averaged over
100 million random executions. The curve labeled "insertion sort" portrays the
same data as the corresponding curve in Figure 1. The curve labeled "unrolled
insertion sort" corresponds to the unrolled version of insertion sort (in the style
of function sort5 unrolled). The other three curves correspond to code derived
from different types of sorting networks: the "insertion sorting network" from
Figure 3 (a) and function sort5 fig3a; the "compressed insertion sorting net-
work" from Figure 3 (b) and function sort5 fig3b; and the "optimal sorting
network", corresponding to the use of a best (smallest) known sorting network.
From the figure, it is clear that standard sorting network optimizations such
as reordering of independent comparators [9] give a slight performance boost.
But there is another clear message: even going beyond standard program trans-
formations by breaking data-dependence and obtaining a sequence of conditional
swaps (i.e., a sorting network), we do not manage to make any significant im-
provements of the performance of sorting implementations for small numbers of
inputs. Furthermore, even when using size-optimal sorting networks, we obtain
no real benefit over compiler-optimized insertion sort. This is in line with the
theoretical results on average case complexity discussed in the previous section.
5
Implementing Sorting Networks Efficiently
The results in the previous two sections explained the rather discouraging re-
sults obtained by a naive attempt to use sorting networks as the base case of a
divide-and-conquer sorting algorithm: they are simply not faster than e.g. inser-
tion sort -- at least when implemented naively. In this section we show how to
exploit two main properties of sorting networks, together with features of mod-
ern CPU architectures, and obtain speed-ups of a factor higher than 3 compared
to unrolled insertion sort.
We first observe that, as sorting networks are data-oblivious, the order of
comparisons is fully determined at compile time, i.e., their implementation is
free of any control-flow branching. Unfortunately, the naive implementation of
each comparator involves branching to decide whether to perform a swap. The
path taken depends entirely on the specific inputs to be sorted, and as such
branch prediction necessarily does not perform very well.
Luckily, we can also implement comparators without branching. To this end,
we use a conditional assignment (defined by the macro COND below), which can
be compiled to the conditonal move (CMOV) instruction available on modern CPU
architectures. This approach proved to be very fruitful. For illustration, from the
optimal-size sorting network for 5 inputs portrayed in Figure 2, we synthesize
the following C function sort5 best, where each row in the code corresponds
to a layer in the sorting network:
# define COND (c ,x , y ) { x = ( c ) ? y : x ; }
# define COMP (x , y ) { int ax = a [ x ]; COND ( a [ y ] < ax , a [ x ] , a [ y ]) ; \
) ; }
COND ( a [ y ] < ax , a [ y ] , ax
static inline void s o r t 5 _ b e s t ( int * a ) {
COMP (0 , 1)
COMP (2 , 4)
COMP (2 , 3)
COMP (0 , 3)
COMP (0 , 2)
COMP (1 , 2)
COMP (3 , 4)
COMP (1 , 4)
COMP (1 , 3)
}
The comparator macro that compares and conditionally swaps the values at
indices x and y works as follows:
1. Keep a copy of the value at index x.
2. Compare (once) the value at index y with the stored value from x.
3. If the value was greater, copy the value at index y to index x. Otherwise, do
nothing.
4. If the value was greater, write the old copied value from x to index y. Oth-
erwise, do nothing.
Correctness follows directly by case analysis. If the value at index y was not
greater than the value at index x, the two conditional assignments do not change
anything, and all we did was an unnecessary copy of the valued at index x. If the
value at index y was greater than the value at index y, we essentially perform a
classic swap using ax as the temporary variable.
Given a sufficient optimization level (-O2 and above), the above code is com-
piled by the LLVM (or GNU) C compiler to use two conditional move (CMOV)
instructions, resulting in a totally branching free code for sort5 best. As can
be expected, the other two instructions are a move (MOV) and a compare (CMP)
instruction. In other words, each comparator is implemented by exactly four
non-branching machine code instructions.
Alternatively, we could implement the comparator applying the folklore idea
of swapping values using XORs to eliminate one conditional assignment:2
# define COND (c ,x , y ) { x = ( c ) ? y : x ; }
# define COMP (x , y ) { int ax = a [ x ]; COND ( a [ y ] < ax , a [ x ] , a [ y ]) ; \
a [ y ] ^= ax ^ a [ x ]; }
This alternative comparator performs a conditional swap as follows:
1. Keep a copy of the value at index x.
2. If the value at index y is greater than the value at index x, copy the value
at index y to index x.
3. Bitwise XOR the value at index y with the copied old and the new value at
index x.
Step 3 works because, if the condition holds, then ax and the value at index x
cancel out, leaving the value at y unchanged, while otherwise the value at y and
ax cancel out, effectively assigning the original value from index x to index y.
We also implemented this variant, and observed that it compiles down to five
instructions (MOV, CMP, CMOV, and two XORs). We benchmarked the two variants
and observed that they are indistinguishable in practice, with differences well
within the margin of measurement error. Thus, we decided to continue with this
second version, as the XOR instructions are more "basic" and can therefore be
expected to behave better w.r.t. e.g. instruction level parallelism.
A third approach would be to define branching-free minimum and maximum
operations,3 and use them to assign the minimum to the upper channel and
the maximum to the lower channel of the comparator. We tested this approach,
but found that it did not compile to branching-free code. Even if it did, the
number of instructions involved would be rather large, eliminating any chance
of competing with the two previous variants.
The reader might wonder whether a different SWAP macro could similarly
speed up the working of standard insertion sort. The answer is a clear no, as the
standard swapping operation is implemented by only three operations. Tricks
like using XORs only increase the number of instructions to execute, while not
reducing branching in the code. We implemented and benchmarked several alter-
native SWAP macros, finding only detrimental effects on measured performance.
Figure 5 compares three sorting algorithms for small numbers of inputs: (1)
the unrolled insertion sort (also plotted in Figure 4); (2) code derived from
a standard insertion sorting network (also plotted in Figure 4); (3) the same
insertion sorting network but with a non-branching version of the COMP macro.
We compare the number of branches encountered and mispredicted (averaged
2 See
https://graphics.stanford.edu/~{}seander/bithacks.html#
SwappingValuesXOR
3 See
IntegerMinOrMax
https://graphics.stanford.edu/~{}seander/bithacks.html#
140
120
100
80
60
40
20
0
unrolled insertion sort #branches
insertion sorting network #branches
non-branching insertion sorting network #branches
unrolled insertion sort #mispredicts
insertion sorting network #mispredicts
non-branching insertion sorting network #mispredicts
+
×
∗
(cid:3)
(cid:4)
◦
×
∗
+
(cid:4)
◦
(cid:3)
×
∗
◦
(cid:4)
(cid:3)
+
2
×
∗
+
(cid:4)
(cid:3)
◦
4
×
∗
+
(cid:4)
(cid:3)
◦
×
+
∗
(cid:4)
(cid:3)
◦
6
×
+
∗
(cid:4)
(cid:3)
◦
×
+
∗
(cid:4)
(cid:3)
◦
10
×
+
∗
(cid:4)
(cid:3)
◦
×
+
∗
(cid:4)
(cid:3)
◦
8
×
+
∗
(cid:4)
(cid:3)
◦
×
+
∗
(cid:4)
(cid:3)
◦
14
×
+
∗
(cid:4)
(cid:3)
◦
×
+
∗
(cid:4)
(cid:3)
◦
12
Fig. 5. Comparing the number of branches, encountered and mispredicted, in optimized
sorting algorithms for small numbers of inputs.
over 100 million random executions). From the figure it is clear that the number
of branches encountered (and mispredicted) is larger for both unrolled insertion
sort and a naive implementation of sorting networks. In contrast, the branching-
free implementation exhibits a nearly constant level of branches encountered
and mispredicted. These branches actually originate from the surrounding test
code (filling an array with random numbers, computing random numbers, and
checking that the result is actually sorted).
Our second observation is that sorting networks are inherently parallel, i.e.,
comparators at the same level can be performed simultaneously. This parallelism
can be mapped directly to instruction level parallelism (ILP). The ability to
make use of ILP has further performance potential. In order to demonstrate
this potential, we constructed artificial test cases with varying levels of data
dependency. Given a natural number m, we construct a comparator network of
size 1000 consisting of subsequences of m parallel comparators. We would expect
that, as m grows, we would see more use of ILP.
5000
4500
4000
3500
3000
2500
2000
1500
1000
+
+
0
1
2
+
3
+
4
+
5
+
6
+
7
+
8
9
Fig. 6. ILP on comparator networks of length 1000 with differing levels of parallelism.
700
600
500
400
300
200
100
0
unrolled insertion sort
insertion sorting network
compressed insertion sorting network
optimal sorting network
non-branching insertion sorting network
non-branching compressed insertion sorting network
non-branching optimal sorting network
×
∗
(cid:3)
•
◦
+
(cid:4)
×
∗
◦
•
(cid:3)
(cid:4)
+
2
×
∗
(cid:3)
+
◦
•
(cid:4)
4
×
∗
(cid:3)
+
◦
•
(cid:4)
×
∗
(cid:3)
+
◦
(cid:4)
•
6
×
∗
(cid:3)
+
◦
(cid:4)
•
×
∗
(cid:3)
+
(cid:4)
◦
•
8
+
×
∗
(cid:3)
(cid:4)
◦
•
∗
×
+
(cid:3)
(cid:4)
◦
•
×
∗
(cid:3)
+
(cid:4)
◦
•
×
∗
(cid:3)
+
(cid:4)
◦
•
×
∗
(cid:3)
+
(cid:4)
◦
•
×
∗
(cid:3)
+
(cid:4)
◦
•
12
14
×
∗
(cid:3)
+
(cid:4)
◦
•
10
Fig. 7. Comparison of sorting networks for small numbers of inputs: non-branching
sorting networks are fastest.
In Figure 6, the values for m are represented on the x-axis, while the y-axis (as
usual) indicates the averaged number of CPU cycles. Indeed, we see significant
performance gains when going from m = 1 to m = 2 and m = 3. From this value
onwards, performance stays unchanged. This is the result of each comparator
being compiled to 5 assembler instructions when using optimization level -O3.
Then we obtain slightly under 2 CPU cycles per comparator.
Combining the gains from ILP with the absence of branching, we obtain
large speed-ups for small inputs when comparing to both insertion sort and
naive implementations of sorting networks. In Figure 7, we show the magni-
tude of the improvements obtained. Once again we plot the number of inputs
on the x-axis against the number of cycles required to sort then on the y-axis,
averaged over 100 million random executions. We consider the unrolled inser-
tion sort, the three sorting networks from Figure 4 (insertion sorting network,
compressed insertion sorting network, and optimal sorting network), and these
same three sorting networks using non-branching comparators (non-branching
insertion sorting network, non-branching compressed insertion sorting network,
and non-branching optimal sorting network). The figure shows that using the
best known (optimal) sorting networks in their non-branching forms results in a
speed-up by a factor of more than 3.
6 Quicksort with Sorting Network Base Case
We now demonstrate that optimizing the code in the base case of a Quicksort
algorithm translates to real-world savings when applying the sorting function. To
this end, we use as base cases (1) the (empirically) best variant of insertion sort
unrolled by applying program transformations to the algorithm from [14], and
(2) the fastest non-branching code derived from optimal (size) sorting networks.
170,000
160,000
150,000
140,000
130,000
120,000
110,000
unrolled insertion sort
non-branching optimal sorting network
+
×
+
×
+
×
+
×
+
×
+
×
+
×
2
4
6
+
×
+
×
8
+
×
+
×
10
+
×
12
+
+
×
×
14
Fig. 8. Quicksort: comparing insertion sort at the base case with non-branching optimal
sorting networks at the base case. Plotting base case size (x axis) and number of cycles
(averaged over one million random runs).
In Figure 8 we depict the results of sorting lists of 10,000 elements. The y-axis
measures the number of cycles (averaged over one million random runs), and the
x-axis specifies the limit at which Quicksort reverts to a base case. For example,
the value 8 indicates that the algorithm uses a base case whenever it is required
to sort a sequence of length at most 8 elements. The value 2 corresponds to the
case where the base case has no impact. To quantify the impact of the choice of
base case, we compare to the case for value 2 (on the x-axis). For insertion sort we
see a 2 -- 12% reduction in runtime depending on the limit, and for non-branching
sorting networks we achieve instead 7 -- 23% reduction in runtime.
7 Conclusion
In this paper, we showed, both theoretically and empirically, that using code
derived naively from sorting networks is not advantageous to sort small numbers
of inputs, compared to the use of standard data-dependent sorting algorithms
like insertion sort. Furthermore, we showed that program transformations are of
only limited utility for improving insertion sort on small numbers of inputs.
By contrast, we showed how to synthesize simple yet efficient implementa-
tions of sorting networks, and gave insight into the microarchitectural features
that enable this implementation. We demonstrated that we do obtain significant
speed-ups compared to naive implementations such as [10]. A further empirical
comparison between our implementation and the one described in [7] (not de-
tailed in this paper) shows similar performance and scaling behavior. However,
our approach allows the exploitation of instruction-level parallelism without the
need for a complex instruction set-specific algorithm, as required by [7]. We also
provided further evidence that efficient sorting networks are useful as a base case
in divide-and-conquer sorting algorithms such as, e.g., Quicksort.
Our results also show that using different sorting networks has measurable
impact on the efficiency of the synthesized C code. While previous research on
finding optimal sorting networks has focused on optimal depth or optimal size,
in the future we plan to identify criteria that will lead to optimal performance
in this context. What are the parameters that determine real-world efficiency of
the synthesized code, and how can we find sorting networks that optimize these
parameters? We also plan to explore other target architectures, such as GPUs,
and to benchmark our approach as base case for other sorting algorithms, such
as merge sort.
References
1. D. Bundala and J. Z´avodn´y. Optimal sorting networks. In A.-H. Dediu, C. Mart´ın-
Vide, J.L. Sierra-Rodr´ıguez, and B. Truthe, editors, LATA 2014, volume 8370 of
LNCS, pages 236 -- 247. Springer, 2014.
2. M. Codish, L. Cruz-Filipe, M. Frank, and P. Schneider-Kamp. Twenty-five com-
parators is optimal when sorting nine inputs (and twenty-nine for ten). In ICTAI
2014, pages 186 -- 193. IEEE, December 2014.
3. M. Codish, L. Cruz-Filipe, and P. Schneider-Kamp. The quest for optimal sorting
networks: Efficient generation of two-layer prefixes. In F. Winkler, V. Negru, T. Ida,
T. Jebelan, D. Petcu, S.M. Watt, and D. Zaharie, editors, SYNASC 2014, pages
359 -- 366. IEEE, 2015.
4. M. Codish, L. Cruz-Filipe, and P. Schneider-Kamp. Sorting networks: the end
game. In A.-H. Dediu, E. Formenti, C. Mart´ın-Vide, and B. Truthe, editors, LATA
2015, volume 8977 of LNCS, pages 664 -- 675. Springer, 2015.
5. T. Ehlers and M. Muller. New bounds on optimal sorting networks. In A. Beck-
mann, V. Mitrana, and M.I. Soskova, editors, CiE 2015, volume 9136 of LNCS,
pages 167 -- 176. Springer, 2015.
6. J.A. Fisher, P. Faraboschi, and C. Young. Embedded Computing: A VLIW Ap-
proach to Architecture, Compilers, and Tools. Morgan Kaufman, 2005.
7. T. Furtak, J.N. Amaral, and R. Niewiadomski. Using SIMD registers and instruc-
tions to enable instruction-level parallelism in sorting algorithms. In SPAA '07,
pages 348 -- 357. ACM, 2007.
8. C.A.R. Hoare. Quicksort. Comput. J., 5(1):10 -- 15, 1962.
9. D.E. Knuth. The Art of Computer Programming, Volume III: Sorting and Search-
ing. Addison-Wesley, 1973.
10. B. Lopez and N. Cruz-Cortes. On the usage of sorting networks to big data. In H.R.
Arabnia, M.Q. Yang, G. Jandieri, J.J. Park, A.M.G. Solo, and F.G. Tinetti, editors,
Advances in Big Data Analytics: The 2014 WorldComp International Conference
Proceedings. Mercury Learning and Information, 2014.
11. I. Parberry. A computer-assisted optimal depth lower bound for nine-input sorting
networks. Mathematical Systems Theory, 24(2):101 -- 116, 1991.
12. R. Sedgewick. The analysis of quicksort programs. Acta Inf., 7:327 -- 355, 1977.
13. R. Sedgewick and P. Flajolet. An introduction to the analysis of algorithms.
Addison-Wesley-Longman, 1996.
14. R. Sedgewick and K. Wayne. Algorithms. Addison-Wesley, 2011. 4th Edition.
15. J. Silc, B. Robic, and T. Ungerer. Processor Architecture: From Dataflow to Su-
perscalar and Beyond. Springer, 1999.
|
1608.06462 | 1 | 1608 | 2016-08-23T11:10:10 | On Low-High Orders of Directed Graphs: Incremental Algorithms and Applications | [
"cs.DS"
] | A flow graph $G=(V,E,s)$ is a directed graph with a distinguished start vertex $s$. The dominator tree $D$ of $G$ is a tree rooted at $s$, such that a vertex $v$ is an ancestor of a vertex $w$ if and only if all paths from $s$ to $w$ include $v$. The dominator tree is a central tool in program optimization and code generation and has many applications in other diverse areas including constraint programming, circuit testing, biology, and in algorithms for graph connectivity problems. A low-high order of $G$ is a preorder $\delta$ of $D$ that certifies the correctness of $D$ and has further applications in connectivity and path-determination problems. In this paper, we first consider how to maintain efficiently a low-high order of a flow graph incrementally under edge insertions. We present algorithms that run in $O(mn)$ total time for a sequence of $m$ edge insertions in an initially empty flow graph with $n$ vertices.These immediately provide the first incremental certifying algorithms for maintaining the dominator tree in $O(mn)$ total time, and also imply incremental algorithms for other problems. Hence, we provide a substantial improvement over the $O(m^2)$ simple-minded algorithms, which recompute the solution from scratch after each edge insertion. We also show how to apply low-high orders to obtain a linear-time $2$-approximation algorithm for the smallest $2$-vertex-connected spanning subgraph problem (2VCSS). Finally, we present efficient implementations of our new algorithms for the incremental low-high and 2VCSS problems and conduct an extensive experimental study on real-world graphs taken from a variety of application areas. The experimental results show that our algorithms perform very well in practice. | cs.DS | cs |
On Low-High Orders of Directed Graphs: An Incremental
Algorithm and Applications
Loukas Georgiadis∗
Aikaterini Karanasiou∗
Giannis Konstantinos∗
Luigi Laura†
October 15, 2018
Abstract
A flow graph G = (V, E, s) is a directed graph with a distinguished start vertex s. The dom-
inator tree D of G is a tree rooted at s, such that a vertex v is an ancestor of a vertex w if and
only if all paths from s to w include v. The dominator tree is a central tool in program optimiza-
tion and code generation, and has many applications in other diverse areas including constraint
programming, circuit testing, biology, and in algorithms for graph connectivity problems. A
low-high order of G is a preorder δ of D that certifies the correctness of D, and has further
applications in connectivity and path-determination problems. In this paper we first consider
how to maintain efficiently a low-high order of a flow graph incrementally under edge insertions.
We present algorithms that run in O(mn) total time for a sequence of m edge insertions in
an initially empty flow graph with n vertices. These immediately provide the first incremental
certifying algorithms for maintaining the dominator tree in O(mn) total time, and also imply
incremental algorithms for other problems. Hence, we provide a substantial improvement over
the O(m2) simple-minded algorithms, which recompute the solution from scratch after each edge
insertion. We also show how to apply low-high orders to obtain a linear-time 2-approximation
algorithm for the smallest 2-vertex-connected spanning subgraph problem (2VCSS). Finally, we
present efficient implementations of our new algorithms for the incremental low-high and 2VCSS
problems, and conduct an extensive experimental study on real-world graphs taken from a va-
riety of application areas. The experimental results show that our algorithms perform very well
in practice.
Introduction
1
A flow graph G = (V, E, s) is a directed graph (digraph) with a distinguished start vertex s ∈ V .
A vertex v is reachable in G if there is a path from s to v; v is unreachable if no such path exists.
The dominator relation in G is defined for the set of reachable vertices as follows. A vertex v is
a dominator of a vertex w (v dominates w) if every path from s to w contains v; v is a proper
dominator of w if v dominates w and v (cid:54)= w. The dominator relation in G can be represented by a
tree rooted at s, the dominator tree D, such that v dominates w if and only if v is an ancestor of
w in D. If w (cid:54)= s is reachable, we denote by d(w) the parent of w in D. Lengauer and Tarjan [35]
presented an algorithm for computing dominators in O(mα(m, n)) time for a flow graph with n
vertices and m edges, where α is a functional inverse of Ackermann's function [45]. Subsequently,
several linear-time algorithms were discovered [3, 9, 14, 15]. The dominator tree is a central
∗Department of Computer Science & Engineering, University of Ioannina, Greece. E-mail: [email protected],
†Dipartimento di Ingegneria Informatica, Automatica e Gestionale, "Sapienza" Universit`a di Roma, Italy. E-mail:
[email protected], giannis [email protected].
[email protected].
Figure 1: A flow graph G, its dominator tree D, and two strongly divergent spanning trees B and
R. The numbers correspond to a preorder numbering of D that is a low-high order of G.
tool in program optimization and code generation [12], and it has applications in other diverse
areas including constraint programming [40], circuit testing [5], theoretical biology [2], memory
profiling [36], the analysis of diffusion networks [29], and in connectivity problems [19, 20, 21, 22,
24, 30, 31, 32, 33].
A low-high order δ of G [25] is a preorder of the dominator tree D such for all reachable vertices
v (cid:54)= s, (d(v), v) ∈ E or there are two edges (u, v) ∈ E, (w, v) ∈ E such that u and w are reachable, u
is less than v (u <δ v), v is less than w (v <δ w), and w is not a descendant of v in D. See Figure 1.
Every flow graph G has a low-high order, computable in linear-time [25]. Low-high orders provide
a correctness certificate for dominator trees that is straightforward to verify [48]. By augmenting
an algorithm that computes the dominator tree D of a flow graph G so that it also computes a
low-high order of G, one obtains a certifying algorithm to compute D. (A certifying algorithm [37]
outputs both the solution and a correctness certificate, with the property that it is straightforward
to use the certificate to verify that the computed solution is correct.) Low-high orders also have
applications in path-determination problems [47] and in fault-tolerant network design [6, 7, 26].
A notion closely related to low-high orders is that of divergent spanning trees [25]. Let Vr be
the set of reachable vertices, and let G[Vr] be the flow graph with start vertex s that is induced by
Vr. Two spanning trees B and R of G[Vr], rooted at s, are divergent if for all v, the paths from s
2
4𝑠𝑎𝑏𝑐𝑖𝑔ℎ𝑒𝑑𝑓𝐷15326789104𝑠𝑎𝑏𝑔ℎ𝑐𝑓𝑑𝑒𝑖𝐺19105738264𝑠𝑎𝑏𝑔ℎ𝑐𝑓𝑑𝑒𝑖𝐵19105738264𝑠𝑎𝑏𝑔ℎ𝑐𝑓𝑑𝑒𝑖𝑅1910573826 to v in B and R share only the dominators of v; B and R are strongly divergent if for every pair
of vertices v and w, either the path in B from s to v and the path in R from s to w share only the
common dominators of v and w, or the path in R from s to v and the path in B from s to w share
only the common dominators of v and w. In order to simplify our notation, we will refer to B and
R, with some abuse of terminology, as strongly divergent spanning trees of G. Every flow graph
has a pair of strongly divergent spanning trees. Given a low-high order of G, it is straightforward
to compute two strongly divergent spanning trees of G in O(m) time [25]. Divergent spanning trees
can be used in data structures that compute pairs of vertex-disjoint s-t paths in 2-vertex connected
digraphs (for any two query vertices s and t) [19], in fast algorithms for approximating the smallest
2-vertex-connected spanning subgraph of a digraph [20], and in constructing sparse subgraphs of a
given digraph that maintain certain connectivity requirements [21, 32, 33].
Figure 2: The flow graph of Figure 1 after the insertion of edge (g, d), and its updated dominator
tree D(cid:48) with a low-high order, and two strongly divergent spanning trees B(cid:48) and R(cid:48).
In this paper we consider how to update a low-high order of a flow graph through a sequence
of edge insertions. See Figure 2. The difficulty in updating the dominator tree and a low-high
order is due to the following facts. An affected vertex can be arbitrarily far from the inserted edge,
and a single edge insertion may cause O(n) parent changes in D. Furthermore, since a low-high
order is a preorder of D, a single edge insertion may cause O(n) changes in this order, even if
there is only one vertex that is assigned a new parent in D after the insertion. More generally,
3
𝑠5𝑐𝐷′1𝑑4𝑎𝑔32𝑓6𝑒7𝑖8𝑏ℎ9105𝑠𝑎𝑏𝑔ℎ𝑐𝑓𝑑𝑒𝑖𝐺′19104738265𝑠𝑎𝑏𝑔ℎ𝑐𝑓𝑑𝑒𝑖𝐵′19104738265𝑠𝑎𝑏𝑔ℎ𝑐𝑓𝑑𝑒𝑖𝑅′1910473826 we note that the hardness of dynamic algorithms on digraphs has been recently supported also by
conditional lower bounds [1]. Our first contribution is to show that we can maintain a low-high
order of a flow graph G with n vertices through a sequence of m edge insertions in O(mn) total
time. Hence, we obtain a substantial improvement over the naive solution of recomputing a low-
high order from scratch after each edge insertion, which takes O(m2) total time. Our result also
implies the first incremental certifying algorithms [37] for computing dominators in O(mn) total
time, which answers an open question in [25]. We present two algorithms that achieve this bound, a
simple and a more sophisticated. Both algorithms combine the incremental dominators algorithm
of [23] with the linear-time computation of two divergent spanning trees from [25]. Our more
sophisticated algorithm also applies a slightly modified version of a static low-high algorithm from
[25] on an auxiliary graph. Although both algorithms have the same worst-case running time, our
experimental results show that the sophisticated algorithm is by far superior in practical scenarios.
We note that the incremental dominators problem arises in various applications, such as in-
cremental data flow analysis and compilation [11, 18, 41, 43], distributed authorization [38], and
in incremental algorithms for maintaining 2-connectivity relations in directed graphs [24]. In Sec-
tion 4 we show how our result on incremental low-high order maintenance implies the following
incremental algorithms that also run in O(mn) total time for a sequence of m edge insertions.
• First, we give an algorithm that maintains, after each edge insertion, two strongly divergent
spanning trees of G, and answers the following queries in constant time: (i) For any two
query vertices v and w, find a path πsv from s to v and a path πsw from s to w, such that
πsv and πsw share only the common dominators of v and w. We can output these paths in
O(πsv + πsw) time. (ii) For any two query vertices v and w, find a path πsv from s to v
that avoids w, if such a path exists. We can output this path in O(πsv) time.
• Then we provide an algorithm for an incremental version of the fault-tolerant reachability
problem [6, 7]. We maintain a flow graph G = (V, E, s) with n vertices through a sequence of
m edge insertions, so that we can answer the following query in O(n) time. Given a spanning
forest F = (V, EF ) of G rooted at s, find a set of edges E(cid:48) ⊆ E \ EF of minimum cardinality,
such that the subgraph G(cid:48) = (V, EF ∪ E(cid:48), s) of G has the same dominators as G.
• Finally, given a digraph G, we consider how to maintain incrementally a spanning subgraph
of G with O(n) edges that preserves the 2-edge-connectivity relations in G.
We also revisit the problem of computing a smallest 2-vertex-connected spanning subgraph
(2VCSS) of a directed graph [10, 20]. We present a linear-time algorithm that computes a 2-
approximation of the smallest 2-vertex-connected spanning subgraph (2VCSS) of a 2-vertex-connected
digraph. This improves significantly the best previous approximation ratio achievable in linear time
for this problem, which was 3 [20]. Our new algorithm is also shown to achieve better performance
in practice.
2 Preliminaries
Let G = (V, E, s) be a flow graph with start vertex s, and let D be the dominator tree of G. A
spanning tree T of G is a tree with root s that contains a path from s to v for all reachable vertices
v. We refer to a spanning subgraph F of T as a spanning forest of G. Given a rooted tree T , we
denote by T (v) the subtree of T rooted at v (we also view T (v) as the set of descendants of v). Let
T be a tree rooted at s with vertex set VT ⊆ V , and let t(v) denote the parent of a vertex v ∈ VT
in T . If v is an ancestor of w, T [v, w] is the path in T from v to w. In particular, D[s, v] consists of
4
the vertices that dominate v. If v is a proper ancestor of w, T (v, w] is the path to w from the child
of v that is an ancestor of w. Tree T is flat if its root is the parent of every other vertex. Suppose
now that the vertex set VT of T consists of the vertices reachable from s. Tree T has the parent
property if for all (v, w) ∈ E with v and w reachable, v is a descendant of t(w) in T . If T has the
parent property and has a low-high order, then T = D [25]. For any vertex v ∈ V , we denote by
C(v) the set of children of v in D. A preorder of T is a total order of the vertices of T such that,
for every vertex v, the descendants of v are ordered consecutively, with v first. Let ζ be a preorder
of D. Consider a vertex v (cid:54)= s. We say that ζ is a low-high order for v in G, if (d(v), v) ∈ E or
there are two edges (u, v) ∈ E, (w, v) ∈ E such that u <ζ v and v <ζ w, and w is not a descendant
of v in D. Given a graph G = (V, E) and a set of edges S ⊆ V × V , we denote by G ∪ S the graph
obtained by inserting into G the edges of S.
3
Incremental low-high order
In this section we describe two algorithms to maintain a low-high order of a digraph through a
sequence of edge insertions. We first review some useful facts for updating a dominator tree after
an edge insertion [4, 23, 41]. Let (x, y) be the edge to be inserted. We consider the effect of this
insertion when both x and y are reachable. Let G(cid:48) be the flow graph that results from G after
inserting (x, y). Similarly, if D is the dominator tree of G before the insertion, we let D(cid:48) be the the
dominator tree of G(cid:48). Also, for any function f on V , we let f(cid:48) be the function after the update. We
say that vertex v is affected by the update if d(v) (its parent in D) changes, i.e., d(cid:48)(v) (cid:54)= d(v). We
let A denote the set of affected vertices. Note that we can have D(cid:48)[s, v] (cid:54)= D[s, v] even if v is not
affected. We let nca(x, y) denote the nearest common ancestor of x and y in the dominator tree D.
We also denote by depth(v) the depth of a reachable vertex v in D. There are affected vertices after
the insertion of (x, y) if and only if nca(x, y) is not a descendant of d(y) [42]. A characterization of
the affected vertices is provided by the following lemma, which is a refinement of a result in [4].
Lemma 3.1. ([23]) Suppose x and y are reachable vertices in G. A vertex v is affected after the
insertion of edge (x, y) if and only if depth(nca(x, y)) < depth(d(v)) and there is a path π in G
from y to v such that depth(d(v)) < depth(w) for all w ∈ π. If v is affected, then it becomes a child
of nca(x, y) in D(cid:48), i.e., d(cid:48)(v) = nca(x, y).
The algorithm (DBS) in [23] applies Lemma 3.1 to identify affected vertices by starting a search
from y (if y is not affected, then no other vertex is). To do this search for affected vertices, it suffices
to maintain the outgoing and incoming edges of each vertex. These sets are organized as singly
linked lists, so that a new edge can be inserted in O(1) time. The dominator tree D is represented
by the parent function d. We also maintain the depth in D of each reachable vertex. We say that
a vertex v is scanned, if the edges leaving v are examined during the search for affected vertices,
and that it is visited if there is a scanned vertex u such that (u, v) is an edge in G. By Lemma 3.1,
a visited vertex v is scanned if depth(nca(x, y)) < depth(d(v)).
Lemma 3.2. ([23]) Let v be a scanned vertex. Then v is a descendant of an affected vertex in D.
3.1 Simple Algorithm
In this algorithm we maintain, after each insertion, a subgraph H = (V, EH ) of G with O(n)
edges that has the same dominator tree as G. Then, we can compute a low-high order δ of H in
O(EH) = O(n) time. Note that δ is also a valid low-high order of G. Subgraph H is formed
by the edges of two divergent spanning trees B and R of G. After the insertion of an edge (x, y),
5
where both x and y are reachable, we form a graph H(cid:48) by inserting into H a set of edges Last(A)
found during the search for affected vertices. Specifically, Last(A) contains edge (x, y) and, for
each affected vertex v (cid:54)= y, the last edge on a path πyv that satisfies Lemma 3.1. Then, we set
H(cid:48) = H ∪ Last(A). Finally, we compute a low-high order and two divergent spanning trees of H(cid:48),
which are also valid for G(cid:48). Algorithm SimpleInsertEdge describes this process.
Algorithm 1: SimpleInsertEdge(G, D, δ, B, R, e)
Input: Flow graph G = (V, E, s), its dominator tree D, a low-high order δ of G, two
Output: Flow graph G(cid:48) = (V, E ∪ (x, y), s), its dominator tree D(cid:48), a low-high order δ(cid:48) of G(cid:48),
divergent spanning trees B and R of G, and a new edge e = (x, y).
and two divergent spanning trees B(cid:48) and R(cid:48) of G(cid:48).
1 Insert e into G to obtain G(cid:48).
2 if x is unreachable in G then return (G(cid:48), D, δ, B, R) else if y is unreachable in G then
3
(D(cid:48), δ(cid:48), B(cid:48), R(cid:48)) ← ComputeLowHigh(G(cid:48))
return (G(cid:48), D(cid:48), δ(cid:48), B(cid:48), R(cid:48))
4
5 end
6 Let H = B ∪ R.
7 Compute the updated dominator tree D(cid:48) of G(cid:48) and return a list A of the affected vertices,
8 Compute the subgraph H(cid:48) = H ∪ Last(A) of G(cid:48).
9 Compute (D(cid:48), δ(cid:48), B(cid:48), R(cid:48)) ← ComputeLowHigh(H(cid:48))
10 return (G(cid:48), D(cid:48), δ(cid:48), B(cid:48), R(cid:48))
and a list Last(A) of the last edge entering each v ∈ A in a path of Lemma 3.2.
Note that when only x is reachable before the insertion, we re-initialize our algorithm by running
the linear-time algorithm of [25, Section 6], which returns both a low-high order and two divergent
spanning trees.
Lemma 3.3. Algorithm SimpleInsertEdge is correct.
Proof. It suffices to show that subgraph H(cid:48) of G(cid:48), computed in line 9, has the same dominator tree
with G(cid:48). Note that graph H, formed by two divergent spanning trees B and R of G in line 7, has
the same dominator tree D as G. Hence, since Last(A) contains (x, y), the immediate dominator
of y is the same in H(cid:48) and in G(cid:48). Let A be the set of affected vertices in G after the insertion of
edge (x, y). Since H(cid:48) is a subgraph of G(cid:48), any vertex in V \ A has the same immediate dominator
in H(cid:48) and in G(cid:48). It remains to argue that for each vertex v ∈ A \ y, there is a path(cid:98)πyv in H(cid:48) that
vertices performed by the algorithm of [23]. We give a corresponding path (cid:98)πyv in H(cid:48). Recall that
satisfies Lemma 3.1. Let πyv be the path from y in v in G that was found by the search for affected
every vertex on πyv is scanned and that every scanned vertex is a descendant in D of an affected
vertex. We argue that for every two successive affected vertices u and w on πyv there is a path πuw
from u to w in H(cid:48) that consists of vertices of depth at least depth(w). Note that, by properties of
the depth-based search, depth(w) ≤ depth(u). Indeed, let (p, w) be the edge entering w from πyv.
Then (p, w) ∈ Last(A) and p is a descendant of u in D. Also, since u dominates p in G, u is an
ancestor of p in both spanning trees B and R. We let(cid:98)πuw = B[u, p] · (p, w). All vertices on B[u, p]
are dominated by u, since otherwise there would be a path from s to p avoiding u. So,(cid:98)πuw is path
from u to w in H(cid:48) that consists of vertices with depth at least depth(w).
Lemma 3.4. Algorithm SimpleInsertEdge maintains a low-high order of a flow graph G with n
vertices through a sequence of edge insertions in O(mn) total time, where m is the total number of
edges in G after all insertions.
6
Proof. Consider the insertion of an edge (x, y). If y was unreachable in G then we compute D, two
divergent spanning trees B and R, and a low-high order in O(m) time [25]. Throughout the whole
sequence of m insertions, such an event can happen O(n) times, so all insertions to unreachable
vertices are handled in O(mn) total time.
Now we consider the cost of executing SimpleInsertEdge. when both x and y are reachable in
G. Let ν be the number of scanned vertices, and let µ be the number of their adjacent edges. We
can update the dominator tree and locate the affected vertices (line 8) in O(ν + µ + n) time [23].
At the same time we can compute the edge set Last(A). Computing H(cid:48) in line 9 takes O(n) time
since B ∪ R ∪ Last(A) contains at most 3(n − 1) edges. Also, computing the dominator tree, two
divergent spanning trees,, and a low-high order of H(cid:48) in O(n) time [25]. So SimpleInsertEdge runs
in O(ν + µ + n) time. . The O(n) term gives a total cost of O(mn) for the whole sequence of m
insertions. We distribute the remaining O(ν + µ) cost to the scanned vertices and edges, that is
O(1) per scanned vertex or edge. Since the depth in D of every scanned vertex decreases by at
least one, a vertex and an edge can be scanned at most O(n) times. Hence, each vertex and edge
can contribute at most O(n) total cost through the whole sequence of m insertions. The O(mn)
bound follows.
3.2 Efficient Algorithm
Here we develop a more practical algorithm that maintains a low-high order δ of a flow graph
G = (V, E, s) through a sequence of edge insertions. Our algorithm uses the incremental dominators
algorithm of [23] to update the dominator tree D of G after each edge insertion. We describe a
process to update δ based on new results on the relation among vertices in D that are affected by
the insertion. . These results enable us to identify a subset of vertices for which we can compute
a "local" low-high order, that can be extended to a valid low-high order of G after the update.
We show that such a "local" low-high order can be computed by a slightly modified version of an
algorithm from [25]. We apply this algorithm on a sufficiently small flow graph that is defined by
the affected vertices, and is constructed using the concept of derived edges [46].
3.2.1 Derived edges and derived flow graphs
Derived graphs, first defined in [46], reduce the problem of finding a low-high order to the case
of a flat dominator tree [25]. By the parent property of D, if (v, w) is an edge of G, the parent
d(w) of w is an ancestor of v in D. Let (v, w) be an edge of G, with w not an ancestor of
v in D. Then, the derived edge of (v, w) is the edge (v, w), where v = v if v = d(w), v is
the sibling of w that is an ancestor of v if v (cid:54)= d(w).
If w is an ancestor of v in D, then the
derived edge of (v, w) is null. Note that a derived edge (v, w) may not be an original edge of G.
For any vertex w ∈ V such that C(w) (cid:54)= ∅, we define the derived flow graph of w, denoted by
Gw = (Vw, Ew, w), as the flow graph with start vertex w, vertex set Vw = C(w)∪{w}, and edge set
Ew = {(u, v) v ∈ Vw and (u, v) is the non-null derived edge of some edge in E}. By definition,
Gw has flat dominator tree, that is, w is the only proper dominator of any vertex v ∈ Vw \ w.
We can compute a low-high order δ of G by computing a low-high order δw in each derived flow
graph Gw. Given these low-high orders δw, we can compute a low-high order of G in O(n) time
by a depth-first traversal of D. During this traversal, we visit the children of each vertex w in
their order in δw, and number the vertices from 1 to n as they are visited. The resulting preorder
of D is low-high on G. Our incremental algorithm identifies, after each edge insertion, a specific
derived flow graph Gw for which a low-high order δw needs to be updated. Then, it uses δw to
update the low-high order of the whole flow graph G. Still, computing a low-high order of Gw can
7
be too expensive to give us the desired running time. Fortunately, we can overcome this obstacle
by exploiting a relation among the vertices that are affected by the insertion, as specified below.
This allows us to compute δw in a contracted version of Gw.
3.2.2 Affected vertices
Let (x, y) be the inserted vertex, where both x and y are reachable. Consider the execution of
algorithm DBS [23] that updates the dominator tree by applying Lemma 3.1. Suppose vertex v is
scanned, and let q be the nearest affected ancestor of v in D. Then, by Lemma 3.1, vertex q is a
child of nca(x, y) in D(cid:48), i.e., d(cid:48)(q) = nca(x, y), and v remains a descendant of q in D(cid:48).
Lemma 3.5. Let u and v be vertices such that u ∈ D(v). Then, any simple path from v to u in G
contains only vertices in D(v).
Proof. Since u ∈ D(v), v dominates u, so all paths from s to u contain v. Let πvu be a simple
path from v to u. Suppose, for contradiction, that πvu contains a vertex w (cid:54)∈ D(v). Let πwu be the
part of πvu from w to u. Since w (cid:54)∈ D(v), there is a path πsw from s to w that avoids v. But then
πsw · πwu is a path from s to u that avoids v, a contradiction.
Lemma 3.6. Let v be vertex that is affected by the insertion of (x, y), and let w be a sibling of v
in D. If there is an edge (u, w) with u a descendant of v in D then w is also affected.
Proof. Since v is affected, there is a path πyv from y to v in G that satisfies Lemma 3.1. By Lemma
3.5 and the fact that u is a descendant of v in D, there is a simple path πvu from v to u in G
that contains only vertices in D(v). Thus, πyv · πvu · (u, w) is a path from y to w that also satisfies
Lemma 3.1. Hence, w is affected.
Lemma 3.7. Let v be an ancestor of w in D, and let u be a vertex that is not a descendant of v
in D. Then any path from u to w contains v.
Proof. Let πuw be a path from u to w. Since u is not a descendant of v, there is a path πsu from s
to u that avoids v. Hence, if πuw does not contain v, then πsu · πuw is path from s to w that avoids
v, a contradiction.
Our next lemma provides a key result about the relation of the affected vertices in D.
Lemma 3.8. All vertices that are affected by the insertion of (x, y) are descendants of a common
child c of nca(x, y).
Proof. Let z = nca(x, y), and let c be the child of z that is an ancestor of y in D. We claim that
all affected vertices are descendants of c in D. Suppose, for contradiction, that there is an affected
vertex v that is not a descendant of c in D. By Lemma 3.1, v must be a descendant z in D. Also,
since the children of z are not affected, v is not a child of z. Hence, v is a proper descendant of
another child q of z in D (q (cid:54)= c). Let πyv be a path from y to v in G that satisfies Lemma 3.1.
Since y is not a descendant of q, by Lemma 3.7 path πyv must contain q. But then πyv contains a
vertex of depth depth(d(v)) or less, which contradicts Lemma 3.1.
We shall apply Lemma 3.8 to construct a flow graph GA for the affected vertices. Then, we
shall use GA to compute a "local" low-high order that we extend to a valid low-high order of G(cid:48).
8
3.2.3 Low-high order augmentation
Let δ be a low-high order of G, and let δ(cid:48) be a preorder of the dominator tree D(cid:48) of G(cid:48). We say
that δ(cid:48) agrees with δ if the following condition holds for any pair of siblings u, v in D that are not
affected by the insertion of (x, y): u <δ(cid:48) v if and only if u <δ v. Our goal is to show that there is a
low-high order δ(cid:48) of G(cid:48) that agrees with δ.
Lemma 3.9. Let δ be a low-high order of G before the insertion of (x, y). There is a preorder δ(cid:48)
of D(cid:48) that agrees with δ.
Proof. By Lemma 3.1, all affected vertices become children of z in D(cid:48). Hence, C(cid:48)(z) ⊇ C(z), and
for any v (cid:54)= z, C(cid:48)(v) ⊆ C(v). Then, for each vertex v, we can order the children of v in C(cid:48)(v) that
are not affected according to δ. Finally, we insert the affected vertices in any order in the list of
children of z. Let δ(cid:48) be the preorder of D(cid:48) that is constructed by a depth-first traversal of D(cid:48) that
visits the children of each vertex w in the order specified above. Then, δ(cid:48) agrees with δ.
Lemma 3.10. Let δ(cid:48) be a preorder of D(cid:48) that agrees with δ. Let v be a vertex that is not a child
of nca(x, y) and is not affected by the insertion of (x, y). Then δ(cid:48) is a low-high order for v in G(cid:48).
Proof. Since v is not affected, d(v) is still the parent of v in D after the insertion. So, if (d(v), v) ∈ E,
then δ(cid:48) is a low-high order for v in G(cid:48). Now suppose that (d(v), v) (cid:54)∈ E. Then there are two edges
(u, v) and (w, v) in E such that u <δ v <δ w, where w is not a descendant of v in D. Let (u, v)
and (w, v) be the derived edges of (u, v) and (w, v), respectively, in D. Then u and w are siblings
of v in D. Siblings u and w exist and are distinct by the fact that (d(v), v) (cid:54)∈ E and by the parent
property of D. Hence, u <δ v <δ w. We argue that after the insertion of (x, y), u (resp., w) remains
a sibling of v, and an ancestor of u (resp., w). If this is not the case, then there is an affected
vertex q on D[u, u]. But then, Lemma 3.6 implies that v is also be affected, a contradiction. So,
both u and w remain siblings of v in D(cid:48), and (u, v) and (w, v) remain the derived edges of (u, v)
and (w, v), respectively, in D(cid:48). Then, since δ(cid:48) agrees with δ, δ(cid:48) is a low-high order for v in G(cid:48).
We shall use Lemmata 3.1 and 3.10 to show that in order to compute a low-high order of G(cid:48),
it suffices to compute a low-high order for the derived flow graph G(cid:48)
z, where z = nca(x, y). Still,
the computation of a low-high order of G(cid:48)
z is too expensive to give us the desired running time.
Fortunately, as we show next, we can limit these computations for a contracted version of G(cid:48)
z,
defined by the affected vertices.
Let δ be a low-high order of G before the insertion of (x, y). Also, let z = nca(x, y), and let
δz be a corresponding low-high order of the derived flow graph Gz. That is, δz is the restriction of
δ to z and its children in D. Consider the child c of z that, by Lemma 3.8, is an ancestor of all
the affected vertices. Let α and β, respectively, be the predecessor and successor of c in δz. Note
z of C(cid:48)(z) ∪ {z} that results from δz
that α or β may be null. An augmentation of δz is an order δ(cid:48)
by inserting the affected vertices arbitrarily around c, that is, each affected vertex is placed in an
arbitrary position between α and c or between c and β.
z. Then, for each child v of z in D, δ(cid:48) is a low-high order for v in G(cid:48).
Lemma 3.11. Let z = nca(x, y), and let δz be a low-high order of the derived flow graph Gz before
the insertion of (x, y). Also, let δ(cid:48)
z be an augmentation of δz, and let δ(cid:48) be a preorder of D(cid:48) that
extends δ(cid:48)
Proof. Since v is a child of z in D it is not affected. Hence, d(cid:48)(v) = d(v) = z. Let G(cid:48)
flow graph of z after the insertion of (x, y). It suffices to show that δ(cid:48)
G(cid:48)
z.
z be the derived
z is a low-high order for v in
9
If (z, v) ∈ E, then (z, v) is an edge in G(cid:48)
z. So, in this case, δ(cid:48)
z is a low-high order for v in G(cid:48)
z.
Now suppose that (z, v) (cid:54)∈ E. Let δ be a preorder of D that extends δz. Then, there are two edges
(u, v) and (w, v) in G such that u <δ v <δ w, where w is not a descendant of v in D. The fact
that (z, v) is not an edge implies that u (cid:54)= z and w (cid:54)= z. Let u(cid:48) (resp., w(cid:48)) be the nearest ancestor
of u (resp., w) in D(cid:48) that is a child of z. We argue that u(cid:48) exists and satisfies u(cid:48) <δ(cid:48)
z v. Let u be
the nearest ancestor of u in D. If no vertex on D(z, u] is affected, then u(cid:48) = u. Also, since u <δz v
and by the fact that δ(cid:48)
z v. Suppose now that there is an
affected vertex q on D(z, u]. By Lemma 3.1, q becomes a child of z in D(cid:48), hence u(cid:48) = q. Also, by
Lemma 3.8, q is a proper descendant of c, so u = c. Then c <δz v, and by the construction of δ(cid:48)
we have u(cid:48) <δ(cid:48)
for v in G(cid:48)
z.
An analogous argument shows that w(cid:48) exists and satisfies v <δ(cid:48)
z is an augmentation of δz, we have u(cid:48) <δ(cid:48)
z w(cid:48). Thus, δ(cid:48)
z is a low-high order
z v.
z
3.2.4 Algorithm
Now we are ready to describe our incremental algorithm for maintaining a low-high order δ of G.
For each vertex v that is not a leaf in D, we maintain a list of its children C(v) in D, ordered by
δ. Also, for each vertex v (cid:54)= s, we keep two variables low (v) and high(v). Variable low (v) stores an
edge (u, v) such that u (cid:54)= d(v) and u <δ v; low (v) = null if no such edge exists. Similarly, high(v)
stores an edge (w, v) such that and v <δ w and w is not a descendant of v in D; high(v) = null if
no such edge exists. These variables are useful in the applications that we mention in Section 4.
Finally, we mark each vertex v such that (d(v), v) ∈ E. For simplicity, we assume that the vertices
of G are numbered from 1 to n, so we can store the above information in corresponding arrays low ,
high, and mark . Note that for a reachable vertex v, we can have low (v) = null or high(v) = null (or
both) only if mark (v) = true. Before any edge insertion, all vertices are unmarked, and all entries
in arrays low and high are null. We initialize the algorithm and the associated data structures by
executing a linear-time algorithm to compute the dominator tree D of G [3, 9] and a linear-time
algorithm to compute a low-high order δ of G [25]. So, the initialization takes O(m + n) time for
a digraph with n vertices and m edges. Next, we describe the main routine to handle an edge
insertion. We let (x, y) be the inserted edge. Also, if x and y are reachable before the insertion, we
let z = nca(x, y).
Algorithm 2: Initialize(G)
1 Compute the dominator tree D and a low-high order δ of G.
2 foreach reachable vertex v ∈ V \ s do
3
if (d(v), v) ∈ E then set mark (v) ← true find edges (u, v) and (w, v) such that
u <δ v <δ w and w (cid:54)∈ D(v)
set low (v) ← u and high(v) ← w
4
5 end
6 return (D, δ, mark , low , high)
Our main task now is to order the affected vertices according to a low-high order of D(cid:48). To do
this, we use an auxiliary flow graph GA = (VA, EA, z), with start vertex z, which we refer to as
the derived affected flow graph. Flow graph GA is essentially a contracted version of the derived
flow graph G(cid:48)
z (i.e., the derived graph of z after the insertion) as we explain later. The vertices of
the derived affected flow graph GA are z, the affected vertices of G, their common ancestor c in
D that is a child of z (from Lemma 3.8), and two auxiliary vertices α∗ and β∗. Vertex α∗ (resp.,
10
Algorithm 3: InsertEdge(G, D, δ, mark , low , high, e)
Input: Flow graph G = (V, E, s), its dominator tree D, a low-high order δ of G, arrays
Output: Flow graph G(cid:48) = (V, E ∪ (x, y), s), its dominator tree D(cid:48), a low-high order δ(cid:48) of G(cid:48),
mark , low and high, and a new edge e = (x, y).
and arrays mark(cid:48), low(cid:48) and high(cid:48).
1 Insert e into G to obtain G(cid:48).
2 if x is unreachable in G then return (G(cid:48), D, δ, mark , low , high) else if y is unreachable in
G then
3
(D(cid:48), δ(cid:48), mark(cid:48), low(cid:48), high(cid:48)) ← Initialize(G(cid:48))
return (G(cid:48), D(cid:48), δ(cid:48), mark(cid:48), low(cid:48), high(cid:48))
4
5 end
6 Compute the nearest common ancestor z of x and y in D.
7 Compute the updated dominator tree D(cid:48) of G(cid:48) and return a list A of the affected vertices.
8 foreach vertex v ∈ A do mark(cid:48)(y) ← false if z = x then mark(cid:48)(y) ← true
9 Execute DerivedLowHigh(z, A, mark(cid:48)).
10 Make a dfs traversal of the subtrees of D(cid:48) rooted at each vertex v ∈ A ∪ {c} to compute δ(cid:48).
11 foreach vertex v ∈ A ∪ {c} do
12
find edges (u, v) and (w, v) such that u <δ(cid:48) v <δ(cid:48) w and w (cid:54)∈ D(cid:48)(v)
set low(cid:48)(v) ← u and high(cid:48)(v) ← w
13
14 end
15 return (G(cid:48), D(cid:48), δ(cid:48), mark(cid:48), low(cid:48), high(cid:48))
β∗) represents vertices in C(z) with lower (resp., higher) order in δ than c. We include in GA the
edges (z, α∗) and (z, β∗). If c is marked then we include the edge (z, c) into GA, otherwise we add
the edges (α∗, c) and (β∗, c) into GA. Also, for each edge (u, c) such that u is a descendant of an
affected vertex v, we add in GA the edge (v, c). Now we specify the edges that enter an affected
vertex w in GA. We consider each edge (u, w) ∈ E entering w in G. We have the following cases:
(a) If u is a descendant of an affected vertex v, we add in GA the edge (v, w).
(b) If u is a descendant of c but not a descendant of an affected vertex, then we add in GA the
edge (c, w).
(c) If u (cid:54)= z is not a descendant of c, then we add the edge (α∗, w) if u <δ c, or the edge (β∗, w)
if c <δ u.
(d) Finally, if u = z, then we add the edge (z, w). (In cases (c) and (d), u = x and w = y.)
See Figure 3. Recall that α (resp., β) is the siblings of c in D immediately before (resp., after) c
in δ, if it exists. Then, we can obtain GA from G(cid:48)
z by contracting all vertices v with v <δ c into
α = α∗, and all vertices v with c <δ v into β = β∗.
Lemma 3.12. The derived affected flow graph GA = (VA, EA, z) has flat dominator tree.
Proof. We claim that for any two distinct vertices v, w ∈ VA\z, v does not dominate w. The lemma
follows immediately from this claim. The claim is obvious for w ∈ {α∗, β∗}, since GA contains the
edges (z, α∗) and (z, β∗). The same holds for w = c, since GA contains the edge (z, c), or both the
edges (α∗, c) and (β∗, c). Finally, suppose w ∈ VA \ {z, α∗, β∗}. Then, by the construction of GA,
vertex w is affected. By Lemma 3.8, w ∈ D(c), so Lemma 3.5 implies that there is a path in G
11
Figure 3: The derived affected flow graph GA that corresponds to the flow graph of Figure 1 after
the insertion of edge (g, d).
from c to w that contains only vertices in D(c). Hence, by construction, GA contains a path from
c to w that avoids α∗ and β∗, so α∗ and β∗ do not dominate w. It remains to show that w is not
dominated in GA by c or another affected vertex v. Let (x, y) be the inserted edge. Without loss of
generality, assume that c <δ x. Since w is affected, there is a path π in G from y to w that satisfies
Lemma 3.1. Then π does not contain any vertex in D[c, d(w)]. Also, by the construction of GA, π
corresponds to a path πA in GA from β∗ to y that avoids any vertex in A ∩ D[c, d(w)]. Hence, w
is not dominated by any vertex in A ∩ D[c, d(w)]. It remains to show that w is not dominated by
any affected vertex v in A \ D[c, d(w)]. Since both v and w are in D(c) and v is not an ancestor of
w in D, there is a path π(cid:48) in G from c to w that avoids v. By Lemma 3.5, π(cid:48) contains only vertices
in D(c). Then, by the construction of GA, π(cid:48) corresponds to a path π(cid:48)
A in GA from c to w that
avoids v. Thus, v does not dominate w in GA.
Lemma 3.13. Let ν and µ, respectively, be the number of scanned vertices and their adjacent
edges. Then, the derived affected flow graph GA has ν + 4 vertices, at most µ + 5 edges, and can be
constructed in O(ν + µ) time.
Proof. The bound on the number of vertices and edges in GA follows from the definition of the
derived affected flow graph. Next, we consider the construction time of GA. Consider the edges
entering the affected vertices. Let w be an affected vertex, and let (u, w) (cid:54)= (x, y) be an edge of G(cid:48).
Let q be nearest ancestor u in C(cid:48)(z). We distinguish two cases:
• u is not scanned. In this case, we argue that q = c. Indeed, it follows from the parent property
of D and Lemma 3.8 that both u and w are descendants of c in D. Since u is not scanned,
no ancestor of u in D is affected, so u remains a descendant of c in D(cid:48). Thus, q = c.
• u is scanned. Then, by Lemma 3.2, q is the nearest affected ancestor of u in D.
So we can construct the edges entering the affected vertices in GA in two phases. In the first phase
we traverse the descendants of each affected vertex q in D(cid:48). At each descendant u of q, we examine
the edges leaving u. When we find an edge (u, w) with w affected, then we insert into GA the edge
(q, w). In the second phase we examine the edges entering each affected vertex w. When we find
an edge (u, w) with u not visited during the first phase (i.e., u was not scanned during the update
of D), we insert into GA the edge (c, w). Note that during this construction we may insert the
same edge multiple times, but this does not affect the correctness or running time of our overall
algorithm. Since the descendants of an affected vertex are scanned, it follows that each phase runs
in O(ν + µ) time.
12
𝑠=𝑧𝑐𝐺𝐴𝑑𝛼∗𝑓𝑒𝛽∗ Finally, we need to consider the inserted edge (x, y). Let f be the nearest ancestor of x that is
in C(z). Since y is affected, c (cid:54)= f . Hence, we insert into GA the edge (β∗, y) if c <δ f , and the
edge (α∗, y) if f <δ c. Note that f is found during the computation of z = nca(x, y), so this test
takes constant time.
We use algorithm DerivedLowHigh, shown below, to order the vertices in C(cid:48)(z) according to a
low-high order of ζ of GA. After computing GA, we construct two divergent spanning trees BA and
RA of GA. For each vertex v (cid:54)= z, if (z, v) is an edge of GA, we replace the parent of v in BA and
in RA, denoted by bA(v) and rA(v), respectively, by z. Then we use algorithm AuxiliaryLowHigh
to compute a low-high order ζ of GA. Algorithm AuxiliaryLowHigh is a slightly modified version
of a linear-time algorithm of [25, Section 6.1] to compute a low-high order. Our modified version
computes a low-high order ζ of GA that is an augmentation of δz. To obtain such a low-high order,
we need to assign to α∗ the lowest number in ζ and to β∗ the highest number in ζ. The algorithm
works as follows. While GA contains at least four vertices, we choose a vertex v (cid:54)∈ {α∗, β∗} whose
in-degree in GA exceeds its number of children in BA plus its number of children in RA and remove
it from GA. (From this choice of v we also have that v (cid:54)= z.) Then we compute recursively a
low-high order for the resulting flow graph, and insert v in an appropriate location, defined by
bA(v) and rA(v).
Algorithm 4: DerivedLowHigh(z, A, mark )
1 Compute the derived affected flow graph GA = (VA, EA, z).
2 Compute two divergent spanning trees BA and RA of GA.
3 foreach vertex v ∈ VA \ {z, α∗, β∗} do
if mark (v) = true then set bA(v) ← z and rA(v) ← z
4
5 end
6 Initialize a list of vertices Λ ← ∅.
7 Compute Λ ← AuxiliaryLowHigh(GA, BA, RA, Λ).
8 Order the set of children C(cid:48)(z) of z in D(cid:48) according to Λ.
Lemma 3.14. Algorithm AuxiliaryLowHigh is correct, that is, it computes a low-high order ζ of
GA, such that for all v ∈ VA \ {z, α∗, β∗}, α∗ <ζ v <ζ β∗.
Proof. We first show that algorithm AuxiliaryLowHigh runs to completion, i.e., it selects every vertex
v ∈ VA \ {z, α∗, β∗} at some execution of line 5. The recursive call in line 13 invokes algorithm
AuxiliaryLowHigh on a sequence of smaller flow graphs GA. We claim that the following invariants
hold for each such flow graph GA:
(i) the dominator tree DA of GA is flat;
(ii) the subgraphs BA and RA corresponding to GA are divergent spanning trees of GA rooted at
z;
(iii) for every v (cid:54)= z, either bA(v) = rA(v) = z or bA(v), rA(v), and z are all distinct.
For the initial graph GA the invariants hold by construction. Assume that the invariants hold on
entry to line 5. Suppose, now, that line 5 chooses a vertex v (cid:54)∈ {α∗, β∗}. Since v has in-degree at
most 2 in GA, the choice of v implies that it has at most one outgoing edge. Hence v is a leaf in
either BA or RA. If it is a leaf in both, deleting v and its incoming edges preserves all the invariants.
Suppose v is a leaf in RA but not BA. Then v has in-degree 2 in GA; that is, bA(v) (cid:54)= rA(v), which
13
Algorithm 5: AuxiliaryLowHigh(GA, BA, RA, Λ)
1 if GA contains only three vertices then
2
set Λ ← (cid:104)α∗, β∗(cid:105)
return Λ
3
4 end
5 Let v (cid:54)∈ {α∗, β∗} be a vertex whose in-degree in GA exceeds its number of children in BA
plus its number of children in RA.
let w be the child of v in BA; replace bA(w) by bA(v)
let w be the child of v in RA; and replace rA(w) by rA(v)
6 Delete v and its incoming edges from GA, BA, and RA.
7 if v was not a leaf in BA then
8
9 end
10 else if v was not a leaf in RA then
11
12 end
13 Call AuxiliaryLowHigh(GA, BA, RA, Λ) recursively for the new graph GA.
14 if bA(v) = z then
15
16 end
17 else
18
19 end
20 return Λ
insert v anywhere between α∗ and β∗ in Λ
insert v just before bA(v) in Λ if rA(v) is before bA(v) in Λ, just after bA(v) otherwise
implies by (iii) that bA(v), rA(v), and v are distinct siblings in DA. Let w be the child of v in BA.
Since rA(w) (cid:54)= v, v, rA(w), and z are distinct by (iii). Also rA(w) (cid:54)= bA(v), since rA(w) = bA(v)
would imply that rA(w) dominates w by (ii). Finally, bA(v) (cid:54)= z, since bA(v) is a sibling of v
and hence of w in DA. We conclude that replacing bA(w) by bA(v) in line 8 preserves (iii). This
replacement preserves (i) since v does not dominate w, it preserves (ii) since it removes v from the
path in BA from s to w. Replacing bA(w) makes v a leaf in BA, after which its deletion preserves
(i)-(iii).
Now we show that the invariants imply that line 5 can always choose a vertex v. All vertices
in VA \ z are leaves in DA. Let X be the subset of VA that consists of the vertices x such that
bA(x) (cid:54)= rA(x). Each vertex in X has in-degree 2 in GA, so there are 2X edges that enter a vertex
in X. By invariant (iii), each edge leaving a vertex in X enters a vertex in X. Invariant (iii) also
implies that at least two edges enter X from VA \ X. Hence, there are at most 2(X − 1) edges
that leave a vertex in X, so there must be a vertex v in X with out-degree at most 1. We claim
that v can be selected in line 5. First note that the in-degree of v in GA exceeds its out-degree in
GA. If v is a leaf in both BA and RA then it can be selected. If not, then v must be a leaf in either
BA or RA, since otherwise its common child w in BA and RA would violate (ii). Hence v can be
selected in this case also.
Finally, we claim that the computed order is low-high for GA, such that α∗ is first and β∗ is last
in this order. The latter follows by the assignment in line 2. So the claim is immediate if GA has
three vertices. Suppose, by induction, that this is true if GA has k ≥ 3 vertices. Let GA have k + 1
vertices and let v be the vertex chosen for deletion. The insertion position of v guarantees that v
has the low-high property. All vertices in GA after the deletion of v have the low-high property
in the new GA \ z by the induction hypothesis, so they have the low-high property in the old GA
14
with the possible exception of w, one of whose incoming edges differs in the old and the new GA.
Suppose bA(w) differs; the argument is symmetric if rA(w) differs. Now we have that v, w, bA(v),
and rA(w) are distinct children of z in DA. Since w has the low-high property in the new GA, it
occurs in Λ between rA(w) and bA(v). Insertion of v next to bA(v) leaves w between rA(w) and v,
so it has the low-high property in the old GA as well.
The correctness of algorithm InsertEdge follows from Lemmata 3.10, 3.11 and 3.14.
Lemma 3.15. Algorithm InsertEdge is correct.
Proof. Let (G(cid:48), D(cid:48), δ(cid:48), mark(cid:48), low(cid:48), high(cid:48)) be the output of InsertEdge(G, D, δ, mark , low , high, e). We
only need to consider the case where both endpoints of the inserted edge e = (x, y) are reachable
in G. Let A be the set of affected vertices, and let z = nca(x, y). Also, let c be the child of z in
D that is a common ancestor of all vertices in A. We will show that the computed order δ(cid:48) is a
low-high order of G(cid:48) that agrees with δ. This fact implies that the arrays mark(cid:48), low(cid:48), high(cid:48) were
z) be the restriction of δ (resp., δ(cid:48)) to C(z)
(resp., C(cid:48)(z)). Then, by Lemma 3.14, δ(cid:48)
z is an augmentation of δz. So, by Lemmata 3.10 and 3.11,
δ(cid:48) is a low-high order in G(cid:48) for any vertex v (cid:54)∈ A ∪ {c}. Finally, Lemma 3.14 implies that δ(cid:48) is also
a low-high order in G(cid:48) for the vertices in A ∪ {c}.
updated correctly, since their entries did not change for the vertices in V \(cid:0)A ∪ {c}(cid:1).
By construction, δ(cid:48) agrees with δ. Let δz (resp., δ(cid:48)
Theorem 3.16. Algorithm InsertEdge maintains a low-high order of a flow graph G with n vertices
through a sequence of edge insertions in O(mn) total time, where m is the total number of edges in
G after all insertions.
Proof. Consider the insertion of an edge (x, y). If y was unreachable in G then we compute D and
a low-high order in O(m) time. Throughout the whole sequence of m insertions, such an event can
happen O(n) times, so all insertions to unreachable vertices are handled in O(mn) total time.
Now we consider the cost of executing InsertEdge when both x and y are reachable in G. Let ν
be the number of scanned vertices, and let µ be the number of their adjacent edges. We can update
the dominator tree and locate the affected vertices (line 8) in O(ν + µ + n) time [23]. Computing
z = nca(x, y) in line 7 takes O(n) time just by using the parent function d of D. Lines 9 -- 10 and
12 are also executed in O(n) time. The for loop in lines 13 -- 16 takes O(ν + µ) since we only need
to examine the scanned edges. (Variables low (c) and high(c) need to be updated only if there is
a scanned edge entering c.) It remains to account for time to compute GA and a low-high order
of it. From Lemma 3.13, the derived affected flow graph can be constructed in O(ν + µ) time. In
algorithm AuxiliaryLowHigh, we represent the list Λ with the off-line dynamic list maintenance data
structure of [25], which supports insertions (in a given location) and order queries in constant time.
With this implementation, AuxiliaryLowHigh runs in linear-time, that is O(ν + µ). So InsertEdge
runs in O(ν + µ + n) time. The O(n) term gives a total cost of O(mn) for the whole sequence of
m insertions. We distribute the remaining O(ν + µ) cost to the scanned vertices and edges, that
is O(1) per scanned vertex or edge. Since the depth in D of every scanned vertex decreases by at
least one, a vertex and an edge can be scanned at most O(n) times. Hence, each vertex and edge
can contribute at most O(n) total cost through the whole sequence of m insertions. The O(mn)
bound follows.
15
4 Applications of incremental low-high orders
4.1 Strongly divergent spanning trees and path queries
We can use the arrays mark , low , and high to maintain a pair of strongly divergent spanning trees,
B and R, of G after each update. Recall that B and R are strongly divergent if for every pair of
vertices v and w, we have B[s, v]∩ R[s, w] = D[s, v]∩ D[s, w] or R[s, v]∩ B[s, w] = D[s, v]∩ D[s, w].
Moreover, we can construct B and R so that they are also edge-disjoint except for the bridges of
G. A bridge of G is an edge (u, v) that is contained in every path from s to v. Let b(v) (resp.,
r(v)) denote the parent of a vertex v in B (resp., R). To update B and R after the insertion of an
edge (x, y), we only need to update b(v) and r(v) for the affected vertices v, and possibly for their
common ancestor c that is a child of z = nca(x, y) from Lemma 3.8. We can update b(v) and r(v)
of each vertex v ∈ A∪{c} as follows: set b(v) ← d(v) if low (v) = null , b(v) ← low (v) otherwise; set
r(v) ← d(v) if high(v) = null , r(v) ← high(v) otherwise. If the insertion of (x, y) does not affect y,
then A = ∅ but we may still need to update b(y) and r(y) if x (cid:54)∈ D(y) in order to make B and R
maximally edge-disjoint. Note that in this case z = d(y), so we only need to check if both low (y)
and high(y) are null. If they are, then we set low (y) ← x if x <δ y, and set high(y) ← x otherwise.
Then, we can update b(y) and r(y) as above.
Now consider a query that, given two vertices v and w, asks for two maximally vertex-disjoint
paths, πsv and πsw, from s to v and from s to w, respectively. Such queries were used in [47] to give
a linear-time algorithm for the 2-disjoint paths problem on a directed acyclic graph. If v <δ w, then
we select πsv ← B[s, v] and πsw ← R[s, w]; otherwise, we select πsv ← R[s, v] and πsw ← B[s, w].
Therefore, we can find such paths in constant time, and output them in O(πsv + πsw) time.
Similarly, for any two query vertices v and w, we can report a path πsv from s to v that avoids w.
Such a path exists if and only if w does not dominate v, which we can test in constant time using
the ancestor-descendant relation in D [44]. If w does not dominate v, then we select πsv ← B[s, v]
if v <δ w, and select πsv ← R[s, v] if w <δ v.
4.2 Fault tolerant reachability
Baswana et al. [6] study the following reachability problem. We are given a flow graph G = (V, E, s)
and a spanning tree T = (V, ET ) rooted at s. We call a set of edges E(cid:48) valid if the subgraph
G(cid:48) = (V, ET ∪ E(cid:48), s) of G has the same dominators as G. The goal is to find a valid set of minimum
cardinality. As shown in [26], we can compute a minimum-size valid set in O(m) time, given the
dominator tree D and a low-high order of δ of it. We can combine the above construction with our
incremental low-high algorithm to solve the incremental version of the fault tolerant reachability
problem, where G is modified by edge insertions and we wish to compute efficiently a valid set for
any query spanning tree T . Let t(v) be the parent of v in T . Our algorithm maintains, after each
edge insertion, a low-high order δ of G, together with the mark , low , and high arrays. Given a
query spanning tree T = (V, ET ), we can compute a valid set of minimum cardinality E(cid:48) as follows.
For each vertex v (cid:54)= s, we apply the appropriate one of the following cases: (a) If t(v) = d(v) then
we do not insert into E(cid:48) any edge entering v. (b) If t(v) (cid:54)= d(v) and v is marked then we insert
(d(v), v) into E(cid:48). (c) If v is not marked then we consider the following subcases: If t(v) >δ v, then
we insert into E(cid:48) the edge (x, v) with x = low (v). Otherwise, if t(v) <δ v, then we insert into E(cid:48)
the edge (x, v) with x = high(v). Hence, can update the minimum valid set in O(mn) total time.
We note that the above construction can be easily generalized for the case where T is forest,
i.e., when ET is a subset of the edges of some spanning tree of G. In this case, t(v) can be null
for some vertices v (cid:54)= s. To answer a query for such a T , we apply the previous construction with
the following modification when t(v) is null. If v is marked then we insert (d(v), v) into E(cid:48), as in
16
case (b). Otherwise, we insert both edges entering v from low (v) and high(v). In particular, when
ET = ∅, we compute a subgraph G(cid:48) = (V, E(cid:48), s) of G with minimum number of edges that has the
same dominators as G. This corresponds to the case k = 1 in [7].
4.3 Sparse certificate for 2-edge-connectivity
Let G = (V, E) be a strongly connected digraph. We say that vertices u, v ∈ V are 2-edge-connected
if there are two edge-disjoint directed paths from u to v and two edge-disjoint directed paths from
v to u. (A path from u to v and a path from v to u need not be edge-disjoint.) A 2-edge-connected
block of a digraph G = (V, E) is defined as a maximal subset B ⊆ V such that every two vertices
in B are 2-edge-connected. If G is not strongly connected, then its 2-edge-connected blocks are
the 2-edge-connected blocks of each strongly connected component of G. A sparse certificate for
the 2-edge-connected blocks of G is a spanning subgraph C(G) of G that has O(n) edges and
maintains the same 2-edge-connected blocks as G. Sparse certificates of this kind allow us to speed
up computations, such as finding the actual edge-disjoint paths that connect a pair of vertices (see,
e.g., [39]). The 2-edge-connected blocks and a corresponding sparse certificate can be computed
in O(m + n) time [21]. An incremental algorithm for maintaining the 2-edge-connected blocks is
presented in [24]. This algorithm maintains the dominator tree of G, with respect to an arbitrary
start vertex s, and of its reversal GR, together with the auxiliary components of G and GR, defined
next.
Recall that an edge (u, v) is a bridge of a flow graph G with start vertex s if all paths from s to
v include (u, v). After deleting from the dominator tree D the bridges of G, we obtain the bridge
decomposition of D into a forest D. For each root r of a tree in the bridge decomposition D we
define the auxiliary graph Gr = (Vr, Er) of r as follows. The vertex set Vr of Gr consists of all the
vertices in Dr. The edge set Er contains all the edges of G among the vertices of Vr, referred to as
ordinary edges, and a set of auxiliary edges, which are obtained by contracting vertices in V \ Vr,
as follows. Let v be a vertex in Vr that has a child w in V \ Vr. Note that w is a root in the bridge
decomposition D of D. For each such child w of v, we contract w and all its descendants in D into
v. The auxiliary components of G are the strongly connected components of each auxiliary graph
Gr.
We sketch how to extend the incremental algorithm of [24] so that it also maintains a sparse
certificate C(G) for the 2-edge-connected components of G, in O(mn) total time.
It suffices to
maintain the auxiliary components in G and GR, and two maximally edge-disjoint divergent span-
ning trees for each of G and GR. We can maintain these divergent spanning trees as described in
Section 4.1. To identify the auxiliary components, the algorithm of [24] uses, for each auxiliary
graph, an incremental algorithm for maintaining strongly connected components [8]. It is easy to
extend this algorithm so that it also computes O(n) edges that define these strongly connected
components. The union of these edges and of the edges in the divergent spanning trees are the
edges of C(G).
5 2-vertex-connected spanning subgraph
Let G = (V, E) be a strongly connected digraph. A vertex x of G is a strong articulation point
if G \ x is not strongly connected. A strongly connected digraph G is 2-vertex-connected if it has
at least three vertices and no strong articulation points [19, 31]. Here we consider the problem of
approximating a smallest 2-vertex-connected spanning subgraph (2VCSS) of G. This problem is
NP-hard [17]. We show that algorithm LH-Z (given below), which uses low-high orders, achieves a
linear-time 2-approximation for this problem. The best previous approximation ratio achievable in
17
Algorithm 6: LH-Z(G)
Input: 2-vertex-connected digraph G = (V, E)
Output: 2-approximation of a smallest 2-vertex-connected spanning subgraph H = (V, EH )
of G
1 Choose an arbitrary vertex s of G as start vertex.
2 Compute a strongly connected spanning subgraph H = (V \ s, EH ) of G \ s.
3 Set H ← (V, EH ).
4 Compute a low-high order δ of flow graph G with start vertex s.
5 foreach vertex v (cid:54)= s do
6
7
8
9
10
11
12
13
14
15
if there are two edges (u, v) and (w, v) in EH such that u <δ v and v <δ w then
do nothing
end
else if there is no edge (u, v) ∈ EH such that u <δ v then
find an edge e = (u, v) ∈ E with u <δ v
set EH ← EH ∪ {e}
end
else if there is no edge (w, v) ∈ EH such that v <δ w then
find an edge e = (w, v) ∈ E with v <δ w or w = s
set EH ← EH ∪ {e}
end
16
17 end
18 Execute the analogous steps of lines 4 -- 17 for the reverse flow graph GR with start vertex s.
19 return H = (V, EH )
linear-time was 3 [20], so we obtain a substantial improvement. The best approximation ratio for
√
2VCSS is 3/2, and is achieved by the algorithm of Cheriyan and Thurimella [10] in O(m2) time,
n + n2) by a combination of [10] and [20]. Computing small spanning subgraphs is of
or in O(m
particular importance when dealing with large-scale graphs, e.g., with hundreds of million to billion
edges. In this framework, one big challenge is to design linear-time algorithms, since algorithms
with higher running times might be practically infeasible on today's architectures. Let G = (V, E)
be a strongly connected digraph. In the following, we denote by GR = (V, ER) the reverse digraph
of G that results from G after reversing all edge directions.
Lemma 5.1. Algorithm LH-Z computes a 2-vertex-connected spanning subgraph of G.
Proof. We need to show that the computed subgraph H is 2-vertex-connected. From [31], we have
that a digraph H is 2-vertex connected if and only if it satisfies the following property: For an
arbitrary start vertex s ∈ V , flow graphs H = (V, E, s) and H R = (V, ER, s) have flat dominator
trees, and H \ s is strongly connected. The digraph H computed by algorithm LH-Z satisfies the
latter condition because of line 2. It remains to show that H has flat dominator tree. The same
argument applies for H R, thus completing the proof. Let δ be the low-high order δ of G, computed
in line 3. We argue that after the execution of the for loop in lines 5 -- 17, δ is also a low-high order
for all vertices in H. Consider an arbitrary vertex v (cid:54)= s. Let (x, v) be an edge entering v in the
strongly connected spanning subgraph of G computed in line 2. If x >δ v, then, by the definition of
δ, there is at least one edge (y, v) ∈ E such that y <δ v. Hence, after the execution of the for loop
for v, the edge set EH will contain at least two edges (u, v) and (w, v) such that u <δ v <δ w. On
the other hand, if x <δ v, then the definition of δ implies that there an edge (y, v) ∈ E such that
18
Graph
Largest SCC
n
m avg. δ
rome99
twitter-higgs-retweet
enron
web-NotreDame
3352
13086
8271
48715
8855
63537
147353
267647
soc-Epinions1
Amazon-302
WikiTalk
web-Stanford
web-Google
Amazon-601
web-BerkStan
32220
241761
111878
150475
434818
395230
334857
442768
1131217
1477665
1576157
3419124
3301051
4523232
2.64
4.86
17.82
5.49
13.74
4.68
13.21
10.47
7.86
8.35
13.51
n
2249
1099
4441
1409
1462
1416
17117
55414
49430
5179
10893
77480
276049
1106
4927
12795
29145
2VCCs
Type
m avg. δ
6467
9290
123527
6856
7279
13226
395183
241663
1254898
129897
162295
840829
2461072
8206
28142
347465
439148
2.88
8.45
27.82
4.87
4.98
9.34
23.09
4.36
road network
twitter
enron mails
web
trust network
co-purchase
web
25.39 Wiki communications
25.08
14.90
10.85
8.92
7.42
5.71
27.16
15.07
web
co-purchase
web
Table 1: Real-world graphs used in the experiments, sorted by the file size of their largest SCC. We
used both the largest SCC and the some of the 2VCCs (inside the largest SCC) in our experiments.
y >δ v or y = s. Notice that in either case y (cid:54)= x. So, again, after the execution of the for loop for
v, the edge set EH will contain at least two edges (u, v) and (w, v) such that either u <δ v <δ w,
or u <δ v and w = s. It follows that δ is a low-high order for all vertices v (cid:54)= s in H. By [25],
this means that H contains two strongly divergent spanning trees B and R of G. Since G has flat
dominator tree, we have that B[s, v] ∩ R[s, v] = {s, v} for all v ∈ V \ s. Hence, since H contains B
and R, the dominator tree of H is flat.
We remark that the construction of H in algorithm LH-Z guarantees that s will have in-degree
and out-degree at least 2 in H. (This fact is implicit in the proof of Lemma 5.1.) Indeed, H will
contain the edges from s to the vertices in V \ s with minimum and maximum order in δ, and the
edges entering s from the vertices in V \ s with minimum and maximum order in δR.
Theorem 5.2. Algorithm LH-Z computes a 2-approximation for 2VCSS in linear time.
Proof. We establish the approximation ratio of LH-Z by showing that EH ≤ 4n. The approxi-
mation ratio of 2 follows from the fact that any vertex in a 2-vertex-connected digraph must have
in-degree at least two. In line 2 we can compute an approximate smallest strongly connected span-
ning subgraph of G \ s [34]. For this, we can use the linear-time algorithm of Zhao et al. [49],
which selects at most 2(n − 1) edges. Now consider the edges selected in the for loop of lines 5 -- 17.
Since after line 2, H \ s is strongly connected, each vertex v ∈ V \ s has at least one entering edge
(x, v). If x <δ v then lines 10 -- 11 will not be executed; otherwise, v <δ x and lines 14 -- 15 will not
be executed. Thus, the for loop of lines 5 -- 17 adds at most one edge entering each vertex v (cid:54)= s.
The same argument implies that the analogous steps executed for GR add at most one edge leaving
each vertex v (cid:54)= s. Hence, EH contains at most 4(n − 1) at the end of the execution.
6 Empirical Analysis
For the experimental evaluation we use the graph datasets shown in Table 1. We wrote our
implementations in C++, using g++ v.4.6.4 with full optimization (flag -O3) to compile the code.
We report the running times on a GNU/Linux machine, with Ubuntu (12.04LTS): a Dell PowerEdge
19
Figure 4: Incremental low-high order: dynamized 2VC graphs (top) and edge insertion in strongly
connected graphs (bottom). Running times, in seconds, and number of edges both shown in loga-
rithmic scale.
R715 server 64-bit NUMA machine with four AMD Opteron 6376 processors and 128GB of RAM
memory. Each processor has 8 cores sharing a 16MB L3 cache, and each core has a 2MB private
L2 cache and 2300MHz speed. In our experiments we did not use any parallelization, and each
algorithm ran on a single core. We report CPU times measured with the getrusage function,
averaged over ten different runs. In Table 1 we can see some statistics about the real-world graphs
we used in our experimental evaluation.
6.1 Incremental low-high order.
We compare the performance of four algorithms. As a baseline, we use a static low-high order
algorithm from [25] based on an efficient implementation of the Lengauer-Tarjan algorithm for
computing dominators [35] from [27]. Our baseline algorithm, SLT, constructs, as intermediary,
two divergent spanning trees. After each insertion of an edge (x, y), SLT recomputes a low-high
order if x is reachable. An improved version of this algorithm, that we refer to as SLT-NCA, tests
if the insertion of (x, y) affects the dominator tree by computing the nearest common ancestor
If this is the case, then SLT-NCA recomputes a low-high order as SLT. The other
of x and y.
two algorithms are the ones we presented in Section 3. For our simple algorithm, DBS-DST, we
extend the incremental dominators algorithm DBS of [23] with the computation of two divergent
spanning trees and a low-high order, as in SLT. Algorithm DBS-DST applies these computations
on a sparse subgraph of the input digraph that maintains the same dominators. Finally, we tested
an implementation of our more efficient algorithm, DBS-AUX, that updates the low-high order by
computing a local low-high order of an auxiliary graph.
We compared the above incremental low-high order algorithms in two different field tests. In
the first one, we considered 2-vertex connected graphs, and we dynamized them in the following
manner: we removed a percentage of edges (i.e., 5%, 10%, and 20% respectively), selected uni-
formly at random, that were incrementally added to the graph. Note that during the execution
of the algorithms some vertices may be unreachable at first. Also, at the end of all insertions,
the final graph has flat dominator tree. In Figure 4 (top) we can see that the algorithms are well
20
0.0010.010.11101001.00010.000100.0001+e41+e51+e6SLTSLT-NCADBS-AUXDBS-DST0.0010.010.11101001.00010.000100.0001+e41+e51+e6SLTSLT-NCADBS-AUXDBS-DST Figure 5: Smallest 2-vertex-connected spanning subgraph. Top: running times, in seconds, and
number of edges both shown in logarithmic scale. Bottom: relative size of the resulting 2VCSS.
distinguished: our DBS-AUX performs consistently better than the other ones (with the exception
of two NotreDame instances). The total running times are given in Table 2. On average, DBS-DST
is about 2.84 times faster than SLT-NCA, with their relative performance depending on the density
of the graph (the higher the average degree the better DBS-DST performs w.r.t. SLT-NCA.) As
we mentioned, the naive SLT is the worst performer. The above observed behavior of the algo-
rithms is similar also in the second test. Here, we consider the strongly connected graphs, and
we incrementally insert random edges up to a certain percentage of the original number of edges
(i.e., as before, 5%, 10%, and 20% respectively). We use strongly connected graphs only in order
to guarantee that all vertices are reachable from the selected source. (Strong connectivity has no
other effect in these tests.) The endpoints of each new edge are selected uniformly at random, and
the edge is inserted if it is not a loop and is not already present in the current graph. The ranking
of the algorithms does not change, as we can see in Figure 4 (bottom), but the difference is bigger:
we note a bigger gap of more than two orders of magnitude, in particular, between DBS-AUX and
the couple SLT-NCA and DBS-DST. This is expected because, unlike the first test, here all edges
connect already reachable vertices. This means that DBS-DST and DBS-AUX do not execute a full
restart for any of these insertions. The total running times are given in Table 3.
6.2 2-vertex-connected spanning subgraph.
In this experimental evaluation we compared four algorithms for computing the (approximated)
smallest 2-vertex-connected spanning subgraph. Specifically, we tested two algorithms from [20],
FAST which computes a 3-approximation in linear-time by using divergent spanning trees, and
FAST-CT which combines FAST with the 3/2-approximation algorithm of Cheriyan and Thurimella [10].
In the experiments reported in [20], the former algorithm achieved the fastest running times, while
the latter the best solution quality. We compare these algorithms against our new algorithm LH-Z
of Section 5, and a new hybrid algorithm LH-Z-CT, that combines LH-Z with the algorithm of
Cheriyan and Thurimella [10].
Algorithm LH-Z-CT works as follows. First, it computes a 1-matching M in the input graph
G [16], using bipartite matching as in [10]. Let H be the subgraph of G\s, for arbitrary start vertex
s, that contains only the edges in M . We compute the strongly connected components C1, . . . , Ck
21
0.0010.010.11101001.00010.000100.0001.000.0001+e41+e51+e6FASTLH-ZLH-Z-CTFAST-CT1.01.21.41.61.82.01+e41+e51+e6FASTLH-ZLH-Z-CTFAST-CT Graph
rome05
rome10
rome20
twitter05
twitter10
twitter20
NotreDame05
NotreDame10
NotreDame20
enron05
enron10
enron20
webStanford05
webStanford10
webStanford20
Amazon05
Amazon10
Amazon20
WikiTalk05
WikiTalk10
WikiTalk20
nodes
2249
2249
2249
1099
1099
1099
1416
1416
1416
4441
4441
4441
5179
5179
5179
55414
55414
55414
49430
49430
49430
starting edges
6144
5820
5174
8826
8361
7432
12565
11903
10581
117351
111174
98822
123402
116907
103918
229580
217497
193330
1192153
1129408
1003918
final edges
6467
6467
6467
9290
9290
9290
13226
13226
13226
123527
123527
123527
129897
129897
129897
241663
241663
241663
1254898
1254898
1254898
SLT SLT-NCA DBS-AUX DBS-DST
0.16457
0.319025
0.639627
0.012953
0.027945
0.070981
0.003421
0.003341
0.003374
0.152272
0.388753
1.40979
0.370119
0.439147
0.329067
26.5169
18.4656
26.7402
20.2353
45.7164
99.6058
0.120026
0.231678
0.646389
0.051123
0.085982
0.198226
0.003981
0.003997
0.01446
0.811483
1.5252
3.08813
0.936905
1.02925
1.04364
37.4155
41.4565
44.8627
113.946
247.601
500.581
0.060632
0.242326
0.231463
0.004682
0.006498
0.018012
0.002727
0.005094
0.004571
0.033019
0.204307
0.999617
0.236135
0.316648
0.323679
8.91418
8.80395
8.66914
4.7007
17.5997
45.9101
0.216091
0.734963
0.772791
0.320313
0.996879
2.06744
0.012942
0.012958
0.019733
51.5453
109.719
158.83
42.1674
51.2838
51.6162
185.868
214.185
230.7
15026.2
24846.2
51682
Table 2: Running times of the plot shown in Figure 4 (top) .
Graph
rome05
rome10
rome20
twitter05
twitter10
twitter20
enron05
enron10
enron20
NotreDame05
NotreDame10
NotreDame20
Amazon05
Amazon10
Amazon20
WikiTalk05
WikiTalk10
WikiTalk20
webStanford05
webStanford10
webStanford20
nodes
3352
3352
3352
13086
13086
13086
8271
8271
8271
48715
48715
48715
241761
241761
241761
111878
111878
111878
150475
150475
150475
starting edges
8855
8855
8855
63537
63537
63537
147353
147353
147353
267647
267647
267647
1131217
1131217
1131217
1477665
1477665
1477665
1576157
1576157
1576157
final edges
9298
9741
10626
66714
69891
76244
154721
162088
176824
281029
294412
321176
1187778
1244339
1357460
1551548
1625432
1773198
1654965
1733773
1891388
SLT SLT-NCA DBS-AUX DBS-DST
0.07962
0.147052
0.259299
4.2933
8.73372
11.1431
1.72764
3.97011
4.94514
234.757
359.028
585.807
3094.69
5628.51
9157.5
1157.84
1666.28
2124.6
2295.55
3749.12
5381.12
0.185072
0.366822
0.410448
9.94862
25.189
31.7239
11.9889
25.0999
20.1978
375.356
610.29
1168.5
6386.97
11905.7
15871
3414.28
5301.51
7296.72
8403.03
11503.4
15792.1
0.002457
0.005531
0.008154
0.073755
0.120719
0.186917
0.068994
0.09557
0.106017
0.70628
0.79135
1.31932
26.5493
45.1881
60.197
10.3364
14.5151
19.5778
7.028
13.7287
12.7093
0.662001
1.06533
2.74201
26.5965
55.4924
96.3205
82.6084
180.222
353.174
785.012
1691.05
3593.09
>24h
>24h
>24h
>24h
>24h
>24h
>24h
>24h
>24h
Table 3: Running times of the plot shown in Figure 4 (bottom) .
22
Graph
Size
FAST
LH-Z
LH-Z-CT
FAST-CT
rome99
web-NotreDame1
web-NotreDame2
web-BerkStan1
twitter-higgs-retweet
web-NotreDame3
web-BerkStan2
web-Stanford
Amazon-302
web-BerkStan3
web-Google
WikiTalk
Amazon-601
nodes
2249
1409
1462
1106
1099
1416
4927
5179
55414
12795
77480
49430
276049
edges
6467
6856
7279
8206
9290
13226
28142
129897
241663
347465
840829
1254898
2461072
time
0.003581
0.002697
0.002784
0.001753
0.000885
0.003182
0.009555
0.022056
0.11804
0.055141
0.182113
0.338172
0.977274
edges
5691
3796
3949
3423
3553
4687
13391
17940
164979
45111
256055
176081
932989
time
0.004513
0.001211
0.005436
0.00393
0.005976
0.005515
0.018985
0.031001
0.164081
0.049862
0.401427
0.548573
1.607812
edges
5370
3356
3545
2795
3143
3990
12296
14583
141467
36328
221327
161128
744345
time
2.099891
0.927683
0.999031
0.525704
1.087169
0.751014
11.223287
22.141856
2986.022813
149.561649
7668.066338
5883.002974
140894.0119
edges
4837
3029
3189
2433
2745
3560
10646
11556
123095
29307
191616
138030
612760
time
5.042213
2.119423
2.262008
0.949573
1.153608
2.143551
23.887951
31.346059
3935.135495
203.913794
20207.08225
4770.705853
298775.2092
edges
5057
3153
3300
2440
2879
3768
11750
12920
132847
32989
211529
143958
688159
Table 4: Running times and number of edges in the resulting 2-vertex-connected spanning subgraph;
plots shown in Figure 5 .
in H, and form a contracted version G(cid:48) of G as follows. For each strongly connected component
Ci of H, we contract all vertices in Ci into a representative vertex ui ∈ Ci. (Contractions are
performed by union-find [45] and merging lists of out-edges of G.) Then, we execute the linear-
time algorithm of Zhao et al. [49] to compute a strongly connected spanning subgraph of G(cid:48), and
store the original edges of G that correspond to the selected edges by the Zhao et al. algorithm.
Let Z be this set of edges. We compute a low-high order of G with root s, and use it in order to
compute a 2-vertex-connected spanning subgraph W of G using as many edges from Z and M as
possible, as in LH-Z. Then, we run the filtering phase of Cheriyan and Thurimella. For each edge
(x, y) of W that is not in M , we test if x has two vertex-disjoint paths to y in W \ (x, y). If it
does, then we set W ← W \ (x, y). We remark that, similarly to FAST-CT, LH-Z preserves the 3/2
√
approximation guarantee of the Cheriyan-Thurimella algorithm for k = 2 and improves its running
time from O(m2) to O(m
n+n2), for a digraph with n vertices and m arcs. In our implementation,
the bipartite matching is computed via max-flow, using an implementation of the Goldberg-Tarjan
push-relabel algorithm [28] from [13], which is very fast in practice. (This implementation was
provided by the authors of [13].)
In Figure 5 (top) we can see the running times of the four algorithms. (See also Table 4.) It is
easy to observe that the algorithms belong to two distinct classes, with FAST and LH-Z being faster
than the other two by approximately five orders of magnitude. In the bottom part of Figure 5 we
can see the relative size of the smallest spanning subgraph computed by the four algorithms. In
all of our experiments, the smallest subgraph was the one computed by our new hybrid algorithm
LH-Z-CT. One the other hand, on average LH-Z is only twice as slow as FAST but improves the
solution quality by more than 13%. Summing up, if one wants a fast and good solution LH-Z is the
right choice.
Acknowledgement. We would like to thank Bob Tarjan for valuable comments and suggestions.
References
[1] A. Abboud and V. Vassilevska Williams. Popular conjectures imply strong lower bounds for
In Proc. 55th IEEE Symposium on Foundations of Computer Science,
dynamic problems.
FOCS, pages 434 -- 443, 2014.
23
[2] S. Allesina and A. Bodini. Who dominates whom in the ecosystem? Energy flow bottlenecks
and cascading extinctions. Journal of Theoretical Biology, 230(3):351 -- 358, 2004.
[3] S. Alstrup, D. Harel, P. W. Lauridsen, and M. Thorup. Dominators in linear time. SIAM
Journal on Computing, 28(6):2117 -- 32, 1999.
[4] S. Alstrup and P. W. Lauridsen. A simple dynamic algorithm for maintaining a dominator
tree. Technical Report 96-3, Department of Computer Science, University of Copenhagen,
1996.
[5] M. E. Amyeen, W. K. Fuchs, I. Pomeranz, and V. Boppana. Fault equivalence identification
using redundancy information and static and dynamic extraction. In Proceedings of the 19th
IEEE VLSI Test Symposium, March 2001.
[6] S. Baswana, K. Choudhary, and L. Roditty. Fault tolerant reachability for directed graphs.
In Yoram Moses, editor, Distributed Computing, volume 9363 of Lecture Notes in Computer
Science, pages 528 -- 543. Springer Berlin Heidelberg, 2015.
[7] S. Baswana, K. Choudhary, and L. Roditty. Fault tolerant reachability subgraph: Generic and
optimal. In Proc. 48th ACM Symp. on Theory of Computing, pages 509 -- 518, 2016. To appear.
[8] M. A. Bender, J. T. Fineman, S. Gilbert, and R. E. Tarjan. A new approach to incremental
cycle detection and related problems. ACM Transactions on Algorithms, 12(2):14:1 -- 14:22,
December 2015.
[9] A. L. Buchsbaum, L. Georgiadis, H. Kaplan, A. Rogers, R. E. Tarjan, and J. R. Westbrook.
Linear-time algorithms for dominators and other path-evaluation problems. SIAM Journal on
Computing, 38(4):1533 -- 1573, 2008.
[10] J. Cheriyan and R. Thurimella. Approximating minimum-size k-connected spanning subgraphs
via matching. SIAM J. Comput., 30(2):528 -- 560, 2000.
[11] S. Cicerone, D. Frigioni, U. Nanni, and F. Pugliese. A uniform approach to semi-dynamic
problems on digraphs. Theor. Comput. Sci., 203:69 -- 90, August 1998.
[12] R. Cytron, J. Ferrante, B. K. Rosen, M. N. Wegman, and F. K. Zadeck. Efficiently comput-
ing static single assignment form and the control dependence graph. ACM Transactions on
Programming Languages and Systems, 13(4):451 -- 490, 1991.
[13] D. Delling, A. V. Goldberg, I. Razenshteyn, and R. F. Werneck. Graph partitioning with nat-
ural cuts. In 25th International Parallel and Distributed Processing Symposium (IPDPS'11),
2011.
[14] W. Fraczak, L. Georgiadis, A. Miller, and R. E. Tarjan. Finding dominators via disjoint set
union. Journal of Discrete Algorithms, 23:2 -- 20, 2013.
[15] H. N. Gabow. The minset-poset approach to representations of graph connectivity. ACM
Transactions on Algorithms, 12(2):24:1 -- 24:73, February 2016.
[16] H. N. Gabow and R. E. Tarjan. Faster scaling algorithms for general graph matching problems.
J. ACM, 38:815 -- 853, 1991.
[17] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of
NP-Completeness. W. H. Freeman & Co., New York, NY, USA, 1979.
24
[18] K. Gargi. A sparse algorithm for predicated global value numbering. SIGPLAN Not., 37(5):45 --
56, May 2002.
[19] L. Georgiadis. Testing 2-vertex connectivity and computing pairs of vertex-disjoint s-t paths
In Proc. 37th Int'l. Coll. on Automata, Languages, and Programming, pages
in digraphs.
738 -- 749, 2010.
[20] L. Georgiadis. Approximating the smallest 2-vertex connected spanning subgraph of a directed
graph. In Proc. 19th European Symposium on Algorithms, pages 13 -- 24, 2011.
[21] L. Georgiadis, G. F. Italiano, L. Laura, and N. Parotsidis. 2-edge connectivity in directed
graphs. In Proc. 26th ACM-SIAM Symp. on Discrete Algorithms, pages 1988 -- 2005, 2015.
[22] L. Georgiadis, G. F. Italiano, L. Laura, and N. Parotsidis. 2-vertex connectivity in directed
graphs. In Proc. 42nd Int'l. Coll. on Automata, Languages, and Programming, pages 605 -- 616,
2015.
[23] L. Georgiadis, G. F. Italiano, L. Laura, and F. Santaroni. An experimental study of dynamic
In Proc. 20th European Symposium on Algorithms, pages 491 -- 502, 2012. Full
dominators.
version: CoRR, abs/1604.02711.
[24] L. Georgiadis, G. F. Italiano, and N. Parotsidis. Incremental 2-edge-connectivity in directed
In Proc. 43nd Int'l. Coll. on Automata, Languages, and Programming, 2016. To
graphs.
appear.
[25] L. Georgiadis and R. E. Tarjan. Dominator tree certification and divergent spanning trees.
ACM Transactions on Algorithms, 12(1):11:1 -- 11:42, November 2015.
[26] L. Georgiadis and R. E. Tarjan. Addendum to "Dominator tree certification and divergent
spanning trees". ACM Transactions on Algorithms, 12(4):56:1 -- 56:3, August 2016.
[27] L. Georgiadis, R. E. Tarjan, and R. F. Werneck. Finding dominators in practice. Journal of
Graph Algorithms and Applications (JGAA), 10(1):69 -- 94, 2006.
[28] A. V. Goldberg and R. E. Tarjan. A new approach to the maximum-flow problem. Journal of
the ACM, 35:921 -- 940, October 1988.
[29] M. Gomez-Rodriguez and B. Scholkopf. Influence maximization in continuous time diffusion
networks. In 29th International Conference on Machine Learning (ICML), 2012.
[30] M. Henzinger, S. Krinninger, and V. Loitzenbauer. Finding 2-edge and 2-vertex strongly
connected components in quadratic time. In Proc. 42nd Int'l. Coll. on Automata, Languages,
and Programming, pages 713 -- 724, 2015.
[31] G. F. Italiano, L. Laura, and F. Santaroni. Finding strong bridges and strong articulation
points in linear time. Theoretical Computer Science, 447:74 -- 84, 2012.
[32] R. Jaberi. Computing the 2-blocks of directed graphs. RAIRO-Theor. Inf. Appl., 49(2):93 -- 119,
2015.
[33] R. Jaberi. On computing the 2-vertex-connected components of directed graphs. Discrete
Applied Mathematics, 204:164 -- 172, 2016.
25
[34] S. Khuller, B. Raghavachari, and N. E. Young. Approximating the minimum equivalent di-
graph. SIAM J. Comput., 24(4):859 -- 872, 1995. Announced at SODA 1994, 177-186.
[35] T. Lengauer and R. E. Tarjan. A fast algorithm for finding dominators in a flowgraph. ACM
Transactions on Programming Languages and Systems, 1(1):121 -- 41, 1979.
[36] E. K. Maxwell, G. Back, and N. Ramakrishnan. Diagnosing memory leaks using graph min-
ing on heap dumps. In Proceedings of the 16th ACM SIGKDD international conference on
Knowledge discovery and data mining, KDD '10, pages 115 -- 124, 2010.
[37] R. M. McConnell, K. Mehlhorn, S. Naher, and P. Schweitzer. Certifying algorithms. Computer
Science Review, 5(2):119 -- 161, 2011.
[38] M. Mowbray and A. Lain. Dominator-tree analysis for distributed authorization. In Proceedings
of the Third ACM SIGPLAN Workshop on Programming Languages and Analysis for Security,
PLAS '08, pages 101 -- 112, New York, NY, USA, 2008. ACM.
[39] H. Nagamochi and T. Ibaraki. A linear-time algorithm for finding a sparse k-connected span-
ning subgraph of a k-connected graph. Algorithmica, 7:583 -- 596, 1992.
[40] L. Quesada, P. Van Roy, Y. Deville, and R. Collet. Using dominators for solving constrained
In Proc. 8th International Conference on Practical Aspects of Declarative
path problems.
Languages, pages 73 -- 87, 2006.
[41] G. Ramalingam and T. Reps. An incremental algorithm for maintaining the dominator tree
In Proc. 21st ACM SIGPLAN-SIGACT Symp. on Principles of
of a reducible flowgraph.
Programming Languages, pages 287 -- 296, 1994.
[42] G. Ramalingam and T. Reps. An incremental algorithm for maintaining the dominator tree
of a reducible flowgraph. In Proceedings of the 21st ACM SIGPLAN-SIGACT symposium on
Principles of programming languages, pages 287 -- 296, 1994.
[43] V. C. Sreedhar, G. R. Gao, and Y. Lee. Incremental computation of dominator trees. ACM
Transactions on Programming Languages and Systems, 19:239 -- 252, 1997.
[44] R. E. Tarjan. Finding dominators in directed graphs. SIAM Journal on Computing, 3(1):62 -- 89,
1974.
[45] R. E. Tarjan. Efficiency of a good but not linear set union algorithm. Journal of the ACM,
22(2):215 -- 225, 1975.
[46] R. E. Tarjan. Fast algorithms for solving path problems. Journal of the ACM, 28(3):594 -- 614,
1981.
[47] T. Tholey. Linear time algorithms for two disjoint paths problems on directed acyclic graphs.
Theoretical Computer Science, 465:35 -- 48, 2012.
[48] J. Zhao and S. Zdancewic. Mechanized verification of computing dominators for formalizing
In Proc. 2nd International Conference on Certified Programs and Proofs, pages
compilers.
27 -- 42. Springer, 2012.
[49] L. Zhao, H. Nagamochi, and T. Ibaraki. A linear time 5/3-approximation for the minimum
strongly-connected spanning subgraph problem. Information Processing Letters, 86(2):63 -- 70,
2003.
26
|
1704.04163 | 3 | 1704 | 2019-01-03T15:37:07 | Spectrum Approximation Beyond Fast Matrix Multiplication: Algorithms and Hardness | [
"cs.DS",
"cs.LG",
"math.NA"
] | Understanding the singular value spectrum of a matrix $A \in \mathbb{R}^{n \times n}$ is a fundamental task in countless applications. In matrix multiplication time, it is possible to perform a full SVD and directly compute the singular values $\sigma_1,...,\sigma_n$. However, little is known about algorithms that break this runtime barrier.
Using tools from stochastic trace estimation, polynomial approximation, and fast system solvers, we show how to efficiently isolate different ranges of $A$'s spectrum and approximate the number of singular values in these ranges. We thus effectively compute a histogram of the spectrum, which can stand in for the true singular values in many applications.
We use this primitive to give the first algorithms for approximating a wide class of symmetric matrix norms in faster than matrix multiplication time. For example, we give a $(1 + \epsilon)$ approximation algorithm for the Schatten-$1$ norm (the nuclear norm) running in just $\tilde O((nnz(A)n^{1/3} + n^2)\epsilon^{-3})$ time for $A$ with uniform row sparsity or $\tilde O(n^{2.18} \epsilon^{-3})$ time for dense matrices. The runtime scales smoothly for general Schatten-$p$ norms, notably becoming $\tilde O (p \cdot nnz(A) \epsilon^{-3})$ for any $p \ge 2$.
At the same time, we show that the complexity of spectrum approximation is inherently tied to fast matrix multiplication in the small $\epsilon$ regime. We prove that achieving milder $\epsilon$ dependencies in our algorithms would imply faster than matrix multiplication time triangle detection for general graphs. This further implies that highly accurate algorithms running in subcubic time yield subcubic time matrix multiplication. As an application of our bounds, we show that precisely computing all effective resistances in a graph in less than matrix multiplication time is likely difficult, barring a major algorithmic breakthrough. | cs.DS | cs | Spectrum Approximation Beyond Fast Matrix Multiplication:
Algorithms and Hardness
Cameron Musco
MIT
Praneeth Netrapalli
Microsoft Research
Aaron Sidford
Stanford University
[email protected]
[email protected]
[email protected]
Shashanka Ubaru
David P. Woodruff
University of Minnesota
Carnegie Mellon University
[email protected]
[email protected]
Abstract
Understanding the singular value spectrum of a matrix A ∈ Rn×n is a fundamental task in
countless numerical computation and data analysis applications. In matrix multiplication time,
it is possible to perform a full SVD of A and directly compute the singular values σ1, . . . , σn.
However, little is known about algorithms that break this runtime barrier.
Using tools from stochastic trace estimation, polynomial approximation, and fast linear sys-
tem solvers, we show how to efficiently isolate different ranges of A's spectrum and approximate
the number of singular values in these ranges. We thus effectively compute an approximate
histogram of the spectrum, which can stand in for the true singular values in many applications.
We use our histogram primitive to give the first algorithms for approximating a wide class
of symmetric matrix norms and spectral sums faster than the best known runtime for matrix
multiplication. For example, we show how to obtain a (1 + ǫ) approximation to the Schatten
1-norm (i.e. the nuclear or trace norm) in just O((nnz(A)n1/3 +n2)ǫ−3) time for A with uniform
row sparsity or O(n2.18ǫ−3) time for dense matrices. The runtime scales smoothly for general
Schatten p-norms, notably becoming O(p nnz(A)ǫ−3) for any real p ≥ 2.
At the same time, we show that the complexity of spectrum approximation is inherently
tied to fast matrix multiplication in the small ǫ regime. We use fine-grained complexity to
give conditional lower bounds for spectrum approximation, showing that achieving milder ǫ
dependencies in our algorithms would imply triangle detection algorithms for general graphs
running in faster than state of the art matrix multiplication time. This further implies, through
a reduction of [WW10], that highly accurate spectrum approximation algorithms running in
subcubic time can be used to give subcubic time matrix multiplication. As an application of
our bounds, we show that precisely computing all effective resistances in a graph in less than
matrix multiplication time is likely difficult, barring a major algorithmic breakthrough.
9
1
0
2
n
a
J
3
]
S
D
.
s
c
[
3
v
3
6
1
4
0
.
4
0
7
1
:
v
i
X
r
a
1
Introduction
Given A ∈ Rn×d, a central primitive in numerical computation and data analysis is to compute A's
spectrum: the singular values σ1(A) ≥ . . . ≥ σd(A) ≥ 0. These values can reveal matrix structure
and low effective dimensionality, which can be exploited in a wide range of spectral data analysis
methods [Jol02, US16]. The singular values are also used as tuning parameters in many numerical
algorithms performed on A [GVL12], and in general, to determine some of the most well-studied
matrix functions [Hig08]. For example, for any f : R+ → R+, we can define the spectral sum:
def
=
Sf (A)
dXi=1
f (σi(A)).
Spectral sums often serve as snapshots of A's spectrum and are important in many applications.
They encompass, for example, the log-determinant, the trace inverse, the Schatten p-norms, in-
cluding the important nuclear norm, and general Orlicz norms (see Section 1.2 for details).
While the problem of computing a few of the largest or smallest singular values of A has been
exhaustively studied [Par98, Saa11], much less is known about algorithms that approximate the
full spectrum, and in particular, allow for the computation of summary statistics such as spectral
sums. In nω time, it is possible to perform a full SVD and compute the singular values exactly.1
Here, and throughout, ω ≈ 2.3729 denotes the current best exponent of fast matrix multiplication
[Wil12]. However, even if one simply desires, for example, a constant factor approximation to
the nuclear norm kAk1, no o(nω) time algorithm is known. We study the question of spectrum
approximation, asking whether obtaining an accurate picture of A's spectrum is truly as hard as
matrix multiplication, or if it is possible to break this barrier. We focus on spectral sums as a
motivating application.
1.1 Our Contributions
Upper Bounds: On the upper bound side, we show that significant information about A's
spectrum can be determined in o(nω) time, for the current value of ω. We show how to compute
a histogram of the spectrum, which gives approximate counts of the number of squared singular
values in the ranges [(1− α)tσ2
1(A)] for some width parameter α and for t ranging
from 0 to some maximum T . Specifically our algorithm satisfies the following:
Theorem 1 (Histogram Approximation -- Informal). Given A ∈ Rn×d, let bt be the number of
squared singular values of A on the range [(1 − α)tσ2
1(A)]. Then given error
parameter ǫ > 0, with probability 99/100, Algorithm 1 outputs for all t ∈ {0, ..., T}, bt satisfying:
1(A), (1 − α)t−1σ2
1(A), (1− α)t−1σ2
(1 − ǫ)bt ≤ bt ≤ (1 + ǫ)bt + ǫ(bt−1 + bt+1).
and κ
k(A)+Pd
i=k+1 σ2
def
=
kσ2
For input parameter k ∈ {1, ..., d}, let ¯κ
maximum number of nonzeros in a row of A. The algorithm's runtime can be bounded by:
d·(1−α)T
O nnz(A)k + dkω−1 +pnnz(A)[d · ds(A) + dk]¯κ
for sparse A or O(cid:16) ndγ−1+n1/2d3/2√¯κ)
poly(ǫ, α)
poly(ǫ,α)
def
=
i (A)
σ2
k+1(A)
(1−α)T . Let ds(A) be the
!
√κ⌉
(cid:17) for dense A, where dγ is the time it takes to multiply a
O nnz(A)k + dkω−1 + (nnz(A) + dk)⌈
! or
poly(ǫ, α)
d × d matrix by a d × k matrix using fast matrix multiplication.
1 Note that an exact SVD is incomputable even with exact arithmetic [TB97]. Nevertheless, direct methods for
the SVD obtain superlinear convergence rates and hence are often considered to be 'exact'.
1
This primitive is useful on its own, giving an accurate summary of A's spectrum which can be
used in many downstream applications. Setting the parameter k appropriately to balance costs (see
overview in Section 1.3), we use it to give the first o(nω) algorithms for computing (1 ± ǫ) relative
error approximations to a broad class of spectral sums for functions f , which are a) smooth and
b) quickly growing, so that very small singular values cannot make a significant contribution to
Sf (A). This class includes for example the Schatten p-norms for all p > 0, the SVD entropy, the
Ky Fan norms, and many general Orlicz norms.
For a summary of our p-norm results see Table 1. Focusing for simplicity on square matrices,
p in
with uniformly sparse rows, and assuming ǫ, p are constants, our algorithms approximate kAkp
O(nnz(A)) time for any real p ≥ 2.2 For p ≤ 2, we achieve O(cid:16)nnz(A)n
2/p+1pnnz(A)(cid:17)
runtime. In the important case of p = 1, this becomes O(cid:16)nnz(A)n1/3 + npnnz(A)(cid:17). Note that
npnnz(A) ≤ n2, and for sparse enough A, this bound is subquadratic. For dense A, we use fast
1+.0435p (cid:17) for all p < 2. For p = 1, this gives O(n2.18).
matrix multiplication, achieving time O(cid:16)n
Even without fast matrix multiplication, the runtime is O(n2.33), and so o(nω) for ω ≈ 2.3729.
Lower Bounds: On the lower bound side, we show that obtaining o(nω) time spectrum approxi-
mation algorithms with very high accuracy may be difficult. Our runtimes all depend polynomially
on the error ǫ, and we show that improving this dependence, e.g., to log(1/ǫ), or even to a better
polynomial, would give faster algorithms for the well studied Triangle Detection problem.
2.3729−.0994p
1/p−1/2
1/p+1/2 + n
4/p−1
Specifically, for a broad class of spectral sums, including all Schatten p-norms with p 6= 2, SVD
entropy, log det(A), tr(A−1), and tr(exp(A)), we show that any (1 ± ǫ) approximation algorithm
running in O(nγǫ−c) time yields an algorithm for triangle detection running in O(nγ+O(c)) time.
For γ < ω and sufficiently small c, such an algorithm would improve the state of the art in
triangle detection, which currently requires Θ(nω) time on dense graphs. Furthermore, through
a reduction of [WW10], any subcubic time triangle detection algorithm yields a subcubic time
algorithm for Boolean Matrix Multiplication (BMM). Thus, any spectral sum algorithm achieving
subcubic runtime and 1
ǫc accuracy for small enough constant c, must (implicitly) implement fast
matrix multiplication. This is in stark contrast to the fact that, for c = 3, for many spectral sums,
including all Schatten p-norms with p ≥ 1/2, we obtain subcubic, and in fact o(nω) for ω = 2.3729,
runtimes without using fast matrix multiplication (see Table 1 for precise ǫ dependencies).
Our lower bounds hold even for well-conditioned matrices and structured matrices like sym-
metric diagonally dominant (SDD) systems, both of which admit nearly linear time algorithms for
system solving [ST04]. This illustrates a dichotomy between linear algebraic primitives like ap-
plying A−1 to a vector and spectral summarization tasks like precisely computing tr(A−1), which
in some sense require more global information about the matrix. Our analysis has ramifications
regarding natural open problems in graph theory and numerical computation. For example, for
graph Laplacians, we show that accurately computing all effective resistances yields an accurate
algorithm for computing tr(A−1) of certain matrices, which is enough to give triangle detection.
1.2 Related Work on Spectral Sums
The applications of approximate spectral sum computation are very broad. When A is positive
semidefinite (PSD) and f (x) = log(x), Sf (A) is the log-determinant, which is important in machine
learning and inference applications [Ras04, DKJ+07, FHT08]. For f (x) = 1/x, Sf (A) is the trace
of the inverse, used in uncertainty quantification [BCF09] and quantum chromodynamics [SLO13].
2For any A ∈ Rn×d, nnz(A) denotes the number of nonzero entries in A.
2
p range Sparsity Approx.
p > 2
p ≤ 2
p ≤ 2
p > 0
p > 2
p < 2
uniform
uniform
(1 + ǫ)
(1 + ǫ)
dense
(1 + ǫ)
general
general
general
(1 + ǫ)
1/γ
1/γ
Runtime
4/p−1
3+p/2
· n
f (p,ǫ)
O(cid:16) 1
f (p,ǫ)
n
2.3729−.0994p
O(cid:18) 1
1/p−1/2
1/p+1/2 + n
2/p+1pnnz(A)(cid:21)(cid:19)
O(cid:0)nnz(A) · p/ǫ3(cid:1)
f (p,ǫ) (cid:20)nnz(A)n
1+.0435p (cid:17) or O(cid:18) 1
f (p,ǫ)hnnz(A)n
O(cid:16) 1
p3hnnz(A)n
O(cid:16) 1
1+p/2(cid:19) w/o FMM
1+pi(cid:17)
O(p nnz(A) · nγ)
1/p+1/2 +γ/2 +pnnz(A) · n
4/p−1
2/p−1i(cid:17)
1
1+p + n1+ 2
1/p−1/2
Theorem
Thm 32
Thm 33
Thm 31
Thms 32, 33
Thm 34
Thm 34
Table 1: Summary of our results for approximating the Schatten p-norms. We define f (p, ǫ) = min{1, p3} ·
ǫmax{3,1+1/p}, which appears as a factor in many of the bounds. The uniform sparsity assumption is that the
maximum row sparsity ds(A) ≤ ξ
n nnz(A) for some constant ξ. In our theorems, we give general runtimes,
parameterized by ξ. When we do not have the uniform sparsity assumption, we are still able to give a (1 + ǫ)
approximation in O(ǫ−3 nnz(A)√n + n2) time for example for kAk1. We can also give 1/γ approximation
for any constant γ < 1 by paying an nγ/2 factor in our runtime. Note that for dense matrices, for all p we
obtain o(nω) runtime, or o(n3) runtime if we do not use fast matrix multiplication.
When f (x) = xp, Sf (A) = kAkp
p where kAkp is the Schatten p-norm of A. Computation of
the Schatten 1-norm, also known as the nuclear or trace norm, is required in a wide variety of
applications. It is often used in place of the matrix rank in matrix completion algorithms and other
convex relaxations of rank-constrained optimization problems [CR12, DTV11, JNS13, NNS+14]. It
appears as the 'graph energy' in theoretical chemistry [Gut92, Gut01], the 'singular value bound'
in differential privacy [HLM12, LM12], and in rank aggregation and collaborative ranking [LN15].
Similar to the nuclear norm, general Schatten p-norms are used in convex relaxations for
rank-constrained optimization [NHD12, NWC+12]. They also appear in image processing appli-
cations such as denoising and background subtraction [XGL+16], classification [LYCG14], restora-
tion [XQT+16], and feature extraction [DHJZ15].
When f (x) = −x log x (after A is normalized by kAk1), Sf (A) is the SVD entropy [ABB00],
which is used in feature selection [VGLH06, BP14], financial data analysis [Car14, GXL15], and
genomic data [ABB00] applications.
Despite their importance, prior to our work, few algorithms for fast computation of spectral sums
existed. Only a few special cases of the Schatten p-norms were known to be computable efficiently
in o(nω) time. These include the Frobenius norm (p = 2) which is trivially computed in O(nnz(A))
time, the spectral norm (p = ∞) which can be estimated via the Lanczos method in O(nnz(A)ǫ− 1
2 )
time [KW92], and the Schatten-p norms for even integers p > 2, or general integers with PSD A.
These norms can be approximated in O(nnz(A)ǫ−2) time via trace estimation [Woo14, BCKY16],
since when p is even or A is PSD, Ap is PSD and so its trace equals kAkp
p.
There are a number of works which consider estimating matrix norms in sublinear space and with
a small number of passes over A [LNW14, AKR15, LW16a, LW16b, BCKY16, LW17]. However, in
these works, the main focus is on space complexity, and no non-trivial runtime bounds are given.
We seem to be the first to tackle the arguably more fundamental problem of obtaining the best time
complexity for simple norms like the Schatten-p norms, irrespective of space and pass complexity.
Another interesting line of works tries to estimate the Schatten-p norms of an underlying co-
variance matrix from a small number of samples from the distribution [KV16], or from entrywise
sampling under various incoherence assumptions [KO17]. This model is different from ours, as we
do not assume an underlying distribution or any incoherence properties. Moreover, even with such
3
assumptions, these algorithms also only give non-trivial sample complexity when either A is PSD
and p is an integer, or A is a general matrix but p is an even integer, which as mentioned above
are easy to handle from the perspective of time complexity alone.
A number of works have focused on computing spectral sums when A has bounded condition
number, and relative error results exist for example for the log-determinant, tr(exp(A)), tr(A−1),
and the Schatten p-norms [BDKZ15, HMAS17, UCS17]. We are the first to give relative error re-
sults in o(ω) time for general matrices, without the condition number dependency. Our histogram
approach resembles spectral filtering and spectral density estimation techniques that have been con-
sidered in the numerical computation literature [ZWJ15, DNPS16, LSY16, US16, UCS17, USS17].
However, this literature typically requires assuming gaps between the singular values and existing
work is not enough to give relative error spectral sum approximation for general matrices
1.3 Algorithmic Approach
def
Spectral Sums via Trace Estimation: A common approach to spectral sum approximation is
to reduce to a trace estimation problem involving the PSD matrix AT A, using the fact that the trace
of this matrix equals the sum of its singular values. In fact, this has largely been the only known
technique, other than the full SVD, for obtaining aggregate information about A's singular values
[Hut90, SLO13, WLK+16, WWZ14, RKA15, FORF16, BDKZ15, HMAS17]. The idea is, letting
g(x) = f (x1/2), we have Sf (A) = Sg(AT A). Writing the SVD AT A = UΛUT , and defining the
matrix function g(AT A)
= Ug(Λ)UT where [g(Λ)]i,i = g([Λ]i,i), we have Sg(AT A) = tr(g(AT A))
since, if g(·) is nonnegative, g(AT A) is PSD and its trace equals the sum of its singular values.
It is well known that this trace can be approximated up to (1± ǫ) accuracy by averaging O(ǫ−2)
samples of the form xT g(AT A)x where x is a random Gaussian or sign vector [Hut90, AT11]. While
g(AT A) cannot be explicitly computed without a full SVD, a common approach is to approximate
g with a low-degree polynomial φ [BDKZ15, HMAS17]. If φ has degree q, one can apply φ(AT A)
to any vector x in O(nnz(A) · q) time, and so estimate its trace in just O(nnz(A) · q
ǫ2 ) time.
Unfortunately, for many of the functions most important in applications, e.g., f (x) = xp for odd
p, f (x) = x log x, f (x) = x−1, g(x) has a discontinuity at x = 0 and cannot be approximated
well by a low-degree polynomial near zero. While the approximation only needs to be good on the
1(A)
range [σ2
n(A) is
σ2
the condition number, which can be unbounded for general matrices.
1(A)], the required degree q will still typically depend on √κ where κ def= σ2
n(A), σ2
Singular Value Deflation for Improved Conditioning: Our first observation is that, for
many functions, it is not necessary to approximate g(x) on the full spectral range. For example,
for g(x) = xp/2 (i.e., when Sg(AT A) = kAkp
p), setting λ = ( ǫ
p)1/p, we have:
X{iσi(A)≤λ}
n kAkp
p ≤ ǫkAkp
p .
σi(A)p ≤ n ·
ǫ
n kAkp
Hence we can safely 'ignore' any σi(A) ≤ λ and still obtain a relative error approximation to
Sg(AT A) = kAkp
p. The larger p is, the larger we can set λ (corresponding to (1−α)T in Theorem 1)
to be, since, after powering, the singular values below this threshold do not contribute significantly
p, our 'effective condition number' for approximating g(x) becomes κ = σ2
1 (A)
to kAkp
p. For kAkp
λ2 =
. Unfortunately, in the worst case, we may have σ1(A) ≈ kAkp and hence √κ = ( n
ǫ )2/p· σ2
1(A)
( n
ǫ )1/p.
kAk2
Hiding ǫ dependences, this gives runtime O(nnz(A) · n) when p = 1.
p
4
To improve the effective condition number, we can apply singular vector deflation. Our above
bound on κ is only tight when the first singular value is very large and so dominates kAkp. We can
remedy this by flattening A's spectrum by deflating off the top k singular vectors (corresponding
to k in Theorem 1), and including their singular values in the spectral sum directly.
Specifically, letting Pk be the projection onto the top k singular vectors of A, we consider the
k kAkp
deflated matrix ¯A
p,
and thus this singular value cannot dominate the p-norm. As an example, considering p = 1 and
ignoring ǫ dependencies, our effective condition number after deflation is
= A(I − Pk), which has σ1( ¯A) = σk+1(A). Importantly, σp
k+1(A) ≤ 1
def
κ =
n2 · σ2
k+1(A)
kAk2
1
n2
k2
≤
(1)
The runtime required to approximate Pk via an iterative method (ignoring possible gains from fast
matrix multiplication) is roughly O(nnz(A)k + nk2). We then require O(nnz(A)√κ + nk√κ) time
to approximate the polynomial trace of ¯AT ¯A. The nk√κ term comes from projecting off the top
singular directions with each application of ¯AT ¯A. Setting k = √n to balance the costs, we obtain
runtime O(nnz(A)√n + n2).
1
p+1 + n2+ 1
For p 6= 1 a similar argument gives runtime O(nnz(A)n
p+1 ). This is already a
significant improvement over a full SVD. As p grows larger, the runtime approaches O(nnz(A))
reflecting the fact that for larger p we can ignore a larger and larger portion of the small singular
values in A and correspondingly deflate off fewer and fewer top values.
Unfortunately, we get stuck here. Considering the important Schatten-1 norm, for a matrix
with √n singular values each equal to √n and Θ(n) singular values each equal to 1, the tail of
small singular values contributes a constant fraction of kAk1 = Θ(n). However, there is no good
polynomial approximation to g(x) = x1/2 on the range [1, n] with degree o(√n) (recall that we
pick this function since Sg(AT A) = kAk1). So to accurately approximate g(AT A), we either
must deflate off all √n top singular values, requiring Θ(nnz(A)√n) time, or apply a Θ(√n) degree
polynomial approximation, requiring the same amount of time.
Further Improvements with Stochastic Gradient Descent: To push beyond this barrier,
we look to stochastic gradient methods for linear systems. When using polynomial approximation,
our bounds depend on the condition number of the interval over which we must approximate
g(AT A), after ignoring the smallest singular values and deflating off the largest. This is analogous
to the condition number dependence of iterative linear system solvers like conjugate gradient or
accelerated gradient descent, which approximate f (AT A) for f = 1/x using a polynomial of AT A.
However, recent advances in convex optimization offer an alternative. Stochastic gradient meth-
ods [JZ13, SSZ14] sample one row, ai, of A at a time, updating the current iterate by adding a
multiple of ai. They trade a larger number of iterations for updates that take O(nnz(ai)) time,
rather than O(nnz(A)) time to multiply A by a vector. At the same time, these methods give much
finer dependencies on the singular value spectrum. Specifically, it is possible to approximately apply
(AT A)−1 to a vector with the number of iterations dependent on the average condition number:
¯κ =
1
nPn
i=1 σ2
σ2
n(A)
i (A)
.
¯κ is always at most the standard condition number, κ = σ2
1(A)
n(A) . It can be significantly smaller when
σ2
A has a quickly decaying spectrum, and hence 1
i=1 σ2
i (A) ≪ σ2
1(A). Further, the case of a
quickly decaying spectrum with a few large and many small singular values is exactly the hard case
nPn
5
for our earlier approach. If we can understand how to translate improvements on linear system
solvers to spectral sum approximation, we can handle this hard case.
From Linear System Solvers to Histogram Approximation: The key idea to translating
the improved average condition number bounds for linear systems to our problem of approximating
Sf (A) is to note that linear system solvers can be used to apply threshold functions to AT A.
Specifically, given any vector y, we can first compute AT Ay. We can then apply a fast system
solver to approximate (AT A + λI)−1AT Ay. The matrix function rλ(AT A) def= (AT A + λI)−1AT A
has a number of important properties. All its singular values are between 0 and 1. Further, any
singular value in AT A with value ≥ λ is mapped to a singular value in rλ(AT A) which is ≥ 1/2.
Correspondingly, any singular value < λ is mapped to < 1/2.
Thus, we can apply a low degree polynomial approximation to a step function at 1/2 to rλ(AT A)
to obtain sλ(AT A), which approximates a step function at λ [FMMS16]. For some steepness
parameter γ which affects the degree of the polynomial approximation, for x ≥ (1 + γ)λ we have
sλ(x) ≈ 1 and for x < (1 − γ)λ, sλ(x) ≈ 0. On the intermediate range x ∈ [(1 − γ)λ, (1 + γ)λ],
sλ(x) falls somewhere between 0 and 1.
By composing these approximate threshold functions at different values of λ, it is possible to
'split' our spectrum into a number of small spectral windows. For example, sa(AT A)·(I−sb(AT A))
is ≈ 1 on the range [a, b] and ≈ 0 outside this range, with some ambiguity near a and b.
Splitting our spectrum into windows of the form [(1 − α)t−1, (1 − α)t] for a width parameter α,
and applying trace estimation on each window lets us produce an approximate spectral histogram.
Of course, this histogram is not exact and in particular, the 'blurring' of our windows at their
boundaries can introduce significant error. However, by applying a random shifting technique and
setting the steepness parameter γ small enough (i.e., 1/ poly(α, ǫ)), we can ensure that most of the
spectral weight falls outside these boundary regions with good probability, giving Theorem 1.
From Histogram Approximation to Spectral Sums:
If α is small enough, and f (·) (the
function in the spectral sum) and correspondingly g(·) (where g(x) = f (x1/2)) are smooth enough,
we can approximate Sf (A) = Sg(AT A) by simply summing over each bucket in the histogram,
approximating g(x) by its value at one end of the bucket.
This technique can be applied for any spectral sum. The number of windows required (controlled
by α) and the histogram accuracy ǫ scale with the smoothness of f (·) and the desired accuracy in
computing the sum, introducing polynomial runtime dependencies on these parameters.
However, the most important factor determining the final runtime is the smallest value λ (corre-
sponding to (1−α)T in Theorem 1) which we must include in our histogram in order to approximate
Sf (A). The cost of computing the last bucket of the histogram is proportional to the cost of ap-
plying sλ(AT A), and hence of approximately computing (AT A + λI)−1AT Ay. Using stochastic
gradient descent this depends on the average condition number of (AT A + λI).
Again considering the Schatten 1-norm for illustration, we can ignore any singular values with
n kAk1. Hiding ǫ dependence, this means that in our histogram, we must include any
1. This gives us effective average
σi(A) ≤ ǫ
singular values of AT A with value σi(AT A) = σ2
condition number after deflating off the top k singular values:
¯κ =
n2Pn
i=k+1 σ2
nkAk2
1
i (A)
≤
n2 kAk2
i (A) ≥ 1
n2σk+1(A) ·Pn
nkAk2
1
where the last inequality follows from the observation that σk+1(A) ≤ 1
kAk1. Comparing to (1), this bound is better by an n/k factor.
k kAk1 andPn
i=k+1 σi(A) ≤
i=k+1 σi(A)
n
k
≤
(2)
6
Ignoring details and using a simplification of the runtimes in Theorem 1, we obtain an algorithm
time to approximate the spectral sum over the deflated matrix. Choosing k to balance these costs,
gives our final runtimes. For the nuclear norm, using the bound on ¯κ from (2), we set k = n1/3 which
running in O(nnz(A)k+nk2) time to deflate k singular vectors, plus O(cid:16)nnz(A)√¯κ +pnnz(A)nk¯κ(cid:17)
gives ¯κ = n2/3 and runtime O(nnz(A)n1/3 + n3/2√ds) where ds ≤ n is the maximum row sparsity.
For dense A this is O(n2.33), which is faster than state of the art matrix multiplication time. It
can be further accelerated using fast matrix multiplication methods. See details in Section 7.
Returning to our original hard example for intuition, we have A with √n singular values at √n
and Θ(n) singular values at 1. Even without deflation, we have (again ignoring ǫ dependencies)
1 = Θ(n2), this gives ¯κ = Θ(√n). Thus,
¯κ = Pn
we can actually approximate kAk1 in just O(nnz(A)n1/4) time for this matrix.
With average condition number dependence, our performance is limited by a new hard case.
Consider A with n1/3 singular values at n2/3 and Θ(n) at 1. The average condition number without
deflation is nkAk2
kAk2
deflate off nearly all n1/3 top singular vectors, we do not improve this bound significantly.
n (cid:17) = Θ(n2/3) giving √¯κ = Θ(n1/3). Further, we can see that unless we
. Since kAk2
F = Θ(n3/2) and kAk2
= Θ(cid:16) n5/3
= nkAk2
kAk2
1
F
i (A)
i=1 σ2
nλ
F
1
1.4 Lower Bound Approach
We now shift focus to our lower bounds, which explore the fundamental limits of spectrum approx-
imation using fine-grained complexity approaches. Fine-grained complexity has had much success
for graph problems, string problems, and problems in other areas (see, e.g., [Wil15] for a survey),
and is closely tied to understanding the complexity of matrix multiplication. However, to the best
of our knowledge it has not been applied broadly to problems in linear algebra.
Existing hardness results for linear algebraic problems tend to apply to restricted computational
models such as arithmetic circuits [BS83], bilinear circuits or circuits with bounded coefficients and
number of divisions [Mor73, Raz03, Shp03, RS03], algorithms for dense linear systems that can
only add multiples of rows to each other [KKS65, KS70], and algorithms with restrictions on the
dimension of certain manifolds defined in terms of the input [Win70, Win87, Dem13]. In contrast,
we obtain conditional lower bounds for arbitrary polynomial time algorithms by showing that faster
algorithms for them imply faster algorithms for canonical hard problems in fine-grained complexity.
From Schatten 3-norm to Triangle Detection: We start with the fact that the number
of triangles in any unweighted graph G is equal to tr(A3)/6, where A is the adjacency matrix.
Any algorithm for approximating tr(A3) to high enough accuracy therefore gives an algorithm for
detecting if a graph has at least one triangle.
A is not PSD, so tr(A3) is actually not a function of A's singular values -- it depends on the
signs of A's eigenvalues. However, the graph Laplacian given by L = D−A where D is the diagonal
degree matrix, is PSD and we have:
kLk3
3 = tr(L3) = tr(D3) − 3 tr(D2A) + 3 tr(DA2) − tr(A3).
tr(D2A) = 0 since A has an all 0 diagonal. Further, it is not hard to see that tr(DA2) = tr(D2). So
this term and tr(D3) are easy to compute exactly. Thus, if we approximate kLk3
3 up to additive error
6, we can determine if tr(A3) = 0 or tr(A3) ≥ 6 and so detect if G contains a triangle. kLk3
3 ≤ 8n4
for any unweighted graph on n nodes, and hence computing this norm up to (1 ± ǫ) relative error
for ǫ = 3/(6n4) suffices to detect a triangle. If we have an O(nγǫ−c) time (1 ± ǫ) approximation
algorithm for the Schatten 3-norm, we can thus perform triangle detection in O(nγ+4c) time.
7
Our strongest algorithmic result for the Schatten 3-norm requires just O(n2/ǫ3) time for dense
Improving the ǫ dependence to o(1/ǫ(ω−2)/4) = O(1/ǫ.09) for the current value of ω,
matrices.
would yield an algorithm for triangle detection running in o(nω) time for general graphs, breaking
a longstanding runtime barrier for this problem. Even a 1/ǫ1/3 dependence would give a sub-
cubic time triangle detection algorithm, and hence could be used to give a subcubic time matrix
multiplication algorithm via the reduction of [WW10].
Generalizing to Other Spectral Sums We can generalize the above approach to the Schatten
4-norm by adding λ self-loops to each node of G, which corresponds to replacing A with λI+A. We
then consider tr((λI + A)4) = kλI + Ak4
4. This is the sum over all vertices of the number of paths
that start at vi and return to vi in four steps. All of these paths are either (1) legitimate four cycles,
(2) triangles combined with self loops, or (3) combinations of self-loops and two-step paths from
a vertex vi to one of its neighbors and back. The number of type (3) paths is exactly computable
using the node degrees and number of self loops. Additionally, if the number of self loops λ is large
enough, the number of type (2) paths will dominate the number of type (1) paths, even if there is
just a single triangle in the graph. Hence, an accurate approximation to kλI + Ak4
4 will give us the
number of type (2) paths, from which we can easily compute the number of triangles.
This argument extends to a very broad class of spectral sums by considering a power series
expansion of f (x) and showing that for large enough λ, tr (f (λI + A)) is dominated by tr(A3)
along with some exactly computable terms. Thus, an accurate approximation to this spectral sum
allows us to determine the number of triangles in G. This approach works for any f (x) that can
be represented as a power series, with reasonably well-behaved coefficients on some interval of R+,
giving bounds for all kAkp with p 6= 2, the SVD entropy, log det(A), tr(A−1), and tr(exp(A)).
We further show that approximating tr(A−1) for the A used in our lower bound can be reduced
to computing all effective resistances of a certain graph Laplacian up to (1± ǫ) error. Thus, we rule
out highly accurate (with 1/ǫc dependence for small c) approximation algorithms for all effective
resistances, despite the existence of linear time system solvers (with log(1/ǫ) error dependence) for
Laplacians [ST04]. Effective resistances and leverage scores are quantities that have recently been
crucial to achieving algorithmic improvements to fundamental problems like graph sparsification
[SS08] and regression [LMP13, CLM+15]. While crude multiplicative approximations to the quan-
tities suffice for these problems, more recently computing these quantities has been used to achieve
breakthroughs in solving maximum flow and linear programming [LS14], cutting plane methods
[LSW15], and sampling random spanning trees [KM09, MST15]. In each of these settings having
more accurate estimates would be a natural route to either simplify or possibly improve existing
results; we show that this is unlikely to be successful if the precision requirements are two high.
1.5 Paper Outline
Section 2: Preliminaries. We review notations that will be used throughout.
Section 3: Spectral Window Approximation. We show how to approximately restrict the
spectrum of a matrix to a small window, which will be our main primitive for accessing the spectrum.
Section 4: Spectral Histogram Approximation. We show how our spectral window algorithms
can be used to compute an approximate spectral histogram. We give applications to approximating
general spectral sums, including the Schatten p-norms, Orlicz norms, and Ky Fan norms.
Section 5: Lower Bounds. We prove lower bounds showing that highly accurate spectral sum
algorithms can be used to give algorithms for triangle detection and matrix multiplication.
8
Section 6: Improved Algorithms via Polynomial Approximation. We demonstrate how to
tighten ǫ dependencies in our runtimes using a more general polynomial approximation approach.
Section 7: Optimized Runtime Bounds. We instantiate the techniques of Section 6 give our
best runtimes for the Schatten p-norms and SVD entropy.
2 Preliminaries
Here we outline notation and conventions used throughout the paper.
Matrix Properties: For A ∈ Rn×d we assume without loss of generality that d ≤ n. We let
σ1(A) ≥ . . . ≥ σd(A) ≥ 0 denote the matrix's singular values, nnz(A) denote the number of non-
zero entries, and ds(A) denote the maximum number of non-zero entries in any row. Note that
ds(A) ∈ [nnz(A)/n, d].
Fast Matrix Multiplication: Let ω ≈ 2.3729 denote the current best exponent of fast ma-
trix multiplication [Wil12, GU17]. Additionally, let ω(γ) denote the exponent such that it takes
O(cid:0)dω(γ)(cid:1) time to multiply a d × d matrix by a d × dγ matrix for any γ ≤ 1. ω(γ) = 2 for γ < α
where α > 0.31389 and ω(γ) = 2 + (ω − 2) γ−α
Asymptotic Notation: We use O(·) notation to hide poly-logarithmic factors in the input pa-
rameters, including dimension, failure probability, and error ǫ. We use 'with high probability' or
'w.h.p.' to refer to events happening with probability at least 1 − 1/dc for some constant c, where
d is our smaller input dimension.
1−α for γ ≥ α [LG12, GU17]. For γ = 1, ω(γ) = ω.
Other: We denote [d] def= {0, . . . , d}. For any y ∈ Rd and PSD N ∈ Rd×d, we denote kykN
pyT Ny.
def=
3 Approximate Spectral Windows via Ridge Regression
In this section, we give state-of-the-art results for approximating spectral windows over A. As
discussed, our algorithms will split A's spectrum into small slices using these window functions,
performing trace estimation to estimate the number of singular values on each window and pro-
ducing an approximate spectral histogram.
In Section 3.1 we show how to efficiently apply smooth approximations to threshold functions of
the spectrum provided we have access to an algorithm for solving regularized regression problems
with the matrix. In Section 3.2 we then provide the fastest known algorithms for the regression
problems in the given parameter regimes using both stochastic gradient methods and traditional
solvers. Departing from our high level description in Section 1.3, we actually incorporate singular
vector deflation directly into our system solvers to reduce condition number. This simplifies our
final algorithms but has the same effect as the deflation techniques discussed in Section 1.3. Finally,
in Section 3.3 we provide algorithms and runtime analysis for applying smooth approximations to
window functions of the spectrum, which is the main export of this section.
3.1 Step Function Approximation
To compute a window over A's spectrum, we will combine two threshold functions at the boundaries
of the window. We begin by discussing how to compute these threshold functions.
Let sλ : [0, 1] → [0, 1] be the threshold function at λ. sλ(x) = 1 for x ∈ [λ, 1] and 0 for x ∈ [0, λ).
For some gap γ we define a soft step function by:
9
Definition 2 (Soft Step Function). sγ
λ : [0, 1] → [0, 1] is a γ-soft step at λ > 0 if:
sγ
λ(x) =(0 for x ∈ [0, (1 − γ)λ]
1 for x ∈ [λ, 1]
and sγ
λ(x) ∈ [0, 1] for x ∈ [(1 − γ)λ, λ].
(3)
We use the strategy from [FMMS16], which, for A with kAk2 ≤ 1 shows how to efficiently
multiply a γ-soft step sγ
λ(AT A) by any y ∈ Rd using ridge regression. The trick is to first ap-
proximately compute AT A(AT A + λI)−1y = rλ(AT A)y where rλ(x)
x+λ . Then, note that
s1/2(rλ(x)) = sλ(x). Additionally, the symmetric step function s1/2 can be well approximated with
a low degree polynomial. Specifically, there exists a polynomial of degree O(γ−1 log(1/(γǫ))) that
is within additive ǫ of a true γ-soft step at 1/2 and can be applied stably such that any error in
computing rλ(AT A) remains bounded. The upshot, following from Theorem 7.4 of [AZL16] is:
Lemma 3 (Step Function via Ridge Regression). Let A(A, y, λ, ǫ) be an algorithm that on input
probability. Then there is an algorithm B(A, y, λ, γ, ǫ) which on input A ∈ Rn×d with kAk2 ≤ 1,
y ∈ Rd, λ ∈ (0, 1), and γ, ǫ > 0, returns x ∈ Rd with
A ∈ Rn×d, y ∈ Rd, λ, ǫ > 0 returns x ∈ Rd such that (cid:13)(cid:13)x − (AT A + λI)−1y(cid:13)(cid:13)2 ≤ ǫkyk2 with high
def
= x
(cid:13)(cid:13)x − sγ
λ(AT A)y(cid:13)(cid:13)2 ≤ ǫkyk2
λ(AT A)y for sγ
λ is a γ-soft step at λ (i.e. satisfies Definition 2). B(A, y, λ, γ, ǫ) requires O(γ−1 log(1/ǫγ))
where sγ
calls to A(A, y, λ, ǫ′) along with O(nnz(A)) additional runtime, where ǫ′ = poly(1/(γǫ)).
3.2 Ridge Regression
Given Lemma 3, to efficiently compute sγ
λ(·) satisfying Definition 2, it suffices to
quickly approximate (AT A+λI)−1y (i.e. to provide the algorithm A(A, y, λ, ǫ) used in the lemma).
In this section we provide two theorems which give the state-of-the-art ridge regression running
times achievable in our parameter regime, using sampling, acceleration, and singular value deflation.
Naively, computing (AT A + λI)−1y using an iterative system solver involves a dependence on
the condition number σ2
1(A)/λ. In our theorems, this condition number is replaced by a deflated
condition number depending on σ2
k(A) for some input parameter k ∈ [d]. We achieve this improved
dependence following the techniques presented in [GOSS16]. We first approximate the top k singular
vectors of A and then construct a preconditioner based on this approximation, which significantly
flattens the spectrum of the matrix. By using this preconditioner in conjunction with a stochastic
gradient based linear system solver, we further enjoy an average condition number dependence.
The following theorem summarizes the results.
Theorem 4 (Ridge Regression -- Accelerated Preconditioned SVRG). For any A ∈ Rn×d and
where k ∈ [d] is an input parameter. There
λ > 0, let Mλ
is an algorithm that builds a preconditioner for Mλ using precomputation time O(nnz(A)k +dkω−1)
for sparse A or O(ndω(logd k)−1) time for dense A, and for any input y ∈ Rd, returns x such that
λ y(cid:13)(cid:13)Mλ ≤ ǫkykM−1
O(cid:16)nnz(A) +pnnz(A)[d · ds(A) + dk]¯κ(cid:17)
with high probability (cid:13)(cid:13)x − M−1
time for sparse A or O(cid:0)nd + n1/2d3/2√¯κ)(cid:1) time for dense A.
Proof. We give a proof in Appendix A. Note that the ǫ dependence in the runtime is log(1/ǫ) and
so is hidden by the O(·) notation.
def= AT A+λI. Let ¯κ def=
i=k+1 σ2
dλ
kσ2
k(A)+Pd
i (A)
in
λ
10
When A is dense, the runtime of Theorem 4 is essentially the best known. Due to its average
condition number dependence, the method always outperforms traditional iterative methods, like
conjugate gradient, up to log factors. However, in the sparse case, traditional approaches can give
faster runtimes if the rows of A are not uniformly sparse and ds(A) is large. We have the following,
also proved in Appendix A using the same deflation-based preconditioner as in Theorem 4:
Theorem 5 (Ridge Regression -- Preconditioned Iterative Method). For any A ∈ Rn×d and λ >
0, let Mλ
where k ∈ [d] is an input parameter. There is an
algorithm that builds a preconditioner for Mλ using precomputation time O(nnz(A)k + dkω−1), and
in
for any input y ∈ Rd, returns x such that with high probability (cid:13)(cid:13)x − M−1
O(cid:16)(nnz(A) + dk)⌈√κ⌉(cid:17) time.
λ y(cid:13)(cid:13)Mλ ≤ ǫkykM−1
def
= AT A + λI and κ
σ2
k+1(A)
def
=
λ
λ
3.3 Overall Runtimes For Spectral Windows
Combined with Lemma 3, the ridge regression routines above let us efficiently compute soft step
functions of A's spectrum. Composing step functions then gives our key computational primitive:
the ability to approximate soft window functions that restrict A's spectrum to a specified range.
We first define our notion of soft window functions and then discuss runtimes. The corresponding
Theorem 7 is our main tool for spectrum approximation in the remainder of the paper.
Definition 6 (Soft Window Function). Given a, b > 0 with a < b, and γ ∈ [0, 1], hγ
[0, 1] is a γ-soft window for the range [a, b] if:
[a,b] : [0, 1] →
[a,b](x) =(1 for x ∈ [a, b]
hγ
0 for x ∈ [0, (1 − γ)a] ∪ [(1 + γ)b, 1]
and hγ
[a,b](x) ∈ [0, 1] for x ∈ [(1 − γ)a, a] ∪ [b, (1 + γ)b].
Theorem 7 (Spectral Windowing). For any A ∈ Rn×d with kAk2 ≤ 1, y ∈ Rd, and a, b, γ, ǫ ∈
(0, 1], with a < b, there is an algorithm W(A, y, a, b, γ, ǫ) that returns x satisfying w.h.p.:
(cid:13)(cid:13)(cid:13)x − hγ
[a,b](AT A)y(cid:13)(cid:13)(cid:13)2 ≤ ǫkyk2
[a,b] is a soft window function satisfying Def. 6. Let ¯κ
where hγ
where k ∈ [d] is an input parameter. The algorithm uses precomputation time O(nnz(A)k + dkω−1)
for sparse A or O(ndω(logd k)−1) for dense A after which given any y it returns x in time:
and κ
k(A)+Pd
i=k+1 σ2
d·a
σ2
k+1(A)
i (A)
def
=
def
=
kσ2
a
O (nnz(A) + dk)⌈
γ
√κ⌉
!
γ
! or
O nnz(A) +pnnz(A)[d · ds(A) + dk]¯κ
(cid:17) for dense A.
for sparse A or O(cid:16) nd+n1/2d3/2√¯κ)
Proof. If b ≥ 1/(1 + γ) then we can simply define hγ
2. Otherwise, given soft steps sγ
a(x) · (1 − sγ/2
sγ
for [a, b] (i.e. satisfy Definition 6). Further, we have for any y ∈ Rd:
(1+γ)b(x)). Since γ
[a,b](x) = sγ
2 ≤ γ
γ
a(x) for any sγ
a satisfying Definition
(1+γ)b satisfying Definition 2, we can define hγ
a and sγ/2
[a,b](x) =
1+γ we can verify that this will be a valid soft window function
hγ
[a,b](AT A)y = sγ
a(AT A)(I − sγ/2
(1+γ)b(AT A))y.
(4)
11
We can compute sγ
a(AT A)y and sγ/2
(1+γ)b(AT A)y each up to error ǫkyk2 via Lemma 3. This gives
a(AT A)(cid:13)(cid:13)2 ≤ 1 and(cid:13)(cid:13)(cid:13)I − sγ/2
(1+γ)b(AT A)(cid:13)(cid:13)(cid:13)2 ≤ 1
the error bound in the theorem, since we have both(cid:13)(cid:13)sγ
so the computation in (4) does not amplify error. Our runtime follows from combining Theorems
4 and 5 with λ = a, b with Lemma 3. The errors in these theorems are measured with respect
. To obtain the error in k·k2 as used by Lemma 3, we simply apply the theorems with
to k·kMλ
ǫ′ = ǫκ(Mλ) which incurs an additional log(κ(Mλ)) cost. Since a < b the runtime is dominated by
i (A)
the computation of sγ
a(AT A)y, which depends on the condition number ¯κ def=
kσ2
k(A)+Pd
i=k+1 σ2
d·a
when using SVRG (Theorem 4) or κ
def
=
σ2
k+1(A)
a
for a traditional iterative solver (Theorem 5).
4 Approximating Spectral Sums via Spectral Windows
We now use the window functions discussed in Section 3 to compute an approximate spectral
histogram of A. We give our main histogram algorithm and approximation guarantee in Section
4.1. In Section 4.2 we show how this guarantee translates to accurate spectral sum approximation
for any smooth and sufficiently quickly growing function f (x). In Section 4.3 we apply this general
result to approximating the Schatten p-norms for all real p > 0, bounded Orlicz norms, and the
Ky Fan norms.
4.1 Approximate Spectral Histogram
Our main histogram approximation method is given as Algorithm 1. The algorithm is reasonably
simple. Assuming kAk2 ≤ 1 (this is w.l.o.g. as we can just scale the matrix), and given cutoff λ,
below which we will not evaluate A′s spectrum, we split the range [λ, 1] into successive windows
R0, ..., RT where Rt = [a1(1 − α)t, a1(1 − α)t−1]. Here α determines the width of our windows.
In our final spectral approximation algorithms, we will set α = Θ(ǫ). a1 is a random shift, which
insures that, in expectation, the boundaries of our soft windows do not overlap too many singular
values. This argument requires that most of the range [λ, 1] is not covered by boundary regions.
Thus, we set the steepness parameter γ = Θ(ǫ2α) where ǫ2 will control the error introduced by the
boundaries. Finally, we iterate over each window, applying trace estimation to approximate the
singular value count in each window.
In our final algorithms, the number of windows and samples required for trace estimation will
be O(poly(1/ǫ)). The dominant runtime cost will come from computing the window for the lowest
range RT , which will incur a dependence on the condition number of AT A + aT I with aT = Θ(λ).
Theorem 8 (Histogram Approximation). Let a1, b0, ..., bT be output by Algorithm 1. Let R0 =
[a1, 1], Rt = [a1(1 − α)t, a1(1 − α)t−1] for t ≥ 1, and bt = (cid:12)(cid:12){i : σ2
squared singular values of A on the range Rt. Then, for sufficiently small constants c1, c2, c3, with
probability 99/100, for all t ∈ {0, ...,⌈log(1−α) λ⌉}, bt output by Algorithm 1 satisfies:
i (A) ∈ Rt}(cid:12)(cid:12) be the number of
(1 − ǫ1)bt ≤ bt ≤ (1 + ǫ1)bt + ⌈log(1−α) λ⌉ · ǫ2(bt−1 + bt + bt+1).
That is, bt approximates the number of singular values of the range Rt up to multiplicative
(1±ǫ1) error and additive error ⌈log(1−α) λ⌉·ǫ2(bt−1+bt+bt+1). Note that by setting ǫ2 ≤
,
⌈log(1−α) λ⌉
the error on each bucket is just multiplicative on its size plus the size of the two adjacent buckets,
which contain singular values in nearby ranges. For simplicity we assume A passed to the algorithm
has kAk2 ≤ 1. This is without loss of generality: we can estimate kAk2 in O(nnz(A)) time via the
power or Lanczos methods [KW92, MM15], and scale down the matrix appropriately.
ǫ1
12
Algorithm 1 Approximate Spectral Histogram
Input: A ∈ Rn×d with kAk2 ≤ 1, accuracy parameters ǫ1, ǫ2 ∈ (0, 1), width parameter α ∈ (0, 1),
and minimum singular value parameter λ ∈ (0, 1).
Output: Set of range boundaries aT +1 < aT < ... < a1 < a0 and counts {b0, ..., bT } where bt
approximates the number of squared singular values of A on [at+1, at].
Set γ = c1ǫ2α, T = ⌈log(1−α) λ⌉, and S = log n
Set a0 = 1 and choose a1 uniformly at random in [1 − α/4, 1].
Set at = a1(1 − α)t−1 for 2 ≤ t ≤ T + 1.
for t = 0 : T do
c2ǫ2
1
.
Set bt = 0.
for s = 1 : S do
⊲ Iterate over histogram buckets.
⊲ Initialize bucket size estimate.
⊲ Estimate bucket size via trace estimation.
Choose y ∈ {−1, 1}d uniformly at random.
Set bt = bt + 1
If bt ≤ 1/2 set bt = 0.
S · yTW(AT A, y, at+1, at, γ, c3ǫ2
end for
1/n).
⊲ Apply soft window via Thm 7.
⊲ Round small estimates to ensure relative error.
end for
return a1 and bt for t = 0 : T .
⊲ Output histogram representation.
Ignoring logarithmic and ǫ dependencies, the runtime of Algorithm 1 is dominated by the calls
to W for the bucket corresponding to the smallest singular values, with aT = Θ(λ). This runtime is
given by Theorem 7. Since balancing the deflation parameter k in that theorem with the minimum
squared singular value λ considered can be complex, we wait to instantiate full runtimes until
employing Algorithm 1 for specific spectral sum computations.
Proof. We use the notation of Algorithm 1, where a0 = 1, a1 is chosen uniformly in [1− α/4, 1] and
at = a1(1 − α)t−1. With this notation, we have Rt = [at+1, at]. Let hγ
be a γ-soft window for Rt
(Definition 6) and let ¯Rt = [(1 − γ)at+1, (1 + γ)at] be the interval on which hγ
is nonzero. bt is an
estimation of the trace of such a window applied to AT A. We first show that, if these traces are
computed exactly, they accurately estimate the singular value counts in each range Rt. We have:
Rt
Rt
tr(hγ
Rt
(AT A)) = Xσ2
i (A)∈Rt
hγ
Rt
(σ2
i (A)) + Xσ2
i (A)∈ ¯Rt\Rt
hγ
Rt
(σ2
i (A))
= bt + Xσ2
i (A)∈ ¯Rt\Rt
hγ
Rt
(σ2
i (A)).
(5)
We can bound the second term using the random shift a1. Since γ = c1ǫ2α < α, each singular
value falls within at most two extended ranges: ¯Rt and ¯Rt±1 for some t. Let I be the set of indices
whose singular values fall within two ranges. We have i ∈ I only if a1(1 − α)t ∈ (1 ± γ)σ2
i (A) for
(1−α)d(cid:17), which occurs with
some t. Letting d = ⌈logσ2
probability at most 8γ
α since a1 is chosen uniformly in the range [1−α/4, 1]. Thus we have P[i ∈ I] ≤
def= a1(1− α)t−1 is distributed uniformly in the range [(1− α/4)(1− α)t−1 , (1− α)t−1],
8γ
α . Further, at
so we know for certain that if the constant c1 on γ is set small enough:
i (A)(1 − α)⌉, this holds only if a1 ∈ (1 ± γ)(cid:16) σ2
i (A)
¯Rt ⊂ [(1 − γ)(1 − α/4)(1 − α)t, (1 + γ)(1 − α)t−1]
⊂ [(1 − α/3)(1 − α)t, (1 + α/3)(1 − α)t−1]
⊂ [(1 − α)t+1, (1 − α/4)(1 − α)t−2].
13
(6)
def
= [(1 − α)t+1, (1 − α/4)(1 − α)t−2]. Note that Mt is fixed (i.e. not a random variable).
Let Mt
Regardless of the random shift a1, by (6), we always have ¯Rt ⊂ Mt. We also have Mt ⊂ Rt−1 ∪
Rt ∪ Rt+1.
Let I[i ∈ I] be 1 if i ∈ I and 0 otherwise. We can upper bound the second term of (5) by:
Xσ2
i (A)∈ ¯Rt\Rt
hγ
Rt
(σ2
I[i ∈ I]
I[i ∈ I]
i (A)∈ ¯Rt\Rt
i (A)) ≤ Xσ2
≤ Xσ2
≤ Xσ2
i (A)∈Mt
i (A)∈Mt\Rt
I[i ∈ I].
(7)
The first bound follows from the fact that for σi(A)2 ∈ ¯Rt \ Rt, we have σi(A)2 ∈ I and that
hγ
i (A)) ≤ 1. The second bound follows from (6), which shows that ¯Rt ⊂ Mt. Let mt = {i :
(σ2
Rt
σ2
i (A) ∈ Mt} be the number of squared singular values falling in Mt. Note that like Mt, mt is
fixed (i.e., not a random variable.) Thus, by linearity of expectation, we have:
Letting T = ⌈log(1−α) λ⌉ as in Algorithm 1, by a Markov bound, with probability 1 −
1
200(T +1) :
i (A)∈Mt
E Xσ2
Xσ2
i (A)∈Mt
I[i ∈ I] = Xσ2
i (A)∈Mt
P[i ∈ I] =
8γ
α · mt.
I[i ∈ I] ≤
1600(T + 1)γ
α
· mt ≤ T ǫ2 · mt
if c1 is set small enough. By a union bound this holds for all t ∈ {0, ..., T} simultaneously with
probability ≥ 199/200. Plugging back into (7), we have, with probability ≥ 199/200, simultaneously
for all t:
Xσ2
i (A)∈ ¯Rt\Rt
hγ
Rt
(σ2
i (A)) ≤ T ǫ2 · mt
≤ T ǫ2 · (bt−1 + bt + bt+1)
(8)
where the second bound follows from the fact that, regardless of the setting of the shift a1, Mt ⊂
Rt−1 ∪ Rt ∪ Rt+1 so mt ≤ (bt−1 + bt + bt+1). Plugging (8) into (5) we have:
bt ≤ tr(hγ
Rt
(AT A)) ≤ bt + T ǫ2(bt−1 + bt + bt+1).
(9)
We conclude by showing that, before the final rounding step of Algorithm 1, bt ∈ (1±ǫ1) tr(hγ
√c3ǫ1 with high probability. In the final rounding step, if tr(hγ
(AT A)) ≤ 1/4 (which can only
occur if bt = 0) and c1, c3 are sufficiently small, then we will have bt ≤ 1/2 and so will round down
to bt = 0 = bt. Otherwise, the √c3ǫ1 term will be absorbed into the relative error on tr(hγ
(AT A)).
Rt
Rt
(AT A))+
Rt
Overall, combined with (9) we will have, with probability ≥ 99/100 for all t:
(1 − ǫ)bt ≤ bt ≤ (1 + 2ǫ1)bt + (1 + 2ǫ1)T ǫ2(bt−1 + bt + bt+1)
which gives the theorem if we adjust ǫ1, ǫ2 by making c1, c2, c3 sufficiently small.Thus we conclude by
showing that in fact bt ∈ (1±ǫ1) tr(hγ
as in
(AT A))+√c3ǫ1 with high probability. Setting S = log n
c2ǫ2
1
Rt
14
Rt
Algorithm 1, for y1, . . . , yS chosen from {−1, 1}d, with high probability 1
(1 ± ǫ1) tr(hγ
(AT A)) by a standard trace estimation result [AT11].
Further, let xi = W(AT A, y, at+1, at, γ, c3ǫ2
(AT A)yi(cid:13)(cid:13)2 ≤
(AT A)yi ≤pc3/nǫ1 kyik2 = √c3ǫ1 since
n2 kyik2 which by Cauchy-Schwarz gives yT
kyik2 = √n. Thus, overall we have, before the rounding step in which bt is set to 0 if bt < 1/2:
i xi − yT
(AT A)yi ∈
SPS
1/n)yi. By Theorem 7: (cid:13)(cid:13)xi − hγ
i hγ
Rt
i=1 yT
i hγ
Rt
c3ǫ2
1
Rt
bt =
1
S
SXi=1
yT
i W(AT A, y, at+1, at, γ, c3ǫ2
1/n)yi
∈ (1 ± ǫ1) tr(hγ
Rt
(AT A)) + √c3ǫ1.
4.2 Application to General Spectral Sums
While Theorem 8 is useful in its own right, we now apply it to approximate a broad class of spectral
sums. We need two assumptions on the sums that we approximate. First, for the histogram
discretization to be relatively accurate, we need our function to be relatively smooth. Second, it is
expensive to compute the histogram over very small singular values of A (i.e. with λ very small in
Algorithm 1) as this makes the condition number in Theorem 7 large. So it is important that small
singular values cannot contribute significantly to our sum. We start with the following definition:
Definition 9 (Multiplicative Smoothness). f : R+ → R+ is δf -multiplicatively smooth if for some
δf ≥ 1, for all x, f′(x) ≤ δf
f (x)
x .
We have the following claim, proven in Appendix D.
Claim 10. Let f : R+ → R+ be a δf -multiplicatively smooth function. For all x, y ∈ R+ and
c ∈ (0, 1
3δf
)
y ∈ [(1 − c)x, (1 + c)x] ⇒ f (y) ∈ [(1 − 3δf c)f (x), (1 + 3δf c)f (x)].
multiplicatively smooth.
For the example of the Schatten p-norm, for f (x) = xp, f′(x) = p · xp−1 and so f (x) is p-
We now give our general approximation theorem, showing that any spectral sum depending on
sufficiently smooth and rapidly growing f can be computed using Algorithm 1:
Theorem 11 (Spectral Sums Via Approximate Histogram). Consider any A ∈ Rn×d and any
function f : R+ → R+ satisfying:
• Multiplicative Smoothness: For some δf ≥ 1, f is δf -multiplicatively smooth (Defn. 9).
• Small Tail: For any ǫ > 0 there exists λf (ǫ) such that for x ∈ [0, λf (ǫ)], f (x) ≤ ǫ
nSf (A)
Given error parameter ǫ ∈ (0, 1) and spectral norm estimate M ∈ [kAk2 , 2kAk2], for suf-
ficiently small constant c, if we run Algorithm 1 on 1
M A with input parameters ǫ1, ǫ2 = cǫ,
α = cǫ/δf and λ = λf (cǫ)2/M 2 then with probability 99/100, letting a1, b0, ..., bT be the outputs
of the algorithm and g(x) = f (x1/2):
(1 − ǫ)Sf (A) ≤
TXt=0
g(M 2 · a1(1 − α)t) · bt ≤ (1 + ǫ)Sf (A).
15
For parameter k ∈ [d], letting ¯κ def=
kσ2
k(A)+Pd
i=k+1 σ2
d·λ
i (A)
and κ def=
σ2
k+1(A)
λ
, the algorithm runs in
O nnz(A)k + dkω−1 +
!
nnz(A) +pnnz(A)[d · ds(A) + dk]¯κ
f log(1/λ)) !
√κ⌉
f log(1/λ))
(nnz(A) + dk)⌈
ǫ5/(δ2
ǫ5/(δ2
or O nnz(A)k + dkω−1 +
time for sparse A or O(cid:18)ndω(logd k)−1 + nd+n1/2d3/2√¯κ)
ǫ5/(δ2
f log(1/λ))(cid:19) for dense A.
That is, we accurately approximate Sf (A) by discretizing over the histogram output by Algo-
rithm 1. Note that we can boost our probability of success to 1 − δ by repeating the algorithm
Θ(log(1/δ)) times and taking the median of the outputs.
Proof. Let ¯A
1 and Theorem 8. Recall that we use the notation: R0 = [a1, 1], Rt = [a1(1 − α)t, a1(1 − α)t−1] for
M A. Note that(cid:13)(cid:13) ¯A(cid:13)(cid:13)2 ≤ 1 and so it is a valid matrix on which to apply Algorithm
def
= 1
t ≥ 1, T = ⌈log(1−α) λ⌉, and bt =(cid:12)(cid:12){i : σ2
Since g(x) = f (x1/2), by chain rule:
i ( ¯A) ∈ Rt}(cid:12)(cid:12).
g′(x) =
f′(x1/2)
2x1/2 ≤ δf
f (x1/2)
2x
=
δf
2
g(x)
x
.
So g is δf /2 multiplicatively smooth. By this smoothness, and Claim 10, for any i with σ2
i ( ¯A) ∈ Rt:
g(M 2 · a1(1 − α)t) ∈(cid:18)1 ±
3δf α
2 (cid:19) g(σ2
i (A))
∈ (1 ± ǫ/4) · f (σi(A))
if the constant c on α is set small enough. Small enough c also ensures g(M 2a1(1 − α)t−1) ∈
(1 ± ǫ/4)f (σi(A)) and g(M 2a1(1 − α)t+1) ∈ (1 ± ǫ/4)f (σi(A)). So, both the multiplicative and
additive error terms in Theorem 8 do not hurt us significantly. For now, assume that ǫ2 = cǫ/T . We
will first prove the result with this assumption and then show that it can be relaxed to ǫ2 = cǫ as
i ( ¯A) ∈ Rt for some t ∈ {0, ..., T}.
given in the theorem statement. Let I be the set of indices with σ2
That is, the set of singular values covered by our histogram. Let IT +1 be the set of indices with
i ( ¯A) ∈ RT +1 -- that is, singular values which are not included in the histogram, but may be
σ2
included in the additive error for the last bucket covering RT . Applying Theorem 8:
TXt=0
g(M 2a1(1 − α)t) · bt ≤
TXt=0
g(M 2a1(1 − α)t) · [(1 + ǫ1)bt + ǫ2T (bt−1 + bt + bt+1)]
f (σi(A)) + 3cǫ Xσ2
(1 + cǫ) Xσ2
i ( ¯A)∈Rt
i ( ¯A)∈Rt
f (σi(A))
f (σi(A))
≤ (1 + ǫ/4)
T +1Xt=0
≤ (1 + ǫ) Xi∈I∪IT +1
≤ (1 + ǫ)Sf (A)
16
(10)
if we set c small enough. The second inequality arises because each σi(A) contributing to bt appears
at most three times as an additive error term for bt−1, bt, and bt+1. In this inequality we include
bucket RT +1 in the histogram, which only increases the right hand side.
On the lower bound side, we use our small tail assumption, that for i with σi(A) < λf (cǫ),
n Sf (A). Using the notation of Algorithm 1, we have aT +1 = a1(1 − α)T ≤ (1 −
i ( ¯A) > aT +1 and thus falls in
f (σi(A)) ≤ cǫ
α)⌈log(1−α) λ⌉ ≤ λ def= λf (cǫ)2/M 2. So for any i with σi(A) ≥ λf (cǫ), σ2
some bucket of our histogram so i ∈ I. We thus have:
Xi∈I
f (σi(A)) ≥ Sf (A) − Xi:σi(A)<λf (cǫ)
f (σi(A))
≥ Sf (A) − n ·
cǫ
n Sf (A) = (1 − cǫ)Sf (A).
Applying Theorem 8 again, we have for sufficiently small c:
TXt=0
g(a1(1 − α)t) · bt ≥ (1 − ǫ/4)(1 − cǫ)
= (1 − ǫ/4)(1 − cǫ)Xi∈I
TXt=0 Xσ2
i ( ¯A)∈Rt
f (σi(A))
f (σi(A))
≥ (1 − ǫ)Sf (A).
We conclude the theorem by noting that we can actually set ǫ2 = cǫ instead of ǫ2 = cǫ/T as used
above. The additive error term on each bucket was bounded using a Markov bound, and to union
bound over T buckets, we lost a factor of T . However, in expectation, the total contribution of
so by a Markov bound is ≤ ǫ/2Sf (A) with probability 99/100 if we set our constants small enough.
It just remains to discuss runtime, i.e. to calculate the runtime of Algorithm 1 with inputs
ǫ1, ǫ2 = cǫ, α = cǫ/δf and λ = λf (cǫ)2/M 2. The number of outer loops of the algorithm is
f (σ1(A))(cid:17) = O (Sf (A)) and
ǫ log(1/λ)(cid:17). The number of inner loops is S = O(1/ǫ2). And finally,
additive error to our spectral sum estimation is just O(cid:16)ǫPi∈I∪IT +1
T = ⌈log(1−α) λ⌉ = Θ(cid:16) δf
within each loop the most expensive operation is computing W(AT A, y, at+1, at, γ, c3ǫ2
1/n). Our
final runtime follows from plugging this into Theorem 7 noting that γ = Θ(ǫα) = Θ(ǫ2/δf ) and
at+1 = Ω(λ) for all t ∈ {0, ..., T}. Note that we perform the precomputation step to construct a
preconditioner for AT A + λI just once, incurring cost O(nnz(A) + dkω−1) or O(ndω(logd k)−1).
4.3 Application to Schatten p, Orlicz, and Ky Fan Norms
Theorem 11 is very general, allowing us to approximate any function satisfying a simple smoothness
condition as long as the smaller singular values of A cannot contribute significantly to Sf (A). We
now give some specific applications of the result.
Schatten p-Norms
Theorem 11 already gives the fastest known algorithms for Schatten p-norm estimation. We will
not go into all runtime tradeoffs now as our best runtimes will be worked out in detail in Sections
6 and 7, but as an example:
Corollary 12 (Schatten p-norms via Histogram Approximation). For any A ∈ Rn×n with uni-
formly sparse rows (i.e. ds(A) = O(nnz(A)/n)), given error parameter ǫ ∈ (0, 1) and M ∈
17
1
[kAk2 , 2kAk2], if we run Algorithm 1 on 1
M 2(cid:16) cǫ
p(cid:17)2/p
n kAkp
be the outputs of the algorithm we have:
M A with ǫ1, ǫ2 = cǫ, α = cǫ/ max{1, p} and λ =
for sufficiently small constant c then with probability 99/100, letting a1, b0, ..., bT
and
p ≤
(1 − ǫ)kAkp
Further the algorithm runs in time:
· bt ≤ (1 + ǫ)kAkp
p .
O(cid:18) nnz(A)p2
ǫ5+1/p (cid:19) for p ≥ 2
TXt=0(cid:2)M 2a1(1 − α)t(cid:3)p/2
O nnz(A)n
For dense inputs this can be sped up to O(cid:18) n
p·ǫ5+1/p! without fast matrix multiplication.
O n
stant ǫ the second runtime gives O(nnz(A)n1/3+n3/2pds(A)) which is at worst O(nnz(A)n1/3+n2).
for p ≤ 2.
p·ǫ5+1/p (cid:19) using fast matrix multiplication or
For dense matrices, the nuclear norm estimation time is O(n2.18) using fast matrix multiplication.
It is already O(n2.33) = o(nω), for the current ω, without using fast matrix multiplication.
For constant ǫ, p > 2 the first runtime is O(nnz(A)), and for the nuclear norm (p = 1), for con-
2/p+1 pds(A)
1/p−1/2
1/p+1/2 + n
p · ǫ5+1/p
2.3729−.1171p
3+p/2
1+p/2
1+.0346p
5/p−1/2
Note that we can compute the spectral norm approximation used to scale A via the Lanczos or
power method in O(nnz(A)) time. λ depends on kAkp which we are estimating. However, as we
will discuss in the proof, we can use a rough estimate for kAkp which suffices. As a more general
strategy, λ can be identified via binary search. We can start with λ = σ1(A)2/M 2 and successively
decrease λ running Algorithm 1 up to the stated runtime bounds. If it does not finish in the allotted
time, we know that we have set λ too small. Thus, we can output the result with the smallest λ
such that the algorithm completes within in the stated bounds.
Proof. We invoke Theorem 11 with f (x) = xp. We have f′(x) = p f (x)
x
setting of α = cǫ/ max{1, p} suffices. Additionally, for any c, we can set λf (cǫ) =(cid:16) cǫ
so δf = max{1, p} and our
=
n1/p kAkp and so our setting of λ suffices. Thus the accuracy bound follows from Theorem 11.
We now consider runtime. For p ≥ 2:
p(cid:17)1/p
n kAkp
c1/pǫ1/p
¯κ =
kσ2
k(A) +Pn
nλ
i=k+1 σ2
i (A)
n2/p−1
ǫ2/p
≤
F
· kAk2
kAk2
p
.
We can bound kAkF ≤ n1/2−1/p kAkp and so have ¯κ ≤ 1
σ2−p
k
i=k+1 σ2
i (A)
kσ2
¯κ =
n2/p−1
ǫ2/p
·
k(A) +Pn
kAk2
k(A) ≤ 1
p
k kAkp
Using the fact that σp
p we have the tradeoff between k and ¯κ:
ǫ2/p . For p < 2 we have:
n2/p−1
ǫ2/p
·
≤
i=1 σp
i (A)
(A)Pn
kAk2
p
=
n2/p−1
ǫ2/p
·
σ2−p
(A)
k
kAk2−p
p
.
¯κ ≤
1
ǫ2/p(cid:16) n
k(cid:17)2/p−1
18
.
(11)
As mentioned, λ depends on the value of kAkp. We can simply lower bound kAkp
k(A),
which we estimate up to multiplicative error when performing deflation. We can use this lower
bound to set λ. Our estimated λ will only be smaller than the true value, giving a better approxi-
mation guarantee and the above condition number bound will still hold.
p by kσp
We now analyze our runtime. Recall that for f (x) = xp, δf = max{1, p}. Correspondingly,
log(1/λ) = O(max{1, 1/p}) and so δ2
f log(1/λ) = max{p2, 1/p}. Plugging into the first runtime of
Theorem 11, using the uniform sparsity assumption and the fact that √x + y ≤ √x + √y we have:
O nnz(A)k + nkω−1 +
nnz(A)√¯κ +pnnz(A)nk¯κ
ǫ5/(max{p2, 1/p}
! .
For p ≥ 2 we just set k = 0 and have O(nnz(A)p2/ǫ5+1/p) runtime by our bound ¯κ ≤ 1
p ≤ 2, not trying to optimize poly(1/ǫ) terms, we write the runtime as
ǫ2/p . For
O nds(A)k + nkω−1 +
nds(A)√¯κ + npds(A)k¯κ
ǫ5p
!
1/p−1/2
1/p+1/2 which gives √¯κ = n
and balancing the first two coefficients on n, set k = n
Finally, for dense A we apply the third runtime which gives
and so nds(A)k = nds(A)√¯κ. We then have npds(A)k¯κ = npds(A)k3/2 = n
Finally, the nkω−1 is dominated by the npds(A)k3/2 term so we drop it.
k1/p−1/2 · ǫ5+1/p · p!
O(cid:18)nω(logn k) +
n2√¯κ
n3/2+1/p
ǫ5p (cid:19) = O nω(logn k) +
= O nω(logn k) +
1/p−1/2
1/p+1/2 by (11)
5/p−1/2
2/p+1 pds(A).
n3/2+1/p−(logn k)(1/p−1/2)
ǫ5+1/p · p
! .
We now balance the terms, again not optimizing ǫ dependence. Writing γ = logn k, ω(γ) = 2 for
γ < α where α > 0.31389 and 2 + (ω − 2) γ−α
1−α for γ ≥ α [GU17]. Assuming γ > α we set:
γ − α
1 − α
1
p −
γ
p
γ
2
3
2
+
=
+
2 + (ω − 2)
which gives γ ≈ 1/p−.3294
yields total runtime O(cid:16)n
1/p+.0435 > α for all p < 2 (so our assumption that γ ≥ α was valid.) This
1+.0435p (cid:17). Without using fast matrix multiplication, the first term in
the runtime becomes n2k and so we balance costs by setting: n2+γ = n3/2+1/p−γ/p+γ/2 which gives
γ = 1/p−1/2
2.3729−.0994p
3+p/2
1/p+1/2 and total runtime O(cid:18)n
1+p/2(cid:19).
Bounded Orlicz Norms
An Orlicz norm of a vector x ∈ Rn is given by G(x) :=Pn
i=1 g(xi), where g(·) is convex, nonneg-
ative, and has g(0) = 0. It is easy to see that applied to the vector of singular values, an Orlicz
matrix norm is a special case of a spectral sum and can be approximated with Theorem 11 under
sufficient conditions. A simple example shows that Orlicz norms for g(·) bounded by an envelope
of xp1 and xp2 can be approximated in a similar time as kAkp2
in Corollary 12.
p2
19
Corollary 13 (Bounded Orlicz norms via Histogram Approximation). For any convex function
g(a) .
a(cid:1)p1 ≥ g(b)
g : R+ → R+, let p1 be the minimal real for which for all arguments a, b to g with a ≤ b,(cid:0) b
Let p2 be the maximum positive real for which for all a ≤ b we have(cid:0) b
Then for any A ∈ Rn×n with uniformly sparse rows (i.e. ds(A) = O(nnz(A)/n)), given error
parameter ǫ ∈ (0, 1) and M ∈ [kAk2 , 2kAk2], if we run Algorithm 1 on 1
M A with ǫ1, ǫ2 = cǫ, α =
cǫ/ max{1, p1} and λ = 1
for sufficiently small constant c then with probability
99/100, letting a1, b0, ..., bT be the outputs of the algorithm we have:
a(cid:1)p2 ≤ g(b)
M 2(cid:16) cǫ/2
n kAkp2
g(a) .
(1 − ǫ)Sg(A) ≤
(1 − α)t/2(cid:17) · bt ≤ (1 + ǫ)Sg(A).
1
p2(cid:17)2/p2
g(cid:16)M a1/2
TXt=0
O (p1 + 1)2 · nnz(A)n
Further the algorithm runs in time:
and
ǫ5+1/p2
(cid:19) for p2 ≥ 2
O(cid:18) nnz(A)(p1 + 1)2
For dense inputs this can be sped up to O (p1 + 1)2 · n
Note that λ depends on kAkp2
binary search as described in the proof of Corollary 12.
p2
5/p2−1/2
1/p2 −1/2
1/p2+1/2 + n
p2 · ǫ5+1/p2
2/p2+1 pds(A)
for p2 ≤ 2.
p2·ǫ5+1/p2 ! via fast matrix multiplication.
1+.0435p2
2.3729−.0994p2
which can be estimated for example using Corollary 12, or via
Proof. We invoke Theorem 11 with f (x) = g(x). While we do not show that g(·) exactly satisfies
the multiplicative smoothness condition, the upper bound on g(·) directly implies the result of
Claim 10 with δf = p1. So our setting of α = cǫ/ max{1, p1} suffices as the bound in this Claim is
the only smoothness condition used to prove Theorem 11.
Additionally, by the lower bound on g(·), for any c, we can set λf (cǫ) =(cid:16) cǫ/2
n kAkp2
p2(cid:17)1/p2
. For
any σi(A) ≥ λf (cǫ):
λf (cǫ)(cid:19)p2
g(σi(A)) ≥(cid:18) σi(A)
σp2
i (A)
kAkp2
2n
cǫ ·
≥
g(λf (cǫ))
p2 · g(λf (cǫ)).
which gives:
Xσi(A)≥λf (cǫ)
g(σi(A)) ≥
≥
2n · g(λf (cǫ))
cǫ · kAkp2
2n · g(λf (cǫ))
p2
cǫ
σp2
i (A)
· Xσi(A)≥λf (cǫ)
· (1 − cǫ/2)
where the last bound follows from the fact that by the setting of λf (cǫ), Pσi(A)≥λf (cǫ) σ
kAkp2
1 − cǫ/2. Finally for any σi(A) ≤ λf (cǫ), g(σi(A)) ≤ g(λf (cǫ)). So overall we have:
p2
p2
i (A)
≥
Xσi(A)<λf (cǫ)
g(σi(A)) ≤ n · g(λf (cǫ)) ≤
cǫ/2
1 − cǫ/2 · Xσi(A)≥λf (cǫ)
g(σi(A)) ≤
cǫ/2
1 − cǫ
Sg(A) ≤ cǫSg(A).
20
Thus our setting of λ suffices and the accuracy bound follows from Theorem 11.
It remains to discuss runtime. We have log(1/λ) = O (max{1, 1/p2}) and δ2
f = max{1, p2
(p1 + 1)2. The runtimes follow from Theorem 11 via the same arguments used in Corollary 12.
1} ≤
Ky Fan Norms
i=1 σi(A)
The Ky Fan w-norm of a matrix is the sum of its top w singular values: kAkKF (w)
(note that these norms are typically called the 'Ky Fan k-norms', however we use w to avoid
overloading notation on k). Such a norm is not strictly speaking a spectral sum. However it can
still be approximated using our histogram method. We have the following corollary of Theorem 8:
Corollary 14 (Ky Fan norms via Histogram Approximation). For any A ∈ Rn×n, given rank w,
error parameter ǫ ∈ (0, 1) and M ∈ [kAk2 , 2kAk2], if we run Algorithm 1 on 1
M A with ǫ1 = cǫ,
ǫ2 = c2ǫ2
for sufficiently small constants c, c2 then with
probability 99/100, letting a1, b0, ..., bT be the outputs of the algorithm we have:
log(1/λ) , α = cǫ and λ = (cid:16) cǫ
def=Pw
M w kAkKF (w)(cid:17)2
wXi=1
where t(i) is the smallest integer with Pt(i)
(1 − ǫ)kAkKF (w) ≤
M a1/2
1
(and hence the ith term of the sum is just 0). Further the algorithm runs in time:
t=0(1 + 2cǫ)bt ≥ i or t(i) = ∞ if no such integer exists
(1 − α)t(i)/2 ≤ (1 + ǫ)kAkKF (w)
O(cid:18)nnz(A)√w + nw
ǫ7
(cid:19) .
Note that to compute the top w singular values explicitly would require O(cid:0)nnz(A)w + nwω−1(cid:1)
time using for example a block Krylov method [MM15]. Also note that λ depends on the norm we
are attempting to compute. As discussed in Corollary 12, we can approximate λ via binary search,
successively decreasing it until our call to Algorithm 1 exceeds the stated runtime.
Proof. We can assume without loss of generality that kAk2 ≤ 1 and M = 1 since rescaling will
not effect our approximation factor. By Theorem 8, since we set ǫ1 = cǫ, ǫ2 = c2ǫ2
log(1/λ) , and have
T = ⌈log1−α λ⌉ = Θ(cid:16) log(1/λ)
ǫ
(cid:17):
(1 − cǫ)bt ≤ bt ≤ bt + 2cǫ(bt−1 + bt + bt+1)
(12)
bt ≤ (1 + 2cǫ)bt if c is set small enough.
if we set c2 small enough compared to c. This give bt ≤ 1
1−cǫ
Thus, since these scaled bucket sizes strictly overestimate the true bucket sizes we have t(i) ≤ t(i),
where t(i) is the smallest integer withPt(i)
t=0 bt ≥ i. This gives, since α = cǫ and since by our setting
wXi=1
of λ, at most an cǫ fraction of kAkKF (w) falls outside of the ranges R0, ..., RT :
(1 − α)t(i)/2 ≤ (1 + 4cǫ)
σi(A) ≤ (1 + 4cǫ)
(1 − α)t(i)/2.
kAkKF (w) =
wXi=1
wXi=1
M a1/2
M a1/2
1
1
On the other side, let v be a vector that, for each t ∈ {0, ..., T} has (1 + 4cǫ)bt entries each set to
M a1/2
(1 − α)t/2. Define the top-w norm of v to be the sum of its largest w entries, denoted by
1
21
kvkT (w). Our estimate of kAkKF (w) is equal to kvkT (w). Note that we can also add arbitrary zeros
as padding entries to v and not change this norm.
Similarly, let v′ be a vector with (1 + 10cǫ)bt values each set to (1 + 2cǫ) · a1/2
(1 − α)t/2 for
t ∈ {0, ..., T} and 2ǫc · bT +1 entries set to (1 + 2cǫ) · a1/2
By (12), (1 + 2cǫ)bt ≤ (1 + 6cǫ)bt + 2cǫbt−1 + 2cǫbt+1. This fact combined with the fact that the
entries in v′ are scaled up by a (1 + 2cǫ) factor ensure that the entries of v′ dominate those of v
and so kv′kT (w) ≥ kvkT (w). Further, kv′kT (w) ≤ (1 + 13cǫ)kAkKF (w) since we have scaled up each
entry by at most (1 + 2cǫ) factor and scaled up each bucket size bt by a (1 + 10cǫ) factor. This
bound gives our final multiplicative approximation after adjusting the constant c on ǫ.
It remains to discuss runtime. We invoke Theorem 7, setting k = w1/2. This gives:
(1 − α)T /2 ≤ λ.
1
1
=
λ
w
κ =
σ2
k+1(A)
σ2
k+1(A)w2
KF (w) (ǫc)2 ≤
kAk2
where we us the fact that for k = w1/2, σk(A) ≤ 1
w1/2 kAkKF (w). With our settings of λ, α and
1 · 1/α) = Θ(1/ǫ3) calls to W, each which requires Θ(1/γ) =
ǫ1, ǫ2, Algorithm 1 performs Θ(1/ǫ2
Θ(ǫ2α) = Θ(1/ǫ3) regression calls due to our setting of α and ǫ2. Plugging κ into the second
runtime of Theorem 7 gives the corollary.
(cǫ)2
5 Lower Bounds
In this section we give hardness results for high accuracy spectrum approximation. Specifically, we
show how to detect if an undirected unweighted graph contains a triangle using accurate approxima-
tion algorithms for various important spectral sums such as the Schatten p-norms, log-determinant,
the SVD entropy, and the trace inverse. Our spectral sum bounds further imply hardness for the
important primitives of computing effective resistances in a graph or leverage scores in a matrix.
In the seminal work of [WW10] it was shown that any truly subcubic time algorithm for tri-
angle detection yields a truly subcubic time algorithms for Boolean matrix multiplication (BMM).
Consequently, these results show that computing any of these quantities too precisely is in a sense
as difficult as BMM. Furthermore, as it is a longstanding open question whether or not there is
any subcubic time combinatorial algorithm for BMM, i.e. an algorithm which avoids the powerful
algebraic manipulations inherent in the fastest algorithms for BMM, these results can be viewed
as showing that we do not expect simple iterative methods to yield precise algorithms for these
problems without an immense breakthrough in linear algebra.
In Section 5.1 we give a general result on reducing spectral sums to triangle detection. Then
in Section 5.2 we use this to show hardness for computing various well studied spectral sums. In
Section 5.3 we conclude by showing hardness for computing graph and numerical linear algebra
primitives, i.e. effective resistances and leverage scores.
5.1 Reductions From Triangle Detection
Here we provide our main technical tool for reducing spectral sum computation to triangle detec-
tion. As discussed in Section 1.4, our reduction leverages the well known fact that the number
of triangles in any unweighted graph G is equal to tr(A3)/6 where A is the adjacency matrix for
G. Consequently, given any function f : R+ → R+ whose power series is reasonably behaved, we
can show that for suitably small δ the quantity tr(f (I + δA)) is dominated by the contribution of
tr(Ak) for k ∈ (0, 3). Therefore computing tr(f (I + δA)) approximately lets us distinguish between
whether or not tr(A3) = 0 or tr(A3) ≥ 6.
22
We formalize this in the following theorem. As it simplifies the result, we focus on the case where
f is defined on the interval (0, 2), however, this suffices for our purposes and can be generalized via
shifting and scaling of x.
Theorem 15 (Many Spectral Sums are as Hard as Triangle Detection). Let f : R+ → R+ be an
arbitrary function, such that for x ∈ (0, 2) we can express it as
(13)
Then given the adjacency matrix A ∈ Rn×n of any simple graph G that has no self-loops and
spectral sum estimate X ∈ (1 ± ǫ1/9)Pn
i=1 f (σi(I − δA)) for scaling δ and accuracy ǫ satisfying
f (x) =
∞Xk=0
δ = min(cid:26) 1
n
,
ck
ck(x − 1)k where (cid:12)(cid:12)(cid:12)(cid:12)
10n4h(cid:27) and ǫ1 = min(cid:26)1 ,
c3(cid:12)(cid:12)(cid:12)(cid:12) ≤ hk−3 for all k > 3
c2n2(cid:12)(cid:12)(cid:12)(cid:12)(cid:27)
c0n(cid:12)(cid:12)(cid:12)(cid:12) ,
(cid:12)(cid:12)(cid:12)(cid:12)
c3δ3
(cid:12)(cid:12)(cid:12)(cid:12)
c3δ
1
we can detect if G has a triangle in O(nnz(A)) time.
Consequently, given an algorithm which on input B ∈ Rn×n outputs Y ∈ (1 ± ǫ)Pn
in O(nγǫ−c) time we can produce an O(n2 + nγǫ−c
1 )) time triangle detection algorithm.
i=1 f (σi(B))
Proof. Let A, G, δ, ǫ1, and X be as in the theorem statement and let B
= I− δA. By Gershgorin's
circle theorem, kAk2 ≤ n − 1 and since δ ≤ 1/n, kδAk2 < 1. Consequently B is symmetric PSD,
σi(B) = λi(B) ∈ (0, 2) for all i ∈ [n], and therefore applying (13) yields:
def
ckδk tr(Ak) .
10n4h is enough to insure that the first three terms of this power series dominate. Specifically:
∞Xk=0
1
ck
c3
f (σi(B)) =
nXi=1
nXi=1
ck(δλi(A))k =
=(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
f (1 − δλi(A)) =
c3δ3 ∞Xk=4
δ ≤ 1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∞Xk=4
where the last inequality uses the fact that tr(Ak) ≤ kAkk−2
Further, since tr(A0) = n, tr(A) = 0, and tr(A2) = kAk2
∞Xk=0
nXi=1
δk−3 tr(Ak)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ c3δ3 ∞Xk=4
ckδk tr(Ak)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
c3δ3(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12)
0 ≤ c0 tr(A0) + c1δ tr(A) + c2δ2(tr(A2)) ≤ c3δ3 ·(cid:18)(cid:12)(cid:12)(cid:12)(cid:12)
Now, clearly in O(nnz(A)) time we can compute tr(A2) = kAk2
9 (cid:18)c3δ3
3(cid:21)
X − c0n − c2δ2 tr(A2) = c3δ3 tr(A3) ± c3δ3
9 ±
= c3δ3(cid:20)tr(A3)(cid:18)1 ±
1
10k−3(cid:18) 1
kAk2
20(cid:19) ±
c0n
F ≤ n2 we have:
ǫ1
1
2
tr(Ak) ≤ c3δ3
n4(cid:19)k−3
F ≤ nk ≤ n4(k−3) for all k > 3.
9
.
ǫ1
c2n2
c3δ (cid:12)(cid:12)(cid:12)(cid:12)(cid:19) ≤ c3δ3
F in O(nnz(A)) as well as:
ǫ1 (cid:19)
+ c3δ3 tr(A3) + c3δ3
9
So we can detect if tr(A3) = 0 or if tr(A3) ≥ 6 and hence whether or not G has a triangle.
Note that in this reduction, so long as δ is small (i.e. ≤ 2n) then B = I − δA is a very well
conditioned matrix (its condition number is at most a constant). Consequently, our lower bounds
apply even when approximately applying for example B−1 or B1/2 to high precision is inexpensive.
23
The theorem (and the results in Section 5.3) suggests that the difficulty in computing spectral sums
arises more from the need to measure the contribution from multiple terms precisely, than from
the difficulty in manipulating B for the purposes of applying it to a single vector.
Also, note that the matrix B in this reduction is symmetrically diagonally dominant (SDD). So,
even for these highly structured matrices which admit near linear time application of B−1 [ST04]
as well as approximate factorization [KS16], accurate spectral sums are difficult. We leverage this
in Section 5.3.
5.2 Hardness for Computing Spectral Sums
Here we use Theorem 15 to show hardness for various spectral sum problems. To simplify our pre-
sentation, we focus on the case of dense matrices, showing bounds of the form Ω(nγǫ−c). However,
note that Theorem 15 and the approach we use also yields lower bounds on the running time for
sparse matrices and can be stated in terms of nnz(A).
Schatten p-norm for all p 6= 2
For x ∈ (0, 2), using the Taylor Series about 1 we can write
xp =
∞Xk=0
ck(1 − x)k where ck = Qk−1
k!
i=0 (p − i)
(14)
This series converges since ck ≤ 1 for k > p and for x ∈ (0, 2), (1 − x) < 1. Note that when p is a
non-negative integer, only the first p terms of the expansion are nonzero. When p is non-integral,
the sum is infinite. We will apply Theorem 15 slightly differently for different values of p. We first
give our strongest result:
Corollary 16 (Schatten 3-Norm Hardness). Given an algorithm which on input B ∈ Rn×n returns
X ∈ (1± ǫ)kBk3
3 in O(nγǫ−c) we can produce an algorithm that detects if an n-node graph contains
a triangle in O(nγ+4c) time.
Proof. For p = 3, ck = 0 for k > 3. So we apply Theorem 15 with h = 0 and hence δ = 1/n and
ǫ1 = c3δ3
c0n = 1
n4 .
Note that for p very close to 3 a similar bound holds as h ≈ 0. If p = 3 Theorem 31 gives an
algorithm running in O(n2/ǫ3) time. Significant improvement to the ǫ dependence in this algorithm
therefore either requires loss in the n dependence or would lead to o(nw) time triangle detection
for the current value of ω. We next extend to all p 6= 1, 2.
Corollary 17 (Schatten p-Norm Hardness, p 6= 1, 2). For any p > 0, p 6= 1, 2, given algorithm A
which for any B ∈ Rn×n returns X ∈ (1 ± ǫ)kBkp
p in O(nγǫ−c) time, we can detect if an n-node
graph contains a triangle in O(cid:16)nγ+13c ·
p(p−1)(p−2)(cid:12)(cid:12)(cid:12) ≤ (cid:12)(cid:12)(cid:12)
c3 ≤ pk−3 for all k > 3 as well as (cid:12)(cid:12)(cid:12) c0
and similarly (cid:12)(cid:12)(cid:12) c2
c3(cid:12)(cid:12)(cid:12) ≤ (cid:12)(cid:12)(cid:12)
2 min{p,(p−1)}(cid:12)(cid:12)(cid:12). We apply Theorem 15 with δ = Θ(cid:16) 1
Θ(cid:16)min{p,(p−1),(p−2)}
(cid:17), which gives the result.
min{p,(p−1),(p−2)}c(cid:17) time.
c3(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)
Proof. We have ck
n13p3
1
2 min{p,(p−1),(p−2)}(cid:12)(cid:12)(cid:12)
n4p(cid:17) and ǫ1 = c3δ3
c0n =
p3c
1
1
24
In the typical case when p << n, the p3c term above is negligible. The
min{p,(p−1),(p−2)}c term
is meaningful however. Our bound becomes weak as p approaches 2 (and meaningless when p = 2).
This is unsurprising, as for p very close to 2, kBkp
F , which can be computed exactly in
nnz(B) time. The bound also becomes weak for p ≈ 1, which is natural as our reduction only uses
PSD B, for which kBk1 = tr(B) which can be computed in n time. However, we can remedy this
issue by working with a (non-PSD) square root of B which is easy to compute:
p ≈ kBk2
1
Note that for p ≈ 1,
p3c/2
Corollary 18 (Schatten p-Norm Hardness, p ≈ 1). For any p, given an algorithm which for any
p in O(cid:0)f (nnz(B), n) · 1
B ∈ Rp×n returns X ∈ (1± ǫ)kBkp
with m edges contains a triangle in O(cid:16)f (m, n) · n13c ·
ǫc(cid:1) time, we can detect if an n-node graph
min{p/2,(p/2−1),(p/2−2)}c + m(cid:17) time.
p3c/2
min{p/2,(p/2−1),(p/2−2)}c is just a constant. Again, the bound is naturally
weak when p ≈ 2 as (p/2 − 1) goes to 0
Proof. For B = I − δA as in Theorem 15. Let L = D − A be the Laplacian of G where D is
the diagonal degree matrix. We can write B = δL + bD where bD = I − δD is PSD since δ ≤ 1/n.
Letting M ∈ R(n
2)×n be the vertex edge incidence matrix of A, and N = [δ1/2MT ,bD1/2], we have
NNT = B. Thus, kNkp = kBkp/2
p/2 and so approximating this norm gives triangle detection by
Corollary 17. Note that nnz(N) = O(nnz(A)) and further N matrix can be formed in this amount
of time, giving our final runtime claim.
Note that for p = 1, since N has maximum row sparsity 2, we obtain a runtime via Theorem
33 of O(ǫ−3(cid:0)mn1/3 + n3/2(cid:1)) = o(nω) for the current value of ω, even when m = n2, implying that
significantly improving this ǫ dependence would either improve matrix multiplication or come at a
cost in the polynomials of the other parameters.
SVD Entropy:
Corollary 19 (SVD Entropy Hardness). Given algorithm A which for any B ∈ Rn×n returns
X ∈ (1 ± ǫ)Pn
i=1 f (σi(B)) for f (x) = x log x in O(nγǫ−c) time, we can detect if an n-node graph
contains a triangle in O(nγ+6c) time.
Proof. For x ∈ (0, 2), using the Taylor Series about 1 we can write x log x =P∞k=0 ck(x− 1)k where
c0 = 1 log(1) = 0, c1 = log(1) + 1 = 1, and ck = (k−2)!
3 . Applying Theorem 15 with δ = 1
k! ≤ 1 for k ≥ 2. ck < c3 for all k > 3 and
10n4 and ǫ1 = δ
30n6 gives the result.
= 0 while c2
c3
3n2 = 1
= 1
c0
c3
Log Determinant:
Corollary 20 (Log Determinant Hardness). Given algorithm A which for any B ∈ Rn×n returns
X ∈ (1 ± ǫ) log(det(B)) in O(nγǫ−c) time, we can detect if an n-node graph contains a triangle in
O(nγ+6c) time.
Proof. Using the Taylor Series about 1 we can write log x =P∞k=0 ck(x−1)k where c0 = 0, ci = 1/i
for i ≥ 1. Therefore ck < c3 for all k > 3, c0
and ǫ1 = δ
2 . Applying Theorem 15 with δ = 1
10n4
= 0, and c2
c3
20n6 gives the result.
2n2 = 1
= 3
c3
In Appendix B, Lemma 52 we show that a similar result holds for computing det(B) =
In [BS83] it is shown that, given an arithmetic circuit for computing det(B), one
i=1 λi(B).
can generate a circuit of the same size (up to a constant) that computes B−1. This also yields a
Qn
25
circuit for matrix multiplication by a classic reduction.3 Our results, combined with the reduction
of [WW10] of Boolean matrix multiplication to triangle detection, show that a sub-cubic time al-
gorithm for the approximating log(det(B)) or det(B) up to sufficient accuracy, yields a sub-cubic
time matrix multiplication algorithm, providing a reduction based connection between determinant
and matrix multiplication analogous to the circuit based result of [BS83].
Trace of Exponential:
Corollary 21 (Trace of Exponential Hardness). Given algorithm A which for any B ∈ Rn×n
returns X ∈ (1 ± ǫ) tr(exp(B)) in O(nγǫ−c) time, we can detect if an n-node graph contains a
triangle in O(nγ+13c) time.
Proof. Using the Taylor Series about 1 we can write ex =P∞k=0
and for all k ≥ 3, ck < c3. Applying Theorem 15 with δ = 1
result.
. We have c0
= 3,
c3
1
6000n13 gives the
10n4 and ǫ1 = c3δ3
= 6, c2
c3
e(x−1)k
c0n =
k!
5.3 Leverage Score and Effective Resistance Hardness
Here we show hardness for precisely computing all effective resistances and leverage scores of a
graph. Our main tool is the following result (which in turn is an easy corollary of Theorem 15) for
an algorithm that precisely computes the trace inverse of a strictly symmetric diagonally dominant
(SDD) B, i.e. B = B⊤ and Bii >Pj6=i Bij.
Corollary 22 (Trace of Inverse Hardness). Given an algorithm which for any strictly SDD B ∈
Rn×n with non-positive off-diagonal entries returns X ∈ (1±ǫ) tr(B−1) in O(nγǫ−c) time for γ ≥ 2,
we can produce an algorithm which detects if an n-node graph contains a triangle in O(nγ+13c) time.
x =P∞k=0(1 − x)k, and then apply Theorem 15 with δ = 1
Proof. For x ∈ (0, 2) we can write 1
and ǫ1 = δ3
off-diagonal entries yields the result.
10n4
1000n13 . Checking that the B in Theorem 15 is strictly SDD with non-positive
n =
1
Using this we prove hardness for precisely computing effective resistances in a graph. Recall
that for a weighted undirected graph G = (V, E, w) its Laplacian, L ∈ RV ×V is given by Lij = −wij
if there is an edge between i and j and 0 otherwise and Lii = −Pi6=j Lij or equivalently L = D− A
where D is the diagonal degree matrix and A is the weighted adjacency matrix associated with
G. Note that this describes a clear bijection between a Laplacian and its associated graph and we
therefore use them fairly interchangeably in the remainder of this section.
The effective resistance between vertices i and j is given by (1i−1j)⊤L†(1i−1j) where † denotes
the Moore-Penrose pseudoinverse. In the following lemma we prove that compute all the effective
resistances between a vertex and its neighbors in the graph can be used to compute the trace of the
inverse of any strictly SDD matrix with non-positive off-diagonals and therefore doing this precisely
is as hard as triangle detection via Corollary 22. Our proof is based off a fairly standard reduction
between solving strictly SDD matrices with negative off-diagonals and solving Laplacian systems.
Lemma 23 (Effective Resistance Yields Trace). Suppose we have an algorithm which given Lapla-
cian L ∈ Rn×n with m-non-zero entries, entry i ∈ [n], and error ǫ ∈ (0, 1) computes a 1 ± ǫ
3Matrix multiplication reduces to inversion by the fact that
0
0
I A 0
I B
0
I
−1
=
I −A AB
−B
0
I
0
0
I
. See [Isa08].
26
approximation to the total effective resistance between i and the neighbors of i in the graph associ-
ated with L, that is
X ∈ (1 ± ǫ) Xj∈[n]:Lij6=0
(1i − 1j)⊤L†(1i − 1j)
in time O(mγǫ−c). Then there is an algorithm that computes the trace of the inverse of n× n strict
SDD matrix with m non-zero entries and non-positive off-diagonals in O(mγǫ−c) time.
Proof. Let M ∈ Rn×n be an arbitrary strictly SDD matrix with non-positive off-diagonals, i.e.
M = M⊤, Mii >Pj6=i Mij, and Mij ≤ 0 for all i 6= j. Let v
def=(cid:18) M −v
α (cid:19) .
def
= 1⊤M1, and
def
= M1, α
−v⊤
L
Now, clearly by our assumptions on M we have that v > 0 entrywise and therefore α > 0. Therefore,
the off-diagonal entries of L are non-positive and by construction L1 = 0. Consequently, L is a
(n + 1) × (n + 1) symmetric Laplacian matrix with nnz(M) + 2n + 1 non-zero entries.
Now, consider any x ∈ Rn and y ∈ R that satisfy the following for some i ∈ [n]
(cid:18) M −v
−v⊤
α (cid:19)(cid:18) x
y (cid:19) =(cid:18) 1i
−1 (cid:19) .
Since L is a symmetric Laplacian and the associated graph is connected by construction we know
that ker(L) = span({1}) and there there must exist such x and y. Furthermore, since M is strictly
SDD it is invertible and since M1 = v we have that
x = M−1 (y · v + 1i) = y · 1 + M−11i
and consequently
(1i − 1n+1)⊤L†(1i − 1n+1) = 1⊤i x − y = 1⊤i M−11i .
Consequently, if we used the algorithm to get a multiplicative approximation X as stated then
X ∈ (1 ± ǫ) tr(M−1) and the result follows.
Using this, we also show that computing leverage scores of matrix, a powerful and prevalent
notion in convex analysis and numerical linear algebra, is also difficult to compute. This follows
from the well known fact that effective resistances in graphs and leverage scores of matrices are the
same up to scaling by known quantities.
Corollary 24 (Leverage Score Hardness). Suppose we have an algorithm which given A ∈ Rn×d
can compute eσ that is a 1 ± ǫ multiplicative approximation to the leverage scores of A, i.e.
eσi ∈ (1 ± ǫ)1⊤i A(A⊤A)†A⊤1i for all i ∈ [n]
in time O(nnz(A)γǫ−c). Then there is a O(n2γ+13c) time algorithm for detecting if an n-node graph
contains a triangle.
Proof. Let L ∈ Rn×n be a symmetric Laplacian. Let E = {{i, j} ⊆ [n] × [n] : Lij 6= 0}, i.e. the
set of edges in the graph associated with L. Let m = E and B ∈ Rm×n be the incidence matrix
associated with L, i.e. for all e = {i, j} ∈ E we have Be,i =p−Lij and Be,j = −p−Lij for some
canonical choice of ordering of i and j and let all other entries of B = 0. Clearly nnz(B) = nnz(L)
and we can form B in O(nnz(L)) time.
27
It is well known and easy to check that L = B⊤b. Consequently, for all e = {i, j} ∈ E we have
1⊤e B(B⊤B)†B⊤1e = (−Lij) · (1i − 1j)⊤L†(1i − 1j)
Now if we computeeσ using the assumed algorithm in O(nnz(L)γǫ−c) = O(n2γǫ−c) time, then since
−Lij is non-negative in an additional O(nnz(L) = O(n2) time this yields a 1 ± ǫ multiplicative
approximation to the total effective resistance between any i and all its neighbors in the graph
associated with L. Consequently, by Lemma 23 and Corollary 22 the result follows.
6
Improved Error Dependence via Polynomial Approximation
For constant ǫ, Theorem 11 and the resulting Corollary 12 matches our fastest runtimes for the
Schatten p-norms. However, it is possible to significantly improve the ǫ dependence by generalizing
our approach. Instead of splitting our spectrum into many small spectral windows, we split into
windows of constant multiplicative width and approximate xp via a low degree polynomial over
each window. The degree of this polynomial only varies with log(1/ǫ).
In Theorem 11, fixing δf to be constant for illustration, each window has width α = Θ(ǫ) and so
there are T = ⌈log1−α λ⌉ = Θ(log λ/ǫ) windows. Additionally, we must set the steepness parameter
to be γ = Θ(ǫα) = Θ(ǫ2). This loses us a total 1/ǫ2 factor in our runtime as compared to our
improved algorithm which will set α = Θ(1) and so γ = Θ(ǫ).
We begin by showing that if we well approximate f on each window, then we well approximate
In the following theorems for simplicity we work with PSD matrices, as we will
Sf (A) overall.
eventually apply these lemmas to AT A.
Lemma 25. Consider f : R → R+, parameters α, ǫ ∈ (0, 1), and gap parameter γ ∈ (0, α). Set
a0 = 1, pick a1 uniformly at random from [1 − α, 1], and set at = a1(1 − α)t−1 for t ≥ 2. Let
Rt = [at+1, at] for all t ≥ 0. Let hγ
be a γ-soft window for Rt (Definition 6) and let ¯Rt =
[(1 − γ)at+1, (1 + γ)at] be the interval on which hγ
is nonzero. Furthermore, for t ≥ 0, let qt be a
'well-behaved' approximation to f on ¯Rt in the following sense:
Rt
Rt
• Multiplicative Approximation: qt(x) − f (x) ≤ ǫf (x) for x ∈ ¯Rt.
• Approximate Monotonicity: qt(y) ≤ c1qt(x) for all y ≤ x for some c1 ≥ 1.
• Range Preserving: qt(0) = 0.
40c1γ
40c1γ
(cid:18)1 − ǫ −
Then, for any PSD A ∈ Rd×d with kAk2 ≤ 1, with probability 9/10:
(A)A) ≤(cid:18)1 + ǫ +
Proof. Note that the restriction that qt(0) = 0 along with the approximate monotonicity property
implies that q(x) is nonnegative, so Sqt is a valid spectral sum and Sqt(hγ
(A)A)),
Rt
so we will be able to estimate this sum via stochastic trace estimation.
α (cid:19)Sf (A) ≤
α (cid:19)Sf (A).
(A)A) = tr(qt(hγ
Rt
∞Xt=0
Sqt(hγ
Rt
Let ti denote the unique index such that σi(A) ∈ Rti. Since γ < α, each σi(A) lies in the
support of at most two overlapping soft windows (in at most two ranges ¯Rti and ¯Rti±1). Let T be
the set of indices whose singular values fall in the support of exactly one soft window and ¯T be its
28
complement. We first bound the error introduced on singular values with indices in T .
∞Xt=0
Sqt(hγ
Rt
(A)A) =
Rt
dXi=1
=Xi∈T
∈ (1 ± ǫ)Xi∈T
∞Xt=0
qt(cid:0)σi(A) · hγ
(σi(A))(cid:1)
∞Xt=0
qti(σi(A)) +Xi∈ ¯T
qt(cid:0)σi(A) · hγ
∞Xt=0
f (σi(A)) +Xi∈ ¯T
(σi(A))(cid:1)
qt(cid:0)σi(A) · hγ
Rt
Rt
(16)
(17)
(σi(A))(cid:1) .
ǫf (x) for x ∈ ¯Rt. Let us now consider a particular i ∈ ¯T and boundP∞t=0 qt(cid:0)σi(A) · hγ
The last inequality follows from the multiplicative approximation assumption on qt that qt(x) − f (x) ≤
Note that there are precisely two non-zero terms in this summation -- one corresponding to ti and
the other, to ti ± 1, which we denote by t¯i. Using the above hypothesis on qt again, we see that
(σi(A))(cid:1).
(15)
Rt
qti(cid:16)σi(A) · hγ
Rti
(σi(A))(cid:17) = qti (σi(A)) ∈ (1 ± ǫ)f (σi(A)).
For the term involving t¯i, we have σi(A) ∈ ¯R¯ti and by the approximate monotonicity requirement
that y ≤ x ⇒ qt(y) ≤ c1qt(x) have:
qt¯i(cid:16)σi(A) · hγ
Rt¯i
(σi(A))(cid:17) ≤ c1qt¯i(σi(A)) ∈ c1(1 ± ǫ)f (σi(A)).
Due to the random positioning of the windows, in expectation, the total contribution of the
singular values lying at the intersection of two windows is small. Specifically,
E
α1Xi∈ ¯T
f (σi(A)) =
dXi=1
P[i ∈ ¯T ] · f (σi(A)) ≤
2γ
α
dXi=1
f (σi(A)) =
2γ
α Sf (A)
(18)
σi(A)
where the bound on P[i ∈ ¯T ] follows from the fact that i ∈ ¯T only if a1(1 − α)t ∈ (1 ± γ)σi(A) for
some t. This holds only if a1 ∈ (1 ± γ)(cid:18)
α since
a1 is chosen uniformly in the range [1− α, 1]. By a Markov bound applied to (18), with probability
9/10 we have Pi∈ ¯T f (σi(A)) ≤ 20γ
α Sf (A). Combining with (15), (16), and(17), we obtain, with
probability 9/10:
α (cid:19)Sf (A) ≤
⌈logσi(A) 1−α⌉(cid:19), which occurs with probability 2γ
(A)A) ≤(cid:18)1 + ǫ +
α (cid:19)Sf (A).
(cid:18)1 − ǫ −
Sqt(hγ
(1−α)
40c1γ
40c1γ
Rt
∞Xt=0
We now show that, as long as the contribution of the smaller singular values of A to Sf (A) is
not too large, we can truncate this sum and still obtain an accurate approximation. Specifically:
Corollary 26. For any PSD A ∈ Rd×d with kAk2 ≤ 1, let f : R → R+ be a function such that,
for any ǫ > 0 there exists λf (ǫ) such that for x ∈ [0, λf (ǫ)], f (x) ≤ ǫ
nSf (A). Given parameters
α, ǫ ∈ (0, 1), and gap parameter γ < α, for t ≥ 0 define Rt, hγ
and, ¯Rt as in Lemma 25. Let qt be
a well-behaved approximation to f on ¯Rt as in Lemma 25. With probability 9/10:
(cid:18)1 −
α − 5c1ǫ(cid:19)Sf (A) ≤
+ 5c1ǫ(cid:19)Sf (A).
(A)A) ≤(cid:18)1 +
Sqt(hγ
40c1γ
40c1γ
(19)
Rt
Rt
α
⌈log1−α λf (ǫ)⌉Xt=0
29
Proof. This follows from Lemma 25 along with the small tail assumption. Specifically, by (17):
∞X⌈log1−α λf (ǫ)⌉+1
Sqt(hγ
Rt
(A)A) ≤ X{iσi(A)≤λf (ǫ)}
4c1f (σi(A)) ≤ 4c1n ·
ǫ
nSf (A) = 4c1ǫSf (A).
We now show that well behaved polynomial approximations exist for the function xp for general
real p, whose spectral sums give the Schatten p-norms.
Lemma 27 (Polynomial Approximation of Power Function). For all p ∈ [−1, 0) and k ≥ 0 let
def
=
ak
kYj=1(cid:18)1 −
p + 1
j (cid:19) and qk(x) =
kXj=0
aj(1 − x)j .
Then for all x ∈ (0, 1] and k ≥ 0 we have 0 ≤ ak ≤ 1 and
0 ≤ xp − qk(x) ≤
exp(−kx)
x
.
Proof. Induction shows that the k-th derivative of f (x) = xp at x ∈ R, is given by
Furthermore, direct calculation reveals that for all x, t ∈ R and k ≥ 1
(p + 1 − j) xp−k .
f (k)(x) =
kYj=1
(x − t)k =
· tp−k · (t − x)k · (−1)k = aktp(cid:16)1 −
kYj=1
p + 1 − j
j
f (k)(t)
k!
x
t(cid:17)k
.
(20)
In the last step we took the worst case of t = 1 in the integral and used that since p ∈ [−1, 0) it is
the case that 1− (p + 1)/j ∈ (0, 1] and therefore 0 ≤ ak ≤ 1. Consequently, limi→∞ qk(x) = xp and
0 ≤ xp − qk(x) =
(1 − x)j ≤
exp(−kx)
x
.
1
(x − t)kdt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∞Xj=k+1
= ak · p − k ·(cid:12)(cid:12)(cid:12)(cid:12)Z x
aj(1 − x)j ≤ (1 − x)k+1 ∞Xj=0
Consequently, qk(x) is the degree k Taylor expansion of xp about 1 evaluated at x. Furthermore,
since f (k+1)(t) = p−k
· f (k)(t), the integral form of the remainder of a Taylor series expansion shows
t
f (k+1)(t)
x
tp−1(cid:16)1 −
t(cid:17)k
dt(cid:12)(cid:12)(cid:12)(cid:12) ≤
k − p
x
exp(−kx) .
xp − qk(x) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z x
1
k!
Corollary 28. For any p, ǫ > 0, a, b ∈ (0, 1] with a < b, there is a polynomial q of degree
aǫ(cid:1) · b
a + p(cid:1) such that:
O(cid:0)log(cid:0) b
• Multiplicative Approximation: q(x) − xp ≤ ǫxp for x ∈ [a, b].
• Monotonicity: q(y) < q(x) for all y < x.
30
• Range Preserving: q(0) = 0.
Proof. Set p′ = p − ⌈p⌉ and i = c1 log(cid:0) b
a for large enough constant c1. Instantiate Lemma 27
with p′ and i to obtain qi(x). Set q(x) = bp · (x/b)⌈p⌉ · qi(x/b). It is clear that q(0) = 0. Further,
aǫ(cid:1) · b
xp − q(x) = bp(x/b)⌈p⌉ · (x/b)p′
− qi(x/b)
aǫ(cid:19) b
· exp(cid:18)−c1 log(cid:18) b
For x ∈ [a, b], x/b ≥ a/b. Further, (b/x)1+p′
− qi(x) =P∞j=i+1 aj(1 − x)j. All aj
xp − q(x) < ǫxp. We finally show monotonicity. We have xp′
are positive, so this difference is monotonically decreasing in x. We thus have, for y < x:
≤ b/a and so if we choose c1 large enough we have
≤ xp(b/x)1+p′
b(cid:19)
a ·
x
q(y)
q(x)
=
(y/b)⌈p⌉qi(y/b)
(x/b)⌈p⌉qi(x/b) ≤
y
x ·
qi(y/b)
qi(x/b) ≤
y
x ·
(y/b)p′
(x/b)p′ ≤ 1
since p′ ∈ [0, 1]. This gives us monotonicity.
We now combine the approximations of Corollary 28 with Corollary 26 to give our improved
Schatten norm estimation algorithm.
Lemma 29 (Schatten Norms via Polynomial Approximation). For any A ∈ Rn×d with kAk2 ≤ 1,
d2/p (cid:19)kAk2
p > 0 and error parameter ǫ ∈ (0, 1), for sufficiently small c1, c2, let α = 1/2, λ =(cid:18) c2/p
p,
and γ = c2ǫ. For t ≥ 0 define Rt, hγ
, and ¯Rt as in Lemma 25. Let qt(x) be as defined in Corollary
28 with p′ = p/2, ǫ′ = c3ǫ for sufficiently small c3 and [a, b] = Rt. With probability 9/10:
⌈log 1/λ⌉Xt=0
(AT A)AT A) ∈ (1 ± ǫ)kAkp
p .
Sqt(hγ
(21)
Proof. We apply Corollary 26 to AT A, with f (x) = xp/2, ǫ = c1ǫ, γ = c2ǫ and α = 1/2. qt
ǫ2/p
Rt
Rt
1
Plugging in α = 1/2, γ = c2ǫ into (19) we have:
satisfies the necessary conditions by Corollary 28. Further, we have λf (c1ǫ) =(cid:18) c2/p
1/2 ± c1ǫ(cid:19)Sf (AAT ) = kAkp
(AT A)AT A) ∈(cid:18)1 ± c3ǫ ±
Sqt(hγ
40c2ǫ
Rt
p
1
⌈log 1/λ⌉Xt=0
ǫ2/p
d2/p (cid:19)kAk2
p = λ.
max{p, 1/p3}
which gives (1 ± ǫ) approximation if we set c1, c2, c3 small enough.
Theorem 30 (Schatten Norm Polynomial Approximation Algorithm). For any A ∈ Rn×d, p > 0
and k ∈ [d] there is an algorithm returning X ∈ (1 ± ǫ)kAkp
O(cid:18)nnz(A)k + dkω−1 +
ǫmax{3,1+1/p} ·hnnz(A) +pnnz(A)[d · ds(A) + dk] · (d/k)max{0,1/p−1/2}i(cid:19)
ǫmax{3,1+1/p} ·h(nnz(A) + dk)(d/k)1/pi(cid:19)
kmax{0,1/p−1/2}#!
or O(cid:18)nnz(A)k + dkω−1 +
ǫmax{3,1+1/p} ·"nd +
max{p, 1/p3}
O ndω(logd k)−1 +
max{p, 1/p3}
p that runs in time:
time for sparse A or
n1/2d1+1/p
for dense A.
31
¯κt =
Similarly, κt =
σ2
k+1(A)
at
≤ 1
δ2/p
t
dat
i (A)
p for some δt ≥ ǫ,
k=1 σ2
1
δ2/p
d2/p · kAk2
k(A) +Pd
kσ2
k(cid:1)2/p
(cid:0) d
(AT A)AT A)) up to multiplicative error 1 ± O(cid:16)
k(cid:19)2/p−1
t (cid:18) d
≤
.
. We can further optimize our ǫ dependence by noting that it
suffices to approximate tr(qt(hγ
Rt
since even if there are d singular values below at−1, they will contribute at most a δt fraction of
kAkp
p. So our total cost of approximating the trace for each of our windows using the first runtime
of Theorem 7 is:
ǫ
(AT A)AT A) = tr(qt(hγ
Rt
Proof. We apply Lemma 29, first scaling A so that kAk2 ≤ 1. We apply a trace estimator to
approximate Sqt(hγ
(AT A)AT A)) up to (1 ± ǫ) multiplicative error plus
very small additive error for each of our O(log 1/λ) = O(max{1, 1/p}) windows. The trace estima-
tion requires O(cid:16)⌈p⌉ǫ2(cid:17) applications of hγ
(AT A)AT A, since the degree of qt as given in Corollary
28 is O(1) + p.
Rt
Rt
(AT A)AT A to a vector is given by Theorem 7. Following the argument
The cost to apply hγ
Rt
in Corollary 12, if we write at = δ2/p
t
min{1,δt} log(1/λ)(cid:17)
1
δ1/p
t
·
t
ǫ2
⌈log 1/λ⌉Xt=0
min{1, δt}2 · ⌈p⌉(log 1/λ)2
nnz(A) +pnnz(A)[d · ds(A) + dk] · (d/k)1/p−1/2
O
Factoring out the nnz(A) +pnnz(A)[d · ds(A) + dk] · (d/k)1/p−1/2 term, we have:
O
⌈log 1/λ⌉Xt=0
= O(cid:18) max{p, 1/p2}
min{1, δt}2 · max{p, 1/p2}
ǫ1/p−2(cid:18)1 +
+ X{t:δt≥1}
+ ... + ǫ1/p−2(cid:19) + log(1/λ)(cid:21)(cid:19) .
= O max{p, 1/p2}
· X{t:δt<1}
ǫ · δ1/p
δ1/p−2
t
·(cid:20)
41/p−2
21/p−2
ǫ3δ1/p
ǫ3
ǫ3
+
1
1
1
1
t
Note that if p > 2, 1/p − 2 < 0 the terms in the geometric sum are increasing so the largest term
is ǫ1/p−2 and so the whole thing is dominated by O( p log(1/λ)
ǫ3 ). If p < 2, then the largest
term in the geometric sum is 1 as so similarly the whole thing is O( 1/p2 log(1/λ)
ǫmax{3,1+1/p} ) = O(
ǫmax{3,1+1/p} ).
A similar argument gives the runtimes for standard iterative and dense methods. Adding in the
cost for deflation-based preconditioning gives the final runtimes. Note that while λ depends on kAkp
p
which is unknown, it again suffices to lower bound the truncation point using σp
k kAkp
k(A) ≤ 1
p.
This lower bound gives only a better approximation and our bounds on ¯κt and κt still hold.
) = O( p
1/p3
ǫ3
7 Optimized Runtime Results
The runtimes given in Theorem 30 are quite complex, with many tradeoffs depending on the
properties of A. In this section, we instantiate the theorem showing how to optimize k and achieve
our final runtime bounds. For simplicity, we consider the case when n = d.
7.1 Schatten p-Norms
We first tackle the relatively simply case when A is dense.
32
Theorem 31 (Schatten p-Norms for Dense Matrices). For any p ≥ 0 and A ∈ Rn×n there is an
algorithm returning X ∈ (1 ± ǫ)kAkp
O(cid:18)
p which runs in O(p · n2/ǫ3) time for p ≥ 2 and
p3 · ǫmax{3,1+1/p} · n
1+.0435p (cid:19)
2.3729−.0994p
1
time for p < 2. If we do not use fast matrix multiplication the runtime is O(cid:18)
1
p3·ǫmax{3,1+1/p} · n
3+p/2
1+p/2(cid:19).
Note that for the important case of p = 1 our runtime is O(n2.18/ǫ3) or O(n2.33/ǫ3) if we do not
use fast matrix multiplication. As p approaches 0, our runtime approaches O(nω) which is natural,
as p = 0 gives the matrix rank, which seems difficult to determine. As p approaches 2 the runtime
smoothly approaches O(n2/ǫ3), which is then required for all p ≥ 2.
Proof. The bound for p ≥ 2 follows immediately from Theorem 30 with k set to 0. For p < 2 we
have runtime:
O nω(logn k) +
p3 · ǫmax{3,1+1/p} ·" n3/2+1/p
k1/p−1/2#!
1
We can partially optimize this term, ignoring p and ǫ factors for simplicity and setting k to equalize
1/p−1/2
1/p+1/2
the coefficients on n. Our optimization is identical to the argument in Corollary 12 (set k = n
when not using fast matrix multiplication), giving the stated runtimes.
We now tackle the more complex case when A is sparse. We first consider p ≥ 2.
Theorem 32 (Schatten p-norms, p ≥ 2, for Sparse Matrices). For any p ≥ 2, and A ∈ Rn×n there
is an algorithm returning X ∈ (1 ± ǫ)kAkp
p with high probability in time:
O(cid:16) p
ǫ3 ·pnnz(A)nds(A)(cid:17) or O(cid:16) p
ǫ3hnnz(A)n
1
1+p + n1+ 2
1+pi(cid:17)
Note that if our matrix is uniformly sparse, ds(A) = O(nnz(A)/n) and so the first runtime
becomes O(p · nnz(A)/ǫ3). The second runtime can be better when A is not uniformly sparse.
Proof. The first runtime follows by setting k = 0 in Theorem 30. Note that nnz(A) ≤pnnz(A)nds(A).
For the second, we use the second runtime of Theorem 30 which gives
O(cid:16)nnz(A)k + nkω−1 +
ǫ3 ·h(nnz(A) + nk)(n/k)1/pi(cid:17)
k(cid:1)1/p to balance the coefficients on the nnz(A) terms gives k = n
1
1+p which yields the
result. Note that the nkω−1 term is dominated by the cost of the regressions, even if we do not use
fast matrix multiplication this term will be nk2.
Setting k =(cid:0) n
p
Finally, we consider the most complex cost, p ≤ 2 for sparse matrices. We have:
Theorem 33 (Schatten p-norms, p ≤ 2, for Sparse Matrices). For any p ∈ (0, 2], and A ∈ Rn×n
there is an algorithm returning X ∈ (1 ± ǫ)kAkp
p3 · ǫmax{3,1/p} ·(cid:20)nnz(A)n
1/p+1/2√γs +pnnz(A) · n
p with high probability in time:
O(cid:18)
1/p−1/2
1
or O(cid:18)
1
p3 · ǫmax{3,1/p}hnnz(A)n
1
1+p + n1+ 2
4/p−1
2/p+1(cid:21)(cid:19)
1+pi(cid:19)
where γs = ds(A)n
nnz(A) ≥ 1.
33
Note that in the special case of p = 1 the first runtime gives O(cid:16) 1
The second term here is at worst n2, and could be much smaller for sparse A.
ǫ3hnnz(A)n1/3√γs +pnnz(A)ni(cid:17).
Proof. For the second runtime, we use the second runtime of Theorem 30, balancing costs exactly
1
as in Theorem 32 (setting k = n
1+p ). For the first runtime, we consider the first runtime of
Theorem 30 which gives:
O(cid:18)nnz(A)k + nkω−1 +
O nds(A)k
γs
1
p3 · ǫmax{3,1/p} ·hnnz(A) +pnnz(A)[n · ds(A) + nk] · (n/k)1/p−1/2i(cid:19)
k(cid:17)1/p−1/2!
# ·(cid:16) n
p3 · ǫmax{3,1/p} ·" nds(A) + npds(A)√k
+ nkω−1 +
√γs
1
Ignoring ǫ, p, and γs dependence, one way we can balance the costs is by setting:
nds(A)k = nds(A)(cid:16) n
k(cid:17)1/p−1/2
1/p−1/2
1/p+1/2 and final runtime:
which gives k = n
1
p3 · ǫmax{3,1/p} ·(cid:20)nnz(A)n
O(cid:18)
p3 · ǫmax{3,1/p} ·(cid:20)nnz(A)n
O(cid:18)
1/p−1/2
1/p+1/2√γs + npds(A)/γs · k3/2(cid:21)(cid:19) =
2/p+1 (cid:21)(cid:19) .
1/p+1/2√γs + npds(A)/γs · n
3/p−3/2
1/p−1/2
1
Note that we drop the nkω−1 term as it is dominated by the last term with k3/2 in it. This gives
our final result by noting that ds(A)/γs = nnz(A)/n, the average row sparsity.
1/p−1/2
p3(cid:20)nnz(A)n
7.2 Constant Factor Approximation without Row-Sparsity Dependence
Theorem 34 (Removing the Uniform Sparsity Constraint). Let γ ∈ (0, 1), p ≥ 1, and A ∈ Rn×n.
There is an algorithm returning X with kAkp ≤ X = O(1/γ)kAkp with high probability in time
2/p+1(cid:21)(cid:19) for p < 2.
O(p nnz(A)nγ) for p ≥ 2, and in time O(cid:18) 1
1/p+1/2 +γ/2 +pnnz(A) · n
Proof. For i ∈ {1, 2, . . . ,⌈ 1
γ⌉}, let Ai ∈ Rn×n be the matrix whose rows agree with that of A
, nγ(i+1)· nnz(A)
provided the number of non-zero entries in these rows are in the interval [nγi· nnz(A)
),
while the remaining rows of Ai are set to 0n. Let A0 = A −P⌈ 1
γ ⌉
i=1 Ai. Since p ≥ 1 we can apply
i kAikp.
1
γ ⌉ + 1) max
the triangle inequality,
kAikp ≤ (⌈
kAkp ≤
(22)
⌈ 1
4/p−1
n
n
γ ⌉Xi=0
For any matrix B obtained from a matrix A by replacing some rows of A with 0n and preserving
the remaining rows, kBkp ≤ kAkp. This follows from the fact that kBxk2 ≤ kAxk2 for all vectors
x, together with the min-max theorem for singular values. Hence,
kAkp ≥ max
i kAikp.
34
(23)
Combining (22) and (23), an O(1)-approximation to maxi kAikp is an O( 1
γ )-approximation to kAkp.
We remove the zero rows from the Ai, obtaining corresponding matrices Bi. By definition of Ai,
the number of rows in Bi is at most n1−γi. For each Bi, which is an si × n matrix for si ≤ n1−γi,
we can right-multiply it by an n × ti OSNAP matrix Ti, with ti = O(n1−γi log n) columns and
O(log n) non-zero entries per column, so that kBiTikp = (1 ± 1/2)kBikp for all i with probability
1 − 1/ poly(n), see [Coh16] (the fact that kBiTikp = (1 ± 1/2)kBikp follows from the fact that Ti
is a subspace embedding of the row space of Bi together with Lemma C.2 of [LNW14]). The time
to compute Ci = BiTi is O(nnz(A)). Since Ti has O(log n) non-zero entries per row, each row of
Ci has a number of non-zero entries within an O(log n) factor of the corresponding row of Bi.
Let di
s denote the parameter ds when run on Ci. Then di
We compute O(1)-approximations to the kCikp for each i, and then take the maximum over i.
To do so, for p ≥ 2 we run the algorithm of Theorem 32 on each Ci, while for p < 2 we run the
algorithm of Theorem 33.
log n). We can
assume Ci is a square ti × ti matrix by padding with zeros, where ti = O(n1−γi log n). Hence,
s = O(nnz(A)nγ log n). Also, nnz(Ci) = O(nnz(Bi) log n) = O(nnz(A) log n).
tidi
The total time to apply Theorem 32 for p ≥ 2 across all Ci is O(p nnz(A)nγ). The total time
2/p+1(cid:21)(cid:19),
to apply Theorem 33 for p < 2 across all Ci is O(cid:18) 1
p3(cid:20)nnz(A)n
1/p+1/2 +γ/2 +pnnz(A) · n
using that the γs of that theorem is O(nγ log n).
s = O(nγi+γ nnz(A)
1/p−1/2
4/p−1
n
7.3 SVD Entropy
In this section, we will show how to approximately estimate the SVD entropy of a matrix A
def
= σ1(A)/σn(A) to be bounded by nc1. Recall that the SVD
assuming its condition number K
is the matrix
entropy of a matrix A is given by H(B)
A scaled inversely by its Schatten 1-norm. Using the results in Section 7.1, we obtain a (1 + ǫ)
def
= A
kAk1
def
= −Pi σi(B) log σi(B), where B
approximation Z to kAk1 i.e., kAk1 ≤ Z ≤ (1 + ǫ)kAk1. Define eB
Z (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xi
kAk1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)H(eB) − H(B)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xi (cid:18)σi(A)
≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xi
σi(A)(cid:18) 1
1 + ǫ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xi
kAk1 −
Z(cid:19) log
kAk1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
The following thoerem gives our main result for SVD entropy.
+ ǫ ≤ 2ǫ log n.
kAk1 −
σi(A)
kAk1
kAk1
def= A
σi(A)
σi(A)
σi(A)
σi(A)
σi(A)
≤
log
log
Z
1
log
ǫ
Z . This means that
σi(A)
Z
log kAk1
Z (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Theorem 35. Given any PSD A ∈ Rn×n and error parameter ǫ ∈ (0, 1). Then, there exists
an algorithm that computes the approximate SVD entropy of A and outputs S, such that, with
probability 9/10
S ∈ (1 ± O(ǫ))H(B),
(24)
where B = A
kAk1
. The runtime of the algorithm will be
2√ds
3
O nnz(A)n
1
3 + n
ǫ6
log(1/ǫ)! ,
35
Algorithm 2 SVD entropy estimation via multi-point interpolation.
for i = 1 : k1 do
Z , where Z is a (1 + ǫ) approximation to kAk1.
Input: A ∈ Rn×n, ǫ ∈ (0, 1).
Output: Approximate SVD entropy S.
Compute eB = A
Choose k1 points α1,··· , αk1 as in (25), and seteǫ = ǫ/(12c1(k1 + 1)3 log n).
Compute Z1+αi, a (1 +eǫ) -approximation for Schatten norm(cid:13)(cid:13)(cid:13)eB(cid:13)(cid:13)(cid:13)1+αi
1+αi (eB).
end for
Return S an estimate of T (0) by interpolation using the points T SV D
1+αi (eB) = (1 − Z1+αi/Z 1+αi)/αi.
Compute T SV D
.
where ds(A) = O(nnz(A)/n), and ǫ = O( ǫ
log n ).
The algorithm depends on the βth SVD Tsallis entropy T SV D
The algorithm we consider to estimate the SVD entropy of eB follows the techniques of [HNO08].
(eB) of matrix eB, which is given by
β
β
1 −(cid:13)(cid:13)(cid:13)eB(cid:13)(cid:13)(cid:13)
β − 1
β
.
T SV D
β
(eB) =
We can describe the multi-point interpolation method for SVD entropy estimation with additive
and multiplicative error approximations using [HNO08, Algorithm 1], see Algorithm 2.
Here, given an error parameter ǫ, we compute an approximate SVD Tsallis entropy T SV D
def
= log 1
1+αi at k1
ǫ + log c1 + log log n. α1,··· , αk1 are defined
different points {1 + α0, . . . , 1 + αk1}, where k1
as follows. Let ℓ
def
= 1/(2c1(k1 + 1) log n) and let g(·) be defined as:
g(y) =
k2
1ℓy − ℓ(k2
1 + 1
2k2
1 + 1)
, then, αi
def
= g(cos(iπ/k1)).
(25)
2c1(k1+1) log n . We have
smallest power Schatten norm. We get the smallest power when αi =
k1 = log(1/ǫ) such Schatten norms to be estimated. Thus, we obtain the runtime in Theorem 35 by
Algorithm 2 gives the stepwise algorithm. The runtime of the algorithm is dominated by the
, particularly the
−1
Now define an error parametereǫ def= ǫ/(12c1(k1 + 1)3 log n).
cost of computing Z1+αi, a (1 +eǫ) -approximation for Schatten norm (cid:13)(cid:13)(cid:13)eB(cid:13)(cid:13)(cid:13)1+αi
using p = α = O(cid:16)1 − 1
and observing that n1/ log(n/ǫ) is between 1 and 2 (we set p ≈ 1).
Additive approximation : Section 3.3.2 of [HNO08] tells us that if in Algorithm 2, the ap-
proximate T SV D
log(n/ǫ)(cid:17) in the runtime for Schatten norm estimation given in Corollary 12,
1+αi (eB) is an additive eǫ approximation to T SV D
use these to compute an additive-ǫ approximation to H(eB). Since
and since(cid:13)(cid:13)(cid:13)eB(cid:13)(cid:13)(cid:13)1
1 log n multiplicative approximation to(cid:13)(cid:13)(cid:13)eB(cid:13)(cid:13)(cid:13)1+αk1
approximation can be obtained using the results of Section 7.1.
1+αi (eB) for every i ∈ [k1], then we can
−1
1 log n ,
16c1k3
suffices. This
2c1k1 log n < αi <
= 1, obtaining
1
16c1k3
−1
36
1+αi (B) at each αi.
If the matrix has reasonable (large) SVD entropy (meaning the singular values are uniform and
T SV D
using results from section 7.1.
This issue is equivalent to the 'heavy element' issue (one of the entries in the vector is very
large) discussed in [HNO08]. Hence, to overcome the above issue, we can follow the workaround
However, if the matrix has very small entropy, i.e., we have one singular value with very large mag-
nitude and remaining singular values are very small, then achieving a multiplicative approximation
will be difficult. This is because, we are approximating T SV D
Multiplicative approximation : Article [HNO08] further extends the multi-point interpola-
tion method to achieve multiplicative approximation for Shannon entropy (equivalently for SVD
entropy) using the following modifications. We set the number of interpolation points k1 =
max{5, log(1/ǫ)}. Then, section 3.4 in [HNO08] shows that if T SV D
i ∈ [k1] computed are to be (1 +eǫ)-multiplicative approximation to T SV D
above), then we can achieve multiplicative approximation for H(B) using the multi-point interpo-
lation method with these T SV D
the T SV D
1+αi (eB) (in Algorithm 2) for every
1+αi (B) (whereeǫ is as defined
1+αi (eB). So, we need to obtain (1 +eǫ)-relative error approximations to
there is no one singular value that is very large), then we can obtain (1 ±eǫ) approximation to the
1+αi (B) by simply computing (1 ±eǫ) approximation to (cid:13)(cid:13)(cid:13)eB(cid:13)(cid:13)(cid:13)1+αi
1+αi of eB, which is matrix A scaled by
1+αi (eB)
a (1 ± ǫ) approximation of its Schatten 1-norm. The (1 +eǫ) approximation obtained for T SV D
proposed in [HNO08]. Specifically, Lemma 5.5 in [HNO08] shows that a (1 +eǫ)-approximation to
1 − σ1 together with a (1 +eǫ)-approximation to Pj>1 σ1+αi
singular values of B withPi σi = 1, suffices to get a (1 +eǫ)-approximation to T SV D
follows that we just need to obtain (1 +eǫ)-approximations for these latter two quantities for each
spectrum) is preserved, see Appendix C for the proof. This means we can get (1+eǫ)-approximations
to both kA−1k1 and (kA−1k1+αi)1+αi , where A−1 is matrix A with the top singular vector exactly
deflated off, by running the algorithm presented in this paper. That is, get the Schatten 1 and
Schatten (1 + αi) norms for the matrix A with the top singular vector deflated off. We can also
approximate kAk1 up to (1 +eǫ) relative error using our algorithm. Then, the two quantities
above (1− σ1 andPj>1 σ1+αi
) are exactly kA−1k1/kAk1 and (kA−1k1+αi)1+αi/kAk1, respectively.
Since we have relative (1 +eǫ)-approximations to the numerators and denominators of both these
quantities, we obtain then up to (1 +eǫ)-relative error the quantities 1 − σ1 and Pj>1 σ1+αi
, as
needed to achieve a multiplicative approximation to the SVD entropy. Note that we do not need to
compute kAk1 exactly here even when the matrix has very low SVD entropy. A similar argument
can be seen in section 6 of [HNO08].
It can be shown that when the Krylov method [MM15] (or power method) is used to deflate
the top singular vector, we have that any unitarily invariant norm of the tail (remaining part of the
j
j
might not be close to T SV D
1+αi (B) in this case.
αi.
j
, where σ1 ≥ σ2 ≥ . . . ≥ σn are the
1+αi (B) at αi. It
Acknowledgements
We thank Vladimir Braverman for pointing out an error in our original proof of Theorem 8, which
has been corrected.
References
[ABB00] Orly Alter, Patrick O Brown, and David Botstein. Singular value decomposition for
genome-wide expression data processing and modeling. Proceedings of the National
37
Academy of Sciences, 97(18):10101 -- 10106, 2000.
[AKR15] Alexandr Andoni, Robert Krauthgamer, and Ilya P. Razenshteyn. Sketching and em-
bedding are equivalent for norms. In Proceedings of the 47th Annual ACM Symposium
on Theory of Computing (STOC), pages 479 -- 488, 2015.
[AT11]
[AZ17]
Haim Avron and Sivan Toledo. Randomized algorithms for estimating the trace of an
implicit symmetric positive semi-definite matrix. Journal of the ACM, 58(2):8, 2011.
Zeyuan Allen-Zhu. Katyusha: The first direct acceleration of stochastic gradient
methods. Proceedings of the 49th Annual ACM Symposium on Theory of Computing
(STOC), 2017.
[AZL16]
Zeyuan Allen-Zhu and Yuanzhi Li. Faster principal component regression via optimal
polynomial approximation to sgn(x). arXiv:1608.04773, 2016.
[BCF09]
Constantine Bekas, Alessandro Curioni, and I Fedulova. Low cost high performance
uncertainty quantification. In Proceedings of the 2nd Workshop on High Performance
Computational Finance, page 8. ACM, 2009.
[BCKY16] Vladimir Braverman, Stephen R Chestnut, Robert Krauthgamer, and Lin F Yang.
Sketches for matrix norms: Faster, smaller and more general. arXiv:1609.05885, 2016.
[BDKZ15] Christos Boutsidis, Petros Drineas, Prabhanjan Kambadur, and Anastasios Zouzias. A
randomized algorithm for approximating the log determinant of a symmetric positive
definite matrix. arXiv:1503.00374, 2015.
[BP14]
[BS83]
[Car14]
Monami Banerjee and Nikhil R Pal. Feature selection with SVD entropy: some modi-
fication and extension. Information Sciences, 264:118 -- 134, 2014.
Walter Baur and Volker Strassen. The complexity of partial derivatives. Theoretical
computer science, 22(3):317 -- 330, 1983.
Petre Caraiani. The predictive power of singular value decomposition entropy for stock
market dynamics. Physica A: Statistical Mechanics and its Applications, 393:571 -- 578,
2014.
[CLM+15] Michael B. Cohen, Yin Tat Lee, Cameron Musco, Christopher Musco, Richard Peng,
and Aaron Sidford. Uniform sampling for matrix approximation. In Proceedings of the
6th Conference on Innovations in Theoretical Computer Science (ITCS), pages 181 --
190, 2015.
[Coh16]
Michael B. Cohen. Nearly tight oblivious subspace embeddings by trace inequalities.
In Proceedings of the 27th Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA), pages 278 -- 287, 2016.
[CR12]
Emmanuel J. Cand`es and Benjamin Recht. Exact matrix completion via convex opti-
mization. Communications of the ACM, 55(6):111 -- 119, 2012.
[Dem13]
James Demmel. An arithmetic complexity lower bound for computing rational func-
tions, with applications to linear algebra. submitted to SIMAX, 2013.
38
[DHJZ15] Haishun Du, Qingpu Hu, Manman Jiang, and Fan Zhang. Two-dimensional principal
component analysis based on Schatten p-norm for image feature extraction. Journal of
Visual Communication and Image Representation, 32:55 -- 62, 2015.
[DKJ+07] Jason V Davis, Brian Kulis, Prateek Jain, Suvrit Sra, and Inderjit S Dhillon.
Information-theoretic metric learning. In Proceedings of the 24th International Confer-
ence on Machine Learning (ICML), pages 209 -- 216, 2007.
[DNPS16] Edoardo Di Napoli, Eric Polizzi, and Yousef Saad. Efficient estimation of eigenvalue
counts in an interval. Numerical Linear Algebra with Applications, 2016.
[DTV11] Amit Deshpande, Madhur Tulsiani, and Nisheeth K. Vishnoi. Algorithms and hardness
for subspace approximation. In Proceedings of the 22nd Annual ACM-SIAM Symposium
on Discrete Algorithms (SODA), pages 482 -- 496, 2011.
[FGKS15] Roy Frostig, Rong Ge, Sham M Kakade, and Aaron Sidford. Un-regularizing: approxi-
mate proximal point and faster stochastic algorithms for empirical risk minimization. In
Proceedings of the 32nd International Conference on Machine Learning (ICML), 2015.
[FHT08]
Jerome Friedman, Trevor Hastie, and Robert Tibshirani. Sparse inverse covariance
estimation with the graphical lasso. Biostatistics, 9(3):432 -- 441, 2008.
[FMMS16] Roy Frostig, Cameron Musco, Christopher Musco, and Aaron Sidford. Principal com-
ponent projection without principal component analysis. 2016.
[FORF16] JK Fitzsimons, MA Osborne, SJ Roberts, and JF Fitzsimons.
Improved stochastic
trace estimation using mutually unbiased bases. arXiv:1608.00117, 2016.
[GHJ+16] Dan Garber, Elad Hazan, Chi Jin, Sham M. Kakade, Cameron Musco, Praneeth Ne-
trapalli, and Aaron Sidford. Faster eigenvector computation via shift-and-invert pre-
conditioning. In Proceedings of the 33rd International Conference on Machine Learning
(ICML), 2016. Full version at arXiv:1605.08754.
[GOSS16] Alon Gonen, Francesco Orabona, and Shai Shalev-Shwartz. Solving ridge regression us-
ing sketched preconditioned SVRG. In Proceedings of the 33rd International Conference
on Machine Learning (ICML), 2016.
[GU17]
[Gut92]
[Gut01]
Fran¸cois Le Gall and Florent Urrutia. Improved rectangular matrix multiplication using
powers of the Coppersmith-Winograd tensor. arXiv:1708.05622, 2017.
Ivan Gutman. Total π-electron energy of benzenoid hydrocarbons. In Advances in the
Theory of Benzenoid Hydrocarbons II, pages 29 -- 63. Springer, 1992.
Ivan Gutman. The energy of a graph: old and new results. In Algebraic combinatorics
and applications, pages 196 -- 211. Springer, 2001.
[GVL12] Gene H. Golub and Charles F Van Loan. Matrix computations, volume 3. JHU Press,
2012.
[GXL15]
Rongbao Gu, Wei Xiong, and Xinjie Li. Does the singular value decomposition entropy
have predictive power for stock market? evidence from the shenzhen stock market.
Physica A: Statistical Mechanics and its Applications, 439:103 -- 113, 2015.
39
[Hig08]
Nicholas J Higham. Functions of matrices: theory and computation. SIAM, 2008.
[HLM12] Moritz Hardt, Katrina Ligett, and Frank McSherry. A simple and practical algorithm
for differentially private data release. In Advances in Neural Information Processing
Systems 25 (NIPS), pages 2348 -- 2356. 2012.
[HMAS17] Insu Han, Dmitry Malioutov, Haim Avron, and Jinwoo Shin. Approximating the spec-
tral sums of large-scale matrices using Chebyshev approximations. SIAM Journal on
Scientific Computing, 39(4), 2017.
[HNO08] Nicholas JA Harvey, Jelani Nelson, and Krzysztof Onak. Sketching and streaming
entropy via approximation theory. In Proceedings of the 49th Annual IEEE Symposium
on Foundations of Computer Science (FOCS), pages 489 -- 498, 2008.
[Hut90]
[Isa08]
[JNS13]
Michael F Hutchinson. A stochastic estimator of the trace of the influence matrix for
Laplacian smoothing splines. Communications in Statistics-Simulation and Computa-
tion, 19(2):433 -- 450, 1990.
Garth Isaak.
http://www.lehigh.edu/~gi02/m242/08linstras.pdf, 2008.
Fast matrix multiplication and inversion, math 242 notes.
Prateek Jain, Praneeth Netrapalli, and Sujay Sanghavi. Low-rank matrix completion
using alternating minimization. In Proceedings of the 45th Annual ACM Symposium
on Theory of Computing (STOC), pages 665 -- 674, 2013.
[Jol02]
Ian Jolliffe. Principal component analysis. Wiley Online Library, 2002.
[JZ13]
[KKS65]
[KM09]
[KO17]
[KS70]
[KS16]
Rie Johnson and Tong Zhang. Accelerating stochastic gradient descent using predictive
variance reduction. In Advances in Neural Information Processing Systems 26 (NIPS),
pages 315 -- 323, 2013.
V. V. Klyuev and N. I. Kokovkin-Shcherbak. Minimization of the number of arithmetic
operations in the solution of linear algebra systems of equations. 5(1):25-43, 1965.
translated by H. F. Cleaves.
Jonathan A. Kelner and Aleksander Madry. Faster generation of random spanning
trees. In Proceedings of the 50th Annual IEEE Symposium on Foundations of Computer
Science (FOCS), pages 13 -- 21, 2009.
Ashish Khetan and Sewoong Oh.
arXiv:1703.06327, 2017.
Spectrum estimation from a few entries.
N. I. Kokovkin-Shcherbak. Minimization of numerical algorithms for solving arbitrary
systems of linear equations. 22(4):494-502, 1970.
Rasmus Kyng and Sushant Sachdeva. Approximate gaussian elimination for Laplacians:
Fast, sparse, and simple.
In Proceedings of the 57th Annual IEEE Symposium on
Foundations of Computer Science (FOCS), 2016.
[KV16]
Weihao Kong and Gregory Valiant.
arXiv:1602.00061, 2016.
Spectrum estimation from samples.
40
[KW92]
[LG12]
J Kuczy´nski and H Wo´zniakowski. Estimating the largest eigenvalue by the power
and Lanczos algorithms with a random start. SIAM Journal on Matrix nalysis and
Applications, 13(4):1094 -- 1122, 1992.
In Pro-
Fran¸cois Le Gall. Faster algorithms for rectangular matrix multiplication.
ceedings of the 53rd Annual IEEE Symposium on Foundations of Computer Science
(FOCS), pages 514 -- 523, 2012.
[LM12]
Chao Li and Gerome Miklau. Measuring the achievable error of query sets under
differential privacy. arXiv:1202.3399, 2012.
[LMP13] Mu Li, Gary L. Miller, and Richard Peng. Iterative row sampling. In Proceedings of the
54th Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages
127 -- 136, 2013.
[LN15]
Yu Lu and Sahand N Negahban. Individualized rank aggregation using nuclear norm
regularization. In Communication, Control, and Computing (Allerton), 2015 53rd An-
nual Allerton Conference on, pages 1473 -- 1479. IEEE, 2015.
[LNW14] Yi Li, Huy L. Nguyen, and David P. Woodruff. On sketching matrix norms and the
top singular vector. In Proceedings of the 46th Annual ACM Symposium on Theory of
Computing (STOC), pages 1562 -- 1581, 2014.
[LS14]
Yin Tat Lee and Aaron Sidford. Path finding methods for linear programming: Solving
linear programs in o(vrank) iterations and faster algorithms for maximum flow.
In
Proceedings of the 55th Annual IEEE Symposium on Foundations of Computer Science
(FOCS), pages 424 -- 433, 2014.
[LSW15] Yin Tat Lee, Aaron Sidford, and Sam Chiu-wai Wong. A faster cutting plane method
and its implications for combinatorial and convex optimization. In Proceedings of the
56th Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages
1049 -- 1065, 2015.
[LSY16]
Lin Lin, Yousef Saad, and Chao Yang. Approximating spectral densities of large ma-
trices. SIAM Review, 58(1):34 -- 65, 2016.
[LW16a]
[LW16b]
Yi Li and David P. Woodruff. On approximating functions of the singular values in a
stream. In Proceedings of the 48th Annual ACM Symposium on Theory of Computing
(STOC), 2016.
Yi Li and David P. Woodruff. Tight bounds for sketching the operator norm, Schatten
norms, and subspace embeddings. In Proceedings of the 20th International Workshop
on Randomization and Computation (RANDOM), 2016.
[LW17]
Yi Li and David P. Woodruff. Embeddings of Schatten norms with applications to data
streams. arXiv:1702.05626, 2017.
[LYCG14] Lei Luo, Jian Yang, Jinhui Chen, and Yicheng Gao. Schatten p-norm based matrix re-
gression model for image classification. In Pattern Recognition, pages 140 -- 150. Springer,
2014.
41
[MM15]
Cameron Musco and Christopher Musco. Randomized block krylov methods for
stronger and faster approximate singular value decomposition. In Advances in Neu-
ral Information Processing Systems 28 (NIPS), pages 1396 -- 1404, 2015.
[Mor73]
Jacques Morgenstern. Note on a lower bound on the linear complexity of the fast
Fourier transform. Journal of the ACM (JACM), 20(2):305 -- 306, 1973.
[MST15] Aleksander Madry, Damian Straszak, and Jakub Tarnawski. Fast generation of random
spanning trees and the effective resistance metric. In Proceedings of the 26th Annual
ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 2019 -- 2036, 2015.
[Nes13]
Yurii Nesterov. Introductory lectures on convex optimization: A basic course, volume 87.
Springer Science & Business Media, 2013.
[NHD12]
Feiping Nie, Heng Huang, and Chris Ding. Low-rank matrix recovery via efficient Schat-
ten p-norm minimization. In Twenty-Sixth AAAI Conference on Artificial Intelligence,
2012.
[NNS+14] Praneeth Netrapalli, UN Niranjan, Sujay Sanghavi, Animashree Anandkumar, and
Prateek Jain. Non-convex robust PCA. In Advances in Neural Information Processing
Systems 27 (NIPS), pages 1107 -- 1115, 2014.
[NWC+12] Feiping Nie, Hua Wang, Xiao Cai, Heng Huang, and Chibiao Ding. Robust matrix com-
pletion via joint Schatten p-norm and lp-norm minimization. In Data Mining (ICDM),
2012 IEEE 12th International Conference on, pages 566 -- 574. IEEE, 2012.
[Par98]
Beresford N Parlett. The symmetric eigenvalue problem. SIAM, 1998.
[Ras04]
[Raz03]
[RKA15]
Carl Edward Rasmussen. Gaussian processes in machine learning. In Advanced lectures
on machine learning, pages 63 -- 71. Springer, 2004.
Ran Raz. On the complexity of matrix product. SIAM Journal on Computing,
32(5):1356 -- 1369, 2003.
Farbod Roosta-Khorasani and Uri Ascher. Improved bounds on sample size for implicit
matrix trace estimators. Foundations of Computational Mathematics, 15(5):1187 -- 1212,
2015.
[RS03]
Ran Raz and Amir Shpilka. Lower bounds for matrix product in bounded depth circuits
with arbitrary gates. SIAM Journal on Computing, 32(2):488 -- 513, 2003.
[Saa03]
Yousef Saad. Iterative methods for sparse linear systems. SIAM, 2003.
[Saa11]
[Shp03]
[SLO13]
Yousef Saad. Numerical Methods for Large Eigenvalue Problems: Revised Edition,
volume 66. SIAM, 2011.
Amir Shpilka. Lower bounds for matrix product. SIAM Journal on Computing,
32(5):1185 -- 1200, 2003.
Andreas Stathopoulos, Jesse Laeuchli, and Kostas Orginos. Hierarchical probing for es-
timating the trace of the matrix inverse on toroidal lattices. SIAM Journal on Scientific
Computing, 35(5):S299 -- S322, 2013.
42
[SS08]
[SSZ14]
[ST04]
Daniel A. Spielman and Nikhil Srivastava. Graph sparsification by effective resistances.
In Proceedings of the 40th Annual ACM Symposium on Theory of Computing (STOC),
pages 563 -- 568, 2008.
Shai Shalev-Shwartz and Tong Zhang. Accelerated proximal stochastic dual coordi-
nate ascent for regularized loss minimization. In Proceedings of the 31st International
Conference on Machine Learning (ICML), pages 64 -- 72, 2014.
Daniel A Spielman and Shang-Hua Teng. Nearly-linear time algorithms for graph
partitioning, graph sparsification, and solving linear systems.
In Proceedings of the
36th Annual ACM Symposium on Theory of Computing (STOC), pages 81 -- 90, 2004.
[TB97]
Lloyd N. Trefethen and David Bau. Numerical Linear Algebra. SIAM, 1997.
[UCS17]
Shashanka Ubaru, Jie Chen, and Yousef Saad. Fast estimation of tr(f(a)) via stochastic
Lanczos quadrature. 2017. manuscript.
[US16]
Shashanka Ubaru and Yousef Saad. Fast methods for estimating the numerical rank
of large matrices.
In Proceedings of The 33rd International Conference on Machine
Learning, pages 468 -- 477, 2016.
[USS17]
Shashanka Ubaru, Yousef Saad, and Abd-Krim Seghouane. Fast estimation of approx-
imate matrix ranks using spectral densities. Neural Computation, 2017.
[VGLH06] Roy Varshavsky, Assaf Gottlieb, Michal Linial, and David Horn. Novel unsupervised
feature filtering of biological data. Bioinformatics, 22(14):e507 -- e513, 2006.
[Wil12]
[Wil15]
Virginia Vassilevska Williams. Multiplying matrices faster than Coppersmith-
Winograd. In Proceedings of the 44th Annual ACM Symposium on Theory of Computing
(STOC), pages 887 -- 898, 2012.
Virginia Vassilevska Williams. Hardness of easy problems: Basing hardness on popular
conjectures such as the strong exponential time hypothesis (invited talk). In 10th Inter-
national Symposium on Parameterized and Exact Computation, IPEC 2015, September
16-18, 2015, Patras, Greece, pages 17 -- 29, 2015.
[Win70]
Shmuel Winograd. On the number of multiplications necessary to computer certain
functions. 23:165-179, 1970.
[Win87]
Shmuel Winograd. Arithmetic complexity of computations. 1987.
[WLK+16] Lingfei Wu, Jesse Laeuchli, Vassilis Kalantzis, Andreas Stathopoulos, and Efstratios
Gallopoulos. Estimating the trace of the matrix inverse by interpolating from the
diagonal of an approximate inverse. Journal of Computational Physics, 326:828 -- 844,
2016.
[Woo14]
David P. Woodruff. Sketching as a tool for numerical linear algebra. Foundations and
Trends in Theoretical Computer Science, 10(1-2):1 -- 157, 2014.
[WW10]
Virginia Vassilevska Williams and Ryan Williams. Subcubic equivalences between path,
matrix and triangle problems. In Proceedings of the 51st Annual IEEE Symposium on
Foundations of Computer Science (FOCS), pages 645 -- 654, 2010.
43
[WWZ14] Karl Wimmer, Yi Wu, and Peng Zhang. Optimal query complexity for estimating the
trace of a matrix. In Proceedings of the 41st International Colloquium on Automata,
Languages and Programming (ICALP), pages 1051 -- 1062, 2014.
[XGL+16] Yuan Xie, Shuhang Gu, Yan Liu, Wangmeng Zuo, Wensheng Zhang, and Lei Zhang.
Weighted Schatten p-norm minimization for image denoising and background subtrac-
tion. IEEE transactions on image processing, 25(10):4842 -- 4857, 2016.
[XQT+16] Y. Xie, Y. Qu, D. Tao, W. Wu, Q. Yuan, and W. Zhang. Hyperspectral image restora-
tion via iteratively regularized weighted Schatten p-norm minimization. IEEE Trans-
actions on Geoscience and Remote Sensing, PP(99):1 -- 18, 2016.
[ZWJ15] Yuchen Zhang, Martin J Wainwright, and Michael I Jordan. Distributed estimation of
generalized matrix rank: Efficient algorithms and lower bounds. In Proceedings of the
32nd International Conference on Machine Learning (ICML), 2015.
44
A Linear System Solvers
In this section we give runtimes for solving ridge regression using both traditional iterative methods
and stochastic gradient descent equipped with deflation-based preconditioning. We start with a few
standard notions from convex optimization, which are necessary for our stochastic method bounds.
Definition 36 (Strong convexity). A function f : Rd → R is µ-strongly convex if, for all x, y ∈ Rd,
f (x) − f (y) ≥ ∇ f (y)T (x − y) +
µ
2 kx − yk2
2 .
Equivalently, if ∇2 f (cid:23) µI.
Definition 37 (Smoothness). A function f : Rd → R is β-smooth if for all x, y ∈ Rd,
f (x) − f (y) ≤ ∇ f (y)T (x − y) +
β
2 kx − yk2
2 .
Equivalently, if ∇2 f (cid:22) βI.
The rate of convergence that iterative methods achieve on f typically depends on the ratio β/µ,
which corresponds to the condition number of a linear system. We next show how ridge regression
can be recast as minimizing a convex function f , and show that our error from the true ridge
solution is proportional to error in minimizing f .
Fact 38 (Ridge Regression as Convex Optimization). For any A ∈ Rn×d, b ∈ Rd, and λ > 0, let
Mλ
λ b. x∗ is the minimizer of the convex function:
= AT A + λI and x∗ def
def
= M−1
f (x) =
1
2
xT (AT A + λI)x − bT x,
(26)
which has gradient ∇ f (x) = (AT A + λI)x − b. f is λ-strongly convex and (σ1(A)2 + λ)-smooth.
Proof. To check that x∗ is the minimizer, notice that:
∇ f (x∗) = (AT A + λI)(AT A + λI)−1b − b = 0.
Since the function is quadratic, x∗ is the unique minimizer. ∇2 f = Mλ, so by Definitions 36 and
37, f is λ-strongly convex and (σ1(A)2 + λ)-smooth.
Fact 39 (Function Error Equals Norm Error). For any x ∈ Rd, letting Mλ, x∗ and f be defined
as in Fact 38,
kx − x∗k2
Mλ
= 2[f (x) − f (x∗)].
(27)
Proof. Since x∗ = M−1
λ b:
kx − x∗k2
Mλ
def= (x − x∗)T Mλ(x − x∗)
= xT Mλx − 2xT Mλx∗ + x∗T Mλx∗
= xT Mλx − 2xT b + bT M−1
λ b
= 2 [f (x) − f (x∗)] .
We will focus on making multiplicative progress in [f (x) − f (x∗)] which will lead to multiplica-
M and a close approximation in log(1/ǫ) iterations.
tive progress in kx − x∗k2
45
A.1 Unaccelerated SVRG
We first prove an unaccelerated and unpreconditioned runtime for the Stochastic Variance Reduced
Gradient (SVRG) algorithm, introduced in [JZ13].
Theorem 40 (Standard SVRG Runtime). For any A ∈ Rn×d, b ∈ Rd, and λ > 0, let Mλ
AT A + λI and x∗ def= M−1
in
def=
λ y. There is an algorithm that returns x with: Ekx − x∗kMλ ≤ ǫkx∗kMλ
O "nnz(A) +
ds(A)kAk2
F
λ
# · log(1/ǫ)!
time.
SVRG proceeds in epochs.
In each epoch we will make one full gradient computation --
∇ f (x0) = Mλx0 − b where x0 is the value of our iterate at the beginning of the epoch. We
will then make a number of stochastic gradient updates each 'variance reduced' using ∇ f (x0)
and show that we make constant factor progress on our function value in expectation. Stringing
together log(1/ǫ) epochs yields Theorem 40. We write our function f (x) given by (26) as a sum:
1
2
=
1
2
i +
f (x) =
I! x −
xT (AT A + λI)x − bT x
xT aiaT
nXi=1
λkaik2
kAk2
where ai ∈ Rd is the ith row of A and aiaT
i ∈ Rd×d is the rank-1 matrix which is its contribution
to AT A. We start with a well known lemma showing that it is possible to make constant progress
on the value of f (x) in each epoch of SVRG:
Lemma 41 (SVRG Epoch). Consider a set of convex functions {ψ1, ψ2, ..., ψn} mapping Rd → R.
Let f (x) =Pn
i=1 ψi(x) and x∗ = arg minx∈Rd f (x). Suppose we have a probability distribution p on
[1, 2, ..., n] and that starting from some initial point x0 ∈ Rd in each iteration we select i ∈ [1, ..., n]
with probability pi and set:
bT x.
(28)
1
n
2
F
xk+1 := xk −
η
pi
(∇ ψi(xk) − ∇ ψi(x0)) + η ∇ f (x0)
for some step size η. If f is µ-strongly convex and if for all x ∈ Rd we have
pi k∇ ψi(x) − ∇ ψi(x∗)k2
2 ≤ 2 ¯S [f (x) − f (x∗)]
(29)
(30)
1
nXi=1
f (xk) − f (x∗)# ≤
for some variance parameter ¯S then for all m ≥ 1 we have:
1 − 2η ¯S(cid:20) 1
E" 1
µηm
m
1
mXk=1
+ 2η ¯S(cid:21) · [f (x0) − f (x∗)]
Consequently, if we pick η to be a small multiple of 1/ ¯S, then for m = O( ¯S/µ) we will decrease the
error by a constant multiplicative factor in expectation.
Proof. See for example Theorem 9 of [GHJ+16].
46
n bT x and pi = kaik2
kAk2
F
2
I(cid:17) x − 1
2
F
i + λkaik2
kAk2
2 xT(cid:16)aiaT
2 ≤ 2 ¯S [f (x) − f (x∗)]
n b. Write x − x∗ = y for simplicity of notation.
2
aiaT + λkaik2
F !2
kAk2
nXi=1 kyk2
2 λ2 kaik2
kAk4
I y
!#
i y
+
F
2
To apply the above Lemma we need the variance bound:
Lemma 42 (SVRG Variance Bound). If ψi(x) = 1
then letting ¯S = kAk2
F + 2λ we have:
1
nXi=1
2
2
F
F
1
=
i +
nXi=1
i + λkaik2
kAk2
λkaik2
kAk2
pi k∇ ψi(x) − ∇ ψi(x∗)k2
Proof. We have ∇ ψi(x) =(cid:16)aiaT
I(cid:17) x − 1
pi k∇ ψi(x) − ∇ ψi(x∗)k2
2 ·(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
I! y(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
aiaT
nXi=1
kAk2
kaik2
2 · yT(aiaT
1
nXi=1
λkaik2
= kAk2
kAk2
2 ! + 2
F ·" nXi=1 yT (aiaT
nXi=1
λyT aiaT
kaik2
kAk2
F
F #
F ·"yT AT Ay +
+ kyk2
2 λ2
kAk2
kAk2
F"kyk2
F #
2λkyk2
kAk2
F + 2λ)[f (x) − f (x∗)]
≤ kAk2
≤ 2(kAk2
= kAk2
= kAk2
2λyT AAT y
kaik2
i )2 + 2
i )2y
Mλ
Mλ
+
2
2
2
F
2
F
F
F
where the last step uses Fact 39.
We can now plug this variance bound into Lemma 41 to prove Theorem 40
Proof of Theorem 40. Using Lemma 42 we can instantiate Lemma 41 with ¯S = kAk2
set
F + 2λ. If we
m = O(cid:18) ¯S
µ(cid:19) = O kAk2
F + λ
λ
! = O kAk2
λ !
F
and η = O(cid:16)
1
kAk2
F +λ(cid:17) (which we can compute explicitly) then after an m step SVRG epoch:
E" 1
m
mXk=1
f (xk) − f (x∗)# ≤
1
2
[f (x0) − f (x∗)] .
If we choose k uniformly from [1, ..., m] this gives us E [f (xk) − f (x∗)] ≤ 1
2 [f (x0) − f (x∗)]. So
stringing together log(1/ǫ) of these epochs and letting x0 = 0 in the first epoch gives the theorem.
Each epoch requires m stochastic gradient steps given by (29), taking O(mds) = O(cid:16)kdsAk2
(cid:17) time
plus nnz(A) time to compute ∇ f (x0), giving us the stated runtime.
λ
F
47
Note that, naively, to produce xk+1 we need O(d) time, not O(ds) time since our gradient step
includes adding multiples of b and η ∇ f (x0) both of which might be dense vectors in Rd. However,
in each epoch, we will just keep track of the coefficients of these two components in our iterate,
allowing us to still compute (29) in O(ds) time.
A.2 Unaccelerated SVRG with Deflation-Based Preconditioning
i (A)
1
F
λ
i (A)
i=1 σ2
nλ
i=1 σ2
nλ
If we assume that ds = O(cid:16) nnz(A)
n (cid:17) (i.e. our rows are uniformly sparse) we see that the runtime of
. ¯κ def= Pd
= nnz(A) · Pd
Theorem 40 is dominated by ds(A)kAk2
can be seen as an
= σ2
def
average condition number, which is always smaller than the condition number κ
λ . This average
condition number dependence means that SVRG can significantly outperform traditional iterative
solvers that require a number of iterations depending on κ. However, this advantage is limited if
kAk2
F is very concentrated in a few large singular values, and hence ¯κ ≈ κ. We can perform better
in such situations by deflating off these large singular values and preconditioning with our deflated
matrix, significantly 'flattening' the spectrum of A. This method was used in [GOSS16], and we
follow their approach closely, giving our own proof for completeness and so that we can express
runtimes in terms of the necessary parameters for our downstream results. In particular, we show
that the preconditioning methods of [GOSS16] can be implemented efficiently for sparse systems.
Theorem 43 (Preconditioned SVRG Runtime). For any A ∈ Rn×d, b ∈ Rd, and λ > 0, let
where k ∈ [0, ..., d] is an input parameter. There
Mλ
is an algorithm that uses O(nnz(A)k log d + dkω−1) precomputation time for sparse matrices or
O(ndω(1,1,logd k)−1 log d) for dense matrices after which, given any input y ∈ Rd, letting x∗ def= M−1
λ y
the algorithm returns x with Ekx − x∗kMλ ≤ ǫkx∗kMλ
def= AT A + λI. Let ¯κ def=
i=k+1 σ2
dλ
k(A)+Pd
kσ2
i (A)
in
O (nnz(A) log(1/ǫ) + log(1/ǫ) [d · ds(A) + dk] ¯κ)
time for sparse A or
time for dense A.
O(cid:0)log(1/ǫ)(nd + d2⌈¯κ⌉)(cid:1)
Our ideal algorithm would compute the top k singular vectors of A and deflate these off our
matrix to flatten the spectrum. However, for efficiency we will instead compute approximate
singular vectors, using an iterative method, like simultaneous iteration or a block Krylov iteration.
These algorithms give the following guarantee [MM15]:
Lemma 44. There is an algorithm that, with high probability in O(nnz(A)k log d + dkω−1) time
i AT Azi,
k+1(A). For dense inputs the runtime can be sped up to O(ndω(1,1,logd k)−1 log d)
2 ≤ 2σk+1(A)2 and, for all i ≤ k, letting σ2
returns Z ∈ Rd×k such that(cid:13)(cid:13)A − ZZT A(cid:13)(cid:13)2
(cid:12)(cid:12)σi − σ2
i (A)(cid:12)(cid:12) ≤ 2σ2
by applying fast matrix multiplication at each iteration.4
def= zT
[GOSS16] shows that we can build a good preconditioner from such a Z. Specifically:
i
4Each iteration requires ndω(1,1,logd k)−1 ≥ dω(1,1,logd k) time since we assume n ≥ d, which dominates the O(dkω−1)
time required to orthogonalize the approximate singular directions.
48
Lemma 45 (Theorem 5 of [GOSS16]). For any A ∈ Rn×d, b ∈ Rd, and λ > 0, given Z ∈ Rd×k
satisfying the guarantees of Lemma 44, let
P−1/2 = Z Σ−1/2ZT +
Then we have:
tr(P−1/2(AT A + λI)P−1/2)
λd(P−1/2(AT A + λI)P−1/2)
(I − ZZT )
qσ2
k + λ
= O kσ2
where Σ−1/2
i,i =
.
1
qσ2
i + λ
k(A) +Pd
i=k+1 σ2
λ
i (A)
+ d! .
Intuitively, after applying the preconditioner, all top singular values are capped at σ2
k(A), giving
a much flatter spectrum and better performance when optimizing with SVRG. To make use of the
above bound, we first define a preconditioned ridge regression problem.
Definition 46 (Preconditioned Ridge Regression). For any A ∈ Rn×d, b ∈ Rd, and λ > 0,
λ b. Letting P−1/2 be as described in Lemma 45, let Mλ =
let Mλ
P−1/2MλP−1/2 and define the preconditioned regression objective function by:
= AT A + λI and x∗ def
def
= M−1
f (x) =
1
2
xT Mλx − bT P−1/2x.
λ P−1/2b = P1/2(AT A + λI)−1b.
f is minimized at x∗ = M−1
Fact 47 (Preconditioned Solution). For any x ∈ Rd, if kx − x∗k Mλ ≤ ǫkx∗k Mλ
ǫkx∗kMλ
Proof. We have x∗ = P1/2x∗ and so using the fact that kx − x∗k Mλ ≤ ǫkx∗k Mλ
where x∗ is the minimizer of f (x).
then(cid:13)(cid:13)P−1/2x − x∗(cid:13)(cid:13)Mλ ≤
can write:
(cid:13)(cid:13)(cid:13)x − P1/2x∗(cid:13)(cid:13)(cid:13) Mλ ≤ ǫ(cid:13)(cid:13)(cid:13)P1/2x∗(cid:13)(cid:13)(cid:13) Mλ
(cid:13)(cid:13)(cid:13)P−1/2x − x∗(cid:13)(cid:13)(cid:13)Mλ ≤ ǫkx∗kMλ
.
By Fact 47, we can find a near optimal solution to our preconditioned system and by multiplying
by P−1/2 obtain a near optimal solution to the original ridge regression problem. We now show
that our preconditioned problem can be solved quickly. We first give a variance bound as we did
in Lemma 42 in the non-preconditioned case.
Lemma 48 (Preconditioned Variance Bound). Let BT = [P−1/2AT ,√λP−1/2] so that BT B =
P−1/2(AT A + λI)P−1/2 = Mλ. For i ∈ 1, ..., (n + d), let
i(cid:1) x −
1
xT(cid:0)bibT
bT P−1/2x
ψi(x) =
n + d
1
2
2
then letting ¯S = tr(P−1/2(AT A + λI)P−1/2:
so that Pn+d
i=1
ψi(x) = f (x). Set pi = kbik2
kBk2
F
nXi=1
1
pi(cid:13)(cid:13)(cid:13)∇ ψi(x) − ∇ ψi(x∗)(cid:13)(cid:13)(cid:13)
49
2
2 ≤ 2 ¯S [f (x) − f (x∗)] .
Proof. Write x − x∗ = y for simplicity of notation. Then we have:
n+dXi=1
1
pi(cid:13)(cid:13)(cid:13)∇ ψi(x) − ∇ ψi(x∗)(cid:13)(cid:13)(cid:13)
2
2
= kBk2
F
n+dXi=1 (cid:13)(cid:13)bibT
i y(cid:13)(cid:13)2
kbik2
2
2
i y
F
yT bi(bT
i bi)bT
kbik2
n+dXi=1
= kBk2
= kBk2
F · yT BT By
F · [ f (x) − f (x∗)]
= 2kBk2
2
where the last step uses Fact 39. Finally we write:
¯S = kBk2
F = tr(BT B) = tr(P−1/2(AT A + λI)P−1/2).
We can now combine this variance bound with Lemmas 41 and 45 to prove our preconditioned
SVRG runtime.
Proof of Theorem 43. Using Lemma 48 we can instantiate Lemma 41 with ¯S = kBk2
λI)P−1/2). Letting µ = λd(P−1/2(AT A + λI)P−1/2) be a lower bound on the strong convexity of
f , by Lemma 45 we have:
F = tr(P−1/2(AT A+
i (A)
m = O(cid:18) ¯S
i=k+1 σ2
λ
λd(P−1/2(AT A + λI)P−1/2)!
µ(cid:19) = O tr(P−1/2(AT A + λI)P−1/2)
+ d! .
= O kσ2
k(A) +Pd
¯S(cid:1) = O(cid:16) 1
F(cid:17) which we will compute explicitly, then after an m
2 xT(cid:0)bibT
and hence, E(cid:13)(cid:13)P−1/2x − x∗(cid:13)(cid:13)Mλ ≤ ǫkx∗kMλ
n+d bT P−1/2x each selected
, we make constant progress in expectation on f (x). After log(1/ǫ)
i(cid:1) x − 1
step SVRG epoch running on the functions ψi(x) = 1
with probability pi = kbik2
kBk2
iterations we have x with Ekx − x∗k Mλ ≤ ǫkx∗k Mλ
by Fact 47. It remains to bound the cost of each SVRG epoch applied to f (x).
If we set the step size η = O(cid:0) 1
For dense inputs, the argument is relatively simple. We can compute the approximate top
singular vector space Z ∈ Rd×k in O(ndω(1,1,logd k)−1 log d) time by Lemma 44. Using the factored
structure of P−1/2, we can explicitly form AP−1/2 and P−1/2 also in (O(ndω(1,1,logd k)−1) time. This
allows us to compute a full gradient in O(nd) time (which we do once per epoch) and perform each
stochastic gradient step in O(d) time (which we do m times per epoch). This gives final runtime
kBk2
F
2
O ((nd + md) log(1/ǫ)) = O(cid:0)log(1/ǫ)(nd + d2⌈¯κ⌉)(cid:1)
i (A)
where ¯κ
def
=
kσ2
k(A)+Pd
i=k+1 σ2
dλ
.
For sparse A we have to be much more careful to fully exploit sparsity when preconditioning.
Computing Z takes time O(nnz(A)k log d+dkω−1) by Lemma 44. We will not explicitly form P−1/2
but will show how to apply it when needed. Recall that P−1/2 = Z Σ−1/2ZT + δ(I − ZZT ) where
we denote δ def= 1/(σ2
i + λ). First, in order to determine our step size η we must compute kBk2
F .
kBk2
2
2
F
F =(cid:13)(cid:13)(cid:13)AP−1/2(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)AZ Σ−1/2ZT(cid:13)(cid:13)(cid:13)
+ λ(cid:13)(cid:13)(cid:13)P−1/2(cid:13)(cid:13)(cid:13)
+ δ(cid:13)(cid:13)A(I − ZZT )(cid:13)(cid:13)2
F
F
2
50
F + λ(cid:13)(cid:13)(cid:13)Z Σ−1/2ZT(cid:13)(cid:13)(cid:13)
2
F
+ λδ(cid:13)(cid:13)I − ZZT(cid:13)(cid:13)2
F .
which can be computed in O(nnz(A)k) time with just a single multiplication of A by Z.
Applying Pythagorean theorem and submultiplicativity we have:
kBk2
We can similarly compute our sampling probabilities pi = kbik2
B2
F
2
2
2
F
+ δ kAk2
F − δ kAZk2
F =(cid:13)(cid:13)(cid:13)AZ Σ−1/2(cid:13)(cid:13)(cid:13)
i Z Σ−1/2(cid:13)(cid:13)(cid:13)
2 = (cid:13)(cid:13)(cid:13)aT
2)(cid:19). This again requires just O(nnz(A)k) time.
F + λ(cid:13)(cid:13)(cid:13)Z Σ−1/2(cid:13)(cid:13)(cid:13)
2 − δ(cid:13)(cid:13)aT
i Z(cid:13)(cid:13)2
+ δ kaik2
+ δ(1 − kzik2
F
2
2
In each SVRG epoch we must compute one full gradient of the form:
i = 1, ..., n, kbik2
λ(cid:18)(cid:13)(cid:13)(cid:13)zT
i
2
2
Σ−1/2(cid:13)(cid:13)(cid:13)
+ λδ(d − k)
quickly by noting that for
2 and for i = n + 1, ..., n + d, kbik2
2 =
∇ f (x0) = BT Bx0 − P−1/2b = P−1/2AT AP−1/2x0 + λP−1x0 − P−1/2b
which takes O(nnz(A) + dk) time as P−1/2 can be applied to a vector in O(dk) time. Naively, we
then must make m stochastic gradient steps, each requiring O(dk) time. However we can do better.
Instead of storing our iterate xk explicitly, we will store it as the sum:
k (cid:17) + Zx
(1)
(0)
xk =(cid:16)x
k − Zx
k ∈ Rk and x(3)
0 = ZT x0 and x(3)
(2)
(2)
k + x(3)
k
· ∇ f (x0).
k
(0)
(1)
k ∈ Rd, x
k , x
0 = ZT x0, x(2)
Note that x
x(0)
0 = x0, x(1)
is a scalar. At the beginning of each epoch we will set
0 = 0. We can compute ZT x0 in O(dk) time. We will
(2)
k will give
k − Zx
the component of xk within this span.
maintain the invariant that (cid:16)x
For ease of notation let an+i denote √λei where ei ∈ Rd the ith standard basis vector. In this
way we have bi = P−1/2ai for all i ∈ 1, ..., n + d. Each stochastic gradient step is of the form:
k (cid:17) is perpendicular to the span of Z while x
(1)
(0)
xk+1 = xk −
= xk +
= xk +
η
P−1/2(aiaT
pi(cid:16)∇ ψi(xk) − ∇ ψi(x0)(cid:17) + η ∇ f (x0)
(P−1/2ai)(cid:16)(aT
i P−1/2)xk − aT
i )P−1/2(xk − x0) + η ∇ f (x0)
η
pi
η
pi
i (P−1/2x0)(cid:17) + η ∇ f (x0).
We can precompute P−1/2x0 in O(dk) time per epoch and then can compute the dot product
(1)
k (cid:17) is always perpendic-
aT
ular to the span of Z, we can write the dot product (aT
i (P−1/2x0) in nnz(ai) = O(ds(A)) time. Using the fact that(cid:16)x
i (cid:16)Z Σ−1/2ZT + δ(I − ZZT )(cid:17)(cid:16)(cid:16)x
k (cid:17) + Zx
k − Zx
k + x(3)
aT
(1)
(0)
(1)
(2)
(0)
k ∇ f (x0)(cid:17)
i P−1/2)xk as:
(0)
k − Zx
i (cid:16)x
k (cid:17) + aT
(2)
k aT
= δaT
k + x(3)
k − Zx
i Z Σ−1/2x
i Z Σ−1/2ZT ∇ f (x0) + δx(3)
i (I − ZZT )∇ f (x0).
We can precompute AZ in O(nnz(A)k) time. We can also precompute Z Σ−1/2ZT ∇ f (x0) and
(I − ZZT )∇ f (x0) in O(dk) time. With these values in hand, computing the above dot product
takes just O(nnz(ai) + k) = O(ds(A) + k) time.
Now, we can write P−1/2ai = Z( ΣZT ai) − δ(I − ZZT )ai. Since we have precomputed AZ,
(2)
k with the appropriate weight in O(k) time. We can then
ΣZT ai can be computed and added to x
k aT
51
compute for the appropriate weight w:
(cid:16)x
(0)
k+1 − Zx
(1)
k+1(cid:17) =(cid:16)x
=(cid:16)x
(0)
(1)
k − Zx
k (cid:17) + w · (I − ZZT )ai
k + w · ZT ai(cid:17)
k + wai(cid:17) − Z(cid:16)x
(0)
(1)
which takes time O(ds(A) + k). Finally, we set x(3)
O(nnz(A) + dk + m(ds(A) + k)) and so our total runtime is:
k+1 = x(3)
k + η. Overall, our runtime per epoch is
O (nnz(A) log(1/ǫ) + log(1/ǫ) [d · ds(A) + dk] ¯κ)
i (A)
kσ2
k(A)+Pd
i=k+1 σ2
dλ
.
where ¯κ
def
=
A.3 Accelerated and Preconditioned SVRG
We now combine the deflation-based preconditioning described above with accelerated gradient
methods to give our strongest runtime bound for ridge regression using via stochastic solvers.
Theorem 49 (Accelerated Preconditioned SVRG Runtime). For any A ∈ Rn×d, b ∈ Rd, and λ >
0, let Mλ
where k ∈ [0, ..., d] is an input parameter. There
is an algorithm that uses O(nnz(A)k log d + dkω−1) precomputation time for sparse matrices or
O(ndω(1,1,logd k)−1 log d) for dense matrices after which, given any input y ∈ Rd, letting x∗ def
= M−1
λ y
the algorithm returns x with Ekx − x∗kMλ ≤ ǫkx∗kMλ
def= AT A+λI and ¯κdef=
i=k+1 σ2
dλ
k(A)+Pd
i (A)
kσ2
in
O(cid:16)nnz(A) log(1/ǫ) + log(1/ǫ) log(d¯κ) ·pnnz(A)[d · ds(A) + dk]¯κ(cid:17)
time for sparse A or
time for dense A.
O(cid:16)log(1/ǫ)(nd + n1/2d3/2 log(¯κ)√¯κ)(cid:17)
The above runtime will follow from applying a blackbox technique for accelerating the runtime
of convex optimization methods. While there are a number of improvements over this method
[AZ17], we at most lose logarithmic factors and gain significant simplicity in our proofs.
Lemma 50 (Theorem 1.1 of [FGKS15]). Let f (x) be µ-strongly convex and let x∗def= arg minx∈Rd f (x).
2 kx − x0k2
2. Let x∗γ,x0
For any γ > 0 and any x0 ∈ Rd, let fγ,x0(x)
= arg minx∈Rd fγ,x0(x).
Suppose that, for all x0 ∈ Rd, c > 0, γ > 2µ, we can compute xc such that
c(cid:2)fγ,x0 − fγ,x0(x∗γ,x0)(cid:3)
E(cid:2)fγ,x0(xc) − fγ,x0(x∗γ,x0)(cid:3) ≤
def
= f (x)+ γ
def
1
in time Tc. Then we can compute x1 such that E [f (x1) − f (x∗)] ≤ 1
c [f (x0) − f (x∗)] in time
O T4(cid:16) 2γ+µ
µ (cid:17)3/2pγ/µ log c! .
52
Proof of Theorem 49. We focus again on optimizing the preconditioned function f in Definition
46 as by Fact 47 a near optimal minimizer for this function yields a near optimal solution for
our original ridge regression problem. We split fγ,x0 = 1
2 as
2 xT Mλx − bT P−1/2x + γ
2 kx − x0k2
fγ,x0 =Pn+d
i=1
ψi
γ,x0(x) where
ψi
γ,x0(x) =
1
2
xT (bibT
i )x −
1
n + d
bT P−1/2x +
B is as defined in Lemma 48 with BT = [P−1/2AT ,√λP−1/2].
We have ∇ ψi
γ,x0(x) = (bibT
n+d bP−1/2 + γkbik2
kBk2
F
2
(x − x0). Letting y = x − x∗γ,x0, we can
follow a similar calculation to Lemma 42 to show:
2
γ kbik2
F(cid:13)(cid:13) kx − x0k2
2(cid:13)(cid:13)B2
2 .
i )x − 1
nXi=1
1
2
2
pi(cid:13)(cid:13)(cid:13)∇ ψi
γ,x0(x) − ∇ ψi
n+dXi=1
= kBk2
≤ 2(kBk2
F + 2γ)[f (x) − f (x∗γ,x0)].
γ,x0(x∗γ,x0 )(cid:13)(cid:13)(cid:13)
I)y(cid:13)(cid:13)(cid:13)
i + γkbik2
kBk2
kbik2
2
(cid:13)(cid:13)(cid:13)(bibT
F
F
2
2
2
F and µ = λd(P−1/2(AT A + λI)P−1/2) be an lower bound on the strong convexity
µ = r. The strong convexity of fγ,x0 is lower bounded by µγ = λd(P−1/2(AT A +
Let ¯S = kBk2
of f . Denote γ
λI)P−1/2) + γ = Θ(r · λd(P−1/2(AT A + λI)P−1/2)) which gives:
= O kσ2
tr(P−1/2(AT A + λI)P−1/2) + 2γ
λd(P−1/2(AT A + λI)P−1/2) + γ
¯S
µγ
=
k(A) +Pd
rλ
d
r!
+
i=k+1 σ2
i (A)
kσ2
def
=
i (A)
k(A)+Pd
def= r = dκ
So, by Theorem 43, for dense inputs, ignoring the O(ndω(1,1,logd k)−1 log d) precomputation cost to
compute Z, P−1/2 and AP−1/2, which we only pay once, letting κ
, we have
43. Otherwise, setting γ
µ
Tc = O(cid:16)log c ·(cid:16)nd + d2κ
O T4(cid:16) 2γ+µ
n we can solve f up to ǫ accuracy in time:
r (cid:17)(cid:17). If nd ≥ d2κ then we already have runtime O(nd log(1/ǫ)) by Theorem
µ (cid:17)3/2pγ/µ log(1/ǫ)! = O(cid:16)log(1/ǫ)n1/2d3/2 log(¯κ)√¯κ(cid:17) .
i=k+1 σ2
dλ
to compute Z and AZ. We have:
For sparse inputs, we again ignore the one time precomputation cost of O(cid:0)nnz(A)k log d + dkω−1(cid:1)
Tc = O(cid:16)log c ·(cid:16)nnz(A) + [d · ds(A) + dk]
If nnz(A) ≥ [d · ds(A) + dk] ¯κ) then we already have runtime O(nnz(A) log(1/ǫ)) by Theorem 43.
Otherwise, nnz(A) ≤ [d · ds(A) + dk] ¯κ. So setting r = ¯κ[d·ds(A)+dk]
√nnz(A)
r(cid:17)(cid:17) .
we have:
¯κ
O T4(cid:16) 2γ+µ
µ (cid:17)3/2pγ/µ log(1/ǫ)! = O(cid:16)log(1/ǫ) log(d¯κ) ·pnnz(A)[d · ds(A) + dk]¯κ(cid:17) .
53
A.4 Preconditioned Iterative Methods
Finally, we describe how to combine deflation-based preconditioning with standard iterative meth-
ods, which can give runtime advantages over Theorem 49 in some parameter regimes.
Theorem 51 (Preconditioned Iterative Method). For any A ∈ Rn×d, b ∈ Rd, and λ > 0, let
where k ∈ [0, ..., d] is an input parameter. There is an algorithm
Mλ
that uses O(nnz(A)k log d + dkω−1) precomputation time after which, given any input y ∈ Rd,
letting x∗ def
λ y the algorithm returns x such that with high probability kx − x∗kMλ ≤ ǫkx∗kMλ
in time
def= AT A + λI and κ =
= M−1
σ2
k+1(A)
λ
O(cid:16)log(1/ǫ)(nnz(A) + dk)√κ(cid:17) .
Proof. If we form the preconditioner P−1/2 as in Lemma 45, by Lemmas 2 and 4 of [GOSS16] we
have the preconditioned condition number bound:
def
=
κ
λ1(P−1/2MλP−1/2)
λd(P−1/2MλP−1/2)
= O σ2
k+1(A)
λ
!
Note that the bound in [GOSS16] is actually in terms of σ2
k+1(A) so that
our theorem holds in the case that k = 0. This can be achieved with no effect on the asymptotic
runtime by setting k to k + 1.
k(A), however we write σ2
We can now apply any accelerated linear system solver, such as Chebyshev iteration, Conjugate
Gradient, or Accelerated Gradient Descent [Saa03, Nes13] to obtain x with kx − x∗kMλ ≤ ǫkx∗kMλ
in O(cid:16)log(1/ǫ)√κ · matvec(P−1/2(AT A + λI)P−1/2(cid:17) where matvec(P−1/2(AT A + λI)P−1/2 is the
time required to multiply a single vector by this matrix. matvec(P−1/2(AT A + λI)P−1/2) =
O(nnz(A) + dk) since P−1/2 = Z Σ−1/2ZT + δ(I − ZZT ) can be applied in O(dk) time. This gives
the result combined with the precomputation time for P−1/2 from Lemma 44.
B Additional Proofs: Lower Bounds
Lemma 52 (Determinant Hardness). Given algorithm A which given B ∈ Rn×n returns X ∈
(1± ǫ) det(B) in O(nγǫ−c) time, we can detect if an n-node graph contains a triangle in O(nγ+12c)
time.
Proof. Let A ∈ Rn×n be the adjacency matrix of an n-node graph G. Let λ1, ..., λn denote its
eigenvalues. Let B = I + δA for some δ which we will set later. We can write:
det(B) =
λi(B) =
nYi=1
δk · Xi1<i2<...<ik
The k = 0 term in (31) is 1, and the next two are easy to compute. δPn
δ2Pi<j λiλj = δ2
nYi=1
2 (cid:16)Pi,j λiλj −Pi λ2
i(cid:17) = δ2
2 Pi λi tr(A) − δ2
2 kAk2
nXk=0
(1 + δλi) =
λi1λi2...λik .
i=1 λi = δ tr(A) = 0, and
F /2. For k = 3 we
F = −δ2 kAk2
(31)
54
have:
δ3 Xi<j<k
λiλjλk =
δ3
3 Xi<j
λiλj tr(A) −Xi6=j
δ3
3 kAk2
tr(A3).
F · tr(A) +
δ3
3
= 0 −
δ3
=
3
λ2
i λj
tr(A3)
1 ≤ (nδλ1)k ≤ (n2δ)k
since λ1 ≤ n. However, in order to obtain a tighter result, we will use stronger bounds for k = 4, 5.
These bounds are very tedious but straightforward. Specifically:
1
λ2
λ3
λ3
λ4
δ4
=
δ4
δ4
λ2
i λ2
δ4
=
λ2
i λ2
=
=
λiλjλk −
λiλjλkλl(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
δ4 Xi<j<k<l
λi1λi2...λik(cid:12)(cid:12)(cid:12) ≤(cid:0)n
k(cid:1)δkλk
i λjλk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2 Xi6=j6=k
i λj(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
j −Xi6=j
i(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i −Xi
We will bound the k > 3 terms by: (cid:12)(cid:12)(cid:12)δk ·Pi1<i2<...<ik
4 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
tr(A) Xi<j<k
8 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
tr(A)Xi6=j
i λj −Xi6=j
8 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
j + tr(A)Xi6=j
Xi6=j
8 (cid:12)(cid:12)(cid:12)kAk2
F − 2 tr(A4)(cid:12)(cid:12)(cid:12) ≤
30(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i λjλkλl(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
λiλjλkλlλm(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xi6=j6=k6=l
30(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i λjλk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
j λk + Xi6=j6=k
2 Xi6=j6=k
30(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i λj(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
5Xi6=j
j +Xi6=j
30(cid:12)(cid:12)(cid:12)5(X λ2
i(cid:12)(cid:12)(cid:12)
i ) − 6X λ5
i )(X λ3
λ1X λ4
δ5 Xi<j<k<l<m
tr(A3) +
λ2
i λ2
δ5n2
6
δ5n2
δ5
5
δ5n5
i λ3
λ2
δ4n4
.
4
δ5
δ5
δ5
=
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
δ5
=
λ3
λ4
i
tr(A3) +
.
5
λ2
λ2
And similarly:
=
=
≤
≤
6
55
Finally, if we set δ = 1
10n4 then we have:
δk · Xi1<i2<...<ik
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
nXk=4
λi1λi2...λik(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
δ4n4
4
≤
+
δ5n5
5
+
δ5n2
6
≤ δ3(cid:18) 1
40
δ3
30
≤
+
1
500
+
1
600
+
δ3
600
tr(A3).
(n2δ)k
tr(A3) +
∞Xk=6
tr(A3) +(cid:18) 1
103 +
1
105 + ...(cid:19)(cid:19)
We then write:
det(B) = 1 −
δ2 kAk2
F
2
+
δ3 tr(A3)
3
δ3
30 ±
δ3
600
±
tr(A3).
Since 1 ≤ δ3 · 103n12 and δ2kAk2
≤ δ3 · 5n6 if we compute X ∈ (1 ± c1/n12) det(B) for sufficiently
small constant c1 and subtract off (cid:16)1 − δ2kAk2
(cid:17), we will be able to determine if tr(A3) > 0 and
hence detect if G has a triangle. So any algorithm approximating det(B) to (1±ǫ) error in O(nγǫ−c)
time yields a triangle detection algorithm running in O(nγ+12c) time.
2
2
F
F
C Krylov and Power Methods preserve Invariant Norms
In this section, we show that when the Krylov method [MM15] or the power method is used to
deflate the top singular vectors of the matrix, any unitarily invariant norm of the tail (remaining
part of the spectrum) is preserved. Let P = ZZ⊤ be the projector obtained for the top k singular
vectors of A using the Krylov method [MM15] or the power method. Then, for any invariant norm
of the tail to be preserved, we just need to show the following:
σi((I − P)A) ≤ (1 + ǫ)σi(A − Ak) +
ǫ
n
σ1(A).
(32)
Since P is a (1 + ǫ) approximation obtained in terms of the spectral norm , we have by the spectral
low rank approximation guarantee of [MM15] that σ1((I− P)A) ≤ (1 + ǫ)σ1(A− Ak), where Ak is
the best rank k approximation of A. Consider first any i for which σi(A−Ak) ≥ (1−ǫ)σ1(A−Ak) =
(1 − ǫ)σk+1(A). Then we have,
σi((I − P)A) ≤ σ1((I − P)A) ≤ (1 + ǫ)σk+1(A) ≤
(1 + ǫ)
(1 − ǫ)
σi(A − Ak) ≤ (1 + 3ǫ)σi(A − Ak).
For any i with σi(A − Ak) < (1 − ǫ)σk+1(A), i.e., with σi+k(A) ≤ (1 − ǫ)σk+1(A) there is a large
(> ǫ) relative gap between this singular value and σk. We have by the min-max characterization
of singular values:
σi((I − P)A) =
Yrank(Y)=n−i+1(cid:18)
min
If we just set Y = Z, then we have:
σn−k+1((I − P)A) ≤
max
y∈span(Y)kyk2=1(cid:13)(cid:13)yT (I − P)A(cid:13)(cid:13)2(cid:19)
y∈span(Z)kyk2=1(cid:13)(cid:13)yT (I − P)A(cid:13)(cid:13)2 = 0.
max
56
It follows that the bottom k singular values are all 0, and so equal to the bottom k singular values
of σi(A − Ak).
Now, for i < n − k + 1, let UU denote the top i + k − 1 singular vectors of A and UL denote
the bottom n − (i + k − 1) singular vectors. We set Y = [Z, UL]. First, we note that Y has
k + n− (i + k − 1) = n− i + 1 columns and also rank(Y) = n− i + 1. This is because, if we consider
YT Y, the top left k × k blocks is ZT Z = I and the bottom right n − i − k + 1 block is UT
LUL = I.
The off-diagonal entries in the top right and bottom left blocks are all bounded by 1/ poly(n); the
proof of this latter statement is given in the latter part of this section, where we bound kZT ULk2.
By the Gershgorin circle theorem [GVL12], all eigenvalues of YT Y are in the range 1 ± 1/ poly(n)
and so the matrix is full rank. So, we have:
σi((I − P)A) ≤
max
y∈span(Y)kyk2=1(cid:13)(cid:13)yT (I − P)A(cid:13)(cid:13)2 .
(33)
Next, we can write y = [Z, UL]w for some w. By our argument above, every singular value of
[Z, UL] lies in 1± 1/ poly(n). Then, we have kwk2 ≤ 1 + 1/ poly(n). Splitting w = w1 + w2, where
w1 contains the first k coordinates of the vector and w2 contains the rest, we have:
(I − P)y = (I − P)Zw1 + (I − P)ULw2
= 0 + (I − P)ULw2
= ULw2 − ZZT ULw2.
Then, we have
(cid:13)(cid:13)ZZT ULw2(cid:13)(cid:13)2 ≤(cid:13)(cid:13)ZT UL(cid:13)(cid:13)2 · kw2k2 ≤ 1/ poly(n),
where the inequality comes from bounding(cid:13)(cid:13)ZT UL(cid:13)(cid:13)2 using the fact that all its entries are at most
1/ poly(n) (see the latter part of this section), and bounding kw2k2 ≤ kwk2 ≤ 1+1/ poly(n). Thus,
finally we obtain:
(cid:13)(cid:13)yT (I − P)A(cid:13)(cid:13)2 ≤(cid:13)(cid:13)wT
2 UT
2 UT
LA(cid:13)(cid:13)2 +(cid:13)(cid:13)wT
≤ (1 + 1/ poly(n)) ·(cid:13)(cid:13)UT
LZZT A(cid:13)(cid:13)2
LA(cid:13)(cid:13)2 + 1/ poly(n) · kAk2
σ1(A).
≤ (1 + 1/ poly(n))σi(A − Ak) +
poly(n)
1
Plugging into (33) gives the proof of (32). Note that we can assume
ǫ = o(1/ poly(n)), we can just compute the SVD.
1
poly(n) ≤ ǫ
n since if
Bound for (cid:13)(cid:13)ZT UL(cid:13)(cid:13)2
To see this, we first write AqΩ = ZTWT in its SVD form. Then, we have kuT
i ZTk2 ≥ kuT
: Let Z be an orthonormal basis for AqΩ from the power method (or
pq(A)Ω from block Krylov method). Then, for any singular vector ui with corresponding singular
value σi(A) with σi(A) ≤ (1 − ǫ)σk(A), we have kuT
i AqΩk2 =
kuT
i Zk2τk,k, where τk,k is the kth diagonal entry of T. On the other hand, we also
have kuT
is the right singular vector corresponding to ui of A.
2 = O(k log n) with probability 1 − 1/n2 for a fixed vi
Since Ω is i.i.d. Gaussian, we have kvT
(see proof of Lemma 2.12 in [Woo14]). So, we can union bound over all n right singular vectors
and the relation holds for all right singular vectors v ∈ V. We condition on this event.
σi(A)q poly(n), thenkuT
i AqΩk2 = O(σi(A)q√k log n). It follows that, if we show τk,k ≥
This event implies that kuT
i Ωk2, where vT
i Ωk2
i AqΩk2 = σi(A)qkvT
i Zk2 ≤ 1/ poly(n).
i Zk2 ≤ 1/ poly(n).
i
57
for q = Θ((log n)/ǫ) (for the block Krylov method, we need to use higher number of iterations than
For this, we have τk,k = σk(UΣqVT Ω) ≥ σk(UkUT
kVT
Next, to show that τk,k ≥ σi(A)q poly(n), it suffices to show τk,k ≥ σk(A)q√k. Indeed, note that
required, which is q = Θ((log n)/√ǫ) [MM15]), by definition of i, we have σk(Aq)/σi(Aq) ≥ poly(n).
k UΣqVT Ω) since σi(PA) ≤ σi(A) for any
projection matrix P and any matrix A. Then, we have σk(UkUT
k Ω) =
σk(Σq
k Ω) by the definition of the SVD, and that the columns of Uk are orthonormal. So, we
get τk,k ≥ σk(Σq
k Ω is a k × k matrix with i.i.d. Gaussian distribution due to
the rotational invariance of the Gaussian distribution. Since the k-th singular value is the smallest
√k · C with probability
singular value of H and Σk, we then have σk(Σq
at least 9/10, for an arbitrary constant C > 0, using standard properties of minimum singular
values of squared Gaussian matrices (see e.g., Fact 6 in section 4.3 of [Woo14]). Thus, we have
k UΣqVT Ω) = σk(UkΣq
k(A)σk(H) ≥ σq
kH), where H = VT
kH) ≥ σq
kVT
k
τk,k ≥ C√kσq
k(A), which completes the proof.
D Additional Proofs
Claim 10. Let f : R+ → R+ be a δf -multiplicatively smooth function on the range [a, b]. For any
x, y ∈ [a, b] and c ∈ (0, 1
3δf
)
y ∈ [(1 − c)x, (1 + c)x] ⇒ f (y) ∈ [(1 − 3δf c)f (x), (1 + 3δf c)f (x)].
Proof. Let R denote the range [min(x, y), max(x, y)]. For y ∈ [(1 − c)x, (1 + c)x] we have:
f (x) − f (y) ≤ x − y · sup
z∈R f′(z) ≤ cx ·
δf supz∈R f (z)
min(x, y) ≤
cδf supz∈R f (z)
1 − c
.
(34)
Similarly, letting Z = supz∈R f (z) we have:
and hence Z ≤
δf Z
cδf Z
min(x, y) ≤
Z − f (x) ≤ x − y
1 − c ≤
1−(δf +1)c f (x). So plugging into (34) we have:
cδf
1−c
cδf
c
δf
1 − 1/3 − 1/3
f (x) − f (y) ≤
1 − (δf + 1)c · f (x) ≤
1 − 1/3 − 1/3
f (x) = 3δf f (x)
which gives the result.
58
|
1905.00566 | 1 | 1905 | 2019-05-02T03:55:51 | Graph Coloring via Degeneracy in Streaming and Other Space-Conscious Models | [
"cs.DS"
] | We study the problem of coloring a given graph using a small number of colors in several well-established models of computation for big data. These include the data streaming model, the general graph query model, the massively parallel computation (MPC) model, and the CONGESTED-CLIQUE and the LOCAL models of distributed computation. On the one hand, we give algorithms with sublinear complexity, for the appropriate notion of complexity in each of these models. Our algorithms color a graph $G$ using about $\kappa(G)$ colors, where $\kappa(G)$ is the degeneracy of $G$: this parameter is closely related to the arboricity $\alpha(G)$. As a function of $\kappa(G)$ alone, our results are close to best possible, since the optimal number of colors is $\kappa(G)+1$.
On the other hand, we establish certain lower bounds indicating that sublinear algorithms probably cannot go much further. In particular, we prove that any randomized coloring algorithm that uses $\kappa(G)+1$ many colors, would require $\Omega(n^2)$ storage in the one pass streaming model, and $\Omega(n^2)$ many queries in the general graph query model, where $n$ is the number of vertices in the graph. These lower bounds hold even when the value of $\kappa(G)$ is known in advance; at the same time, our upper bounds do not require $\kappa(G)$ to be given in advance. | cs.DS | cs |
Graph Coloring via Degeneracy in Streaming and Other
Space-Conscious Models
Suman K. Bera∗
Amit Chakrabarti∗
Prantar Ghosh∗
May 3, 2019
Abstract
We study the problem of coloring a given graph using a small number of colors in several well-
established models of computation for big data. These include the data streaming model, the general
graph query model, the massively parallel communication (MPC) model, and the CONGESTED-CLIQUE
and the LOCAL models of distributed computation. On the one hand, we give algorithms with sublinear
complexity, for the appropriate notion of complexity in each of these models. Our algorithms color a
graph G using about κ(G) colors, where κ(G) is the degeneracy of G: this parameter is closely related to
the arboricity α(G). As a function of κ(G) alone, our results are close to best possible, since the optimal
number of colors is κ(G) + 1.
On the other hand, we establish certain lower bounds indicating that sublinear algorithms probably
cannot go much further. In particular, we prove that any randomized coloring algorithm that uses at most
κ(G) + 1 colors would require Ω(n2) storage in the one pass streaming model, and Ω(n2) many queries in
the general graph query model, where n is the number of vertices in the graph. These lower bounds hold
even when the value of κ(G) is known in advance; at the same time, our upper bounds do not require κ(G)
to be given in advance.
1 Introduction
Graph coloring is a fundamental topic in combinatorics and the corresponding algorithmic problem of
coloring an input graph with few colors is a basic and heavily studied problem in computer science. It has
numerous applications including in scheduling [TZP18, LS86, Lei79], air traffic flow management [BB04],
frequency assignment in wireless networks [BB06, PL96], register allocation [Cha82, CH90, CAC+81]. More
recently, vertex coloring has been used to compute seed vertices in social networks that are then expanded to
detect community structures in the network [MOT14].
Given an n-vertex graph G = (V, E), the task is to assign colors to the vertices in V so that no two adjacent
vertices get the same color. Doing so with the minimum possible number of colors -- called the chromatic
number, χ(G) -- is famously hard: it is NP-hard to even approximate χ(G) to a factor of n1−ε for any constant
ε > 0 [FK96, Zuc06, KP06]. In the face of this hardness, it is algorithmically interesting to color G with a
possibly suboptimal number of colors depending upon tractable parameters of G. One such simple parameter
is ∆, the maximum degree: a trivial greedy algorithm colors G with ∆ + 1 colors.
We study graph coloring in a number of space-constrained and data-access-constrained settings, including
the data streaming model and certain distributed computing models. In such settings, finding a coloring with
"about ∆" colors is a fairly nontrivial problem that has been studied from various angles in a flurry of research
over the last decade [ACK19,BCHN18,HSS16,CLP18,PS18,Bar16]. In a recent breakthrough (awarded Best
∗Department of Computer Science, Dartmouth College. Supported in part by NSF under Award CCF-1650992.
1
Paper at SODA 2019), Assadi, Chen, and Khanna [ACK19] gave sublinear algorithms for (∆ + 1)-coloring an
input graph in such models.
In this work, we focus on colorings that use "about κ" colors, where κ = κ(G) is the degeneracy of G, a
parameter that improves upon ∆. It is defined as follows: κ = min{k : every induced subgraph of G has a
vertex of degree at most k}. Clearly, κ (cid:54) ∆. There is a simple greedy algorithm that runs in linear time and
produces (κ + 1)-coloring; see Section 2. However, just as before, when processing a massive graph under
the constraints of either the space-bounded streaming model or certain distributed computing models, the
inherently sequential nature of the greedy algorithm makes it infeasible.
1.1 Our Results and Techniques
We obtain a number of algorithmic results, as well as several lower bound results.
Algorithms. We give new graph coloring algorithms, parametrized by κ, in the following models: (1) the
data streaming model, where the input is a stream of edge insertions and deletions (i.e., a dynamic graph
stream) resulting in the eventual graph to be colored and we are limited to a work space of (cid:101)O(n) bits1, the
so-called semi-streaming setting [FKM+05]; (2) the general graph query model [Gol17], where we may
access the graph using only neighbor queries (what is the ith neighbor of x?) and pair queries (are x and
y adjacent?); (3) the massively parallel communication (MPC) model, where each of a large number of
memory-limited processors holds a sublinear-sized portion of the input data and computation proceeds
using rounds of communication; (4) the congested clique model of distributed computation, where there
is one processor per vertex holding that vertex's neighborhood information and each round allows each
processor to communicate O(log n) bits to a specific other processor; and (5) the LOCAL model of distributed
computation, where there is one processor per vertex holding that vertex's neighborhood information and
each round allows each processor to send an arbitrary amount of information to all its neighbors.
Model
Number of Colors
Streaming
(one pass)
Query
MPC
Congested Clique
LOCAL
∆ + 1
κ + o(κ)
∆ + 1
κ + o(κ)
∆ + 1
κ + o(κ)
∆ + 1
κ + o(κ)(cid:63)
O(αn1/k)
O(αn1/k log n)
(cid:101)O(n) space,
(cid:101)O(n) space,
∆) post-processing time
Complexity Parameters
√
(cid:101)O(n
(cid:101)O(n) post-processing time
(cid:101)O(n3/2) queries
(cid:101)O(n3/2) queries
O(1) rounds, O(n log3 n) bits per processor
O(1) rounds, O(n log2 n) bits per processor
O(1) rounds
O(1) rounds
(cid:112)
for k ∈(cid:2)ω(log log n), O(
log n), o(log n)(cid:3)
for k ∈(cid:2)ω(
(cid:112)
log n)(cid:3)
O(k) rounds,
O(k) rounds,
Source
[ACK19]
this paper
[ACK19]
this paper
[ACK19]
this paper
[CFG+18]
this paper
[KP11]
this paper
Table 1: Summary of our algorithmic results and basic comparison with most related previous work. In the result marked
((cid:63)), we require that κ = ω(log2 n). In the first two results, the number of colors can be improved to min{∆ + 1, κ + o(κ)}
by running our algorithm alongside that of [ACK19]; in the streaming setting, this would require knowing ∆ in advance.
Table 1 summarizes our algorithmic results and provides, in each case, a basic comparison with the most
related result from prior work; more details appear in Section 1.2. As we have noted, κ (cid:54) ∆ in every case;
1The (cid:101)O(·) notation hides factors polylogarithmic in n.
2
indeed, κ could be arbitrarily better than ∆ as shown by the example of a star graph, where κ = 1 whereas
∆ = n − 1. From a practical standpoint, it is notable that in many real-world large graphs drawn from various
application domains -- such as social networks, web graphs, and biological networks -- the parameter κ is often
significantly smaller than ∆. See Table 2 for some concrete numbers. That said, κ + o(κ) is mathematically
incomparable with ∆ + 1.
Graph Name
soc-friendster
fb-uci-uni
soc-livejournal
soc-orkut
web-baidu-baike
web-hudong
web-wikipedia2009
web-google
bio-mouse-gene
bio-human-gene1
bio-human-gene2
bio-WormNet-v3
∆
E
2B
V
κ
305
66M
5K
17
59M 92M 5K
4M
214
28M 3K
3M 106M 27K 231
18M 98K
2M
83
2M
15M 62K 529
67
3K
5M
2M
65
5M
916K
6K
1K
14M 8K
43K
22K
12M 8K
2K
2K
7K
9M
14K
16K
763K
1K
165
Table 2: Statistics of several large real-world graphs taken from the application domains of social networks, web graphs,
and biological networks, showing that the degeneracy, κ, is often significantly smaller than the maximum degree, ∆.
Source: http://networkrepository.com [RA15].
The parameter κ is also closely related to the arboricity α = α(G), defined as the minimum number of
forests into which the edges of G can be partitioned. It is an easy exercise to show that α (cid:54) κ (cid:54) 2α − 1.
Perhaps even more than these results, our key contribution is a conceptual idea and a corresponding
technical lemma underlying all our algorithms. We show that every graph admits a "small" sized low
degeneracy partition (LDP), which is a partition of its vertex set into "few" blocks such that the subgraph
induced by each block has low degeneracy, roughly logarithmic in n. Moreover, such an LDP can be computed
by a very simple and distributed randomized algorithm: for each vertex, choose a "color" independently and
uniformly at random from a suitable-sized palette (this is not to be confused with the eventual graph coloring
we seek; this random assignment is most probably not a proper coloring of the graph). The resulting color
classes define the blocks of such a partition, with high probability. Theorem 3.1, the LDP Theorem, makes
this precise.
Given an LDP, a generic graph coloring algorithm is to run the aforementioned minimum-degree-based
greedy algorithm on each block, using distinct palettes for the distinct blocks. We obtain algorithms achieving
our claimed results by suitably implementing this generic algorithm in each computational model.
Lower Bounds. Recall that a graph with degeneracy κ admits a proper (κ + 1)-coloring. As Table 1
makes clear, there are several space-conscious (∆ + 1)-coloring algorithms known; perhaps we could aim
for improved algorithms that provide (κ + 1)-colorings? We prove that this is not possible in sublinear
space in either the streaming or the query model. In fact, our lower bounds prove more. We show that
distinguishing n-vertex graphs of degeneracy κ from those with chromatic number κ + 2 requires Ω(n2) space
in the streaming model and Ω(n2) queries in the general graph query model. This shows that it is hard to
produce a (κ + 1)-coloring and in fact even to determine the value of κ. These results generalize to the
problems of producing a (κ + λ)-coloring or estimating the degeneracy up to ±λ; the corresponding lower
bounds are Ω(n2/λ2). Furthermore, the streaming lower bounds hold even in the insertion-only model, where
the input stream is simply a listing of the graph's edges in some order; compare this with our upper bound,
which holds even for dynamic graph streams.
3
A possible criticism of the above lower bounds for coloring is that they seem to depend on it being hard to
estimate the degeneracy κ. Perhaps the coloring problem could become easier if κ was given to the algorithm
in advance? We prove two more lower bounds showing that this is not so: the same Ω(n2/λ2) bounds hold
even with κ known a priori.
Most of our streaming lower bounds use reductions from the index problem in communication complexity
(a standard technique), via a novel gadget that we develop here; one bound uses a reduction from a variant of
disjointness. Our query lower bounds use a related gadget and reductions from basic problems in Boolean
decision tree complexity.
We conclude the paper with a "combinatorial" lower bound that addresses a potential criticism of our
main algorithmic technique: the LDP. Perhaps a more sophisticated graph-theoretic result, such as the Palette
Sparsification Theorem of Assadi et al. (see below), could improve the quality of the colorings obtained? We
prove that this is not so: there is no analogous theorem for colorings with "about κ" colors.
1.2 Related Work and Comparisons
Streaming and Query Models. The work closest to ours in spirit is the recent breakthrough of Assadi,
Chen, and Khanna [ACK19]: they give a one-pass streaming (∆ + 1)-coloring algorithm that uses (cid:101)O(n) space
(i.e., is semi-streaming) and works on dynamic graph streams. Their algorithm exploits a key structural result
that they establish: choosing a random O(log n)-sized palette from {1, . . . , ∆ + 1} for each vertex allows a
stream update quickly, but then spends (cid:101)O(n
√
with the stream updates and has a faster (cid:101)O(n)-time post-processing step. Further, our algorithm is "truly
compatible list coloring. They call this the Palette Sparsification Theorem. Their algorithm processes each
∆) time in post-processing. Our algorithm is similarly quick
(∆ + 1)-coloring algorithm that makes (cid:101)O(n3/2) queries, followed by a fairly elaborate computation that runs
in (cid:101)O(n3/2) time and space. Our algorithm has the same complexity parameters and is arguably much simpler:
one-pass" in that it does not require foreknowledge of κ or any other parameter of the input graph, whereas
the Assadi et al. algorithm needs to know the precise value of ∆ before seeing the stream.
In the same paper, Assadi et al. also consider the graph coloring problem in the query model. They give a
its post-processing is just the straightforward greedy offline algorithm for (κ + 1)-coloring.
Another recent work on coloring in the streaming model is Radhakrishnan et al. [RSV15], which studies
the problem of 2-coloring an n-uniform hypergraph. In the query model, there are a number of works studying
basic graph problems [GR08, PR07, CRT05] but, to the best of our knowledge, Assadi et al. were the first
to study graph coloring in this sense. Also, to the best of our knowledge, there was no previously known
algorithm for O(α)-coloring in a semi-streaming setting, whereas here we obtain (κ + o(κ))-colorings; recall
the bound κ (cid:54) 2α − 1.
MPC and Congested Clique Models. The MapReduce framework [DG04] is extensively used in distributed
computing to analyze and process massive data sets. Beame, Koutris, and Suciu [BKS17] defined the
Massively Parallel Communication (MPC) model to abstract out key theoretical features of MapReduce;
it has since become a widely used setting for designing and analyzing big data algorithms, especially for
graph problems. In this model, an input of size m is distributed among p ≈ m/S processors, each of which is
computationally unbounded and restricted to S bits of space. The processors operate in synchronous rounds;
in each round, a processor may communicate with all others, subject to the space constraint. The focus is on
using a very small number of rounds.
Another well-studied model for distributed graph algorithms is Congested Clique [LPPP05], where there
are n nodes, each holding the local neighborhood information for one of the n vertices of the input graph. The
nodes communicate in synchronous rounds; in a round, every pair of processors may communicate, but each
message is restricted to O(log n) bits. Behnezhad et al. [BDH18] show that Congested Clique is equivalent to
4
the so-called "semi-MPC model," defined as MPC with O(n log n) bits of memory per machine: there are
simulations in both directions preserving the round complexity.
Graph coloring has been studied in these models before. Harvey et al. [HLL18] gave a (∆ + o(∆))-coloring
algorithm in the MapReduce model; it can be simulated in MPC using O(1) rounds and O(n1+c) space per
machine for some constant c > 0. Parter [Par18] gave a Congested Clique algorithm for (∆ + 1)-coloring
using O(log log ∆ · log(cid:63) ∆) rounds; Parter and Su [PS18] improved this to O(log(cid:63) ∆). The aforementioned
paper of Assadi et al. [ACK19] gives an MPC algorithm for (∆ + 1)-coloring using O(1)-round and O(n log3 n)
bits of space per machine. Because this space usage is ω(n log n), the equivalence result of Behnezad et
al. [BDH18] does not apply and this doesn't lead to an O(1)-round Congested Clique algorithm. In contrast,
our MPC algorithm uses only O(n log n) bits of space per machine for graphs with degeneracy ω(log2 n), and
therefore leads to such a Congested Clique algorithm. Chang et al. [CFG+18] have recently designed two
(∆ + 1) list-coloring algorithms: an O(1)-round Congested Clique algorithm, and an O(
log log n)-round
MPC algorithm with o(n) space per machine and (cid:101)O(m) space in total. To the best of our knowledge, no
(cid:112)
O(α)-coloring algorithm was previously known, in either the MPC or the Congested Clique model.
The LOCAL Model. The LOCAL model of distributed computing is "orthogonal" to Congested Clique:
the input setup is similar but, during computation, each node may only communicate with its neighbors in
the input graph, though it may send an arbitrarily long message. As before, the focus is on minimizing the
number of rounds (a.k.a., time). There is a deep body of work on graph coloring in this model. Indeed, graph
coloring is one of the most central "symmetry breaking" problems in distributed computing. We refer the
reader to the monograph by Barenboim and Elkin [BE13] for an excellent overview of the state of the art.
Here, we shall briefly discuss only a few results closely related to our contribution.
There is a long line of work on fast (∆+1)-coloring in the LOCAL model, in the deterministic as well as the
randomized setting [PS96,Bar16,FHK16,Lub86,Joh99,ABI86,SW10,BEPS16] culminating in sublogarithmic
time solutions due to Harris [HSS16] and Chang et al. [CLP18]. Barenboim and Elkin [BE10, BE11] studied
fast distributed coloring algorithms that may use far fewer than ∆ colors: in particular, they gave algorithms
that use O(α) colors and run in O(αε log n) time on graphs with arboricity at most α. Recall again that
κ (cid:54) 2α − 1, so that a 2α-coloring always exists. They also gave a faster O(log n)-time algorithm using O(α2)
colors. Further, they gave a family of algorithms that produce an O(tα2)-coloring in O(logt n + log(cid:63) n), for
every t such that 2 (cid:54) t (cid:54) O(
√
n/α). Our algorithm for the LOCAL model builds on this latter result.
They gave a randomized O(k)-round algorithm that uses O(αn1/k) colors for 2 log log n (cid:54) k (cid:54) (cid:112)
O(α1+1/kn1/k+3/k22−2k) colors for k < 2 log log n. We extend their result to the range k ∈(cid:2)ω(
Kothapalli and Pemmaraju [KP11] focused on arboricity-dependant coloring using very few rounds.
log n and
log n), o(log n)(cid:3),
(cid:112)
using O(αn1/k log n) colors.
Ghaffari and Lymouri [GL17] gave a randomized O(α)-coloring algorithm that runs in time O(log n ·
min{log log n, log α}) as well as an O(log n)-time algorithm using min{(2+ε)α+O(log n log log n), O(α log α)}
colors, for any constant ε > 0. However, their technique does not yield a sublogarithmic time algorithm, even
at the cost of a larger palette.
The LDP Technique. As mentioned earlier, our algorithmic results rely on the concept of a low degeneracy
partition (LDP) that we introduce in this work. Some relatives of this idea have been considered before.
Specifically, Barenboim and Elkin [BE13] define a d-defective (resp. b-arbdefective) c-coloring to be a vertex
coloring using palette [c] such that every color class induces a subgraph with maximum degree at most d
(resp. arboricity at most b). Obtaining such improper colorings is a useful first step towards obtaining proper
colorings. They give deterministic algorithms to obtain good arbdefective colorings [BE11]. However, their
algorithms are elaborate and are based on construction of low outdegree acyclic partial orientations of the
graph's edges: an expensive step in our space-conscious models.
Elsewhere (Theorem 10.5 of Barenboim and Elkin [BE13]), they note that a useful defective (not
arbdefective) coloring is easily obtained by randomly picking a color for each vertex; this is then useful for
5
computing an O(∆)-coloring.
Our LDP technique can be seen as a simple randomized method for producing an arbdefective coloring.
Crucially, we parametrize our result using degeneracy instead of arboricity and we give sharp -- not just
asymptotic -- bounds on the degeneracy of each color class.
Other Related Work. Other work considers coloring in the setting of dynamic graph algorithms: edges
are inserted and deleted over time and the goal is to maintain a valid vertex coloring of the graph that must
be updated quickly after each modification. Unlike in the streaming setting, there is no space restriction.
Bhattacharya et al. [BCHN18] gave a randomized algorithm that maintains a (∆ + 1)-coloring with O(log ∆)
expected amortized update time and a deterministic algorithm that maintains a (∆ + o(∆))-coloring with
O(polylog ∆) amortized update time. Barba et al. [BCK+17] gave tradeoffs between the number of colors
used and update time. However, the techniques in these works do not seem to apply in the streaming setting
due to fundamental differences in the models.
al. [MTVV15] gave a one pass (1 + ε)-approximation algorithm to estimate the arboricity of graph using (cid:101)O(n)
Estimating the arboricity of a graph in the streaming model is a well studied problem. McGregor et
space. Bahmani et al. [BKV12] gave a matching lower bound. Our lower bounds for estimating degeneracy
are quantitatively much larger but they call for much tighter estimates.
2 Preliminaries
Throughout this paper, graphs are simple, undirected, and unweighted. In considering a graph coloring
problem, the input graph will usually be called G and we will put n = V(G). The notation "log x" stands for
log2 x. For an integer k, we denote the set {1, 2, . . . , k} by [k].
For a graph G, we define ∆(G) = max{deg(v) : v ∈ V(G)}. We say that G is k-degenerate if every induced
subgraph of G has a vertex of degree at most k. For instance, every forest is 1-degenerate and an elementary
theorem says that every planar graph is 5-degenerate. The degeneracy κ(G) is the smallest k such that G
is k-degenerate. The arboricity α(G) is the smallest r such that the edge set E(G) can be partitioned into r
forests. When the graph G is clear from the context, we simply write ∆, κ, and α, instead of ∆(G), κ(G), and
α(G).
We note two useful facts: the first is immediate from the definition, and the second is an easy exercise.
Fact 2.1. If an n-vertex graph has degeneracy κ, then it has at most κn edges.
Fact 2.2. In every graph, the degeneracy κ and arboricity α satisfy α (cid:54) κ (cid:54) 2α − 1.
(cid:3)
(cid:3)
In analyzing our algorithms, it will be useful to consider certain vertex orderings of graphs and their
connection with the notion of degeneracy, given by Lemma 2.5 below. Although the lemma is folklore, it is
crucial to our analysis, so we include a proof for completeness.
Definition 2.3. An ordering of G is a list consisting of all its vertices (equivalently, a total order on V(G)).
Given an ordering (cid:67), for each v ∈ V(G), the ordered neighborhood
NG,(cid:67)(v) := {w ∈ V(G) : {v, w} ∈ E(G), v (cid:67) w} ,
i.e., the set of neighbors of v that appear after v in the ordering. The ordered degree odegG,(cid:67)(v) := NG,(cid:67)(v).
Definition 2.4. A degeneracy ordering of G is an ordering produced by the following algorithm: starting
with an empty list, repeatedly pick a minimum degree vertex v (breaking ties arbitrarily), append v to the end
of the list, and delete v from G; continue this until G becomes empty.
6
Lemma 2.5. A graph G is k-degenerate iff there exists an ordering (cid:67) such that odegG,(cid:67)(v) (cid:54) k for all
v ∈ V(G).
Proof. Suppose that G is k-degenerate. Let (cid:67) = (v1, . . . , vn) be a degeneracy ordering. Then, for each i,
odegG,(cid:67)(vi) is the degree of vi in the induced subgraph G \{v1, . . . , vi−1}. By definition, this induced subgraph
has a vertex of degree at most k, so vi, being a minimum degree vertex in the subgraph, must have degree at
most k.
On the other hand, suppose that G has an ordering (cid:67) such that odegG,(cid:67)(v) (cid:54) k for all v ∈ V(G). Let H
be an induced subgraph of G. Let v be the leftmost (i.e., smallest) vertex in V(H) according to (cid:67). Then all
(cid:3)
neighbors of v in H in fact lie in NG,(cid:67)(v), so degH(v) (cid:54) odegG,(cid:67)(v) (cid:54) k. Therefore, G is k-degenerate.
A c-coloring of a graph G is a mapping ψ: V(G) → [c]; it is said to be a proper coloring if it makes no
edge monochromatic: ψ(u) (cid:44) ψ(v) for all {u, v} ∈ E(G). The smallest c such that G has a proper c-coloring is
called the chromatic number χ(G). By considering the vertices of G one at a time and coloring greedily, we
immediately obtain a proper (∆ + 1)-coloring. This idea easily extends to degeneracy-based coloring.
Lemma 2.6. Given unrestricted ("offline") access to an input graph G, we can produce a proper (κ + 1)-
coloring in linear time.
Proof. Construct a degeneracy ordering (v1, . . . , vn) of G and then consider the vertices one by one in the
order (vn, . . . , v1), coloring greedily. Given a palette of size κ + 1, by the "only if" direction of Lemma 2.5,
(cid:3)
there will always be a free color for a vertex when it is considered.
Of course, the simple algorithm above is not implementable directly in "sublinear" settings, such as
space-bounded streaming algorithms, query models, or distributed computing models. Nevertheless, we shall
make use of the algorithm on suitably constructed subgraphs of our input graph.
We shall use the following form of the Chernoff bound.
such that EX (cid:54) µ and 0 (cid:54) δ (cid:54) 1. Then, Pr(cid:2)X (cid:62) (1 + δ)µ(cid:3) (cid:54) exp
Fact 2.7. Let X be a sum of mutually independent indicator random variables. Let µ and δ be real numbers
(cid:3)
(cid:16)−µδ2/3
(cid:17)
.
3 A Generic Framework for Coloring
In this section, we give a generic framework for graph coloring that we later instantiate in various compu-
tational models. As a reminder, our focus is on graphs G with a nontrivial upper bound on the degeneracy
κ = κ(G). Each such graph admits a proper (κ + 1)-coloring; our focus will be on obtaining a proper
(κ + o(κ))-coloring efficiently.
As a broad outline, our framework calls for coloring G in two phases. The first phase produces a low
degeneracy partition (LDP) of G: it partitions V(G) into a "small" number of parts, each of which induces a
subgraph that has "low" degeneracy. This step can be thought of as preprocessing and it is essentially free
(in terms of complexity) in each of our models. The second phase properly colors each part, using a small
number of colors, which is possible because the degeneracy is low. In Section 4, we shall see that the low
degeneracy allows this second phase to be efficient in each of the models we consider.
3.1 A Low Degeneracy Partition and its Application
In this phase of our coloring framework, we assign each vertex a color chosen uniformly at random from [(cid:96)],
these choices being mutually independent, where (cid:96) is a suitable parameter. For each i ∈ [(cid:96)], let Gi denote the
subgraph of G induced by vertices colored i. We shall call each Gi a block of the vertex partition given by
(G1, . . . , G(cid:96)). The next theorem, our main technical tool, provides certain guarantees on this partition given a
suitable choice of (cid:96).
7
Theorem 3.1 (LDP Theorem). Let G be an n-vertex graph with degeneracy κ. Let k ∈ [1, n] be a "guess" for
the value of κ and let s (cid:62) Cn log n be a sparsity parameter, where C is a sufficiently large universal constant.
Put
(1)
and let ψ: V(G) → [(cid:96)] be a uniformly random coloring of G. For i ∈ [(cid:96)], let Gi be the subgraph induced by
ψ−1(i). Then, the partition (G1, . . . , G(cid:96)) has the following properties.
κ(cid:96) log n ,
λ = 3
(cid:96) =
s
,
(cid:38)2nk
(cid:39)
(cid:112)
(i) If k (cid:54) 2κ, then w.h.p., for each i, the degeneracy κ(Gi) (cid:54) (κ + λ)/(cid:96).
(ii) W.h.p., for each i, the block size V(Gi) (cid:54) 2n/(cid:96).
(iii) If κ (cid:54) k (cid:54) 2κ, then w.h.p., the number of monochromatic edges E(G1) ∪ ··· ∪ E(G(cid:96)) (cid:54) s.
In each case, "w.h.p." means "with probability at least 1 − 1/ poly(n)."
It will be convenient to encapsulate the guarantees of this theorem in a definition.
Definition 3.2. Suppose graph G has degeneracy κ. A vertex partition (G1, . . . , G(cid:96)) simultaneously satisfying
the degeneracy bound in item (i), the block size bound in item (ii), and the (monochromatic) edge sparsity
bound in item (iii) in Theorem 3.1 is called an ((cid:96), s, λ)-LDP of G.
It will turn out that an ((cid:96), s, λ)-LDP leads to a proper coloring of G using at most κ + λ + (cid:96) colors. An
instructive setting of parameters is s = Θ((n log n)/ε2), where ε is either a small constant or a slowly vanishing
made, Theorem 3.1 guarantees an LDP that has edge sparsity s = (cid:101)O(n) and that leads to an eventual proper
function of n, such as 1/ log n. Then, a quick calculation shows that when an accurate guess k ∈ [κ, 2κ] is
coloring using (1 + O(ε))κ colors. When ε = o(1), this number of colors is κ + o(κ).
Recall that the second phase of our coloring framework involves coloring each Gi separately, exploiting its
low degeneracy. Indeed, given an ((cid:96), s, λ)-LDP, each block Gi admits a proper (κ(Gi) + 1)-coloring. Suppose
we use a distinct palette for each block; then the total number of colors used is
(cid:96)(cid:88)
i=1
(cid:19)
(cid:18) κ + λ
(cid:96)
(κ(Gi) + 1) (cid:54) (cid:96)
+ 1
= κ + λ + (cid:96) ,
(2)
as claimed above. Of course, even if our first phase random coloring ψ yields a suitable LDP, we still
have to collect each block Gi or at least enough information about each block so as to produce a proper
(κ(Gi) + 1)-coloring. How we do this depends on the precise model of computation. We take this up in
Section 4.
3.2 Proof of the LDP Theorem
We now turn to proving the LDP Theorem from Section 3.1. Notice that when k (cid:54) (C/2) log n, the condition
s (cid:62) Cn log n results in (cid:96) = 1, so the vertex partition is the trivial one-block partition, which obviously satisfies
all the properties in the theorem. Thus, in our proof, we may assume that k > (C/2) log n.
Proof of Theorem 3.1. We start with item (ii), which is the most straightforward. From eq. (1), we have
(cid:96) (cid:54) 4nk/s, so
Each block size V(Gi) has binomial distribution Bin(n, 1/(cid:96)), so a Chernoff bound gives
(cid:34)
V(Gi) >
Pr
n
(cid:96)
(cid:62) s
4k
(cid:35)
2n
(cid:96)
(cid:62) Cn log n
(cid:62) C log n
.
4
(cid:32)
(cid:33)
(cid:54) exp
(cid:54) exp
−C log n
12
(cid:54) 1
n2 ,
(cid:19)
4k
(cid:18)− n
3(cid:96)
8
for sufficiently large C. By a union bound over the at most n blocks, item (ii) fails with probability at most
1/n.
Items (i) and (iii) include the condition k (cid:54) 2κ, which we shall assume for the rest of the proof. By eq. (1)
and the bounds s (cid:62) Cn log n and k > (C/2) log n,
(cid:38)
(cid:39)
(cid:115)
κ ·
(cid:96) (cid:54)
whence, for sufficiently large C,
2k
C log n
(cid:54)
4k
C log n
(cid:54)
8κ
C log n ,
8κ
· log n (cid:54) κ .
λ (cid:54) 3
(3)
We now turn to establishing item (i). Let (cid:67) be a degeneracy ordering for G. For each i ∈ [(cid:96)], let (cid:67)i be the
restriction of (cid:67) to V(Gi). Consider a particular vertex v ∈ V(G) and let j = ψ(v) be its color. We shall prove
that, w.h.p., odegG,(cid:67) j(v) (cid:54) (κ + λ)/(cid:96).
By the "only if" direction of Lemma 2.5, we have odegG,(cid:67)(v) = NG,(cid:67)(v) (cid:54) κ. Now note that
C log n
(cid:88)
u∈NG,(cid:67)(v)
odegG j,(cid:67) j(v) =
1{ψ(u)=ψ(v)}
is a sum of mutually independent indicator random variables, each of which has expectation 1/(cid:96). Therefore,
E odegG j,(cid:67) j(v) = odegG,(cid:67)(v)/(cid:96) (cid:54) κ/(cid:96). Since λ (cid:54) κ by eq. (3), we may use the form of the Chernoff bound in
Fact 2.7, which gives us
(cid:20)
(cid:21)
(cid:32)
(cid:33)
(cid:32)
(cid:33)
Pr
odegG j,(cid:67) j(v) >
κ + λ
(cid:96)
(cid:54) exp
− κ
(cid:96)
λ2
3κ2
= exp
−9κ(cid:96) log n
3κ(cid:96)
(cid:54) 1
n3 ,
where the equality follows from eq. (1). In words, with probability at least 1 − 1/n3, the vertex v has ordered
degree at most (κ + λ)/(cid:96) within its own block. By a union bound, with probability at least 1 − 1/n2, all n
vertices of G satisfy this property. When this happens, by the "if" direction of Lemma 2.5, it follows that
κ(Gi) (cid:54) (κ + λ)/(cid:96) for every i.
Finally, we take up item (iii), which is now straightforward. Assume that the high probability event in
item (i) occurs. Then, by Fact 2.1,
E(G1) ∪ ··· ∪ E(G(cid:96)) (cid:54)
(cid:96)(cid:88)
i=1
(cid:96)(cid:88)
i=1
κ(Gi)V(Gi) (cid:54) κ + λ
(cid:96)
V(Gi) =
n(κ + λ)
(cid:96)
(cid:54) 2nκ
(cid:96)
(cid:54) s ,
where the final inequality uses the condition κ (cid:54) k and eq. (1).
(cid:3)
4 Specific Sublinear Algorithms for Coloring
We now turn to designing graph coloring algorithms in specific models of computation for big data, where the
focus is on utilizing space sublinear in the size of the massive input graph. Such models are sometimes termed
space-conscious. In each case, our algorithm ultimately relies on the framework developed in Section 3.
9
4.1 Data Streaming
We begin with the most intensely studied space-conscious model: the data streaming model. For graph
problems, in the basic model, the input is a stream of non-repeated edges that define the input graph G: this
is called the insertion-only model, since it can be thought of as building up G through a sequence of edge
insertions. In the more general dynamic graph model or turnstile model, the stream is a sequence of edge
updates, each update being either an insertion or a deletion: the net effect is to build up G. Our algorithm will
work in this more general model. Later, we shall give a corresponding lower bound that will hold even in the
insertion-only model (for a lower bound, this is a strength).
(cid:16)n
(cid:17)
We assume that the vertex set V(G) = [n] and the input is a stream σ of at most m = poly(n) updates
to an initially empty graph. An update is a triple (u, v, c), where u, v ∈ V(G) and c ∈ {−1, 1}: when c = 1,
this token represents an insertion of edge {u, v} and when c = −1, it represents a deletion. Let N =
and
[[m]] = Z ∩ [−m, m]. It is convenient to imagine a vector x ∈ [[m]]N of edge multiplicities that starts at zero
and is updated entrywise with each token. The input graph G described by the stream will be the underlying
simple graph, i.e., E(G) will be the set of all edges {u, v} such that xu,v (cid:44) 0 at the end. We shall say that σ
builds up x and G.
2
Our algorithm makes use of two data streaming primitives, each a linear sketch. (We can do away with
these sketches in the insertion-only setting; see the end of this section.) The first is a sketch for sparse
recovery given by a matrix A (say): given a vector x ∈ [[m]]N with sparsity (cid:107)x(cid:107)0 (cid:54) t, there is an efficient
algorithm to reconstruct x from Ax. The second is a sketch for (cid:96)0 estimation given by a random matrix B
(say): given a vector x ∈ [[m]]N, there is an efficient algorithm that takes Bx and computes from it an estimate
of (cid:107)x(cid:107)0 that, with probability at least 1 − δ, is a (1 + γ)-multiplicative approximation. It is known that there
exists a suitable A ∈ {0, 1}d×N, where d = O(t log(N/t)), where A has column sparsity O(log(N/t)); see, e.g.,
Theorem 9 of Gilbert and Indyk [GI10]. It is also known that there exists a suitable distribution over matrices
giving B ∈ {0, 1}d(cid:48)×N with d(cid:48) = O(γ−2 log δ−1 log N(log γ−1 + log log m)). Further, given an update to the ith
entry of x, the resulting updates in Ax and Bx can be effected quickly by generating the required portion of
the ith columns of A and B.
(cid:46) σ builds up x and G; k ∈ [1, n] is a guess for κ(G)
Algorithm 1 One-Pass Streaming Algorithm for Graph Coloring via Degeneracy
1: procedure Color(stream σ, integer k)
choose s, (cid:96) as in eq. (1) and t, d, d(cid:48), A, B as in the above discussion
2:
initialize y ∈ [[m]]d and z ∈ [[m]]d(cid:48)
3:
foreach u ∈ [n] do ψ(u) ← uniform random color in [(cid:96)]
4:
foreach token (u, v, c) in σ do
5:
if ψ(u) = ψ(v) then y ← y + cAu,v; z ← z + cBu,v
6:
if estimate of (cid:107)w(cid:107)0 obtained from z is > 5s/4 then abort
w(cid:48) ← result of t-sparse recovery from y
foreach i ∈ [(cid:96)] do
to zero
7:
8:
9:
10:
11:
Gi ← simple graph induced by {{u, v} : w(cid:48)
color Gi using palette {(i, j) : 1 (cid:54) j (cid:54) κ(Gi) + 1}; cf. Lemma 2.6
u,v (cid:44) 0 and ψ(u) = ψ(v) = i}
(cid:46) we expect that w(cid:48) = w
(cid:46) net effect is to color G
In our description of Algorithm 1, we use Au,v (resp. Bu,v) to denote the column of A (resp. B) indexed by
{u, v}. The algorithm's logic results in sketches y = Aw and z = Bw, where w corresponds to the subgraph of
G consisting of ψ-monochromatic edges only (cf. Theorem 3.1), i.e., w is obtained from x by zeroing out all
entries except those indexed by {u, v} with ψ(u) = ψ(v). We choose the parameter t = 2s, where s (cid:62) Cn log n
is the sparsity parameter from Theorem 3.1, which gives d = O(s log n); we choose γ = 1/4 and δ = 1/n,
giving d(cid:48) = O(log3 n).
10
Notice that Algorithm 1 requires a guess for κ := κ(G), which is not known in advance. Our final one-pass
algorithm runs O(log n) parallel instances of Color(σ, k), using geometrically spaced guesses k = 2, 4, 8 . . . .
It outputs the coloring produced by the non-aborting run that uses the smallest guess.
Theorem 4.1. Set s = (cid:100)ε−2n log n(cid:101), where ε > 0 is a parameter. The above one-pass algorithm processes a
dynamic (i.e., turnstile) graph stream using O(ε−2n log4 n) bits of space and, with high probability, produces
coloring using (cid:101)O(n) space. Each edge update is processed in (cid:101)O(1) time and post-processing at the end of the
stream takes (cid:101)O(n) time.
a proper coloring using at most (1 + O(ε))κ colors. In particular, taking ε = 1/ log n, it produces a κ + o(κ)
Proof. The coloring produced is obviously proper. Let us bound the number of colors used. One of the
parallel runs of Color(σ, k) in 1 will use a value k = k(cid:63) ∈ (κ, 2κ]. We shall prove that, w.h.p., (a) every
non-aborting run with k (cid:54) k(cid:63) will use at most (1 + O(ε))κ colors, and (b) the run with k = k(cid:63) will not abort.
We start with (a). Consider a particular run using k (cid:54) k(cid:63). By item (i) of Theorem 3.1, each Gi has
degeneracy at most (κ + λ)/(cid:96); so if w is correctly recovered by the sparse recovery sketch (i.e., w(cid:48) = w in
Algorithm 1), then each Gi is correctly recovered and the run uses at most κ + λ + (cid:96) colors, as in eq. (2).
Using the values from eq. (1), this number is at most (1 + O(ε))κ. Now, if the run does not abort, then the
estimate of the sparsity (cid:107)w(cid:107)0 is at most 5s/4. By the guarantees of the (cid:96)0-estimation sketch, the true sparsity
is at most (5/4)(5s/4) < 2s = t, so, w.h.p., w is indeed t-sparse and, by the guarantees of the sparse recovery
sketch, w(cid:48) = w. Taking a union bound over all O(log n) runs, the bound on the number of colors holds for all
required runs simultaneously, w.h.p.
We now take up (b). Note that (cid:107)w(cid:107)0 is precisely the number of ψ-monochromatic edges in G. By item (iii)
of Theorem 3.1, we have (cid:107)w0(cid:107) (cid:54) s w.h.p. By the accuracy guarantee of the (cid:96)0-estimation sketch, in this run
the estimate of (cid:107)w(cid:107)0 is at most 5s/4 w.h.p., so the run does not abort.
The space usage of each parallel run is dominated by the computation of y, so it is O(d log m) =
O(s log n log m) = O(ε−2n log3 n), using our setting of s and the assumption m = poly(n). The claims about
the update time and post-processing time follow directly from the properties of a state-of-the-art sparse
recovery scheme, e.g., the scheme based on expander matching pursuit given in Theorem 9 of Gilbert and
(cid:3)
Indyk [GI10].
Simplification for Insertion-Only Streams. Algorithm 1 can be simplified considerably if the input stream
is insertion-only. We can then initialize each Gi to an empty graph and, upon seeing an edge {u, v} in the
stream, insert it into Gi iff ψ(u) = ψ(v) = i. We abort if we collect more than s edges; w.h.p., this will not
happen, thanks to Theorem 3.1. Finally, we color the collected graphs Gi greedily, just as in Algorithm 1.
With this simplification, the overall space usage drops to O(s log n) = O(ε−2n log2 n) bits.
The reason this does not work for dynamic graph streams is that the number of monochromatic edges
could exceed s by an arbitrary amount mid-stream.
4.2 Query Model
We now turn to the general graph query model, a standard model of space-conscious algorithms for big
graphs where the input graph is random-accessible but the emphasis is on the examining only a tiny (ideally,
sublinear) portion of it; for general background see Chapter 10 of Goldreich's book [Gol17]. In this model,
the algorithm starts out knowing the vertex set [n] of the input graph G and can access G only through the
following types of queries.
• A pair query Pair({u, v}), where u, v ∈ [n]. The query returns 1 if {u, v} ∈ E(G) and 0 otherwise. For
better readability, we shall write this query as Pair(u, v).
11
• A neighbor query Neighbor(u, j), where u ∈ [n] and j ∈ [n − 1]. The query returns v ∈ [n] where v is
the jth neighbor of u in some underlying fixed ordering of vertex adjacency lists; if deg(v) < j, so that
there does not exist a jth neighbor, the query returns ⊥.
Naturally, when solving a problem in this model, the goal is to do so while minimizing the number of queries.
By adapting the combinatorial machinery from their semi streaming algorithm, Assadi et al. [ACK19]
gave an (cid:101)O(n3/2)-query algorithm for finding a (∆ + 1)-coloring. Our LDP framework gives a considerably
simpler algorithm using κ + o(κ) colors, where κ := κ(G). We remark here that (cid:101)O(n3/2) query complexity
is essentially optimal, as Assadi et al. [ACK19] proved a matching lower bound for any (c · ∆)-coloring
algorithm, for any constant c > 1.
Theorem 4.2. Given query access to a graph G, there is a randomized algorithm that, with high probability,
produces a proper coloring of G using κ + o(κ) colors. The algorithm's worst-case query complexity, running
time, and space usage are all (cid:101)O(n3/2).
Proof. The algorithm proceeds in two stages. In the first stage, it attempts to extract all edges in G through
neighbor queries alone, aborting when "too many" queries have been made. More precisely, it loops over
all vertices v and, for each v, issues queries Neighbor(v, 1), Neighbor(v, 2), . . . until a query returns ⊥. If
√
this stage ends up making 3n3/2 queries (say) without having processed every vertex, then it aborts and
the algorithm moves on to the second stage. By Fact 2.1, if κ (cid:54)
n, then this stage will not abort and the
algorithm will have obtained G completely; it can then (κ + 1)-color G (as in Lemma 2.6) and terminate,
skipping the second stage.
√
√
n. The algorithm now uses a random coloring ψ to construct
n, with s = Θ(ε−2n log n) and (cid:96), λ given by Equation (1).
an ((cid:96), s, λ)-LDP of G using the "guess" k =
To produce each subgraph Gi in the LDP, the algorithm simply makes all possible queries Pair(u, v) where
ψ(u) = ψ(v). W.h.p., the number of queries made is at most
In the second stage, we know that κ >
(cid:88)
i∈[(cid:96)]
1
2
(cid:32)2n
(cid:33)2
(cid:96)
V(Gi)2 (cid:54) (cid:96)
2
(cid:54) 2n2s
4nk
= Θ
(cid:32)n3/2 log n
(cid:33)
,
ε2
where the first inequality uses Item (ii) of Theorem 3.1. We can enforce this bound in the worst case by
aborting if it is violated.
Clearly, k (cid:54) 2κ, so Item (i) of Theorem 3.1 applies and by the discussion after Definition 3.2, the
algorithm uses (1 + O(ε))κ colors. Setting ε = 1/ log n, this number is at most κ + o(κ) and the overall number
(cid:3)
of queries remains (cid:101)O(n3/2), as required.
4.3 MPC and Congested Clique Models
In the Massively Parallel Communication (MPC) model of Beame et al. [BKS17], an input of size m is
distributed adversarially among p processors, each of which has S bits of working memory: here, p and S are
o(m) and, ideally, p ≈ m/S . Computation proceeds in synchronous rounds: in each round, a processor carries
out some local computation (of arbitrary time complexity) and then communicates with as many of the other
When the input is an n-vertex graph, the most natural and widely studied setting of MPC is S = (cid:101)O(n),
processors as desired, provided that each processor sends and receives no more than S bits per round. The
primary goal in solving a problem is to minimize the number of rounds.
which enables each processor to hold some information about every vertex; this makes many graph problems
tractable. Since the input size m is potentially Ω(n2), it is reasonable to allow p = n many processors. Note
that the input is just a collection of edges, distributed adversarially among these processors, subject to the
memory constraint.
12
Theorem 4.3. There is a randomized O(1)-round MPC algorithm that, given an n-vertex graph G, outputs a
(κ + o(κ))-coloring of G with high probability. The algorithm uses n processors, each with O(n log2 n) bits of
memory.
Proof. Our algorithm will use n processors, each assigned to one vertex. If E(G) = O(n log n), then all of G
can be collected at one processor in a single round using E(G)· 2(cid:100)log n(cid:101) = O(n log2 n) bits of communication
and the problem is solved trivially. Therefore, we may as well assume that E(G) = ω(n log n), which implies
κ = ω(log n), by Fact 2.1. We shall first give an algorithm assuming that κ is known a priori. Our final
algorithm will be a refinement of this preliminary one.
Preliminary algorithm. Take k = κ. Each processor chooses a random color for its vertex, implicitly
producing a partition (G1, . . . , G(cid:96)) that is, w.h.p., an ((cid:96), s, λ)-LDP; we take (cid:96), λ as in eq. (1), s = Θ(ε−2n log n),
and ε = (k−1 log n)1/4. Note that ε = o(1). In Round 1, each processor sends its chosen color to all others --
this is O(n log n) bits of communication per machine -- and as a result every processor learns which of its
vertex's incident edges are monochromatic. Now each color i ∈ [(cid:96)] is assigned a unique machine Mi and, in
Round 2, all edges in Gi are sent to Mi. Each Mi then locally computes a (κ(Gi) + 1)-coloring of Gi using a
palette disjoint from those of other Mis; by the discussion following Definition 3.2, this colors G using at
most (1 + O(ε))κ = κ + o(κ) colors.
The communication in Round 2 is bounded by maxi E(Gi) · 2(cid:100)log n(cid:101). By Fact 2.1, items (i) and (ii) of
Theorem 3.1, and eq. (1), the following holds w.h.p. for each i ∈ [(cid:96)]:
E(Gi) (cid:54) κ(Gi)V(Gi) (cid:54) κ + λ
(cid:96)
2n
(cid:96)
(cid:54) 4nκ
(cid:96)2
(cid:54)
4nk
(2nk/s)2
=
O(ε−2n log n)2
nk
= O
ε4k
(cid:32)n log2 n
(cid:33)
= O(n log n) .
(4)
Thus, the communication per processor in Round 2 is O(n log2 n) bits.
Final algorithm. When we don't know κ in advance, we can make geometrically spaced guesses k, as in
Section 4.1. In Round 1, we choose a random coloring for each such k. In Round 2, we determine the
quantities E(Gi) for each k and each subgraph Gi and thereby determine the smallest k such that eq. (4)
holds for every Gi corresponding to this k. We then run Round 3 for only this one k, replicating the logic of
Round 2 of the preliminary algorithm.
Correctness is immediate. We turn to bounding the communication cost. For Round 3, the previous
analysis shows that the communication per processor is O(n log2 n) bits. For Rounds 1 and 2, let us consider
the communication involved for each guess k: since each randomly-chosen color and each cardinality E(Gi)
can be described in O(log n) bits, each processor sends and receives at most O(n log n) bits per guess. This is
(cid:3)
a total of O(n log2 n) bits, as claimed.
The CONGESTED-CLIQUE model [LPPP05] is a well established model of distributed computing for
graph problems. In this model, there are n nodes, each of which holds the local neighborhood information (i.e.,
the incident edges) of one vertex of the input graph G. In each round, every pair of nodes may communicate,
whether or not they are adjacent in G, but the communication is restricted to O(log n) bits. There is no
constraint on a node's local memory. The goal is to minimize the number of rounds.
Behnezhad et al. [BDH18] built on results of Lenzen [Len13] to show that any algorithm in the semi-MPC
model -- defined as MPC with space per machine being O(n log n) bits -- can be simulated in the Congested
Clique model, preserving the round complexity up to a constant factor. Based on this, we obtain the following
result.
Theorem 4.4. There is a randomized O(1)-round algorithm in the Congested Clique model that, given a
graph G, w.h.p. finds a (κ + O(κ3/4 log1/2 n))-coloring. For κ = ω(log2 n), this gives a (κ + o(κ))-coloring. (cid:3)
13
Proof. We cannot directly use our algorithm in Theorem 4.3 because it is not a semi-MPC algorithm: it uses
O(n log2 n) bits of space per processor, rather than O(n log n). However, with a more efficient implementation
of Round 1, a more careful analysis of Round 2, and a slight tweak of parameters for Round 3, we can
improve the communication (hence, space) bounds to O(n log n), whereupon the theorem of Behnezhad et
al. [BDH18] completes the proof.
For Round 3, the tweak is to set ε = (k−1 log2 n)1/4 but otherwise replicate the logic of the final algorithm
from Theorem 4.3. With this higher value of ε, the bound from eq. (4) improves to E(Gi) = O(n). Therefore
the per-processor communication in Round 3 is only O(n log n) bits. The number of colors used is, w.h.p., at
most (1 + O(ε))κ = κ + O(κ3/4 log1/2 n).
For a tighter analysis of the communication cost of Round 2, note that, for a particular guess k, there is
a corresponding (cid:96) given by eq. (1) such that each processor need only send/receive (cid:96) cardinalities E(Gi),
each of which can be described in O(log n) bits. Consulting eq. (1), we see that (cid:96) = O(n2/s) = O(n/ log n).
Therefore, summing over all O(log n) choices of k, each processor communicates at most
O(n/ log n) · O(log n) · O(log n) = O(n log n) bits.
Round 1 appears problematic at first, since there are O(log n) many random colorings to be chosen, one
for each guess k. However, note that these colorings need not be independent. Therefore, we can choose
just one random (cid:100)log n(cid:101)-bit "master color" φ(v) for each vertex v and derive the random colorings for the
various guesses k by using only appropriate length prefixes of φ(v). This ensures that each processor only
(cid:3)
communicates O(n log n) bits in Round 1.
4.4 Distributed Coloring in the LOCAL Model
In the LOCAL model, each node of the input graph G hosts a processor that knows only its own neighborhood.
The processors operate in synchronous rounds, during which they can send and receive messages of arbitrary
length to and from their neighbors. The processors are allowed unbounded local computation in each round.
The key complexity measure is time, defined as the number of rounds used by an algorithm (expected number,
for a randomized algorithm) on a worst-case input.
Graph coloring in the LOCAL model is very heavily studied and is one of the central problems in
distributed algorithms. Here, our focus is on algorithms that properly color the input graph G using a number
of colors that depends on α := α(G), the arboricity of G. Recall that α (cid:54) κ (cid:54) 2α − 1 (Fact 2.2). Unlike
in previous sections, our results will give big-O bounds on the number of colors, so we may as well state
them in terms of α (following established tradition in this line of work) rather than κ. Our focus will be on
algorithms that run in sublogarithmic time, while using not too many colors. See Section 1.2 for a quick
summary of other interesting parameter regimes and Barenboim and Elkin [BE13] for a thorough treatment
of graph coloring in the LOCAL model.
2 log log n (cid:54) k (cid:54) (cid:112)
such a time/quality tradeoff applies: for k ∈(cid:2)ω(
Kothapalli and Pemmaraju [KP11] gave an O(k)-round algorithm that uses O(αn1/k) colors, for all k with
log n. We give a new coloring algorithm that, in particular, extends the range of k to which
log n), o(log n)(cid:3), we can compute an O(αn1/k log n)-coloring
(cid:112)
in O(k) rounds.
Our algorithm uses our LDP framework to split the input graph into parts with logarithmic degeneracy
(hence, arboricity) and then invokes an algorithm of Barenboim and Elkin. The following theorem records
the key properties of their algorithm.
Lemma 4.5 (Thm 5.6 of Barenboim and Elkin [BE10]). There is a deterministic distributed algorithm in
the LOCAL model that, given an n-vertex graph G, an upper bound b on α(G), and a parameter t with
(cid:3)
2 < t (cid:54) O(
√
n/b), produces an O(tb2)-coloring of G in time O
(cid:16)
(cid:17)
.
logt n + log(cid:63) n
Here is the main result of this section.
14
Theorem 4.6. There is a randomized distributed algorithm in the LOCAL model that, given an n-vertex graph
G, an estimate of its arboricity α up to a constant factor, and a parameter t such that 2 < t (cid:54) O(
n/ log n),
produces an O(tα log n)-coloring of G in time O
.
logt n + log(cid:63) n
(cid:16)
(cid:17)
(cid:112)
Proof. To simplify the presentation, we assume that α = α(G). We assume that every node (vertex) knows n
and α. Consider a ((cid:96), s, λ)-LDP of G, where we put s = Cn log n, for some large constant C, as in Theorem 3.1.
This setting of s gives (cid:96) = O(α/ log n). First, each vertex v chooses a color ψ(v) uniformly at random from
[(cid:96)]. Next, we need to effectively "construct" the blocks Gi, for each i ∈ [(cid:96)]. This is straightforwardly done in
a single round: each vertex v sends ψ(v) to all its neighbors.
At this point, each vertex v knows its neighbors in the block Gψ(v). So it's now possible to run a distributed
algorithm on each Gi. We invoke the algorithm in Lemma 4.5. The algorithm needs each vertex v to know an
upper bound bi on α(Gi), where i = ψ(v). A useful upper bound of bi = O(log n), which holds w.h.p., is given
by item (i) of Theorem 3.1.
By Lemma 4.5, each Gi can be colored using O(t log2 n) colors, within another O
since 2 < t (cid:54) O(
G is at most (cid:96) · O(t log2 n) = O(tα log n), as required.
rounds,
n/ log n). Using disjoint palettes for the distinct blocks, the total number of colors used for
(cid:3)
logt n + log(cid:63) n
(cid:112)
(cid:16)
(cid:17)
The particular form of the tradeoff stated in Table 1 is obtained by setting t = n1/k (for some k (cid:62) 3) in the
above theorem.
Corollary 4.7. There is a randomized LOCAL algorithm that, given graph G, estimate α ≈ α(G), and a
parameter k with 2 < n1/k (cid:54) O(
. (cid:3)
n/ log n), finds an O(αn1/k log n)-coloring of G in time O
k + log(cid:63) n
(cid:112)
(cid:17)
(cid:16)
5 Lower Bounds
Can we improve the guarantees of our algorithms so that they use at most κ + 1 colors, rather than κ + o(κ)?
After all, every graph G does have a proper (κ(G) + 1)-coloring. The main message of this section is that
answer is a strong "No," at least in the data streaming and query models. If we insist on a coloring that good,
we would incur the worst possible space or query complexity: Ω(n2). In fact, this holds even if κ is known to
the algorithm in advance. Moreover, all our streaming lower bounds hold even if the input stream consists of
edge insertions alone.
Our lower bounds generalize to the problem of producing a (κ + λ)-coloring. We show that this requires
Ω(n2/λ2) space or query complexity. Such generalizations are based on the following Blow-Up Lemma.
Definition 5.1. Let G be a graph and λ a positive integer. The blow-up graph Gλ is obtained by replacing
each vertex of G with a copy of the complete graph Kλ and replacing each edge of G with a complete bipartite
graph between the copies of Kλ at its endpoints. More succinctly, Gλ is the lexicographical product G[Kλ].
Lemma 5.2 (Blow-Up Lemma). For all graphs G and positive integers λ, c, if G has a c-clique, then Gλ has
a (cλ)-clique. Also, κ(Gλ) (cid:54) (κ(G) + 1)λ − 1.
Proof. The claim about cliques is immediate. The bound on κ(Gλ) follows by taking a degeneracy ordering
of G and replacing each vertex v by a list of vertices of the clique that replaces v in Gλ, ordering vertices
(cid:3)
within the clique arbitrarily.
Our lower bounds come in two flavors. The first address the hardness of distinguishing low-degeneracy
graphs from high-chromatic-number graphs. This is encapsulated in the following abstract problem.
15
(a)
(b)
(c)
Figure 1: Gadget graphs used in (a) Lemma 5.5; (b) Theorem 5.7; (c) Lemma 5.12 and Theorem 5.14.
Definition 5.3 (graph-dist problem). Consider two graph families: G1 := G1(n, q, λ), consisting of n-vertex
graphs with chromatic number χ (cid:62) (q + 1)λ, and G2 := G2(n, q, λ), consisting of n-vertex graphs with
κ (cid:54) qλ − 1. Then graph-dist(n, q, λ) is the problem of distinguishing G1 from G2; note that G1 ∩ G2 = ∅.
More precisely, given an input graph G on n vertices, the problem is to decide whether G ∈ G1 or G ∈ G2,
with success probability at least 2/3.
We shall prove that graph-dist is "hard" in the insertion-only streaming setting and in the query setting,
thereby establishing that in these models it is hard to produce a (κ + λ)-coloring. In fact, our proofs will show
that it is just as hard to estimate the parameter κ; this goes to show that the hardness of the coloring problem
is not just because of the large output size.
Lower bounds of the above flavor raise the following question: since estimating κ itself is hard, does the
coloring problem become easier if the value of κ(G) is given in advance, before the algorithm starts to read
G? In fact, the (∆ + 1)-coloring algorithms by Assadi et al. [ACK19] assume that ∆ is known in advance.
However, perhaps surprisingly, we prove a second flavor of lower bounds, showing that a priori knowledge
of κ does not help and (κ + 1)-coloring (more generally, (κ + λ)-coloring) remains a hard problem even under
the strong assumption that κ is known in advance.
5.1 Streaming Lower Bounds
In this section, we prove both flavors of lower bounds in the one-pass streaming setting. The next section
takes up the query model.
Our streaming lower bounds use reductions from the index and int-find (intersection finding, a variant
of disjointness) problems in communication complexity. In the indexN problem, Alice is given a vector
x = (x1, . . . , xN) ∈ {0, 1}N and Bob is given an index k ∈ [N]. The goal is for Alice to send Bob a (possibly
random) c-bit message that enables Bob to output xk with probability at least 2/3. The smallest c for
which such a protocol exists is called the one-way randomized communication complexity, R→(indexN). In
int-findN, Alice and Bob hold vectors x, y ∈ {0, 1}N, interpreted as subsets of [N], satisfying the promise that
x∩y = 1. They must find the unique index i where xi = yi = 1, using at most c bits of randomized interactive
communication, succeeding with probability at least 2/3. The smallest c for which such a protocol exists is
the randomized communication complexity, R(int-findN). As is well known, R→(indexN) = Ω(N) [Abl96]
and R(int-findN) = Ω(N); the latter is a simple extension of the disjointness lower bound [Raz92].
16
LRlrCyzlLRLRCrlraiajbbjiAB We shall in fact consider instances of indexN where N = p2, for an integer p. Using a canonical
bijection between [N] and [p] × [p], we reinterpret x as a matrix with entries (xi j)i, j∈[p], and Bob's input as
(y, z) ∈ [p]× [p]. We further interpret this matrix x as the bipartite adjacency matrix of a (2p)-vertex balanced
bipartite graph Hx. Such graphs Hx will be key gadgets in the reductions to follow.
Definition 5.4. For x ∈ {0, 1}p×p, a realization of Hx on a list ((cid:96)1, . . . , (cid:96)p, r1, . . . , rp) of distinct vertices is a
graph on these vertices whose edge set is {{(cid:96)i, r j} : xi j = 1}.
First Flavor: Degeneracy Not Known in Advance. To prove lower bounds of the first flavor, we start by
demonstrating the hardness of the abstract problem graph-dist, from Definition 5.3.
Lemma 5.5. Solving graph-dist(n, q, λ) in one randomized streaming pass requires Ω(n2/λ2) space.
More precisely, there is a constant c > 0 such that for every integer λ (cid:62) 1 and every sufficiently large
integer q, there is a setting n = n(q, λ) for which every randomized one-pass streaming algorithm for
graph-dist(n, q, λ) requires at least cn2/λ2 bits of space.
Proof. Put p = q − 1. We reduce from indexN, where N = p2, using the following plan. Starting with an
empty graph on n = 3λp vertices, Alice adds certain edges based on her input x ∈ {0, 1}p×p and then Bob
adds certain other edges based on his input (y, z) ∈ [p] × [p]. By design, solving graph-dist(n, q, λ) on the
resulting final graph reveals the bit xyz, implying that a one-pass streaming algorithm for graph-dist requires
at least R→(indexN) = Ω(N) = Ω(p2) = Ω(n2/λ2) bits of memory. The details follow.
We first consider λ = 1. We use the vertex set L (cid:93) R (cid:93) C (the notation "(cid:93)" denotes a disjoint union),
where L = {(cid:96)1, . . . , (cid:96)p}, R = {r1, . . . , rp}, and C = p. Alice introduces the edges of the gadget graph Hx
(from Definition 5.4), realized on the vertices ((cid:96)1, . . . , (cid:96)p, r1, . . . , rp). Bob introduces all possible edges within
C ∪ {(cid:96)y, rz}, except for {(cid:96)y, rz}. Let G be the resulting graph.
If xyz = 1, then G contains a clique on C ∪ {(cid:96)y, rz}, whence χ(G) (cid:62) p + 2. If, on the other hand, xyz = 0,
then we claim that κ(G) (cid:54) p. By Lemma 2.5, the claim will follow if we exhibit a vertex ordering (cid:67) such that
odegG,(cid:67)(v) (cid:54) p for all v ∈ V(G). We use an ordering where
L ∪ R \ {(cid:96)y, rz} (cid:67) (cid:96)y (cid:67) {rz} ∪ C
and the ordering within each set is arbitrary. By construction of Hx, each vertex in L ∪ R \ {(cid:96)y, rz} has total
degree at most p. For each vertex v ∈ {rz} ∪ C, we trivially have odegG,(cid:67)(v) (cid:54) p because C = p. Finally,
since xyz = 0, the vertex rz is not a neighbor of (cid:96)y; so odegG,(cid:67)((cid:96)y) = C = p. This proves the claim.
When λ (cid:62) 1, Alice and Bob introduce edges so as to create the blow-up graph Gλ, as in Definition 5.1.
By Lemma 5.2, if xyz = 1, then Gλ has a (p + 2)λ-clique, whereas if xyz = 0, then κ(Gλ) (cid:54) (p + 1)λ − 1. In
the former case, χ(Gλ) (cid:62) (p + 2)λ = (q + 1)λ, so that Gλ ∈ G1(n, q, λ); cf. Definition 5.3. In the latter case,
κ(Gλ) (cid:54) qλ − 1, so that Gλ ∈ G2(n, q, λ). Thus, solving graph-dist(n, q, λ) on Gλ reveals xyz.
(cid:3)
Our coloring lower bounds are straightforward consequences of the above lemma.
Theorem 5.6. Given a single randomized pass over a stream of edges of an n-vertex graph G, succeeding
with probability at least 2/3 at either of the following tasks requires Ω(n2/λ2) space, where λ (cid:62) 1 is an
integer parameter:
(i) produce a proper (κ + λ)-coloring of G;
(ii) produce an estimate κ such that κ − κ (cid:54) λ.
Furthermore, if we require λ = O(cid:0)κ
2−γ(cid:1), where γ > 0, then neither task admits a semi-streaming algorithm.
1
17
Proof. An algorithm for either task (i) and or task (ii) immediately solves graph-dist with appropriate
parameters, implying the Ω(n2/λ2) bounds, thanks to Lemma 5.5. For the "furthermore" statement, note that
with the stated guarantee on λ would require Ω(n1+2γ) space, which is not in (cid:101)O(n).
the graphs in the family G2 constructed in the proof of Lemma 5.5 have κ = Θ(n), so performing either task
(cid:3)
coloring is possible in semi-streaming space whereas producing a (κ + O(cid:0)κ
2−γ(cid:1))-coloring is not. We leave
Combining the above result with the algorithmic result in Theorem 4.1, we see that producing a (κ + o(κ))-
1
open the question of whether this gap can be tightened.
Second Flavor: Degeneracy Known in Advance. We now show that the coloring problem remains just as
hard even if the algorithm knows the degeneracy of the graph before seeing the edge stream.
Theorem 5.7. Given as input an integer κ, followed by a stream of edges of an n-vertex graph G with
degeneracy κ, a randomized one-pass algorithm that produces a proper (κ + λ)-coloring of G requires
Ω(n2/λ2) bits of space. Furthermore, if we require λ = O(cid:0)κ
2−γ(cid:1), where γ > 0, then the task does not admit a
1
semi-streaming algorithm.
Proof. We reduce from indexN, where N = p2, using a plan analogous to the one used in proving Lemma 5.5.
Alice and Bob will construct a graph on n = 5λp vertices, using their respective inputs x ∈ {0, 1}p×p and
(y, z) ∈ [p] × [p].
First, we consider the case λ = 1. We use the vertex set L (cid:93) R (cid:93) L (cid:93) R (cid:93) C, where L = {(cid:96)1, . . . (cid:96)p},
R = {r1, . . . , rp}, L = {(cid:96)1, . . . , (cid:96)p}, R = {r1, . . . , rp}, and C = p. Let x be the bitwise complement of x. Alice
introduces the edges of the gadget graph Hx (from Definition 5.4), realized on L ∪ R, and the edges of Hx
realized on L ∪ R. For ease of notation, put (cid:96) := (cid:96)y, r := rz, (cid:96) := (cid:96)y, r := rz, and S := C ∪ {(cid:96), r, (cid:96), r}. Bob
introduces all possible edges within S , except for {(cid:96), r} and {(cid:96), r}. Let G be the resulting graph.
We claim that the degeneracy κ(G) = p + 2. To prove this, we consider the case xyz = 1 (the other case,
xyz = 0, is symmetric). By construction, G contains a clique on the p + 3 vertices in C ∪ {(cid:96), r, (cid:96)}; therefore, by
definition of degeneracy, κ(G) (cid:62) p + 2. To show that κ(G) (cid:54) p + 2, it will suffice to exhibit a vertex ordering
(cid:67) such that odegG,(cid:67)(v) (cid:54) p + 2 for all v ∈ V(G). To this end, consider an ordering where
V(G) \ S (cid:67) (cid:96) (cid:67) S \ {(cid:96)}
and the ordering within each set is arbitrary. Each vertex v ∈ V(G) \ S has odegG,(cid:67)(v) (cid:54) deg(v) (cid:54) p and each
vertex v ∈ S \ {(cid:96)} has odegG,(cid:67)(v) (cid:54)(cid:12)(cid:12)(cid:12)S \ {(cid:96)}(cid:12)(cid:12)(cid:12) − 1 = p + 2. As for the vertex (cid:96), since xyz = 1 − xyz = 0, by the
construction in Definition 5.4, r is not a neighbor of (cid:96); therefore, odegG,(cid:67)((cid:96)) (cid:54)(cid:12)(cid:12)(cid:12)S \ {(cid:96), r}(cid:12)(cid:12)(cid:12) = p + 2.
Let A be a streaming algorithm that behaves as in the theorem statement. Recall that we are considering
λ = 1. Since κ(G) = p + 2 for every instance of indexN, Alice and Bob can simulate A on their constructed
graph G by first feeding it the number p + 2, then Alice's edges, and then Bob's. When A succeeds, the
coloring it outputs is a proper (p + 3)-coloring; therefore it must repeat a color inside S , as S = p + 4. But S
has exactly one pair of non-adjacent vertices: the pair {(cid:96), r} if xyz = 0, and the pair {(cid:96), r} if xyz = 1. Thus, an
examination of which two vertices in S receive the same color reveals xyz, solving the indexN instance. It
follows that A must use at least R→(indexN) = Ω(N) = Ω(p2) bits of space.
Now consider an arbitrary λ. Alice and Bob proceed as above, except that they simulate A on the blow-up
graph Gλ. Since G always has a (p + 3)-clique and κ(G) = p + 2, the two halves of Lemma 5.2 together imply
κ(Gλ) = (p + 3)λ − 1. So, when A succeeds, it properly colors Gλ using at most (p + 4)λ − 1 colors. For each
A ⊆ V(G), abusing notation, let Aλ denote its corresponding set of vertices in Gλ (cf. Definition 5.1). Since
S λ = (p + 4)λ, there must be a color repetition within S λ. Reasoning as above, this repetition must occur
within {(cid:96), r}λ when xyz = 0 and within {(cid:96), r}λ when xyz = 1. Therefore, Bob can examine the coloring to solve
indexN, showing that A must use Ω(N) = Ω(p2) = Ω(n2/λ2) space.
The "furthermore" part follows by observing that κ(Gλ) = Θ(cid:0)V(Gλ)(cid:1).
(cid:3)
18
Multiple Passes. The streaming algorithm from Section 4.1 is one-pass, as are the lower bounds proved
above. Is the coloring problem any easier if we are allowed multiple passes over the edge stream? We now
give a simple argument showing that, if we slightly generalize the problem, it stays just as hard using multiple
(O(1) many) passes.
The generalization is to allow some edges to be repeated in the stream. In other words, the input is a
multigraph G. Clearly, a coloring is proper for G iff it is proper for the underlying simple graph G, so the
relevant algorithmic problem is to properly (κ + λ)-color G, where κ := κ(G). Note that our algorithm in
Section 4.1 does, in fact, solve this more general problem.
Theorem 5.8. Given as input an integer κ, followed by a stream of edges of an n-vertex multigraph G
whose underlying simple graph has degeneracy κ, a randomized p-pass algorithm that produces a proper
(κ + λ)-coloring of G requires Ω(n2/(λ2 p)) bits of space. This holds even if the stream is insertion-only, with
each edge appearing at most twice.
Proof. As usual, we prove this for λ = 1 and appeal to the Blow-Up Lemma (Lemma 5.2) to generalize.
We reduce from int-findN, with N =
. Let Alice and Bob treat their inputs as (xi j)1(cid:54)i< j(cid:54)n and (yi j)1(cid:54)i< j(cid:54)n
in some canonical way. Alice (resp. Bob) converts their input into an edge stream consisting of pairs (i, j)
such that i < j and xi j = 0 (resp. yi j = 0). The concatenation of these streams defines the multigraph G given
to the coloring algorithm. Let (h, k) be the unique pair such that xhk = yhk = 1. Note that the underlying
simple graph G is Kn minus the edge {h, k}. Therefore, κ = n − 2 and so, in a proper (n − 1)-coloring of G,
there must be a repeated color and this can only happen at vertices h and k.
(2p − 1)s bits of communication. Therefore, s = Ω(N/p) = Ω(n2/p).
Thus, a p-pass (κ + 1)-coloring algorithm using s bits of space leads to a protocol for int-findN using
(cid:3)
(cid:17)
(cid:16)n
2
5.2 Query Complexity Lower Bounds
We now turn to the general graph query model [Gol17]. Recall that our algorithm from Section 4.2 produces
a (κ + o(κ))-coloring while making at most (cid:101)O(n3/2) queries, without needing to know κ in advance. Here, we
shall prove that the number of colors cannot be improved to κ + 1: that would preclude sublinear complexity.
In fact, we prove more general results, similar in spirit to the streaming lower bounds from the previous
section. For these lower bounds, we use another family of gadget graphs.
Definition 5.9. Given a large integer p (a size parameter), the gadgets for that size are (2p + 1)-vertex graphs
on vertex set A (cid:93) B, where A = {a1, . . . , ap+1} and B = {b1, . . . , bp}. Let H be the graph consisting of a clique
on A and a clique on B, with no edges between A and B. For 1 (cid:54) i < j (cid:54) p, let Hi j be a graph on the same
vertex set obtained by slightly modifying H as follows (see Figure 1c):
E(Hi j) = E(H) \(cid:8){ai, a j},{bi, b j}(cid:9) ∪(cid:8){ai, b j},{a j, bi}(cid:9) .
(5)
Notice that the vertex ap+1 is not touched by any of these modifications. The relevant properties of these
gadget graphs are as follows.
Lemma 5.10. For all 1 (cid:54) i < j (cid:54) p, κ(Hi j) = p − 1, whereas the chromatic number χ(H) = p + 1.
Proof. The claim about χ(H) is immediate.
Consider a particular graph Hi j. The subgraph induced by A \ {ai} is a p-clique, so κ(Hi j) (cid:62) p − 1.
Now consider the following ordering (cid:67) for Hi j: B (cid:67) ai (cid:67) A \ {ai}, where the order within each set is
arbitrary. For each v ∈ B, odegHi j,(cid:67)(v) (cid:54) deg(v) = p−1. For each v ∈ A\{ai}, odegHi j,(cid:67)(v) (cid:54) A\{ai}−1 = p−1.
Finally, ai has exactly p−1 neighbors in A\{ai} (by construction, a j is not a neighbor), so odegHi j,(cid:67)(ai) = p−1.
By Lemma 2.5, it follows that κ(Hi j) (cid:54) p − 1.
(cid:3)
19
Our proofs will use these gadget graphs in reductions from a pair of basic problems in decision tree
complexity. Consider inputs that are vectors in {0, 1}N: let 0 denote the all-zero vector and, for i ∈ [N], let ei
denote the vector whose ith entry is 1 while all other entries are 0. Let unique-orN and unique-findN denote
the following partial functions on {0, 1}N:
unique-orN(x) =
if x = 0 ,
if x = ei , for i ∈ [N] ,
1 ,
(cid:63) , otherwise;
unique-findN(x) =
if x = ei , for i ∈ [N] ,
(cid:63) , otherwise.
0 ,
i ,
Informally, these problems capture, respectively, the tasks of (a) determining whether there is a needle in
a haystack under the promise that there is at most one needle, and (b) finding a needle in a haystack under
the promise that there is exactly one needle. Intuitively, solving either of these problems with high accuracy
should require searching almost the entire haystack. Formally, let Rdt
δ ( f ) denote the δ-error randomized query
complexity (a.k.a. decision tree complexity) of f . Elementary considerations of decision tree complexity lead
to the bounds below (for a thorough discussion, including formal definitions, we refer the reader to the survey
by Buhrman and de Wolf [BdW02]).
Fact 5.11. For all δ ∈ (0, 1
δ (unique-findN) (cid:62) (1− δ)N−1. (cid:3)
δ (unique-orN) (cid:62) (1−2δ)N and Rdt
2), we have Rdt
With this setup, we turn to lower bounds of the first flavor.
Lemma 5.12. Solving graph-dist(n, p, λ) in the general graph query model requires Ω(n2/λ2) queries.
More precisely, there is a constant c > 0 such that for every integer λ (cid:62) 1 and every sufficiently large
integer p, there is a setting n = n(p, λ) for which every randomized query algorithm for graph-dist(n, p, λ)
requires at least cn2/λ2 queries in the worst case.
2
Proof. We reduce from unique-orN, where N =
, using the following plan. Put n = (2p + 1)λ. Let
C be a query algorithm for graph-dist(n, p, λ). Based on C, we shall design a 1
3-error algorithm A for
unique-orN that makes at most as many queries as C. By Fact 5.11, this number of queries must be at least
N/3 = Ω(p2) = Ω(n2/λ2).
As usual, we detail our reduction for λ = 1; the Blow-up Lemma (Lemma 5.2) then handles general λ.
By Lemma 5.10, H ∈ G1 whereas each Hi j ∈ G2 (cf. Definition 5.3, taking q = p).
We now design A. Let x ∈ {0, 1}N be the input to A. Using a canonical bijection, let us index the bits of
x as xi j, where 1 (cid:54) i < j (cid:54) p. Algorithm A simulates C and outputs 1 iff C decides that its input lies in G2.
Since C makes queries to a graph, we shall design an oracle for C whose answers, based on query answers for
input x to A, will implicitly define a graph on vertex set V := A (cid:93) B, as in Definition 5.9. The oracle answers
queries as follows.
• For i, j ∈ [p], it answers Pair(ai, a j) and Pair(bi, b j) with 1 − xi j.
• For i, j ∈ [p], it answers Pair(ai, b j) and Pair(a j, bi) with xi j.
• For i ∈ [p], it answers Pair(ap+1, ai) with 1 and Pair(ap+1, bi) with 0.
• For i ∈ [p] and d ∈ [p − 1], it answers Neighbor(ai, d) with a j if xi j = 0 and b j if xi j = 1, where j = d
if d < i, and j = d + 1 otherwise.
• For i, d ∈ [p], it answers Neighbor(ai, p) with ap+1 and Neighbor(ap+1, d) with ad.
• For i ∈ [p] and d ∈ [p − 1], it answers Neighbor(bi, d) with b j if xi j = 0 and a j if xi j = 1, where j = d
if d < i, and j = d + 1 otherwise.
• For all other combinations of v ∈ V and d ∈ N, it answers Neighbor(v, d) = ⊥.
(cid:17)
(cid:16)p
20
By inspection, we see that the graph defined by this oracle is H if x = 0 and is Hi j if x = ei j. Furthermore,
the oracle answers each query by making at most one query to the input x. It follows that A makes at most as
many queries as C and decides unique-orN with error at most 1
3. This completes the proof for λ = 1.
To handle λ > 1, we modify the oracle in the natural way so that the implicitly defined graph is Hλ when
(cid:3)
i j when x = ei j. We omit the details, which are routine.
x = 0 and Hλ
As an immediate consequence of Lemma 5.12, we get the following query lower bounds.
Theorem 5.13. Given query access to an n-vertex graph G, succeeding with probability at least 2/3 at either
of the following tasks requires Ω(n2/λ2) queries, where λ (cid:62) 1 is an integer parameter:
(i) produce a proper (κ + λ)-coloring of G;
(ii) produce an estimate κ such that κ − κ (cid:54) λ.
We now prove a lower bound of the second flavor, where the algorithm knows κ in advance.
(cid:3)
2
Theorem 5.14. Given an integer κ and query access to an n-vertex graph G with κ(G) = κ, an algorithm
that, with probability 2
3, produces a proper (κ + λ)-coloring of G must make Ω(n2/λ2) queries.
Proof. We focus on the case λ = 1; the general case is handled by the Blow-up Lemma, as usual.
(cid:16)p
(cid:17)
Let C be an algorithm for the coloring problem. We design an algorithm A for unique-findN, where
, using the same reduction as in Lemma 5.12, changing the post-processing logic as follows: A
N =
outputs (i, j) as its answer to unique-findN(x), where 1 (cid:54) i < j (cid:54) p is such that ai and a j are colored the same
by C.
To prove the correctness of this reduction, note that when x = ei j, the graph defined by the simulated
oracle is Hi j and κ(Hi j) = p − 1 (Lemma 5.10). Suppose that C is successful, which happens with probability
3. Then C properly p-colors Hi j. Recall that V(Hi j) = A (cid:93) B, where A = p + 1; there must therefore
at least 2
be a color repetition within A. The only two non-adjacent vertices inside A are ai and a j, so A correctly
answers (i, j). By Fact 5.11, A must make Ω(N) = Ω(p2) queries.
(cid:3)
5.3 A Combinatorial Lower Bound
Finally, we explore a connection between degeneracy based coloring and the list coloring problem. In the
latter problem, each vertex has a list of colors and the goal is to find a corresponding list coloring -- i.e., a
proper coloring of the graph where each vertex receives a color from its list -- or to report that none exists.
Assadi et al. [ACK19] proved a beautiful Palette Sparsification Theorem, a purely graph-theoretic result that
connects the (∆ + 1)-coloring problem to the list coloring problem.
Define a graph G to be [(cid:96), r]δ-randomly list colorable (briefly, [(cid:96), r]δ-RLC) if choosing r random colors
per vertex, independently and uniformly without replacement from the palette [(cid:96)], permits a list coloring with
probability at least 1 − δ using these chosen lists.2 Their theorem can be paraphrased as follows.
Fact 5.15 (Assadi et al. [ACK19], Theorem 1). There exists a constant c such that every n-vertex graph G is
(cid:3)
[∆(G) + 1, c log n]1/n-RLC.
Indeed, this theorem is the basis of the various coloring results in their work. Let us outline how things
work in the streaming model, focusing on the space usage. Given an input graph G that is promised to
be [(cid:96), r]1/3-RLC, for some parameters (cid:96), r that may depend on G, we sample r random colors from [(cid:96)] for
each vertex before reading the input. Chernoff bounds imply that the conflict graph -- the subgraph of G
consisting only of edges between vertices whose color lists intersect -- is of size O(E(G)r2/(cid:96)), w.h.p.. Using
2When r (cid:62) l, this procedure simply produces the list [(cid:96)] for every vertex.
21
E(G) (cid:54) n∆/2, taking (cid:96) = ∆ + 1 and r = O(log n) bounds this size by (cid:101)O(n), so a semi-streaming space bound
suffices to collect the entire conflict graph. (For full details, see Lemma 4.1 in [ACK19].) Finding a list
space usage to be (cid:101)O(n), we need r = O(polylog n). This raises the following combinatorial question: what is
coloring of the conflict graph (which exists with probability at least 2/3) yields an (cid:96)-coloring of G.
For a similar technique to work in our setting, we would want (cid:96) ≈ κ. Recalling that E(G) (cid:54) nκ, for the
the smallest λ for which we can guarantee that every graph is [κ + λ, O(polylog n)]1/3-RLC?
not possible with λ = O(κ
By the discussion above, our streaming lower bound in Theorem 5.7 already tells us that such a result is
2−γ). Our final result (Theorem 5.17 below) proves that we can say much more.
Let Jn,t denote the graph Kt + Kn−t, i.e., the graph join of a t-clique and an (n − t)-sized independent set.
1
More explicitly,
Jn,t = (A (cid:93) B, E) , where A = t, B = n − t, E = {{u, v} : u ∈ A, v ∈ A ∪ B, u (cid:44) v} .
Lemma 5.16. For integers 0 < r (cid:54) t < n, if Jn,t is [κ + κ/r, r]δ-RLC, then δ (cid:62) 1 − rn/(r + 1)n−t.
Proof. Take a graph Jn,t with vertices partitioned into A and B as in eq. (6). An ordering with B (cid:67) A shows
that κ = κ(Jn,t) = t. We claim that for every choice of colors lists for vertices in A, taken from the palette
[t + t/r], the probability that the chosen lists for B permit a proper list coloring is at most p := rn/(r + 1)n−t.
This will prove that δ (cid:62) 1 − p.
(6)
To prove the claim, consider a particular choice of lists for A. Fix a partial coloring ψ of A consistent
with these lists. If ψ is not proper, there is nothing to prove. Otherwise, since A induces a clique, ψ must
assign t distinct colors to A. In order for a choice of lists for B to permit a proper extension of ψ to the entire
graph, every vertex of B must sample a color from the remaining t/r colors in the palette. Since r colors are
chosen per vertex, this event has probability at most
(cid:32)
r ·
(cid:33)B
(cid:18)
t/r
t + t/r
=
r
r + 1
(cid:19)n−t
.
The claimed upper bound on p now follows by a union bound over the rt possible partial colorings ψ.
(cid:3)
This easily leads to our combinatorial lower bound, given below. In reading the theorem statement, note
that the restriction on edge density strengthens the theorem.
Theorem 5.17. Let n be sufficiently large and let m be such that n (cid:54) m (cid:54) n2/ log2 n. If every n-vertex graph
G with Θ(m) edges is [κ(G) + λ, c log n]1/3-RLC for some parameter λ and some constant c, then we must
have λ > κ(G)/(c log n).
Proof. Suppose not. Put t = (cid:100)m/n(cid:101), r = c log n, and consider the graph Jn,t defined in eq. (6). By the bounds
on m, E(Jn,t) = t(t − 1)/2 + t(n− t) = Θ(nt) = Θ(m). Put κ := κ(Jn,t). By assumption, Jn,t is [κ + κ/r, r]-RLC,
so Lemma 5.16 implies that
(cid:33)n
(cid:32)
1 − 1
r + 1
(cid:54)
2
3
rn
(r + 1)n−t
=
(cid:18)− n
r + 1
(cid:19)
(r + 1)t (cid:54) exp
+ t ln(r + 1)
.
Since t = O(n/ log2 n) and r = c log n, this is a contradiction for sufficiently large n.
(cid:3)
We remark that the above result rules out the possibility of using a palette sparsification theorem along
the lines of Assadi et al. [ACK19] to obtain a semi-streaming coloring algorithm that uses fewer colors than
Algorithm 1 (with the setting ε = 1/ log n).
More generally, suppose we were willing to tolerate a weaker notion of palette sparsification by sampling
O(logd n) colors per vertex, for some d (cid:62) 1: this would increase the space complexity of an algorithm based
22
on such sparsification by a polylog n factor. By Lemma 5.16, arguing as in Theorem 5.17, we would need to
spend at least κ + κ/Θ(logd n) colors. This is no better than the number of colors obtained using Algorithm 1
with the setting ε = 1/ logd n, which still maintains semi-streaming space. In fact, palette sparsification does
not immediately guarantee a post-processing runtime that is better than exponential, because we need to color
the conflict graph in post-processing. Meanwhile, recall that Algorithm 1 has (cid:101)O(n) post-processing time via a
straightforward greedy algorithm. Furthermore, since there exist "hard" graphs Jn,t at all edge densities from
Θ(n) to Θ(n2/ log2 n), we cannot even hope for a semi-streaming palette-sparsification-based algorithm that
might work only for sparse graphs or only for dense graphs.
Acknowledgement
We gratefully acknowledge several helpful discussions we have had with Sepehr Assadi (especially those that
called to our attention a nuance with the Congested Clique algorithm) and Deeparnab Chakrabarty.
References
[ABI86]
[Abl96]
Noga Alon, L´aszl´o Babai, and Alon Itai. A fast and simple randomized parallel algorithm for
the maximal independent set problem. Journal of algorithms, 7(4):567 -- 583, 1986.
Farid Ablayev. Lower bounds for one-way probabilistic communication complexity and their
application to space complexity. Theor. Comput. Sci., 175(2):139 -- 159, 1996.
[ACK19]
Sepehr Assadi, Yu Chen, and Sanjeev Khanna. Sublinear algorithms for (∆+ 1) vertex coloring.
In Proc. 30th Annual ACM-SIAM Symposium on Discrete Algorithms, page To Appear, 2019.
[Bar16]
[BB04]
[BB06]
Leonid Barenboim. Deterministic (∆+ 1)-coloring in sublinear (in ∆) time in static, dynamic,
and faulty networks. Journal of the ACM (JACM), 63(5):47, 2016.
Nicolas Barnier and Pascal Brisset. Graph coloring for air traffic flow management. Annals of
operations research, 130(1-4):163 -- 178, 2004.
Balabhaskar Balasundaram and Sergiy Butenko. Graph domination, coloring and cliques in
telecommunications. In Handbook of Optimization in Telecommunications, pages 865 -- 890.
Springer, 2006.
[BCHN18] Sayan Bhattacharya, Deeparnab Chakrabarty, Monika Henzinger, and Danupon Nanongkai.
In Proc. 39th Annual ACM-SIAM Symposium on
Dynamic algorithms for graph coloring.
Discrete Algorithms, pages 1 -- 20, 2018.
[BCK+17] Luis Barba, Jean Cardinal, Matias Korman, Stefan Langerman, Andr´e van Renssen, Marcel
Roeloffzen, and Sander Verdonschot. Dynamic graph coloring. In Workshop on Algorithms and
Data Structures, pages 97 -- 108, 2017.
[BDH18]
Soheil Behnezhad, Mahsa Derakhshan, and Mohammad Taghi Hajiaghayi. Brief announcement:
Semi-mapreduce meets congested clique. CoRR, abs/1802.10297, 2018.
[BdW02] Harry Buhrman and Ronald de Wolf. Complexity measures and decision tree complexity: a
survey. Theor. Comput. Sci., 288(1):21 -- 43, 2002.
[BE10]
Leonid Barenboim and Michael Elkin. Sublogarithmic distributed mis algorithm for sparse
graphs using nash-williams decomposition. Distributed Computing, 22(5-6):363 -- 379, 2010.
23
[BE11]
[BE13]
Leonid Barenboim and Michael Elkin. Deterministic distributed vertex coloring in polylogarith-
mic time. Journal of the ACM (JACM), 58(5):23, 2011.
Leonid Barenboim and Michael Elkin. Distributed Graph Coloring: Fundamentals and Recent
Developments. Synthesis Lectures on Distributed Computing Theory. Morgan & Claypool
Publishers, 2013.
[BEPS16] Leonid Barenboim, Michael Elkin, Seth Pettie, and Johannes Schneider. The locality of
distributed symmetry breaking. Journal of the ACM (JACM), 63(3):20, 2016.
[BKS17]
Paul Beame, Paraschos Koutris, and Dan Suciu. Communication steps for parallel query
processing. Journal of the ACM (JACM), 64(6):40, 2017.
[BKV12] Bahman Bahmani, Ravi Kumar, and Sergei Vassilvitskii. Densest subgraph in streaming and
mapreduce. International Conference on Very Large Data Bases, 5(5):454 -- 465, 2012.
[CAC+81] Gregory J Chaitin, Marc A Auslander, Ashok K Chandra, John Cocke, Martin E Hopkins, and
Peter W Markstein. Register allocation via coloring. Computer languages, 6(1):47 -- 57, 1981.
[CFG+18] Yi-Jun Chang, Manuela Fischer, Mohsen Ghaffari, Jara Uitto, and Yufan Zheng. Simple
graph coloring algorithms for congested clique and massively parallel computation. CoRR,
abs/1808.08419, 2018.
[CH90]
[Cha82]
[CLP18]
[CRT05]
[DG04]
Fred C Chow and John L Hennessy. The priority-based coloring approach to register allocation.
ACM Transactions on Programming Languages and Systems (TOPLAS), 12(4):501 -- 536, 1990.
Gregory J Chaitin. Register allocation & spilling via graph coloring. In ACM Sigplan Notices,
volume 17, pages 98 -- 105, 1982.
Yi-Jun Chang, Wenzheng Li, and Seth Pettie. An optimal distributed (∆+ 1)-coloring algorithm.
In Proc. 50th Annual ACM Symposium on the Theory of Computing, pages 445 -- 456, 2018.
Bernard Chazelle, Ronitt Rubinfeld, and Luca Trevisan. Approximating the minimum spanning
tree weight in sublinear time. SIAM Journal on computing, 34(6):1370 -- 1379, 2005.
Jeffrey Dean and Sanjay Ghemawat. Mapreduce: Simplified data processing on large clusters. In
6th Symposium on Operating System Design and Implementation (OSDI 2004), San Francisco,
California, USA, December 6-8, 2004, pages 137 -- 150, 2004.
[FHK16]
Pierre Fraigniaud, Marc Heinrich, and Adrian Kosowski. Local conflict coloring. In Proc. 57th
Annual IEEE Symposium on Foundations of Computer Science, pages 625 -- 634, 2016.
[FK96]
Uriel Feige and Joe Kilian. Zero knowledge and the chromatic number.
Conference on Computational Complexity, page 278, 1996.
In Annual IEEE
[FKM+05] Joan Feigenbaum, Sampath Kannan, Andrew McGregor, Siddharth Suri, and Jian Zhang. On
graph problems in a semi-streaming model. Theor. Comput. Sci., 348(2 -- 3):207 -- 216, 2005.
Preliminary version in Proc. 31st International Colloquium on Automata, Languages and
Programming, pages 531 -- 543, 2004.
[GI10]
Anna C. Gilbert and Piotr Indyk. Sparse recovery using sparse matrices. Proceedings of the
IEEE, 98(6):937 -- 947, 2010.
24
[GL17]
Mohsen Ghaffari and Christiana Lymouri. Simple and near-optimal distributed coloring for
sparse graphs. In 31st International Symposium on Distributed Computing (DISC 2017), page 20,
2017.
[Gol17]
Oded Goldreich. Introduction to Property Testing. Cambridge University Press, 2017.
[GR08]
[HLL18]
[HSS16]
[Joh99]
[KP06]
[KP11]
[Lei79]
[Len13]
Oded Goldreich and Dana Ron. Approximating average parameters of graphs. Random
Structures & Algorithms, 32(4):473 -- 493, 2008.
Nicholas J. A. Harvey, Christopher Liaw, and Paul Liu. Greedy and local ratio algorithms in the
mapreduce model. In Proceedings of the 30th on Symposium on Parallelism in Algorithms and
Architectures, SPAA 2018, Vienna, Austria, July 16-18, 2018, pages 43 -- 52, 2018.
David G Harris, Johannes Schneider, and Hsin-Hao Su. Distributed (∆+ 1)-coloring in sublog-
arithmic rounds. In Proc. 48th Annual ACM Symposium on the Theory of Computing, pages
465 -- 478, 2016.
Ojvind Johansson. Simple distributed ∆+ 1-coloring of graphs. Information Processing Letters,
70(5):229 -- 232, 1999.
Subhash Khot and Ashok Kumar Ponnuswami. Better inapproximability results for maxclique,
chromatic number and min-3lin-deletion. In International Colloquium on Automata, Languages
and Programming, pages 226 -- 237, 2006.
Kishore Kothapalli and Sriram Pemmaraju. Distributed graph coloring in a few rounds. In Proc.
30th ACM Symposium on Principles of Distributed Computing, pages 31 -- 40, 2011.
Frank Thomson Leighton. A graph coloring algorithm for large scheduling problems. Journal
of research of the national bureau of standards, 84(6):489 -- 506, 1979.
Christoph Lenzen. Optimal deterministic routing and sorting on the congested clique. In Proc.
32nd ACM Symposium on Principles of Distributed Computing, pages 42 -- 50, 2013.
[LPPP05] Zvi Lotker, Boaz Patt-Shamir, Elan Pavlov, and David Peleg. Minimum-weight spanning tree
construction in O(log log n) communication rounds. SIAM J. Comput., 35(1):120 -- 131, 2005.
[LS86]
[Lub86]
[MOT14]
Vahid Lotfi and Sanjiv Sarin. A graph coloring algorithm for large scale scheduling problems.
Computers & operations research, 13(1):27 -- 32, 1986.
Michael Luby. A simple parallel algorithm for the maximal independent set problem. SIAM J.
Comput., 15(4):1036 -- 1053, 1986.
Farnaz Moradi, Tomas Olovsson, and Philippas Tsigas. A local seed selection algorithm for
overlapping community detection. In 2014 IEEE/ACM International Conference on Advances in
Social Networks Analysis and Mining (ASONAM 2014), pages 1 -- 8, 2014.
[MTVV15] Andrew McGregor, David Tench, Sofya Vorotnikova, and Hoa T Vu. Densest subgraph in
dynamic graph streams. In International Symposium on Mathematical Foundations of Computer
Science, pages 472 -- 482, 2015.
[Par18]
Merav Parter.
Colloquium on Automata, Languages and Programming, pages 160:1 -- 160:14, 2018.
(∆+1) coloring in the congested clique model.
In Proc. 45th International
25
[PL96]
[PR07]
[PS96]
[PS18]
[RA15]
[Raz92]
Taehoon Park and Chae Y Lee. Application of the graph coloring algorithm to the frequency
assignment problem. Journal of the Operations Research society of Japan, 39(2):258 -- 265, 1996.
Michal Parnas and Dana Ron. Approximating the minimum vertex cover in sublinear time and a
connection to distributed algorithms. Theoretical Computer Science, 381(1-3):183 -- 196, 2007.
Alessandro Panconesi and Aravind Srinivasan. On the complexity of distributed network
decomposition. Journal of Algorithms, 20(2):356 -- 374, 1996.
Merav Parter and Hsin-Hao Su. Randomized (Delta+1)-Coloring in O(log* Delta) Congested
Clique Rounds. In Proc. 32nd International Symposium on Distributed Computing, pages
39:1 -- 39:18, 2018.
Ryan A. Rossi and Nesreen K. Ahmed. The network data repository with interactive graph
analytics and visualization. In Proceedings of the Twenty-Ninth AAAI Conference on Artificial
Intelligence, 2015.
Alexander Razborov. On the distributional complexity of disjointness. Theor. Comput. Sci.,
106(2):385 -- 390, 1992. Preliminary version in Proc. 17th International Colloquium on Automata,
Languages and Programming, pages 249 -- 253, 1990.
[RSV15]
Jaikumar Radhakrishnan, Saswata Shannigrahi, and Rakesh Venkat. Hypergraph two-coloring
in the streaming model. arXiv preprint arXiv:1512.04188, 2015.
[SW10]
[TZP18]
[Zuc06]
Johannes Schneider and Roger Wattenhofer. A new technique for distributed symmetry breaking.
In Proc. 29th ACM Symposium on Principles of Distributed Computing, pages 257 -- 266, 2010.
Simon Thevenin, Nicolas Zufferey, and Jean-Yves Potvin. Graph multi-coloring for a job
scheduling application. Discrete Applied Mathematics, 234:218 -- 235, 2018.
David Zuckerman. Linear degree extractors and the inapproximability of max clique and
chromatic number. In Proc. 38th Annual ACM Symposium on the Theory of Computing, pages
681 -- 690, 2006.
26
|
1506.08392 | 1 | 1506 | 2015-06-28T12:56:32 | A Linear-Size Logarithmic Stretch Path-Reporting Distance Oracle for General Graphs | [
"cs.DS"
] | In 2001 Thorup and Zwick devised a distance oracle, which given an $n$-vertex undirected graph and a parameter $k$, has size $O(k n^{1+1/k})$. Upon a query $(u,v)$ their oracle constructs a $(2k-1)$-approximate path $\Pi$ between $u$ and $v$. The query time of the Thorup-Zwick's oracle is $O(k)$, and it was subsequently improved to $O(1)$ by Chechik. A major drawback of the oracle of Thorup and Zwick is that its space is $\Omega(n \cdot \log n)$. Mendel and Naor devised an oracle with space $O(n^{1+1/k})$ and stretch $O(k)$, but their oracle can only report distance estimates and not actual paths. In this paper we devise a path-reporting distance oracle with size $O(n^{1+1/k})$, stretch $O(k)$ and query time $O(n^\epsilon)$, for an arbitrarily small $\epsilon > 0$. In particular, our oracle can provide logarithmic stretch using linear size. Another variant of our oracle has size $O(n \log\log n)$, polylogarithmic stretch, and query time $O(\log\log n)$.
For unweighted graphs we devise a distance oracle with multiplicative stretch $O(1)$, additive stretch $O(\beta(k))$, for a function $\beta(\cdot)$, space $O(n^{1+1/k} \cdot \beta)$, and query time $O(n^\epsilon)$, for an arbitrarily small constant $\epsilon >0$. The tradeoff between multiplicative stretch and size in these oracles is far below girth conjecture threshold (which is stretch $2k-1$ and size $O(n^{1+1/k})$). Breaking the girth conjecture tradeoff is achieved by exhibiting a tradeoff of different nature between additive stretch $\beta(k)$ and size $O(n^{1+1/k})$. A similar type of tradeoff was exhibited by a construction of $(1+\epsilon,\beta)$-spanners due to Elkin and Peleg. However, so far $(1+\epsilon,\beta)$-spanners had no counterpart in the distance oracles' world.
An important novel tool that we develop on the way to these results is a {distance-preserving path-reporting oracle}. | cs.DS | cs |
A Linear-Size Logarithmic Stretch
Path-Reporting Distance Oracle for General Graphs∗
Michael Elkin †
Seth Pettie ‡
July 22, 2018
Abstract
In [33] Thorup and Zwick came up with a landmark distance oracle. Given an n-vertex
undirected graph G = (V, E) and a parameter k = 1, 2, . . ., their oracle has size O(kn1+1/k),
and upon a query (u, v) it constructs a path Π between u and v of length δ(u, v) such that
dG(u, v) ≤ δ(u, v) ≤ (2k − 1)dG(u, v). The query time of the oracle from [33] is O(k) (in
addition to the length of the returned path), and it was subsequently improved to O(1)
[36, 13]. A major drawback of the oracle of [33] is that its space is Ω(n · log n). Mendel and
Naor [23] devised an oracle with space O(n1+1/k) and stretch O(k), but their oracle can only
report distance estimates and not actual paths. In this paper we devise a path-reporting
distance oracle with size O(n1+1/k), stretch O(k) and query time O(nǫ), for an arbitrarily
small ǫ > 0. In particular, for k = log n our oracle provides logarithmic stretch using linear
size. Another variant of our oracle has size O(n log log n), polylogarithmic stretch, and query
time O(log log n).
For unweighted graphs we devise a distance oracle with multiplicative stretch O(1), ad-
ditive stretch O(β(k)), for a function β(·), space O(n1+1/k · β), and query time O(nǫ), for
an arbitrarily small constant ǫ > 0. The tradeoff between multiplicative stretch and size in
these oracles is far below Erdos's girth conjecture threshold (which is stretch 2k − 1 and size
O(n1+1/k)). Breaking the girth conjecture tradeoff is achieved by exhibiting a tradeoff of
different nature between additive stretch β(k) and size O(n1+1/k). A similar type of tradeoff
was exhibited by a construction of (1 + ǫ, β)-spanners due to Elkin and Peleg [18]. However,
so far (1 + ǫ, β)-spanners had no counterpart in the distance oracles' world.
An important novel tool that we develop on the way to these results is a distance-
preserving path-reporting oracle. We believe that this oracle is of independent interest.
∗A preliminary version of this paper was published in SODA'15 [19].
†Department of Computer Science, Ben-Gurion University of the Negev, Beer-Sheva, 84105,
Israel,
[email protected]
This research has been supported by the Israeli Academy of Science, grant 593/11, and by the Binational Science
Foundation, grant 2008390. In addition, this research has been supported by the Lynn and William Frankel Center
for Computer Science. A part of this research was performed while visiting the the Center for Massive Algorithms
(MADALGO), which is supported by Danish National Research Foundation grant DNRF84.
‡Department of Computer Science, University of Michigan, Ann Arbor.
This research has been supported by the Binational Science Foundation, grant 2008390. A part of this research was
performed while visiting the Center for Massive Algorithms (MADALGO), which is supported by Danish National
Research Foundation grant DNRF84.
1
1
Introduction
1.1 Distance Oracles for General Graphs
In the distance oracle problem we wish to preprocess a weighted undirected n-vertex graph G =
(V, E). As a result of this preprocessing we construct a compact data structure (which is called
distance oracle) D(G), which given a query pair (u, v) of vertices will efficiently return a distance
estimate δ(u, v) of the distance dG(u, v) between u and v in G. Moreover, the distance oracle
should also compute an actual path Π(u, v) of length δ(u, v) between these vertices in G. We say
that a distance oracle is path-reporting if it does produce the paths Π(u, v) as above; otherwise we
say that it is not path-reporting.
The most important parameters of a distance oracle are its stretch, its size, and its worst-case
query time.1 The stretch α of a distance oracle D(G) is the smallest (in fact, infimum) value such
that for every u, v ∈ V , dG(u, v) ≤ δ(u, v) ≤ α · dG(u, v).
The term distance oracle was coined by Thorup and Zwick [33]. See their paper also for a
very persuasive motivation of this natural notion. In their seminal paper Thorup and Zwick [33]
devised a path-reporting distance oracle (henceforth, TZ oracle). The TZ oracle with a parameter
k = 1, 2, . . . has size O(k · n1+1/k), stretch 2k − 1 and query time O(k). As argued in [33], this
tradeoff between size and stretch is essentially optimal for k ≤ log n
log log n , as Erdos' girth conjecture
implies that Ω(n1+1/k) space is required for any k. Note, however, that k · n1+1/k = Ω(n · log n),
and Thorup and Zwick [33] left it open if one can obtain meaningful distance oracles of linear size
(or, more generally, size o(n log n)).
A partial answer to this question was provided by Mendel and Naor [23], who devised a
distance oracle with size O(n1+1/k), stretch O(k) and query time O(1). Alas, their distance oracle
is inherently not path-reporting. Specifically, the oracle of [23] stores a collection of O(k · n1/k)
hierarchically-separated trees (henceforth, HSTs; see [8] for its definition), whose sizes sum up
to O(n1+1/k). The query algorithm for this oracle can return paths from these HSTs, i.e., paths
which at best can belong to the metric closure of the original graph. These paths will typically
not belong to the graph itself.
One can try to convert this collection into a collection of low-stretch spanning trees of the input
graph G using star-decomposition or petal-decomposition techniques (see [16, 3]). However, each
of this spanning trees is doomed to have n − 1 edges, making the size of the entire structure as
large as Ω(k · n1+1/k). (In addition, with the current state-of-the-art techniques with low-stretch
spanning trees one can only achieve bounds which are somewhat worse than the optimal ones
achievable with HSTs. Hence the approach that we have just outlined will probably produce an
oracle with stretch ω(k), while using space O(k · n1+1/k).)
For a parameter t ≥ 1 their oracle uses space O(n·t) and provides stretch O(√t·n2/√t) for weighted
graphs. The query time of their oracle is O(log t · logn wmax), where wmax is the aspect ratio of
the graph, i.e., the ratio between the heaviest and the lightest edge. For unweighted graphs their
oracle exhibits roughly the same behavior. For a parameter ǫ > 0 it uses space O(n · t/ǫ) and
provides stretch O(t · n1/t(t + nǫ/t)).
Another result in this direction was recently obtained by Elkin, Neiman and Wulff-Nilsen [17].
1The query time of all path-reporting distance oracles that we will discuss is of the form O(q + Π), where Π is
the path returned by the query algorithm. To simplify the notation we will often omit the additive term of O(Π).
2
log n
The distance oracles of [17] are the first path-reporting oracles that use o(n log n) space and
provide non-trivial stretch. However, their stretch is by far larger than that of the oracles of
[33, 23]. Therefore the tantalizing problem of whether one can have a linear-size path-reporting
distance oracle with logarithmic stretch remained wide open. In the current paper we answer this
log log n ≤ k ≤ log n, and any arbitrarily small constant
question in the affirmative. For any k,
ǫ > 0, our path-reporting distance oracle has stretch O(k), size O(n1+1/k) and query time O(nǫ).
(When ǫ > 0 is subconstant the stretch becomes O(k) · (1/ǫ)O(1).) Hence our oracle achieves an
optimal up to constant factors tradeoff between size and stretch in the range
log log n ≤ k ≤ log n,
i.e., in the range "missing" in the Thorup-Zwick's result. Though our query time is nǫ for an
arbitrarily small constant ǫ > 0 is much larger than Thorup-Zwick's query time, we stress that all
existing path-reporting distance oracles either use space Ω(n · log n) [33, 36, 13] or have stretch
nΩ(1) [17]. (The query time of the TZ oracle was recently improved to O(1) in [36, 13].) The
only previously existing path-reporting distance oracle that achieves the optimal tradeoff in this
range of parameters can be obtained by constructing a (2k − 1)-spanner2 with O(n1+1/k) edges
and answering queries by conducting Dijkstra explorations in the spanner. However, with this
approach the query time is O(n1+1/k). Our result is a drastic improvement of this trivial bound
from O(n1+1/k) to O(nǫ), for an arbitrarily small constant ǫ > 0.
log n
We also can trade between the stretch and the query time. Specifically, a variant of our oracle
uses O(n log log n) space, has stretch O(loglog4/3 7 n) ≈ O(log6.76 n) and query time O(log log n). For
log n
a comparison, the path-reporting distance oracle of [17] with this stretch uses space Ω(n ·
log log n )
and has query time O(log log n · logn wmax).
We also remark that using a super-constant (but not trivial) query time is a common place by
now in the distance oracles literature. In particular, this is the case in the oracles of Porat and
Roditty [30], Agarwal and Godfrey [5] and of Agarwal et al. [6].
1.2 Distance Oracles with Stretch (α, β) for Unweighted Graphs
We say that a distance oracle D(G) provides stretch (α, β) for a pair of parameters α ≥ 1, β ≥ 0
if for any query (u, v) it constructs a path Π(u, v) of length δ(u, v) which satisfies dG(u, v) ≤
δ(u, v) ≤ α· dG(u, v) + β. The notion of (α, β)-stretch is originated from the closely related area of
spanners. A subgraph G′ = (V, H) is said to be an (α, β)-spanner of a graph G = (V, E) , H ⊆ E,
if for every pair u, v ∈ V , it holds that dH(u, v) ≤ α · dG(u, v) + β.
This notion was introduced in [18], where it was shown that for any ǫ > 0 and k = 1, 2, . . .,
for any n-vertex unweighted graph G = (V, E) there exists a (1 + ǫ, β)-spanner with O(β · n1+1/k)
edges, where β = β(ǫ, k) is independent of n. Later a number of additional constructions of
(1 + ǫ, β)-spanners with similar properties were devised in [15, 35, 29].
It is natural to attempt converting these constructions of spanners into distance oracles with
a similar tradeoff between stretch and size. However, generally so far such attempts were not
successful. See, e.g., the discussion titled "Additive Guarantees in Distance Oracles" in the intro-
duction of [25]. Patra¸scu and Roditty [25] devised a distance oracle with stretch (2, 1) and size
O(n5/3), and query time O(1). Abraham and Gavoille [1] generalized the result of [25] to devise
a distance oracle with stretch (2k − 2, 1) and space O(n1+(2/(2k−1))). (The query time in [1] is
unspecified.)
2For a parameter t ≥ 1, G′ = (V, H) is a t-spanner of a graph G = (V, E), H ⊆ E, if dH (u, v) ≤ t · dG(u, v).
3
Note, however, that neither of these previous results achieves multiplicative stretch o(k) with
size O(n1+1/k), at the expense of an additive stretch. (This is the case with the result of [18] in the
context of spanners, where the multiplicative stretch becomes as small as 1 + ǫ, for an arbitrarily
small ǫ > 0.) In this paper we devise the first distance oracles that do achieve such a tradeoff.
Specifically, our path-reporting distance oracle has stretch (O(1), β(k)), space O(β(k) · n1+1/k),
β(k) = kO(log log k), and query time O(nǫ), for an arbitrarily small ǫ > 0. The multiplicative stretch
O(1) here is a polynomial function of 1/ǫ, but it can be made much smaller than k. (Think, e.g.,
of ǫ > 0 being a constant and k being a slowly growing function of n.) We can also have stretch
(o(k), β(k)), space O(β(k) · n1+1/k) and query time nO(k−γ ), where γ > 0 is a universal constant.
(Specifically, the theorem holds, e.g., for γ = 1/7.)
In both these results the tradeoff between multiplicative stretch and size of the oracle is below
Erdos' girth conjecture barrier (which is stretch 2k − 1 and space O(n1+1/k)). In fact, it is known
that when the additive stretch is 0, distance oracles for general n-vertex graphs that have size
O(n1+1/k) must have multiplicative stretch Ω(k) [33, 22, 21]. Our results, like the results of [18]
for spanners, break this barrier by introducing an additive stretch β(k). To the best of our
knowledge, our distance oracles are the first distance oracles that exhibit this behavior.
Using known lower bounds we also show that there exist no distance labeling schemes with
stretch (O(1), β(k)) and maximum label size O(β(k)· n1/k). (Rather one needs labels of size nΩ(1)
for this.) This is also the case for routing schemes. (See Section 2 for relevant definitions.) We also
show that in the cell-probe model of computation any distance oracle for unweighted undirected
n-vertex graphs with stretch (O(1), β(k)) and space O(β(k) · n1+1/k) has query time Ω(k). This
is in contrast to distance oracles with multiplicative stretch, which can have constant query time
[23, 13].
1.3 Distance Oracles for Sparse Graphs
A central ingredient in all our distance oracles is a new path-reporting distance oracle for graphs
with O(n) edges. The most relevant result in this context is the paper by Agarwal et al. [6]. In
this paper the authors devised a (not path-reporting)3 linear-size distance oracle which given a
parameter k = 1, 2, . . . provides distance estimates with stretch 4k − 1, uses linear space and has
time O(n1/(k+1)). (Their result is, in fact, more general than this. We provide this form of their
result to facilitate the comparison.) In this paper we present the first path-reporting linear-size
distance oracle for this range of parameters. Specifically, our linear-size oracle (see Corollary
6.4) has stretch O(klog4/3 7) and query time O(n1/k), for any constant parameter k of the form
k = (4/3)h, h = 1, 2, . . ..
1.4 A Distance-Preserving Path-Reporting Distance Oracle
In [14] the authors showed that for any n-vertex graph G = (V, E) and a collection P of P pairs
of vertices there exists a subgraph G′ = (V, H) of size O(max{n + √n · P,√P · n}) so that for
every (u, v) ∈ P, dH(u, v) = dG(u, v). In this paper we devise the first distance-oracle counterpart
of this result. Specifically, our distance oracle uses O(n + P 2) space, and for any query (u, v) ∈ P
3It was erroneously claimed in [6] that all their distance oracles are path-reporting. While their distance oracles
with stretch smaller than 3 are path-reporting (albeit their space requirement is superlinear), this is not the case
for their oracles with stretch 4k − 1, k ≥ 1 [4].
4
it produces the exact shortest path Π between u and v in O(Π) time, where Π is the number
of edges in Π.
We employ this distance oracle very heavily in all our other constructions.
Remark: The construction time of our distance-preserving oracle is O(n·P 2)+ O(m·min{n, P}).
The construction time of our path-reporting oracle for sparse graphs is O(m · n) = O(n2λ), where
λ = m/n. The construction time of our oracles with nearly-linear space for general graphs is
O(n2+1/k). Finally, the construction time of our oracle for unweighted graphs with a hybrid
multiplicative-additive stretch is O(β(k)n2+1/k) = kO(log log k) O(n2+1/k). (In both cases k is the
stretch parameter of the respective oracle.)
1.5 Related Work
There is a huge body of literature about distance oracles by now. In addition to what we have
already surveyed there are probe-complexity lower bounds by Sommer et al. [32]. There is an
important line of work by Patra¸scu et al. [26, 25] on oracles with rational stretch. Finally, Baswana
and Sen [11], Baswana and Kavitha [10] and Baswana et al. [9] improved the preprocessing time
of the TZ oracle.
1.6 Structure of the Paper
We start with describing our distance preserving oracle (Section 3). We then proceed with devis-
ing our basic path-reporting oracle for sparse graphs (Section 4). This oracle can be viewed as a
composition of an oracle from Agarwal et al. [6] with our distance-preserving oracle from Section
3. The oracle is described for graphs with small arboricity. Its extension to general sparse graphs
(based on a reduction from [6]) is described in Section 5. Then we devise a much more elaborate
multi-level path-reporting oracle for sparse graphs. The oracle of [6] and our basic oracle from Sec-
tion 4 both use just one set of sampled vertices. Our multi-level oracle uses a carefully constructed
hierarchy of sampled sets which enables us to get the query time down from n1/2+ǫ to nǫ. Next
we proceed (Section 6) to using this multi-level oracle for a number of applications. Specifically,
we use it to construct a linear-size logarithmic stretch path-reporting oracle with query time nǫ,
linear-size polylogarithmic stretch path-reporting oracle with query time O(log log n), and finally,
oracles that break the girth barrier for unweighted graphs. Our lower bounds can be found in
Section 7.
2 Preliminaries
For a pair of integers a ≤ b, we denote [a, b] = {a, a + 1, . . . , b}, and [b] = [1, b]. The arboricity
of a graph G is given by λ(G) = maxU⊆V,U≥2 E(U )
U−1 , where E(U) is the set of edges induced
by the vertex set U. We denote by degG(u) the degree of a vertex u in G; we omit G from
this notation whenever G can be understood from the context. We use the notation O(f (n)) =
O(f (n)polylog(f (n))) and Ω(f (n) = Ω(f (n)/polylog(f (n))). We say that a function f () is quasi-
polynomial if f (n) ≤ nlogO(1) n.
A distance-labeling scheme for a graph G = (V, E) assigns every vertex v ∈ V a short label
ϕ(v). Given a pair of labels ϕ(u), ϕ(v) of a pair of vertices u, v ∈ V , the scheme computes an
5
estimate δ(ϕ(u), ϕ(v)). This estimate has to be within a factor α, for some α ≥ 1, from the actual
distance dG(u, v) between u and v in G. The parameter α is called the stretch of the labeling
scheme, and the maximum number of bits employed by one of the labels is called the (maximum)
label size of the scheme.
A closely related notion is that of compact routing scheme. Here each vertex v is assigned
a label ϕ(v) and a routing table ψ(v). Given a label ϕ(u) of routing destination u and its own
routing table ψ(v), the vertex v = v0 needs to be able to compute the next hop v1. Given the table
ψ(v1) of v1 and the destination's label ϕ(u), the vertex v1 computes the next hop v2, etc. The
resulting path v = v0, v1, v2, . . . has to end up eventually in u, and its length needs to be at most
α times longer than the length of the shortest u − v path in G, for a stretch parameter α ≥ 1. In
addition to stretch, another important parameter in this context is the maximum number of bits
used by the label and the routing table (together) of any individual vertex. This parameter will
be referred to as maximum memory requirement of a routing scheme.
3 A Distance-Preserving Path-Reporting Oracle
Consider a directed weighted n-vertex graph G = (V, E, ω).
(The result given in this section
applies to both directed and undirected graphs. However, our other distance oracles apply only
to undirected graphs.) Let Pairs ⊆ (cid:0)V
2(cid:1) be a subset of ordered pairs of vertices. We denote
its cardinality by P = Pairs. In this section we describe a distance oracle which given a pair
(u, v) ∈ Pairs returns a shortest path Πu,v from u to v in G. The query time of the oracle is
proportional to the number of edges (hops) Πu,v in Πu,v. The oracle uses O(n + P 2) space.
The construction of the oracle starts with computing a set Paths = {Πu,v (u, v) ∈ Pairs} of
shortest paths between pairs of vertices from Pairs. This collection of shortest paths is required
to satisfy the property that if two distinct paths Π, Π′ ∈ Paths traverse two common vertices x
and y in the same order (i.e., e.g., both traverse first x and then y), then they necessarily share
the entire subpath between x and y. It is argued in [14] that this property can be easily achieved.
We will need the following definitions from [14].
For a path Π = (u0, u1, . . . , uh) and a vertex ui ∈ V (Π), the predecessor of ui in Π, denoted
predΠ(ui), is the vertex ui−1 (assuming that i ≥ 1; otherwise it is defined as NULL), and the
successor of ui in Π, denoted succΠ(ui), is the vertex ui+1 (again, assuming that i ≤ h − 1;
otherwise it is NULL).
Definition 3.1 [14] A branching event (Π, Π′, x) is a triple with Π, Π′ ∈ Paths being two distinct
paths and x ∈ V (Π)∩V (Π′) be a vertex that belongs to both paths and such that {predΠ(x), succΠ(x)} 6=
{predΠ′(x), succΠ′(x)}. We will also say that the two paths Π, Π′ branch at the vertex x.
Note that under this definition if Π traverses edges (ui−1, ui), (ui, ui+1) and Π′ traverses edges
(ui+1, ui), (ui, ui−1) then (Π, Π′, ui) is not a branching event.
It follows directly from the above property of the collection Paths (see also [14], Lemma 7.5,
for a more elaborate discussion) that for every pair of distinct paths Π, Π′ ∈ Paths, there are at
most two branching events that involve that pair of paths. Let B denote the set of branching
events. The overall number of branching events for the set Paths is B ≤ P aths2 = P 2. Our
oracle will keep O(1) data for each vertex, O(1) data for each branching event, and O(1) data for
each path. Hence the oracle stores O(n + B + P ) data in total.
6
Specifically, in our oracle for every vertex v ∈ V we keep an identity of some path Π ∈ Paths
(If there is no path
that contains v as an internal point, and two edges of Π incident on v.
Π ∈ Paths that contains v as an internal point, then our oracle stores nothing for v in this data
structure.) The path Π stored for v will be referred to as the home path of v.
In addition, for every branching event (Π, Π′, v) we keep the (at most four) edges of Π and Π′
incident on v. Finally, for every pair (x, y) ∈ Pairs we also store the first and the last edges of the
path Πx,y. Observe that the resulting space requirement is at most O(n+B+P ) = O(n+P 2). We
assume that the branching events are stored in a hash table of linear size, which allows membership
queries in O(1) time per query.
The query algorithm proceeds as follows. Given a pair (x, y) ∈ Pairs, we find the first edge
(x, x′) of the path Πx,y, and "move" to x′. Then we check if (x′, y) is the last edge of Πx,y. If it is
then we are done. Otherwise let Π(x′) denote the home path of x′. (Observe that since the vertex
x′ is an internal vertex in Πx,y, it follows that there exists a home path Π(x′) for x′.)
Next, we check if Π(x′) = Πx,y. (This test is performed by comparing the identities of the two
paths.) If it is the case then we fetch the next edge (x′, x′′) of Π(x′), and move to x′′. Otherwise
(if Π(x′) 6= Π(x, y)) then we check if the triple (Π(x′), Πx,y, x′) is a branching event. This check is
performed by querying the branching events' hash table.
If there is no branching event (Π(x′), Πx,y, x′) then we again fetch the next edge (x′, x′′) of
Π(x′), and move to x′′. (In fact, the algorithm does not need to separate between this case and
the case that Π(x′) = Πx,y. We distinguished between these cases here for clarity of presentation.)
Finally, if there is a branching event (Π(x′), Πx,y, x′) then we fetch from our data structure all
the information associated with this event. In particular, we fetch the next edge (x′, x′′) of Πx,y,
and move to x′′.
In all cases the procedure then recurses with x′′. It is easy to verify that using appropriate
hash tables all queries can be implemented in O(1) time per vertex, and in total O(Πx,y) time.
We summarize this section with the following theorem.
Theorem 3.2 Given a directed weighted graph G = (V, E, ω) and a collection Pairs ⊆(cid:0)V
2(cid:1) of pairs
of vertices, our distance-preserving path-reporting oracle (shortly, DPPRO) reports shortest paths
Πx,y for query pairs (x, y) ∈ Pairs in O(Πx,y) time. The oracle employs O(n+B+P ) = O(n+P 2)
space, where B is the set of branching events for a fixed set of shortest paths between pairs of vertices
from Pairs, and P = Pairs.
One can construct the shortest paths in O(m · min{P, n}) time. Then for each vertex v one
keeps the list of paths that traverse v. For every such path one keeps the two edges of this path
which are incident on v. In overall O(n · P 2) additional time one can use these lists to create the
list of branching events. A hash table with them can be constructed in additional O(P 2) time.
Hence the overall construction time of this oracle is O(m · min{P, n}) + O(n · P 2).
Observe that if one is given a set S, S = O(n1/4), of terminals, then Theorem 3.2 provides
a linear-size DPPRO (i.e., O(1) words per vertex on average) which can report shortest paths
between all pairs of terminals. It is well-known that any distance labeling scheme which is guar-
anteed to return exact distances between all pairs of n1/4 terminals must use maximum label size
Ω(n1/4) [33]. This is also the case for compact routing schemes [34]. (In the latter case the lower
bound of Ω(n1/4) is on the maximum memory requirement of any individual vertex.)
7
We remark that our DPPRO here employs O(n + B + P ) space, whereas the underlying
distance preserver has O(n +pn · B) edges [14]. It is plausible that there exists a DPPRO of
size O(n +pn · B). We leave this question open.
4 A Basic Distance Oracle for Graphs with
Bounded Arboricity
In this section we describe a basic variant of our path-reporting distance oracle for weighted
undirected graphs G = (V, E, ω) of arboricity λ(G) ≤ λ, for some parameter λ. (We will mostly
use this oracle for constant or small values of λ. On the other hand, the result is meaningful for
higher values of λ as well.) Our oracle reports paths of stretch O(k), for some positive integer
parameter k. Unlike the partial oracle from Section 3, the oracle in this section is a full one,
i.e., it reports paths for all possible queries (u, v) ∈ (cid:0)V
2k+2 · λ). (Whp4, the query time is O(n1/2+ 1
2(cid:1). This is the case also for all our other
oracles, which will be described in consequent sections. The expected query time of our oracle is
O(n1/2+ 1
2k+2 · log n· λ).) The oracle requires O(n) space,
in addition to the space required to store the graph G itself. Observe that for λ = O(1) the query
time is O(n1/2+ǫ), for an arbitrarily small constant ǫ > 0, while the stretch is O( 1
ǫ ) = O(1). In
Section 5 we extend this oracle to general m-edge n-vertex graphs with λ = m
n .
Our basic oracle employs just one level of sampled vertices, which we (following the terminology
of [6]) call landmarks. Each v ∈ V is sampled independently at random with probability ρ
n , where
ρ is a parameter which will be determined in the sequel. Denote by L the set of sampled vertices
(landmarks). Note that IE(L) = ρ.
For every vertex v ∈ V we keep the path Π(v) to its closest landmark vertex ℓ(v), breaking ties
arbitrarily. Denote by D(v) the length w(Π(v)) of this path. This is a collection of vertex-disjoint
shortest paths trees (shortly, SPTs) {T (u) u ∈ L}, where each T (u) is an SPT rooted at u for the
subset {v dG(u, v) ≤ dG(u′, v),∀u′ 6= u, u, u′ ∈ L}. (Ties are broken arbitrarily.) This collection
is a forest, and storing it requires O(n) space.
The oracle also stores the original graph G. For the set of landmarks we compute the complete
graph L = (L,(cid:0)L
2(cid:1), dGL). Here dGL stands for the metric of G restricted to the point set L. (In
other words, in the landmarks graph L, for every pair u, u′ ∈ L of distinct landmarks the weight
ωL(u, u′) of the edge (u, u′) connecting them is defined by ωL(u, u′) = dG(u, u′).)
Next we invoke Thorup-Zwick's distance oracle [33] with a parameter k. (Henceforth we will
call it the TZ oracle.) One can also use here Mendel-Naor's oracle [23], but the resulting tradeoff
will be somewhat inferior to the one that is obtained via the TZ oracle. Denote by H the TZ
distance oracle for the landmarks graph L. The oracle requires O(k·L1+1/k) space, and it provides
(2k − 1)-approximate paths Πu,u′ in L for pairs of landmarks u, u′ ∈ L. The query time is O(k)
(plus O(Πu,u′)). Observe that some edges of Πu,u′ may not belong to the original graph G. We
note also that by using more recent oracles [13, 36] one can have here query time O(1), but this
improvement is immaterial for our purposes.
The TZ oracle H has a useful property that the union H =S{Πu,u′ (u, u′) ∈(cid:0)L
2(cid:1)} of all paths
that the oracle returns forms a sparse (2k − 1)-spanner. Specifically, IE(H) = O(k · L1+1/k).
4Here and thereafter we use the shortcut "whp" for "with high probability". The meaning is that the probability
is at least 1 − n−c, for some constant c ≥ 2.
8
(This property holds for Mendel-Naor's oracle as well, but there the stretch of the spanner is
O(k), where the constant hidden by the O-notation is greater than 2. On the other hand, their
space requirement is O(L1+1/k), rather than O(k · L1+1/k).) Fix an oracle H as above for
H = O(k · L1+1/k). Whp such an H will be computed by running the procedure that computes
the TZ oracle for O(log n) times. We will view the spanner H as a collection of pairs of vertices
of our original graph G.
Finally, we invoke our distance preserving oracle (shortly, DPPRO) from Section 3 on the graph
G and set Pairs = H. We will refer to this oracle as D(G, H). Its size is, with high probability,
O(n + H2) = O(n + k2 · L2+2/k). Upon a query (y, y′) ∈ H, this oracle returns a shortest path
Πy,y′ between y and y′ in G in time O(Πy,y′).
Observe that L is the sum of identical independent indicator random variables L =Pv∈V Iv,
where Iv is the indicator random variable of the event {v ∈ L}. Hence, by Chernoff's inequality,
for any constant ǫ > 0,
IP(L > (1 + ǫ)IE(L)) = IP(L > (1 + ǫ) · ρ) < exp(−Ω(ρ)) .
We will set the parameter ρ to be at least c log n, for a sufficiently large constant c. This will
ensure that whp L = O(ρ), and so L2+2/k = O(ρ2+2/k). Set ρ so that k2 · ρ2+2/k = Θ(n), i.e.,
ρ = n
k . This guarantees that aside from the storage needed for the original graph, the total
space used by our oracle is O(n).
2k+2 · 1
k
This completes the construction algorithm of our oracle. Next we describe its query algorithm.
We need the following definition. For a vertex v ∈ V , let Ball(v) = {x dG(v, x) < dG(v, ℓ(v))}
denote the set of all vertices x which are closer to v than the closest landmark vertex ℓ(v) to v.
Given a pair u, v of vertices of G, our oracle starts with testing if u ∈ Ball(v) and if v ∈ Ball(u).
To test if u ∈ Ball(v) we just conduct a Dijkstra exploration rooted at v in the graph G, until we
discover either u or ℓ(v). (Recall that G is stored in our oracle.) If u is discovered before ℓ(v)
we conclude that u ∈ Ball(v), and return the (exact) shortest path between them. Otherwise we
conclude that u 6∈ Ball(v). Analogously, the algorithm tests if v ∈ Ball(u).
Henceforth we assume that u 6∈ Ball(v) and v 6∈ Ball(u), and therefore the two searches returned
u′ = ℓ(u), v′ = ℓ(v), and the shortest paths Π(u) and Π(v) between u and u′ and between v and
v′, respectively. (In fact, using the forest of SPTs rooted at landmarks that our oracle stores,
the query algorithm can compute shortest paths between u and u′ and between v and v′ in time
proportional to the lengths of these paths.) Observe that dG(u′, v′) ≤ dG(u′, u)+dG(u, v)+dG(v, v′),
and dG(u′, u), dG(v, v′) ≤ dG(u, v). Hence dG(u′, v′) ≤ 3 · dG(u, v).
Then the query algorithm invokes the query algorithm of the oracle H for the landmarks graph
L. The latter algorithm returns a path Π′ = (u′ = z0, z1, . . . , zh = v′) in L between u′ and v′. The
length ωL(Π′) of this path is at most (2k− 1)· dG(u′, v′) ≤ (6k− 3)· dG(u, v). The time required for
this computation is O(k + h), where Π′ = h. For each edge (zi, zi+1) ∈ Π′, i ∈ [0, h− 1], we invoke
the query algorithm of the DPPRO D(G, H). (The edges (zi, zi+1) of the path Π′ are typically not
edges of the original graph. H is a (2k − 1)-spanner of L produced by the oracle H. Observe that
Π′ ⊆ H, and so (zi, zi+1) ∈ H, for every index i ∈ [0, h − 1].) The oracle D(G, H) returns a path
Πi between zi and zi+1 in G of length ωL(zi, zi+1) = dG(zi, zi+1). Let Π = Π0 · Π1 · . . .· Πh−1 be the
concatenation of these paths. Observe that Π is a path in G between z0 = u′ and zh = v′, and
ω( Π) =
ω( Πi) =
h−1
Xi=0
h−1
Xi=0
dG(zi, zi+1) =
h−1
Xi=0
9
ωL(zi, zi+1) = ωL(Π′) ≤ (6k − 3) · dG(u, v) .
Finally, the query algorithm returns the concatenated path Π = Π(u)· Π· Π(v) as the approximate
path for the pair u, v. This completes the description of the query algorithm of our basic oracle.
Observe that
ω( Π) = ω(Π(u)) + ω( Π) + ω(Π(v)) ≤ dG(u, v) + (6k − 3) · dG(u, v) + dG(u, v) = (6k − 1) · dG(u, v) .
Next, we analyze the running time of the query algorithm. First, consider the step that tests if
v ∈ Ball(u) and if u ∈ Ball(v). Denote by X the random variable that counts the number of
vertices discovered by some fixed Dijkstra exploration originated at u before the landmark ℓ(u)
is discovered. We order all graph vertices by their distance from u in a non-decreasing order,
i.e., u = u0, u1, . . . , un−1, such that dG(u, ui) ≤ dG(u, uj) for i ≤ j. (This is the order in which
the aforementioned Dijkstra exploration originated at u discovers them.) For an integer value
1 ≤ t ≤ n−1, the probability that X = t is equal to the probability that the vertices u0, u1, . . . , ut−1
are all not sampled and the vertex ut is sampled. Hence X is distributed geometrically with the
parameter p = ρ/n. Hence
IE(X) =
n−1
Xt=1
(1 − p)t · p · t ≤
1
p
=
n
ρ
.
(1)
ρ log n).
ρ λ) edges, and whp O( n
ρ (λ + log n)) in expectation, and O( n
Also, obviously for any positive constant c, IP(X > n
X = O( n
ρ c ln n) ≤ (1 − ρ/n)(n/ρ)c ln n ≤ n−c, i.e., whp
Recall that the graph G has arboricity at most λ, and thus any set of n′ ≤ n vertices induces
O(n′ · λ) edges. Hence Dijkstra algorithm traverses expected O( n
ρ λ log n)
edges. In an unweighted graph such exploration requires time linear in the number of edges, and
in weighted5 graphs the required time is O( n
ρ λ · log n) whp.
(Recall that Dijkstra algorithm that scans a subgraph (V ′, E′) requires time O(E′+V ′ log V ′).)
The second step of our query algorithm queries the distance oracle H for the landmarks graph
L. (The query is (u′, v′), u′ = ℓ(u), v′ = ℓ(v).) This query returns a path Π′ between u′ and v′
in L in time O(Π′ + k). Finally, for each of the h = Π′ edges (zi, zi+1), i = 0, 1, . . . , h − 1 of
the path Π′, the query algorithm invokes our DPPRO D(G, H) with the query (zi, zi+1). This
oracle returns the shortest path Πi between zi and zi+1 in G within time O( Πi). Finally, the
algorithm returns the concatenated path Π = Π(u) · Π0 · Π1 · . . . · Πh−1 · Π(v). The running time
i=0 Πi) = O( Π), and Π′ ≤ Π. Hence
required for producing the path Π0 · . . . · Πh−1 is O(Ph−1
ρ · λ + Π) for unweighted graphs, and
the overall expected running time of the algorithm is O( n
ρ · (λ + log n) + Π) for weighted. (Observe that the additive term of O(k) is dominated by
is O( n
O( n
ρ · λ). Specifically, we will be using ρ ≤ n/ log n, and k ≤ O(log n).) For the high-probability
bounds one needs to multiply the first term of the running time by an additional O(log n) factor
in both the unweighted and the weighted cases.
1
2 + 1
2k+2 ·
Now we substitute ρ = 1
2k+2 . The resulting expected query time becomes O(k · n
λ) + O( Π). We summarize the properties of our basic oracle in the following theorem.
k · n
k
5One subtlety: we have to avoid scanning too many edges with just one endpoint in Ball(u). We store the edges
incident to each vertex x in increasing order of their weights, and relax them in that order when x is scanned. As
soon as an edge (x, y) is relaxed such that the tentative distance to y is greater than dG(u, ℓ(u)) we can dispense
with relaxing the remaining edges. Alternatively, a modification of the sampling rule which we describe in Section
5 also resolves this issue.
10
Theorem 4.1 For an undirected n-vertex graph G of arboricity λ and a positive integer parameter
k = 1, 2, . . ., there exists a path-reporting distance oracle of size (whp) O(n) (in addition to the
size required to store the input graph G) that returns (6k − 1)-approximate shortest paths Π. The
expected query time is O(n
2k+2 · k · (λ + log n)) in
weighted ones. (The same bounds on the query time apply whp if one multiplies them by O(log n).
In addition, in all cases the query time contains the additive term O( Π).)
2k+2 · k · λ) in unweighted graphs and O(n
2 + 1
2 + 1
1
1
In particular, Theorem 4.1 implies that for any constant ǫ > 0 one can have a path-reporting
oracle with query time O(n1/2+ǫλ), which provides O(1)-approximate shortest paths for weighted
undirected graphs. Observe also that for k = 1 we obtain a 5-approximate path-reporting oracle
with query time O(n3/4λ). We remark that to get the latter oracle one does not need to use the TZ
oracle for the landmarks graph L. Rather one can build a DPPRO H for all pairs of landmarks.
2(cid:1) = O(ρ2) = O(√n), and so the size of the oracle
(In this case ρ = n1/4, L = O(ρ), Pairs = (cid:0)L
H is O(Pairs2 + n) = O(n).)
One can build the forest of SPTs rooted at the landmarks in O(m) time. In additional O(m·ρ) =
O(k · m · n1/2− 1
2k+2 ) time one can construct the metric closure of L, i.e., the graph L. This graph
has n′ = ρ vertices and m′ ≤ ρ2 edges. In O(km′ · n′1/k) = O(kρ2+1/k) = O(k · n
2k+2 ) time one
can construct the TZ oracle for it. To construct the DPPRO with P = O(k · ρ1+1/k) = O(k · n1/2)
pairs one needs O(n· P 2) + O(k · m · n1/2− 1
2k+2 ) time. Hence the
overall construction time of this oracle is O(k2 · n2) + O(k · m · n1/2− 1
m = λ · n edges.
In Section 5 we show (see Corollary 5.1) that Theorem 4.1 extends to general graphs with
2k+2 ) = O(k2 · n2) + O(k · m· n1/2− 1
2k+2 ).
2k+1
5 An Extension to General Graphs
In this section we argue that Theorem 4.1 can be extended to general n-vertex graphs G = (V, E, ω)
with m = λn edges. In its current form the theorem only applies to graphs of arboricity at most
λ. While this is sufficient for our main application, i.e., for Theorem 6.7, our another application
(Theorem 6.8) requires a more general result. Our extension is based on the reduction of Agarwal
et al. [6] of the distance oracle problem in general graphs to the same problem in bounded-degree
graphs. Our argument is somewhat more general than the one from [6], as it also applies to
path-reporting distance oracles. We provide our extension for the sake of completeness.
Given an m-edge n-vertex graph G with λ = m/n, we split each vertex ui into d(u) = ⌈ deg(u)
⌉
copies u(1), u(2), . . . , u(d(u)). Each copy is now selected independently at random with probability
ρ/n, for a parameter ρ determined in the same way as in Section 4. The original vertex u is
selected to the landmarks' set if and only if at least one of its copies (which will also be called
virtual nodes) is selected. Observe that the rule that we have described is equivalent to selecting
u with probability d(u) · ρ
n = ⌈ deg(u)
⌉ · ρ
n .
The expected number of selected virtual nodes is
λ
λ
d(v) ·
ρ
n
=
Xv∈V
ρ
n ·Xv∈V
⌈
deg(u)
λ
⌉ ≤
deg(v)
λ
(
ρ
nXv∈V
+ 1) = ρ +
ρ
λnXv∈V
deg(v) = 3ρ .
The number L of landmarks is at most the number of selected virtual nodes, and so IE(L) ≤ 3ρ.
11
By Chernoff's bound, the number of selected virtual nodes is whp O(ρ), and so, whp, L2+2/k =
O(ρ2+2/k) as well. Hence the size of our oracle remains O(n).
The rest of the construction algorithm for our distance oracle is identical to that of Section 4.
(The only change is the distribution of selecting landmarks.) The query algorithm is identical to
the query algorithm from Section 4. In particular, note that the virtual nodes have no effect on
the computation, i.e., the returned paths contain only original vertices.
Next we argue that the expected query time of the modified oracle is still at most O( n
ρ · λ) in
unweighted graphs, and O( n
ρ · λ log n) in weighted ones. (As usual, we omit the additive term of
the number of edges of the returned path.) Specifically, we argue that the tests if v ∈ Ball(u) and
if u ∈ Ball(v) can be carried out within the above expected time.
Let u = u0, u1, . . . , un−1 be all graph vertices ordered by a Dijkstra exploration originated
, . . . , u(d(ui))
from u, and replace each vertex ui by its d(ui) copies u(1)
. The copies appear in an
i
arbitrary order. Since each virtual node has probability ρ
n to be selected independently of other
vertices, it follows by a previous argument that the number N of virtual nodes that the algorithm
encounters before seeing a selected virtual node is O( n
ρ ). (The algorithm actually explores only
original vertices. For the sake of this argument we imagine that when the algorithm reaches a
vertex y it reaches its first copy y(1). Right after that it reaches the next copy y(2), etc., and
then reaches y(d(y)). After "reaching" all these copies the algorithm continues to the next original
vertex.)
i
i be a
selected copy of ui. (We assume that all copies of uj, for j < i, are not selected, and all copies
uh′
Denote the original vertices explored by the algorithm u1, u2, . . . , ui−1, ui, and let uh
i , h′ < h, are also not selected.) It follows that N =Pi−1
j=1 d(uj) + h. Hence
Hence
as well. Thus
d(uj)! ≤ IE(N) = O(cid:18)n
ρ(cid:19) .
⌉! = O(cid:18)n
ρ(cid:19)
IE i−1
Xj=1
IE i−1
Xj=1
deg(uj)! = O(cid:18)λn
ρ (cid:19) = O(cid:18) m
ρ(cid:19) .
deg(uj)
⌈
λ
IE i−1
Xj=1
Observe that the number of edges explored by the algorithm before reaching ui is at most
Pi−1
j=1 deg(uj). (The only edges incident on ui explored by the algorithm are edges (uj, ui), for
j < i. These edges are accounted for in the above sum of degrees.) Hence the expected number of
edges explored by the algorithm is O( m
ρ ) (respectively,
O( m
ρ · log n)) in unweighted (resp., weighted) graphs. The bounds that hold with high probability
are higher by a factor of O(log n).
ρ ). Hence its expected running time is O( m
Corollary 5.1 Up to constant factors, the result of Theorem 4.1 holds for general undirected
unweighted m-edge n-vertex graphs with m = λn. For undirected weighted graphs the expected
query time becomes O(n1/2+ 1
n · log n), and the same bound
applies whp if one multiplies it by another log n factor.
2k+2 · k · λ · log n) = O(n1/2+ 1
2k+2 · k · m
12
Since IE(L) = O(ρ), the construction time of the oracle is, up to constant factors, the same as in
Section 4.
This result provides a path-reporting analogue of the result of Agarwal et al. [6], which provides
stretch O(k) and query time (nλ)O(1/k). Their oracle is not path-reporting. Our oracle is path-
reporting, but its query time is significantly higher, specifically it is n1/2+O(1/k) · k · λ.
6 Oracles with Smaller Query Time
In this section we devise two path-reporting oracles with improved query time. The first oracle has
size O(m + n) (it stores the original graph), and query time λ · nǫ, for an arbitrarily small ǫ > 0.
The stretch parameter of this oracle grows polynomially with ǫ−1. For the time being we will focus
on graphs of arboricity at most λ. The argument extends to general graphs with m = λn in the
same way as was described in Section 5. Our second oracle has size O(n log log n) (independent of
the size of the original graph) and reports stretch-O(loglog4/3 7 n) paths in O(log log n) time. Both
draw on techniques used in sublinear additive spanner constructions of [29]. We will later build
upon the first oracle to construct additional oracles that work for dense graphs as well. Like the
second oracle, these later oracles will not have to store the input graph.
6.1 Construction of an Oracle with time O(λ · nǫ)
It will use a hierarchy of
In this section we describe the construction algorithm of our oracle.
landmarks' sets L1, L2, . . . , Lh, for a positive integer parameter h that will be determined later.
For each index i ∈ [h], every vertex v is selected into Li independently at random with probability
pi = ρi
n , ρ1 > ρ2 > . . . > ρh. The sequence ρ1, ρ2, . . . , ρh will be determined in the sequel. The
vertices of Li will be called the i-level landmarks, or shortly, the i-landmarks. For convenience of
notation we also denote L0 = V .
i
(v) denote the one-third-ball of v, and Balli(v) = B1
For each vertex v ∈ V and index i ∈ [h], let ℓi(v) denote the closest i-landmark to v, where
ties are broken in an arbitrary consistent way. Denote ri(v) = dG(v, ℓi(v)) the distance between
v and its closest i-landmark ℓi(v). Following [29], for a real number 0 < c ≤ 1, let Bc
i = {u
dG(v, u) < c·ri(v)} denote the ith c-fraction-ball of v. In our analysis c will be set to either 1/3 or 1.
Specifically, let B1/3
i (v) = {u dG(v, u) < ri(v)}
denote the ith ball of v.
For each vertex v ∈ V we keep a shortest path between v and ℓ1(v). (This is a forest of vertex-
disjoint SPTs rooted at 1-landmarks. For each 1-landmark u′, its SPT spans all vertices v ∈ V
which are closer to u′ than to any other 1-landmark.) Similarly, for each i ∈ [h − 1] and every
i-landmark u we keep a shortest path between u and its closest (i + 1)st landmark ℓi+1(u) = u(i+1).
Again, this entails storing a forest of vertex-disjoint SPTs rooted at (i + 1)-landmarks, for each
each index i ∈ [h − 1]. Overall this part of the oracle requires O(n · h) space.
For the hth-level landmarks' set Lh we build a DPPRO Lh described in Section 3. Given a pair
u, v of h-landmarks this oracle returns a shortest path Π(u, v) between them in time proportional
to the number of edges in this path, i.e., O(Π(u, v)). The space requirement of the oracle Lh is
O(n + Lh4), and thus we will select ρh to ensure that Lh4 = O(n), i.e., ρh will be roughly n1/4.
Denote also Ph =(cid:0)Lh
2(cid:1) be the set of all pairs of h-landmarks.
13
For each index i ∈ [h − 1], we also build a DPPRO Di for the following set Pi of pairs of
i+1(v) is inserted
i-landmarks. Each pair of i-landmarks u, v such that either v ∈ B1/3
into Pi.
Similarly to the DPPRO Lh, given a pair (u, v) ∈ Pi for some i ∈ [h− 1], the oracle Di returns
a shortest path Π(u, v) between u and v in time O(Π(u, v)). Our oracle also stores the graph G
itself. We will later show a variant of this oracle that does not store G (Theorem 6.6). The size of
the oracle is O(n + Branchi), where Branchi is the set of branching events for the set Pi. Since
we aim at a linear size bound, we will ensure that Branchi = O(n), for every i ∈ [h − 1]. We will
also construct a hash table Hi for Pi of size O(Pi) that supports membership queries to Pi in
O(1) time per query. The resulting h-level oracle will be denoted Λh.
i+1(u) or u ∈ B1/3
6.2 The Query Algorithm
Next, we describe the query algorithm of our oracle Λh. The query algorithm is given a pair
u = u(0), v = v(0) of vertices. The algorithm starts with testing if u ∈ Ball1(v) and if v ∈ Ball1(u).
For this test the algorithm just conducts a Dijkstra search from v until it discovers either v(1) or
u (and, symmetrically, also conducts a search from u).
Observe that by Equation (1), the expected size of Ball1(v) and of Ball1(u) is O( n
ρ1
), and whp
both these sets have size O( n
ρ1 · λ).
(Specifically, it is O( n
ρ1 · log n · (λ + log n)) in weighted
ones. The expected running time of this step is smaller by a factor of log n than the above bound.)
If the algorithm discovers that v ∈ Ball1(u) or that u ∈ Ball1(v) then it has found the shortest
In this case the algorithm returns this path. Otherwise it has found
ρ1 · log n). Hence the running time of this step is, whp, O( n
ρ1 · λ · log n) in unweighted graphs, and O( n
path between u and v.
u(1) = ℓ1(u(0)) and v(1) = ℓ1(v(0)).
In general consider a situation when for some index j, 1 ≤ j ≤ h, the algorithm has already
computed u(j) and v(j). In this case, inductively, the algorithm has already computed shortest
paths Π(u(0), u(1)), Π(u(1), u(2)), . . . , Π(u(j−1), u(j)) and Π(v(0), v(1)), Π(v(1), v(2)), . . . , Π(v(j−1), v(j))
between u(0) and u(1), u(1) and u(2), . . ., u(j−1) and u(j), v(0) and v(1), v(1) and v(2), . . ., v(j−1) and
v(j), respectively. (Note that the base case j = 1 has been just argued.)
For j < h, the query algorithm of our oracle Λh then queries the hash table Hj whether the pair
(u(j), v(j)) ∈ Pj. If it is the case then the algorithm queries the oracle Dj, which, in turn, returns
the shortest path Π(u(j), v(j)) between u(j) and v(j) in time O(Π(u(j), v(j))). The algorithm then
reports the concatenated path
Π(u, v) = Π(u(0), u(1)) · Π(u(1), u(2)) · . . . Π(u(j−1), u(j)) · Π(u(j), v(j))
· Π(v(j), v(j−1)) · . . . · Π(v(2), v(1)) · Π(v(1), v(0)) .
Computing this concatenation requires O(j) ≤ O(Π(u, v)) time.
In the complementary case when (u(j), v(j)) 6∈ Pj, the algorithm fetches the prerecorded paths
Π(u(j), u(j+1)) and Π(v(j), v(j+1)), and invokes itself recursively on the pair (u(j+1), v(j+1)). (Recall
that for each index j, 1 ≤ j ≤ h− 1, the algorithm stores a forest of vertex-disjoint SPTs rooted at
(j + 1)-landmarks Lj+1. These SPTs enable us to compute the paths Π(u(j), u(j+1)), Π(v(j), v(j+1))
for all j ∈ [h − 1], in time proportional to the number of edges in these paths.)
Finally, if j = h then we query the DPPRO Lh of the graph Lh with the query (u(h), v(h)).
(Note that it is not necessary to query if (u(h), v(h)) is in the DPPRO Lh, since, by construction, all
14
such pairs are there.) The query returns the shortest path between them in time O(Π(u(h), v(h))).
It follows that the overall running time of the query algorithm is dominated by the time required
to compute Π(u(0), u(1)) and Π(v(0), v(1)). Specifically, it is
j−1
O(
n
ρ1 · λ) +
Remark:
Xi=0 (cid:0)Π(u(i), u(i+1)) + Π(v(i), v(i+1))(cid:1) + Π(u(j), v(j)),
where 1 ≤ j ≤ h is the smallest index such that (u(j), v(j)) ∈ Pj. (Recall that for j = h, Ph =(cid:0)Lh
2(cid:1),
i.e., all pairs of h-landmarks belong to Ph.) Hence the overall query time is O( n
ρ1 ·λ)+O(Π(u, v)+
h), where Π(u, v) is the path that the algorithm ultimately returns.
If for each index 0 ≤ j ≤ h−1 at least one of the subpaths Π(u(j), u(j+1)), Π(v(j), v(j+1))
is not empty then h ≤ Π(u, v), and the resulting query time is O( n
λ) + O(Π(u, v)). One
can artificially guarantee that all these subpaths will not be empty, i.e., that u(j) 6= u(j+1) and
v(j) 6= v(j+1), for every j. To do this one can modify the construction slightly so that the set of
i-landmarks and the set of j-landmarks will be disjoint for all i 6= j. Under this modification
of the algorithm the query time is O( n
ρ1 · λ) + O(Π(u, v)), while the stretch guarantee of the
oracle (which will be analyzed in Section 6.3) stays the same. This modification can make oracle's
performance only worse than it is without this modification, but the bounds on the query time of
the modified oracle in terms of the number of edges in the returned path become somewhat nicer.
(See Theorem 6.6.)
ρ1
6.3 The Stretch Analysis
Recall that in the case that v ∈ Ball1(u) or u ∈ Ball1(v) our algorithm returns the exact shortest
path between u = u(0) and v = v(0). Hence we next consider the situation when v 6∈ Ball1(u)
and u 6∈ Ball1(v). For brevity let d = d(0) = dG(u, v). At this point the algorithm also has
already computed u(1) and v(1), along with the shortest paths Π(u(0), u(1)) and Π(v(0), v(1)) be-
tween u(0) and u(1) and between v(0) and v(1), respectively. Observe that in this scenario we have
dG(u(0), u(1)), dG(v(0), v(1)) ≤ d, and so
dG(u(1), v(1)) ≤ dG(u(1), u(0)) + dG(u(0), v(0)) + dG(v(0), v(1)) ≤ 3 · d.
Hence if (u(1), v(1)) ∈ P1 then the path Π(u(0), u(1)) · Π(u(1), v(1)) · Π(v(1), v(0)) returned by the
algorithm is a 5-approximate path between u and v. Indeed, its length is at most
dG(u(0), u(1)) + dG(u(1), v(1)) + dG(v(1), v(0)) ≤ d + 3 · d + d = 5 · d.
More generally, suppose the query algorithm reached the j-level landmarks u(j), v(j), for some
j, 1 ≤ j ≤ h − 1, and suppose that (u(j), v(j)) 6∈ Pj. This means that v(j) 6∈ B1/3
j+1(u(j)) and
u(j) 6∈ B1/3
j+1(v(j)). By definition of the one-third-ball it follows that
and
dG(u(j), v(j)) ≥
1
3 · dG(u(j), u(j+1)) =
1
3 · rj+1(u(j)) ,
dG(u(j), v(j)) ≥
1
3 · dG(v(j), v(j+1)) =
1
3 · rj+1(v(j)) ,
15
where u(j+1) (respectively, v(j+1)) is the (j + 1)-landmark closest to u(j) (resp., v(j)).
Hence
dG(u(j+1), v(j+1)) ≤ dG(u(j+1), u(j)) + dG(u(j), v(j)) + dG(v(j), v(j+1)) ≤ 7 · dG(u(j), v(j)) .
Denote by p, 1 ≤ p ≤ h, the index for which the algorithm discovers that (u(p), v(p)) ∈ Pp. (Since
(u(h), v(h)) ∈ Ph for every pair (u(h), v(h)) of h-landmarks, it follows that the index p is well-defined.)
We have seen that dG(u(1), v(1)) ≤ 3d, and for every index j, 1 ≤ j ≤ p− 1, dG(u(j+1), v(j+1)) ≤
7 · dG(u(j), v(j)). Hence for every j, 1 ≤ j ≤ p, it holds that dG(u(j), v(j)) ≤ 3 · 7j−1 · d. Denote
d(j) = 3 · 7j−1 · d, for 0 ≤ j ≤ p. Also, dG(u(0), u(1)), dG(v(0), v(1)) ≤ d = d(0), and for every index
j, 1 ≤ j ≤ p − 1,
dG(u(j), u(j+1)) ≤ 3 · dG(u(j), v(j)) ≤ 3 · d(j) = 32 · 7j−1 · d .
Hence the length of the path
Π(u(0), u(1)) · . . . · Π(u(p−1), u(p)) · Π(u(p), v(p)) · Π(v(p), v(p−1)) · . . . Π(v(1), v(0))
returned by the algorithm is at most
d(0) + 3 · p−1
d(j)! + d(p) + 3 · p−1
Xj=1
Xj=1
d · 2 · 1 + 3 · p−1
Xj=1
d(j)! + d(0) =
3 · 7j−1!! + 3 · 7p−1! = d · (6 · 7p−1 − 1) .
Since p ≤ h we conclude that the oracle has stretch at most 6 · 7h−1 − 1.
6.4 The Size of the Oracle
We next set the values ρ1 > ρ2 > . . . > ρh so that each of the auxiliary oracles D1,D2, . . . ,Dh−1,Lh
For each index i ∈ [h], our oracle stores a forest of (vertex-disjoint) SPTs rooted at i-landmarks.
Each of these forests requires O(n) space, i.e., together these h forests require O(n · h) space.
requires O(n) space. Each of the hash tables H1,H2, . . . ,Hh associated with these oracles requires
less space than its respective oracle. Recall that the parameter ρ1 also determines the query time.
(It is O( n
λ) + O(Π), where Π the path returned by the algorithm. In the sequel we will often
ρ1
skip the additive term of O(Π) when stating the query time.)
For each i ∈ [h] we write ρi = nαi, where αi = 1 − (3/4)h−i+1. Observe that αh = 1/4, i.e.,
ρh = n1/4.
Hence IE(Lh) = ρh = n1/4, and by Chernoff's bound, whp, Lh = O(n1/4). (Recall that Lh is
a Binomial random variable.) Hence the DPPRO Lh for Ph =(cid:0)Lh
2(cid:1) requires space O(Lh4 + n) =
Next we analyze the space requirements of the oracles D1,D2, . . . ,Dh−1. Fix an index i ∈ [h−1],
and recall that the space requirement of the DPPRO Di is O(n +Branchi +Pi), where Branchi
is the set of branching events for the set Pi of pairs of vertices. Next we argue that (whp)
Branchi = O(n). Recall that the set Pi contains all pairs of i-landmarks (u(i), v(i)) such that
either v(i) ∈ B1/3
O(n), whp.
i+1(u(i)) or u(i) ∈ B1/3
i+1(v(i)).
16
The following two lemmas from [29] are the key to the analysis of the oracle's size. The first
says that with our definition of Pi+1 all branching events are confined to (i + 1)st level balls. The
second bounds the expected number of branching events in terms of the sampling probabilities.
For completeness, the proofs of these lemmas are provided in Appendix A.
Lemma 6.1 Suppose that v ∈ B1/3
i+1(u). Then if (x, y) ∈ Pi+1 and there is a branching event
between the pairs (u, v) and (x, y) then necessarily x, y ∈ Balli+1(u).
i+1(cid:17). Moreover, whp
i+1 · log3 n(cid:17), and IE(Branchi) = O(cid:16) ρ4
Lemma 6.2 Whp, Branchi = O(cid:16) ρ4
Pi = O(cid:16) ρ2
Observe that with our choice of ρi (ρi = nαi, αi = 1 − (3/4)h−i+1, for every i ∈ [h]), it holds
ρi+1(cid:17) = O(n2αi−αi+1) =
for every i ∈ [h − 1] that O(cid:16) ρ4
2 ( 3
O(n1− 1
4 )h−i). Hence by Lemma 6.2, for each i ∈ [h − 1], the oracle Di requires expected space
O(n + Branchi + Pi) = O(n). Thus the overall expected space required by our h-level oracle
oracle Λh (in addition to the space required to store the original graph G) is O(n · h). Recall that
the query time is (whp) O((n/ρ1)λ) = O(n(3/4)h · λ).
The argument described in Section 5 enables us to extend these results to general m-edge
n-vertex graphs.
i+1(cid:17) = O(n4αi−3αi+1) = O(n), and O(cid:16) ρ2
ρ3
ρ3
i
ρi+1(cid:17).
ρi+1 · log n(cid:17), and IE(Pi) = O(cid:16) ρ2
i
i
i
i
ρ3
i
Theorem 6.3 For any parameter h = 1, 2, . . . and any n-vertex undirected possibly weighted graph
G with arboricity λ, the path-reporting distance oracle Λh uses expected space O(n· h), in addition
to the space required to store G. Its stretch is (6·7h−1−1), and its query time is (whp) O(n(3/4)hλ).
The same result applies for any m-edge n-vertex graph with λ = m/n.
n · log2 n).
Specifically, in unweighted graphs with arboricity λ the query time is O((n/ρ1) · λ · log n) =
O(n(3/4)h · λ · log n), while in weighted graphs it is O(n(3/4)h · (λ + log n) log n). In unweighted
m-edge n-vertex graphs the query time is O(n(3/4)h · m
n · log n), while in m-edge n-vertex weighted
graphs it is O(n(3/4)h · m
By introducing a parameter t = (4/3)h we get query time O(n1/tλ), space O(n · log t), and
stretch at most tlog4/3 7. (The exponent is ≈ 6.76.)
Corollary 6.4 For any constant t of the form t = (4/3)h (for a positive integer h) and an n-
vertex graph G with arboricity λ, our path-reporting distance oracle Λh uses expected space O(n)
(in addition to the space needed to store G). It provides stretch at most tlog4/3 7, and its query time
is (whp) O(n1/tλ). (For a non-constant t the space requirement becomes O(n · log t).) The same
result applies for any m-edge n-vertex graph with λ = m/n.
Yet better bounds can be obtained if one is interested in small expected query time. The
expected query time is dominated by the time required to test if v ∈ Ball1(u) and if u ∈ Ball1(v).
For unweighted graphs these tests require O( n
ρ1
λ) = O(n(3/4)hλ) expected time.
Corollary 6.5 For any t of the form t = (4/3)h, for a positive integer h, and an n-vertex m-edge
graph G, our path-reporting oracle Λh uses expected O(n·h) space in addition to the space required to
store G. It provides stretch at most tlog4/3 7, and its expected query time is O(n1/t·(m/n)+log t) for
unweighted graphs. In the case of weighted graphs the expected query time is O(n1/t(m/n) · log n).
17
Consider now the oracle Λh for a superconstant number of levels h = log4/3(log n + 1). Then
ρ1 = (2n)α1 = n. In other words, all vertices V of G are now defined as the first level landmarks
(1-landmarks), i.e., L1 = V . (For levels i = 2, 3, . . . , h, landmarks Li are still selected at random
from V with probability ρi/n < 1, independently. For level 1 this probability is 1.) Recall that our
oracle starts with testing if v ∈ Ball1(u) and if u ∈ Ball1(v). Now both these balls are empty sets,
because all vertices belong to L1. Thus with this setting of parameters the oracle Λh no longer
needs to conduct this time-consuming test. Rather it proceeds directly to querying the oracle D1.
Remarkably, this variant of our oracle does not require storing the graph G. (Recall that the graph
was only used by the query algorithm for testing if v ∈ Ball1(u) and if u ∈ Ball1(v).) The query
time of the new oracle is now dominated by the h queries to the oracles D1,D2, . . . ,Dh−1,Lh, i.e.,
O(h) = O(log log n). Recall that, by the remark at the end of Section 6.2, one can always make
our oracle to return paths with at least h edges, and thus the O(h) = O(log log n) additive term
in the query time can be swallowed by O(Π), where Π is the path that our oracle returns.
6 · 7h−1 − 1 = O(loglog4/3 7 n).
Theorem 6.6 The oracle Λ is a path-reporting oracle with expected space O(n log log n), where
n is the number of vertices of its input undirected weighted graph G. Its stretch is O(loglog4/3 7 n)
and its query time is O(log log n). (It can be made O(1), but the paths returned by the oracle will
then contain Ω(log log n) edges.)
Denote by Λ the oracle which was just described. The stretch of Λ is (by Theorem 6.3)
Note that by Markov's inequality, Theorem 6.6 implies that one can produce a path-reporting
oracle with space O(n log log n), query time O(log log n) and polylogarithmic stretch by just re-
peating the above oracle-constructing algorithm for O(log n) times. Whp, in one of the executions
the oracle's space will be O(n log log n). Similarly, by the same Markov's argument, Corollary 6.4
implies that whp one can have the space of the oracle Λh bounded by O(n) (in addition to the
space required to store the input graph).
i
i
Next we analyze the construction time of our oracle. The h forests rooted at landmarks can
be constructed in O(m · h) time. We also spend O(m · n) = O(n2λ) time to compute all-pairs-
shortest-paths (henceforth, APSP). Then for each ball Bi+1(u), u ∈ Li, we store all i-landmarks
that belong to it. They can be fetched from the APSP structure in O(1) time per i-landmark.
The expected size of this data structure is O(Pi) = O( ρ2
) = O(n). Then we produce all
possible quadruples u, v, x, y with v, x, y ∈ Balli+1(u) ∩ Li, u ∈ Li. By the proof of Lemma 6.2,
there are expected O( ρ4
) = O(n) such quadruples. For each of these quadruples we check if
ρ3
i+1
the involved shortest paths intersect, and compute the corresponding branching events. Since
the length of each such path is whp O( n
ρi+1 · log n), it follows that the entire computation can be
) expected time. Recall that ρi+1 = Ω(n1/4), and thus this running time is
carried out in O( n2
ρi+1
O(n7/4). In O(n · P 2) = O(n2) additional time we construct the DPPRO Lh for the set of all
pairs of h-landmarks. The total expected construction time is therefore dominated by the APSP
computation, i.e., it is O(m · n).
ρi+1
6.5 Spanner-Based Oracles
While the query time of our oracle Λ is close to optimal (there is an additive slack of O(log log n)),
its space requirement O(n log log n) is slightly suboptimal, and also its stretch requirement is
18
O(loglog4/3 7 n), instead of the desired O(log n). Next we argue that one can get an optimal space
O(n) and optimal stretch O(log n), at the expense of increasing the query time to O(nǫ), for an
arbitrarily small constant ǫ > 0.
Given an n-vertex weighted graph G = (V, E, ω) we start with constructing an O(log n)-
spanner G′ = (V, H, ω) of G with O(n) edges. (See [7]; a faster algorithm was given in [31]. For
unweighted graphs a linear-time construction can be found in [27], and a linear-time construction
with optimal stretch-space tradeoff can be found in [20].) Then we build the oracle Λh for the
spanner G′. The space required by the oracle is (by Corollary 6.4) O(n), plus the space required
to store the spanner G′, i.e., also O(n). Hence the total space required for this spanner-based
oracle is O(n). Its stretch is the product of the stretch of the oracle, i.e., at most tlog4/3 7, with
t = (4/3)h for an integer h, and the stretch of the spanner, i.e., O(log n). Hence the oracle's stretch
is O(tlog4/3 7 · log n). The oracle reports paths in G′ = (V, H), but since H ⊆ E, these paths belong
to G as well. Observe also that the query time of the spanner-based oracle is O(n1/t · m′
n ), where
m′ = H is the number of edges in the spanner. Since m′ = O(n), it follows that the query time
is, whp, O(n1/t). We remark also that the spanners produced by [7, 31] have constant arboricity,
and thus one does not really need the reduction described in Section 5 for this result.
Theorem 6.7 For any constant ǫ > 0, the oracle obtained by invoking the oracle Λh with h =
⌈log4/3 ǫ−1⌉ from Corollary 6.4 on a linear-size O(log n)-spanner is a path-reporting oracle with
space O(n), stretch O(log n), and query time O(nǫ).
log n
log log n ≤ k ≤ log n with O(n1+1/k) edges. As a
result we obtain a path-reporting distance oracle with space O(n1+1/k), stretch O(k) and query
time O(nǫ+1/k) = O(nǫ+o(1)).
Generally, we can use an O(k)-spanner,
Observe that Theorem 6.7 exhibits an optimal (up to constant factors) tradeoff between the
stretch and the oracle size in the range
log log n ≤ k ≤ log n. The only known oracle that exhibits
this tradeoff is due to Mendel and Naor [23]. However, the oracle of [23] is not path-reporting,
while our oracle is.
log n
The construction time of this oracle consists of the time required to build the O(log n)-spanner
(which is O(n2) [31]) and the construction time of the oracle Λh in G′ (which is also O(n2), because
G′ has O(n) edges). Hence its overall construction time is O(n2).
In the context of unweighted graphs the same idea of invoking our oracle from Corollary 6.4
on a spanner can be used in conjunction with (1 + ǫ, β)-spanners. Given an unweighted n-vertex
graph G = (V, E), let G′ = (V, H) be its (1 + δ, β)-spanner, β = β(δ, k) = (cid:0) log k
, with
H = O(β · n1+1/k) edges, for a pair of parameters δ > 0, k = 1, 2, . . .. (Such a construction
was devised in [18].) For the sake of the following application one can set δ = 1.
Invoke the
distance oracle from Corollary 6.4 with a parameter t on top of this spanner. We obtain a path-
Its stretch is (O(tlog4/3 7), β = β(t, k)),
reporting distance oracle with space O(βn1+1/k) (whp).
β(t, k) = O(tlog4/3 7 · β(1, k)) = tlog4/3 7 · kO(log log k), and its query time is O(n1/t+1/k), whp. As long
log4/3 7 ), the multiplicative stretch is o(k), the additive stretch is still β(k) = kO(log log k),
as t = o(k
log4/3 7+η!, for an
while the space is O(βn1+1/k).
arbitrarily small constant η > 0, stretch (o(k), kO(log log k)), and space O(kO(log log k)n1+1/k).
In particular, one can have query time n
δ (cid:1)O(log k)
1
1
−
O k
Another variant of this construction has a higher query time O(nǫ), for some arbitrarily small
constant ǫ > 0, but its multiplicative stretch is O(1). We just set t to be a large fixed constant
19
and consider k ≫ tlog4/3 7. Then the query time is O(nǫ) whp (ǫ = t−1), stretch is (O(1), poly(1/ǫ)·
kO(log log k)), and space O(β · n1+1/k).
Theorem 6.8 For any unweighted undirected n-vertex graph G, any arbitrarily small constant
ǫ > 0 and any parameter k = 1, 2, . . ., our path-reporting distance oracle has query time O(nǫ)
(whp), stretch (O(1), β(k))) and space O(β(k) · n1+1/k) (whp), where β(k) = kO(log log k). Another
log4/3 7+η! whp, for an arbitrarily small constant η > 0,
variant of this oracle has query time n
stretch (o(k), kO(log log k)), and space O(kO(log log k) · n1+1/k) whp.
O k
1
−
To our knowledge these are the first distance oracles whose tradeoff between multiplicative
stretch and space is better than the classical tradeoff, i.e., 2k − 1 versus O(n1+1/k). Naturally,
we pay by having an additive stretch. By lower bounds from [33], an additive stretch of Ω(k) is
inevitable for such distance oracles.
One can also use a (5 + ǫ, kO(1))-spanner with O(n1+1/k) edges from [29] instead of (1 +
ǫ )O(log k))-spanner with ( log k
ǫ, ( log k
ǫ )O(log k)n1+1/k edges from [18] for our distance oracle. As a result
the oracle's space bound decreases to O(n1+1/k), its additive stretch becomes polynomial in k,
but the multiplicative stretch grows by a factor of 5 + ǫ. In general, any construction of (α, β)-
spanners with size O(S · n) can be plugged in our oracle. The resulting oracle will have stretch
(tlog4/3 7 · α, tlog4/3 7 · β), size O(Sn + n · log t), and query time O(S · n1/t).
The construction time of this oracle is the time needed to construct the (1 + ǫ, β)-spanner G′,
plus the construction of Λh on G′. The construction time of [18] is O(n2+1/k). The construction
time of the oracle Λh on G′ is O(m′ · n′), where m′ = O(β · n1+1/k) is the number of edges in G′,
and n′ = n is the number of vertices in G′. Hence the overall construction time in this case is
O(β(k) · n2+1/k) = kO(log log k)n2+1/k.
7 Lower Bounds
In this section we argue that one cannot expect to obtain distance labeling or routing schemes (see
Section 2 for their definitions) with properties analogous to those of our distance oracles (given
by Theorem 6.8 and Corollary 6.5). We also employ lower bounds of Sommer et al. [32] to show
that a distance oracle with stretch (O(1), β(k)) and space O(β(k)· n1+1/k) for unweighted n-vertex
graphs (like the distance oracle given by Theorem 6.8) must have query time Ω(k).
7.1 Distance Labeling and Routing
We start with discussing distance labeling schemes. Suppose for contradiction that there were a
t+4 ) and
distance labeling scheme D for unweighted n-vertex graphs with maximum label size O(n
stretch (t, t · β(k)), for some fixed function β(·), and any parameter k. Consider an infinite family
of n-vertex unweighted graphs Gn = (V, En) with girth at least t + 2 and En = Θ(n1+ 1
t+2 ). (Such
a family can be easily constructed by probabilistic method; see, e.g., [12], Theorem 3.7(a). Denser
t+2 ) different subgraphs of each Gn.
extremal graphs can be found in [22, 21].) There are 2Θ(n
To achieve stretch t, one would need 2Θ(n1+ 1
t+2 ) distinct encodings for these graphs, i.e., the total
1+ 1
1
20
label size for this task is Ω(n1+ 1
[33], Chapter 5, for this lower bound.)
t+2 ), and the maximum individual label size is Ω(n
1
t+2 ). (See. e.g.,
Replace every edge of G = Gn by a path of length 10t · β(k), consisting of new vertices. The
new graph G′n has N = O(n1+ 1
t+2 · t · β(k)) vertices. Invoke the distance labeling scheme D on
G′n. For a pair of original vertices u, v (vertices of Gn), the distance between them in G′n is
d′(u, v) = 10tβ(k) · dG(u, v). Given their labels ϕ(u) and ϕ(v), the labeling scheme D provides us
with an estimate δ(ϕ(u), ϕ(v)) of the distance between them in G′n which satisfies:
δ(ϕ(u), ϕ(v)) ≤ t · d′(u, v) + t · β(k) = (10tβ(k) · dG(u, v)) · t + t · β(k) .
On the other hand, a path of length dG(u, v) · t + 1 in G between u and v translates into a path
of length at most
10t · β(k)(dG(u, v) · t + 1) = 10t2β(k)dG(u, v) + 10tβ(k)
between them in G′n. Hence the estimate provided by D corresponds to a path between u and v
of length at most dG(u, v)· t in Gn, i.e., via D we obtain a t-approximate distance labeling scheme
for Gn.
The maximum label size used by D is
O(N
1
t+4 ) = O((n
t+3
t+2 · t · β(k))
1
t+4 ) = O(n
t+3
(t+2)(t+4) · (β(k))
1
t+4 ) .
However, by the above argument, this label size must be Ω(n
1
t+2 ). Note that
t+3
(t+2)(t+4) (β(k))
n
1
t+4 < n
1
t+2 ,
as long as β(k) < n. This condition holds for any constant k and fixed function β(·), and also for
any k = O(log n) and quasi-polynomial function β(·). (Recall that in all relevant upper bounds for
spanners/distance oracles/distance labeling schemes, it is always the case that k = O(log n) and
β(·) is at most a quasi-polynomial function of k.) Hence this is a contradiction, and there can be
t+4 ) and stretch (t, t· β(k)),
no distance labeling scheme for unweighted graphs with label size O(n
for any parameter k.
The same argument clearly applies to routing schemes as well. The only difference is that one
needs to use lower bounds on the tradeoff between space and multiplicative stretch for routing
due to [28, 34, 2], instead of analogous lower bounds of [33] for distance labeling.
1
To summarize, while Theorem 6.8 provides a distance oracle with stretch (t, t·β(k)) and average
space per vertex of O(β(k) · n1/k) for k ≫ tlog4/3 7, for distance labeling or routing one needs at
least nΩ(1/t) space per vertex to achieve the same stretch guarantee.
Similarly, one cannot have a distance labeling scheme for sparse graphs (graphs G = (V, E)
with O(n1+1/k) edges, for some k ≥ 1) with maximum label size O(n1/k) and stretch O(t), for a
parameter t ≪ k. 6 A distance labeling scheme as above requires maximum label size of nΩ(1/t),
as otherwise one would get a distance labeling with stretch (t, t · poly(k)) for general graphs with
maximum label size no(1/t), contradiction.
6Recall that by Corollary 6.5, a path-reporting distance oracle of total size O(n1+1/k) with stretch O(t) and
query time O(n
1
tc + 1
k + Π(u, v)) (for a query u, v; the constant c is given by c = log7 4/3) does exist.
21
7.2 Distance Oracles
Next we argue that in the cell-probe model of computation (cf., [24]), any distance oracle with size
and stretch like in Theorem 6.8 (i.e., size O(n1+1/k) and stretch (O(1), β(k)), for a fixed function
β(·)) must have query time Ω(k). We rely on the following lower bound of [32].
Theorem 7.1 [32] A distance oracle with stretch t using query time q requires space S ≥ n1+ c
in the cell-probe model with w-bit cells, even on unweighted undirected graphs with maximum degree
at most (t · q · w)O(1), where t = o(
log w+log log n ), and c is a positive constant.
log n
t·q / log n
Suppose for a contradiction that there exists a distance oracle with stretch (t, t · β(k)), for a
t·q / log n (and query
pair of parameters t ≪ k and a fixed function β(·), with space at most n1+ c/2
time q) for general unweighted graphs.
Let G = (V, E) be an n-vertex unweighted graph with maximum degree at most (t · q · w)O(1),
and let G′ be the graph obtained from G by replacing each edge of G by a path of length 10t· β(k).
The graph G′ has N ≤ (t · q · w)O(1) · β(k)· n vertices, and an oracle with stretch (t, t · β(k)) for G′
can be used also as a stretch-t oracle for G. The size of this oracle is, by our assumption, at most
(n · (t · q · w)O(1) · β(k))1+ c/2
t·q
log N
<
t·q
n1+ c/2
log n · ((t · q · w)O(1)β(k))1+ c/2
t·q .
As long as ((t · q · w)O(1) · β(k))1+ c/2
c/2
t·q , i.e., as long as
t·q < n
((t · q · w)O(1) · β(k))
2
c t·q+1 < n ,
(2)
we have a contradiction to Theorem 7.1. (As the oracle uses less than n1+ c
stretch t and query time q.)
t·q / log n space and has
For k being at most a mildly growing function of n (specifically, k ≤ logζ n, ζ < 1/2), t = o(k),
q ≤ k, w = O(log n), and β(·) being a polynomial (or even a quasi-polynomial) function, the
condition (2) holds. Hence in this range of parameters, any distance oracle for unweighted graphs
with stretch (t, t · β(k)) and query time q requires space S ≥ n1+ c/2
t·q / log n in the cell-probe model
log n
with w-bit cells, assuming t = o(
log w+log log n ).
So if this oracle uses S = O(n1+1/k · β(k)) space, then it holds that n1+1/k · log n· β(k) ≥ n1+ c/2
t·q ,
i.e.,
and so q = Ω(k/t).
1 + 1/k +
log log n + log β(k)
log n
≥ 1 +
,
c/2
t · q
We summarize this lower bound in the next theorem.
Theorem 7.2 Let k ≤ logζ n, for any constant ζ < 1/2, t = o(k), w = O(log n), and β(·) being
a polynomial or a quasi-polynomial function. In the cell-probe model with w-bit cells any distance
oracle for general unweighted undirected n-vertex graphs with space O(β(k) · n1+1/k) and stretch
(t, t · β(k)) has query time q = Ω(k/t) = Ω(k).
Theorem 7.2 states that in contrast to distance oracles with multiplicative stretch which can
have constant query time (see [23, 13]), a distance oracle with stretch (O(1), β(k)) (like the one
given by our Theorem 6.8) must have query time Ω(k).
22
Acknowledgements
The first-named author wishes to thank Ofer Neiman and Christian Wulff-Nilsen for helpful dis-
cussions, and Elad Verbin for explaining him the lower bounds from [32].
References
[1] I. Abraham and C. Gavoille. On approximate distance labels and routing schemes with affine
stretch. In DISC, pages 404 -- 415, 2011.
[2] I. Abraham, C. Gavoille, and D. Malkhi. On space-stretch trade-offs: lower bounds. In SPAA
2006: Proceedings of the 18th Annual ACM Symposium on Parallelism in Algorithms and
Architectures, Cambridge, Massachusetts, USA, July 30 - August 2, 2006, pages 207 -- 216,
2006.
[3] I. Abraham and O. Neiman. Using petal-decompositions to build a low stretch spanning tree.
In STOC, pages 395 -- 406, 2012.
[4] R. Agarwal. Personal communication, 2014.
[5] R. Agarwal and P. B. Godfrey. Distance oracles for stretch less than 2. In Proceedings of the
Twenty-Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2013, New
Orleans, Louisiana, USA, January 6-8, 2013, pages 526 -- 538, 2013.
[6] R. Agarwal, P. B. Godfrey, and S. Har-Peled. Approximate distance queries and compact
routing in sparse graphs. In INFOCOM, pages 1754 -- 1762, 2011.
[7] I. Althofer, G. Das, D. P. Dobkin, and D. Joseph. Generating sparse spanners for weighted
graphs. In SWAT, pages 26 -- 37, 1990.
[8] Y. Bartal. Probabilistic approximations of metric spaces and its algorithmic applications. In
FOCS, pages 184 -- 193, 1996.
[9] S. Baswana, A. Gaur, S. Sen, and J. Upadhyay. Distance oracles for unweighted graphs:
Breaking the quadratic barrier with constant additive error. In ICALP (1), pages 609 -- 621,
2008.
[10] S. Baswana and T. Kavitha. Faster algorithms for approximate distance oracles and all-pairs
small stretch paths. In FOCS, pages 591 -- 602, 2006.
[11] S. Baswana and S. Sen. Approximate distance oracles for unweighted graphs in expected
o(n2) time. ACM Transactions on Algorithms, 2(4):557 -- 577, 2006.
[12] B. Bollobas. Extremal Graph Theory. Springer-Verlag, 1998.
[13] S. Chechik. Approximate distance oracles with constant query time. In Symposium on Theory
of Computing, STOC 2014, New York, NY, USA, May 31 - June 03, 2014, pages 654 -- 663,
2014.
23
[14] D. Coppersmith and M. Elkin. Sparse source-wise and pair-wise distance preservers.
In
SODA: ACM-SIAM Symposium on Discrete Algorithms, pages 660 -- 669, 2005.
[15] M. Elkin. Computing almost shortest paths.
In Proc. 20th ACM Symp. on Principles of
Distributed Computing, pages 53 -- 62, 2001.
[16] M. Elkin, Y. Emek, D. A. Spielman, and S.-H. Teng. Lower-stretch spanning trees. In STOC,
pages 494 -- 503, 2005.
[17] M. Elkin, O. Neiman, and C. Wulff-Nilsen. Space-efficient path-reporting distance oracles.
CoRR, abs/1410.0768, 2014.
[18] M. Elkin and D. Peleg. Spanner constructions for general graphs. In Proc. of the 33th ACM
Symp. on Theory of Computing, pages 173 -- 182, 2001.
[19] M. Elkin and S. Pettie. A linear-size logarithmic stretch path-reporting distance oracle for
general graphs. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Dis-
crete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015, pages 805 -- 821,
2015.
[20] S. Halperin and U. Zwick. Inpublished manuscript, 2000.
[21] F. Lazebnik and V. A. Ustimenko. Explicit construction of graphs with an arbitrary large
girth and of large size. Discrete Applied Mathematics, 60(1-3):275 -- 284, 1995.
[22] A. Lubotsky, R. Phillips, and P. Sarnak. Ramanujan graphs. Combinatorica, 8:261 -- 277, 1988.
[23] M. Mendel and A. Naor. Ramsey partitions and proximity data structures. In FOCS, pages
109 -- 118, 2006.
[24] P. B. Milterson. Cell probe complexity - a survey. In Invited talk and paper in Advances in
Data Structures (Preconference workshop of FSTTCS), 1999.
[25] M. Patra¸scu and L. Roditty. Distance oracles beyond the Thorup-Zwick bound. In FOCS,
pages 815 -- 823, 2010.
[26] M. Patra¸scu , L. Roditty, and M. Thorup. A new infinity of distance oracles for sparse graphs.
In FOCS, pages 738 -- 747, 2012.
[27] D. Peleg and A. Schaffer. Graph spanners. J. Graph Theory, 13:99 -- 116, 1989.
[28] D. Peleg and E. Upfal. A tradeoff between size and efficiency for routing tables. J. of the
ACM, 36:510 -- 530, 1989.
[29] S. Pettie. Low distortion spanners. ACM Transactions on Algorithms, 6(1), 2009.
[30] E. Porat and L. Roditty. Preprocess, set, query! Algorithmica, 67(4):516 -- 528, 2013.
[31] L. Roditty, M. Thorup, and U. Zwick. Deterministic constructions of approximate distance
oracles and spanners. In ICALP, pages 261 -- 272, 2005.
24
[32] C. Sommer, E. Verbin, and W. Yu. Distance oracles for sparse graphs.
In FOCS, pages
703 -- 712, 2009.
[33] M. Thorup and U. Zwick. Approximate distance oracles. In Proc. of the 33rd ACM Symp.
on Theory of Computing, pages 183 -- 192, 2001.
[34] M. Thorup and U. Zwick. Compact routing schemes. In Proc. of the 13th Symp. on Parallelism
in Algorithms and Architectures, pages 1 -- 10, 2001.
[35] M. Thorup and U. Zwick. Spanners and emulators with sublinear distance errors. In Proc.
of Symp. on Discr. Algorithms, pages 802 -- 809, 2006.
[36] C. Wulff-Nilsen. Approximate distance oracles with improved preprocessing time. In SODA,
pages 202 -- 208, 2012.
25
Appendix
A Missing proofs
In this section we provide proofs of Lemmas 6.1 and 6.2.
Proof of Lemma 6.1: Suppose for contradiction that there exists a pair (x, y) ∈ Pi+1 such that
the pairs (u, v), (x, y) participate in a branching event β, and such that either x 6∈ Balli+1(u) or
y 6∈ Balli+1(u). Then β = (Π(u, v), Π(x, y), z), where Π(u, v) (respectively, Π(x, y)) is a shortest
path between u and v (respectively, between x and y), and z is a node at which these two paths
branch. Since (x, y) ∈ Pi+1 it follows that either y ∈ B1/3
i+1(y). Without loss of
generality suppose that y ∈ B1/3
The proof splits into two cases. In the first case we assume that x 6∈ Balli+1(u), and in the
second we assume that y 6∈ Balli+1(u). (Note that roles of x and y are not symmetric.) In both
cases we reach a contradiction.
3 · ri+1(x) and
dG(u, z) ≤ dG(u, v) < 1
3 · ri+1(u). Denote δ = dG(u, u(i+1)) = ri+1(u), where u(i+1) = ℓi+1(u).
Denote also δ′ = dG(u, x). Observe that ri+1(x) ≤ dG(x, u(i+1)) ≤ δ + δ′, and also (since x 6∈
Balli+1(u)) δ′ = dG(u, x) ≥ δ = ri+1(u). Then
We start with the case x 6∈ Balli+1(u). Observe that dG(x, z) ≤ dG(x, y) < 1
i+1(x) or x ∈ B1/3
i+1(x).
dG(u, z) + dG(z, x) <
1
3 · ri+1(u) +
1
3 · ri+1(x) ≤
δ
3
+
1
3 · (δ + δ′) ≤ δ′ = dG(u, x) .
Hence dG(u, z) + dG(z, x) < dG(u, x), contradicting the triangle inequality.
We are now left with the case that x ∈ Balli+1(u), but y 6∈ Balli+1(u). Then dG(y, z) ≤
3 · ri+1(u). In addition, ri+1(x) ≤ dG(x, u(i+1)) ≤
dG(x, y) < 1
dG(x, u) + ri+1(u) ≤ 2δ. (Note that dG(x, u) ≤ δ = ri+1(u), because x ∈ Balli+1(u).) Hence
3 · ri+1(x). Also, dG(u, z) ≤ dG(u, v) < 1
dG(u, z) + dG(z, y) <
1
3 · (ri+1(u) + ri+1(x)) ≤
1
3 · (δ + 2δ) = δ ≤ dG(u, y) .
(The last inequality is because, by an assumption, y 6∈ Balli+1(u).) This is, however, again a
contradiction to the triangle inequality.
Proof of Lemma 6.2: Recall that (see [14], Lemma 7.5) each pair (u, v), (x, y) may produce at
most two branching events. Hence next we focus on providing an upper bound on the number of
intersecting pairs of paths Π(u, v), Π(x, y) for (u, v), (x, y) ∈ Pi.
By the previous lemma, for a pair (u, v), (x, y) to create a branching event there must be one
of these four vertices (without loss of generality we call it u) such that the three other vertices
belong to Balli+1(u). Hence the number of intersecting pairs as above is at most (a constant
factor multiplied by) the number of quadruples (u, v, x, y) with v, x, y ∈ Balli+1(u). For a fixed
ρi+1 · log n(cid:17). (This
i-landmark u, the number of vertices in its (i+1)st ball Balli+1(u(i)) is, whp, O(cid:16) n
ρi+1 · log n(cid:17) i-landmarks in Balli+1(u). (We select
random variable is distributed geometrically with the parameter p = ρi+1
Balli+1(u) has probability ρi
bound, whp, there are ρi
n .) Each of the vertices in
n to belong to Li, independently of other vertices. Hence, by Chernoff's
ρi+1 · log n(cid:17) = O(cid:16) ρi
n · O(cid:16) n
i
the constant c hidden by the O-notation in O(cid:16) n
expectation is c· ρi
Hence the number of triples v, x, y of i-landmarks in Balli+1(u) is, whp, O(cid:16) ρ3
ρi+1 · log n(cid:17) to be sufficiently large. Then the
ρi+1 · log n ≥ c· log n. Hence the Chernoff's bound applies with high probability.)
i+1 · log3 n(cid:17). The
number of i-landmarks u is, by the Chernoff's bound, whp, O(ρi). Hence the number of quadruples
as above is, whp, at most
ρ3
i
O(ρi) · O(cid:18) ρ3
i+1 · log3 n(cid:19) = O(cid:18) ρ4
i+1 · log3 n(cid:19) .
ρ3
ρ3
i
i
i
Next we argue that the expected number of quadruples (u, v, x, y) of i-landmarks such that
Also, the number of pairs Pi is at most the number of i-landmarks (whp, it is O(ρi)) multiplied by
the maximum number of i-landmarks in an (i+ 1)-level ball Balli+1(u) (whp, it is O(cid:16) ρi
ρi+1 · log n(cid:17)),
ρi+1 · log n(cid:17).
i.e., Pi = O(cid:16) ρ2
v, x, y ∈ Balli+1(u) is O(cid:16) ρ4
For a fixed vertex u, write X(u) = I({u ∈ Li}) · Y (u), where Y (u) is the number of triples of
distinct i-landmarks different from u which belong to Balli+1(u), and I({u ∈ Li}) is the indicator
random variable of the event {u ∈ Li}. (Note that the ball is defined even if u 6∈ Li.) Observe
that the random variables I({u ∈ Li}) and Y (u) are independent, and thus
ρi+1(cid:17).
i+1(cid:17) and that IE(Pi) = O(cid:16) ρ2
ρ3
i
i
IE(X(u)) = IE(I({u ∈ Li})) · IE(Y (u)) =
ρi
n · IE(Y (u)) .
Let σ = (v1, v2, . . . , vn−1) be the sequence of vertices ordered by the non-decreasing distance from
u. (They appear in the order in which the Dijkstra algorithm initiated at u discovers them.) For
k = 3, 4, . . . , n − 1, denote by Jk the random variable which is equal to 0 if vk+1 is not the first
vertex in σ which belongs to Li+1. If vk+1 is the first vertex as above then Jk is equal to the
number of triples vj1, vj2, vj3, 1 ≤ j1 < j2 < j3 ≤ k such that vj1, vj2, vj3 ∈ Li. Also, for each
quadruple 1 ≤ j1 < j2 < j3 < j4 ≤ n − 1 of indices, define J(j1, j2, j3, j4) to be the indicator
random variable of the event that vj1, vj2, vj3 ∈ Li, vj4 ∈ Li+1, and for each j, 1 ≤ j < j4, the
vertex vj is not an (i + 1)-landmark. Observe that
IE(J(j1, j2, j3, j4)) =(cid:16) ρi
n(cid:17)3
·(cid:16)1 −
ρi+1
n (cid:17)j4−1
ρi+1
n
·
.
Also,
IE(Jk) = X1≤j1<j2<j3≤k
IE(J(j1, j2, j3, k + 1)) = (cid:18)k
3(cid:19)(cid:16) ρi
n(cid:17)3
·(cid:16)1 −
ρi+1
n (cid:17)k
ρi+1
n
·
.
Note that Y (u) =Pn−2
k=3 Jk, and so
IE(Y (u)) ≤
∞
Xk=3(cid:18)k
3(cid:19)(cid:16)ρi
n(cid:17)3
·(cid:16)1 −
ρi+1
n (cid:17)k
ρi+1
n
·
.
ii
Denote A = 10 n
ρi+1
. For k ≤ A, since (1 − ρi+1
3(cid:19)(cid:16) ρi
ρi+1
n(cid:17)3
n
·(cid:16)1 −
n (cid:17)k
ρi+1
·
A
Xk=3(cid:18)k
n )k = O(1), it follows that
= O(cid:18) ρ3
n4 (cid:19) A
i · ρi+1
Xk=3
k3 = O(cid:18) ρ3
i+1(cid:19) .
ρ3
i
Also,
∞
Xk=A+1(cid:18)k
·(cid:16)1 −
Denote γ = 1 − ρi+1/n. Then
3(cid:19)(cid:16) ρi
n(cid:17)3
ρi+1
n (cid:17)k
·
ρi+1
n ≤ O(cid:18)ρ3
n4 (cid:19) ·
i · ρi+1
∞
Xk=A+1
k3 ·(cid:16)1 −
ρi+1
n (cid:17)k
.
∞
Xk=A+1
k3γk ≤
d3
dγ3
∞
Xk=A+1
γk+3 ≤
d3
dγ3
1
1 − γ
=
Hence
6
(1 − γ)4 = O (cid:18) n
ρi+1(cid:19)4! .
i
i
i
∞
·
ρ3
ρ3
ρi+1
ρi+1
n
n (cid:17)k
·(cid:16)1 −
i+1(cid:19) ,
). Hence IE(X(u)) = ρi
= O(cid:18)ρ3
3(cid:19)(cid:16) ρi
n(cid:17)3
Xk=A+1(cid:18)k
and so IE(Y (u)) = O( ρ3
ρ3
i+1
n4 (cid:19) · O (cid:18) n
ρi+1(cid:19)4! = O(cid:18) ρ3
i · ρi+1
n · IE(Y (u)) = O( ρ4
i+1 · 1
n ).
Finally, the overall expected number of quadruples (u, v, x, y) of i-landmarks such that v, x, y ∈
Balli+1(u) is, by linearity of expectation, at most Pv∈V IE(X(u)) = O( ρ4
A similar argument provides an upper bound of O(cid:16) ρ2
ρi+1(cid:17) on the expected number of pairs Pi.
For a vertex u, let X′(u) = I({u ∈ Li})·Y ′(u), where Y ′(u) is the number of i-landmarks which
belong to Balli+1(u). Clearly, IE(I({u ∈ Li})) = ρi/n, and the two random variables (I({u ∈ Li})
and Y ′(u)) are independent. For every integer k ≥ 1, let J ′k be a random variable which is equal
to 0 if vk+1 is not the first vertex in σ which belongs to Li+1. Otherwise it is the number of
i-landmarks among v1, v2, . . . , vk. For integer j1, j2, 1 ≤ j1 < j2 ≤ n − 1, let J′(j1, j2) be the
indicator random variable of the event that vj1 ∈ Li, vj2 ∈ Li+1, and for every j < j2, it holds that
vj 6∈ Li+1. Then
We shortly sketch it below.
ρ3
i+1
).
i
i
IE(J′(j1, j2)) =
ρi
n ·(cid:16)1 −
ρi+1
n (cid:17)j2−1
ρi · ρi+1
n2
IE(J′(j1, k + 1)) =
IE(J ′k) =
ρi · ρi+1
n2
·
∞
Xk=1
.
ρi+1
n
·
· k ·(cid:16)1 −
k ·(cid:16)1 −
,
ρi+1
n (cid:17)k
n (cid:17)k
.
ρi+1
Hence
and
Write A = 10 n
ρi+1
IE(J ′k) = X1≤j1≤k
Xk=1
IE(Y ′(u)) ≤
, and
∞
∞
Xk=1
k(cid:16)1 −
ρi+1
n (cid:17)k
=
A
Xk=1
k(cid:16)1 −
ρi+1
n (cid:17)k
+Xk>A
k(cid:16)1 −
ρi+1
n (cid:17)k
.
iii
Each term of the first sum is O(1), and thus the first sum is at most O(A2) = O(n2/ρ2
second sum is at most d
i+1) as well. Hence
1
i+1). The
dγ Pk>A γk+1 ≤ d
dγ
IE(Y ′(u)) =
1−γ = O(n2/ρ2
· O(cid:18) n2
ρ2
ρi · ρi+1
n2
i+1(cid:19) = O(cid:18) ρi
ρi+1(cid:19) .
Hence IE(X′(u)) = O(ρ2
Pu∈V IE(X′(u)) = O(ρ2
i /(ρi+1n)), and by linearity of expectation we conclude that IE(Pi) ≤
i /ρi+1).
iv
|
1009.3594 | 2 | 1009 | 2011-08-11T17:52:44 | Center-based Clustering under Perturbation Stability | [
"cs.DS"
] | Clustering under most popular objective functions is NP-hard, even to approximate well, and so unlikely to be efficiently solvable in the worst case. Recently, Bilu and Linial \cite{Bilu09} suggested an approach aimed at bypassing this computational barrier by using properties of instances one might hope to hold in practice. In particular, they argue that instances in practice should be stable to small perturbations in the metric space and give an efficient algorithm for clustering instances of the Max-Cut problem that are stable to perturbations of size $O(n^{1/2})$. In addition, they conjecture that instances stable to as little as O(1) perturbations should be solvable in polynomial time. In this paper we prove that this conjecture is true for any center-based clustering objective (such as $k$-median, $k$-means, and $k$-center). Specifically, we show we can efficiently find the optimal clustering assuming only stability to factor-3 perturbations of the underlying metric in spaces without Steiner points, and stability to factor $2+\sqrt{3}$ perturbations for general metrics. In particular, we show for such instances that the popular Single-Linkage algorithm combined with dynamic programming will find the optimal clustering. We also present NP-hardness results under a weaker but related condition. | cs.DS | cs | Center-based Clustering under Perturbation Stability$
Pranjal Awasthi
Carnegie Mellon University, Pittsburgh, PA 15213-3891
Avrim Blum
Carnegie Mellon University, Pittsburgh, PA 15213-3891
Or Sheffet
Carnegie Mellon University, Pittsburgh, PA 15213-3891
1
1
0
2
g
u
A
1
1
]
S
D
.
s
c
[
2
v
4
9
5
3
.
9
0
0
1
:
v
i
X
r
a
Abstract
Clustering under most popular objective functions is NP-hard, even to approximate well, and so unlikely to be efficiently
solvable in the worst case. Recently, Bilu and Linial [11] suggested an approach aimed at bypassing this computational
barrier by using properties of instances one might hope to hold in practice. In particular, they argue that instances
in practice should be stable to small perturbations in the metric space and give an efficient algorithm for clustering
instances of the Max-Cut problem that are stable to perturbations of size O(n1/2). In addition, they conjecture that
instances stable to as little as O(1) perturbations should be solvable in polynomial time. In this paper we prove that
this conjecture is true for any center-based clustering objective (such as k-median, k-means, and k-center). Specifically,
we show we can efficiently find the optimal clustering assuming only stability to factor-3 perturbations of the underlying
3 perturbations for general metrics. In particular,
metric in spaces without Steiner points, and stability to factor 2 +
we show for such instances that the popular Single-Linkage algorithm combined with dynamic programming will find
the optimal clustering. We also present NP-hardness results under a weaker but related condition.
√
Keywords: Clustering, k-median, k-means, Stability Conditions
1. Introduction
Problems of clustering data arise in a wide range of
different areas -- clustering proteins by function, cluster-
ing documents by topic, and clustering images by who or
what is in them, just to name a few. In this paper we focus
on the popular class of center based clustering objectives,
such as k-median, k-center and k-means. Under these ob-
jectives we not only partition the data into k subsets, but
we also assign k special points, called the centers, one in
each cluster. The quality of a solution is then measured
as a function of the distances between the data points and
their centers. For example, in the k-median objective, the
goal is to minimize the sum of distances of all points from
their nearest center, and in the k-means objective, we min-
imize the sum of the same distances squared. As these are
NP-hard problems [17, 18, 13], there has been substantial
$This work was supported in part by the National Science Founda-
tion under grant CCF-0830540, as well as by CyLab at Carnegie Mel-
lon under grants DAAD19-02-1-0389 and W911NF-09-1-0273 from
the Army Research Office.
Email addresses: [email protected] (Pranjal Awasthi),
[email protected] (Avrim Blum), [email protected] (Or
Sheffet)
work on approximation algorithms [2, 3, 8, 12, 19, 14] with
both upper and lower bounds on approximability of these
and other objective functions. Note that we are especially
interested in the case that k is part of the input and not a
constant.
Recently, Bilu and Linial [11], focusing on the Max-
Cut problem [16], proposed considering instances where
the optimal clustering is optimal not only under the given
metric, but also under any bounded multiplicative pertur-
bation of the given metric. This is motivated by the fact
that in practice, distances between data points are typi-
cally just the result of some heuristic measure (e.g., edit-
distance between strings or Euclidean distance in some
feature space) rather than true "semantic distance" be-
tween objects. Thus, unless the optimal solution on the
given distances is correct by pure luck, it likely is correct
on small perturbations of the given distances as well. Bilu
and Linial [11] analyze Max-Cut instances of this type and
show that for instances that are stable to perturbations
of multiplicative factor roughly O(n1/2), one can retrieve
the optimal Max-Cut in polynomial time. However, they
conjecture that stability up to only constant magnitude
perturbations should be enough to solve the problem in
polynomial time. In this paper we show that this conjec-
Preprint submitted to Elsevier
October 23, 2018
ture is indeed true for k-median and k-means objectives
and in fact for any well-behaved center-based objective
function (see Definition 1.3).
1.1. Main Result
First, let us formally define the notion due to [11] of
stability under multiplicative perturbations, stated in this
context.
Definition 1.1. Given a metric (S, d), and α > 1, we say
a function d(cid:48) : S × S → R≥0 is an α-perturbation of d, if
for any x, y ∈ S it holds that
d(x, y) ≤ d(cid:48)(x, y) ≤ αd(x, y)
Note that d(cid:48) may be any non-negative function, and
need not be a metric.
Definition 1.2. Suppose we have a clustering instance
composed of n points residing in a metric (S, d) and an
objective function Φ we wish to optimize. We call the clus-
tering instance α-perturbation resilient for Φ if for any d(cid:48)
which is an α-perturbation of d, the (only) optimal cluster-
ing of (S, d(cid:48)) under Φ is identical, as a partition of points
into subsets, to the optimal clustering of (S, d) under Φ.
We will in particular be concerned with separable, center-
based clustering objectives Φ (which include k-median, k-
means, and k-center among others).
Definition 1.3. A clustering objective is center-based if
the optimal solution can be defined by k points c∗
k in
the metric space called centers such that every data point is
assigned to its nearest center. Such a clustering objective
is separable if it furthermore satisfies the following two
conditions:
1, . . . , c∗
• The objective function value of a given clustering is
either a (weighted) sum or the maximum of the in-
dividual cluster scores.
• Given a proposed single cluster, its score can be com-
puted in polynomial time.
√
Our main result is that we can efficiently find the opti-
mal clustering for perturbation-resilient instances of sepa-
rable center-based clustering objectives. In particular, we
get an efficient algorithm for 3-perturbation-resilient in-
stances when the metric S is defined only over data points,
3)-perturbation-resiliant instances for gen-
and for (2 +
eral metrics.
Theorem 1.4. For α ≥ 3 (in the case of finite metrics
defined only over the data) or α ≥ 2 +
3 (for general
metrics), there is a polynomial-time algorithm that finds
the optimal clustering of α-perturbation resilient instances
for any given separable center-based clustering objective.
√
2
The algorithm, described in Section 2.2, turns out to
be quite simple. As a first step, it runs the classic single-
linkage algorithm, but unlike the standard approach of
halting when k clusters remain, it runs the algorithm un-
til all points have been merged into a single cluster and
keeps track of the entire tree-on-clusters produced.1 Then,
the algorithm's second step is to apply dynamic program-
ming to this hierarchical clustering to identify the best
k-clustering that is present within the tree. Using a result
of Balcan et al. [6] we show that the resulting clustering
obtained is indeed the optimal one. Albeit being very dif-
ferent, our approach resembles, in spirit, the work of Bar-
tal [7], Abraham et al [1] and Racke [22] in the sense that
we reduce the problem of retrieving an optimal solution
from a general instance to a tree-like instance (where it is
poly-time solvable).
Our algorithms use only a weaker property, which we
call center-proximity (see Section 2.1), that is implied by
perturbation-resilience. We then complement these results
with a lower bound showing that for the problem of k-
median on general metrics, for any > 0, there exist NP-
hard instances that satisfy (3 − )-center proximity.2
1.2. Related work
There have been a number of investigations of differ-
ent notions of stability for the problem of clustering. For
example, Ostrovsky et al. [21] consider a k-means instance
to be stable if the optimal k-clustering is substantially
cheaper than the optimal (k − 1)-clustering under this ob-
jective. They present an efficient algorithm for finding
near-optimal k-means clusterings when this gap is large,
and these results were subsequently strengthened to apply
to smaller gaps in [4]. Balcan et al. [5] consider instead
a clustering instance to be stable if good approximations
to the given objective are guaranteed to be close, as clus-
terings, to a desired ground-truth partitioning. This is
motivated by the fact that when the true goal is to match
some unknown correct answer (e.g., to correctly cluster
proteins by their function), this is an implicit assumption
already being made when viewing approximation ratio as
a good performance measure. Balcan et al. [5] show that
in fact this condition can be used to bypass approximation
hardness results for a number of clustering objectives in-
cluding k-median and k-means. Here they show that if all
(1 + α)-approximations to the objective are δ-close to the
desired clustering in terms of how points are partitioned,
then one can efficiently get O(δ/α)-close to the desired
clustering. Ben-David et al. [10, 9] consider a notion of
1The example depicted in Figure 3 proves that indeed, halting
the Single-Linkage algorithm once k clusters are formed may fail on
certain α-perturbation resilient instances.
2We note that while our belief was that allowing Steiner points
in the lower bound was primarily a technicality, Balcan et al. (M.F.
Balcan, personal communication) have recently shown this is not the
case, giving a clever algorithm that finds the optimal clustering for
k-median instances in finite metrics when α = 1 +
√
2.
stability of a clustering algorithm, which is called stable
if it outputs similar clusters for different sets of n input
points drawn from the same distribution. For k-means,
the work of Meila [20] discusses the opposite direction --
classifying instances where an approximated solution for
k-means is close to the target clustering.
2. Proof of Main Theorem
2.1. Properties of Perturbation Resilient Instances
We begin by deriving other properties which every α-
perturbation resilient clustering instance must satisfy.
Definition 2.1. Let p ∈ S be an arbitrary point, let c∗
i be
the center p is assigned to in the optimal clustering, and let
j (cid:54)= c∗
c∗
i be any other center in the optimal clustering. We
say a clustering instance satisfies the α-center proximity
property if for any p it holds that
d(p, c∗
j ) > αd(p, c∗
i )
i and C∗
Fact 2.2. If a clustering instance satisfies the α-perturbation
resilience property, then it also satisfies the α-center prox-
imity property.
Proof. Let C∗
j be any two clusters in the opti-
mal clustering and pick any p ∈ C∗
i . Assume we blow up
all the pairwise distances within cluster C∗
i by a factor
of α. As this is a legitimate perturbation of the metric,
it still holds that the optimal clustering under this per-
turbation is the same as the original optimum. Hence, p
is still assigned to the same cluster. Furthermore, since
the distances within C∗
i were all changed by the same
constant factor, c∗
i will still remain an optimal center of
cluster i. The same holds for cluster C∗
j . It follows that
even in this perturbed metric, p prefers c∗
j . Hence
αd(p, c∗
j ) = d(p, c∗
j ).
Corollary 2.3. For every point p and its center c∗
i , and
for every point p(cid:48) from a different cluster, it follows that
d(p, p(cid:48)) > (α − 1)d(p, c∗
i ).
Proof. Denote by c∗
longs to. Now, consider two cases. Case (a): d(p(cid:48), c∗
d(p, c∗
i ).
d(p, p(cid:48)) ≥ d(p(cid:48), c∗
stable to α-perturbations, Fact 2.2 gives us that d(p(cid:48), c∗
αd(p(cid:48), c∗
≥ (α − 1)d(p, c∗
by traingle inequality we get that d(p, p(cid:48)) ≥ d(p, c∗
d(p(cid:48), c∗
j ) > αd(p, c∗
j the center of the cluster that p(cid:48) be-
j ) ≥
In this case, by traingle inequality we get that
i ). Since the data instance is
i ) >
j )−d(p, c∗
i )
i ). Again
j ) −
j ). Hence we get that d(p, p(cid:48)) > αd(p(cid:48), c∗
j ) < d(p, c∗
j ) > (α − 1)d(p, c∗
i ).
i ). Case (b): d(p(cid:48), c∗
i ) − d(p(cid:48), c∗
i ) − d(p, c∗
i ) = d(cid:48)(p, c∗
i ) < d(cid:48)(p, c∗
i to c∗
A key ingredient in the proof of Theorem 1.4 is the tree-
clustering formulation of Balcan et. al [6]. In particular,
we prove that if an instance satisfies α-center proximity for
α ≥ 3 (in the case of finite metrics without Steiner points)
or for α ≥ 2 +
3 (for general metrics) then it also sat-
isfies the "min-stability property" (defined below). The
√
3
min-stability property, as shown in [6], is sufficient (and
necessary) for the Single-Linkage algorithm to produce a
tree such that the optimal clustering is some pruning of
this tree. In order to define the "min-stability" property,
we first introduce the following notation. For any two sub-
sets A, B ⊂ S, we denote the minimum distance between
A and B as dmin(A, B) = min{d(a, b) a ∈ A, b ∈ B}.
Definition 2.4. A clustering instance satisfies the min-
stability property if for any two clusters C and C(cid:48) in the
optimal clustering, and any subset A (cid:40) C, it holds that
dmin(A, C \ A) ≤ dmin(A, C(cid:48)).
In words, the min-stability property means that for
any set A that is a strict subset of some cluster C in the
optimal clustering, the closest point to A is a point from
C \ A, and not from some other cluster. The next two
lemmas lie at the heart of our algorithm.
i , C∗
i and C∗
j respectively.
Lemma 2.5. A clustering instance in which centers must
be data points that satisfies α-center proximity for α ≥ 3
(for a center-based clustering objective), also satisfies the
min-stability property.
Proof. Let C∗
tering. Let A and A(cid:48) be any two subsets s.t. A (cid:40) C∗
A(cid:48) ⊆ C∗
obtain the minimum distance dmin(A, A(cid:48)). Let q ∈ C∗
be the nearest point to p. Also, denote by c∗
i and c∗
centers of clusters C∗
A) ≥ dmin(A, A(cid:48)). Suppose c∗
d(p, p(cid:48)) = dmin(A, A(cid:48)) ≤ dmin(A, C∗
α ≥ 3, this contradicts Corollary 2.3.
d(p, p(cid:48)) > (3−1)d(p, c∗
i ) = 2d(p, c∗
We therefore have that d(p(cid:48), c∗
3d(q, c∗
d(q, c∗
d(p(cid:48), c∗
j be any two clusters in the target clus-
i and
j . Let p ∈ A and p(cid:48) ∈ A(cid:48) be the two points which
i \ A
j the
i \
For the sake of contradiction, assume that dmin(A, C∗
i /∈ A. This means that
i ). As
i ) ≥
i )/2.
i ) ≤
i )/α <
i , p)+d(p, p(cid:48))+
i ). This contradicts Fact 2.2.
i )/2. This implies that d(p(cid:48), c∗
i )/2, and thus d(q, c∗
j ) < 3d(q, c∗
i ∈ A. It follows that d(q, c∗
i ), so d(p, c∗
i ) < d(q, c∗
i ) ≤ d(p, p(cid:48)) + d(p, c∗
Thus we may assume c∗
i \ A) ≤ d(p, c∗
i ) ≤ αd(q, c∗
j ) < d(p(cid:48), c∗
j ) ≤ d(q, c∗
i )+d(c∗
√
Lemma 2.6. A clustering instance in which centers need
not be data points that satisfies α-center proximity for
α ≥ 2 +
3 (for a center-based clustering objective), also
satisfies the min-stability property.
Proof. As in the proof of Lemma 2.5, let C∗
i , C∗
j be any
two clusters in the target clustering and let A and A(cid:48) be
j . Let p ∈ A
i and A(cid:48) ⊆ C∗
any two subsets s.t. A (cid:40) C∗
and p(cid:48) ∈ A(cid:48) be the two points which obtain the minimum
distance dmin(A, A(cid:48)) and let q ∈ C∗
i \ A be the nearest
point to p. Also, as in the proof of Lemma 2.5, let c∗
i and
i and C∗
c∗
j denote the centers of clusters C∗
j respectively
(though these need not be datapoints).
By definition of center-proximity, we have the following
inequalities:
d(p, p(cid:48)) + d(p(cid:48), c∗
d(p, p(cid:48)) + d(p, c∗
d(p, p(cid:48)) + d(p(cid:48), c∗
j ) > αd(p, c∗
i )
[c.p. applied to p]
i ) > αd(p(cid:48), c∗
j ) [c.p. applied to p(cid:48)]
j ) + d(p, q) > α(d(q, p) − d(p, c∗
i ))
[center proximity applied to q and triangle ineq.]
finite metrics that are (3−)-perturbation resilient. For ex-
ample, consider the instance shown in Figure 2. In this in-
stance, the clustering tree produced by single-linkage is not
laminar with the optimal k-median clustering. It is easy
to check that this instance is resilient to α-perturbations
for any α < 3.
Multiplying the first inequality by 1− 1
ond by 1
we get
α+1 , the third by 1
α−1 , the sec-
α−1 , and summing them together
α+1 − 1
d(p, p(cid:48)) > α2−4α+1
d(p, c∗
α−1
i ) + d(q, p),
which for α = 2 +
3 implies d(p, p(cid:48)) > d(q, p) as desired.
√
2.2. The Algorithm
As mentioned, Balcan et al [6] proved (Theorem 2)
that if an instance satisfies min-stability, then the tree on
clusters produced by the single-linkage algorithm contains
the optimal clustering as some k-pruning of it. I.e., the
tree produced by starting with n clusters of size 1 (viewed
as leaves), and at each step merging the two clusters C,C(cid:48)
minimizing dmin(C, C(cid:48)) (viewing the merged cluster as their
parent) until only one cluster remains. Given the struc-
tural results proven above, our algorithm (see Figure 1)
simply uses this clustering tree and finds the best k-pruning
using dynamic programming.
Proof of Theorem 1.4. By Lemmas 2.5 and 2.6, the data
satisfies the min-stability property, which as shown in [6]
is sufficient to guarantee that some pruning of the single-
linkage hierarchy is the target clustering. We then find the
optimal clustering using dynamic programming by exam-
ining k-partitions laminar with the single-linkage cluster-
ing tree. The optimal k-clustering of a tree-node is either
the entire subtree as one cluster (if k = 1), or the mini-
mum over all choices of k1-clusters over its left subtree and
k2-clusters over its right subtree (if k > 1). Here k1, k2 are
positive integers, such that k1 + k2 = k. Therefore, we just
traverse the tree bottom-up, recursively solving the clus-
tering problem for each tree-node. By assumption that the
clustering objective is separable, so each step including the
base-case can be performed in polynomial time. For the
case of k-median in a finite metric, for example, one can
maintain a n × O(n) table for all possible centers and all
possible clusters in the tree, yielding a running time of
O(n2 + nk2). For the case of k-means in Euclidean space,
one can compute the cost of a single cluster by computing
the center as just the average of all its points. In general,
the overall running time is O(n(k2 + T (n))), where T (n)
denotes the time it takes to compute the cost of a single
cluster.
2.3. Some Natural Barriers
We complete this section with a discussion of barriers
of our approach. First, our algorithm indeed fails on some
4
Figure 2: A finite metric k-median instance with 2 < α < 3
where our algorithm fails. The optimal 2-median clustering is
{c, p, q},{c(cid:48), p(cid:48)}. In contrast, when we run our algorithm over
on this instance, single linkage first connects {c, p} with {c(cid:48), p(cid:48)},
and only then merges these 4 points with q.
Second, observe that our analysis, though emanating
from perturbation resilience, only uses center proximity.
We next show that for general metrics, one cannot hope
to solve (in poly-time) k-median instances satisfying α-
center proximity for α < 3. This is close to our upper
bound of 2 +
3 for general metrics.
√
Theorem 2.7. For any α < 3, the problem of solving k-
median instances over general metrics that satisfy α-center
proximity is NP-hard.
Proof. The proof of Theorem 2.7 follows from the classical
reduction of Max-k-Coverage to k-median. In this reduc-
tion, we create a bipartite graph where the right-hand side
vertices represent the elements in the ground set; the left-
hand side vertices represent the given subsets; and the
distance between the set-vertex and each element-vertex
is 1, if the set contains that element. Using shortest-path
distances, it follows that the distance from any element-
vertex to a set-vertex to which it does not belong to is at
least 3. Using the fact that the NP-hardness results for
Max-k-Coverage holds for disjoint sets (i.e. the optimal
solution of Yes-instances is composed of k disjoint sets,
see [15]), the α-center proximity property follows.
Lastly, we comment that using Single-Linkage in the
usual way (namely, stopping when there are k clusters re-
maining) is not sufficient to produce a good clustering. We
demonstrate this using the example shown in Figure 3.
Observe, in this instance, since C contains significantly
less points than A,B, or D, this instance is stable -- even
if we perturb distances by a factor of 3, the cost of any al-
ternative clustering is higher than the cost of the optimal
solution. However, because d(A, C) > d(B, D), it follows
that the usual version of Single-Linkage will unite B and
D, and only then A and C. Hence, if we stop the Single-
Linkage algorithm at k = 3 clusters, we will not get the
desired clustering.
'' 1. Run Single-Linkage until only one cluster remains, producing the entire tree on clusters.
2. Find the best k-pruning of the tree by dynamic programming using the equality
best-k-pruning(T ) = min
0<k(cid:48)<k
{best-k(cid:48)-pruning(T 's left child) + best-(k − k(cid:48))-pruning(T 's right child)}
Figure 1: Algorithm to find the optimal k-clustering of instances satisfying α-center proximity. The algorithm is described for the
case (as in k-median or k-means) that Φ defines the overall score to be a sum over individual cluster scores. If it is a maximum
(as in k-center) then replace "+" with "max" above.
suitable distribution. Can this weaker notion be used to
still achieve positive guarantees?
References
[1] Ittai Abraham, Yair Bartal, T-H. Hubert Chan, Kedar Dhamd-
here Dhamdhere, Anupam Gupta, Jon Kleinberg, Ofer Neiman,
and Aleksandrs Slivkins. Metric embeddings with relaxed guar-
antees. In Proc. 46th Annual IEEE Symp. Foundations of Com-
puter Science (FOCS), 2005.
[2] Sanjeev Arora, Prabhakar Raghavan, and Satish Rao. Ap-
proximation schemes for euclidean k-medians and related prob-
lems. In Proc. 30th Annual ACM Symp. Theory of Computing
(STOC), 1998.
[3] Vijay Arya, Naveen Garg, Rohit Khandekar, Adam Meyerson,
Kamesh Munagala, and Vinayaka Pandit. Local search heuristic
for k-median and facility location problems. In Proc. 33rd ACM
Symp. Theory of Computing (STOC), 2001.
[4] Pranjal Awasthi, Avrim Blum, and Or Sheffet. Stability yields a
ptas for k-median and k-means clustering. In Proc. 51st Annual
IEEE Symp. Foundations of Computer Science (FOCS), 2010.
[5] Maria-Florina Balcan, Avrim Blum, and Anupam Gupta. Ap-
proximate clustering without the approximation. In Proc. 19th
Annual ACM-SIAM Symp. Discrete Algorithms (SODA), 2009.
[6] Maria-Florina Balcan, Avrim Blum, and Santosh Vempala.
A discriminative framework for clustering via similarity func-
tions. In Proc. 40th Annual ACM Symp. Theory of Computing
(STOC), 2008.
[7] Yair Bartal. On approximating arbitrary metrices by tree met-
rics. In Proc. 30th Annual ACM Symp. Theory of Computing
(STOC), 1998.
[8] Yair Bartal, Moses Charikar, and Danny Raz. Approximating
min-sum k-clustering in metric spaces. In Proc. 33rd Annual
ACM Symp. Theory of Computing (STOC), 2001.
[9] Shai Ben-David, D´avid P´al, and Hans-Ulrich Simon. Stability
of k-means clustering. In COLT, pages 20 -- 34, 2007.
[10] Shai Ben-David, Ulrike von Luxburg, and D´avid P´al. A sober
look at clustering stability. In G´abor Lugosi and Hans-Ulrich
Simon, editors, COLT, volume 4005 of Lecture Notes in Com-
puter Science, pages 5 -- 19. Springer, 2006.
[11] Yonatan Bilu and Nati Linial. Are stable instances easy? 1st
Symp. Innovations in Computer Science (ICS), 2010.
[12] Moses Charikar, Sudipto Guha, ´Eva Tardos, and David B.
Shmoys. A constant-factor approximation algorithm for the
k-median problem. In Proc. 31st Annual ACM Symp. Theory
of Computing (STOC), 1999.
[13] Sanjoy Dasgupta. The hardness of k-means clustering. Techni-
cal report, University of California at San Diego, 2008.
[14] W. Fernandez de la Vega, Marek Karpinski, Claire Kenyon,
and Yuval Rabani. Approximation schemes for clustering prob-
lems. In Proc. 35th Annual ACM Symp. Theory of Computing
(STOC), 2003.
[15] Uriel Feige. A threshold of ln n for approximating set cover.
JACM, 45:314 -- 318, 1998.
[16] Michael R. Garey and David S. Johnson. Computers and In-
tractability; A Guide to the Theory of NP-Completeness. W.
H. Freeman & Co., New York, NY, USA, 1990.
Figure 3: An example showing failure of the usual version of
Single-Linkage. The instance is composed of 4 components,
each with inner-distance and outer-distance as described in
the figure. However, components A, B and D each contain 100
points, whereas component C has only 10 points. The optimal
3-median clustering consists of 3 clusters: {A, C},{B},{D}
and has cost OPT = 200 + 300.
3. Open Problems
There are several natural open questions left by this
work. First, can one reduce the perturbation factor α
needed for efficient clustering? As mentioned earlier, re-
cently Balcan et al. (M.F. Balcan, personal communica-
tion) have given a very interesting algorithm that reduces
the α = 3 factor needed by our algorithm for finite met-
rics to 1 +
2. Can one go farther, perhaps by using fur-
ther implications of perturbation-resilience beyond center-
proximity? Alternatively, if one cannot find the optimal
clustering for small values of α, can one still find a near-
optimal clustering, of approximation ratio better than what
is possible on worst-case instances?
√
In a different direction, one can also consider relax-
ations of the perturbation-resilience condition. For exam-
ple, Balcan et al. (personal communication) also consider
instances that are "mostly resilient" to α-perturbations:
under any α-perturbation of the underlying metric, no
more than a δ-fraction of the points get mislabeled un-
der the optimal solution. For sufficiently large constant α
and sufficiently small constant δ, they present algorithms
that get good approximations to the objective under this
condition. A different kind of relaxation would be to con-
sider a notion of resilience to perturbations on average: a
clustering instance whose optimal clustering is likely not
to change, assuming the perturbation is random from some
5
ACDB1020100 [17] Sudipto Guha and Samir Khuller. Greedy strikes back: Im-
proved facility location algorithms. In Journal of Algorithms,
pages 649 -- 657, 1998.
[18] Kamal Jain, Mohammad Mahdian, and Amin Saberi. A new
greedy approach for facility location problems (extended ab-
stract). In Proc. 34th Annual ACM Symp. Theory of Comput-
ing (STOC), pages 731 -- 740, 2002.
[19] Amit Kumar, Yogish Sabharwal, and Sandeep Sen. A simple lin-
ear time (1+ ) -approximation algorithm for k-means clustering
in any dimensions. In Proc. 45th Annual IEEE Symp. Founda-
tions of Computer Science (FOCS), 2004.
[20] Marina Meila. The uniqueness of a good optimum for k-means.
In Proc. 23rd International Conference on Machine Learning
(ICML), pages 625 -- 632, 2006.
[21] Rafail Ostrovsky, Yuval Rabani, Leonard J. Schulman, and
Chaitanya Swamy. The effectiveness of Lloyd-type methods for
the k-means problem. In Proc. 47th Annual IEEE Symp. Foun-
dations of Computer Science (FOCS), pages 165 -- 176, 2006.
[22] Harald Racke. Optimal hierarchical decompositions for con-
gestion minimization in networks. In Proc. 40th Annual ACM
Symp. Theory of Computing (STOC), 2008.
6
|
1604.04471 | 1 | 1604 | 2016-04-15T12:54:09 | On Dynamic Job Ordering and Slot Configurations for Minimizing the Makespan Of Multiple MapReduce Jobs | [
"cs.DS",
"cs.DC"
] | MapReduce is a popular parallel computing paradigm for Big Data processing in clusters and data centers. It is observed that different job execution orders and MapReduce slot configurations for a MapReduce workload have significantly different performance with regarding to the makespan, total completion time, system utilization and other performance metrics. There are quite a few algorithms on minimizing makespan of multiple MapReduce jobs. However, these algorithms are heuristic or suboptimal. The best known algorithm for minimizing the makespan is 3-approximation by applying Johnson rule. In this paper, we propose an approach called UAAS algorithm to meet the conditions of classical Johnson model. Then we can still use Johnson model for an optimal solution. We explain how to adapt to Johnson model and provide a few key features of our proposed method. | cs.DS | cs | IEEE TRANSACTIONS ON SERVICE COMPUTING , VOL. , NO. , 2016
On Dynamic Job Ordering and Slot
Configurations for Minimizing the Makespan
1
Of Multiple MapReduce Jobs
6
1
0
2
r
p
A
5
1
]
S
D
.
s
c
[
1
v
1
7
4
4
0
.
4
0
6
1
:
v
i
X
r
a
Wenhong Tian, Guangchun Luo, Ling Tian, and Aiguo Chen
Abstract-MapReduce is a popular parallel computing paradigm for Big Data processing in clusters and data centers.
It is observed that different job execution orders and MapReduce slot configurations for a MapReduce workload have
significantly different performance with regarding to the makespan, total completion time, system utilization and other
performance metrics. There are quite a few algorithms on minimizing makespan of multiple MapReduce jobs. However,
these algorithms are heuristic or suboptimal. The best known algorithm for minimizing the makespan is 3-approximation by
applying Johnson rule. In this paper, we propose an approach called UAAS algorithm to meet the conditions of classical
Johnson model. Then we can still use Johnson model for an optimal solution. We explain how to adapt to Johnson model
and provide a few key features of our proposed method.
Index Terms-MapReduce; Hadoop; Batch Workload; Optimized Schedule; Minimized Makespan.
!
Job
scheduling
1 INTRODUCTION
With the rapid increase in size and number of
jobs that are being processed in the MapReduce
framework, efficiently scheduling multiple jobs
under this framework is becoming increasingly
important.
in MapReduce
framework brings a new challenge to Cloud
computing [1] such as minimizing the makespan,
load balancing and reduce data skew etc.,
it
has already received much attention. Originally,
MapReduce was designed for periodically running
large batch workloads with a FIFO (First-In-First-
Out) scheduler. As the number of users sharing
the same MapReduce cluster
there
are Capacity scheduler [11] and Hadoop Fair
Scheduler (HFS) [7] which intend to support more
efficient cluster sharing. There are also a few
research prototypes of Hadoop schedulers that
aim to optimize explicitly some given scheduling
metrics, e.g., FLEX [8], ARIA [4]. A MapReduce
simulator called SimMR [5] is also developed to
increased,
• Prof. Tian is in the School of Information and Software Engineering,
University of Electronic Science and Technology of China. E-mail:
tian [email protected];This research is sponsored by the Na-
tional Science Foundation of China with project ID 61450110440;
• G.Luo and L. Tian and A. Chen are with the same University.
service
simulate different workload and performance of
MapReduce. Yao et al. [15] proposed a scheme
which uses slot ratio between Map and Reduce
tasks as a tunable knob for dynamically allocating
slots. However, as pointed out in [1], the existing
schedulers do not provide a support for minimizing
the makespan for a set of jobs.
Starfish project
[9] proposes a workflow-aware
scheduler that correlate data (block) placement
with task scheduling to optimize the workflow
completion time. Zhao et al.
[16] propose a
integrating
reference
for
framework
into
scientific workflow management
systems
various cloud platforms. Moseley et al.
[10]
formulate MapReduce scheduling as a generalized
version of the classical two-stage flexible flow-shop
problem with identical machines; they provide a
12-approximation algorithm for the offline problem
of minimizing the total flow-time, which is the
sum of
the time between the arrival and the
completion of each job. Zhu et al. [15] consider
nonpreemptive case to propose 3
2-approximation
for offline scheduling regarding the makespan
where they did not considering job ordering or
applying Johnson model. In [1] and [2], the authors
propose heuristics to minimize the makespan, the
proposed algorithm called BalancedPools by
considering two pools for a Hadoop cluster. Tang
et al. [17] proposed a new algorithm called MK JR
for minimizing the makespan. The works of [1]
and [17] are closely related to our research in
minimizing the makespan. However, our present
work meets all the requirements of Johnson model
and provide optimal solution to offline scheduling
while Verma et al. [1] did not modify Johnson's
model and provided separating pools
(called
for minimizing the makespan,
BalancedPools)
and BalancedPools is a heuristic approach but not
optimal in many cases. MK JR is a 3-approximation
algorithm for minimizing the makespan. There
is still room for improving the performance of
MapReduce regarding minimize the makespan.
In summary, there is only a small number of
scheduling algorithms with regarding to minimize
the makespan of a set of MapReduce jobs in open
literature and still much room for improving the
performance of MapReduce regarding minimizing
the makespan. Therefore, we propose new model-
ing and scheduling approaches for offline jobs in
the following sections. The major contributions of
this paper include:
1) provided a new modeling and scheduling ap-
proach for multiple MapReduce jobs;
2) proposed an optimal algorithm for offline
scheduling considering Map and Reduce stages by
adapting to classical Johnson's model;
3) introduced a few key features (theorems) of our
proposed algorithm (UAAS).
2 PROBLEM FORMULATION
We consider the following problem as in [1] [17].
Let J= {J1, J2, . . . , Jn} be a set of n MapReduce jobs
with no data dependencies between them. These
jobs can be executed in any order. A MapReduce
job Ji consists of two stages, a map stage M and
reduce stage R. Each stage consists of a number of
tasks. The workload is executed on a MapReduce
cluster under FIFO scheduling by default, consist-
ing of a set of (map and reduce) slots. Let SM
and SR denote the set of map slots and reduce
slots configured by MapReduce administrator (i.e.,
S=SM U SR), so that the number of map slots and
reduce slots are SM and SR, correspondingly. Let
φ denote the job submission order for a MapReduce
workload. We consider the offline case in which
all the jobs are available at time 0. Let ci denote
the completion time of Ji (i.e., the time when Ji's
2
i
and T R
i
, T R
i
i,j and tR
and J R
We denote J M
reduce tasks all finish). The makespan for the work-
load {J1, J2, . . . , Jn} is defined as Cmax =maxi∈[n]ci.
i as the number of tasks
in Ji's map stage and reduce stage, respectively.
Let tM
i,j denote the execution time of Ji's
jth map task and jth reduce task, respectively.
Let T M
i denote the execution time of Ji's
map and reduce stage respectively. Ji requests
i × SR
i MapReduce slots and has Map and Re-
SM
duce stage durations (T M
i ) respectively. The
system scheduler can change a job's MapReduce
slots allocation depending on available resources.
We aim to determine an order (a schedule) of
execution of jobs Ji ∈ J such that the makespan
of all jobs is minimized. Let us set the actually
allocated MapReduce slots for job Ji as AM
×AR
i ,
the max available MapReduce slots in the Hadoop
cluster is SM
i . The original Johnson Rule
[3] considers that There are n items which must
go through one production stage or machine and
then a second one. There is only one machine for
each stage. At most one item can be on a machine
at a given time. We consider MapReduce as two
non-overlapped stages, i.e., map and reduce stage
respectively, the same as in [1][17]. Also we classify
all jobs into Map type and Reduce type. For Map
type jobs, their map durations should be smaller
than reduce durations while Reduce type jobs have
longer reduce durations than map durations. Based
on these assumptions and Johnson algorithm [1],
we can obtain the optimal makespan of a set of
jobs as follows:
× SR
i
i
n(cid:88)
u(cid:88)
i=1
where
Cmax =
Ku =
T R
i + maxn
u=1Ku
i − u−1(cid:88)
T M
T R
i .
(1)
(2)
i=1
i=1
the classic
Johnson algorithm for
Observation 1. If each job utilizes either all map
or all reduce slots during its processing, there
is a perfect match between the assumptions
of
two-
stage production system and MapReduce job
processing,
then Johnson's algorithm can be
applied to find optimal solution for minimizing
the makespan of a set of MapReduce jobs.
Based on our observations and intensive real test
experiences, we propose a new method called
UAAS (Utilizing All Available Slots) algorithm,
with the pesudocode given in Algorithm 2.1. The
following theorem is the key strategy for our
results.
Theorem 1. Based on available MapReduce slots
in the system,
the scheduler can increase or
decrease the number of MapReduce slots to
the job to meet the requirements of JohnSon
Rule, the result obtained by UAAS algorithm
following Johnson rule is optimal regarding to
minimize the makespan.
Proof: The original Johnson Rule [3] considers
that "there are n items which must go through one
production stage or machine and then a second one.
There is only one machine for each stage. At most
one item can be on a machine at a given time". To
adapt the MapReduce model, we treat the Map and
Reduce stage resources as a whole (like a single ma-
chine), i.e., to represent the resources as MapReduce
slots in the whole in our algorithm UAAS. USSA
algorithm allocates all available MapReduce slots
to each job at each stage, so that UAAS meets all
requirements of Johnson Rule. Since Johnson Rule
obtains optimal results with regarding to minimize
the makespan (the proof is provided in [3]), and our
UAAS algorithm meets all requirements of Johnson
Rule, therefore UAAS obtains the optimal result
with regard to minimizing the makespan. (cid:4)
input : the total number of MapReduce slots
(SM,SR) for a Hadoop cluster,
estimated all Jobs' Map and Reduce
durations (T M
available Map and Reduce slots for
each job in the cluster
i ) [1] by utilizing all
, T R
i
output: the scheduled jobs, the makespan Cmax
1 List the Map and Reduce's durations in two
vertical columns (implemented in a list) ;
2 for all Ji ∈ J do
3
4
5
6
7
8
9
10
i
i ));
, T R
Find the shortest one among all durations
(min (T M
In case of ties, for the sake of simplicity,
order the item with the smallest subscript
first. In case of a tie between Map and
Reduce, order the item according to the
Map ;
IF it is the first job of Map type, place the
corresponding item at the first place ;
ELSE it is the first job of Reduce type, place
the corresponding item at the last place ;
IF it is Map type job (and not the first job),
place the corresponding item right next to
the previous job (i.e., in non-decreasing
order of Map durations) ;
ELSE it is Reduce type job (and not the
first job), place the corresponding item left
next to the previous job (i.e., in
non-increasing order of Reduce durations) ;
Remove both durations for that job ;
Repeat these steps on the remaining set of
jobs
11 end
12 Compute the makespan (Cmax)
Algorithm 2.1: Utilizing All Available Slots
(UAAS) algorithm
BalancedPools Algorithm [1]: is way to minimize
the makespan for offline scheduling proposed
in [1], it partitions the Hadoop cluster into two
balanced pools and then allocated each job to a
suitable pool to minimize the makespan.
Observation 2. BalancedPool Algorithm does not
meet the requirement of Johnson model but just
order the job by Johnson rule and is a heuristic
algorithm with computational
complexity of
O(n2lognlogP ) where n is the number of jobs
and P is the number of MapReduce slots.
MK JR algorithm [17]: Divide the jobs set J into
Fig. 1. Five MapReduce Jobs Execution in One
Cluster by MK JR
3 THREE ALGORITHMS COMPARED
In this section, we compare UAAS algorithm with
two best known algorithms (BalancedPools and
MK JR) regarding to minimize the makespan of a
set of offline MapReduce Jobs.
3
4
23
4
5
6
30
4
510
15
20
25
30
35
1
J2
J5
J1
J4
J3
30
40
45
20X20
20X20
10X10
30X30
Time
(in
units)
map
reduce
MapReduce slots. There are five jobs, among them,
require 30×30 MapReduce slots
J1,J2 and J5
while J3 and J4 require 20×20 MapReduce slots.
The total makespan by MK JR algorithm is 47
units, visualized in Fig.1. However, if we allow
that any job can use all available MapReduce
slots in the system when execution (this can be
implemented easily in Hadoop, for example by
splitting the input files based on available number
MapReduce slots), the result is very different from
both MK JR and BalancedPools algorithms. For
the same example, in UAAS algorithm job J3 and
J4 can use all available 30 × 30 MapReduce slots,
then J3 will have Map and Reduce durations (20,
8
3), J4 will have Map and Reduce durations (4, 20)
respectively. Therefore the total makespan will be
35 2
3 as shown in Fig. 2, where X1=1. This result is
smaller (about 31.76%) than the result (47 units)
obtained by MK JR in [1]. The makespan of Pool1
and Pool2 is 39 and 40 time units respectively by
applying BalancedPools algorithm, where Pool1
has configuration of 10×10 MapReduce slots and
Pool2 has configuration of 20×20 MapReduce slots,
and J1,J2 and J5 (short jobs) are with Pool1 while
J2 and J3 (longer jobs) are with Pool2. Therefore,
the UAAS result is about 12.14% smaller than the
result (40 time units) obtained by BalancedPools
algorithm.
Theorem 2. MK JR is an 3-approximation algo-
rithm for the makespan optimization in general
case.
Proof: Applying the intermediate results from [17]
(Equ. (8) in supplementary material for proof of
THOREM 1 in [17]), we have
Cmax ≤ ( Cmax + maxn
k=1
k(cid:88)
k(cid:88)
tR
i )
tM
i + maxn
k=1
i=1
i=1
k=1
i=1
(cid:80)k
tR
i
i
i
k=1
i=1
k=1
i=1
tM
i +maxn
Cmax
(3)
where tM
and tR
is the estimated map and re-
(cid:80)k
duce duration for job Ji, respectively. Let us define
σ= maxn
, the same as in [17],
Cmax is the theoretical optimal makespan
where
given by Equ. (1)-(2). Considering the worst case
2 =C0,
that there are two jobs J1 and J2, T M
(cid:80)k
2 =1; In this case, the optimal order
and T R
tR
is J1-J2 and maxn
i =C0.
i=1
tM
And Cmax=C opt
i =C0+2 by
C0+2 ≤ 2. Therefore
UAAS algorithm, we have σ= 2C0+1
tM
i =maxn
max= maxn
1 =C0, T R
1 =1, T M
(cid:80)k
(cid:80)k
max, C opt
k=1
i=1
k=1
Fig. 2. New Result of Five MapReduce Jobs Execu-
tion by UAAS
TABLE 1
The example of 5 jobs
Job ID SM
r
30
J1
30
J2
20
J3
20
J4
30
J5
SR
r
30
30
20
20
30
i
T M
4
1
30
6
2
T R
i
5
4
4
30
3
i < T R
two disjoint sub-sets JA and JB. Set JA consists
of those jobs Ji for which T M
i . Set JB
contains the remaining jobs. Sequence jobs in JA
in non-decreasing order of T M
and those in JB
i
in non-increasing order of T R
i . The job order is
obtained by appending the sorted set JB to the
end of sorted set JA.
Observation 3. MK JR algorithm does not meet
the requirement of Johnson model but just order
the job by Johnson rule after estimating the map
and reduce durations of each job.
The reason that BalancedPools and MK JR
algorithms do not meet
the requirement of
Johnson model
lies that
they do not utilize
all available MapReduce
slots
each job
in general case,
though they estimate the job
ordering by Johnson rule. Therefore, unlike UAAS
algorithm, BalancedPools and MK JR algorithms
are suboptimal.
for
Table 1 shows an example from [2], where SM
r
and SR
is the requested number of slots for map
r
and reduce stage respectively for job Ji. Example
1. Consider a scenario shown in Table 1 from [1],
where the cluster has a configuration of 30×30
4
4
2
3
4
5
4
20
20
8/3
510
15
20
25
30
35
1m2
m5
m1
m4
m3
r2
r5
r1
r4
r3
X1
Map
Reduce
5me
(in
units)
map
reduce
≈ 3
(4)
the approximation ratio of MK JR is
Cmax(M K JR)
C opt
max
=
C0 + 2 + σ
C0 + 2
=
C0 + 2 + 2C0+1
C0+2
C0 + 2
approximation ratio.
(cid:4)
It worths notice that the worst case is applied
for
(1+σ)-appromixation
algorithm where σ ∈ [0, 2], should be called 3-
approximation algorithm since σ is 3 in the worst
case.
Based on previous results, we have the following
observation.
and MK JR
Observation 4. BalancedPools
algorithms
to
regarding
minimizing the makespan, they may not have the
minimum makespan for a set of jobs; applying
Theorem 1 to single Hadoop cluster always has
optimal total makespan for a set of jobs.
suboptimal
are
Theorem 3. Given a homogeneous environment
where the Hadoop configurations of slave nodes
are identical,
the job order φ1 produced by
UAAS for a batch of jobs are independent of
the number of slave nodes (N) but depends on
the total number of available MapReduce slots
(SM, SR), and is stable with regarding to the
change of the total number of slave nodes.
Proof: Let us set the execution durations of map
and reduce stages for a given job Ji under a given
configuration of Hadoop cluster with SM × SR
MapReduce slots, as T M
i , respectively. If the
MapReduce slots configuration of Hadoop cluster
SM
is changed to SM
x ×SR
x . Applying
SM
UAAS algorithm, the execution duration of map
and reduce stage for a given job Ji will change to
T M(cid:48)
i and T R
x and set ρ0=
and T R(cid:48)
i
i
T M(cid:48)
i = T M
i
. And we have
SM
x = T M
SM
SR
x = T R
SR
i ρ0
T R(cid:48)
i = T R
i
i ρ0
(5)
(6)
Observation 5. The the job ordering of MK JR
and BalancedPools is not stable when the total
number of slave nodes changes.
Let us consider the example given in [17]. There
is a Hadoop cluster with 5 nodes, each configured
with 2 map and 2 reduce slots. Let J1 be defined
as follows: Map stage duration is 9 and requires
10 map slots. Reduce stage duration is 10 and
requires 1 reduce slot. Let J2 be defined as follows:
Map stage duration is 11 and requires 8 map slots
and reduce stage duration is 15 and requires 1
reduce slot. In this case, the optimal job scheduling
order by UAAS is J2-J1, and their corresponding
map and reduce duration is (8.8,1.5) and (9,1)
respectively by utilizing all 10 MapReduce slots
in each stage, with the makespan of 18.8. The
job order produced by MK JR is J1-J2 with the
makespan of 35, which is about 86.17% larger
than optimal result. Now, if one node fails, then
there are only 4 nodes left with 8 map and 8
reduce tasks available in the cluster. In this case,
the optimal job scheduling by UAAS is still J2-J1,
however,
their corresponding map and reduce
duration is (11.25,1.25) and (11, 1.875) respectively
by utilizing all 8 MapReduce slots in each stage,
with makespan of 23.5. The job order generated by
MK JR keeps the same, i.e., J1-J2, with makespan
of 43, about 82.97% larger than the optimal.
Notice that BalancedPools algorithm has following
results. When there are 5 nodes, J1 with duration
(9,10) will be put into Pool1 with 2 nodes of 4
MapReduce slots and J2 will be allocated to Pool2
with 3 nodes of 6 MapReduce slots. Then J1 will
have duration (22.5, 10) and J2 will have duration
(14.67,15). If one node fail, J1 still with Pool1
and J2 with Pool2; J1 and J2 will have duration
(32.5, 10) and (37, 10) respectively. In either case,
BalancedPools is far from optimal results.
This means execution duration of map and
reduce stage for each job will change proportional
to ρ0 but their relative relationship (ordering by
their durations) will not change. Therefore the job
order of UAAS is stable with regarding to the
change of the total number of slave nodes. (cid:4)
Theorem 4. Let ρ be the ratio of map slots to
reduce slots, i.e., ρ=SM
SR . The optimal configu-
ration of ρ for makespan Cmax depends on the
total number of slots (SM,
SR), MapReduce
workload as well as its job submission order φ.
Proof:
5
[4] A. Verma, L. Cherkasova, and R. H. Campbell, ARIA: Au-
tomatic Resource Inference and Allocation for MapReduce
Environments,in Proc. of ICAC,pp.235-244, 2011,Germany.
[5] A. Verma, L. Cherkasova, and R. H. Campbell, Play It
Again, SimMR! in Proc. of Intl. IEEE Cluster'2011, pp. 253-
261, IEEE Computer Society Washington, DC, USA, 2011.
[6] M. Garey and D. Johnson, Computers and Intractability: A
Guide to the Theory of NP-completeness. WH Freeman &
Co., 1979.
[7] M. Zaharia, D. Borthakur, J. Sen Sarma, K. Elmeleegy, S.
Shenker, and I. Stoica, "Delay scheduling: A simple tech-
nique for achieving locality and fairness in cluster schedul-
ing,in Proc. of EuroSys. ACM, 2010, pp. 265-278.
J. Wolf and et al, FLEX: A Slot Allocation Scheduling Op-
timizer for MapReduce Workloads, ACM/IFIP/USENIX
Intl. Middleware Conference, Lecture Notes in Computer
Science Volume 6452, pp.1-20, 2010.
[9] H. Herodotou and S. Babu, Profiling, What-if Analysis, and
Costbased Optimization of MapReduce Programs. in Proc.
of the VLDB Endowment, pp.1111-1122, Vol. 4, No. 11,
2011.
[10] B. Moseley, A. Dasgupta, R. Kumar, and T.Sarl's, On
in Proc. of
scheduling in map-reduce and flow-shops,
SPAA, pp.289-298, ACM New York, NY, USA, 2011.
[8]
[11] Capacity
Scheduler
Guide.
[Online].
Available:
http://hadoop.apache.org/common/docs/r0.20.1/
capacity scheduler.html
[12] Y. Zheng, N. B. Shroff, P. Sinha, A New Analytical Tech-
nique for Designing Provably Efficient MapReduce Sched-
ulers, In the Proceedings of INFOCOM 2013, pp.1600-1608,
14-19 April 2013,Turin.
[13] Wordcount, http://www.cs.cornell.edu/home/llee/data/simple/
[14] http://sortbenchmark.org/YahooHadoop.pdf.
[15] Y. Zhu, Y. Jiang, W. Wu, L. Ding, A. Teredesai, D. Li,
W. Lee, Minimizing makespan and total completion time
in MapReduce-like systems, In Proceedings of INFOCOM
2014, pp.2166-2174, April 27 2014-May 2 2014, Toronto, ON.
[16] Yong Zhao ;Youfu Li ; I. Raicu ; Shiyong Lu ; Cui Lin
; Yanzhe Zhang ; Wenhong Tian ; Ruini Xue, A Ser-
vice Framework for Scientific Workflow Management in
the Cloud, IEEE Trans. Services Computing 8(6): 930-944
(2015).
[17] Shanjiang Tang, Bu-Sung Lee, Bingsheng He, Dynamic Job
Ordering and Slot Configurations for mapreduce Work-
loads, IEEE Trans. Services Computing 9(1): 4-17 (2016).
T M
i +
T R
i
f rom[17]
Cmax =
n
max
k=1
k(cid:88)
i=1
=
=
n
max
(
k=1
1
SM
1
SR n
max
(
k=1
1
ρ
n(cid:88)
i=k
tM
i,j +
n
i
max
i=k
J M
(cid:88)
k(cid:88)
k(cid:88)
(cid:88)
j=1
J M
i=1
i
n(cid:88)
i (cid:88)
i=k
J R
i (cid:88)
J R
tR
i,j)
j=1
tR
i,j) (7)
1
SR
n(cid:88)
tM
i,j +
i=1
j=1
i=k
j=1
This means the optimal configuration of ρ for
makespan Cmax depends on the total number of
slots (SM,SR) MapReduce workload (tM
i,j )
i,j, tR
as well as its job submission order φ(={1,..n}). (cid:4)
When the workload and job order are fixed,
it
is obvious that larger number of total number
of MapReduce slots will lead to smaller value of
Cmax. This is consistent with Theorem 1 and UAAS
algorithm to utilize all available MapReduce slots
(SM, SR).
4 CONCLUSION
Observing that there are quite a few algorithms on
minimizing makespan of multiple MapReduce jobs
and these algorithms are heuristic or suboptimal.
In this paper, we proposed an optimal approach
called UAAS algorithm to minimize the makespan
of a set of MapReduce jobs. The proposed algo-
rithm meets the requirements of classical Johnson
algorithm and therefore is optimal with regarding
to the makespan. We also conducted extensive tests
in real Hadoop environment to validate our theo-
retical results by benchmarks provided in [13][14].
Because this is a short paper, we do not provide the
test results yet. There are future research directions
such as considering minimizing the makespan of
online MapReduce jobs and minimizing the total
completion time and total flow time of a set of
Mapreduce jobs.
REFERENCES
[1] A. Verma, L. Cherkasova, R. H. Campbell, Orchestrating
an Ensemble of MapReduce Jobs for Minimizing Their
Makespan, IEEE Transactions on Dependable and Secure
Computing, April 2013 (online version).
[2] A. Verma, L. Cherkasova, and R. H. Campbell, Two Sides
of a Coin: Optimizing the Schedule of MapReduce Jobs
to Minimize Their Makespan and Improve Cluster Per-
formance. MASCOTS, page 11-18. IEEE Computer Society,
(2012)
S.
Johnson, Optimal Two-and Three-Stage Production
Schedules with Setup Times Included,Naval Res. Log.
Quart., 1954.
[3]
6
|
1205.1924 | 2 | 1205 | 2012-10-05T11:44:14 | Distributed Algorithms for Scheduling on Line and Tree Networks | [
"cs.DS",
"cs.DC"
] | We have a set of processors (or agents) and a set of graph networks defined over some vertex set. Each processor can access a subset of the graph networks. Each processor has a demand specified as a pair of vertices $<u, v>$, along with a profit; the processor wishes to send data between $u$ and $v$. Towards that goal, the processor needs to select a graph network accessible to it and a path connecting $u$ and $v$ within the selected network. The processor requires exclusive access to the chosen path, in order to route the data. Thus, the processors are competing for routes/channels. A feasible solution selects a subset of demands and schedules each selected demand on a graph network accessible to the processor owning the demand; the solution also specifies the paths to use for this purpose. The requirement is that for any two demands scheduled on the same graph network, their chosen paths must be edge disjoint. The goal is to output a solution having the maximum aggregate profit. Prior work has addressed the above problem in a distibuted setting for the special case where all the graph networks are simply paths (i.e, line-networks). Distributed constant factor approximation algorithms are known for this case.
The main contributions of this paper are twofold. First we design a distributed constant factor approximation algorithm for the more general case of tree-networks. The core component of our algorithm is a tree-decomposition technique, which may be of independent interest. Secondly, for the case of line-networks, we improve the known approximation guarantees by a factor of 5. Our algorithms can also handle the capacitated scenario, wherein the demands and edges have bandwidth requirements and capacities, respectively. | cs.DS | cs |
Distributed Algorithms for Scheduling on Line and Tree Networks
Venkatesan T. Chakaravarthy
Sambuddha Roy
Yogish Sabharwal
IBM Research Lab, New Delhi, India
{vechakra,sambuddha,ysabharwal}@in.ibm.com
Abstract
We have a set of processors (or agents) and a set of graph networks defined over some vertex
set. Each processor can access a subset of the graph networks. Each processor has a demand
specified as a pair of vertices hu, vi, along with a profit; the processor wishes to send data
between u and v. Towards that goal, the processor needs to select a graph network accessible to
it and a path connecting u and v within the selected network. The processor requires exclusive
access to the chosen path, in order to route the data. Thus, the processors are competing for
routes/channels. A feasible solution selects a subset of demands and schedules each selected
demand on a graph network accessible to the processor owning the demand; the solution also
specifies the paths to use for this purpose. The requirement is that for any two demands
scheduled on the same graph network, their chosen paths must be edge disjoint. The goal is to
output a solution having the maximum aggregate profit. Prior work has addressed the above
problem in a distibuted setting for the special case where all the graph networks are simply
paths (i.e, line-networks). Distributed constant factor approximation algorithms are known for
this case.
The main contributions of this paper are twofold. First we design a distributed constant
factor approximation algorithm for the more general case of tree-networks. The core compo-
nent of our algorithm is a tree-decomposition technique, which may be of independent interest.
Secondly, for the case of line-networks, we improve the known approximation guarantees by a
factor of 5. Our algorithms can also handle the capacitated scenario, wherein the demands and
edges have bandwidth requirements and capacities, respectively.
1
Introduction
Consider the following fundamental scheduling/routing problem. We have a set V consisting of n
points or vertices. A set of r undirected graphs provide communication networks over these vertices.
All the edges in the graphs provide a uniform bandwidth, say 1 unit. There are m processors (or
agents) each having access to a subset of the communication networks. Each processor P has
a demand/job a specified as a pair of vertices u and v, and a bandwidth requirement (or height)
h(a) ≤ 1. The processor P wishes to send data between u and v, and for this purpose, the processor
can use any of the networks G accessible to it. To send data over a network G, the processor P
requires a bandwidth of h(a) along some path (or route) connecting the pair of vertices u and v
in G. The input specifies a profit for each demand. A feasible solution is to select a subset of
demands and schedule each selected demand on some graph-network. For each selected demand
hu, vi scheduled on a graph-network G, the feasible solution must also specify which path connecting
u and v must be used for transmission. The following conditions must be satisfied: (i) Accessibility
1
requirement: If a demand hu, vi owned by a processor P is scheduled on a graph-network G, then
P should be able to access G; (ii) Bandwidth requirement: For any network G and for any edge e in
G, the sum of bandwidth requirements of selected demands that use the edge e must not exceed 1
unit (the bandwidth offered by the edge). We call this the throughput maximization problem1. We
shall refer to the special case of the problem wherein the heights of all demands is 1 unit as the unit
height case. In this case, we see that the paths of any two demands scheduled on the same network
should be edge disjoint. The general case wherein the heights can be arbitrary will be referred to
as the arbitrary height case.
It is known that the throughput maximization problem is NP-hard to approximate within a
factor of Ω(log1/2−ǫ n), even for the unit height case of a single graph-network [1]. Constant factor
approximations are known for special cases of the throughput maximization problem (c.f.
[10]).
Our goal in this paper is to study the problem in a distributed setting. Prior work has addressed
the problem in a distributed setting for the special case of line networks. In our paper, we present
distributed algorithms for the more general case of tree networks and also improve the known ap-
proximation ratios for the case of line networks. We first discuss the concept of line networks and
summarize the known sequential and distributed algorithms for this case.
Line-Networks: A line-network refers to a graph which is simply a path. Consider the special
case of the throughput maximization problem wherein all the graph-networks are identical paths;
say the path is 1, 2, . . . , n. We can reformulate this special case by viewing the path as a timeline.
We visualize each edge (i, i + 1) as a timeslot so that the number of timeslots is n − 1, say numbered
1, 2, . . . , n − 1; then the timeline consisting of these timeslots becomes a range [1, n − 1]. Each
demand pair hu, vi can be represented by the timeslots u, u + 1, . . . , v − 1 and can be viewed as a
interval [u, v − 1]. Thus, each demand can be assumed to be specified as an interval [s, e], where
s and e are the starting and ending timeslots. Each graph network can be viewed as a resource
offering a uniform bandwidth of 1 unit throughout the timeline. We see that a feasible solution
selects a set of demands and schedules each demand on a resource accessible to the processor
owning the demand such that for any resource and any timeslot, the sum of heights of the demands
scheduled on the resource and active at the timeslot does not exceed 1 unit. The goal is to choose
a subset of demands with the maximum throughput. See Figure 1 for an illustration.
In natural applications, a demand may specify a window [rt, dl] (release time and deadline)
where it can be executed and a processing time ρ. The job can be executed on any time segment
of length ρ contained within the window. The rest of the problem description remains the same
as above. In the new setup, apart from selecting a set of demands and determining the resources
where they must be executed, a feasible solution must also choose a execution segment for each
selected demand. As before, the accessibility and the bandwidth constraints must be satisfied. The
goal is to find a feasible solution having maximum profit.
The throughput maximization problem on line-networks has been well-studied in the realm of
classical, sequential computation. For the arbitrary height case, Bar-Noy et al. [4] presented a 5-
approximation algorithm. For the unit height case, Bar-Noy et al. [4], and independently Berman
and Dasgupta [5] presented 2-approximation algorithms; both these algorithms can also handle the
1The generalization in which the bandwidths offered by edges can vary has also been studied. For the case where
there is only one graph, this is known as the unsplittable flow problem (UFP), which has been well-studied (see survey
[12]). In this paper, we shall only consider the case where the bandwidth offered by all the edges are uniform, say 1
unit
2
0.7
A
C
0.5
B
0.4
Figure 1: Illustration for the problem on line-networks. The bandwidth/capacity offered by the
resource is 1 unit throughout the timeline. The sets of demands {A, C} and {B, C} can be scheduled
on the resource, but both A and B cannot be scheduled on the same resource.
notion of windows. Generalizations and special cases of the problem have also been studied2.
Panconesi and Sozio [15, 16] studied the throughput maximization problem on line-networks
in a distributed setting. In this setup, two processors can communicate with each other, if they
have access to some common resource. We shall assume the standard synchronous, message passing
model of computation: in a given network of processors, each processor can communicate in one
step with all other processors it is directly connected to. The running time of the algorithm is
given by the number of communication rounds. This model is universally used in the context of
distributed graph algorithms. We require that the local computation at any processor takes only
polynomial time. To be efficient, we require the communication rounds to be polylogarithmic in
the input size. We can construct a communication graph taking the processors to be the vertices
and drawing an edge between two processors, if they can communicate (i.e., they share a common
resource). Notice that the diameter of the communication graph can be as large as the number
of processors m. So, there may be a pair of processors such that the path connecting them has a
large number of hops (or edges). Hence, within the stipulated polylogarithmic number of rounds,
it would be infeasible to send information between such a pair of processors. The above fact makes
it challenging to design distributed algorithms with polylogarithmic number of rounds.
Under the above model, Panconesi and Sozio [16] designed distributed approximation algorithms
for the throughput maximization problem on line networks. For the case of unit height demands,
they presented an algorithm with an approximation ratio of (20+ ǫ) (throughout the paper, ǫ > 0 is
a constant fixed arbitrarily). For the general arbitrary height case, they devised an algorithm with
an approximation ratio of (55 + ǫ). Both the above algorithms can also handle the notion of win-
2 For the case where there is only one line-network and there are no windows, improved approximations are known
[4, 7]. The UFP problem on line-networks (where the bandwidth offered varies over the timeline) has also been well
studied (see [3, 2, 8, 10, 11]) and a constant factor approximation algorithm is known [6].
3
dows. The number of communication rounds of these algorithms is: O(cid:16) Time(MIS)
Here, Lmax and Lmin are the maximum and minimum length of any demand, and pmax and pmin
are the maximum and minimum profit of any demand. The value hmin is the minimum height
of any demand (recall that all demand heights are at most 1 unit); in the case of unit height de-
mands, hmin = 1. The value Time(MIS) is the number of rounds needed for computing a maximal
independent set (MIS) in general graphs. The randomized algorithm of Luby [14] can compute
MIS in O(log N ) rounds, where N = nmr (n, m and r are the number of timeslots, demands and
resources, respectively); if this algorithm is used, then the overall distributed algorithm would also
be randomized. Alternatively, via network-decompositions, [17] present a deterministic algorithm
with Time(MIS) = O(2√log N ).
pmin(cid:17).
log Lmax
Lmin
log pmax
ǫ·hmin
Our Contributions: In this paper, we make two important contributions. The first is that we
provide improved approximation ratios for the throughput maximization problems on line-networks
addressed by Panconesi and Sozio [16]. Secondly, we present distributed approximation algorithms
for the more general case of tree-networks. A tree-network refers to a graph which is a tree.
Notice that in a tree, the path between a pair of vertices u and v is unique and so, it suffices if the
feasible solution schedules each selected demand on a tree-network and the paths will be determined
uniquely (see Figure 2).
Prior work has addressed the throughput maximization problem for the scenario where the
input consists of a single tree-network (and all processors have access to the sole tree-network).
Under this setup, Tarjan showed that the unit height case can be solved in polynomial time [18].
Lewin-Eytan et al.
[13] presented a 5-approximation algorithm for the arbitrary height case. In
the setting of multiple tree-networks, the problem is NP-hard even for the unit height case. By
extending the algorithm of Lewin-Eytan et al., we can show that the problem can be approximated
within a factor of 3 and 8, for the unit height and arbitrary height cases, respectively.
One of the main goals of the current paper is to design distributed algorithms for the throughput
maximization problems on tree-networks. Our main result is:
Main result: We present a distributed (7 + ǫ)-approximation algorithm for the unit
height case of the throughput maximization problem on tree-networks.
The number of communication rounds is polylogarithmic in the input size: O(Time(MIS) · (1/ǫ) ·
log n · log(pmax/pmin)). Here, n is the number of vertices; pmax and pmin are the maximum and
minimum profits. Time(MIS) is the number of rounds taken for computing MIS in arbitrary graphs
with N vertices, where N = mr (m is the number of processors/demands and r is the number of
input tree-networks). As in the work of Panconesi and Sozio [16], the size of each message is O(M )
where M is the number of bits needed for encoding the information about a demand (such as its
profit, end-points and height).
Recall that Panconesi and Sozio [16] presented a distributed (20 + ǫ)-approximation algorithm
for the unit height case of the line-networks problem. The main result provides improvements over
the above work along two dimensions: the new algorithm can handle the more general concept of
tree-networks and simultaneously, it offers an improved approximation ratio.
Extending the main result, we design a distributed (80 + ǫ)-approximation algorithm for the
arbitrary height case of the tree-networks problem The number of communication rounds taken by
this algorithm is O(Time(MIS) · (1/ǫ) · (1/hmin) · log n · log(pmax/pmin)). This algorithm assumes
that the value hmin is known to all the processors. Alternatively, we assume that a value hmin is
4
Figure 2: Tree-Networks: There are three demands h1, 10i, h2, 3i and h12, 13i. In the unit height
case, only one of the three demands can be scheduled on the given tree-network (because they all
share the edge h4, 5i). To illustrate the arbitrary height case, suppose their heights are 0.4, 0.7 and
0.3, respectively. Then, the first and third demand can be scheduled together.
fixed a priori and all the demands are required to have height at least hmin.
Next, we provide a improved approximation ratios for the case of line-networks with windows.
We design distributed algorithms with approximation ratios (4 + ǫ) and (23 + ǫ), for the unit height
case and arbitrary height case, respectively 3. The number of communication rounds taken by these
algorithms is the same as that of Panconesi and Sozio [16].
Proof Techniques and Discussion: At a technical level, our paper makes two main contribu-
tions. The algorithms of Panconesi and Sozio [16], as well as our algorithms, go via the primal-dual
method (see [19]). The sequential algorithms of Bar-Noy et al. [4] and Lewin-Eytan et al. [13] use
the local ratio technique, but they can also be reformulated as primal-dual algorithms. Given a de-
mand/job, there are multiple tree-networks (or line-networks) where the demand can be scheduled
and we call each such possibility as a demand instance. All of the above algorithms work in two
phases: in the first phase, a subset of candidate demand instances are identified and an assignment
to dual variables is computed.
In the second phase, the candidate set is pruned and a feasible
solution is constructed. The dual assignment is used as a lowerbound for the optimal solution, by
3The conference version of the paper [9] claimed approximation ratios of (20 + ǫ) and (11 + ǫ) for the arbitrary
height case of tree and line networks, respectively. However, there was a minor error in analyzing the approximation
guarantee of the algorithm. The error is fixed in the current paper with an increase in the ratios
5
appealing to the weak-duality theorem. In fact, approximation algorithms for many other packing
problems utilize the above two-phase strategy.
We first formulate the above two-phase method as a framework. An important feature of the
framework is that any algorithm following the framework must produce an ordering of the demand
instances and also for each demand instance, it must determine the edges along the path whose dual
variables will be increased (or raised). The ordering and the chosen edges should satisfy a certain
property called the "interference property". The number of edges chosen, denoted ∆, is a factor in
determining the approximation ratio. In the case of line-networks, Panconesi and Sozio [16] classify
the demand instances into logarithmic many groups based on their lengths and obtain an ordering
with ∆ = 3. In the case of tree-networks, it is more challenging to design an ordering satisfying the
interference property. Towards that goal, we introduce the notion of "tree-decompositions". The
efficacy of a tree-decomposition is measured by its depth and "pivot size" θ. As it turns out, the
pivot size θ determines the parameter ∆ and the depth determines the number of rounds taken by
the algorithm. Our first main technical contribution is a tree-decomposition with depth O(log n)
and pivot size θ = 2. Using this tree-decomposition, we show how to get an ordering with ∆ = 6.
Our tree-decompositions may be of independent interest.
Another feature of the framework is that an algorithm following the framework should produce
an assignment for the dual variables in the first phase. This assignment need not form a dual feasible
solution, but it should be approximately feasible: the dual assignment divided by a parameter λ
(0 < λ ≤ 1) should yield a feasible solution. The approximation ratio is inversely related to
the parameter λ. The algorithm of Panconesi and Sozio [16] produces a dual assignment with
parameter λ = 1/(5 + ǫ). Our second main technical contribution is a method for constructing a
dual assignment with parameter λ = (1 − ǫ). Thus, we get a improved approximation ratios for the
case of line-networks.
2 Unit Height Case of Tree Networks: Problem Definition
The input consists of a vertex set V containing n vertices, a set of m processors P, a set of m
demands A and a set of r tree-networks T (each defined over the vertex-set V ). A demand a ∈ A is
specified as a pair of vertices a = (u, v) and it is associated with a profit p(a); u and v are called the
end-points of a. Each processor P ∈ P owns a unique demand a ∈ A. For each processor P ∈ P,
the input also provides a set Acc(P ) ⊆ T that specifies the set of tree-networks accessible to P . Let
pmax and pmin be the maximum and minimum profits. We will assume that all the tree-networks
are connected. Note that the tree-networks can have different sets of edges and so, they are allowed
to define different trees.
A feasible solution S selects a set of demands S ⊆ A and schedules each a ∈ S on some tree-
network T ∈ T . The feasible solution must satisfy the following properties: (i) for any a ∈ S, if a
is owned by a processor P and a is scheduled on a tree-network T , then P must be able to access
T (i.e., T ∈ Acc(P )); (ii) for any two selected demands a1 = (u1, v1) and a2 = (u2, v2), if both
a1 and a2 are scheduled on the same tree-network T , then the path between u1 and v1, and the
path between u2 and v2 in the tree-network T must be edge-disjoint (meaning, the two paths must
not share any edge). The profit of solution S is defined to be the sum of profits of the selected
demands; this is denoted p(S). The problem is to find the maximum profit feasible solution.
We next present a reformulation of the problem, which will be more convenient for our discus-
sion. Consider each demand a ∈ A and let P be the processor which owns a. For each tree-network
6
T ∈ Acc(P ), create a copy of a with the same end-points and profit; we call this the demand
instance of a belonging to the tree-network T . Let D denote the set of all demand instances over
all the demands; each demand instance d ∈ D can represented by its two end-points and the tree-
network to which it belongs. For a demand a owned by a processor P , let Inst(a) denote the set
of all instances of a (we have Inst(a) = Acc(P )). The profit of a demand instance d ∈ D is
defined to be the same as that of the demand to which it belongs; we denote this as p(d). A feasible
solution selects a subset of demand instances S ⊆ D such that: (i) for any two demand instances
d1, d2 ∈ S, if d1 and d2 belong to the same tree-network T , then their paths (in the tree-network
T ) do not share any edge; (ii) for any demand a ∈ A, at most one demand instance of a is selected.
The profit of the solution is the sum of profits of the demand instance contained in it. The goal is
to find a feasible solution of maximum profit.
The communication among the processors is governed by the following rule: two processors
P1 and P2 are allowed to communicate, if they have access to some common resource (Acc(P1) ∩
Acc(P2) 6= ∅).
Notation: The following notation will be useful in our discussion. Let E denote the set of all
edges over all the tree-networks; any edge e ∈ E is represented by a triple hu, v, T i, where u and v
are vertices of e and T is the tree-network to which e belongs. For a tree-network T , let D(T ) denote
the set of all demand instances belonging to T . Any demand instance d ∈ D(T ) can be viewed as
a path in T and we denote this as path(d). For a demand instance d ∈ D(T ) and an edge e in T ,
we say that d is active on the edge e, if the path(d) includes e; this is denoted d ∼ e. We say that
two demand instances d1 and d2 are overlapping, if d1 and d2 belong to the same tree-network, and
path(d1) and path(d2) share some edge; the demands are said to non-overlapping, otherwise. Two
demand instances d1 and d2 are said to be conflicting, if both d1 and d2 belong to the same demand
or they overlap; otherwise, the demands are said to be non-conflicting. We shall alternatively use
the term independent to mean a pair of non-conflicting demands. A set of demand instances D is
said to be independent set, if every pair of demand instances in D is independent. Notice that a
feasible solution is nothing but an independent set of demand instances.
3 LP and the Two-phase Framework
Our algorithm uses the well-known primal-dual scheme and goes via a two-phase framework. We
first present the primal and the dual LPs and then discuss the framework.
3.1 LP Formulation
The LP and its dual are presented below. For each demand instance d ∈ D, we introduce a primal
variable x(d). The first set of primal constraints capture the fact that a feasible solution cannot
select two demand instances active on the same edge. Similarly, the second set of primal constraints
capture the fact that a feasible solution can select at most one demand instance belonging to any
demand. For each demand a ∈ A and each edge e ∈ E, the dual includes a variable α(a) and β(e),
respectively. Similarly, for each demand instance d ∈ D, the dual includes a constraint; we call
this the dual constraint of d. Let ad denote the demand to which a demand instance d belongs.
7
max
x(d) · p(d)
Xd∈D
Xd∈D : d∼e
Xd∈Inst(a)
x(d) ≤ 1 (∀e ∈ E)
x(d) ≤ 1 (∀a ∈ A)
x(d) ≥ 0 (∀d ∈ D)
min Xa∈A
α(ad) + Xe : d∼e
α(a) + Xe∈E
β(e)
β(e) ≥ p(d)
(∀d ∈ D)
α(a) ≥ 0
β(e) ≥ 0
(∀a ∈ A)
(∀e ∈ E)
3.2 Two-phase framework
We formulate the ideas implicit in [16, 4, 13] in the form of a two-phase framework, described next.
Our algorithm would follow this framework.
First Phase: The procedure initializes all the dual variables α(·) and β(·) to 0 and constructs
an empty stack, and then it proceeds iteratively. Consider an iteration. Let U be the set of all
demand instances whose dual constraints are still unsatisfied. We select a suitable independent
set I ⊆ U (how to select I is clarified below). For each d ∈ I, we wish to increase (or raise) the
value of the dual variables suitably so that the dual constraint of d is satisfied tightly (i.e., the
LHS becomes equal to the RHS). For this purpose, we adopt the following strategy. Consider each
demand instance d ∈ I. We first determine the slackness s of the constraint, which is the difference
between the LHS and RHS of the constraint: s = p(d) − (α(ad) +Pe : d∼e β(e)). We next select a
suitable subset π(d) consisting of edges on which d is active (how to select π(d) is clarified below).
Next we compute the quantity δ(d) = s/(π(d) + 1). We then raise the value of α(ad) by the
amount δ(d); and for each e ∈ π(d), we raise dual variable β(e) by the amount δ(d). We see that
the dual constraint is satisfied tightly in the process. The edges π(d) are called the critical edges of
d. We say that the demand instance d is raised by the amount δ(d). Finally, the independent set I
is pushed on to the stack (as a single object). This completes an iteration. In the above framework,
in each iteration, we need to select an independent set I and the critical set of edges π(d) for each
d ∈ I. These are left as choices that must be made by the specific algorithm constructed via this
framework. Similarly, the algorithm must also decide the termination condition for the first phase.
Second Phase: We consider the independent sets in the reverse order and construct a solution
S, as follows. We initialize a set D = ∅ and proceed iteratively. In each iteration, the independent
set I on the top of the stack is popped. For each d ∈ I, we add d to D, if doing so does not violate
feasibility (namely, D ∪ {d} is an independent set). The second phase continues until the stack
becomes empty. Let S = D be the feasible solution produced by the second phase. This completes
the description of the framework.
An important aspect of the above framework is that is parallelizable. The set I chosen in each
iteration of the first phase is an independent set. Hence, for any two demand instances d1, d2 ∈ I,
the LHS of the constraints of d1 and d2 do not share any dual variable. Consequently, all the
demand instances d ∈ I can be raised simultaneously.
As we shall see, we can derive an approximation ratio for any algorithm built on the above
framework, provided it satisfies the following condition, which we call the interference property: for
any pair of overlapping demand instances d1 and d2 raised in the first phase, if d1 is raised before
8
d2, then path(d2) must include at least one of the critical edges contained in π(d1).
The following notation is useful in determining the approximation ratio. Let ξ ∈ [0, 1] be any
real number. At any stage of the algorithm, we say that a demand instance d ∈ D is ξ-satisfied, if
in the dual constraint of d, the LHS is at least ξ times the RHS: α(ad) +Pe : d∼e β(e) ≥ ξ · p(d).
If the above condition is not true, then we say that d is ξ-unsatisfied.
We shall measure the efficacy of an algorithm following the above framework using three param-
eters. (1) Critical set size ∆: Let ∆ be the maximum cardinality of π(d), over all demand instances
d raised by the algorithm. (2) Slackness parameter λ: Let λ ∈ [0, 1] be the largest number such that
at the end of the first phase, all the demand instances d ∈ D are λ-satisfied. (3) Round complexity:
The number of iterations taken by the first phase. The parameters ∆ and λ will determine the
approximation ratio of the algorithm; we would like to have ∆ to be small and λ to be close to 1.
The round complexity determines the number of rounds taken by the algorithm when implemented
in a distributed setting. We say that the algorithm is governed by the parameters ∆ and λ.
The following lemma provides an approximation guarantee for any algorithm satisfying the
interference property. The lemma is similar to Lemma 1 in the work of Panconesi and Sozio [15].
Let Opt denote the optimal solution to the input problem instance.
Lemma 3.1 Consider any algorithm satisfying the interference property and governed by parame-
ters ∆ and λ. Then the feasible solution S produced by the algorithm satisfies p(S) ≥(cid:16) λ
∆+1(cid:17)·p(Opt).
Proof: At the end of the first phase, the algorithm produces dual variable assignments α(·) and
β(·). Even though this assignment may not form a dual feasible solution, it ensures that all the
demand instances are λ-satisfied; (intuitively, all the dual constraints are approximately satisfied).
It is easy to convert the assignment hα, βi into a dual feasible solution by scaling the values by an
amount 1/λ: for each demand instance d, setbα(d) = α(d)/λ and for each edge e, set bβ(e) = β(e)/λ.
Notice that the hbα,bβi forms a feasible dual solution.
Let val(α, β) and val(bα,bβ) be the objective value of the dual assignment hα, βi and the dual
feasible solution hbα,bβi, respectively. By the weak duality theorem, val(bα,bβ) ≥ p(Opt). The scaling
process implies that val(bα,bβ) = val(α, β)/λ. We now establish a relationship between val(α, β) and
Let R denote the set of all demand instances that are raised in the first phase. The value
val(α, β) can be computed as follows. For any d ∈ R, at most ∆ + 1 dual variables are raised by an
amount δ(d) (because π(d) ≤ ∆). So, whenever a demand instance d ∈ R is raised, the objective
value raises at most (∆ + 1)δ(d). Therefore,
p(S).
val(α, β) = Xa∈A
α(a) +Xe∈E
≤ (∆ + 1)Xd∈R
β(e)
δ(d).
(1)
We next compute the profit of the solution p(S). For a pair of demand instances d1, d2 ∈ R, we
say that d1 is a predecessor of d2, if the pair is conflicting and d1 is raised before d2; in this case d2
is said to be a successor of d1. For a demand instance d ∈ R, let pred(d) and succ(d) denote the
set of predecessors and successors of d, respectively; we include d in both the sets.
Consider an element d ∈ S. We claim that
p(d) ≥ Xd′∈pred(d)
9
δ(d′).
(2)
To see this claim, consider the iteration in which d is raised. At the beginning of this iteration the
constraint of d is unsatisfied and d is raised to make the constraint tightly satisfied. The interference
property ensures the following condition: any demand instance d′ ∈ pred(d) with d′ 6= d would have
contributed a value of at least δ(d′) to the LHS of the constraint (because the property enforces
that the LHS includes at least one of raising dual variables of d′). Thus, at the beginning of the
iteration, the value of the LHS satisfies:
p(d) ≥ LHS ≥
X
d′ : d′∈pred(d) and d′6=d
δ(d′).
When d is raised, LHS increases at least by δ(d) and the constraint becomes tight. This proves the
claim.
We can now compute a lowerbound on the profit of S:
p(S) = Xd∈S
p(d)
δ(d′)
δ(d′)
≥ Xd∈S Xd′∈pred(d)
= Xd′∈R Xd∈succ(d′)∩S
= Xd′∈R
≥ Xd′∈R
δ(d′).
δ(d′) · succ(d′) ∩ S.
(3)
The second statement follows from (2) and the last statement follows from the fact that for any
d′ ∈ R, either d′ belongs to S or a successor of d′ belongs to S (this is by the construction of the
second phase).
observations made at the beginning of the proof.
Comparing (1) and (3), we see that val(α, β) ≤ (∆ + 1)p(S). The lemma follows from the
(cid:3)
A local-ratio based sequential 3-approximation algorithm for the unit height case of tree-
networks is implicit in the work of Lewin-Eytan [13]. This algorithm can be reformulated in
the two-phase framework with parameters critical set size ∆ = 2 and slackness λ = 1 (however,
the round complexity can be as high as n). We present the above algorithm in Appendix A; the
purpose is to provide a concrete exposition of the two-phase framework.
Panconesi and Sozio [16] designed a distributed algorithm for the throughput maximization
problem restricted to line-networks. In terms of the two-phase framework, their algorithm satisfies
the interference property with critical set size ∆ = 3 and slackness λ = 1/(5 + ǫ). To this end,
they partition the demand instances in to logarithmic number of groups based on their lengths,
wherein the lengths of any pair of demand instances found within the same group differ at most by
a factor of 2. Then they exploit the property that if d1 and d2 are overlapping demand instances
found within the same group , then d2 is active either at the left end-point, the right end-point
or the mid-point of d1. This way, they satisfy the interference property with ∆ = 3. We do not
know how to extend such a length-based ordering to our setting of tree-networks. Consequently,
designing an ordering satisfying the interference property with a constant ∆ turns out to be more
10
challenging. Nevertheless, we show an ordering for which ∆ = 6. Furthermore, we shall present a
method for improving the slackness parameter λ to (1 − ǫ). The notion of tree-decompositions and
layered decompositions form the core components of our algorithms.
4 Tree-Decompositions and Layered Decompositions
We first define the notion of tree-decompositions and show how to construct tree decompositions
with good parameters. Then, we show how to transform tree decompositions into layered decom-
positions.
Let H be a rooted tree defined over the vertex-set V with g as the root. For a node x, define
its depth to be the number of nodes along the path from g to x; the root g itself is defined to have
to depth 1. With respect to H, a node y is said to be an ancestor of x, if y appears along the path
from g to x; in this case, x is said to be a descendent of y. By convention, we do not consider x to
be an ancestor or descendent of itself. For a node z in H, let C(z) be the set consisting of z and
its descendents in H.
4.1 Tree-decomposition: Definition
Let T ∈ T be a tree-network defined over the input vertex-set V consisting of n vertices. A subset
of nodes C ⊆ V is called a component, if C induces a (connected) subtree in T . We say that a
node x ∈ V − C is a neighbor of C, if x is adjacent to some node in C. Let Γ[C] denote the set
of neighbors (or neighborhood) of C. Notice that for any two nodes x ∈ C and y 6∈ C, the path
between x and y must pass through some node in the neighborhood Γ[C].
Let T be a tree-network and H be a rooted-tree defined over V with g as the root. We say
that H is a tree decomposition for T , if the following conditions are satisfied: (i) for any demand
instance d ∈ D(T ), if d passes through nodes x and y then d also passes through LCA(x, y), which
is the least common ancestor of x and y in H; (ii) for any node z in H, C(z) forms a component
in T .
For a node z ∈ H, let χ(z) denote the set of neighbors of the component C(z), i.e., Γ[C(z)].
We call χ(z) the pivot set of z. Clearly, for any nodes x ∈ C(z) and y 6∈ C(z), the path between
x and y in T must pass through one of the nodes in χ(z). We shall measure the efficacy of a tree
decomposition H using two parameters: (i) pivot size θ:
this is the maximum cardinality of χ(z)
over all z ∈ V ; (ii) the depth of the tree.
See Figure 3 for an illustration. This figure shows an example tree-decomposition for the tree-
network shown in Figure 6. The demand instance h4, 13i passes through nodes 2 and 8; it also
passes through LCA(2, 8) = 5. For the node 2, the component C(2) = {2, 4}; its pivot set is
χ(2) = {1, 5}. On the other hand, C(5) = {5, 9, 8, 2, 12, 13, 4} and its pivot set is χ(5) = {1}. This
tree-decomposition has depth 4 and pivot set size θ = 2.
We note that it is not difficult to design tree-decompositions with parameters hdepth = n, θ = 1i
or hdepth = log n, θ = log ni. As it turns out the depth of the tree-decomposition will determine the
number of rounds, whereas the pivot size θ will determine the approximation ratio. Thus, neither
of these two tree-decompositions would yield an algorithm that runs in polylogarithmic number of
rounds, while achieving a constant factor approximation ratio. Our main contribution is a tree-
decomposition with parameters hdepth = 2 log n, θ = 2i (we call this the ideal tree-decomposition).
Interestingly, the ideal tree-decomposition builds on the two simpler tree-decompositions mentioned
11
1
5
8
9
2
11
12
13
4
7
3
6
10
14
Figure 3: Tree-decompositions: Illustration.
above. For the sake of completeness, the two simpler tree-decompositions are discussed next.
4.2 Two Simple Tree-decompositions
Here, we present two tree decompositions called root-fixing tree decomposition and balancing tree
decomposition. The first decomposition has pivot size θ = 1, but its depth can be as high as n.
The second decomposition has depth log n, but its pivot size θ can be as high as log n.
Root-fixing Decomposition: Let T ∈ T be any input tree-network. Convert T into a rooted-
tree by arbitrarily picking a node g ∈ V as the root; let the resulting rooted-tree be H. It is easy
to see that H is a tree decomposition for T . Consider any node z and let z′ be its parent in H; let
C(z) be the descendants of z including z itself. Notice that for any x ∈ C(z) and y 6∈ C(z), the
path between x and z must pass through the parent z′. Thus the component C(z) has only one
neighbor. We see that H has pivot size θ = 1; however, the depth of H can be as high as n. Figure
6 shows a root-fixing decomposition; the chosen root is node 1. The sequential algorithm given in
Section A implicitly uses the root-fixing tree decomposition.
Balancing tree decomposition: Let T ∈ T be a tree-network. Consider a component C ⊆ V
and let T (C) be the (connected) subtree induced by C. Let z be a node in C. If we delete the
node z from T (C), the tree T (C) splits into subtrees T1, T2, . . . , Ts (for some s). Let C1, C2, . . . , Cs
be the vertex-set of these subtrees. Every node in C − {z} is found in some component Ci. We
say that the node z splits C into components C1, C2, . . . Cs. The node z is said to be a balancer for
C, if for all 1 ≤ i ≤ s, Ci ≤ ⌊C/2⌋. The following observation is easy to prove: any component
C ⊆ V contains a balancer z.
Our procedure for constructing the tree decomposition for T works recursively by calling a pro-
cedure BuildBalTD (build balanced tree decomposition). The procedure takes as input a component
C ⊆ V and outputs a rooted-tree having C as the vertex-set. It works as follows. Given a compo-
12
nent C, find a balancer z for C. Then split C by z and obtain components C1, C2, . . . , Cs (for some
s). Each component Ci has size at most ⌊C/2⌋. For 1 ≤ i ≤ s, call the procedure BuildBalTD
recursively on the component Ci and obtain a tree Hi with gi as the root. Construct a tree H by
making z as the root and g1, g2, . . . , gs as its children. Return the tree H.
Given a tree-network T , we obtain a rooted-tree H by calling BuildBalTD with the whole
vertex-set V as the input. It is easy to see that for any node z, C(z) forms a component in T . For
any node z in H with children z1, z2, . . . , zs (for some s), C(z1), C(z2), . . . , C(zs) are nothing but
the components obtained by splitting C(z) by z. This implies that H satisfies the first property
of tree decompositions. Since the size of the input component drops by a factor of two in each
iteration, the depth of H is at most ⌈log n⌉. Consider any node z in H and let C(z) be the set
consisting of descendants of z and z itself. Observe that for any node x ∈ C(z) and y 6∈ C(z),
the path between x and y must pass through one of the ancestors of z in H (because of the first
property of tree decompositions). In other words, the neighborhood of C(z) is contained within
the set of ancestors of z. The number of ancestors is at most ⌈log n⌉ and hence, the pivot size of
H is at most O(log n). Figure 3 shows an example balancing tree-decomposition for the tree given
in Figure 6.
4.3
Ideal Tree-decomposition
In this section, we present the ideal tree-decomposition with parameters hdepth = 2 log n, θ = 2i.
The ideal tree-decomposition also goes via the notion of balancers. Recall that any component
C ⊆ V contains a balancer z.
Fix a tree-network T and we shall construct an ideal tree decomposition H for T with pivot set
size θ = 2 and depth O(log n). Intuitively, the tree H will be constructed recursively. In each level
of the recursion, we will add two nodes to the tree: a balancer and a node that we call a junction.
The output tree-decomposition will have depth at most 2⌈log n⌉.
The construction works via a recursive procedure BuildIdealTD (build ideal tree decomposition).
The procedure BuildIdealTD takes as input a set C ⊆ V forming a component in T . As a
precondition, it requires the component C to satisfy the important property that C has at most
two neighbors in T . It outputs a rooted-tree H with C as the vertex set having depth at most
2⌈log C⌉ such that for any node x ∈ C, the number of neighbors of C(x) is at most 2, where C(x)
is the set consisting of x and its descendants in H.
The procedure BuildIdealTD works as follows. We first find a balancer z for the component C.
The node z splits C into components C1, C2, . . . , Cs. We shall consider two cases based on whether
C has a single neighbor or two neighbors.
Case 1: This is the easier case where C has only one neighbor, say u1. See Figure 4. For this
case, ignore the nodes u2 and u′2. Let u′1 be the node in C which is adjacent to u1 and without loss
of generality, assume that C1 is the component to which u′1 belongs. Observe that Γ(C1) = {u1, z}
and for all i ≥ 2, Γ(Ci) = {z}. In other words, all the components Ci have at most two neighbors.
That is, they all satisfy the precondition set by the procedure. For each 1 ≤ i ≤ s, we recursively
call the procedure BuildIdealTD on the component Ci and obtain a tree Hi with gi as the root. We
construct a tree H by making z as the root and g1, g2, . . . , gs as its children. Then, the rooted-tree
H is returned.
Case 2: Now consider the case where C has two neighbors, say u1 and u2. Let u′1 and u′2 be
the nodes in C which are neighbors of u1 and u2, respectively. We consider two subcases.
13
Figure 4: Illustration for ideal tree-decomposition.
Case 2(a): The first subcase is when u′1 and u′2 lie in two different components, say C1 and
C2, respectively. See Figure 4. Observe that Γ(C1) = {u1, z}, Γ(C2) = {u2, z} and for all i ≥ 3,
Γ(Ci) = {z}. Hence all the components Ci satisfy the precondition set by the procedure. For
each 1 ≤ i ≤ s, we call the procedure BuildIdealTD with Ci as input and obtain a tree Hi. We
construct a tree H by making the balancer z as the root and g1, g2, . . . , gs as its children. Then,
the rooted-tree H is returned.
Case 2(b): Now consider the second and comparatively more involved subcase wherein u′1 and
u′2 belong to the same component, say C1. See Figure 5. Observe that there exists a unique node
j ∈ C1 such that all the three paths u1 u2, u1 z, and u2 z pass through j. We call j
as the junction. Spilt the component C1 by the node j to obtain components C′1, C′2, . . . , C′s′ (for
some s′). Observe that among C′1, C′2, . . . , C′s′, there exists three distinct components such that z
is a neighbor of the first component, and u′1 and u′2 belong to the other two components; without
loss of generality, let these components be C′1, C′2 and C′3, respectively. We see that for 2 ≤ i ≤ s,
Γ(Ci) = {z}; moreover, Γ(C′1) = {j, z}, Γ(C′2) = {u1, j}, Γ(C′3) = {u2, j} and for 4 ≤ i ≤ s′,
Γ(C′i) = {j}. Thus, all the components C2, C3, . . . , Cs and C′1, C′2, . . . , C′s′ satisfy the precondition
set by the procedure. For each 2 ≤ i ≤ s, we call the procedure BuildIdealTD recursively with
Ci as input and obtain a tree Hi with gi as the root. For each 1 ≤ i ≤ s′, we call the procedure
BuildIdealTD recursively with C′i as input and obtain a tree H′i with g′i as the root. Construct a
tree H as follows. Make the junction j as the root; make g′2, g′3, . . . , g′s′ as the children of j; make
z as a child of j; make g′1 and g2, g3, . . . , gs as the children of z. Return the rooted-tree H. This
completes the description of the procedure BuildIdealTD.
By induction, we can argue that BuildIdealTD satisfies the intended property: for any node
x ∈ C, the number of neighbors of C(x) is at most 2. As an example, consider the subcase in which
u′1 and u′2 belong to the same component (the case where a junction j is created). The procedure
creates only two nodes j and z on its own and the rest of the nodes in H are created by the recursive
calls. Consider the node j. It is guaranteed that the input component C has at most two neighbors
(this is the precondition set by the procedure). Since C(j) = C, we see that j satisfies the property.
Now, consider the node z. The component C(z) is the union of C2, C3, . . . , Cs and C′1. We have
that Γ[C(z)] = {j}. Thus, z also satisfies the property. The rest of the nodes satisfy the property
14
Figure 5: Illustration for Case 2(b) of ideal tree-decomposition
by induction.
Let us now analyze the depth of the tree H output by the procedure. Since z is a balancer
for C, the components C1, C2, . . . , Cs have size at most ⌊C/2⌋. Moreover, since C′1, C′2, . . . , C′s′ are
subsets of C1, these components also have size at most ⌊C/2⌋. Thus, all the components input to
the recursive calls have size at most ⌊C/2⌋. Thus, by induction, H has depth at most 2⌈log C⌉.
We next show how to construct a tree decomposition H for the tree-network T . First, find
a balancer g for the entire vertex-set V and split V into components C1, C2, . . . , Cs. For each
component Ci, Γ[Ci] = {g}. For each 1 ≤ i ≤ s, call the procedure BuildIdealTD with Ci as input
and obtain a tree Hi with gi as the root. Construct a tree H by making g as the root and each gi
as its children. Return H.
We can argue that for any node z in H, C(z) forms a component in T . Furthermore, for any
node z in H with children z1, z2, . . . , zs (for some s), C(z1), C(z2), . . . , C(zs) are nothing but the
components obtained by splitting C(z) by z. This implies that H satisfies the first property of
tree decompositions. It follows that H is indeed a tree decomposition. The depth of H is at most
2⌈log n⌉. The properties of the BuildIdealTD procedure ensure that the pivot size of H is at most
2. We have the following result
Lemma 4.1 For any tree-network T ∈ T , there exists a tree decomposition H (called the ideal tree
decomposition) with depth O(log n) and pivot size θ = 2.
15
4.4 Layered Decompositions
In this section, we define the notion of layered decompositions and show how to transform tree
decompositions into layered decompositions.
Let T ∈ T be a tree-network. A layered decomposition of T is a pair σ and π, where σ is a
partitioning of D(T ) into a sequence of groups G1, G2, . . . , Gℓ and π maps each demand instance
d ∈ D(T ) to a subset of edges in path(d). The following property should be satisfied:
for any
1 ≤ i ≤ j ≤ ℓ and for any pair of demand instances d1 ∈ Gi and d2 ∈ Gj, if d1 and d2 are
overlapping, then path(d2) should include at least one of the edges in π(d1). The edges in π(d) are
called the critical edges of d. The value ℓ is called the length (or depth) of the decomposition.
Notice that similarity between the inference property and the notion of layered decompositions.
We shall measure the efficacy of a layered decomposition by two parameters: (i) Critical set size
∆ - this is the maximum cardinality of π(d) over all demand instances d ∈ D(T ); (ii) the length ℓ
of the sequence. Our goal is to construct a layered decomposition with length O(log n) and critical
set size ∆ = 6. Towards that goal we shall show how to transform tree-decompositions into layered
decompositions. The following notations are useful for this purpose.
Let T ∈ T be tree-network and H be a tree-decomposition for T with pivot size θ and depth ℓ.
For a demand instance d, let µ(d) be the node with the least depth in H among all the nodes that
path(d) passes through. The first property of tree decompositions ensure that µ(d) is unique. We
say that d is captured at µ(d). See Figure 3. In this figure, the demand h4, 13i is captured at node
5. Let d ∈ D(T ) be a demand instance and u be a node in T . Observe that there exists a unique
node y belonging to path(d) such that the path from u to y does not pass through any other node
in path(d). We call y as the bending point of d with respect to u. For a node y in path(d), we call
the edges on path(d) adjacent to y as the wings of y on path(d). If y is an end-point of d, there
will be only one wing; otherwise, there will be two wings. See Figure 6. In this figure, with respect
to nodes 3 and 9, the bending points of the demand d = h4, 13i are 2 and 5, respectively. With
respect to path(d), node 4 has only one wing h4, 2i, while node 8 has two wings h5, 8i and h8, 13i.
Lemma 4.2 shows how to transform a tree-decomposition into a layered decomposition. The
lemma is proved by categorizing the demand instances into groups G1, G2, . . . , Gℓ, where Gi consists
of all demand instances captured at a node with depth ℓ − i + 1. For a demand instance d ∈ D(T ),
let z = µ(d). The set π(d) is constructed as follows: (i) we include the wings of z on path(d); (ii)
for each neighbor u of C(z), taking y to be the bending point of path(d) with respect to u, we
include the wings of y on path(d).
Lemma 4.2 Let T ∈ T be a tree-network and H be a tree decomposition for T with pivot size θ
and depth ℓ. Then H can be transformed into a layered decomposition hσ, πi with critical set size
∆ = 2(θ + 1) and length ℓ.
Proof: For 1 ≤ i ≤ ℓ, let eGi to be the set consisting of all demand instances d such that depth of
µ(d) is i. We define σ to be the reverse of eG1, eG2, . . . , eGℓ; namely, let σ = G1, G2, . . . , Gℓ, where
Gi = eGℓ−i+1, for 1 ≤ i ≤ ℓ. Thus, in σ, the demand instances captured at the nodes having the
highest depth are placed in G1 and the demand instances captured at the root are placed in Gℓ. We
now show how to construct the critical set π(d) for each demand instance d ∈ D(T ). Let z = µ(d)
be the node in H where d is captured. Add the wing(s) of z on path(d) to π(d). Then, consider
the component C(z) consisting of z and its descendents in H. Let U = {u1, u2, . . . , us} be the
neighbors of C(z), where s ≤ θ. For 1 ≤ i ≤ s, let yi be the bending point of d with respect to
16
Figure 6: An example tree-network.
ui; add the wing(s) of yi on path(d) to π(d). Notice that π(d) has at most 2(θ + 1) edges. This
completes the construction of σ and π(·).
We now argue that the construction satisfies the properties of layered decompositions. Consider
any two groups Gi and Gj such that i ≤ j. Consider two overlapping demand instances d1 ∈ Gi
and d2 ∈ Gj. Let z1 = µ(d1) and z2 = µ(d2) be the nodes in H where d1 and d2 are captured,
respectively. We consider two cases: (1) z2 ∈ C(z1); (2) z2 6∈ C(z1).
Case 1:
In this case, z2 must be the same as z1 (otherwise, we have depth(z2) > depth(z1);
this would contradict i ≤ j). Therefore, path(d2) should include at least one of the wings of z1 on
path(d1). Recall that the wing(s) of z1 on path(d1) are included in π(d1).
Case 2: By the LCA property of tree-decompositions, path(d1) will be constained within
the component C(z1). We have that path(d2) goes through the node z2 found outside of C(z1);
moreover, it also goes through some node found within C(z1) (since d1 and d2 overlap). By the
second property of tree decompositions, such a path must also pass through one of the neighbors
of C(z1); let u be such a neighbor. Let the bending point of path(d1) with respect to u be y. Since
path(d2) passes through u and overlaps with path(d1), the path(d2) must also pass through the
bending point y. It follows that path(d2) must include one of the wings of y on path(d1). Recall
that the wing(s) of y on path(d1) are included in π(d1).
(cid:3)
By applying Lemma 4.2 for the ideal tree decomposition (given by Lemma 4.1), we establish
the following result.
Lemma 4.3 For any tree-network T ∈ T , we can construct a layered decomposition with critical
set size ∆ = 6 and length at most O(log n).
17
Begin
// Initialize
For all a ∈ A, set α(a) = 0; for all e ∈ E, set β(e) = 0.
Initialize an empty stack.
Let the input set of tree-networks be T = {T1, T2, . . . , Tr}.
For each tree-network Tq
Invoke Lemma 4.3 on Tq and
obtain a layered decomposition σq = G(q)
1 , G(q)
2 , . . . , G(q)
and a mapping πq.
ℓq
Let ℓmax = maxq ℓq.
For each d ∈ D(Tq), define π(d) = πq(d).
For each k = 1 to ℓmax,
q=1G(q)
k .
Define Gk = ∪r
// First phase
For k = 1 to ℓmax
//Epochs
Let b be the smallest integer such that (14/15)b ≤ ǫ
For j = 1 to b
//Stages
While
// Steps or iterations.
Let U = {d ∈ Gk : d is (1 − (14/15)j )-unsatisfied}
If U = ∅, exit the loop.
Find a maximal independent set I contained within U
For each d ∈ I
Compute slackness: s = p(d) − α(ad) −Pe:d∼e β(e).
Compute: δ(d) = s/(π(d) + 1).
Raise the variables: α(ad) ← α(ad) + δ(d); for all e ∈ π(d), β(e) ← β(e) + δ(d).
Push I into the stack (as a single object).
// Second Phase
S = ∅.
While(stack not empty)
Pop the top element I of the stack
For each d ∈ I
If S ∪ {d} is an independent set, then add d to S.
Output S.
End
Figure 7: Pseudocode of the overall algorithm
5 Distributed Algorithm
In this section, we prove the main result of the paper by exhibiting a two-phase procedure with
critical set size ∆ = 6 and slackness parameter λ = (1 − ǫ), for any constant ǫ > 0.
Let the input tree networks be T = {T1, T2, . . . , Tr}. For each tree-network Tq, invoke Lemma
18
4.3 and obtain a layered decomposition σq = G(q)
of length ℓq and a mapping πq.
Let ℓmax = maxq ℓq. The lemma guarantees that ℓmax is O(log n) and all the critical set sizes are
at most ∆ = 6. Let ∆′ = ∆ + 1 and ξ = (2∆′)/(2∆′ + 1) = 14/15. For the ease of exposition, we
combine all the mapping functions into single mapping function π, as follows. For each tree-network
Tq and demand instance d ∈ D(Tq), define π(d) = πq(d).
2 , . . . , G(q)
1 , G(q)
ℓq
q=1G(q)
For each 1 ≤ k ≤ ℓmax, let Gk be union of the kth components of all the layered decompositions:
Gk = ∪r
k . The algorithm would follow the two-phase framework. All the dual variables are
initialized to zero and an empty stack is created. The first phase is split into ℓmax epochs. Epoch
k will process the group Gk. Our goal is to ensure that at the end of the epoch, all the demand
instances in Gk are (1 − ǫ)-satisfied. Each epoch is divided into multiple stages, with each stage
making a gradual progress towards the goal. We will ensure that at the end of stage j, all the demand
instances in Gk are (1 − ξj)-satisfied. Each stage is split into multiple steps (each step corresponds
to an iteration of the two-phase framework). A typical step is explained next. Let U be the set of all
demand instances in Gk that are (1 − ξj)-unsatisfied. Find a maximal independent set I contained
within U . For all demand instances d ∈ I, raise the demand instance d as prescribed by the
framework, taking π(d) to be the critical edges. Namely, for all demand instances d ∈ I, perform the
raising as follows. Compute the slackness s = p(d)−α(ad)−Pe : d∼e β(e) and δ(d) = s/(π(d)+1).
Raise the dual variable α(ad) by the amount δ(d) and for all e ∈ π(d), raise the dual variable β(e)
by δ(d). The stage is completed when all the demand instances in Gk are (1 − ξj)-satisfied and
we proceed to the next stage. The epoch is completed when all the demand instances in Gk are
(1−ǫ)-satisfied. The second phase is the same as that of the two-phase framework. The pseudocode
is provided in Figure 7.
Let us analyze the number of steps (or iterations) taken by the above algorithm. The number of
epochs is ℓmax, which is O(log n). Each epoch has at most logξ ǫ = O(log(1/ǫ)) stages. The lemma
below provides a bound on the number of steps taken by each stage.
It follows that the total
number of communication rounds is at most O(Time(MIS) log n log(1/ǫ) log(pmax/pmin)), where
Time(MIS) is the number of (communication) rounds needed to find a maximal independent set
(see Introduction).
Lemma 5.1 Consider any epoch k and stage j within the epoch. The number of steps taken by the
stage is at most O(log(pmax/pmin)).
Proof: Let the number of steps taken by the stage be L. For 1 ≤ i ≤ L, let Ui be the demand
instances in Gk that are (1 − ξj)-unsatisfied at the beginning of step i. Let I1, I2, . . . , IL be the
sequence of maximal independent sets computed in these steps. For two demand instances d1, d2 ∈
Gk, we say that d1 kills d2 in step i, if d1 ∈ Ii, d2 ∈ Ui+1, and d1 and d2 are conflicting. Intuitively,
both d1 and d2 are present in Ui, and both are contenders for the maximal independent Ii. Of the
two, d1 got selected in Ii and d2 was omitted; even after the demand instances in Ii were raised,
d2 was still (1 − ξj)-unsatisfied. Since d1 and d2 are conflicting, only one of them can be included
in the independent set. We imagine that d1 "kills" d2.
Claim 5.2 Suppose d1 kills d2 in step i. Then, their profits satisfy p(d2) ≥ 2p(d1)
We now prove the claim. Since d1 ∈ Ii, the demand instance is (1 − ξj)-unsatisfied at the
beginning of step i. Hence, the difference between the LHS and RHS of the constraint is at least
19
ξj · p(d1). The number dual variables raised for d1 is at most ∆ + 1. Hence,
δ(d1) ≥
ξj · p(d1)
(∆ + 1)
Since d1 and d2 are conflicting, either it is the case that d1 and d2 belong to the same demand a
or they belong to the same tree-network Tq (for some q) and overlap. In the former case, the dual
constraints of d1 and d2 share the dual variable α(ad). In the latter case, both d1 and d2 belong to
the same group G(q)
k . Hence, the properties of layered decompositions imply that one of the critical
edges in π(d1) also appears in the path(d2). Thus, in either case, when d1 is raised, the LHS of
d2 is also raised by an amount δ(d1). On the other hand, d2 ∈ Ui+1 and so, even after the above
raise in the LHS value, d2 is still (1 − ξj)-unsatisfied. As we are considering stage j, all the demand
instances in Gk are (1 − (ξ)j−1)-satisfied. The gap between (1 − ξj−1)p(d2) and (1 − ξj)p(d2) is
(ξj−1 − ξj)p(d2). We see that even after the value of the LHS of the dual constraint of d2 is raised
by an amount δ(d1), the above gap is not bridged. It follows that
This implies that
(ξj−1 − ξj)p(d2) ≥ δ(d1) ≥
ξj · p(d1)
(∆ + 1)
p(d2)
p(d1)
≥
ξ
(1 − ξ)(∆ + 1)
.
We derive the claim by substituting ξ = 14/15 and ∆ = 6.
Consider any demand instance dL ∈ UL. There must exist a demand instance dL−1 in UL−1
such that dL−1 kills dL. In general, we can find a sequence of demand instances dL, dL−1, . . . , d1
such that for 1 ≤ i ≤ L − 1, di kills di+1. By the above claim, for 1 ≤ i ≤ L − 1, p(di+1) ≥ 2p(di).
It follows that p(dL) ≥ 2L−1p(d1). Hence, L ≤ 1 + log(p(dL)/p(d1)) = O(log(pmax/pmin)).
(cid:3)
The properties of layered decomposition imply that the above two-phase algorithm satisfies the
interference property, governed by parameters ∆ = 6 and λ = (1 − ǫ). Therefore, by Lemma 3.1,
it follows that the algorithm has an approximation ratio of 7/(1 − ǫ). For ǫ′ > 0, we can choose
ǫ suitably and obtain an approximation ratio of (7 + ǫ′). We have proved the main result of the
paper.
Theorem 5.3 There exists a distributed algorithm for the unit height case of the throughput maxi-
mization problem on tree-networks with approximation ratio (7+ǫ) and number of (communication)
rounds is at most O(Time(MIS) log n log(1/ǫ) log(pmax/pmin)), where ǫ > 0 is any constant.
Remark: Recall that Panconesi and Sozio [15] presented an algorithm for the unit height case
of line-networks. Their algorithm follows the two-phase framework with the slackness parameter
λ = 1/(5+ǫ). On the other hand, our algorithm has λ = (1−ǫ). A comparison of the two algorithms
is in order. We reformulate their algorithm to suit our framework. Their algorithm also classifies
the demand instances into groups (based on length) and processes the groups in epochs. However,
each epoch consists of only a single stage. They split the stage into multiple iterations/steps. In any
iteration, a demand instance d which is (1/(5 + ǫ))-satisfied is ignored for the rest of the first phase.
In contrast, our algorithm works in multiple stages, where in each stage, we make gradual progress
towards making the demand instances within the group to be (1 − ǫ)-satisfied. In particular, in
stage j, a demand instance which is (1 − ξj)-satisfied is not ignored; it exits the current stage, but
20
it is included in the MIS computations in the next stage.
Distributed Implementation: Here, we sketch certain aspects of implementing the algorithm
in a distributed manner. Let Mmax be the number of bits needed to encode the information about
any demand (such as its end-points and profit).
For now, assume that the values pmax and pmin are known to all the processors. Under this as-
sumption, we can count the number of epochs, stages and iterations exactly. The number of epochs
is 2⌈log n⌉ (the maximum depth of ideal tree decompositions); the number of stages within each
epoch is ⌈logξ ǫ⌉, where ξ = 14/15; the number of iterations within each stage is c log(pmax/pmin),
where c is a suitable constant.
Each processor P computes the ideal tree-decomposition and the corresponding layered decom-
position for each tree-network T accessible to it. Each processor maintains the values of the dual
variables correspoding to its demand instances. The algorithm proceeds in a synchronous fashion
consisting of multiple communication rounds, where each round corresponds to a tuple hf1, f2, f3i,
where f1, f2 and f3 are the epoch, stage and iteration number of the pseudocode. Given a tuple
hf1, f2, f3i, a processor P can determine which demand instances can participate in the MIS cal-
culation of this communication round. The MIS calculation is performed considering the conflict
graph: the demand instances participating in the MIS computation form the vertices and an edge
is drawn between a pair of vertices, if they are conflicting. The number of vertices is at most
N = mr, where m is the number of demands and r is the number of tree-networks. The MIS can
computed using either the randomized algorithm of Luby [14] or using the deterministic procedure
of network decompositions [17]. In the former case, the number of (communication) rounds needed
in O(log N ), whereas in the latter case, it is O(2√log N ). Each processor P contains at most one
demand instance d belonging to the MIS. The processor P raises the dual variables corresponding
to d as given in the pseudocode. The new dual variables are transmitted to the processors sharing
a common resource with P . Upon receiving the new dual variables, each processor updates the
dual variables of its demand instances suitably. Each processor P raises at most a constant num-
ber of dual variables in each iteration (because the critical set size ∆ = 6 is a constant) and the
amount of increase is at most pmax. Thefore, the message size is bounded by Mmax. The stack is
implemented in a distributed manner. Each processor P maintains its own stack. Whenever the
processor P raises a demand instance d, it pushes d onto its stack along with a tuple t (consisting
of the corresponding epoch, stage and iteration numbers).
The second phase proceeds in the reverse order. Each round of the second phase is associated
with a tuple t (consisting of epoch, stage and iteration numbers). A processor P will compare tuple
t′ on the top of the stack with the tuple t; if they match, then it will pop the demand instance d
on the top of the stack. It will output d, if feasibility is maintained. In this case, the processor P
will inform its neighboring processors that d has been included in the output solution.
Finally, we note that it is not difficult to bypass the assumption that all the processors know
the values of pmax and pmin.
6 Arbitrary Height Case for Tree-Networks
Panconesi and Sozio [15] presented a distributed (20 + ǫ)-approximation algorithm for the unit
height case of the line-networks. Then, they extended this algorithm for the arbitrary height case
and obtained a (55 + ǫ)-approximation algorithm [16].
In this section, we extend our (7 + ǫ)-
21
approximation algorithm for the unit height case of the tree-networks to the arbitrary height case
and derive a (20+ǫ)-approximation algorithm. We note that the extra ideas needed for the extension
roughly follow the theme of Panconesi and Sozio [16]. The main difference is that their algorithm
follows the two-phase framework with the slackness parameter being λ = 1/(5 + ǫ), whereas we aim
for λ = (1 − ǫ). Below we highlight the necessary changes.
We first develop some notation for the arbitrary height case. The problem setup is as before,
except that each demand a ∈ A also has a bandwidth requirement (or height) h(a). All the edges
in all the tree-networks are assumed to provide a uniform bandwidth of 1 unit. For each demand
a ∈ A owned by a processor P and each tree-network T accessible to P , create a copy of a with the
same end-points, height and profit; these copies are called demand instances of a. Let D denote
the set of all demand instances (over all demands). A demand instance d ∈ D belonging to a
tree-network T has a height h(d) and a profit p(d), and it can be viewed as a path between its
end-points in T . A feasible solution selects a subset of demand instances S ⊆ D such that: (i)
for any demand a ∈ A, at most one demand instance of a is selected; (ii) for any edge e in some
tree-network T , among the demand instances in S, the sum of heights of the demand instances
passing through e is at most 1 unit. The profit of the solution is the sum of profits of the demand
instances contained in it. The goal is to choose the solution having the maximum profit.
We classify the input demand instances d ∈ D into two categories based on their height: (i)
narrow instances: d is said to be narrow, if h(d) ≤ 1/2; (ii) wide instances: d is said to be wide,
if h(d) > 1/2. Notice that two wide demand instances which are overlapping cannot be picked
by a feasible solution. Hence, if our input consists only of wide demand instances, we can reuse
the algorithm for the unit height case and get a (7 + ǫ) approximation factor. We next describe
a (13 + ǫ)-approximation algorithm for the special case where in the input consists of only narrow
instances. The final algorithm will be derived by combining the above two algorithms.
6.1 Narrow Instances
Here, we assume that the input consists of only narrow instances and develop a (13+ǫ)-approximation
algorithm.
LP and Dual
The LP and the dual have to be modified suitably to reflect the notion of heights. Recall that for
each demand instance d ∈ D, we have a variable x(d).
max
x(d) · p(d)
Xd∈D
x(d) · h(d) ≤ 1
for edges e ∈ E
x(d) ≤ 1
for all demands a ∈ A
x(d) ≥ 0
for all demand instances d ∈ D
Xd∈D : d∼e
Xd∈Inst(a)
The first set of constraints capture the fact that the cumulative height of the demand instances
active on an edge e cannot exceed one unit. Similarly, the second set of constraints capture the
fact that a feasible solution can select at most one demand instance belonging to any demand.
22
The dual of the LP is as follows. For each demand a ∈ A and each edge e ∈ E, the dual includes
a variable α(a) and β(e), respectively. Similarly, for each demand instance d ∈ D, the dual includes
a constraint; we call this the dual constraint of d. Recall that ad denotes the demand to which a
demand instance d belongs.
min
α(ad) + h(d) Xe : d∼e
Xa∈A
α(a) +Xe∈E
β(e)
β(e) ≥ p(d)
for demand instances d ∈ D
The dual also consists of the non-negativity constraints: for all a ∈ A and e ∈ E, α(a) ≥ 0 and
β(e) ≥ 0.
Two-phase Framework
We modify the two-phase framework as follows. As before, the algorithm proceeds iteratively and
a typical iteration is performed as below. We choose suitable independent set I and raise each
demand instance d ∈ I. The slackness computation is modified to reflect the notion of heights in
the constraints. Define the slackness to be:
s = p(d) − α(ad) + h(d) · Xe : d∼e
β(e)! .
We next select a suitable subset π(d) consisting of critical edges on which d is active. The strategy
for raising the dual variables is modified slightly. We raise α(ad) by δ(d) and for each e ∈ π(d),
raise β(e) by 2π(d)δ(d). Towards that goal, define δ(d) = s/(1 + 2h(d)π(d)2). We see that the
dual constraint is satisfied tightly in the process. The second phase of the algorithm remains the
same.
The parameters critical set size ∆ and slackness λ are defined as before. Similarly, the concept
of interference property remains the same. Lemma 3.1 can be extended as follows, using similar
arguments.
Approximation Guarantee
Lemma 6.1 Suppose the input consists of only narrow instances. Consider any algorithm satis-
fying the interference property and governed by parameters ∆ and λ. Then the feasible solution S
produced by the algorithm satisfies p(S) ≥(cid:16) λ
1+2∆2(cid:17) · p(Opt).
Proof: As before, let α(·) and β(·) be the dual variable assignment produced at the end of the
first phase. Convert the assignment hα, βi into a dual feasible solution by scaling the values by an
amount 1/λ: for each demand instance d, setbα(d) = α(d)/λ and for each edge e, set bβ(e) = β(e)/λ.
Notice that the hbα,bβi forms a feasible dual solution.
Let val(α, β) and val(bα,bβ) be the objective value of the dual assignment hα, βi and the dual
feasible solution hbα,bβi, respectively. By the weak duality theorem, val(bα,bβ) ≥ p(Opt). The scaling
process implies that val(bα,bβ) = val(α, β)/λ. We now establish a relationship between val(α, β) and
p(S).
23
Let R denote the set of all demand instances that are raised in the first phase. The value
val(α, β) can be computed as follows. For any d ∈ R, the variable α(ad) is raised by an amount
δ(d) and for each e ∈ π(d), the variable β(e) is raised by an amount 2π(d)δ(d). We have that
π(d) ≤ ∆. So, whenever a demand instance d ∈ R is raised, the objective value raises at most
(2∆2 + 1)δ(d). Therefore,
val(α, β) ≤ (2∆2 + 1)Xd∈R
δ(d).
(4)
We next compute the profit of the solution p(S). For a pair of demand instances d1, d2 ∈ R, we
say that d1 is a predecessor of d2, if the pair is conflicting and d1 is raised before d2; in this case d2
is called the successor of d1. For a demand instance d ∈ R, let pred(d) and succ(d) denote the set
of predecessors and successors of d, respectively; we exclude d from both the sets.
Define K = R − S; we say that demand instances in K are killed by the procedure. We say that
a demand d′ is killed by a demand d if the following three conditions are true: (i) d ∈ S; (ii) both
d and d′ belong to the same demand; (iii) d′ ∈ pred(d). For a demand d ∈ S, let K1(d) denote the
set of demand instances killed by d; notice that K1(d) ⊆ pred(d). Let K1 be the union of K1(d)
over all demand instances d ∈ S. Let K2 be the demand instances in K that could not be added
to the solution because of bandwidth constraints (i.e., K2 consists of demand instances d such that
for some edge e ∈ path(d), h(d) +Pd′∈S : d′∼e h(d′) > 1).
Consider any demand instance d ∈ S and we shall computer a lowerbound on p(d). Consider
the iteration in which d was raised; after the raise, the LHS of the constraint of d becomes equal to
the RHS (i.e., p(d)). The demand instance d contributes at least δ(d) to the LHS. In the previous
iterations, any demand instance d′ ∈ K1(d) would have contributed δ(d′) to the LHS. Similarly, for
any e ∈ path(d) and any demand instance d′ ∈ pred(d) ∩ K2 such that e ∈ π(d′), d′ would have
contributed 2h(d)π(d′)δ(d′) to the LHS. Thus, we see that:
d′∈pred(d)∩K2 : e∈π(d′)
Summing up p(d) over all the demand instances d ∈ S, we get that
p(d) ≥ δ(d) + Xd′∈K1(d)
δ(d) +Xd∈S Xd′∈K1(d)
p(S) ≥Xd∈S
δ(d′) + Xd′∈K1
p(d) ≥ Xd′∈S
δ(d′) + Xe∈path(d)
X
δ(d′) +Xd∈S Xe∈path(d)
δ(d′) +Xd′∈K2
Rewriting the second and the third terms of the RHS:
Xd∈S
2π(d′)δ(d′) Xe∈π(d′)
Xd∈S∩succ(d′) : d∼e
Now let us analyze the third term in the RHS. Consider any d′ ∈ K2. The demand instance d′ could
not be added to S because of the bandwidth constraint being violated at some edge e′ ∈ path(d′).
Meaning,
2h(d)π(d′)δ(d′) .
X
d′∈pred(d)∩K2 : e∈π(d′)
2h(d)π(d′)δ(d′) .
h(d) .
(5)
h(d′) +
Xd∈S∩succ(d′) : d∼e′
h(d) > 1.
24
Since all demand instances are assumed to be narrow, we have that h(d′) ≤ 1/2. It follows that
Xd∈S∩succ(d′) : d∼e′
h(d) > 1/2.
Let X = {d ∈ S ∩ succ(d′) : d ∼ e′}. By the interference property, all the demand instances in X
h(d) >
1
,
2π(d′)
Xd∈ eX
second summation is over all the edges e ∈ π(d′). We replace this summation by the single quantity
are active at one of the edges in π(d′). It follows that there exists an edgeee ∈ π(d′) such that
where eX = {d ∈ X : d ∼ ee}. Consider the third term in the RHS of the formula (5). The
corresponding toee. We get that
δ(d′) +Xd′∈K1
p(d) ≥ Xd′∈S
p(d) ≥ Xd′∈S
≥ Xd′∈R
h(d)
Xd∈S∩succ(d′) : d∼ee
2π(d′)
δ(d′) + Xd′∈K2
δ(d′)! + Xd′∈K1
δ(d′) + Xd′∈K2
2π(d′)δ(d′) ×
2π(d′)δ(d′)
Xd∈S
Xd∈S
1
Therefore,
δ(d′).
The second statement follows from the fact that R is the union of S, K1 and K2. Comparing (4),
we see that val(α, β) ≤ (2∆2 + 1)p(S). The lemma is proved by the observations made at the
beginning of the proof.
(cid:3)
Distributed Algorithm
Fix any ǫ > 0. Our goal is to design a two-phase procedure with critical set size ∆ = 6 and
slackness parameter λ = (1 − ǫ). The distributed algorithm is similar to that of the unit height
case (Section 5). We use the same layered decompositions given by Lemma 4.3 (with ∆ = 6 and
, for some suitable
length ℓ = log n). The only difference is that we set the parameter ξ =
constant c. Arguments similar that of Lemma 5.1 can be used to show that in any epoch k and
any stage j, the algorithm takes at most O(log(pmax/pmin)) steps (or iterations). The number of
epochs in O(log n), as before. The number of stages within each epoch is logξ(ǫ). The above value
of ξ ensures that the above quantity is at most O(1/hmin) log(1/ǫ). Therefore, the total number of
steps is O(Time(MIS) · (1/hmin) · (log n) · log(1/ǫ) · log(pmax/pmin)). The above algorithm satisfies
the interference property with critical set size ∆ = 6 and slackness parameter λ = (1 − ǫ). By
Lemma 6.1, the approximation ratio is (2∆2 + 1)/λ = 73/(1 − ǫ). We have established the following
lemma.
c+hmin
c
Lemma 6.2 Fix any hmin ≤ 1/2 and ǫ > 0. Consider the special case of the scheduling problem
on tree-networks with heights wherein all the demands a are narrow and satisfy h(a) ≥ hmin. There
exists a distributed algorithm for the above problem with approximation ratio (73 + ǫ). The number
of (communication) rounds is at most O(cid:16) Time(MIS)
hmin
log n log 1
ǫ log pmax
pmin(cid:17).
25
Overall Algorithm
Fix any ǫ′ > 0. We present an algorithm within an approximation ratio of (80+ǫ′), for the arbitrary
height case of tree-networks.
We classify the demand instances into wide and narrow instances. Let Opt1 and Opt2 denote
the optimal solutions considering only the wide and narrow instances, respectively. Notice that
p(Opt) ≤ p(Opt1) + p(Opt2). For the wide instances, we run the algorithm for the unit height case
(Theorem 5.3) and obtain a solution S1 such that p(Opt1) ≤ (7+ǫ′)p(S1). Next we run the algorithm
for the narrow instances (Lemma 6.2) and obtain a solution S2 such that p(Opt2) ≤ (73 + ǫ′)p(S2).
Output a combined solution S as follows. For each tree-network T ∈ T , consider the set of demand
instances scheduled on T by the solution S1 and the set of demand instances scheduled on T by
the solution S2; among the two sets, choose the one with the higher profit.
It is easy to see
that the bandwidth constraints are satisfied by S. Furthermore, for any demand a, either all the
demand instances of a are narrow or all of them are wide. Therefore, S will pick at most one
demand instance from any demand. This shows that S is indeed a feasible solution. We have
that p(S) ≥ max{p(S1), p(S2)}. It follows that p(Opt) ≤ (80 + 2ǫ′)p(S). We have established the
following theorem.
Theorem 6.3 Fix any hmin ≤ 1/2 and ǫ > 0. Consider the scheduling problem on tree-networks
with arbitrary heights wherein all the demands a satisfy h(a) ≥ hmin. There exists a distributed
algorithm for the above problem with approximation ratio (80 + ǫ). The number of (communication)
rounds is at most O(cid:16) Time(MIS)
hmin
pmin(cid:17).
log n log 1
ǫ log pmax
7 Line-Networks
Recall that Panconesi and Sozio [16] presented distributed algorithms for the case of line-networks
with windows. For the unit height case the approximation ratio was (20 + ǫ) and for the arbitrary
height case, the ratio was (55 + ǫ). In this section, we obtain a improved approximation ratios
(4 + ǫ) and (23 + ǫ), respectively. We next explain the new algorithms within our framework.
We first develop some notation. Let P be the set of m processors, T be the set of r resources
and A be the set of n demands. We divide timeline into n discrete timeslots, 1, 2, . . . , n. Each
processor P owns a demand a ∈ A. Each demand a ∈ A is specified by a window [rt(a), dl(a) and
a processing time ρ(a), where rt(a) and dl(a) are the release time and deadline of a. A profit p(a)
and a height h(a) are associated with each demand a. Consider a processor P and the demand
a owned by P . For each resource T accessible by P and each interval of length ρ(a) contained
within [rt(a), dl(a)], create a demand instance d; its profit and height are the same as that of a; the
number of demand instances is at most nAcc(P ), where Acc(P ) is the set of resource accessible
to P . Let D denote the set of all demand instances. Each demand instance d ∈ D is described by
a starting time s(d), and ending time e(d), a profit p(d), a height h(d) and the resource to which
it belongs. Recall that the time-line can be viewed as a tree-network with n + 1 vertices. In other
words, the case of line-networks can be reduced to the case of tree-networks. Therefore, Theorem
5.3 and 6.3 apply to the case of line-networks. We next show how to improve these results in the
case of line networks. The improvements are obtained by designing layered decompositions with
better parameters.
26
Improved Layered Decomposition
In the case of tree-networks, we derived decompositions with critical set size ∆ = 6 and length
ℓ = O(log n). In the case special of line-networks, we show how to construct decompositions with
parameters ∆ = 3 and ℓ = O(log(Lmax/Lmin)) (Lmax and Lmin are the maximum and minimum
length of the demand instances). We note that this decomposition is implicit in [16].
Partition the demand instances in to ℓ = ⌈log(Lmax/Lmin)⌉ categories based on their length,
where the length of a demand instance is len(d) = e(d) − s(d) + 1. For 1 ≤ i ≤ ℓ, define Gi =
{d : 2i−1Lmin ≤ len(d) ≤ 2iLmin. Define an ordering σ = G1, G2, . . . , Gℓ. For each demand
instance d, let mid(d) = ⌊(s(d) + e(d))/2⌋ be the mid-point of d. Define π(d) = {s(d), mid(d), e(d)},
for all d ∈ D. It is not difficult to argue that the pair hσ, πi forms layered decomposition.
Unit Height Case
Fix ǫ > 0. We modify the distributed algorithm given in Section 5 to use the above layered
decomposition. We suitably change the value of ξ to 8/9 (instead of 14/15). This algorithm would
satisfy the interference property with ∆ = 3 and λ = (1−ǫ). So, the approximation ratio is 4/(1−ǫ).
The number of (communication) rounds is O(Time(MIS)·log(1/ǫ)·log(Lmax/Lmin)·log(pmax/pmin).
We have established the following result.
Theorem 7.1 There exists a distributed algorithm for the scheduling problem for the unit height
case of line-networks with windows with approximation ratio (4 + ǫ). The number of (communica-
tion) rounds is at most O(cid:16)Time(MIS) log 1
ǫ log Lmax
Lmin
log pmax
pmin(cid:17),
Arbitrary Height Case
Here, we discuss the arbitrary height case. As in Section 6, we partition the set of demand instances
into narrow and wide categories. For the case of wide instances, Theorem 7.1 applies and yields an
algorithm with an approximation ratio of (4 + ǫ). For the case of narrow instances, the algorithm
, for a
is similar to that of Lemma 6.2; the only change is that we set the parameter ξ =
suitable constant c′. This way we get an algorithm with an approximation ratio of 19 + ǫ (because,
in the current setup ∆ = 3). We obtain an overall algorithm by combining the solutions output
by the above two algorithm; the idea is same as that of Theorem 6.3. We have established the
following result.
c′+hmin
c′
Theorem 7.2 Fix any hmin ≤ 1/2 and ǫ > 0. Consider the scheduling problem on tree-networks
with arbitrary heights wherein all the demands a satisfy h(a) ≥ hmin. There exists a distributed
algorithm for the above problem with approximation ratio (23 + ǫ). The number of (communication)
rounds is at most O(cid:16) Time(MIS)
hmin
log n log 1
ǫ log pmax
pmin(cid:17).
References
[1] M. Andrews, J. Chuzhoy, S. Khanna, and L. Zhang. Hardness of the undirected edge-disjoint
paths problem with congestion. In FOCS, 2005.
[2] N. Bansal, A. Chakrabarti, A. Epstein, and B. Schieber. A quasi-PTAS for unsplittable flow
on line graphs. In ACM Symposium on Theory of Computing (STOC), pages 721 -- 729, 2006.
27
[3] N. Bansal, Z. Friggstad, R. Khandekar, and M. Salavatipour. A logarithmic approximation for
unsplittable flow on line graphs. In ACM-SIAM Symposium on Discrete Algorithms (SODA),
pages 702 -- 709, 2009.
[4] A. Bar-Noy, R. Bar-Yehuda, A. Freund, J. Naor, and B. Schieber. A unified approach to
approximating resource allocation and scheduling. Journal of the ACM, 48(5):1069 -- 1090,
2001.
[5] P. Berman and B. DasGupta. Improvements in throughout maximization for real-time schedul-
ing. In STOC, 2000.
[6] P. Bonsma, J. Schulz, and A. Wiese. A constant factor approximation algorithm for unsplit-
table flow on paths. In FOCS, 2011.
[7] G. Calinescu, A. Chakrabarti, H. J. Karloff, and Y. Rabani. Improved approximation algo-
rithms for resource allocation. In IPCO, 2002.
[8] V. Chakaravarthy, V. Pandit, Y. Sabharwal, and D. Seetharam. Varying bandwidth resource
allocation problem with bag constraints. In IPDPS, 2010.
[9] V. Chakaravarthy, S. Roy, and Y. Sabharwal. Distributed algorithms for scheduling on line
and tree networks. In PODC, 2012.
[10] A. Chakrabarti, C. Chekuri, A. Gupta, and A. Kumar. Approximation algorithms for the
unsplittable flow problem. Algorithmica, 47(1):53 -- 78, 2007.
[11] C. Chekuri, M. Mydlarz, and F. Shepherd. Multicommodity demand flow in a tree and packing
integer programs. ACM Transactions on Algorithms, 3(3), 2007.
[12] S. Kolliopoulos. Edge-disjoint paths and unsplittable flow. In T. Gonzalez, editor, Handbook
of Approximation Algorithms and Metaheuristics. Chapman and Hall/CRC, 2007.
[13] L. Lewin-Eytan, J. Naor, and A. Orda. Admission control in networks with advance reserva-
tions. Algorithmica, 40(4):293 -- 304, 2004.
[14] M. Luby. A simple parallel algorithm for the maximal independent set problem. SIAM Journal
of Computing, 15(4):1036 -- 1053, 1986.
[15] A. Panconesi and M. Sozio. Fast distributed scheduling via primal-dual. In SPAA, 2008.
[16] A. Panconesi and M. Sozio. Fast primal-dual distributed algorithms for scheduling and match-
ing problems. Distributed Computing, 22(4):269 -- 283, 2010.
[17] A. Panconesi and A. Srinivasan. On the complexity of distributed network decomposition.
Journal of Algorithms, 20(2):356 -- 374, 1996.
[18] R. Tarjan. Decomposition by clique separators. Discrete Mathematics, 55(2):221 -- 232, 1985.
[19] D. Williamson and D. Shmoys. The Design of Approximation Algorithms. Cambridge Univer-
sity Press, 2011.
28
Begin
// Initialize
For all a ∈ A, set α(a) = 0; for all e ∈ E, set β(e) = 0.
Initialize an empty stack.
// First phase
For i = 1 to r
While
Let U = {d ∈ D(Ti) : dual constraint of d is unsatisfied}.
If U = ∅, exit the loop.
Let d be the element of U appearing earliest in σ(Ti).
Compute slackness: s = p(d) − α(ad) −Pe:d∼e β(e).
Compute: δ(d) = s/(π(d) + 1).
Raise the variables: α(ad) ← α(ad) + δ(d); for all e ∈ π(d), β(e) ← β(e) + δ(d).
Push d into the stack.
// Second Phase
S = ∅.
While(stack not empty)
Pop the top element d of the stack
If S ∪ {d} is an independent set, then add d to S.
Output S.
End
Figure 8: Sequential Algorithm: Pseudocode
A A Sequential Algorithm for Tree-networks
Here, we present a sequential algorithm satisfying the interference property with critical set size
∆ = 2 and slackness parameter λ = 1.
Let T = {T1, T2, . . . , Tr} be the given set of tree-networks. For 1 ≤ i ≤ r, construct a rooted tree
Hi by arbitrarily selecting a node gi ∈ V and making gi as the root of Ti. Let H = {H1, H2, . . . , Hr}
be the set of rooted trees constructed by the above process.
Consider an input tree-network T ∈ T . Let H ∈ H be the rooted tree corresponding to T with
g as the root. For a node x, define its depth (or height) to be the number of the nodes along the
path from g to x; the root g itself is defined to have to depth 1. For a demand instance d ∈ D(T ),
let µ(d) denote the node in H having the least depth among all the nodes appearing in path(d).
We say that d is captured at the node µ(d); observe that µ(d) is uniquely determined. See Figure
6. A rooted-tree H has been constructed by picking the node 1 as the root. The demand instance
d = h4, 13i will be captured at the node µ(d) = 2. With respect to H, a node y is said to be an
ancestor of x, if y appears along the path from g to x; in this case, x is said to be a descendent of
y. By convention, we do not consider x to be an ancestor or descendent of itself.
Consider a demand instance d ∈ D(T ). We have that path(d) passes through the node µ(d).
Let π(d) denote the set of edges of path(d) that are adjacent to µ(d). If µ(d) is one of the end-points
of d then π(d) will have only one edge; otherwise, it will have two edges. See Figure 6. The set
π(d) for the demand instance d = h4, 13i is given by π(d) = {h2, 4i, h2, 5i}. We can now make the
following important observation.
29
Observation A.1 Consider any two overlapping demand instances d1, d2 ∈ D(T ). If d1 is captured
at a node z and d2 is captured at an ancestor of z, then path(d2) will include one of the edges from
π(d1). Furthermore, if d1 and d2 are captured at the same node z, then also path(d2) will include
one of the edges from π(d1).
Based on the above observation, our algorithm works as follows. In the two-phase framework,
in each iteration, we need to determine an independent set I. We shall perform the above task by
selecting a singleton demand instance d (forming a trivial independent set); we can afford to do
this, since we are designing a sequential algorithm. Towards determining d, we define an ordering
of the demand instances for each tree-network. Consider a tree-network T ∈ T and let H be the
rooted tree corresponding to T with g as the root. Order the demand instances d ∈ D(T ) in the
descending order of the depth of µ(d); ties are broken arbitrarily. Thus, demand instances captured
at the bottom-most leaves of H will be placed first and those captured at the root will be placed
last. Let σ(T ) denote the ordering obtained by the above process.
The first phase works in r rounds, where the ith round will process the tree-network Ti. The
round i works in multiple iterations, where each iteration is performed as described below. Let U
be the demand instances belonging to Ti whose constraint is unsatisfied (recall that λ = 1). Among
demand instances in U , pick the demand instance d appearing earliest in the ordering σ(Ti) and
raise the demand d: namely, raise the dual variable α(ad) and the dual variables of the edges found
in π(d). The ith round is completed when the dual constraints of all the demand instances of D(Ti)
are satisfied. The pseudocode of the algorithm is presented in Figure 8.
Observation A.1 shows that the algorithm satisfies the interference property. Furthermore, for
any demand instance d, π(d) ≤ 2. Thus, we have a two-phase procedure with parameters ∆ = 2
and λ = 1. Lemma 3.1 shows that the algorithm is a 3-approximation algorithm. In the special
case where there is only one tree-network, we do not need to raise the dual variables α(·) and the
approximation ratio can be improved to 2. The resulting algorithm would essentially be the same
as that of Lewin-Eytan et al.
[13]. The round complexity of the algorithm can be as high as n,
since only a single demand instance is raised in each iteration.
30
|
1702.03989 | 3 | 1702 | 2018-09-26T05:59:57 | Selecting with History | [
"cs.DS"
] | We define a new selection problem, \emph{Selecting with History}, which extends the secretary problem to a setting with historical information. We propose a strategy for this problem and calculate its success probability in the limit of a large sequence. | cs.DS | cs |
Selecting with History
Tom Hess and Sivan Sabato
Ben-Gurion University of the Negev
Beer Sheva, Israel
{tomhe,sabatos}@cs.bgu.ac.il
Abstract
We define a new selection problem, Selecting with History, which extends the secretary problem to
a setting with historical information. We propose a strategy for this problem and calculate its success
probability in the limit of a large sequence.
In the classical secretary problem (Dynkin, 1963; Gilbert and Mosteller, 1966), n numbers appear at a
random order. The algorithm is allowed to select a single number. If it decides to select a number, it must
do so immediately, before observing the next numbers, and it cannot later change its decision. The goal
of the algorithm is to select the maximal number with the highest probability, where the set of numbers is
selected by an adversary and the order of their appearance is random. Gilbert and Mosteller (1966) show
that, for any input size n, there is a number tn < n such that the optimal strategy is to observe the first tn
numbers, set θ to be the maximal number among those, and then select the first number in the rest of the
sequence which is larger than θ. They show that limn→∞ tn/n = 1/e and that the probability of success of
the optimal strategy by also tends to 1/e when for n → ∞.
In this note we define a new selection problem, Selecting with History (SwH), which extends the secre-
tary problem to a setting with historical information. We propose a strategy for this problem, and calculate
its success probability in the limit of a large sequence.
Let N, K ≥ 2 be integers, such that K divides N . Let Z be a finite set of real numbers of size N . In
this problem, the numbers in Z are ordered according to a uniformly random order. The algorithm observes
the first N (1 − 1/K) numbers (the history). Then, the algorithm observes the last N/K numbers (the
selection sequence) one by one, and should select the maximal number in the selection sequence with the
highest probability. As in the secretary problem, the algorithm may only select a number immediately after
observing it, and cannot regret this selection later. The secretary problem is thus equivalent to SwH with
K = 1. When K ≥ 2, one can ignore the history and simply apply the optimal secretary problem strategy to
the selection sequence. However, this does not exploit the information from the history. Instead, we propose
the following strategy for SwH. This strategy is parametrized by β ∈ (0, 1).
During the first ⌈βN/K⌉ numbers in the selection sequence, select the first number that exceeds
the Kth-largest value in the history. If no such number was found in this part of the selection
sequence, select from the rest of the sequence the first number that exceeds the maximal number
observed so far in the selection sequence.
1
This strategy is inspired by a strategy proposed in Gilbert and Mosteller (1966) for a setting where a selection
sequence is drawn i.i.d. from a known distribution. Whereas under a known distribution the first threshold
can be set based on this knowledge, here we estimate it based on the history.
As in the secretary problem, the probability of success of this strategy depends only on the rank order
of the numbers, and not on their specific values. For K ≥ 2 that divides N , we denote by R(N, K) the
probability that the proposed strategy succeeds in selecting the maximal number from the selection sequence,
for any Z of size N . This probability depends on β, which we leave as an implicit parameter of R. For
convenience, we also let R(N, 1) := Psp(N ), where Psp(n) denotes the success probability of the optimal
secretary problem strategy on an input sequence of size n.
Define
Q(K) := lim
L→∞
R(LK, K).
By the definition of R, Q(1) = limn→∞ Psp(n) = 1/e. The following lemma gives the value of Q(K) for
K > 1, as a function of β.
Lemma 1. If K > 1, then
Q(K) = β log(1/β)(1 −
1
K
)K +
∞
Xj=1
(cid:18)j + K − 1
K − 1 (cid:19)(1 −
1
K
)K 1
K j (cid:18) 1 − (1 − β)j
j
+ βZ 1
β
(1 − x)j−1
x
dx(cid:19) .
Proof. Let N = LK. We calculate R(N, K) based on its definition, and then take the limit L → ∞. Let
Z = {z1, . . . , zN } be the set of input numbers, where z1 > z2 > . . . > zN . Denote by G the event that
the SwH strategy selects the maximal number in the selection sequence, when the strategy is applied to a
random ordering of Z. Let A1 be the set of numbers in the history, and let A2 = Z \A1 be the set of numbers
in the selection sequence. Let A′
2 ⊆ A2 be the set of first B := ⌈βL⌉ numbers in selection sequence. Let
θ be the K'th largest number in A1, and let θ′ be the largest number in A′
2. The strategy described above
selects the first number observed from A′
2 which is larger than θ if one exists. Otherwise, it selects the first
2 that is larger than θ′ (if one exists). Let J = {z ∈ A2 z > θ}. We have
number observed from A2 \ A′
P[G J = j]P[J = j]. Note that the probabilities all depend (implicitly) on L. Let c > 1.
P[G] = PL
For any L ≥ cK 2,
j=0
P[G] =
cK 2
Xj=0
P[G J = j]P[J = j] + P[G J > cK 2]P[J > cK 2].
Define q(j) := limL→∞ P[G J = j], p(j) := limL→∞ P[J = j], and suppose that for some α : R → R,
α(c) ≥ limL→∞ P[J > cK 2]. Assuming all these limits exist, we have
cK 2
Xj=0
q(j)p(j) ≤ lim
L→∞
P[G] ≤
cK 2
Xj=0
q(j)p(j) + α(c).
If in addition limc→∞ α(c) = 0, then, taking c → ∞ on the inequality above, we get
lim
L→∞
P[G] =
∞
Xj=0
q(j)p(j).
2
(1)
We now give expressions for q(j), p(j) and α(c). First, for p(j), we calculate P[J = j]. Define the
random variable I which satisfies θ = zI . If I = i, this means that out of the numbers z1, . . . , zi−1, exactly
K − 1 are in A1, and also zi ∈ A1. Therefore J = i − K. Since A1 = N − L and its content is allocated
uniformly at random, we have
P[J = i − K] = P[I = i] = (cid:18) i − 1
K − 1(cid:19) K−1
Yl=0
N − L − l
N − l
i−K−1
Yl=0
L − l
N − K − l
.
Therefore
P[J = j] = (cid:18)j + K − 1
K − 1 (cid:19) K−1
Yl=0
N − L − l
N − l
j−1
Yl=0
L − l
N − K − l
.
Taking the limit for L → ∞ (recalling N = LK) we get
p(j) ≡ lim
L→∞
P[J = j] = (cid:18)j + K − 1
K − 1 (cid:19)(1 −
1
K
)K 1
K j .
(2)
Second, to find q(j), we now calculate P[G J = j]. If j = 0, then all z ∈ A2 have z < θ, therefore no
element will be selected from A′
2. The probability of success is thus exactly as the probability of success of
the secretary problem strategy with input size L and threshold B = ⌈βL⌉. Denote this probability PL. We
have, following the analysis in Ferguson (1989) for the secretary problem,
lim
L→∞
PL = lim
L→∞
⌈βL⌉
L
L
Xi=B+1
1
i − 1
= lim
L→∞
⌈βL⌉
L
L
Xi=B+1
1
L (cid:18) L
i − 1(cid:19) = βZ 1
β
1
x
dx = β log(1/β).
Hence ,
q(0) ≡ lim
L→∞
P[G J = 0] = β log(1/β).
(3)
To find q(j) for j > 0, let R be the location in A2 of the maximal number z∗ = max A2. Note that if
the strategy does not select anything before reaching location R, it will certainly select z∗ by the definition
of the strategy. Distinguish two cases:
1. If R ≤ B, then z∗ is selected as long as all other j − 1 items that exceed θ are located after z∗. Hence,
for r ≤ min(B, L − j + 1)
P[G R = r, J = j] =
j−2
Yl=0
L − r − l
L − 1 − l
j−2
(1 −
=
Yl=0
r − 1
L − 1 − l
).
2. If R > B, then z∗ is selected as long as all other j − 1 items that exceed θ are located after z∗, and also
2 that
the maximal item in the first R − 1 items is in the first B items, so that z∗ is the first item in A2 \ A′
is larger than θ′. Hence, for B ≤ r ≤ L − j + 1,
P[G R = r, J = j] =
j−2
Yl=0
L − r − l
L − 1 − l
B
r − 1
j−2
(1 −
=
Yl=0
r − 1
L − 1 − l
)
B
r − 1
.
3
3. Neither of the conditions above can hold if R > L − j + 1, since j − 1 numbers cannot be located after
z∗ in this case. Hence ,P[G R > L − j + 1, J = j] = 0.
Therefore
We have
and B
r−1 = ⌈βL⌉
L
P[G J = j] =
=
L−j−1
Xr=1
Xr=1
1
L
L−j+1
P[R = r]P[G J = j, R = r]
r − 1(cid:19)I[r>B]
(cid:18) B
j−2
(1 −
·
Yl=0
r − 1
L − 1 − l
).
(1 −
r − 1
L + 1 − j
)j−1 ≤
j−2
(1 −
Yl=0
r − 1
L − 1 − l
) ≤ (1 −
r − 1
L − 1
)j−1,
L
r−1 . Therefore
1
L
L−j+1
Xr=1
(cid:18) ⌈βL⌉
L
L
r − 1(cid:19)I[r>B]
· (1 −
r − 1
L + 1 − j
)j−1 ≤ P[G J = j]
Taking the limit L → ∞ on both sides and defining x = r/L, this gives, for j ≥ 1,
≤
1
L
L−j+1
Xr=1
(cid:18) ⌈βL⌉
L
L
r − 1(cid:19)I[r>B]
· (1 −
r − 1
L − 1
)j−1.
q(j) ≡ lim
L→∞
P[G J = j] = Z β
(1 − x)j−1 dx + βZ 1
(1 − x)j−1
dx
0
1 − (1 − β)j
=
j
β
x
(1 − x)j−1
x
+ βZ 1
β
dx.
(4)
Lastly, we are left to show an upper bound α(c) ≥ limL→∞ P[J > cK 2] such that limc→∞ α(c) = 0.
Recall that if θ = zi then J = i − K. For an integer t, denote Bt = {i i < t and zi ∈ A2}.
Note that J ≥ t if and only if θ ≤ zt+K , which occurs if and only if Bt+K ≥ t. Therefore P[J ≥
t] = P[Bt+K ≥ t]. We now give an upper bound on P[Bt+K ≥ t], using a concentration bound for
sampling without replacement from a population. Denote the ordered numbers in the selection sequence by
I[ji < t]. Bt is a sum of L uniformly random draws without replacement
from the sequence x1, . . . , xN , where xi := I[i < t]. We have P[ji < t] = (t − 1)/N . Hence ,E[Bt] =
L(t − 1)/N = (t − 1)/K. Setting t = cK 2 for c > 1, we have t − (t + K − 1)/K ≥ t/4. Hence ,
(zj1, . . . , zjL). Then Bt = PL
i=1
P[Bt+K ≥ t] = P[Bt+K − E[Bt+K ] ≥ t − (t + K − 1)/K] ≤ P[Bt+K − E[Bt+K ] ≥ t/4].
By Bernstein's inequality for sampling without replacement (Boucheron et al., 2013), setting ǫ = t/(4L)
and σ2 = 1
i=1(xi − (t − 1)/N )2,
N PN
P[Bt+K ≥ t] ≤ P[
1
L
Bt+K −
1
L
E[Bt+K] ≥ ǫ] ≤ exp(−Lǫ2/(2σ2 + (2/3)ǫ)).
4
i=1 x2
Noting that σ2 ≤ 1
i = (t − 1)/N ≤ cK/L, and ǫ = cK 2/(4L), we get that for some constant
b, P[J ≥ cK 2] ≤ exp(−bcK 2). Setting the RHS to α(c), we get α(c) ≥ limL→∞ P[J > cK 2] and
limc→∞ α(c) = 0, as required. Combining Eq. (1), Eq. (2), Eq. (3), Eq. (4) and the limit above, we get the
equality in the statement of the lemma.
N PN
The value of Q(K) for a given β can be calculated numerically. We propose to select β := 0.63. This
gives, e.g. Q(2) ≈ 0.47, Q(3) ≈ 0.51, Q(10) ≈ 0.55. Compare this to Q(1) = 1/e ≈ 0.37.
In a previous version of this manuscript we proposed to use the strategy for SwH above to improve
the competitive ratio of the submodular secretary problem under resource constraints. Unfortunately our
analysis turned out to have an error which we have not been able to solve as of yet.
References
S. Boucheron, G. Lugosi, and P. Massart. Concentration inequalities: A nonasymptotic theory of indepen-
dence. Oxford university press, 2013.
E. B. Dynkin. The optimum choice of the instant for stopping a markov process. In Sov. Math. Dokl, volume
4(52), pages 627 -- 629, 1963.
T. S. Ferguson. Who solved the secretary problem? Statistical science, pages 282 -- 289, 1989.
J. P. Gilbert and F. Mosteller. Recognizing the maximum of a sequence. Journal of the American Statistical
Association, 61(313):35 -- 73, 1966.
5
|
1207.7134 | 1 | 1207 | 2012-07-30T23:38:52 | Improved approximation algorithms for low-density instances of the Minimum Entropy Set Cover Problem | [
"cs.DS",
"cs.CC"
] | We study the approximability of instances of the minimum entropy set cover problem, parameterized by the average frequency of a random element in the covering sets. We analyze an algorithm combining a greedy approach with another one biased towards large sets. The algorithm is controled by the percentage of elements to which we apply the biased approach. The optimal parameter choice has a phase transition around average density $e$ and leads to improved approximation guarantees when average element frequency is less than $e$. | cs.DS | cs |
Improved approximation algorithms for
low-density instances of the Minimum Entropy
Set Cover Problem.
Cosmin Bonchiş∗, Gabriel Istrate†
July 31, 2018
Abstract
We study the approximability of instances of the minimum entropy
set cover problem, parameterized by the average frequency of a ran-
dom element in the covering sets. We analyze an algorithm combining
a greedy approach with another one biased towards large sets. The al-
gorithm is controled by the percentage of elements to which we apply
the biased approach. The optimal parameter choice has a phase tran-
sition around average density e and leads to improved approximation
guarantees when average element frequency is less than e.
1 Introduction
The minimum entropy set cover problem (MESC) [Halperin and Karp(2005)]
arose from a maximum likelihood approach to haplotype inference in com-
putational biology (see also [Mãndoiu and Paşaniuc(2005)]). Halperin and
Karp showed that the problem is NP-complete and provided an additive up-
per bound (equal to three) on the performance of the Greedy algorithm. This
was later improved by Cardinal et al. [Cardinal et al.(2008a)], who showed a
∗Department of Computer Science, West University of Timişoara, Bd. V. Pârvan 4,
Timişoara, RO-300223, Romania.
†e-Austria Research Institute, Bd. V. Pârvan 4, cam. 045 B, Timişoara, RO-300223,
Romania.email:[email protected]. corresponding author
1
tight additive upper bound of log2(e). Cardinal et al.
[Cardinal et al.(2012)] also studied several versions of this problem, notably
minimum entropy graph coloring [Cardinal et al.(2004)] and minimum en-
tropy orientation [Cardinal et al.(2008b)], as well as a generalization to ar-
bitrary objective functions [Cardinal and Dumeunier(2008)]. Minimum en-
tropy graph coloring has found applications to problems related to functional
compression in information theory [Cardinal et al.(2004)].
Minimum entropy set cover also lies behind a recently proposed family
of measures of worst-case fairness in cost allocations in cooperative game
theory [Bonchiş and Istrate(2012a)]. This was accomplished by first studying
[Bonchiş and Istrate(2012b)] a minimum entropy version of the well-known
submodular set cover problem [Wolsey(1982), Fujita(2000)]. Submodularity
corresponds in the setting of cooperative game theory to concavity of the
associated game, a property that guarantees many useful features of the game
such as the non-emptiness of the core, membership of the Shapley value in the
core, equivalence between group-strategyproofness and cross-monotonicity in
mechanism design [Moulin(1999)] and so on.
In this paper we further study MESC restricted to sparse instances, that
is to instances of Set Cover parameterized by f (formally defined below), the
average number of sets that cover a random element.
In the spirit of the
minimum entropy orientation problem (a version of MESC for which f = 2)
we aim to provide better approximation guarantees than those valid for the
Greedy algorithm. To accomplish this goal we study the performance of
an approximation algorithm BiasedGreedy(δ) parameterized by a constant
δ ∈ [0, 1].
Our main result can be summarized as follows: we give general upper
bounds on the performance of our proposed algorithm. These bounds im-
prove on the approximation guarantee of the greedy algorithm when average
element frequency is less than the constant e. Furthermore, the best choice of
control parameter δ depends on this frequency: it corresponds to the choice
of a "biased" algorithm below critical value e, and to the greedy algorithm
above it.
The paper is structured as follows: in Section 2 we review basic notions
and define the algorithm BiasedGreedy. The main result is presented and
further discussed in Section 3.
Its proof is given in Section 4. Next we
present several applications of our main result to the Minimum Entropy
Graph Coloring problem.
2
2 Preliminaries
In this paper we need the definition of Shannon entropy and its associated
divergence of two distributions P and Q:
D(P k Q) = Xi
pi log2
pi
qi
.
We recall that D(P k Q) ≥ 0 for all P and Q.
We are concerned with the following problem:
Definition 1. [MINIMUM ENTROPY SET COVER (MESC)]: Let
U = {u1, u2, . . . , un} be an n−element ground set, for some n ≥ 1, and let
P = {P1, P2, . . . , Pm} be a family of subsets of U which cover U. A cover is
a function g : U → [m] such that for every 1 ≤ i ≤ n,
ui ∈ Pg(ui)("ui is covered by set Pg(ui)")
The entropy of cover g is defined by:
Ent(g) = −
m
Xi=1
g−1(i)
U
log2
g−1(i)
U
.
(2.1)
[OBJECTIVE:] Find a cover g of minimum entropy.
Consider an instance (U, P) as above. Define
f = Pm
i=1 Pi
U
,
the average frequency of a random element in U.
In the algorithm below we divide the elements of the ground set into Light
and Heavy elements, based on their frequency of occurrence. Parameter δ
controls this division: the least frequent δn elements are deemed Light, while
the rest are considered Heavy.
Informally, the algorithm will first covers Light elements in a biased man-
ner, simultaneously covering each such elemen by a set of maximum cardinal-
ity containing it. Once this phase is complete all Light elements are deleted
from all sets. The Heavy elements are handled in an incremental manner via
a Greedy approach. The algorithm is formally presented in the following:
3
INPUT: An instance (U, P) of MESC
P H := {P H
i = Pi \ L for all i ∈ [k]
1 , P H
2 , . . . , P H
k } where P H
While (there exists e ∈ L)
choose ie ∈ [k] to maximize Pie where Pie ∋ e;
let g(e) = ie;
L := L \ {e};
While (there exists e ∈ H)
choose ie ∈ [k] to maximize P H
ie where P H
ie ∋ e;
let g(e) = ie;
erase e from all P H
i
;
H := H \ {e};
OUTPUT: the cover g.
Figure 2.1: BiasedGreedy(δ)
3 Main result
Our main result shows that the following upper bound on the performance
of algorithm BiasedGreedy holds:
Theorem 1. Algorithm BiasedGreedy(δ) produces a cover BG : U 7−→ [k]
satisfying:
Ent(BG) ≤ Ent(OP T ) − (1 − δ) log2(cid:18)1 − δ
e (cid:19) + log2 f + o(1).
(3.1)
Corollary 1. The Biased algorithm, defined as the BiasedGreedy algorithm
with δ = 1, produces a cover BI whose entropy satisfies
Ent(BI) ≤ Ent(OP T ) + log2 f.
(3.2)
Observation 1. Optimizing over constant δ in inequality (3.1) reveals an
interesting fact: the optimal choice of δ is always δ ∈ {0, 1}, i.e. the pure
Biased or Greedy algorithms. More precisely
4
• choice δ = 1 (i.e. Biased) is optimal for f < e.
• when f > e choice δ = 0 (i.e. Greedy) becomes best.
Thus the optimal choice for δ has a phase transition from δ = 1 to δ = 0
around average density f = e.
4 Proof of the main result
Proof. Let BG be the cover generated by the BiasedGreedy algorithm, and
denote by p♭
the associated probability distribution.
i = BG−1(i)
If OP T is the optimal solution of the same instance, denote xi = OP T −1(i)
n
and yi = OP T −1(i) ∩ Heavy for all 1 ≤ i ≤ k. By choice of δ, Pk
n − ⌈δn⌉ ≤ (1 − δ)n while Pk
We rewrite the entropy of BG as follows:
i=1 xi = n.
i=1 yi =
Ent(BG) = −
k
Xi=1
p♭
i log2 p♭
i = −
k
Xi=1
p♭
i log2(cid:18)Pi
p♭
i
Pi(cid:19)
Denoting by # = (#i) the distribution #i = Pi
Pk
j=1 Pj
we obtain:
k
+ log2
Pi
Xi=1
Xi=1
k
p♭
i log2 Pi −
k
Xi=1
p♭
i log2
p♭
i
#i
Ent(BG) = −
= −
k
Xi=1
Xi=1
k
p♭
i log2 Pi − D(BG k #) + log2
Pi
(4.1)
Considering now just the first sum we obtain
−
k
Xi=1
p♭
i log2 Pi = −
= −
k
Xi=1
n Xv∈U
1
BG−1(i)
n
log2 Pi = −
1
n
k
Xi=1 Xv∈BG−1(i)
log2 PBG(v)
log2 PBG(v) = −
1
n Xv∈U
log2 av
where av is the size of the set assigned by BiasedGreedy to cover v.
5
Continuing, we infer
−
k
Xi=1
p♭
i log2 Pi = −
= −
1
n
1
n
k
Xi=1 Xv∈OP T −1(i)
log2 av = −
1
n
k
Xi=1
log2 Yv∈OP T −1(i)
av
k
Xi=1
log2
Yv∈OP T −1(i)∩Light
av ·
Yv∈OP T −1(i)∩Heavy
av
From the definition of the algorithm we conclude the following:
• for all v ∈ OP T −1(i) ∩ Light,
av =PBG(v) = max
j,Pj∋v
Pj ≥ POP T (v) ≥ OP T −1(i) = xi
• On the other hand, for v ∈ OP T −1(i) ∩ Heavy we analyze the Greedy
phase of BiasedGreedy algorithm in a manner completely similar to the
analysis of the Greedy algorithm in [Cardinal et al.(2008a)] and infer
that
Yv∈OP T −1(i)∩Heavy
av ≥ yi!
Therefore,
−
k
Xi=1
k
1
n
Yv∈OP T −1(i)∩Light
Xi=1
Xi=1
yi log2 xi −
1
n
k
xi
(yi!) =
Xi=1
1
n
k
k
Xi=1
log2 yi!
(xi − yi) log2 xi −
log2 yi!
p♭
i log2 Pi ≤ −
1
n
k
Xi=1
k
log2
= −
= −
1
n
Xi=1
Xi=1
xi
n
k
log2(cid:0)xxi−yi
i
(cid:1) (yi!) = −
log2
xi
n
− log2 n +
1
n
Applying now the inequality y! ≥ (y/e)y we obtain:
6
−
k
Xi=1
p♭
i log2 Pi ≤ Ent(OP T ) − log2 n +
yi log2 xi −
1
n
1
n
k
Xi=1
Xi=1
k
1
n
1
n
k
Xi=1
Xi=1
k
log2
yyi
i
eyi
yi log2 e
= Ent(OP T ) − log2 n +
≤ Ent(OP T ) − log2 n −
1
n
1
n
k
Xi=1
Xi=1
k
yi log2 xi −
yi log2 yi +
yi log2
yi
xi
+ (1 − δ) log2 e
(4.2)
Considering now distributions xi = xi
n , yi = yi
Pj yj
we obtain
−
1
n
k
Xi=1
yi log2
yi
xi
= −
1
n
k
Xi=1
(n − ⌈δn⌉)yi log2
(n − ⌈δn⌉)yi
nxi
= −
= −
(n − ⌈δn⌉)
n
(n − ⌈δn⌉)
n
k
Xi=1
yi log2
yi
xi
−
(n − ⌈δn⌉)
n
log2
(n − ⌈δn⌉)
n
D(y k x) −
(n − ⌈δn⌉)
n
log2
(n − ⌈δn⌉)
n
Putting all things together:
Ent(BG) ≤Ent(OP T ) − (1 − δ) log2(1 − δ) + (1 − δ) log2 e + log2 f + o(1).
and the proof is complete.
5 Application to minimum entropy graph col-
oring
Just as it is the case with the Greedy algorithm [Cardinal et al.(2012)], our
result has implications for the minimum entropy coloring problem. This
problem can be recast as an implicit set cover problem [Karp(2011)], where
the sets are the maximal independent sets in G. Given the intractability of
the maximum independent set problem, we can only efficiently implement the
Biased algorithm on special classes of graphs, where this problem is easier.
On the other hand algorithm Biased has some nice properties, similar to
those discussed in [Cardinal et al.(2012)]) for the Greedy algorithm:
7
1. it can be implemented in polynomial time on perfect graphs. Indeed,
the largest independent set containing a given vertex can easily be
computed in a perfect graph.
2. it allows the use of η-approximately optimal independent sets (for some
constant η ≥ 1) instead of optimal ones at the expense of introducing
an extra factor of log2(η) in the upper bound of equation (3.1). This
follows easily by simply redoing the proof of the main Theorem in this
setting. We can apply this observation to get a slight improvement of
Theorem 8 from [Cardinal et al.(2012)] when f < e:
Corollary 2. Algorithm Biased produces a coloring of a graph G =
(V, E) with maximum degree ∆ satisfying
Ent(Biased) ≤ Ent(OP T ) + log2(∆ + 2) + log2(f /3).
The proof of the corollary directly parallels that of Theorem 8 from
[Cardinal et al.(2012)].
Applying Theorem 1 to graph coloring problems is rather inconvenient as
parameter f involves maximal independent sets and is not easy to compute.
The situation is slightly better for graphs with independence number α(G) ≤
3.
In this case maximal independent sets correspond either to triangles,
edges, or isolated vertices in the complement graph G. Parameter f also has
an easier interpretation: Let I be the number of isolated vertices in G. Let
T be the number of distinct triangles in G. Finally, let M be the number of
edges that are not contained in any triangle. Then
f =
I + 2M + 3T
n
Furthermore, in this case the algorithm Biased has a very natural inter-
pretation: we create a tentative color cW for any maximal independent set
(triangle, edge or isolated vertex) W and add color cW to a list coloring of all
vertices in W . Then for each vertex we select a random color from its list.
The algorithm Biased can be improved in practice by employing a number
of heuristics such as:
• Attempt to color all elements of a largest independent set with the
same color.
8
• Collapse two colors into one if legal.
These heuristics can only decrease the entropy of the resulting coloring.
There are instances (e.g. edge orientations of a cycle from [Cardinal et al.(2008b)])
where Biased outperforms Greedy. But even when it doesn't, our analysis
may provide better theoretical guarantees than those available for Greedy.
Example 1. Consider graph G = (V, E) from Figure 5.1 (a) (its complement
is displayed in Figure 5.1 (b)).
1
8
a
b
2
7
3
6
4
5
1
8
2
7
3
6
4
5
Figure 5.1: (a) graph G (b) its complement G
a
a
b
b
c
c
b
c
b
c
a
a
a
d
Figure 5.2: Two colorings C1 and C2 of graph G. For convenience the com-
plement graph G is pictured, rather than G. C1 is an optimal solution.
Graph G provides an easy instance where Greedy and Biased (may) pro-
duce different colorings. Indeed, node 4 is colored by Biased with a color cor-
responding to a triangle, whereas 5 takes a color corresponding to an edge, so
9
nodes 4 and 5 must assume different colors in a Biased coloring, whereas they
may have the same color in a Greedy coloring. With the optimizations de-
scribed above both Greedy and Biased produce one of the two colorings C1, C2
from Figure 1, with color classes of cardinalities (3; 3; 2; 0) and (3; 2; 2; 1),
respectively. The first one corresponds to the optimal solution. On the other
hand for this graph the average element frequency f = 3×3+1×2
8 < e, so
the upper bound on the entropy of coloring C2 given by Corollary 3.2 is tighter
than the one provided by the Greedy algorithm in [Cardinal et al.(2012)].
= 11
8
Acknowledgments
Both authors contributed in a substantially equal manner to this work: G.I.
suggested the problem, performed research and wrote the paper. C.B. per-
formed research and wrote the paper.
The first author has been supported by a project on Postdoctoral national
programs, contract CNCSIS PD_575/2010.
The corresponding author has been supported by CNCS IDEI Grant PN-
II-ID-PCE-2011-3-0981 "Structure and computational difficulty in combina-
torial optimization: an interdisciplinary approach".
References
[Halperin and Karp(2005)] E. Halperin, R. Karp, The minimum entropy set
cover problem, Theoretical Computer Science 348 (2005) 340 -- 350.
[Mãndoiu and Paşaniuc(2005)] I. Mãndoiu, B. Paşaniuc, Haplotype infer-
ence by entropy minimization, in: 9th Annual International Conference
on Research in Computational Molecular Biology (RECOMB) (2005),
pp. 221 -- 222.
[Cardinal et al.(2008a)] J. Cardinal, S. Fiorini, G. Joraet, Tight results on
minimum entropy set cover, Algorithmica 51 (2008) 49 -- 60.
[Cardinal et al.(2012)] J. Cardinal, S. Fiorini, G. Joret, Minimum entropy
combinatorial optimization problems, Theory of Computing Systems 51
(2012) 4 -- 21.
10
[Cardinal et al.(2008b)] J. Cardinal, S. Fiorini, G. Joret, Minimum entropy
orientations, Operations Research Letters 36 (2008) 680 -- 683.
[Cardinal and Dumeunier(2008)] J. Cardinal, C. Dumeunier, Set Cover-
ing Problems with General Objective Functions, Technical Report
0802.2184, arXiv.org, 2008.
[Cardinal et al.(2004)] J. Cardinal, S. Fiorini, G. Van Assche, On minimum
in: Proceedings of ISIT'04, the International
entropy graph colorings,
Symposium on Information Theory (2004), p. 43.
[Bonchiş and Istrate(2012a)] C. Bonchiş, G. Istrate, A parametric worst-case
approach to fairness in TU-Cooperative Games, 2012a. Manuscript.
[Bonchiş and Istrate(2012b)] C. Bonchiş, G. Istrate, Minimum entropy sub-
modular set cover and fairness in cooperative games, 2012b. Manuscript.
[Wolsey(1982)] L. Wolsey, An analysis of the greedy algorithm for the sub-
modular set covering problem, Combinatorica 2 (1982) 385 -- 393.
[Fujita(2000)] T. Fujita, Approximation algorithms for submodular set cover
with applications, IEICE Transactions on Information and Systems E83-
D (2000) 480 -- 487.
[Moulin(1999)] H. Moulin,
Incremental cost sharing: Characterization by
coalition strategy-proofness, Social Choice and Welfare 16 (1999) 279 --
320.
[Karp(2011)] R. M. Karp, Heuristic algorithms in computational molecular
biology, Journal of Computer and System Sciences 77 (2011) 122 -- 128.
11
|
1302.5401 | 1 | 1302 | 2013-02-21T20:19:31 | Sparse Fault-Tolerant BFS Trees | [
"cs.DS"
] | This paper addresses the problem of designing a sparse {\em fault-tolerant} BFS tree, or {\em FT-BFS tree} for short, namely, a sparse subgraph $T$ of the given network $G$ such that subsequent to the failure of a single edge or vertex, the surviving part $T'$ of $T$ still contains a BFS spanning tree for (the surviving part of) $G$. Our main results are as follows. We present an algorithm that for every $n$-vertex graph $G$ and source node $s$ constructs a (single edge failure) FT-BFS tree rooted at $s$ with $O(n \cdot \min\{\Depth(s), \sqrt{n}\})$ edges, where $\Depth(s)$ is the depth of the BFS tree rooted at $s$. This result is complemented by a matching lower bound, showing that there exist $n$-vertex graphs with a source node $s$ for which any edge (or vertex) FT-BFS tree rooted at $s$ has $\Omega(n^{3/2})$ edges. We then consider {\em fault-tolerant multi-source BFS trees}, or {\em FT-MBFS trees} for short, aiming to provide (following a failure) a BFS tree rooted at each source $s\in S$ for some subset of sources $S\subseteq V$. Again, tight bounds are provided, showing that there exists a poly-time algorithm that for every $n$-vertex graph and source set $S \subseteq V$ of size $\sigma$ constructs a (single failure) FT-MBFS tree $T^*(S)$ from each source $s_i \in S$, with $O(\sqrt{\sigma} \cdot n^{3/2})$ edges, and on the other hand there exist $n$-vertex graphs with source sets $S \subseteq V$ of cardinality $\sigma$, on which any FT-MBFS tree from $S$ has $\Omega(\sqrt{\sigma}\cdot n^{3/2})$ edges. Finally, we propose an $O(\log n)$ approximation algorithm for constructing FT-BFS and FT-MBFS structures. The latter is complemented by a hardness result stating that there exists no $\Omega(\log n)$ approximation algorithm for these problems under standard complexity assumptions. | cs.DS | cs |
Sparse Fault-Tolerant BFS Trees
Merav Parter ∗†
David Peleg ∗
October 18, 2018
Abstract
A fault-tolerant structure for a network is required to continue functioning fol-
lowing the failure of some of the network's edges or vertices. This paper considers
breadth-first search (BFS) spanning trees, and addresses the problem of designing
a sparse fault-tolerant BFS tree, or FT-BFS tree for short, namely, a sparse sub-
graph T of the given network G such that subsequent to the failure of a single edge
or vertex, the surviving part T (cid:48) of T still contains a BFS spanning tree for (the
surviving part of) G. For a source node s, a target node t and an edge e ∈ G,
the shortest s − t path Ps,t,e that does not go through e is known as a replacement
path. Thus, our FT-BFS tree contains the collection of all replacement paths Ps,t,e
for every t ∈ V (G) and every failed edge e ∈ E(G).
Our main results are as follows. We present an algorithm that for every n-vertex
graph G and source node s constructs a (single edge failure) FT-BFS tree rooted at
s with O(n · min{Depth(s),
n}) edges, where Depth(s) is the depth of the BFS
tree rooted at s. This result is complemented by a matching lower bound, showing
that there exist n-vertex graphs with a source node s for which any edge (or vertex)
FT-BFS tree rooted at s has Ω(n3/2) edges.
√
We then consider fault-tolerant multi-source BFS trees, or FT-MBFS trees for
short, aiming to provide (following a failure) a BFS tree rooted at each source s ∈ S
for some subset of sources S ⊆ V . Again, tight bounds are provided, showing that
there exists a poly-time algorithm that for every n-vertex graph and source set
S ⊆ V of size σ constructs a (single failure) FT-MBFS tree T ∗(S) from each source
√
si ∈ S, with O(
σ · n3/2) edges, and on the other hand there exist n-vertex graphs
with source sets S ⊆ V of cardinality σ, on which any FT-MBFS tree from S has
Ω(
σ · n3/2) edges.
√
∗The Weizmann Institute of Science, Rehovot, Israel. Email: {merav.parter,david.peleg}@
weizmann.ac.il. Supported in part by the Israel Science Foundation (grant 894/09), the United States-
Israel Binational Science Foundation (grant 2008348), the Israel Ministry of Science and Technology
(infrastructures grant), and the Citi Foundation.
†Recipient of the Google Europe Fellowship in distributed computing; research supported in part by
this Google Fellowship.
1
Finally, we propose an O(log n) approximation algorithm for constructing FT-BFS
and FT-MBFS structures. The latter is complemented by a hardness result stating
that there exists no Ω(log n) approximation algorithm for these problems under
standard complexity assumptions. In comparison with the randomized FT-BFS con-
√
struction implicit in [14], our algorithm is deterministic and may improve the num-
n for some instances. All our algorithms can be
ber of edges by a factor of up to
extended to deal with one vertex failure as well, with the same performance.
1
Introduction
Background and motivation Modern day communication networks support a variety
of logical structures and services, and depend on their undisrupted operation. As the
vertices and edges of the network may occasionally fail or malfunction, it is desirable
to make those structures robust against failures. Indeed, the problem of designing fault-
tolerant constructions for various network structures and services has received considerable
attention over the years.
Fault-resilience can be introduced into the network in several different ways. This
paper focuses on a notion of fault-tolerance whereby the structure at hand is augmented
or "reinforced" (by adding to it various components) so that subsequent to the failure
of some of the network's vertices or edges, the surviving part of the structure is still
operational. As this reinforcement carries certain costs, it is desirable to minimize the
number of added components.
To illustrate this type of fault tolerance, let us consider the structure of graph k-
spanners (cf. [17, 19, 20]). A graph spanner H can be thought of as a skeleton structure
that generalizes the concept of spanning trees and allows us to faithfully represent the
underlying network using few edges, in the sense that for any two vertices of the network,
the distance in the spanner is stretched by only a small factor. More formally, consider
a weighted graph G and let k ≥ 1 be an integer. Let dist(u, v, G) denote the (weighted)
distance between u and v in G. Then a k-spanner H satisfies that dist(u, v, H) ≤ k ·
dist(u, v, G) for every u, v ∈ V .
Towards introducing fault tolerance, we say that a subgraph H is an f -edge fault-
tolerant k-spanner of G if dist(u, v, H \ F ) ≤ k · dist(u, v, G \ F ) for any set F ⊆ E of
size at most f , and any pair of vertices u, v ∈ V . A similar definition applies to f -vertex
fault-tolerant k-spanners. Sparse fault-tolerant spanner constructions were presented in
[6, 11].
This paper considers breadth-first search (BFS) spanning trees, and addresses the
problem of designing fault-tolerant BFS trees, or FT-BFS trees for short. By this we mean
a subgraph T of the given network G, such that subsequent to the failure of some of the
2
vertices or edges, the surviving part T (cid:48) of T still contains a BFS spanning tree for the
surviving part of G. We also consider a generalized structure referred to as a fault-tolerant
multi-source BFS tree, or FT-MBFS tree for short, aiming to provide a BFS tree rooted at
each source s ∈ S for some subset of sources S ⊆ V .
The notion of FT-BFS trees is closely related to the problem of constructing replacement
paths and in particular to its single source variant, the single-source replacement paths
problem, studied in [14]. That problem requires to compute the collection Ps of all s − t
replacement paths Ps,t,e for every t ∈ V and every failed edge e that appears on the s − t
shortest-path in G. The vast literature on replacement paths (cf. [4, 14, 23, 25, 28]) focuses
on time-efficient computation of the these paths as well as their efficient maintenance in
data structures (a.k.a distance oracles).
In contrast, the main concern in the current
paper is with optimizing the size of the resulting fault tolerant structure that contains
the collection Ps of all replacement paths given a source node s. A typical motivation
for such a setting is where the graph edges represent the channels of a communication
network, and the system designer would like to purchase or lease a minimal collection of
channels (i.e., a subgraph G(cid:48) ⊆ G) that maintains its functionality as a "BFS tree" with
respect to the source s upon any single edge or vertex failure in G. In such a context, the
cost of computation at the preprocessing stage may often be negligible compared to the
purchasing/leasing cost of the resulting structure. Hence, our key cost measure in this
paper is the size of the fault tolerant structure, and our main goal is to achieve sparse (or
compact) structures.
Most previous work on sparse / compact fault-tolerant structures and services con-
cerned structures that are distance-preserving (i.e., dealing with distances, shortest paths
or shortest routes), global (i.e., centered on "all-pairs" variants), and approximate (i.e.,
settling for near optimal distances), such as spanners, distance oracles and compact rout-
ing schemes. The problem considered here, namely, the construction of FT-BFS trees,
still concerns a distance preserving structure. However, it deviates from tradition with
respect to the two other features, namely, it concerns a "single source" variant, and it
insists on exact shortest paths. Hence our problem is on the one hand easier, yet on the
other hand harder, than previously studied ones. Noting that in previous studies, the
"cost" of adding fault-tolerance (in the relevant complexity measure) was often low (e.g.,
merely polylogarithmic in the graph size n), one might be tempted to conjecture that
a similar phenomenon may reveal itself in our problem as well. Perhaps surprisingly, it
turns out that our insistence on exact distances plays a dominant role and makes the
problem significantly harder, outweighing our willingness to settle for a "single source"
solution.
Contributions We obtain the following results.
In Sec. 2, we define the Minimum
FT-BFS and Minimum FT-MBFS problems, aiming at finding the minimum such structures
3
tolerant against a single edge or vertex fault. Section 3 presents lower bound constructions
for these problems. For the single source case, in Subsec. 3.1, we present a lower bound
stating that for every n there exists an n-vertex graph and a source node s ⊆ V for which
any FT-MBFS tree from s requires Ω(n3/2) edges. In Subsec. 3.2, we then show that there
exist n-vertex graphs with source sets S ⊆ V of size σ, on which any FT-MBFS tree from
the source set S has Ω(
σ · n3/2) edges.
√
These results are complemented by matching upper bounds.
In Subsec. 4.1, we
√
present a simple algorithm that for every n-vertex graph G and source node s, constructs
a (single edge failure) FT-BFS tree rooted at s with O(n · min{Depth(s),
n}) edges. A
similar algorithm yields an FT-BFS tree tolerant to one vertex failure, with the same size
bound. In addition, for the multi source case, in Subsec. 4.2, we show that there exists
a polynomial time algorithm that for every n-vertex graph and source set S ⊆ V of size
S = σ constructs a (single failure) FT-MBFS tree T ∗(S) from each source si ∈ S, with
√
O(
σ · n3/2) edges.
In Sec. 5, we show that the minimum FT-BFS problem is NP-hard and moreover,
cannot be approximated (under standard complexity assumptions) to within a factor of
Ω(log n), where n is the number of vertices of the input graph G. Note that while the
algorithms of Sec. 4 match the worst-case lower bounds, they might still be far from
optimal for certain instances, as illustrated in Sec. 6. Consequently, in Sec. 6, we
complete the upper bound analysis by presenting an O(log n) approximation algorithm
for the Minimum FT-MBFS problem. This approximation algorithm is superior in instances
where the graph enjoys a sparse FT-MBFS tree, hence paying O(n3/2) edges (as does the
algorithm of Sec. 4) is wasteful. In light of the hardness result for these problems (of Sec.
5), the approximability result is tight (up to constants).
Related work To the best of our knowledge, this paper is the first to study the spar-
sity of fault-tolerant BFS structures for graphs. The question of whether it is possible
to construct a sparse fault tolerant spanner for an arbitrary undirected weighted graph,
raised in [9], was answered in the affirmative in [6], presenting algorithms for construct-
ing an f -vertex fault tolerant (2k − 1)-spanner of size O(f 2kf +1 · n1+1/k log1−1/k n) and
an f -edge fault tolerant 2k − 1 spanner of size O(f · n1+1/k) for a graph of size n. A
randomized construction attaining an improved tradeoff for vertex fault-tolerant span-
ners was shortly afterwards presented in [11], yielding (with high probability) for every
graph G = (V, E), odd integer s and integer f , an f -vertex fault-tolerant s-spanner
edges. This should be contrasted with the best stretch-size
with O
tradeoff currently known for non-fault-tolerant spanners [24], namely, 2k − 1 stretch with
O(n1+1/k) edges.
s+1 n1+ 2
(cid:17)
(cid:16)
f 2− 2
s+1 log n
An efficient algorithm that given a set V of n points in d-dimensional Euclidean space
constructs an f -vertex fault tolerant geometric (1 + )-spanner for V , namely, a sparse
4
graph H satisfying that dist(u, v, H \ F ) ≤ (1 + )dist(u, v, G) for any set F ⊆ V of
size f and any pair of points u, v ∈ V \ F , was presented in [15]. A fault tolerant
geometric spanner of improved size was later presented in [16]; finally, a fault tolerant
geometric spanner of optimal maximum degree and total weight was presented in [9]. The
distinction between the stronger type of fault-tolerance obtained for geometric graphs
(termed rigid fault-tolerance) and the more flexible type required for handling general
graphs (termed competitive fault-tolerance) is elaborated upon in [18].
A related network service is the distance oracle [3, 22, 25], which is a succinct data
structure capable of supporting efficient responses to distance queries on a weighted graph
G. A distance query (s, t) requires finding, for a given pair of vertices s and t in V , the
distance (namely, the length of the shortest path) between u and v in G. The query
protocol of an oracle S correctly answers distance queries on G. In a fault tolerant distance
oracle, the query may include also a set F of failed edges or vertices (or both), and the
oracle S must return, in response to a query (s, t, F ), the distance between s and t in
G(cid:48) = G \ F . Such a structure is sometimes called an F -sensitivity distance oracle. The
focus is on both fast preprocessing time, fast query time and low space.
It has been
shown in [10] that given a directed weighted graph G of size n, it is possible to construct
in time O(mn2) a 1-sensitivity fault tolerant distance oracle of size O(n2 log n) capable of
answering distance queries in O(1) time in the presence of a single failed edge or vertex.
The preprocessing time was recently improved to O(mn), with unchanged size and query
time [4]. A 2-sensitivity fault tolerant distance oracle of size O(n2 log3 n), capable of
answering 2-sensitivity queries in O(log n) time, was presented in [12].
Recently, distance sensitivity oracles have been considered for weighted and directed
graphs in the single source setting [14]. Specifically, Grandoni and Williams considered
the problem of single-source replacement paths where one aims to compute the collection
of all replacement paths for a given source node s, and proposed an efficient random-
ized algorithm that does so in (cid:101)O(AP SP (n, M )) where AP SP (n, M ) is the time required
to compute all-pairs-shortest-paths in a weighted graph with integer weights [−M, M ].
Interestingly, although their algorithm does not aim explicitly at minimizing the total
number of edges used by the resulting collection of replacement paths, one can show that
the resulting construction yields a rather sparse path collection, with at most O(n3/2 log n)
edges (although it may also be far from optimal in some instances).
Label-based fault-tolerant distance oracles for graphs of bounded clique-width are
presented in [8]. The structure is composed of a label L(v) assigned to each vertex v, and
handles queries of the form (L(s), L(t), F ) for a set of failures F . For an n-vertex graph
of tree-width or clique-width k, the constructed labels are of size O(k2 log2 n).
A relaxed variant of distance oracles, in which distance queries are answered by ap-
proximate distance estimates instead of exact ones, was introduced in [25], where it was
shown how to construct, for a given weighted undirected n-vertex graph G, an approxi-
5
mate distance oracle of size O(n1+1/k) capable of answering distance queries in O(k) time,
where the stretch (multiplicative approximation factor) of the returned distances is at
most 2k − 1.
An f -sensitivity approximate distance oracle S was presented in [5]. For an integer
parameter k ≥ 1, the size of S is O(kn1+ 8(f +1)
k+2(f +1) log (nW )), where W is the weight of the
heaviest edge in G, the stretch of the returned distance is 2k − 1, and the query time is
O(F · log2 n · log log n · log log d), where d is the distance between s and t in G \ F .
A fault-tolerant label-based (1+)-approximate distance oracle for the family of graphs
with doubling dimension bounded by α is presented in [2]. For an n-vertex graph G(V, E)
in this family, and for desired precision parameter > 0, the distance oracle constructs and
stores an O(log n/2α)-bit label at each vertex. Given the labels of two end-vertices s and
t and of collections FV and FE of failed (or "forbidden") vertices and edges, the oracle
computes, in time polynomial in the length of the labels, an estimate for the distance
between s and t in the surviving graph G(V \ FV , E \ FE), which approximates the true
distance by a factor of 1 + .
Our final example concerns fault tolerant routing schemes. A fault-tolerant routing
protocol is a distributed algorithm that, for any set of failed edges F , enables any source
vertex s to route a message to any destination vertex d along a shortest or near-shortest
path in the surviving network G \ F in an efficient manner (and without knowing F in
advance).
In addition to route efficiency, it is often desirable to optimize also the amount of
memory stored in the routing tables of the vertices, possibly at the cost of lower route
efficiency, giving rise to the problem of designing compact routing schemes (cf. [1, 7, 17,
21, 24]).
Label-based fault-tolerant routing schemes for graphs of bounded clique-width are
presented in [8]. To route from s to t, the source needs to specify the labels L(s) and
L(t) and the set of failures F , and the scheme efficiently calculates the shortest path
between s and t that avoids F . For an n-vertex graph of tree-width or clique-width k,
the constructed labels are of size O(k2 log2 n).
Fault-tolerant compact routing schemes are considered in [5], for up to two edge fail-
ures. Given a message M destined to t at a source vertex s, in the presence of a failed edge
set F of size F ≤ 2 (unknown to s), the scheme presented therein routes M from s to t in
a distributed manner, over a path of length at most O(k) times the length of the optimal
path (avoiding F ). The total amount of information stored in vertices of G on average is
bounded by O(kn1+1/k). This should be compared with the best memory-stretch tradeoff
currently known for non-fault-tolerant compact routing [24], namely, 2k − 1 stretch with
O(n1+1/k) memory per vertex.
A compact routing scheme capable of handling multiple edge failures is presented
6
in [7]. The scheme routes messages (provided their source s and destination t are still
connected in the surviving graph G \ F ) over a path whose length is proportional to the
distance between s and t in G\ F , to F3 and to some poly-log factor. The routing table
required at a node v is of size proportional to v's degree and some poly-log factor.
A routing scheme with stretch 1 + for graphs of bounded doubling dimension is also
presented in [2]. The scheme can be generalized also to the family of weighted graphs of
bounded doubling dimension and bounded degree. In this case, the label size will also
depend linearly on the maximum vertex degree ∆, and this is shown to be necessary.
2 Preliminaries
paths (or BFS) tree rooted at s. For a source node set S ⊆ V , let T0(S) =(cid:83)
Notation Given a graph G = (V, E) and a source node s, let T0(s) ⊆ G be a shortest
s∈S T0(s) be
a union of the single source BFS trees. Let π(s, v, T ) be the s − v shortest-path in tree
T , when the tree T = T0(s), we may omit it and simply write π(s, v). Let Γ(v, G) be
the set of v neighbors in G. Let E(v, G) = {(u, v) ∈ E(G)} be the set of edges incident
to v in the graph G and let deg(v, G) = E(v, G) denote the degree of node v in G.
When the graph G is clear from the context, we may omit it and simply write deg(v).
Let depth(s, v) = dist(s, v, G) denote the depth of v in the BFS tree T0(s). When the
source node s is clear from the context, we may omit it and simply write depth(v). Let
Depth(s) = maxu∈V {depth(s, u)} be the depth of T0(s). For a subgraph G(cid:48) = (V (cid:48), E(cid:48)) ⊆ G
(where V (cid:48) ⊆ V and E(cid:48) ⊆ E) and a pair of nodes u, v ∈ V , let dist(u, v, G(cid:48)) denote the
shortest-path distance in edges between u and v in G(cid:48). For a path P = [v1, . . . , vk], let
LastE(P ) be the last edge of path P . Let P denote the length of the path and P [vi, vj]
be the subpath of P from vi to vj. For paths P1 and P2, P1 ◦ P2 denote the path obtained
by concatenating P2 to P1. Assuming an edge weight function W : E(G) → R+, let
SP (s, vi, G, W ) be the set of s − vi shortest-paths in G according to the edge weights of
W . Throughout, the edges of these paths are considered to be directed away from the
source node s. Given an s − v path P and an edge e = (x, y) ∈ P , let dist(s, e, P ) be the
distance (in edges) between s and e on P . In addition, for an edge e = (x, y) ∈ T0(s),
define dist(s, e) = i if depth(x) = i − 1 and depth(y) = i.
Definition 2.1 A graph T ∗ is an edge (resp., vertex) FT-BFS tree for G with respect to
a source node s ∈ V , iff for every edge f ∈ E(G) (resp., vertex f ∈ V ) and for every
v ∈ V , dist(s, v, T ∗ \ {f}) = dist(s, v, G \ {f}).
A graph T ∗ is an edge (resp., vertex) FT-MBFS tree for G with respect to source set
S ⊆ V , iff for every edge f ∈ E(G) (resp., vertex f ∈ V ) and for every s ∈ S and v ∈ V ,
dist(s, v, T ∗ \ {f}) = dist(s, v, G \ {f}).
7
To avoid cumbersome notation, we refer to edge FT-BFS (resp., edge FT-MBFS) trees
simply by FT-BFS (resp., FT-MBFS) trees. Throughout, we focus on edge fault, yet the
entire analysis extends trivially to the case of vertex fault as well.
The Minimum FT-BFS problem Denote the set of solutions for the instance (G, s) by
T (s, G) = {(cid:98)T ⊆ G (cid:98)T is an FT-BFS tree w.r.t. s}. Let Cost∗(s, G) = min{E((cid:98)T )
T (s, G)} be the minimum number of edges in any FT-BFS subgraph of G. These definitions
naturally extend to the multi-source case where we are given a source set S ⊆ V of size
σ. Then
T (S, G) = {(cid:98)T ⊆ G (cid:98)T is a FT-MBFS with respect to S}
(cid:98)T ∈
and
∗(S, G) = min{E((cid:98)T )
(cid:98)T ∈ T (S, G)}.
Cost
In the Minimum FT-BFS problem we are given a graph G and a source node s and
the goal is to compute an FT-BFS (cid:98)T ∈ T (s, G) of minimum size, i.e., such that E((cid:98)T ) =
source node set S and the goal is to compute an FT-MBFS (cid:98)T ∈ T (S, G) of minimum size
i.e., such that E((cid:98)T ) = Cost∗(S, G).
Cost∗(s, G). Similarly, in the Minimum FT-MBFS problem we are given a graph G and a
3 Lower Bounds
In this section we establish lower bounds on the size of the FT-BFS and FT-MBFS structures.
In Subsec. 3.1 we consider the single source case and in Subsec. 3.2 we consider the case
of multiple sources.
3.1 Single Source
We begin with a lower bound for the case of a single source.
Theorem 3.1 There exists an n-vertex graph G(V, E) and a source node s ∈ V such that
any FT-BFS tree rooted at s has Ω(n3/2) edges, i.e., Cost∗(s, G) = Ω(n3/2).
Proof: Let us first describe the structure of the graph G = (V, E). Set d = (cid:98)√
n/2(cid:99).
The graph consists of four main components. The first is a path π = [s = v1, . . . , vd+1 =
v∗] of length d. The second component consists of a node set Z = {z1, . . . , zd} and a collec-
tion of d disjoint paths of deceasing length, P1, . . . , Pd, where Pj = [vj = pj
1, . . . , zj = pj
tj ]
connects vj with zj and its length is tj = Pj = 6 + 2(d − j), for every j ∈ 1,··· , d.
j=1 V (Pj), is of size Q = d2 + 7d. The
Altogether, the set of nodes in these paths, Q =(cid:83)d
8
to Z. Overall, V = X ∪ Q and E = E ∪ E(π) ∪(cid:83)d
third component is a set of nodes X of size n − (d2 + 7d), all connected to the terminal
node v∗. The last component is a complete bipartite graph B = (X, Z, E) connecting X
j=1 E(Pj). Note that n/4 ≤ Q ≤ n/2
for sufficiently large n. Consequently, X = n − Q ≥ n/2, and E = Q · X ≥ n3/2/4.
Figure 1: Lower bound construction for FT-BFS The original BFS tree consists of the non-
dashed edges. The dashed edges are the ones necessary to make it an FT-BFS tree. For
example, the bold dashed edge (xj, zi) is required upon failure of the edge ei.
A BFS tree T0 rooted at s for this G (illustrated by the solid edges in Fig. 1) is given
by
E(T0) = {(xi, zi) i ∈ {1, . . . , d}} ∪ d(cid:91)
E(Pj) \ {(pj
(cid:96)j
(cid:96)j−1)},
, pj
j=1
where (cid:96)j = tj − (d − j) for every j ∈ {1, . . . , d}. We now show that every FT-BFS tree
T (cid:48) ∈ T (s, G) must contain all the edges of B, namely, the edges ei,j = (xi, zj) for every
i ∈ {1, . . . ,X} and j ∈ {1, . . . , d} (the dashed edges in Figure 1). Assume, towards
contradiction, that there exists a T (cid:48) ∈ T (s, G) that does not contain ei,j (the bold dashed
edge (xi, zj) in the figure). (the bold dashed edge (xi, zj) in Figure 1). Note that upon the
failure of the edge ej = (vj, vj+1) ∈ π, the unique s − xi shortest-path connecting s and
xi in G \ {ej} is P (cid:48)
j = π[v1, vj] ◦ Pj ◦ [zj, xi], and all other alternatives are strictly longer.
Since ei,j /∈ T (cid:48), also P (cid:48)
(cid:42) T (cid:48), and therefore dist(s, xi, G \ {ej}) < dist(s, xi, T (cid:48) \ {ej}), in
contradiction to the fact that T (cid:48) is an FT-BFS tree. It follows that every FT-BFS tree T (cid:48)
must contain at least E = Ω(n3/2) edges. The theorem follows.
j
9
X Z xi zj 𝑑=𝑂(𝑛) ej vj Pj v* B S z1 vd zd 3.2 Multiple Sources
We next consider an intermediate setting where it is necessary to construct a fault-tolerant
subgraph FT-MBFS containing several FT-BFS trees in parallel, one for each source s ∈ S,
for some S ⊆ V . We establish the following.
Theorem 3.2 There exists an n-vertex graph G(V, E) and a source set S ⊆ V of cardi-
√
σ · n3/2) edges, i.e.,
nality σ, such that any FT-MBFS tree from the source set S has Ω(
Cost∗(S, G) = Ω(
σ · n3/2).
√
Proof: Our construction is based on the graph G(d) = (V1, E1), which consists of three
components: (1) a set of vertices U = {u1, . . . , ud} connected by a path P1 = [u1, . . . , ud],
(2) a set of terminal vertices Z = {z1, . . . , zd} (viewed by convention as ordered from left
to right), and (3) a collection of d vertex disjoint paths Qi of length Qi = 6 + 2(d − i)
d. The vertex
connecting ui and zi for every i ∈ {1, . . . , d}. Thus Q1
r(G(d)) = ud is fixed as the root of G(d), hence the edges of the paths Qi are viewed
as directed away from ui, and the terminal vertices of Z are viewed as the leaves of the
graph, denoted Leaf(G(d)) = Z. See Fig. 2 for illustration.
1 > . . . > Q1
Overall, the vertex and edge sets of G(d) are V1 = U ∪ Z ∪(cid:83)d
E(P1) ∪(cid:83)d
i=1 E(Qi).
Figure 2: The graph G1(d).
i=1 V (Qi) and E1 =
Observation 3.3 (a) The number of leaves in G(d) is Leaf(G(d)) = d.
(b) V1 = c · d2 for some constant c.
10
u11 u12 u13 u14 z1 z2 z3 z4 Take σ copies, G(cid:48)
1, . . . , G(cid:48)
1), . . . , Leaf(G(cid:48)
σ, of G(d), where d = O((n/σ)1/2). Note that Obs. 3.3,
each copy G(cid:48)
i consists of O(n/σ) nodes. Let yi be the node ud and si = r(G(cid:48)
i) in the
ith copy G(cid:48)
i. Add a node v∗ connected to a set X of Ω(n) nodes and connect v∗ to
each of the nodes yi, for i ∈ {1, . . . , d}. Finally, connect the set X to the σ leaf sets
Leaf(G(cid:48)
σ) by a complete bipartite graph, adjusting the size of the set X
in the construction so that V (G) = n. Since nLeaf(G(cid:48)
i) = Ω((n/σ)1/2) (see Obs. 3.3),
overall E(G) = Ω(n· σ· nLeaf(G1(d))) = Ω(n· (σn)1/2). Since the path from each source
j for j (cid:54)= i, the analysis of the single-source case can be
si to X cannot aid the nodes of G(cid:48)
applied to show that each of the bipartite graph edges in necessary upon a certain edge
fault. See Fig. 3 for an illustration.
Figure 3:
Illustration of the lower bound for the multi-source case.
4 Upper Bounds
In this section we provide tight matching upper bounds to the lower bounds presented in
Sec. 3.
4.1 Single Source
For the case of FT-BFS trees, we establish the following.
11
s X 𝑂(𝑛) s 𝑑 s s𝜎 Z𝜎 v* G'𝜎 𝑑 s2 s1 G'2 G'1 𝑑=𝑂(𝑛/𝜎) Z2 Z1 √
Theorem 4.1 There exists an O(nm) time algorithm that for every n-vertex graph G
and source node s constructs an FT-BFS tree rooted at s with O(n · min{Depth(s),
n})
edges.
To prove the theorem, we first describe a simple algorithm for the problem and then prove
its correctness and analyze the size of the resulting FT-BFS tree. We note that using the
sparsity lemma of [23] and the tools of [14], one can provide a randomized construction for
an FT-BFS tree with O(n3/2 log n) edges with high probability. In contrast, the algorithm
presented in this paper is deterministic and achieve an FT-BFS tree with O(n3/2) edges,
matching exactly the lower bound established in Sec. 3.
The Algorithm To avoid complications due to shortest-paths of the same length, we
assume all shortest-path are computed with a weight assignment W that guarantees the
uniqueness of the shortest-paths. This can be achieved by considering a weight function
W defined so as to ensure that the shortest paths are also of minimal number of edges but
at the same time guarantees the uniqueness of the u− v shortest-path, for every u, v ∈ V .
Let e1, . . . , em be some arbitrary ordering of E(G). Then set W (ek) = 2m+1 + 2k. Let
T0 = BF S(s, G) be the BFS tree rooted at s in G, computed according to the weight
assignment W . For every ej ∈ T0, let T0(ej) be the BFS tree rooted at s in G \ {ej}.
Then the final FT-BFS tree is given by
T ∗(s) = T0 ∪ (cid:91)
T0(ej).
ej∈T0
The correctness is immediate by construction.
Observation 4.2 T ∗(s) is an FT-BFS tree.
Proof: Consider a vertex v and an edge e. If e /∈ π(s, v), then π(s, v) ⊆ T ∗(s)\{e}, hence
dist(s, v, T ∗(s) \ {e}) = π(s, v) = dist(s, v, G \ {e}). Otherwise, e ∈ π(s, v) ⊆ T0. Then
by construction, T0(e) ⊆ T ∗(s). By definition, dist(s, v, T ∗(s) \ {e}) = dist(s, v, T0(e)) =
dist(s, v, G \ {e}). The observation follows.
Due to [26] each of the n− 1 BFS trees T0(ej) can be constructed in O(m) time, hence
O(nm) rounds are required in total. It therefore remains to bound the size of T ∗(s).
Size Analysis We first provide some notation. For a path P , let Cost(P ) =(cid:80)
e∈P W (e)
be the weighted cost of P , i.e., the sum of its edge weights. An edge e ∈ G is defined as
new if e /∈ E(T0). For every vi ∈ V and ej ∈ T0, let P ∗
i,j = π(s, vi, T0(ej)) ∈ SP (s, vi, G \
{ej}, W ) be the optimal replacement path of s and vi upon the failure of ej ∈ T0. Let
New(P ) = E(P ) \ E(T0) and
New(vi) = {LastE(P ∗
i,j) ej ∈ T0} \ E(T0)
12
be the set of vi new edges appearing as the last edge in the replacement paths P ∗
i,j of vi
and ej ∈ T0. It is convenient to view the edges of T0(ej) as directed away from s. We
then have that
T ∗ = T0 ∪ (cid:91)
New(vi).
vi∈V \{s}
I.e., the set of new edges that participate in the final FT-BFS tree T ∗ are those that appear
as a last edge in some replacement path.
i,j) /∈ E(T0), then ej ∈ π(s, vi).
We now upper bound the size of the FT-BFS tree T ∗. Our goal is to prove that New(vi)
n) edges for every vi ∈ V . The following observation is crucial in
√
contains at most O(
this context.
Observation 4.3 If LastE(P ∗
i,j ⊆ T0(ej) be the s−vi
Proof: Assume, towards contradiction, that ej /∈ π(s, vi) and let P ∗
replacement path in G \ {ej} according to the weight assignment W . Since LastE(P ∗
i,j) /∈
E(T0), we have two different s − vi shortest paths in G \ {ej}, namely, π(s, vi) and
i,j. By the optimality of π(s, vi) in G, i.e., π(s, vi) ∈ SP (s, vi, G, W ), it holds that
P ∗
i,j in G \ {ej}, i.e.,
Cost(π(s, u)) < Cost(P ∗
i,j ∈ SP (s, vi, G\{ej}, W ), we have that Cost(π(s, u)) > Cost(P ∗
P ∗
Obs. 4.3 also yields the following.
i,j). On the other hand, by the optimality of P ∗
i,j). Contradiction.
Corollary 4.4 (1) New(vi) = {LastE(P ∗
(2) New(vi) ≤ min{depth(vi), deg(vi)}.
i,j) ej ∈ π(s, vi)} \ E(T0) and
This holds since the edges of New(vi) are coming from at most depth(vi) replacement
i,j (one for every ej ∈ π(s, vi)), and each such path contributes at most one edge
paths P ∗
incident to vi.
For the reminder of the analysis, let us focus on one specific node u = vi and let
π = π(s, u), N = New(u). For every edge ek ∈ New(u), we define the following parameters.
Let f (ek) ∈ π be the failed edge such that ek ∈ T0(f (ek)) appears in the replacement path
Pk = π(s, u, T (cid:48)) for T (cid:48) = T0(f (ek)). (Note that ek might appear as the last edge on the
path π(s, u, T0(e(cid:48))) for several edges e(cid:48) ∈ π; in this case, one such e(cid:48) is chosen arbitrarily).
Let bk be the last divergence point of Pk and π, i.e., the last vertex on the replacement
path Pk that belongs to V (π) \ {u}. Since LastE(Pk) /∈ E(T0), it holds that bk is not the
neighbor of u in Pk.
Let New(u) = {e1, . . . , eN} be sorted in non-decreasing order of the distance between
bk and u, dist(bk, u, π) = π(bk, u). I.e.,
dist(b1, u, π) ≤ dist(b2, u, π) . . . ≤ dist(bN , u, π).
(1)
We consider the set of truncated paths P (cid:48)
k = Pk[bk, u] and show that these paths are
13
vertex-disjoint except for the last common endpoint u. We then use this fact to bound
the number of these paths, hence bound the number N of new edges. The following
observation follows immediately by the definition of bk.
Observation 4.5 (V (P (cid:48)
k) ∩ V (π)) \ {bk, u} = ∅.
Lemma 4.6 (cid:0)V (P (cid:48)
i ) ∩ V (P (cid:48)
j)(cid:1) \ {u} = ∅ for every i, j ∈ {1, . . . , N}, i (cid:54)= j.
u(cid:48) ∈(cid:0)V (P (cid:48)
j)(cid:1) \ {u}
i ) ∩ V (P (cid:48)
i ) (cid:54)= LastE(P (cid:48)
Proof: Assume towards contradiction that there exist i (cid:54)= j, and a node
j), by Obs. 4.5 we have that P (cid:48)
j ⊆
in the intersection. Since LastE(P (cid:48)
G \ E(π). The faulty edges f (ei), f (ej) belong to E(π). Hence there are two distinct
u(cid:48) − u shortest paths in G \ {f (ei), f (ej)}. By the optimality of P (cid:48)
in T0(f (ei)), (i.e.,
Pi ∈ SP (s, u, G\{f (ei)}, W )), we have that Cost(P (cid:48)
j[u(cid:48), u]). In addition,
i [u(cid:48), u]) < Cost(P (cid:48)
j in T0(f (ej)), (i.e., Pj ∈ SP (s, u, G \ {f (ej)}, W )), we have that
by the optimality of P (cid:48)
i [u(cid:48), u]). Contradiction.
j[u(cid:48), u]) < Cost(P (cid:48)
Cost(P (cid:48)
We are now ready to prove our key lemma.
i , P (cid:48)
i
Lemma 4.7 New(u) = O(n1/2) for every u ∈ V .
Proof: Assume towards contradiction that N = New(u) >
2n. By Lemma 4.6, we
have that b1, . . . , bN are distinct and by definition they all appear on the path π. There-
fore, by the ordering of the P (cid:48)
k, we have that the inequalities of Eq. (1) are strict, i.e.,
dist(b1, u, π) < dist(b2, u, π) < . . . < dist(bN , u, π). Since b1 (cid:54)= u (by definition), we also
have that dist(b1, u, π) ≥ 1. We Conclude that
√
Next, note that each P (cid:48)
π(bk, u), implying that P (cid:48)
dist(bk, u, π) = π(bk, u) ≥ k .
(2)
k is a replacement bk − u path and hence it cannot be shorter than
k ≥ π(bk, u). Combining, with Eq. (2), we have that
k ≥ k for every k ∈ {1, . . . , N} .
P (cid:48)
(3)
Since by Lemma 4.6, the paths P (cid:48)
k are vertex disjoint (except for the common vertex u),
we have that(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) N(cid:91)
k=1
(V (P (cid:48)
k) \ {u})
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =
N(cid:88)
k) \ {u} ≥ N(cid:88)
V (P (cid:48)
k=1
k=1
(k − 1) > n,
where the first inequality follows by Eq. (3) and the last inequality by the assumption
that N >
√
2n. Since there are a total of n nodes in G, we end with contradiction.
Turning to the case of a single vertex failure, the entire proof goes through almost
without change, yielding the following.
14
Theorem 4.8 There exists a polynomial time algorithm that for every n-vertex graph
and source node s constructs an FT-BFS tree from s tolerant to one vertex failure, with
O(n · min{Depth(s),
n}) edges.
√
4.2 Multiple Sources
For the case of multiple sources, we establish the following upper bound.
Theorem 4.9 There exists a polynomial time algorithm that for every n-vertex graph
G = (V, E) and source set S ⊆ V of size S = σ constructs an FT-MBFS tree T ∗(S)
from each source si ∈ S, tolerant to one edge or vertex failure, with a total number of
√
si∈S depth(si), O(
σn)} edges.
n · min{(cid:80)
The algorithm As in the single source case, to avoid complications due to shortest-
paths of the same length, all shortest path distances in G are computed using a weight
function W defined so as to ensure the uniqueness of a single u − v shortest-path. For
every si ∈ S and every ej ∈ T0(si), let T (si, ej) be the BFS tree rooted at si in G \ {ej}.
Let
(cid:91)
si∈S
(cid:91)
the FT-BFS tree for si is T ∗(si) = T0 ∪(cid:83)
T0(S) =
T0(si)
be the joint structure containing all the BFS trees of S. Then by the previous section,
ej∈T0(si) T (si, ej). Define the FT-MBFS for S as
(cid:91)
si∈S
T ∗(S) =
T ∗(si) =
T (si, ej).
si∈S,ej∈T0(si)
Analysis The correctness follows immediately by the single source case. It remains to
bound the number of edges of T ∗(S). An edge e is new if e /∈ T0(S). For every vi ∈ V ,
define its new edge set in the graph T ∗(S) by
New(S, vi) = {LastE(π(s, vi, T (si, ej))) si ∈ S, ej ∈ T0(si)} \ E(T0(S)).
To bound the size of T ∗(S), we focus on node u = vi, and bound its new edges New(S, u) =
{e1, . . . , eN}. Obs. 4.3 yields the following.
Corollary 4.10 New(S, u) ≤(cid:80)
si∈S depth(si).
√
σn) new
Towards the end of this section, we prove that New(S, u) contains at most O(
edges. For ease of notation, let π(si) = π(si, u) for every i ∈ {1, . . . , σ}. For every edge
ek ∈ New(S, u), we define the following parameters. Let s(ek) ∈ S and f (ek) ∈ T0(s(ek))
be such that ek ∈ T (s(ek), f (ek)). I.e., the edge ek appears in the replacement s(ek) − u
15
P =(cid:83)
path Pk = π(s, u, T (cid:48)), where T (cid:48) = T (s(ek), f (ek)) is the BFS tree rooted at s(ek) in
G \ {f (ek)}. By Obs. 4.3, f (ek) ∈ π(s(ek)). (Note that for a given new edge ek there
might be several s(cid:48) and e(cid:48) such that ek = LastE(π(s(cid:48), u, T (s(cid:48), e(cid:48)))); in this case one such
pair s(cid:48), e(cid:48) is chosen arbitrarily.) For every replacement path Pk (whose last edge is ek),
denote by bk the last divergence point of Pk and the collection of shortest si − u paths
si∈S π(si, u) \ {u}. I.e., bk is the last point on Pk that belongs to V (P) \ {u}. Let
P (cid:48)
k = Pk[bk, u] be the truncated path from the divergence point bk to u. Note that since
e = (x, u) = LastE(Pk) /∈ E(T0(S)) is a new edge, it holds that x /∈ V (P) \ {u} and bk is
in V \ {u}. The following observation is useful.
Observation 4.11 P (cid:48)
We now show that the paths P (cid:48)
regardless of their respective source s(ek)).
k ⊆ G \ E(P) for every k ∈ {1, . . . , N}.
k are vertex disjoint except for their endpoint u (this is
Lemma 4.12 (cid:0)V (P (cid:48)
i ) ∩ V (P (cid:48)
Proof: Assume towards contradiction that there exists i (cid:54)= j, and a node
j)(cid:1) \ {u} = ∅ for every i (cid:54)= j ∈ {1, . . . , N}.
u(cid:48) ∈(cid:0)V (P (cid:48)
j)(cid:1) \ {u}
i ) ∩ V (P (cid:48)
i ) (cid:54)= LastE(P (cid:48)
j ⊆ G \ E(P),
in the intersection. Since LastE(P (cid:48)
and the faulty edges f (ei), f (ej) ∈ P, we have two distinct u(cid:48) − u replacement paths in
i in T (s(ei), f (ei)), (i.e., Pi ∈ SP (s(ei), u, G \
G \ {f (ei), f (ej)}. By the optimality of P (cid:48)
{f (ei)}, W )), we have that Cost(P (cid:48)
j). Similarly, by the optimality of P (cid:48)
j
in T (s(ej), f (ej)), (i.e., Pj ∈ SP (s(ej), u, G \ {f (ej)}, W )), we have that Cost(P (cid:48)
j) <
Cost(P (cid:48)
j) and by Obs. 4.11, P (cid:48)
i ), contradiction. The lemma follows.
i ) < Cost(P (cid:48)
i , P (cid:48)
We are now ready to state and prove our main lemma.
√
Lemma 4.13 N = New(S, u) = O(
We begin by classifying the set of new edges ei ∈ New(S, u) into σ classes according to the
position of the divergence point bi. For every ei ∈ New(S, u), let(cid:98)s(ei) ∈ S be some source
node such that the divergence point bi ∈ π((cid:98)s(ei), u) appears on its(cid:98)s(ei)− u shortest path
σn).
T0(S). If there are several such sources for the edge ei, one is chosen arbitrarily.
For every sj ∈ S, let
New(sj) = {ei ∈ New(S, u) (cid:98)s(ei) = sj}
(cid:83)
be the set of new edges in New(S, u) that are mapped to sj ∈ S. Then, New(S, u) =
sj∈S New(sj). Let xj = New(sj).
We now focus on sj. For every ejk ∈ New(sj), k = {1, . . . , xj}, let Pjk = π(s(ejk), u, T (cid:48))
for T (cid:48) = T (s(ejk), f (ejk)) be the replacement path such that LastE(Pjk) = ejk and bjk
16
be its corresponding (last) divergence point with π(sj, u) (sj =(cid:98)s(ejk)). In addition, the
truncated path is given by P (cid:48)
= Pjk[bjk, u]. Note that LastE(Pjk) = ejk.
jk
Consider the set of divergence points bj1, . . . , bjxj
sorted in non-decreasing order of the
distance between bjk and sj on the shortest sj − u path π(sj) i.e., π(bjk, u, T0(sj)), where
π(bj1, u, T0(sj)) ≤ π(bj2, u, T0(sj)) . . . ≤ π(bjxj
(4)
Note that by Lemma 4.12, bj(cid:96) (cid:54)= bj(cid:96)(cid:48) for every (cid:96), (cid:96)(cid:48) ∈ {1, . . . , xj}. In addition, since each
∈ π(sj), combining with Eq. (4)
bj(cid:96) (cid:54)= u, π(bj1, u, T0(sj)) ≥ 1. Hence, since bj1, . . . , bjxj
, u, T0(sj)) .
we get that
1 ≤ π(bj1, u, T0(sj)) < π(bj2, u, T0(sj)) . . . < π(bjxj
is an alternative bj(cid:96) − u replacement path, we have that
, u, T0(sj)) .
Since P (cid:48)
j(cid:96)
P (cid:48)
≥ π(bj(cid:96), u, T0(sj)) ≥ (cid:96).
where the last inequality follows by Eq. (4). Hence, since all P (cid:48)
for the last node u, we get the total number of nodes V (sj) =(cid:83) V (P (cid:48)
j(cid:96)
j(cid:96)
are vertex disjoint, except
)\{u} occupied by
j(cid:96)
(5)
(6)
P (cid:48)
j(cid:96)
paths is
xj(cid:88)
V (P (cid:48)
j(cid:96)
) = V (sj) = O(x2
j ).
Since the nodes of V (sj1) and V (sj2) are disjoint for every sj1, sj2 ∈ S, by Lemma
j ) ≤ n. Therefore,
j=1 V (sj) = O(x2
4.12, it follows that New(S, u) = (cid:80)σ
New(S, u) =(cid:80)σ
√
j=1 xj ≤ O(
j=1 xj but (cid:80)σ
σn).
(cid:96)=1
As there are n nodes, combining with Cor. 4.10, we get that the total number of edges
in T ∗(S) is given by
E(T ∗(S)) ≤ E(T0(S)) +
(cid:88)
u∈V
New(S, u) ≤ σn + n · min{(cid:88)
si∈S
√
depth(si), O(
σn)},
as required. Thm. 4.9 is established. The analysis for the case of vertex faults follows
with almost no changes.
5 Hardness of Approximation of the Minimum FT-BFS
Problem
In this section we establish the following.
Theorem 5.1 The Minimum FT-BFS problem is NP-complete and cannot be approxi-
mated to within a factor c log n for some constant c > 0 unless NP ⊆ T IME(npoly log(n)).
17
We prove Theorem 5.1 by showing a gap preserving reduction from the Set-Cover problem
to the Minimum FT-BFS problem. An instance (cid:104)U, F(cid:105) of the Set-Cover problem consists
of a set of N elements U = {u1, . . . , uN} and a collection of M sets F = {S1, . . . , SM}
such that Si ⊆ U and (cid:83) Si = U . The task is to choose the minimal number of sets in
F whose union covers all of U . Fiege [13] showed that the Set Cover problem cannot be
approximated within a ratio of (1 − o(1)) ln n unless NP ⊆ T IME(M poly log(M )).
The Transformation. Given a Set-Cover instance (cid:104)U, F(cid:105), we construct a Minimum
FT-BFS instance I(U, F) = (G, s) as follows. Let X = {x1, . . . , xM} (resp., Z = {z1, . . . , zN})
be the vertex set corresponding to the collection of sets F (resp., elements of U ). Let
BXZ = (X, Z, EXZ) be the bipartite graph corresponding to the input (cid:104)U, F(cid:105), where
EXZ = {(xj, zi)
ui ∈ Sj, j ∈ {1, . . . , M} and i ∈ {1, . . . , N}}. Embed the bi-
partite graph BXZ in G in the following manner. Construct a length-(N + 1) path
P = [s = p0, p1 . . . , pN , pN +1], connect a vertex v(cid:48) to pN and connect a set of ver-
tices Y = {y1, . . . , yR} for R = O((M N )3) to the vertex pN +1 by the edges of EpY =
{(pN +1, yi)
i ∈ {1, . . . , R}}. Connect these vertices to the bipartite graph BXZ as
follows. For every i ∈ {1, . . . , N}, connect the node pi−1 of P to the node zi of Z
= zi] where ti = Qi = 6 + 2(N − i). Thus the
by a path Qi = [pi−1 = qi
paths Qi are monotonely decreasing and vertex disjoint. In addition, connect the vertices
v(cid:48) and pN +1 to every vertex of X, adding the edge sets EvX = {(v(cid:48), xi) xi ∈ X}
and EpX = {(pN +1, xj) xj ∈ X}. Finally, construct a complete bipartite graph
BXY = (X, Y, EXY ) where EXY = {(y(cid:96), xj) xj ∈ X, y(cid:96) ∈ Y }. This completes the
description of G. For illustration, see Fig. 4. Overall,
0, . . . , qi
ti
V (G) = X ∪ Z ∪ V (P ) ∪ N(cid:91)
V (Qi) ∪ {v(cid:48)} ∪ Y,
i=1
and
E(G) = EXZ ∪ E(P ) ∪ N(cid:91)
i=1
E(Qi) ∪ {(pN , v(cid:48))} ∪ EpY ∪ EvX ∪ EpX ∪ EXY .
Note that V (G) = O(R) and that E(G) = O(EXZ + N 2 + M R) = O(M R).
First, note the following.
Observation 5.2 Upon the failure of the edge ei = (pi−1, pi), i ∈ {1, . . . , N}, the follow-
ing happen:
(a) the unique s − zi shortest path in G \ {ei} is given by (cid:101)Pi = P [s, pi−1] ◦ Qi.
(b) the shortest-paths connecting s and the vertices of {pN , pN +1, v(cid:48)} ∪ X ∪ Y disconnect
and hence the replacement paths in G \ {ei} must go through the Z nodes.
18
Figure 4: Schematic illustration of the reduction from Set-Cover to Minimum FT-BFS. In
this example F = {S1, S2, . . . , S5} where S1 = {u1, u3, u4}, S2 = {u1, u3}, S3 = {a2, a4},
S4 = {a3} and S5 = {a1, a4}. Thus, N = 4 and M = 5. The minimal vertex cover is
given by S2 and S3. The vertex set Y is fully connected to X. In the optimal FT-BFS
T ∗, Y is required to be connected to the xj nodes that corresponds to the sets appearing
in the optimal cover. For example, y(cid:96) is connected to x2 and x3 which "covers" the Z
nodes. The red edges are necessary upon the fault of e3. All edges described except for
the (xj, y(cid:96)) edges are required in any FT-BFS tree.
We begin by observing that all edges except those of BXY are necessary in every
FT-BFS tree (cid:98)T ∈ T (s, G). Let (cid:101)E = E(G) \ EXY .
Observation 5.3 (cid:101)E ⊆ (cid:98)T for every (cid:98)T ∈ T (s, G).
Proof: The edges of the paths P and the edges of EpY ∪ {(pN , v(cid:48))} are trivially part of
every FT-BFS tree. The edges of the path Qi are necessary, by Obs. 5.2(a), upon the
failure of ei for every i ∈ {1, . . . , N}. To see that the edges of EvX are necessary, note
that upon the failure of the edge (pN , pN +1) or the edge (pN +1, xj), the unique s − xj
replacement path goes through v(cid:48) for every j ∈ {1, . . . , M}. Similarly, the edges EpX are
necessary upon the failure of (pN , v(cid:48)) or (v(cid:48), xj).
It remains to consider the edges of EXZ. Assume, towards contradiction, that there
exists some T (cid:48) ∈ T (s, G) that does not contain ej,i = (xj, zi) ∈ EXZ. Note that by Obs.
19
e3 p5 x1 x2 x3 x4 x5 z4 z3 z2 z1 s v' P Qi Y p1 p2 p3 p4 yl p0 BXY BXZ Z X EpY EpX EvX
(7)
5.2(a), upon the failure of the edge ei = (pi−1, pi) ∈ P , the unique s − xj shortest-path
in G \ {ei} is P (cid:48)
i = π[p0, pi−1] ◦ Qi ◦ [zi, xj], and all other alternatives are strictly longer.
Since ej,i /∈ T (cid:48), also P (cid:48)
(cid:42) T (cid:48), and therefore dist(s, xj, G \ {ei}) < dist(s, xj, T (cid:48) \ {ei}), in
contradiction to the fact that T (cid:48) ∈ T (s, G). The observation follows.
i
We now prove the correctness of the reduction and then consider gap-preservation.
E((cid:98)T ) ≥ (cid:101)E + κ((cid:98)T ) · R .
there exists some ui ∈ U not covered by F(cid:48). Consider the graph G(cid:48) = G \ {ei} where
Let (cid:98)T ∈ T (s, G) and define by Γ(y(cid:96),(cid:98)T ) = {xj
(xj, y(cid:96)) ∈ (cid:98)T} the X nodes that are
connected to y(cid:96) in (cid:98)T , for every y(cid:96) ∈ Y . Let κ((cid:98)T ) = min{Γ(y(cid:96),(cid:98)T ) y(cid:96) ∈ Y }. Note that
since the edges of (cid:101)E are necessary in every (cid:98)T ∈ T (s, G) it follows that
Lemma 5.4 If (cid:98)T ∈ T (s, G) then there exists a Set-Cover for (cid:104)U, F(cid:105) of size at most κ((cid:98)T ).
Proof: Consider (cid:98)T ∈ T (s, G) and let y(cid:96) ∈ Y be such that Γ(y(cid:96),(cid:98)T ) = κ((cid:98)T ). A cover F(cid:48)
for U for size κ((cid:98)T ) is constructed as follows. Let F(cid:48) = {Sj
xj ∈ Γ(y(cid:96),(cid:98)T )}. By definition,
F(cid:48) = κ((cid:98)T ). We now claim that it is a cover for U . Assume, towards contradiction, that
ei = (pi−1, pi). Recall that by Obs. 5.2(a), (cid:101)Pk = P [s, pk−1] ◦ Qk is the s − zk path in
G \ {ek}. Note that (cid:101)Pk (cid:42) G(cid:48) for every k > i and (cid:101)Pk > (cid:101)Pi for every k < i. Hence
j = (cid:101)Pi ◦ (zi, xj). Therefore the s − y(cid:96) shortest-paths in G(cid:48) are all given by
j ◦ (xj, y(cid:96)), for every xj ∈ Γ(zi). But since (xj, y(cid:96)) /∈ (cid:98)T for every xj ∈ Γ(zi), we have that
dist(s, y(cid:96), G(cid:48)) < dist(s, y(cid:96),(cid:98)T \ {ei}), in contradiction to the fact that (cid:98)T ∈ T (s, G).
Lemma 5.5 If there exists a Set-Cover of size κ then Cost∗(s, G) ≤ (cid:101)E + κ · R.
Proof: Given a cover F(cid:48) ⊆ F, F(cid:48) = κ, construct a FT-BFS tree (cid:98)T ∈ T (s, G) with (cid:101)E+κ·R
edges as follows. Add (cid:101)E to (cid:98)T . In addition, for every Sj ∈ F(cid:48), add the edge (y(cid:96), xj) to (cid:98)T
for every y(cid:96) ∈ Y . Clearly, E((cid:98)T ) = (cid:101)E + κ · R. It remains to show that (cid:98)T ∈ T (s, G).
denoting the set of neighbors of zi in X by Γ(zi) = {xj
(zi, xj) ∈ EXZ}, by Obs.
5.2(b), the unique s − xj shortest-path, for every xj ∈ Γ(zi) such that (zi, xj) ∈ EXY ,
is given by P (cid:48)
P (cid:48)
Note that there is no s − u replacement path that uses any y(cid:96) ∈ Y as a relay, for any
u ∈ V (G) and y(cid:96) ∈ Y ; this holds as X is connected by two alternative shortest-paths to
both pN +1 and to v(cid:48) and the path through y(cid:96) is strictly longer. In addition, if the edge
e ∈ {(pN , pN +1), (pN +1, y(cid:96))} fails, then the s − y(cid:96) shortest path in G \ {e} goes through
any neighbor xj of y(cid:96). Since each y(cid:96) has at least one X node neighbor in (cid:98)T , it holds that
dist(s, y(cid:96),(cid:98)T \ {e}) = dist(s, y(cid:96), G \ {e}).
Since the only missing edges of (cid:98)T , namely, E(G)\E((cid:98)T ), are the edges of EXY , it follows
path in G \ {ei}, which is by Obs. 5.2(b), P (cid:48) = (cid:101)Pi ◦ (zi, xj) ◦ (xj, y(cid:96)), exists in (cid:98)T \ {ei} for
that it remains to check the edges ei = (vi−1, vi) for every i ∈ {1, . . . , N}. Let Sj ∈ F(cid:48) such
that ui ∈ Sj. Since F(cid:48) is a cover, such Sj exists. Hence, the optimal s − y(cid:96) replacement
20
every y(cid:96) ∈ Y . It follows that (cid:98)T ∈ T (s, G), hence Cost∗(s, G) ≤ E((cid:98)T ) = (cid:101)E + κ· R. The
lemma follows.
Let κ∗ be the cost of the optimal Set-Cover for the instance (cid:104)U, F(cid:105). We have the
following.
Corollary 5.6 Cost∗(s, G) = (cid:101)E + κ∗ · R.
Proof: Let T ∗ ∈ T (s, G) be such that E(T ∗) = Cost∗(s, G). It then holds that
(cid:101)E + κ(T ∗) · R ≤ E(T ∗) = Cost
∗(s, G) ≤ (cid:101)E + κ∗ · R,
where the first inequality holds by Eq. (7) and the second inequality follows by Lemma
5.5. Hence, κ(T ∗) ≤ κ∗. Since by Lemma 5.4, there exists a cover of size κ(T ∗), we have
that κ∗ ≤ κ(T ∗). It follows that κ∗ = κ(T ∗) and Cost∗(s, G) = (cid:101)E + κ∗ · R as desired.
We now show that the reduction is gap-preserving. Assume that there exists an
α approximation algorithm A for the Minimum FT-BFS problem. Then applying our
transformation to an instance I(U, F) = (G, s) would result in an FT-BFS tree (cid:98)T ∈ T (s, G)
such that
(cid:101)E + κ((cid:98)T ) · R < E((cid:98)T ) ≤ α((cid:101)E + κ∗ · R) ≤ 3α · κ∗ · R ,
where the first inequality follows by Eq. (7), the second by the approximation guarantee
of A and by Cor. 5.6, and the third inequality follows by the fact that (cid:101)E ≤ 2R. By
Lemma 5.4, a cover of size κ((cid:98)T ) ≤ 3ακ∗ can be constructed given (cid:98)T , which results in a 3α
approximation to the Set-Cover instance. As the Set-Cover problem is inapproximable
within a factor of (1 − o(1)) ln n, under an appropriate complexity assumption [13], we
get that the Minimum FT-BFS problem is inapproximable within a factor of c · log N for
some constant c > 0. This complete the proof of Thm. 5.1.
6 O(log n)-Approximation for FT-MBFS Trees
In Sec. 4.1, we presented an algorithm that for every graph G and source s constructs
an FT-BFS tree (cid:98)T ∈ T (s, G) with O(n3/2) edges. In Sec. 3.1, we showed that there exist
FT-BFS (cid:98)T ∈ T (s(cid:48), G(cid:48)) which is denser by a factor of Ω(
FT-BFS tree, i.e., such that E((cid:98)T ) ≥ Ω(
graphs G and s ∈ V (G) for which Cost∗(s, G) = Ω(n3/2), establishing tightness of our
algorithm in the worst-case. Yet, there are also inputs (G(cid:48), s(cid:48)) for which the algorithm of
√
Sec. 4, as well as algorithms based on the analysis of [14] and [23], might still produce an
n) than the size of the optimal
n) · Cost∗(s(cid:48), G(cid:48)). For an illustration of such a
case consider the graph G(cid:48) = (V, E) which is a modification of the graph G described in
√
21
Sec. 3.1. The modifications are as follows. First, add a node z0 to Z and connect it to
every xi ∈ X. Replace the last edge e(cid:48)
i = LastE(Pi) of the vi − zi path Pi by a vertex ri
that is connected to the endpoints of the edge e(cid:48)
i be the
s − zi modified path where LastE(P (cid:48)
i ) = (ri, zi). Finally, connect the node z0 to all nodes
ri for every i ∈ {1, . . . , d}. See Fig. 5 for illustration.
i for every i ∈ {1, . . . , d}. Let P (cid:48)
Figure 5: Bad example for the algorithm of Sec. 4. The weights of the z0 edges are larger
than those of the other edges. Thus, the entire complete bipartite graph B(X, Z \ {z0})
algorithm. However, an FT-BFS tree T ∗ of O(n) edges can be given by including the edges
of (z0, xi) for every xi ∈ X. The red edges are two optional edges necessary upon the
failure of ei. Adding the edge (xj, z0) is better, yet the algorithm of Sec. 4 adds (xj, zi)
of size Ω(n3/2) is included in the resulting FT-BFS tree (cid:98)T ∈ T (s, G) returned by the
to (cid:98)T for every xj ∈ X.
as the edges of the complete bipartite graph B that are required in every (cid:98)T ∈ T (s, G) are
certain weight assignments the algorithm of Sec. 4 constructs an FT-BFS tree (cid:98)T of size
Observe that whereas Cost∗(s, G) = Ω(n3/2), the modified G(cid:48) has Cost∗(s, G(cid:48)) = O(n),
no longer required in every T (cid:48) ∈ T (s, G(cid:48)); it is sufficient to connect the nodes of X to z0
only, and by that "save" the Ω(n3/2) edges of B in T (cid:48). Nevertheless, as we show next, for
O(n3/2). Specifically, let W be such that each of the edges of
E(cid:48) = {(z0, ri) i ∈ {1, . . . , d}} ∪ {(z0, xi) xi ∈ X}
is assigned a weight which is strictly larger than the weights of the other edges. That
is, W (ek) > W (e(cid:96)) for every ek ∈ E(cid:48) and e(cid:96) ∈ E(G(cid:48)) \ E(cid:48). Note that for every edge
ei = (vi, vi+1) ∈ π, i ∈ {1, . . . , d}, there are two alternative s − xj replacement paths
of the same length, namely, Qi,j = π[s, vi] ◦ P (cid:48)
π[s, vi]◦P (cid:48)
i ◦ (zi, xj) that goes through zi and (cid:98)Qi,j =
i [s, ri]◦(ri, z0)◦(z0, xi) that goes through z0. Although Qi,j = (cid:98)Qi,j, the weight
22
ei v2 vi vi+1 r1 s vd P'i 𝑑=𝑂(𝑛) z0 z1 zi ri rd zd v* X xj Z R B assignment implies that Cost(Qi,j) < Cost((cid:98)Qi,j) and hence (cid:98)Qi,j /∈ SP (s, xj, G \ {ei}, W )
for every i ∈ {1, . . . , d} and every xj ∈ X. Therefore, E(B) ⊆ (cid:98)T , for every FT-BFS
tree (cid:98)T computed by the algorithm of Sec. 4 with the weight assignment W . Hence
E((cid:98)T ) = Θ(n3/2) while Cost∗(s, G(cid:48)) = O(n).
Clearly, a universally optimal algorithm is unlikely given the hardness of approximation
result of Thm. 5.1. Yet the gap can be narrowed down. The goal of this section is to
present an O(log n) approximation algorithm for the Minimum FT-BFS Problem (hence
also to its special case, the Minimum FT-BFS Problem, where S = 1).
To establish this result, we first describe the algorithm and then bound the number of
edges. Let ApproxSetCover(F, U ) be an O(log n) approximation algorithm for the Set-
Cover problem, which given a collection of sets F = {S1, . . . , SM} that covers a universe
U = {u1, . . . , uN} of size N , returns a cover F(cid:48) ⊆ F that is larger by at most O(log N )
than any other F(cid:48)(cid:48) ⊆ F that covers U (cf. [27]).
The Algorithm Starting with (cid:98)T = ∅, the algorithm adds edges to (cid:98)T until it becomes
an FT-MBFS tree.
Set an arbitrary order on the vertices V (G) = {v1, . . . , vn} and on the edges E+ =
E(G) ∪ {e0} = {e0, . . . , em} where e0 is a new fictitious edge whose role will be explained
later on. For every node vi ∈ V , define
Ui = {(cid:104)sk, ej(cid:105) sk ∈ S \ {vi}, ej ∈ E+}.
The algorithm consists of n rounds, where in round i it considers vi. Let Γ(vi, G) =
{u1, . . . , udi} be the set of neighbors of vi in some arbitrary order, where di = deg(vi, G).
For every neighbor uj, define a set Si,j ⊆ Ui containing certain source-edge pairs (cid:104)sk, e(cid:96)(cid:105) ∈
Ui. Informally, a set Si,j contains the pair (cid:104)sk, e(cid:96)(cid:105) iff there exists an sk−vi shortest path in
G \ {e(cid:96)} that goes through the neighbor uj of vi. Note that Si,j contains the pair (cid:104)sk, e0(cid:105)
iff there exists an sk − vi shortest-path in G \ {e0} = G that goes through uj. I.e., the
fictitious edge e0 is meant to capture the case where no fault occurs, and thus we take
care of true shortest-paths in G. Formally, every pair (cid:104)sk, e(cid:96)(cid:105) ∈ Ui is included in every set
Si,j satisfying that
dist(sk, uj, G \ {e(cid:96)}) = dist(sk, vi, G \ {e(cid:96)}) − 1.
(8)
Let Fi = {Si,1, . . . , Si,di}. The edges of vi that are added to (cid:98)T in round i are now selected
i, add the edge (uj, vi) to (cid:98)T . We now turn to prove the correctness of
by using algorithm ApproxSetCover to generate an approximate solution for the set cover
problem on the collection F = {Si,j
i = ApproxSetCover(Fi, Ui).
For every Si,j ∈ F(cid:48)
this algorithm and establish Thm. 6.5.
uj ∈ Γ(vi, G)}. Let F(cid:48)
23
bound its size.
Analysis We first show that algorithm constructs an FT-MBFS (cid:98)T ∈ T (S, G) and then
Lemma 6.1 (cid:98)T ∈ T (S, G).
Proof: Assume, towards contradiction, that (cid:98)T /∈ T (S, G). Let s ∈ S be some source
node such that (cid:98)T /∈ T (s, G) is not an FT-BFS tree with respect to s. By the assumption,
such s exists. Let
BP = {(i, k) vi ∈ V, ek ∈ E+ and dist(s, vi,(cid:98)T \ {ek}) > dist(s, vi, G \ {ek})}
distance in (cid:98)T \{ek} is greater than that in G\{ek}. (By the assumption that (cid:98)T /∈ T (s, G),
be the set of "bad pairs," namely, vertex-edge pairs (i, k) for which the s−vi shortest path
it holds that BP (cid:54)= ∅.) For every vertex-edge pair (i, k), where vi ∈ V \{s} and ek ∈ E+,
define an s − vi shortest-path P ∗
i,k in G \ {ek} in the following manner. Let uj ∈ Γ(vi, G)
in the cover returned by the algorithm ApproxSetCover in round i. Thus, (uj, vi) ∈ (cid:98)T
be such that the pair (cid:104)s, ek(cid:105) ∈ Si,j is covered by the set Si,j of uj and Si,j ∈ F(cid:48)
i is included
and dist(s, uj, G\{ek}) = dist(s, vi, G\{ek})− 1. Let P (cid:48) ∈ SP (s, uj, G\{ek}) and define
i,k = P (cid:48) ◦ (uj, vi).
P ∗
By definition, P ∗
BE(i, k) = P ∗
i,k \ E((cid:98)T ) to be the set of "bad edges," namely, the set of P ∗
i,k = dist(s, vi, G \ {ek}) and by construction, LastE(P ∗
i,k) ∈ (cid:98)T . Define
are missing in (cid:98)T . By definition, BE(i, k) (cid:54)= ∅ for every bad pair (i, k) ∈ BP . Let
i,k edges that
i,k)} be the maximal depth of a missing edge in BE(i, k),
d(i, k) = maxe∈BE(i,k){dist(s, e, P ∗
and let DM (i, k) denote that "deepest missing edge" for (i, k), i.e., the edge e on P ∗
i,k). Finally, let (i(cid:48), k(cid:48)) ∈ BP be the pair that minimizes
satisfying d(i, k) = dist(s, e, P ∗
d(i, k), and let e1 = (v(cid:96)1, vi1) ∈ BE(i(cid:48), k(cid:48)) be the deepest missing edge on P ∗
i(cid:48),k(cid:48), namely,
e1 = DM (i(cid:48), k(cid:48)). Note that e1 is the shallowest "deepest missing edge" over all bad
pairs (i, k) ∈ BP . Let P1 = P ∗
i(cid:48),k(cid:48)[vi1, vi(cid:48)]; see Fig. 6 for
(i1, k(cid:48)) /∈ BP , then any s−vi1 shortest-path P (cid:48) ∈ SP (s, vi1,(cid:98)T\{ek(cid:48)}) , where P (cid:48) = P ∗
illustration. Note that since (i(cid:48), k(cid:48)) ∈ BP , it follows that also (i1, k(cid:48)) ∈ BP . (Otherwise, if
can be appended to P3 resulting in P (cid:48)(cid:48) = P (cid:48) ◦ P3 such that (1) P (cid:48)(cid:48) ⊆ (cid:98)T \ {ek(cid:48)} and (2)
i1,k(cid:48),
we conclude that (i1, k(cid:48)) ∈ BP . Finally, note that LastE(P1) ∈ (cid:98)T by definition, and
i(cid:48),k(cid:48), contradicting the fact that (i(cid:48), k(cid:48)) ∈ BP .) Thus
P (cid:48)(cid:48) = P (cid:48) + P3 = P2 + P3 = P ∗
i1,k(cid:48), P2 = P ∗
i(cid:48),k(cid:48)[s, vi1] and P3 = P ∗
therefore the deepest missing edge of (i, k) must be shallower, i.e., d(i1, k(cid:48)) < d(i(cid:48), k(cid:48)).
However, this is in contradiction to our choice of the pair (i(cid:48), k(cid:48)). The lemma follows.
i,k
Let W : E(G) → R>0 be the weight assignment that guarantees the uniqueness
of shortest-paths. Note that the algorithm did not use W in the computation of the
as considered by the algorithm. For every FT-MBFS tree (cid:101)T ∈ T (S, G), vi ∈ V, e(cid:96) ∈ E+ and
shortest-paths. For every node vi, let Γ(vi, G) = {u1, . . . , udi} be its ordered neighbor set
24
Figure 6: Red solid lines correspond to new edges. The "deepest missing edge" for (i(cid:48), k(cid:48)),
edge e1, is the shallowest such edge over all bad pairs in BP . Yet the pair (i1, k(cid:48)) is bad
too. As the last (green) edge of P1 is included in the FT-MBFS tree, and since P1 and P2
are of the same length, it follows that P1 has a shallower "deepest missing edge".
sk ∈ S, let (cid:101)Pi(sk, e(cid:96)) ∈ SP (sk, vi,(cid:101)T \ {e(cid:96)}, W ) be an sk − vi shortest-path in (cid:101)T \ {e(cid:96)}. Let
Ai((cid:101)T ) = {LastE((cid:101)Pi(sk, e(cid:96))) e(cid:96) ∈ E+, sk ∈ S \ {vi}}
be the edges of vi that appear as last edges in the shortest-paths and replacement paths
(9)
from S to vi in (cid:101)T . Define
We then have that
Fi((cid:101)T ) = {Si,j
(uj, vi) ∈ Ai((cid:101)T )}.
Fi((cid:101)T ) = Ai((cid:101)T ) .
25
s vi' vi1 e1 P2 P3 ',*11kiPP (cid:104)sk, e(cid:96)(cid:105) ∈ Ui \ (cid:91)
Si,j∈Fi((cid:101)T )
Si,j .
(10)
show the reverse direction.
The correctness of the algorithm (see Lemma 6.1) established that if a subgraph (cid:101)T ⊆ G
satisfies that Fi((cid:101)T ) is a cover of Ui for every vi ∈ V , then (cid:101)T ∈ T (S, G). We now turn to
Lemma 6.2 For every (cid:101)T ∈ T (S, G), the collection Fi((cid:101)T ) is a cover of Ui, namely,
(cid:83)
Si,j∈Fi((cid:101)T ) Si,j = Ui,
Proof: Assume, towards contradiction, that there exists an FT-MBFS tree (cid:101)T ∈ T (S, G)
and a vertex vi ∈ V whose corresponding collection of sets Fi((cid:101)T ) does not cover Ui. Hence
for every vi ∈ V .
there exists at least one uncovered pair (cid:104)sk, e(cid:96)(cid:105) ∈ Ui, i.e.,
By definition sk (cid:54)= vi. We next claim that (cid:101)T does not contain an optimal sk − vi path
when the edge e(cid:96) fails, contradicting the fact that (cid:101)T ∈ T (S, G). That is, we show that
dist(sk, vi,(cid:101)T \ {e(cid:96)}) > dist(sk, vi, G \ {e(cid:96)}).
i,(cid:96)) where P ∗
assumption, P ∗
Towards contradiction, assume otherwise, and let (uj, vi) = LastE(P ∗
SP (sk, vi,(cid:101)T \ {e(cid:96)}, W ), hence (uj, vi) ∈ Ai((cid:101)T ) and Si,j ∈ Fi((cid:101)T ). By the contradictory
i,(cid:96) ∈
{e(cid:96)}) − 1. This implies that (cid:104)sk, e(cid:96)(cid:105) ∈ Si,j ∈ Fi((cid:101)T ), in contradiction to Eq. (10), stating
i,(cid:96) = dist(sk, vi, G \ {e(cid:96)}) and hence dist(sk, uj, G \ {e(cid:96)}) = dist(sk, vi, G \
that (cid:104)sk, e(cid:96)(cid:105) is not covered by Fi((cid:101)T ). The lemma follows.
We now turn to bound that number of edges in (cid:98)T .
Lemma 6.3 E((cid:98)T ) ≤ O(log n) · Cost∗(S, G).
Proof: Let δ = c log n be the approximation ratio guarantee of ApproxSetCover. For ease
of notation, let Oi = Ai(T ∗) for every vi ∈ V . Let Fi = {Si,1, . . . , Si,di} be the collection
of vi sets considered at round i where Si,j ⊆ Ui is the set of the neighbor uj ∈ Γ(vi, G)
computed according to Eq. (8).
i = ApproxSetCover(Si, Ui) be the cover returned by the algorithm and define
i} as the collection of edges whose corresponding sets are included
Let F(cid:48)
Ai = {(uj, vi) Si,j ∈ F(cid:48)
in S(cid:48)
i. Thus, by Eq. (9), Oi = Fi(T ∗) and Ai = F(cid:48)
Observation 6.4 Ai ≤ δOi for every vi ∈ V \ {s}.
Proof: Assume, towards contradiction, that there exists some i such that Ai > δOi.
(9) and by the approximation guarantee of ApproxSetCover where in
Then by Eq.
i ⊆ Fi that covers Ui, it follows that Fi(T ∗) is not
particular Fi((cid:101)T ) ≤ δF(cid:48)(cid:48)
i for every vi ∈ V .
i for every F(cid:48)(cid:48)
26
a cover of Ui. Consequently, it follows by Lemma 6.2 that T ∗ /∈ T (S, G), contradiction.
The observation follows.
Since(cid:83) Ai contains precisely the edges that are added by the algorithm to the con-
structed FT-MBFS tree (cid:98)T , we have that
E((cid:98)T ) ≤ (cid:88)
i Oi/2 (as every edge in(cid:83)
(cid:80)
where the second inequality follows by Obs. 6.4 and the third by the fact that E(T ∗) ≥
vi∈V Oi can be counted at most twice, by both its endpoints).
Ai ≤ δ
Oi ≤ 2δ · Cost
∗(S, G) ,
(cid:88)
i
i
The lemma follows.
The following theorem is established.
Theorem 6.5 There exists a polynomial time algorithm that for every n-vertex graph G
and source node set S ⊆ V constructs an FT-MBFS tree (cid:98)T ∈ T (S, G) such that E((cid:98)T ) ≤
O(log n) · Cost∗(S, G).
Acknowledgment We are grateful to Gilad Braunschvig, Alon Brutzkus, Adam Sealfon
and Oren Weimann for helpful discussions.
27
References
[1] B. Awerbuch, A. Bar-Noy, N. Linial, and D. Peleg. Compact distributed data struc-
tures for adaptive network routing. In Proc. 21st ACM Symp. on Theory of Com-
puting, 230 -- 240, 1989.
[2] I. Abraham, S. Chechik, C. Gavoille and D. Peleg. Forbidden-Set Distance Labels for
Graphs of Bounded Doubling Dimension. In Proc. 29th ACM Symp. on Principles
of Distributed Computing, 2010, 192 -- 200.
[3] S. Baswana and S. Sen. Approximate distance oracles for unweighted graphs in
expected O(n2) time. ACM Trans. Algorithms, 2(4):557 -- 577, 2006.
[4] A. Bernstein and D. Karger. A nearly optimal oracle for avoiding failed vertices and
edges. In Proc. 41st ACM Symp. on Theory of Computing, 101 -- 110, 2009.
[5] S. Chechik, M. Langberg, D. Peleg, and L. Roditty. f -sensitivity distance oracles
and routing schemes. Algorithmica, 861 -- 882, 2012.
[6] S. Chechik, M. Langberg, D. Peleg, and L. Roditty. Fault-tolerant spanners for
general graphs. In Proc. 41st ACM Symp. on Theory of computing, 435 -- 444, 2009.
[7] S. Chechik. Fault-Tolerant Compact Routing Schemes for General Graphs. In Proc.
38th Int. Colloq. on Automata, Languages & Prog., 101 -- 112, 2011.
[8] B. Courcelle and A. Twigg. Compact forbidden-set routing. In Proc. 24th Symp. on
Theoretical Aspects of Computer Science, 37 -- 48, 2007.
[9] A. Czumaj and H. Zhao. Fault-tolerant geometric spanners. Discrete & Computa-
tional Geometry, 32, 2003.
[10] C. Demetrescu, M. Thorup, R. Chowdhury, and V. Ramachandran. Oracles for
distances avoiding a failed node or link. SIAM J. Computing, 37:1299 -- 1318, 2008.
[11] M. Dinitz and R. Krauthgamer. Fault-tolerant spanners: better and simpler.
In
Proc. ACM Symp. on Principles of Distributed Computing, 2011, 169-178.
[12] R. Duan and S. Pettie. Dual-failure distance and connectivity oracles. In Proc. 20th
ACM-SIAM Symp. on Discrete Algorithms, 2009.
[13] U. Feige. A Threshold of ln n for Approximating Set Cover. J. ACM, 634 -- 652, 1998.
[14] F. Grandoni and V.V Williams.
Single-Source Replacement Paths.
Computer Science, 2012.
Improved Distance Sensitivity Oracles via Fast
In Proc. 53rd IEEE Symp. on Foundations of
28
[15] C. Levcopoulos, G. Narasimhan, and M. Smid. Efficient algorithms for constructing
fault-tolerant geometric spanners. In Proc. 30th ACM Symp. on Theory of computing,
186 -- 195, 1998.
[16] T. Lukovszki. New results of fault tolerant geometric spanners. In Proc. 6th Workshop
on Algorithms and Data Structures, London 193 -- 204, 1999.
[17] D. Peleg. Distributed Computing: A Locality-Sensitive Approach. SIAM, 2000.
[18] D. Peleg. As good as it gets: Competitive fault tolerance in network structures.
In Proc. 11th Symp. on Stabilization, Safety, and Security of Distributed Systems,
LNCS 5873, 2009, 35 -- 46.
[19] D. Peleg and A.A. Schaffer. Graph spanners. J. Graph Theory, 13:99 -- 116, 1989.
[20] D. Peleg and J.D. Ullman. An optimal synchronizer for the hypercube. SIAM J.
Computing, 18(2):740 -- 747, 1989.
[21] D. Peleg and E. Upfal. A trade-off between space and efficiency for routing tables.
J. ACM, 36:510 -- 530, 1989.
[22] L. Roditty, M. Thorup, and U. Zwick. Deterministic constructions of approximate
distance oracles and spanners. In Proc. 32nd Int. Colloq. on Automata, Languages
& Prog., 261 -- 272, 2005.
[23] L. Roditty and U. Zwick. Replacement paths and k simple shortest paths in un-
weighted directed graphs. ACM Trans. Algorithms ,2012.
[24] M. Thorup and U. Zwick. Compact routing schemes. In Proc. 14th ACM Symp. on
Parallel Algorithms and Architecture, Hersonissos, Crete, 1 -- 10, 2001.
[25] M. Thorup and U. Zwick. Approximate distance oracles. J. ACM, 52:1 -- 24, 2005.
[26] M. Thorup. Undirected single-source shortest paths with positive integer weights in
linear time. J. ACM, 362 -- 394, 1999.
[27] V. Vazirani. Approximation Algorithms. College of Computing, Georgia Institute of
Technology, 1997.
[28] O. Weimann and R. Yuster. Replacement paths via fast matrix multiplication. In
Proc. 51th IEEE Symp. on Foundations of Computer Science, 2010.
29
|
1508.01657 | 2 | 1508 | 2016-03-24T16:52:18 | A parameterized complexity view on non-preemptively scheduling interval-constrained jobs: few machines, small looseness, and small slack | [
"cs.DS",
"cs.DM",
"math.CO"
] | We study the problem of non-preemptively scheduling $n$ jobs, each job $j$ with a release time $t_j$, a deadline $d_j$, and a processing time $p_j$, on $m$ parallel identical machines. Cieliebak et al. (2004) considered the two constraints $|d_j-t_j|\leq \lambda p_j$ and $|d_j-t_j|\leq p_j +\sigma$ and showed the problem to be NP-hard for any $\lambda>1$ and for any $\sigma\geq 2$. We complement their results by parameterized complexity studies: we show that, for any $\lambda>1$, the problem remains weakly NP-hard even for $m=2$ and strongly W[1]-hard parameterized by $m$. We present a pseudo-polynomial-time algorithm for constant $m$ and $\lambda$ and a fixed-parameter tractability result for the parameter $m$ combined with $\sigma$. | cs.DS | cs | Journal of Scheduling manuscript No.
(will be inserted by the editor)
A parameterized complexity view on non-preemptively scheduling
interval-constrained jobs: few machines, small looseness, and small slack
Ren´e van Bevern · Rolf Niedermeier · Ondrej Such´y
6
1
0
2
r
a
M
4
2
]
S
D
.
s
c
[
2
v
7
5
6
1
0
.
8
0
5
1
:
v
i
X
r
a
Submitted: August 7, 2015
Accepted: March 23, 2016
Abstract We study the problem of non-preemptively sched-
uling n jobs, each job j with a release time t j, a deadline d j,
and a processing time p j, on m parallel identical machines.
Cieliebak et al (2004) considered the two constraints d j −
t j ≤ λ p j and d j −t j ≤ p j + σ and showed the problem to
be NP-hard for any λ > 1 and for any σ ≥ 2. We comple-
ment their results by parameterized complexity studies: we
show that, for any λ > 1, the problem remains weakly NP-
hard even for m = 2 and strongly W[1]-hard parameterized
by m. We present a pseudo-polynomial-time algorithm for
constant m and λ and a fixed-parameter tractability result for
the parameter m combined with σ.
Keywords release times and deadlines · machine mini-
mization · sequencing within intervals · shiftable intervals ·
fixed-parameter tractability · NP-hard problem
1 Introduction
Non-preemptively scheduling jobs with release times and
deadlines on a minimum number of machines is a well-stud-
ied problem both in offline and online variants (Chen et al
Ren´e van Bevern is supported by grant 16-31-60007 mol a dk of the
Russian Foundation for Basic Research (RFBR).
Ondrej Such´y is supported by grant 14-13017P of the Czech Science
Foundation.
Ren´e van Bevern
Novosibirsk State University, Novosibirsk, Russian Federation, E-mail:
[email protected]
Rolf Niedermeier
Institut fur Softwaretechnik und Theoretische Informatik, TU Berlin,
Germany, E-mail: [email protected]
Ondrej Such´y
Faculty of Information Technology, Czech Technical University in
Prague, Prague, Czech Republic, E-mail: [email protected]
2016; Chuzhoy et al 2004; Cieliebak et al 2004; Malucelli
and Nicoloso 2007; Saha 2013). In its decision version, the
problem is formally defined as follows:
INTERVAL-CONSTRAINED SCHEDULING
Input: A set J := {1, . . . ,n} of jobs, a number m ∈ N of
machines, each job j with a release time t j ∈ N, a dead-
line d j ∈ N, and a processing time p j ∈ N.
Question: Is there a schedule that schedules all jobs onto
m parallel identical machines such that
1. each job j is executed non-preemptively for p j time units,
2. each machine executes at most one job at a time, and
3. each job j starts no earlier than t j and is finished by d j.
For a job j ∈ J, we call the half-open interval [t j,d j) its time
window. A job may only be executed during its time window.
The length of the time window is d j −t j.
We study INTERVAL-CONSTRAINED SCHEDULING with
two additional constraints introduced by Cieliebak et al (2004).
These constraints relate the time window lengths of jobs to
their processing times:
Looseness If all jobs j ∈ J satisfy d j −t j ≤ λ p j for some
number λ ∈ R, then the instance has looseness λ . By λ -
LOOSE INTERVAL-CONSTRAINED SCHEDULING we denote
the problem restricted to instances of looseness λ .
Slack If all jobs j ∈ J satisfy d j − t j ≤ p j + σ for some
number σ ∈ R, then the instance has slack σ. By σ -SLACK
INTERVAL-CONSTRAINED SCHEDULING we denote the
problem restricted to instances of slack σ.
Both constraints on INTERVAL-CONSTRAINED SCHEDUL-
ING are very natural: clients may accept some small deviation
of at most σ from the desired start times of their jobs. More-
over, it is conceivable that clients allow for a larger deviation
2
Ren´e van Bevern et al.
for jobs that take long to process anyway, leading to the case
of bounded looseness λ .
Cieliebak et al (2004) showed that, even for constant λ >
1 and constant σ ≥ 2, the problems λ -LOOSE INTERVAL-
CONSTRAINED SCHEDULING and σ -SLACK INTERVAL-
CONSTRAINED SCHEDULING are strongly NP-hard.
Instead of giving up on finding optimal solutions and
resorting to approximation algorithms (Chuzhoy et al 2004;
Cieliebak et al 2004), we conduct a more fine-grained com-
plexity analysis of these problems employing the frame-
work of parameterized complexity theory (Cygan et al 2015;
Downey and Fellows 2013; Flum and Grohe 2006; Nieder-
meier 2006), which so far received comparatively little atten-
tion in the field of scheduling with seemingly only a handful
of publications (van Bevern et al 2015a,b; Bodlaender and
Fellows 1995; Cieliebak et al 2004; Fellows and McCartin
2003; Halld´orsson and Karlsson 2006; Hermelin et al 2015;
Mnich and Wiese 2015). In particular, we investigate the
effect of the parameter m of available machines on the pa-
rameterized complexity of interval-constrained scheduling
without preemption.
Related work INTERVAL-CONSTRAINED SCHEDULING is
a classical scheduling problem and strongly NP-hard already
on one machine (Garey and Johnson 1979, problem SS1).
Besides the task of scheduling all jobs on a minimum number
of machines, the literature contains a wide body of work
concerning the maximization of the number of scheduled
jobs on a bounded number of machines (Kolen et al 2007).
For the objective of minimizing the number of machines,
(cid:112)logn/loglogn)-
Chuzhoy et al (2004) developed a factor-O(
approximation algorithm. Malucelli and Nicoloso (2007) for-
malized machine minimization and other objectives in terms
of optimization problems in shiftable interval graphs. Online
algorithms for minimizing the number of machines have been
studied as well and we refer to recent work by Chen et al
(2016) for an overview.
Our work refines the following results of Cieliebak et al
(2004), who considered INTERVAL-CONSTRAINED SCHED-
ULING with bounds on the looseness and the slack. They
showed that INTERVAL-CONSTRAINED SCHEDULING is
strongly NP-hard for any looseness λ > 1 and any slack σ ≥
2. Besides giving approximation algorithms for various spe-
cial cases, they give a polynomial-time algorithm for σ = 1
and a fixed-parameter tractability result for the combined
parameter σ and h, where h is the maximum number of time
windows overlapping in any point in time.
Our contributions We analyze the parameterized complexity
of INTERVAL-CONSTRAINED SCHEDULING with respect
to three parameters: the number m of machines, the loose-
ness λ , and the slack σ. More specifically, we refine known
results of Cieliebak et al (2004) using tools of parameterized
complexity analysis. An overview is given in Table 1.1.
In Section 3, we show that, for any λ > 1, λ -LOOSE
INTERVAL-CONSTRAINED SCHEDULING remains weakly
NP-hard even on m = 2 machines and that it is strongly W[1]-
hard when parameterized by the number m of machines. In
Section 4, we give a pseudo-polynomial-time algorithm for λ -
LOOSE INTERVAL-CONSTRAINED SCHEDULING for each
fixed λ and m. Finally, in Section 5, we give a fixed-para-
meter algorithm for σ -SLACK INTERVAL-CONSTRAINED
SCHEDULING when parameterized by m and σ. This is in
contrast to our result from Section 3 that the parameter com-
bination m and λ presumably does not give fixed-parameter
tractability results for λ -LOOSE INTERVAL-CONSTRAINED
SCHEDULING.
2 Preliminaries
Basic notation We assume that 0 ∈ N. For two vectors u =
(u1, . . . ,uk) and v = (v1, . . . ,vk), we write u ≤ v if ui ≤ vi
for all i ∈ {1, . . . ,k}. Moreover, we write u (cid:8) v if u ≤ v and
u(cid:54)= v, that is, u and v differ in at least one component. Finally,
1k is the k-dimensional vector consisting of k 1-entries.
Computational complexity We assume familiarity with the
basic concepts of NP-hardness and polynomial-time many-
one reductions (Garey and Johnson 1979). We say that a
problem is (strongly) C-hard for some complexity class C if it
is C-hard even if all integers in the input instance are bounded
from above by a polynomial in the input size. Otherwise, we
call it weakly C-hard.
In the following, we introduce the basic concepts of pa-
rameterized complexity theory, which are in more detail
discussed in corresponding text books (Cygan et al 2015;
Downey and Fellows 2013; Flum and Grohe 2006; Nieder-
meier 2006).
Fixed-parameter algorithms The idea in fixed-parameter al-
gorithmics is to accept exponential running times, which are
seemingly inevitable in solving NP-hard problems, but to
restrict them to one aspect of the problem, the parameter.
Thus, formally, an instance of a parameterized prob-
lem Π is a pair (x,k) consisting of the input x and the
parameter k. A parameterized problem Π is fixed-parame-
ter tractable (FPT) with respect to a parameter k if there
is an algorithm solving any instance of Π with size n in
f (k)· poly(n) time for some computable function f . Such
an algorithm is called a fixed-parameter algorithm. It is po-
tentially efficient for small values of k, in contrast to an
algorithm that is merely running in polynomial time for each
fixed k (thus allowing the degree of the polynomial to de-
pend on k). FPT is the complexity class of fixed-parameter
tractable parameterized problems.
A parameterized complexity view on non-preemptively scheduling interval-constrained jobs
3
Table 1.1 Overview of results on INTERVAL-CONSTRAINED SCHEDULING for various parameter combinations. The parameterized complexity
with respect to the combined parameter λ + σ remains open.
Combined
with
λ
σ
m
looseness λ
NP-hard for any λ > 1
(Cieliebak et al 2004)
slack σ
?
NP-hard for any σ ≥ 2
(Cieliebak et al 2004)
Parameter
number m of machines
W[1]-hard for parameter m for any λ > 1 (Theorem 3.1),
weakly NP-hard for m = 2 and any λ > 1 (Theorem 3.1),
pseudo-polynomial time for fixed m and λ (Theorem 4.1)
fixed-parameter tractable for parameter σ + m (Theorem 5.1)
NP-hard for m = 1 (Garey and Johnson 1979)
We refer to the sum of parameters k1 +k2 as the combined
parameter k1 and k2.
Parameterized intractability To show that a problem is pre-
sumably not fixed-parameter tractable, there is a parameter-
ized analog of NP-hardness theory. The parameterized analog
of NP is the complexity class W[1] ⊇ FPT, where it is con-
jectured that FPT (cid:54)= W[1]. A parameterized problem Π with
parameter k is called W[1]-hard if Π being fixed-parameter
tractable implies W[1] = FPT. W[1]-hardness can be shown
using a parameterized reduction from a known W[1]-hard
problem: a parameterized reduction from a parameterized
problem Π1 to a parameterized problem Π2 is an algorithm
mapping an instance I with parameter k to an instance I(cid:48) with
parameter k(cid:48) in time f (k)· poly(I) such that k(cid:48) ≤ g(k) and
I(cid:48) is a yes-instance for Π1 if and only if I is a yes-instance
for Π2, where f and g are arbitrary computable functions.
3 A strengthened hardness result
In this section, we strengthen a hardness result of Cieliebak
et al (2004), who showed that λ -LOOSE INTERVAL-CON-
STRAINED SCHEDULING is NP-hard for any λ > 1. This
section proves the following theorem:
Theorem 3.1 Let λ : N → R be such that λ (n) ≥ 1 + n−c
for some integer c ≥ 1 and all n ≥ 2.
Then λ (n)-LOOSE INTERVAL-CONSTRAINED SCHED-
ULING of n jobs on m machines is
(i) weakly NP-hard for m = 2, and
(ii) strongly W[1]-hard for parameter m.
Note that Theorem 3.1, in particular, holds for any constant
function λ (n) > 1.
We remark that Theorem 3.1 cannot be proved using the
NP-hardness reduction given by Cieliebak et al (2004), which
reduces 3-SAT instances with k clauses to INTERVAL-CON-
STRAINED SCHEDULING instances with m = 3k machines.
Since 3-SAT is trivially fixed-parameter tractable for the
parameter number k of clauses, the reduction of Cieliebak
et al (2004) cannot yield Theorem 3.1.
Instead, to prove Theorem 3.1, we give a parameterized
polynomial-time many-one reduction from BIN PACKING
with m bins and n items to λ (mn)-LOOSE INTERVAL-CON-
STRAINED SCHEDULING with m machines and mn jobs.
BIN PACKING
Input: A bin volume V ∈ N, a list a1, . . . ,an ∈ N of items,
and a number m ≤ n of bins.
Question: Is there a partition S1 (cid:93)···(cid:93) Sm = {1, . . . ,n} such
that ∑i∈Sk ai ≤ V for all 1 ≤ k ≤ m?
Since BIN PACKING is weakly NP-hard for m = 2 bins and
W[1]-hard parameterized by m even if all input numbers are
polynomial in n (Jansen et al 2013), Theorem 3.1 will follow.
Our reduction, intuitively, works as follows: for each
of the n items ai in a BIN PACKING instance with m bins
of volume V , we create a set Ji := { j1
i } of m jobs
that have to be scheduled on m mutually distinct machines.
Each machine represents one of the m bins in the BIN PACK-
ING instance. Scheduling job j1
i on a machine k corresponds
to putting item ai into bin k and will take B + ai time of
machine k, where B is some large integer chosen by the
reduction. If j1
i is not scheduled on machine k, then a job
in Ji \{ j1
i } has to be scheduled on machine k, which will
take only B time of machine k. Finally, we choose the latest
deadline of any job as nB +V . Thus, since all jobs have to
be finished by time nB +V and since there are n items, for
each machine k, the items ai for which j1
i is scheduled on
machine k must sum up to at most V in a feasible schedule.
This corresponds to satisfying the capacity constraint of V of
each bin.
i , . . . , jm
Formally, the reduction works as follows and is illustrated
in Figure 3.1.
Construction 3.2 Given a BIN PACKING instance I with
n ≥ 2 items a1, . . . ,an and m ≤ n bins, and λ : N → R such
that λ (n) ≥ 1 + n−c for some integer c ≥ 1 and all n ≥ 2,
we construct an INTERVAL-CONSTRAINED SCHEDULING
4
Ren´e van Bevern et al.
Fig. 3.1 Reduction from BIN PACKING with four items a1 = 1,a2 = a3 = 2,a4 = 3, bin volume V = 3, and m = 3 bins to 3/2-LOOSE INTERVAL-
CONSTRAINED SCHEDULING. That is, Construction 3.2 applies with c = 1, A = 8, and B = 3· 4· 8 = 96. The top diagram shows (not to scale) the
jobs created by Construction 3.2. Herein, the processing time of each job is drawn as a rectangle of corresponding length in an interval being the
job's time window. The bottom diagram shows a feasible schedule for three machines M1,M2, and M3 that corresponds to putting items a1 and a3
into the first bin, item a2 into the second bin, and a4 into the third bin.
instance with m machines and mn jobs as follows. First, let
A :=
n∑
i=1
ai
and
B := (mn)c · A ≥ 2A.
If V > A, then I is a yes-instance of BIN PACKING and we
return a trivial yes-instance of INTERVAL-CONSTRAINED
SCHEDULING.
Otherwise, we have V ≤ A and construct an instance of
INTERVAL-CONSTRAINED SCHEDULING as follows: for
i } of
each i ∈ {1, . . . ,n}, we introduce a set Ji := { j1
jobs. For each job j ∈ Ji, we choose the release time
t j := (i− 1)B,
i , . . . , jm
the processing time
p j :=
B + ai
B
if j = j1
i ,
if j (cid:54)= j1
i ,
(cid:40)
(cid:40)
and the deadline
d j :=
iB + A if i < n,
iB +V if i = n.
This concludes the construction.
(3.1)
(cid:117)(cid:116)
Remark 3.3 Construction 3.2 outputs an INTERVAL-CON-
STRAINED SCHEDULING instance with agreeable deadlines,
that is, the deadlines of the jobs have the same relative order
as their release times. Thus, in the offline scenario, all hard-
ness results of Theorem 3.1 will also hold for instances with
agreeable deadlines.
In contrast, agreeable deadlines make the problem sig-
nificantly easier in the online scenario: Chen et al (2016)
showed an online-algorithm with constant competitive ratio
for INTERVAL-CONSTRAINED SCHEDULING with agree-
able deadlines, whereas there is a lower bound of n on the
competitive ratio for general instances (Saha 2013).
In the remainder of this section, we show that Construc-
tion 3.2 is correct and satisfies all structural properties that
allow us to derive Theorem 3.1.
First, we show that Construction 3.2 indeed creates an IN-
TERVAL-CONSTRAINED SCHEDULING instance with small
looseness.
Lemma 3.4 Given a BIN PACKING instance with n≥ 2 items
and m bins, Construction 3.2 outputs an INTERVAL-CON-
STRAINED SCHEDULING instance with
(i) at most m machines and mn jobs and
(ii) looseness λ (mn).
Proof It is obvious that the output instance has at most
mn jobs and m machines and, thus, (i) holds.
0BB+a1B+A2B2B+a22B+A3B3B+a33B+A4B4B+a44B+Vj11j21j31j12j22j32j13j23j33j14j24j34createdinstancefeasiblescheduleM1M2M33B+a23B+a1+a3j11j22j132B+a14B+a1+a3=4B+a4j24j21j12j23j34j31j32j33j14 A parameterized complexity view on non-preemptively scheduling interval-constrained jobs
5
Towards (ii), observe that mn ≥ n ≥ 2, and hence, for
each i ∈ {1, . . . ,n} and each job j ∈ Ji, (3.1) yields
d j −t j
A
B
B
−c ≤ λ (mn).
≤ (iB + A)− (i− 1)B
= 1 + (mn)
= 1 +
= 1 +
B
A
B + A
p j
=
(mn)c · A
(cid:117)(cid:116)
We now show that Construction 3.2 runs in polynomial time
and that, if the input BIN PACKING instance has polynomially
bounded integers, then so has the output INTERVAL-CON-
STRAINED SCHEDULING instance.
Lemma 3.5 Let I be a BIN PACKING instance with n ≥
2 items a1, . . . ,an and let amax := max1≤i≤n ai. Construc-
tion 3.2 applied to I
(i) runs in time polynomial in I and
(ii) outputs an INTERVAL-CONSTRAINED SCHEDULING
instance whose release times and deadlines are bounded
by a polynomial in n + amax.
Proof We first show (ii), thereafter we show (i).
(ii) It is sufficient to show that the numbers A and B
in Construction 3.2 are bounded polynomially in n + amax
since all release times and deadlines are computed as sums
and products of three numbers not larger than A, B, or n.
i=1 ai ≤ n·max1≤i≤n ai, which is polynomially
Clearly, A = ∑n
bounded in n + amax. Since mn ≤ n2, also B = (mn)c · A is
polynomially bounded in n + amax.
1=1 ai is clearly computable in time
polynomial in the input length. It follows that also B = (mn)c·
(cid:117)(cid:116)
A is computable in polynomial time.
(i) The sum A = ∑n
It remains to prove that Construction 3.2 maps yes-instances
of BIN PACKING to yes-instances of INTERVAL-CONSTRAINED
SCHEDULING, and no-instances to no-instances.
Lemma 3.6 Given a BIN PACKING instance I with m bins
and the items a1, . . . ,an, Construction 3.2 outputs an IN-
TERVAL-CONSTRAINED SCHEDULING instance I(cid:48) that is a
yes-instance if and only if I is.
Proof (⇒) Assume that I is a yes-instance for BIN PACKING.
Then, there is a partition S1 (cid:93)···(cid:93) Sm = {1, . . . ,n} such that
∑i∈Sk ai ≤ V for each k ∈ {1, . . . ,m}. We construct a feasible
schedule for I(cid:48) as follows. For each i ∈ {1, . . . ,n} and k such
(cid:34)
that i ∈ Sk, we schedule j1
(i− 1)B +∑
iB +∑
a j
j∈Sk, j<i
j∈Sk, j<i
i } on a distinct machine (cid:96) ∈
and each of the m− 1 jobs Ji \{ j1
(cid:34)
(cid:33)
{1, . . . ,m}\{k} in the interval
(i− 1)B +∑
a j
j∈S(cid:96), j<i
i on machine k in the interval
iB +∑
a j
j∈S(cid:96), j<i
a j + ai
(cid:33)
.
,
,
i
ai > V.
It is easy to verify that this is indeed a feasible schedule.
(⇐) Assume that I(cid:48) is a yes-instance for INTERVAL-CON-
STRAINED SCHEDULING. Then, there is a feasible schedule
for I(cid:48). We define a partition S1 (cid:93)···(cid:93) Sm = {1, . . . ,n} for I
as follows. For each k ∈ {1, . . . ,m}, let
i is scheduled on machine k}. (3.2)
Sk := {i ∈ {1, . . . ,n} j1
Since, for each i ∈ {1, . . . ,n}, the job j1
is scheduled on
exactly one machine, this is indeed a partition. We show
that ∑i∈Sk ai ≤ V for each k ∈ {1, . . . ,m}. Assume, towards a
contradiction, that there is a k such that
∑
i∈Sk
By (3.1), for each i ∈ {1, . . . ,n}, the jobs in Ji have the same
release time, each has processing time at least B, and the
length of the time window of each job is at most B + A ≤
B + B/2 < 2B. Thus, in any feasible schedule, the execution
times of the m jobs in Ji mutually intersect. Hence, the jobs
in Ji are scheduled on m mutually distinct machines. By the
pigeonhole principle, for each i ∈ {1, . . . ,n}, exactly one
i ∈ Ji is scheduled on machine k. We finish the proof by
job j∗
showing that,
∀i ∈ {1, . . . ,n},
time iB +∑
a j.
j∈Sk, j≤i
is not finished before
job j∗
(3.3)
(3.4)
i
This claim together with (3.3) then yields that job j∗
finished before
nB +∑
a j > nB +V,
a j = nB +∑
j∈Sk
j∈Sk, j≤n
n is not
1 is B if j∗
1 = j1
which contradicts the schedule being feasible, since jobs in Jn
have deadline nB +V by (3.1). It remains to prove (3.4). We
proceed by induction.
The earliest possible execution time of j∗
1 is, by (3.1),
1 (cid:54)= j1
time 0. The processing time of j∗
1, and B + a1
otherwise. By (3.2), 1 ∈ Sk if and only if j1
1 is scheduled
on machine k, that is, if and only if j∗
1. Thus, job j∗
1 is
not finished before B + ∑ j∈Sk, j≤i a j and (3.4) holds for i = 1.
Now, assume that (3.4) holds for i−1. We prove it for i. Since
i−1 is not finished before (i− 1)B + ∑ j∈Sk, j≤i−1 a j, this is
j∗
the earliest possible execution time of j∗
i . The processing
time of j∗
i and B + ai otherwise. By (3.2),
i ∈ Sk if and only if j∗
i . Thus, job j∗
is not finished
(cid:117)(cid:116)
before iB + ∑ j∈Sk, j≤i a j and (3.4) holds.
We are now ready to finish the proof of Theorem 3.1.
(cid:54)= j1
i = j1
is B if j∗
i
i
i
Proof (of Theorem 3.1) By Lemmas 3.4 to 3.6, Construc-
tion 3.2 is a polynomial-time many-one reduction from BIN
PACKING with n ≥ 2 items and m bins to λ (mn)-LOOSE IN-
TERVAL-CONSTRAINED SCHEDULING, where λ : N → R
6
such that λ (n) ≥ 1 +n−c for some integer c ≥ 1 and all n ≥ 2.
We now show the points (i) and (ii) of Theorem 3.1.
(i) follows since BIN PACKING is weakly NP-hard for m =
2 (Jansen et al 2013) and since, by Lemma 3.4(i), Construc-
tion 3.2 outputs instances of λ (mn)-LOOSE INTERVAL-CON-
STRAINED SCHEDULING with m machines.
(ii) follows since BIN PACKING is W[1]-hard parameter-
ized by m even if the sizes of the n items are bounded by a
polynomial in n (Jansen et al 2013). In this case, Construc-
tion 3.2 generates λ (mn)-LOOSE INTERVAL-CONSTRAINED
SCHEDULING instances for which all numbers are bounded
polynomially in the number of jobs by Lemma 3.5(ii). More-
over, Construction 3.2 maps the m bins of the BIN PACKING
instance to the m machines of the output INTERVAL-CON-
(cid:117)(cid:116)
STRAINED SCHEDULING instance.
Concluding this section, it is interesting to note that Theo-
rem 3.1 also shows W[1]-hardness of λ -LOOSE INTERVAL-
CONSTRAINED SCHEDULING with respect to the height
parameter considered by Cieliebak et al (2004):
Definition 3.7 (Height) For an INTERVAL-CONSTRAINED
SCHEDULING instance and any time t ∈ N, let
St := { j ∈ J t ∈ [t j,d j)}
denote the set of jobs whose time window contains time t.
The height of an instance is
t∈N St.
h := max
Proposition 3.8 Let λ : N → R be such that λ (n) ≥ 1 +n−c
for some integer c ≥ 1 and all n ≥ 2.
Then λ (n)-LOOSE INTERVAL-CONSTRAINED SCHED-
ULING of n jobs on m machines is W[1]-hard parameterized
by the height h.
Proof Proposition 3.8 follows in the same way as Theo-
rem 3.1; one additionally has to prove that Construction 3.2
outputs INTERVAL-CONSTRAINED SCHEDULING instances
of height at most 2m. To this end, observe that, by (3.1), for
each i ∈ {1, . . . ,n}, there are m jobs released at time (i− 1)B
whose deadline is no later than iB + A < (i + 1)B since
A≤ B/2. These are all jobs created by Construction 3.2. Thus,
St contains only the m jobs released at time (cid:98)t/B(cid:99)· B and the
m jobs released at time (cid:98)t/B− 1(cid:99)· B, which are 2m jobs in
(cid:117)(cid:116)
total.
Remark 3.9 Proposition 3.8 complements findings of Cie-
liebak et al (2004), who provide a fixed-parameter tractabil-
ity result for INTERVAL-CONSTRAINED SCHEDULING pa-
rameterized by h + σ: our result shows that their algorithm
presumably cannot be improved towards a fixed-parameter
tractability result for INTERVAL-CONSTRAINED SCHEDUL-
ING parameterized by h alone.
Ren´e van Bevern et al.
4 An algorithm for bounded looseness
In the previous section, we have seen that λ -LOOSE IN-
TERVAL-CONSTRAINED SCHEDULING for any λ > 1 is
strongly W[1]-hard parameterized by m and weakly NP-hard
for m = 2. We complement this result by the following the-
orem, which yields a pseudo-polynomial-time algorithm for
each constant m and λ .
Theorem 4.1 λ -LOOSE INTERVAL-CONSTRAINED SCHED-
ULING is solvable in (cid:96)O(λ m) · n + O(nlogn) time, where (cid:96) :=
max j∈J d j −t j.
The crucial observation for the proof of Theorem 4.1 is the
following lemma. It gives a logarithmic upper bound on the
height h of yes-instances (as defined in Definition 3.7). To
prove Theorem 4.1, we will thereafter present an algorithm
that has a running time that is single-exponential in h.
Lemma 4.2 Let I be a yes-instance of λ -LOOSE INTER-
VAL-CONSTRAINED SCHEDULING with m machines and
(cid:18)
(cid:96) := max j∈J d j −t j. Then, I has height at most
2m·
(cid:19)
log (cid:96)
+ 1
.
logλ − log(λ − 1)
Proof Recall from Definition 3.7 that the height of an INTER-
VAL-CONSTRAINED SCHEDULING instance is maxt∈NSt.
We will show that, in any feasible schedule for I and at
any time t, there are at most N jobs in St that are active on
the first machine at some time t(cid:48) ≥ t, where
N ≤
log (cid:96)
(4.1)
+ 1.
logλ − log(λ − 1)
By symmetry, there are at most N jobs in St that are active
on the first machine at some time t(cid:48) ≤ t. Since there are
m machines, the total number of jobs in St at any time t, and
therefore the height, is at most 2mN.
It remains to show (4.1). To this end, fix an arbitrary
time t and an arbitrary feasible schedule for I. Then, for
any d ≥ 0, let J(t +d) ⊆ St be the set of jobs that are active on
the first machine at some time t(cid:48) ≥ t but finished by time t +d.
We show by induction on d that
(cid:40)
J(t + d) ≤
0
− logd
log(1−1/λ ) + 1
if d = 0,
if d ≥ 1.
(4.2)
If d = 0, then J(t + 0) = 0 and (4.2) holds. Now, consider
the case d ≥ 1. If no job in J(t + d) is active at time t + d − 1,
then J(t + d) = J(t + d − 1) and (4.2) holds by the induction
hypothesis. Now, assume that there is a job j ∈ J(t + d) that
is active at time t + d − 1. Then, d j ≥ t + d and, since j ∈ St,
t j ≤ t. Hence,
p j ≥ d j −t j
≥ t + d −t
=
λ
λ
d
λ .
A parameterized complexity view on non-preemptively scheduling interval-constrained jobs
7
It follows that
J(t + d) ≤ 1 +J(t + d −(cid:100)d/λ(cid:101)).
(4.3)
Thus, if d −(cid:100)d/λ(cid:101) = 0, then J(t + d) ≤ 1 +J(t) ≤ 1 and
(4.2) holds. If d −(cid:100)d/λ(cid:101) > 0, then, by the induction hypoth-
esis, the right-hand side of (4.3) is
Algorithm 4.4 We solve INTERVAL-CONSTRAINED SCHED-
ULING using dynamic programming. First, for an INTERVAL-
CONSTRAINED SCHEDULING instance, let (cid:96) := max j∈J d j−
t j, let St be as defined in Definition 3.7, and let S<
t ⊆ J be
the set of jobs j with d j ≤ t, that is, that have to be finished
by time t.
We compute a table T that we will show to have the
following semantics. For a time t ∈ N, a subset S ⊆ St of jobs
and a vector b = (b1, . . . ,bm) ∈ {−(cid:96), . . . , (cid:96)}m,
1
0
if all jobs in S∪ S<
t can be scheduled so
that machine i is idle from time t + bi
for each i ∈ {1, . . . ,m},
otherwise.
To compute T , first, set T [0, /0,b] := 1 for every vector b ∈
{−(cid:96), . . . , (cid:96)}m. Now we compute the other entries of T by
increasing t, for each t by increasing b, and for each b by S
with increasing cardinality. Herein, we distinguish two cases.
(a) If t ≥ 1 and S ⊆ St−1, then set T [t,S,b] := T [t −1,S(cid:48),b(cid:48)],
where
S(cid:48) := S∪ (St−1 ∩ S<
b(cid:48)
i := min{bi + 1, (cid:96)} for each i ∈ {1, . . . ,m}.
t ) and b(cid:48) := (b(cid:48)
1, . . . ,b(cid:48)
m) with
(b) Otherwise, set T [t,S,b] := 1 if and only if at least one of
the following two cases applies:
i) there is a machine i ∈ {1, . . . ,m} such that bi > −(cid:96)
and T [t,S,b(cid:48)] = 1, where b(cid:48) := (b(cid:48)
1, . . . ,b(cid:48)
m) with
(cid:40)
bi − 1
bi(cid:48)
b(cid:48)
i(cid:48) :=
if i(cid:48) = i,
if i(cid:48) (cid:54)= i,
(cid:40)
bi − p j
bi(cid:48)
or
ii) there is a job j ∈ S and a machine i ∈ {1, . . . ,m}
such that bi > 0, t + bi ≤ d j, t + bi − p j ≥ t j, and
T [t,S\{ j},b(cid:48)] = 1, where b(cid:48) := (b(cid:48)
m) with
1, . . . ,b(cid:48)
if i(cid:48) = i,
if i(cid:48) (cid:54)= i.
b(cid:48)
i(cid:48) :=
Note that, since j ∈ St, one has t j ≥ t − (cid:96) by defi-
i ≥ −(cid:96) is within the allowed
nition of (cid:96). Hence, b(cid:48)
range {−(cid:96), . . . , (cid:96)}.
Finally, we answer yes if and only if T [tmax,Stmax ,1m · (cid:96)] = 1,
(cid:117)(cid:116)
where tmax := max j∈J t j.
Lemma 4.5 Algorithm 4.4 correctly decides INTERVAL-CON-
STRAINED SCHEDULING.
Proof We prove the following two claims: For any time 0 ≤
t ≤ tmax, any set S ⊆ St, and any vector b = (b1, . . . ,bm) ∈
{−(cid:96), . . . , (cid:96)}m,
if T [t,S,b] = 1, then all jobs in S∪S<
t can be sched-
uled so that machine i is idle from time t + bi for
each i ∈ {1, . . . ,m},
(4.4)
+ 1
≤ 1− log(d −(cid:100)d/λ(cid:101))
log(1− 1/λ )
≤ 1− log(d(1− 1/λ ))
log(1− 1/λ )
= 1− logd + log(1− 1/λ )
= −
log(1− 1/λ )
logd
+ 1,
log(1− 1/λ )
+ 1
+ 1
T [t,S,b] =
and (4.2) holds. Finally, since (cid:96) = max1≤ j≤nd j −t j, no job
in St is active at time t + (cid:96). Hence, we can now prove (4.1)
using (4.2) by means of
N ≤ J(t + (cid:96)) ≤ −
log (cid:96)
+ 1
log(1− 1/λ )
log(cid:0) λ−1
= − log (cid:96)
λ
(cid:1) + 1
= −
=
log (cid:96)
log(λ − 1)− logλ + 1
+ 1.
logλ − log(λ − 1)
log (cid:96)
(cid:117)(cid:116)
1
≤ λ .
The following proposition gives some intuition on how the
bound behaves for various λ .
Proposition 4.3 For any λ ≥ 1 and any b ∈ (1,e], it holds
that
logb λ − logb(λ − 1)
Proof It is well-known that (1− 1/λ )λ < 1/e for any λ ≥ 1.
Hence, λ logb(1− 1/λ ) = logb(1− 1/λ )λ < logb 1/e ≤ −1,
that is, −λ logb(1− 1/λ ) ≥ 1. Thus,
−λ logb(1− 1/λ )
Finally,
−logb(1− 1/λ )
≤ λ .
≤ 1
and
1
1
1
−logb(1− 1/λ )
=
=
1
−logb( λ−1
λ )
1
−logb(λ − 1) + logb λ .
(cid:117)(cid:116)
Towards our proof of Theorem 4.1, Lemma 4.2 provides a
logarithmic upper bound on the height h of yes-instances of
INTERVAL-CONSTRAINED SCHEDULING. Our second step
towards the proof of Theorem 4.1 is the following algorithm,
which runs in time that is single-exponential in h. We first
present the algorithm and, thereafter, prove its correctness
and running time.
8
t
(4.5)
t = S(cid:48)∪S<
and
if all jobs in S ∪ S<
can be scheduled so that
machine i is idle from time t + bi for each i ∈
{1, . . . ,m}, then T [t,S,b] = 1.
From (4.4) and (4.5), the correctness of the algorithm easily
follows: observe that, in any feasible schedule, all machines
are idle from time tmax + (cid:96) and all jobs J ⊆ Stmax ∪ S<
tmax are
scheduled. Hence, there is a feasible schedule if and only if
T [tmax,Stmax ,1m · (cid:96)] = 1. It remains to prove (4.4) and (4.5).
First, we prove (4.4) by induction. For T [0, /0,b] = 1,
(4.4) holds since there are no jobs to schedule. We now
prove (4.4) for T [t,S,b] under the assumption that it is true
for all T [t(cid:48),S(cid:48),b(cid:48)] with t(cid:48) < t or t(cid:48) = t and b(cid:48) (cid:8) b.
If T [t,S,b] is set to 1 in Algorithm 4.4(a), then, for S(cid:48)
and b(cid:48) as defined in Algorithm 4.4(a), T [t − 1,S(cid:48),b(cid:48)] = 1.
By the induction hypothesis, all jobs in S(cid:48) ∪ S<
t−1 can be
scheduled so that machine i is idle from time t − 1 + b(cid:48)
i ≤
t +bi. Moreover, S∪S<
t−1 since S(cid:48) = S∪(St−1∩S<
t ).
Hence, (4.4) follows.
If T [t,S,b] is set to 1 in Algorithm 4.4(bi), then one has
T [t,S,b(cid:48)] = 1 for b(cid:48) as defined in Algorithm 4.4(bi). By the
induction hypothesis, all jobs in S∪ S<
t can be scheduled so
i(cid:48) ≤ t + bi(cid:48), and (4.4)
that machine i(cid:48) is idle from time t + b(cid:48)
follows.
If T [t,S,b] is set to 1 in Algorithm 4.4(bii), then T [t,S\
{ j},b(cid:48)] = 1 for j and b(cid:48) as defined in Algorithm 4.4(bii).
By the induction hypothesis, all jobs in (S\{ j})∪ S<
t can
be scheduled so that machine i(cid:48) is idle from time t + b(cid:48)
i(cid:48). It
remains to schedule job j on machine i in the interval [t +
b(cid:48)
i,t +bi), which is of length exactly p j by the definition of b(cid:48).
Then, machine i is idle from time t +bi and any machine i(cid:48) (cid:54)= i
is idle from time t + b(cid:48)
It remains to prove (4.5). We use induction. Claim (4.5)
clearly holds for t = 0, S = /0, and any b ∈ {−(cid:96), . . . , (cid:96)}m by
the way Algorithm 4.4 initializes T . We now show (4.5)
provided that it is true for t(cid:48) < t or t(cid:48) = t and b(cid:48) (cid:8) b.
t−1 for S(cid:48) as defined in
Algorithm 4.4(a). Moreover, since no job in S(cid:48) ∪ S<
t−1 can be
active from time t − 1 + (cid:96) by definition of (cid:96), each machine i
is idle from time t − 1 + min{bi + 1, (cid:96)} = t − 1 + b(cid:48)
i, for b(cid:48) =
m) as defined in Algorithm 4.4(a). Hence, T [t −
(b(cid:48)
1, . . . ,b(cid:48)
1,S(cid:48),b(cid:48)] = 1 by the induction hypothesis, Algorithm 4.4(a)
applies, sets T [t,S,b] := T [t − 1,S(cid:48),b(cid:48)] = 1, and (4.5) holds.
If some machine i is idle from time t + bi − 1, then, by
the induction hypothesis, T [t,S,b(cid:48)] = 1 in Algorithm 4.4(bi),
the algorithm sets T [t,S,b] := 1, and (4.5) holds.
In the remaining case, every machine i is busy at time t +
bi − 1 and K := S \ St−1 (cid:54)= /0. Thus, there is a machine i
executing a job from K. For each job j(cid:48) ∈ K, we have t j(cid:48) ≥ t.
Since machine i is idle from time t + bi and executes j(cid:48), one
has bi > 0. Let j be the last job scheduled on machine i. Then,
since machine i is busy at time t + bi − 1, we have d j ≥ t +
If S ⊆ St−1, then S∪ S<
i(cid:48) = t + bi(cid:48), and (4.4) follows.
t = S(cid:48) ∪ S<
Ren´e van Bevern et al.
bi > t and j /∈ S<
t . Hence, j ∈ St. Since machine i is idle from
time t +bi, we also have t +bi− p j ≥ t j. Now, if we remove j
from the schedule, then machine i is idle from time t +bi− p j
and each machine i(cid:48) (cid:54)= i is idle from time t + b(cid:48)
i(cid:48) = t + bi(cid:48).
Thus, by the induction hypothesis, T [t,S \ { j},b(cid:48)] = 1 in
Algorithm 4.4(bii), the algorithm sets T [t,S,b] := 1, and
(cid:117)(cid:116)
(4.5) holds.
Lemma 4.6 Algorithm 4.4 can be implemented to run in
O(2h· (2(cid:96) +1)m· (h2m +hm2)·n(cid:96) +nlogn) time, where (cid:96) :=
max j∈J d j −t j and h is the height of the input instance.
Proof Concerning the running time of Algorithm 4.4, we first
bound tmax. If tmax > n(cid:96), then there is a time t ∈ {0, . . . ,tmax}
such that St = /0 (cf. Definition 3.7). Then, we can split the
instance into one instance with the jobs S<
t and into one
instance with the jobs J \ S<
t . We answer "yes" if and only if
both of them are yes-instances. Henceforth, we assume that
tmax ≤ n(cid:96).
In a preprocessing step, we compute the sets St and St−1∩
S<
t , which can be done in O(nlogn + hn + tmax) time by
sorting the input jobs by deadlines and scanning over the
input time windows once: if no time window starts or ends at
time t, then St is simply stored as a pointer to the St(cid:48) for the
last time t(cid:48) where a time window starts or ends.
Now, the table T of Algorithm 4.4 has at most (tmax +
1)· 2h · (2(cid:96) + 1)m ≤ (n(cid:96) + 1)· 2h · (2(cid:96) + 1)m entries. A table
entry T [t,S,b] can be accessed in O(m + h) time using a
carefully initialized trie data structure (van Bevern 2014)
since S ≤ h and since b is a vector of length m.
To compute an entry T [t,S,b], we first check, for each
job j ∈ S, whether j ∈ St−1. If this is the case for each j, then
Algorithm 4.4(a) applies. We can prepare b(cid:48) in O(m) time
and S(cid:48) in O(h) time using the set St−1 ∩ S<
t computed in the
preprocessing step. Then, we access the entry T [t − 1,S(cid:48),b(cid:48)]
in O(h + m) time. Hence, (a) takes O(h + m) time.
If Algorithm 4.4(a) does not apply, then we check whether
Algorithm 4.4(bi) applies. To this end, for each i∈{1, . . . ,m},
we prepare b(cid:48) in O(m) time and access T [t,S,b(cid:48)] in O(h + m)
time. Hence, it takes O(m2 + hm) time to check (bi).
To check whether Algorithm 4.4(bii) applies, we try
each j ∈ S and each i ∈ {1, . . . ,m} and, for each, prepare b(cid:48)
in O(m) time and check T [t,S \{ j},b(cid:48)] in O(h + m) time.
(cid:117)(cid:116)
Thus (bii) can be checked in O(h2m + hm2) time.
With the logarithmic upper bound on the height h of yes-
instances of INTERVAL-CONSTRAINED SCHEDULING given
by Lemma 4.2 and using Algorithm 4.4, which, by Lemma 4.6,
runs in time that is single-exponential in h for a fixed num-
ber m of machines, we can now prove Theorem 4.1.
(cid:18)
Proof (of Theorem 4.1) We use the following algorithm. Let
h := 2m·
log (cid:96)
+ 1
.
logλ − log(λ − 1)
(cid:19)
A parameterized complexity view on non-preemptively scheduling interval-constrained jobs
If, for any time t ∈ N, we have St > h, then we are facing
a no-instance by Lemma 4.2 and immediately answer "no".
This can be checked in O(nlogn) time: one uses the interval
graph coloring problem to check whether we can schedule
the time windows of all jobs (as intervals) onto h machines.
Otherwise, we conclude that our input instance has height
at most h. We now apply Algorithm 4.4, which, by Lemma 4.6,
runs in O(2h · (2(cid:96) + 1)m · (h2m + hm2) · n(cid:96) + nlogn) time.
Since, by Proposition 4.3, h ∈ O(λ mlog (cid:96)), this running time
is (cid:96)O(λ m)h· n + O(nlogn).
(cid:117)(cid:116)
A natural question is whether Theorem 4.1 can be generalized
to λ = ∞, that is, to INTERVAL-CONSTRAINED SCHED-
ULING without looseness constraint. This question can be
easily answered negatively using a known reduction from
3-PARTITION to INTERVAL-CONSTRAINED SCHEDULING
given by Garey and Johnson (1979):
Proposition 4.7 If there is an (cid:96)O(m) · poly(n)-time algorithm
for INTERVAL-CONSTRAINED SCHEDULING, where (cid:96) :=
max j∈J d j −t j, then P = NP.
Proof Garey and Johnson (1979, Theorem 4.5) showed that
INTERVAL-CONSTRAINED SCHEDULING is NP-hard even
on m = 1 machine. In their reduction, (cid:96) ∈ poly(n). A sup-
posed (cid:96)O(m) · poly(n)-time algorithm would solve such in-
(cid:117)(cid:116)
stances in polynomial time.
Lemma 5.2 At any time t in any feasible schedule for σ -
SLACK INTERVAL-CONSTRAINED SCHEDULING, each job
j ∈ St is active at some time in the interval [t − σ ,t + σ ].
Proof If the time window of j is entirely contained in [t −
σ ,t + σ ], then, obviously, j is active at some time during the
interval [t − σ ,t + σ ].
Now, assume that the time window of j is not contained
in [t − σ ,t + σ ]. Then, since j ∈ St, its time window con-
tains t by Definition 3.7 and, therefore, one of t − σ or t + σ.
Assume, for the sake of contradiction, that there is a sched-
ule such that j is not active during [t − σ ,t + σ ]. Then j is
inactive for at least σ + 1 time units in its time window --
(cid:117)(cid:116)
a contradiction.
Now that we know that each job in St has to occupy machine
resources around time t, we can bound the size of St in the
amount of resources available around that time.
Lemma 5.3 Any yes-instance of σ -SLACK INTERVAL-CON-
STRAINED SCHEDULING has height at most (2σ + 1)m.
9
5 An algorithm for bounded slack
So far, we considered INTERVAL-CONSTRAINED SCHED-
ULING with bounded looseness λ . Cieliebak et al (2004)
additionally considered INTERVAL-CONSTRAINED SCHED-
ULING for any constant slack σ.
Recall that Cieliebak et al (2004) showed that λ -LOOSE
INTERVAL-CONSTRAINED SCHEDULING is NP-hard for
any constant λ > 1 and that Theorem 3.1 shows that hav-
ing a small number m of machines does make the problem
significantly easier.
Similarly, Cieliebak et al (2004) showed that σ -SLACK
INTERVAL-CONSTRAINED SCHEDULING is NP-hard al-
ready for σ = 2. Now we contrast this result by showing
that σ -SLACK INTERVAL-CONSTRAINED SCHEDULING is
fixed-parameter tractable for parameter m + σ. More specifi-
cally, we show the following:
Theorem 5.1 σ -SLACK INTERVAL-CONSTRAINED SCHED-
ULING is solvable in time
(σ + 1)(2σ +1)m · n· σm· logσm + nlogn
.
O
(cid:16)
(cid:17)
Similarly as in the proof of Theorem 4.1, we first give an
upper bound on the height of yes-instances of INTERVAL-
CONSTRAINED SCHEDULING as defined in Definition 3.7.
To this end, we first show that each job j ∈ St has to occupy
some of the (bounded) machine resources around time t.
Proof Fix any feasible schedule for an arbitrary yes-instance
of σ -SLACK INTERVAL-CONSTRAINED SCHEDULING and
any time t. By Lemma 5.2, each job in St is active at some
time in the interval [t−σ ,t +σ ]. This interval has length 2σ +
1. Thus, on m machines, there is at most (2σ + 1)m available
processing time in this time interval. Consequently, there can
(cid:117)(cid:116)
be at most (2σ + 1)m jobs with time intervals in St.
We finally arrive at the algorithm to prove Theorem 5.1.
Proof (of Theorem 5.1) Let h := (2σ + 1)m. In the same
way as for Theorem 4.1, in O(nlogn) time we discover that
we face a no-instance due to Lemma 5.3 or, otherwise, that
our input instance has height at most h. In the latter case,
we apply the O(n· (σ + 1)h · hlogh)-time algorithm due to
(cid:117)(cid:116)
Cieliebak et al (2004).
6 Conclusion
Despite the fact that there are comparatively few studies on
the parameterized complexity of scheduling problems, the
field of scheduling indeed offers many natural parameteri-
zations and fruitful challenges for future research. Notably,
Marx (2011) saw one reason for the lack of results on "pa-
rameterized scheduling" in the fact that most scheduling
problems remain NP-hard even for a constant number of ma-
chines (a very obvious and natural parameter indeed), hence
destroying hope for fixed-parameter tractability results with
respect to this parameter. In scheduling interval-constrained
jobs with small looseness and small slack, we also have been
confronted with this fact, facing (weak) NP-hardness even
for two machines.
10
Ren´e van Bevern et al.
The natural way out of this misery, however, is to consider
parameter combinations, for instance combining the parame-
ter number of machines with a second one. In our study, these
were combinations with looseness and with slack (see also
Table 1.1). In a more general perspective, this consideration
makes scheduling problems a prime candidate for offering
a rich set of research challenges in terms of a multivariate
complexity analysis (Fellows et al 2013; Niedermeier 2010).
Herein, for obtaining positive algorithmic results, research
has to go beyond canonical problem parameters, since basic
scheduling problems remain NP-hard even if canonical pa-
rameters are simultaneously bounded by small constants, as
demonstrated by Kononov et al (2012).1
Natural parameters to be studied in future research on IN-
TERVAL-CONSTRAINED SCHEDULING are the combination
of slack and looseness -- the open field in our Table 1.1 -- and
the maximum and minimum processing times, which were
found to play an important role in the online version of the
problem (Saha 2013).
Finally, we point out that our fixed-parameter algorithms
for INTERVAL-CONSTRAINED SCHEDULING are easy to
implement and may be practically applicable if the looseness,
slack, and number of machines is small (about three or four
each). Moreover, our algorithms are based on upper bounds
on the height of an instance in terms of its number of ma-
chines, its looseness, and slack. Obviously, this can also be
exploited to give lower bounds on the number of required ma-
chines based on the structure of the input instance, namely, on
its height, looseness, and slack. These lower bounds may be
of independent interest in exact branch and bound or approx-
imation algorithms for the machine minimization problem.
References
van Bevern R (2014) Towards optimal and expressive kernelization for
d-Hitting Set. Algorithmica 70(1):129 -- 147
van Bevern R, Chen J, Huffner F, Kratsch S, Talmon N, Woeginger GJ
(2015a) Approximability and parameterized complexity of multi-
cover by c-intervals. Information Processing Letters 115(10):744 --
749
van Bevern R, Mnich M, Niedermeier R, Weller M (2015b) Interval
scheduling and colorful independent sets. Journal of Scheduling
18(5):449 -- 469
Bodlaender HL, Fellows MR (1995) W[2]-hardness of precedence
constrained k-processor scheduling. Operations Research Letters
18(2):93 -- 97
Chen L, Megow N, Schewior K (2016) An O(logm)-competitive algo-
rithm for online machine minimization. In: Proceedings of the
27th Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA'16), SIAM, pp 155 -- 163
1 The results of Kononov et al (2012) were obtained in context of a
multivariate complexity analysis framework described by Sevastianov
(2005), which is independent of the framework of parameterized com-
plexity theory considered in our work: it allows for systematic classifica-
tion of problems as polynomial-time solvable or NP-hard given concrete
constraints on a set of instance parameters. It is plausible that this frame-
work is applicable to classify problems as FPT or W[1]-hard as well.
Chuzhoy J, Guha S, Khanna S, Naor J (2004) Machine minimization
for scheduling jobs with interval constraints. In: Proceedings of
the 45th Annual Symposium on Foundations of Computer Science
(FOCS'04), pp 81 -- 90
Cieliebak M, Erlebach T, Hennecke F, Weber B, Widmayer P (2004)
Scheduling with release times and deadlines on a minimum number
of machines. In: Exploring New Frontiers of Theoretical Informat-
ics, IFIP International Federation for Information Processing, vol
155, Springer, pp 209 -- 222
Cygan M, Fomin FV, Kowalik L, Lokshtanov D, Marx D, Pilipczuk M,
Pilipczuk M, Saurabh S (2015) Parameterized Algorithms. Springer
Downey RG, Fellows MR (2013) Fundamentals of Parameterized Com-
plexity. Springer
Fellows MR, McCartin C (2003) On the parametric complexity of
schedules to minimize tardy tasks. Theoretical Computer Science
298(2):317 -- 324
Fellows MR, Jansen BMP, Rosamond FA (2013) Towards fully mul-
tivariate algorithmics: Parameter ecology and the deconstruction
of computational complexity. European Journal of Combinatorics
34(3):541 -- 566
Flum J, Grohe M (2006) Parameterized Complexity Theory. Springer
Garey MR, Johnson DS (1979) Computers and Intractability: A Guide
to the Theory of NP-Completeness. Freeman
Halld´orsson MM, Karlsson RK (2006) Strip graphs: Recognition and
scheduling. In: Proceedings of the 32nd International Workshop on
Graph-Theoretic Concepts in Computer Science (WG'06), Springer,
LNCS, vol 4271, pp 137 -- 146
Hermelin D, Kubitza JM, Shabtay D, Talmon N, Woeginger G (2015)
Scheduling two competing agents when one agent has significantly
fewer jobs. In: Proceedings of the 10th International Symposium
on Parameterized and Exact Computation (IPEC'15), Leibniz In-
ternational Proceedings in Informatics (LIPIcs), vol 43, Schloss
Dagstuhl -- Leibniz-Zentrum fur Informatik, pp 55 -- 65
Jansen K, Kratsch S, Marx D, Schlotter I (2013) Bin packing with fixed
number of bins revisited. Journal of Computer and System Sciences
79(1):39 -- 49
Kolen AWJ, Lenstra JK, Papadimitriou CH, Spieksma FCR (2007)
Interval scheduling: A survey. Naval Research Logistics 54(5):530 --
543
Kononov A, Sevastyanov S, Sviridenko M (2012) A complete 4-
parametric complexity classification of short shop scheduling prob-
lems. Journal of Scheduling 15(4):427 -- 446
Malucelli F, Nicoloso S (2007) Shiftable intervals. Annals of Operations
Research 150(1):137 -- 157
Marx D (2011) Fixed-parameter tractable scheduling problems. In:
Packing and Scheduling Algorithms for Information and Commu-
nication Services (Dagstuhl Seminar 11091)
Mnich M, Wiese A (2015) Scheduling and fixed-parameter tractability.
Mathematical Programming 154(1-2):533 -- 562
Niedermeier R (2006) Invitation to Fixed-Parameter Algorithms. Oxford
University Press
Niedermeier R (2010) Reflections on multivariate algorithmics and
problem parameterization. In: Proceedings of the 27th Interna-
tional Symposium on Theoretical Aspects of Computer Science
(STACS'10), Schloss Dagstuhl -- Leibniz-Zentrum fur Informatik,
Leibniz International Proceedings in Informatics (LIPIcs), vol 5,
pp 17 -- 32
Saha B (2013) Renting a cloud. In: Annual Conference on Founda-
tions of Software Technology and Theoretical Computer Science
(FSTTCS) 2013, Schloss Dagstuhl -- Leibniz-Zentrum fur Infor-
matik, Leibniz International Proceedings in Informatics (LIPIcs),
vol 24, pp 437 -- 448
Sevastianov SV (2005) An introduction to multi-parameter complexity
analysis of discrete problems. European Journal of Operational
Research 165(2):387 -- 397
|
1312.1764 | 1 | 1312 | 2013-12-06T04:23:26 | Optimal Error Rates for Interactive Coding I: Adaptivity and Other Settings | [
"cs.DS",
"cs.IT",
"cs.IT"
] | We consider the task of interactive communication in the presence of adversarial errors and present tight bounds on the tolerable error-rates in a number of different settings.
Most significantly, we explore adaptive interactive communication where the communicating parties decide who should speak next based on the history of the interaction. Braverman and Rao [STOC'11] show that non-adaptively one can code for any constant error rate below 1/4 but not more. They asked whether this bound could be improved using adaptivity. We answer this open question in the affirmative (with a slightly different collection of resources): Our adaptive coding scheme tolerates any error rate below 2/7 and we show that tolerating a higher error rate is impossible. We also show that in the setting of Franklin et al. [CRYPTO'13], where parties share randomness not known to the adversary, adaptivity increases the tolerable error rate from 1/2 to 2/3. For list-decodable interactive communications, where each party outputs a constant size list of possible outcomes, the tight tolerable error rate is 1/2.
Our negative results hold even if the communication and computation are unbounded, whereas for our positive results communication and computation are polynomially bounded. Most prior work considered coding schemes with linear amount of communication, while allowing unbounded computations. We argue that studying tolerable error rates in this relaxed context helps to identify a setting's intrinsic optimal error rate. We set forward a strong working hypothesis which stipulates that for any setting the maximum tolerable error rate is independent of many computational and communication complexity measures. We believe this hypothesis to be a powerful guideline for the design of simple, natural, and efficient coding schemes and for understanding the (im)possibilities of coding for interactive communications. | cs.DS | cs | Optimal Error Rates for Interactive Coding I:
Adaptivity and Other Settings
Mohsen Ghaffari
Bernhard Haeupler
MIT
Microsoft Research
Madhu Sudan
Microsoft Research
[email protected]
[email protected]
[email protected]
Abstract
We consider the task of interactive communication in the presence of adversarial errors and
present tight bounds on the tolerable error-rates in a number of different settings.
Most significantly, we explore adaptive interactive communication where the communicating
parties decide who should speak next based on the history of the interaction.
In particular,
this allows this decision to depend on estimates of the amount of errors that have occurred so
far. Braverman and Rao [STOC'11] show that non-adaptively one can code for any constant
error rate below 1/4 but not more. They asked whether this bound could be improved using
adaptivity. We answer this open question in the affirmative (with a slightly different collection
of resources): Our adaptive coding scheme tolerates any error rate below 2/7 and we show that
tolerating a higher error rate is impossible. We also show that in the setting of Franklin et al.
[CRYPTO'13], where parties share randomness not known to the adversary, adaptivity increases
the tolerable error rate from 1/2 to 2/3. For list-decodable interactive communications, where
each party outputs a constant size list of possible outcomes, the tight tolerable error rate is 1/2.
Our negative results hold even if the communication and computation are unbounded,
whereas for our positive results communication and computation are polynomially bounded.
Most prior work considered coding schemes with linear amount of communication, while al-
lowing unbounded computations. We argue that studying tolerable error rates in this relaxed
context helps to identify a setting's intrinsic optimal error rate. We set forward a strong working
hypothesis which stipulates that for any setting the maximum tolerable error rate is independent
of many computational and communication complexity measures. We believe this hypothesis to
be a powerful guideline for the design of simple, natural, and efficient coding schemes and for
understanding the (im)possibilities of coding for interactive communications.
3
1
0
2
c
e
D
6
]
S
D
.
s
c
[
1
v
4
6
7
1
.
2
1
3
1
:
v
i
X
r
a
1
Introduction
"Interactive Coding" or "Coding for Interactive Communication" studies the task of protecting
an interaction between two parties in the presence of communication errors. This line of work
was initated by Schulman [10] who showed, surprisingly at the time, that protocols with n rounds
of communication can be protected against a (small) constant fraction of adversarial errors while
incurring only a constant overhead in the total communication complexity.
In a recent powerful result that revived this area, Braverman and Rao [5] explored the maximal
rate of errors that could be tolerated in an interactive coding setting. They showed the existence
of a protocol that handles a 1/4 − ǫ error rate and gave a matching negative result under the
assumption that the coding scheme is non-adaptive in deciding which player transmits (and which
one listens) at any point of time. They left open the questions whether the 1/4 error rate can
be improved by allowing adaptivity (see [3, Open Problem 7] and [5, Conclusion]) or by reducing
the decoding requirement to list decoding (see [3, Open Problem 9] and [5, Conclusion]), that is,
requiring each party only to give a small list of possible outcomes of which one has to be correct.
In this work we answer both questions in the affirmative (in a somewhat different regime of
computational and communication resources): We give a rate adaptive coding scheme that tolerates
any error rate below 2/7. We furthermore show matching impossibility result which strongly rules
out any coding scheme achieving an error rate of 2/7.
Moreover, we also consider the adaptive coding schemes in the setting of [8] in which both
parties share some randomness not known to the adversary. While non-adaptive coding schemes
can tolerate any error rate below 1/2 this bound increases to 2/3 using adaptivity, which we show
is also best possible.
Lastly, we initiate the study of list decodable interactive communication. We show that allowing
both parties to output a constant size list of possible outcomes allows non-adaptive coding schemes
that are robust against any error rate below 1/2, which again is best possible for in both the
adaptive and non-adaptive setting.
All our coding schemes are deterministic and work with communication and computation being
polynomially bounded in the length of the original protocol. We note that most previous works
considered the more restrictive setting of linear amount of communication (often at the cost of
exponential time computations).
Interestingly, our matching negative results hold even if the
communication and computation are unbounded. We show that this sharp threshold behavior
extends to many other computational and communication complexity measures and is common to
all settings of interactive communication studied in the literature. In fact, an important conceptual
contribution of this paper is the formulation of a strong working hypothesis that stipulates that
maximum tolerable error rates are invariable with changes in complexity and efficiency restrictions
on the coding scheme. Throughout this paper this hypothesis lead us to consider the simplest
setting for positive results and then expanding on the insights derived to get the more general
positive results. We believe that in this way, the working hypothesis yields a powerful guideline for
the design of simple and natural coding schemes as also the search for negative results. This has
been already partially substantiated by subsequent results (see [9] and Appendix B).
Organization In what follows, we briefly introduce the interactive communication model more
formally in Section 2. We also introduce the model for adaptive interaction there. Then, in
Section 3, we explain our results as well as the underlying high-level ideas and techniques.
In
1
Section 4 we describe the simple Exchange problem and give an adaptive protocol tolerating 2/7-
fraction error in Section 4.1. (Combined with Section 3 this section introduces all the principal
ideas of the paper. Rest of the paper may be considered supplemental material.) In the remainder
of Section 4, we prove that error-rate of 2/7 is the best achievable for the Exchange problem and
thus also for the general case of interactive communication. In Section 5, we give interactive coding
schemes over large alphabets tolerating 2/7 error rate for general interactions. In Section 6 we then
convert these to coding schemes over constant size alphabets. Finally, in Section 7, we give protocols
tolerating an 2/3 error rate in the presence of shared randomness. The appendix contains some
technical proofs, as well as some simple impossibility results showing tightness of our protocols.
2 New and Old Settings for Interactive Coding
In this section, we define the classical interactive coding setup as well as all new settings considered
in this work, namely, list decoding, the shared randomness setting, and adaptive protocols.
We start with some standard terminology: An n-round interactive protocol Π between two
players Alice and Bob is given by two functions ΠA and ΠB. For each round of communication,
these functions map (possibly probabilistically) the history of communication and the player's
private input to a decision on whether to listen or transmit, and in the latter case also to a symbol
of the communication alphabet. All protocols studied prior to this work are non-adaptive1 in that
the decision of a player to listen or transmit deterministically depends only on the round number,
ensuring that exactly one party transmits in each round. In this case, the channel delivers the
chosen symbol of the transmitting party to the listening party, unless the adversary interferes and
alters the symbol arbitrarily. In the adversarial channel model with error rate ρ, the number of such
errors is at most ρn. The outcome of a protocol is defined to be the transcript of the interaction.
A protocol Π′ is said to robustly simulate a protocol Π for an error rate ρ if the following holds:
Given any inputs to Π, both parties can uniquely decode the transcript of an error free execution
of Π on these inputs from the transcript of any execution of Π′ in which at most a ρ fraction of
the transmissions were corrupted. This definition extends easily to list decoding by allowing both
parties to produce a small (constant size) list of transcripts that is required to include the correct
decoding, i.e., the transcript of Π. We note that the simulation Π′ typically uses a larger alphabet
and a larger number of rounds. While our upper bounds are all deterministic, we strengthen the
scope of our lower bounds by also considering randomized protocols in which both parties have
access to independent private randomness. We also consider the setting of [8] in which both parties
have access to shared randomness.
In both cases we assume that the adversary does not know
the randomness and we say a randomized protocol robustly simulates a protocol Π with failure
probability p if, for any input and any adversary, the probability that both parties correctly (list)
decode is at least 1 − p.
We now present the notion of an adaptive protocol. It turns out that defining a formal model
for adaptivity leads to several subtle issues. We define the model first and discuss these issues later.
In an adaptive protocol, the communicating players are allowed to base their decision on whether
to transmit or listen (probabilistically) on the communication history. In particular, this allows
players to base their decision on estimates of the amount of errors that have happened so far (see
Theorem 3.2 for why this kind of adaptivity is a natural and useful approach). This can lead to
rounds in which both parties transmit or listen simultaneously. In the first case no symbols are
1Braverman and Rao [5] referred to protocols with this property as robust.
2
delivered while in the latter case the symbols received by the two listening parties are chosen by
the adversary, without it being counted as an error.
It was shown in [5] that protocols which under no
Discussion on the adaptivity model.
circumstances have both parties transmit or listen simultaneously are necessarily non-adaptive.
Any model for adaptivity must therefore allow parties to simultaneously transmit or listen and
specify what happens in either case. Doing this and also deciding on how to measure the amount
of communication and the number of errors leads to several subtle issues.
While it seems pessimistic to assume that the symbols received by two simultaneously listening
parties are determined by the adversary this is a crucial assumption. If, e.g., a listening party could
find out without doubt if the other party transmitted or listened by receiving silence in the latter
case then uncorrupted communication could be arranged by simply using the listen/transmit state
as an incorruptible one-bit communication channel. More subtle points arise when considering how
to define the quantity of communication on which the adversaries budget of corruptions is based.
The number of transmissions performed by the communicating parties, for example, seems like a
good choice. This however would make the adversaries budget a variable (possibly probabilistic)
quantity that, even worse, non-trivially depends on when and how this budgets is spent. It would
furthermore allow parties to time-code, that is, encode a large number (even an encoding of all
answers to all possible queries) in the time between two transmissions. While time-coding strategies
do not seem to lead to very efficient algorithms they would prevent strong lower bounds which show
that even over an unbounded number of rounds no meaningful communication is possible (see, e.g.,
Theorem 3.2 which proves exactly this for an error rate of 2/7).
Our model avoids all these complications. For non-adaptive protocols that perfectly coordinate
a designated sender and receiver in each round our model matches the standard setting. For the
necessary case that adaptive parties fail to coordinate our model prevents any signaling or time-
sharing capabilities and in fact precludes any information exchange. This matches the intuition
that in a conversation no advantage can be derived from both parties speaking or listening at the
same time. It also guarantees that the product between the number of rounds and the bit-size of
the communication alphabet is a clean and tight information theoretic upper bound on the total
amount of information or entropy that can be exchanged (in either direction) between the two
parties. This makes the number of rounds the perfect quantity to base the adversaries budget
on. All this makes our model, in hindsight, the arguably cleanest and most natural extension of
the standard setting to adaptive protocols (see also Appendix A for a natural interpretation as a
wireless channel model with bounded unpredictable noise). The strongest argument for our model
however is the fact that it allows to prove both strong and natural positive and negative results,
implying that our model does not overly restrict or empower the protocols or the adversary.
3 Overview
In this section we state our results and explain the high level ideas and insights behind them.
3.1 Adaptivity
A major contribution of this paper is to show that adaptive protocols can tolerate more than the
1/4 error rate of the non-adaptive setting:
3
Theorem 3.1. Suppose Π is an n-round protocol over a constant bit-size alphabet. For any ǫ > 0,
there is a deterministic computationally efficient protocol Π′ that robustly simulates Π for an error
rate of 2/7 − ǫ using O(n2) rounds and an O(1)-bit size alphabet.
The proof is presented in Section 6. Furthermore, in Section 4.2, we show a matching impossi-
bility result which even applies to the arguably simplest interactive protocol, namely, the Exchange
Protocol. In the Exchange Protocol each party simply gets an input -- simply a one bit input in our
impossibility results -- and sends this input to the other party.
Theorem 3.2. There is no (deterministic or randomized) protocol that robustly simulates the
Exchange Protocol for an error rate of 2/7 with an o(1) failure probability even when allowing com-
putationally unbounded protocols that use an arbitrarily large number of rounds and an unbounded
alphabet.
Why Adaptivity is Natural and Helpful Next, we explain why it should not be surprising
that adaptivity leads to a higher robustness. We also give some insights for why the 2/7 error rate
is the natural tolerable error for adaptive protocols.
It is helpful to first understand why the 1/4 error rate was thought of as a natural barrier. The
intuitive argument, presented in [5], for why one should not be able to cope with an error rate of 1/4
is as follows: During any N round interaction one of the parties, w.l.o.g. Alice, is the designated
sender for at most half of the rounds. With an error rate of 1/4 the adversary can corrupt half of
the symbols Alice sends out. This makes it impossible for Alice to (reliably) deliver even a single
input bit x because the adversary can always make the first half of her transmissions consistent
with x = 0 and the second half with x = 1 without Bob being able to know which of the two is real
and which one is corrupted.
While this intuition is quite convincing at the first glance, it silently assumes that it is a
priori clear which of the two parties transmits less often. This in turn essentially only holds for
non-adaptive protocols for which the above argument can also be made into a formal negative
result [5, Claim 10]. On the other hand, we show that assuming this a priori knowledge is not just
a minor technical assumption but indeed a highly nontrivial restriction which is violated in many
natural settings of interaction. For example, imagine a telephone conversation on a connection
that is bad/noisy in one direction. One person, say Alice, clearly understands Bob while whatever
Alice says contains so much noise that Bob has a hard time understanding it. In a non-adaptive
conversation, Bob would continue to talk half of the time (even though he has nothing to say given
the lack of understandable responses from Alice) while Alice continues to talk little enough that she
can be completely out-noised. This is of course not how it would go in real life. There, Bob would
listen more in trying to understand Alice and by doing this give Alice the opportunity to talk more.
Of course, as soon as this happens, the adversary cannot completely out-noise Alice anymore and
the conversation will be able to progress. In fact, similar dynamic rate adaptation mechanisms that
adapt the bitrate of a senders to channel conditions and the communication needs of other senders
are common in many systems, one prominent example being IEEE 802.11 wireless networks.
Even if one is convinced that adaptive algorithms should be able to beat the 1/4 error rate, it is
less clear at this point what the maximum tolerable error rate should be. In particular, 2/7 seems
like a quite peculiar bound. Next, without going into details of the proofs, we want to give at least
some insight why 2/7 is arguably the right and natural error rate for the adaptive setting.
We first give an intuitive argument why even adaptive protocols cannot deal with an error rate
of 1/3. For this, the adversary runs the same strategy as above which concentrates all errors on
4
one party only. In particular, given a 3N rounds conversation and a budget of N corruptions, the
adversary picks one party , say Alice, and makes her first N transmissions sound like as if her input
is x = 1. The next N transmissions are made to sound like Alice has input x = 0. During the first
N responses, regardless of whether x = 1 (resulting in Alice talking herself) or x = 0 (resulting
in the adversary imitating the same transmissions), the whole conversation will sound legitimate.
This prevents any rate adaptation, in this case on Bob's side, to kick in before 2N rounds of back
and forth have passed. Only then it becomes apparent to the receiver of the corruptions, in this
case Bob, that the adversary is trying to fool him. Knowing that the adversary will only try to
fool one party, Bob can then stop talking and listen to Alice for the rest of the conversation. Still,
even if Bob listens exclusively from this point on, there are only N rounds left which is just enough
for all of them to be corrupted. Having received N transmission from Alice claiming x = 1 and
equally many claiming x = 0, Bob is again left puzzled. This essentially proves the impossibility
of tolerating an error rate of 1/3. But even this 1/3 error rate is not achievable. To explain why
even a lower fraction of errors, namely 2/7, leads to a negative result, we remark that the radical
immediate back-off we just assumed for Bob is not possible. The reason is that if both parties are
so sensitive and radical in their adjustment, the adversary can fool both parties simultaneously by
simply corrupting a few transmissions of both parties after round 2N . This would lead to both
parties assuming that the transmissions of the other side are being corrupted. The result would be
both parties being silent simultaneously which wastes valuable communication rounds. Choosing
the optimal tradeoff for how swift and strong protocols are able to adaptively react without falling
into this trap naturally leads to an error rate between 1/4 and 1/3, and what rate in this range
could be more natural than the mediant 2/7.
Other Settings We also give results on other settings that have been suggested in the literature,
in particular, list decoding and the shared randomness setting of [8]. We briefly describe these
results next.
Franklin et al. [8] showed that if both parties share some random string not known to the
adversary, then non-adaptive protocols can boost the tolerable error rate from 1/4 to 1/2. We
show that also in this setting adaptivity helps to increase the tolerable error rate. In particular, in
Section 7, we prove that an error rate of 2/3 − ǫ is achievable and best possible2:
Theorem 3.3. In the shared randomness setting of [8], there exists a efficient robust coding scheme
for an error rate of 2/3 − ǫ while no such scheme exists for an error rate of 2/3. That is, the
equivalents of Theorem 3.1 and Theorem 3.2 hold for an error rate of 2/3. The number of rounds
of the robust coding scheme can furthermore be reduced to O(n) if one allows exponential time
computations.
We also give the first results for list decodable coding schemes (see Section 2 for their definition).
The notion of list decodability has been a somewhat recent but already widely successful addition
to the study of error correcting codes. It is known that for error correcting codes such a relaxation
leads to being able to efficiently [11] tolerate any constant error rate below 1, which is a factor
of two higher than the 1/2 − ǫ error rate achievable with unique decoding. It has been an open
question whether list decoding can also lead to higher tolerable error rates in interactive coding
2It is interesting to note that similarly to the 2/7 bound (see Theorem 3.2), 2/3 is the mediant between 1/2 and 1,
that is, the mediant between the error rate for non-adaptive protocols and the hypothetical error rate of immediately
reacting/adapting protocols.
5
schemes (see [3, Open Problem 9] and [5, Conclusion]). We show that this is indeed the case. In
particular, for the non-adaptive setting the full factor of two improvement can also be realized in
the interactive setting:
Theorem 3.4. Suppose Π is an n-round protocol over a constant bit-size alphabet. For any ǫ > 0
there is a O(1)-list decodable, non-adaptive, deterministic, and computationally efficient protocol
Π′ that robustly simulates Π for an error rate of 1/2 − ǫ using O(n2) rounds and an O(1)-bit size
alphabet.
The proof of this theorem is presented in Section 6. We also show that the 1/2 − ǫ error rate is
best possible even for adaptive coding schemes. That is, no adaptive or non-adaptive coding scheme
can achieve an error rate of 1/2. We prove these impossibility results formally in Appendix C.
Taken together, our results provide tight negative and matching positive results for any of the
eight interactive coding settings given by the three Boolean attributes, {unique decoding / list de-
coding}, {adaptive / non-adaptive}, and {without shared randomness / with shared randomness}
(at least when allowing a linear size alphabet or quadratic number of rounds in the simulation).
Table 1 shows the maximum tolerable error rate for each of these settings:
Non-adaptive
Adaptive
unique dec. (UD) UD & shared rand.
1/2 ( [8] )
2/3
1/4 ( [5] )
2/7
list dec. (LD) LD & shared rand.
1/2
1/2
1/2
2/3
Table 1: Unless marked with a citation all results in this table are new and presented in this
paper. Matching positive and negative results for each setting show that the error rates are tight.
Even more, the error rates are invariable of assumptions on the communication and computational
complexity of the coding scheme and the adversary (see 3.5, Corollary 3.6, and Appendix B).
3.2
Invariability Hypothesis: A Path to Natural Interactive Coding Schemes
In this section, we take a step back and propose a general way to understand the tolerable error
rates specific to each setting and to design interactive coding schemes achieving them. We first
formulate a strong working hypothesis which postulates that tolerable error rates are invariable
regardless of what communication and computational resources are given to the protocol and to
the adversary. We then use this hypothesis to determine the tight tolerable error rate for any
setting by looking at the simplest setup. Finally, we show how clean insights coming from these
simpler setups can lead to designs for intuitive, natural, and easily analyzable interactive coding
schemes for the more general setup.
Invariability Hypothesis
In this section we formulate our invariability hypothesis.
Surveying the literature for what error rates could be tolerated by different interactive cod-
ing schemes, the maximum tolerable error rate appears to vary widely depending on the setting
and more importantly, depending on what kind of efficiency one strives for. For example, even
for the standard setting -- that is, for non-adaptive unique decoding coding schemes using a large
alphabet -- the following error rates apply:
for unbounded (or exponential time) computations,
Schulman [10] tolerates a 1/240 error rate; Braverman and Rao [5] improved this to 1/4; for sub-
exponential time computations, Braverman [4] gave a scheme working for any error rate below 1/40;
6
for randomized polynomial time algorithms, Brakerski and Kalai [1] got an error rate of 1/16; for
randomized linear time computations, Brakerski and Naor [2] obtained an unspecified constant
error rate smaller than 1/32; lastly, assuming polynomially bounded protocols and adversaries and
using a super-linear number of rounds, Chung et al. [6] gave coding schemes tolerating an error
rate of 1/6 (with additional desirable properties).
We believe that this variety is an artifact of the current state of knowledge rather then being
the essential truth.
In fact, it appears that any setting comes with exactly one tolerable error
rate which furthermore exhibits a strong threshold behavior: For any setting, there seems to be
one error rate for which communication is impossible regardless of the resources available, while for
error rates only minimally below it simple and efficient coding schemes exist. In short, the tolerable
error rate for a setting seems robust and completely independent of any communication resource
or computational complexity restrictions made to the protocols or to the adversary.
Taking this observation as a serious working hypothesis was a driving force in obtaining, under-
standing, and structuring the results obtained in this work. As we will show, it helped to identify
the simplest setup for determining the tolerable error rate of a setting, served as a good pointer to
open questions, and helped in the design of new, simple, and natural coding schemes. We believe
that these insights and schemes will be helpful in future research to obtain the optimal, and effi-
cient coding schemes postulated to exist. In fact, we already have a number of subsequent works
confirming this (e.g., the results of [1,4,6] mentioned above can all be extended to have the optimal
1/4 error rate; see also Appendix B). All in all, we believe that identifying and formulating this
hypothesis is an important conceptual contribution of this work:
Hypothesis 3.5 (Invariability Hypothesis). Given any of the eight settings for interactive com-
munication (discussed above) the maximum tolerable error rates is invariable regardless:
1. whether the protocol to be simulated is an arbitrary n-round protocol or the much simpler
(n-bit) exchange protocol, and
2. whether only O(1)-bit size alphabets are allowed or alphabets of arbitrary size, and
3. whether the simulation has to run in O(n) rounds or is allowed to use an arbitrary number
of rounds, and
4. whether the parties are restricted to polynomial time computations or are computationally
unbounded, and
5. whether the coding schemes have to be deterministic or are allowed to use private randomness
(even when only requiring an o(1) failure probability), and
6. whether the adversary is computationally unbounded or is polynomially bounded in its com-
putations (allowing simulation access to the coding scheme if the coding scheme is not com-
putationally bounded)
We note that our negative results are already as strong as stipulated by the hypothesis, for all
eight settings. The next corollary furthermore summarizes how far these negative results combined
with the positive results presented in this work (see Table 1) already imply and prove two weaker
versions of the hypothesis:
7
Corollary 3.6. The Invariability Hypothesis holds if one weakens point 3. to "3'. whether only
O(n)-bit size alphabets are allowed or alphabets of arbitrary size". The Invariability Hypothesis
also holds if one weakens point 4. to "4'. whether the simulation has to run in O(n2) rounds or is
allowed to use an arbitrary number of rounds".
We also refer the reader to [9] and Appendix B for further (subsequent) results supporting the
hypothesis, such as, a proof that the hypothesis holds if point 4. is replaced by "4'. whether the
simulation has to run in O(n(log∗ n)O(log∗ n)) rounds3 or is allowed to use an arbitrary number of
rounds".
Determining and Understanding Tolerable Error Rates in the Simplest Setting Next,
we explain how we use the invariability hypothesis in finding the optimal tolerable error rates.
Suppose that one assumes, as a working hypothesis, the invariability of tolerable error rates to
hold regardless of the computational setup and even the structure of the protocol to be simulated.
Under this premise, the easiest way to approach determining the best error rate is in trying to
design robust simulations for the simplest possible two-way protocol, the Exchange Protocol. This
protocol simply gives each party n bits as an input and has both parties learn the other party's
input bits by exchanging them (see also Section 4). Studying this setup is considerably simpler.
For instance, for non-adaptive protocols, it is easy to see that both parties sending their input
in an error correcting code (or for n = 1 simply repeating their input bit) leads to the optimal
robust protocol which tolerates any error rate below 1/4 but not more. The same coding scheme
with applying any ECC list decoder in the end also gives the tight 1/2 bound for list decoding.
For adaptive protocols (both with and without shared randomness), finding the optimal robust
1-bit exchange protocol was less trivial but clearly still better than trying to design highly efficient
coding schemes for general protocols right away. Interestingly, looking at simpler setup actually
crystallized out well what can and cannot be done with adaptivity, and why. These insights, on the
one hand, lead to the strong lower bounds for the exchange problem but, on the other hand, were
also translated in a crucial manner to the same tradeoffs for robustly simulating general n-round
protocols.
Natural Interactive Coding Schemes The invariability working hypothesis was also helpful in
finding and formalizing simple and natural designs for obtaining robust coding schemes for general
protocols.
Before describing these natural coding schemes we first discuss the element of "surprise/magic"
in prior works on interactive coding. The existence of an interactive coding scheme that tolerates
a constant error rate is a fairly surprising outcome of the prior works, and remains so even in
hindsight. One reason for this is that the simplest way of adding redundancy to a conversation,
namely encoding each message via an error correcting code, fails dramatically because the adversary
can use its budget non-uniformly and corrupt the first message(s) completely. This derails the
interaction completely and makes all further exchanges useless even if no corruptions happens from
there on. While prior works, such as [10] or [5], manage to overcome this major challenge, their
solution remains a technically intriguing works, both in terms of the ingredients they involve (tree
3Here log∗ n stands for the iterated logarithm which is defined to be the number of times the natural logarithm
function needs to be applied to n to obtain a number less than or equal to 1. We note that this round blowup is
smaller than any constant times iterated logarithm applied to n, that is, 2O(log∗ n · log log∗ n) = o(
8
constant times
{
}
z
log log . . . log n).
codes, whose existence is again a surprising event) and the recipe for converting the ingredients
into a solution to the interactive coding problem. As a result it would appear that the challenge of
dealing with errors in interactive communication is an inherently complex task.
In contrast to this, we aim to give an intuitive and natural strategy which lends itself nicely
to a simple explanation for the possibility of robust interactive coding schemes and even for why
their tolerable error rates are as they are. This strategy simply asserts that if there is no hope to
fully trust messages exchanged before one should find ways to put any response into the assumed
common context by (efficiently) referring back to what one said before. Putting this idea into a
high-level semi-algorithmic description gives the following natural outline for a robust conversation:
Algorithm 1 Natural Strategy for a Robust Conversation (Alice's Side)
B ← What you heard Bob say last (or so far)
Listen
E ′
EA ← What you said last (or so far)
if EA and E ′
1: Assume nothing about the context of the conversation
2: loop
3:
4:
5:
6:
7:
8:
9:
10:
11: Assume / Output the conversation outcome(s) that seem most likely
B makes sense together then
Determine the most relevant response r
Send the response r but also include an (efficient) summary of what you said so far (EA)
else
Repeat what you said last (EA)
At first glance the algorithm may appear vague. In particular notions like "making sense", and
"most relevant response", seem ambiguous and subject to interpretation. It turns out that this is
not the case. In fact, formalizing this outline into a concrete coding scheme turns out to be straight
forward. This is true especially if one accepts the invariability working hypothesis and allows oneself
to not be overly concerned with having to immediately get a highly efficient implementation. In
particular, this permits to use the simplest (inefficiently) summary, namely referring back word by
word to everything said before. This straight-forward formalization leads to Algorithm 2. Indeed,
a reader that compares the two algorithms side-by-side will find that Algorithm 2 is essentially a
line-by-line formalization of Algorithm 1.
In addition to being arguably natural, Algorithm 2 is also easy to analyze. Simple counting
arguments show that the conversation outcome output by most parties is correct if the adversary
interferes at most a 1/4 − ǫ fraction of the time, proving the tight tolerable error rate for the
robust (while somewhat still inefficient) simulation of general n-round protocols. Maybe even more
surprisingly, the exact same simple counting argument also directly shows our list decoding result,
namely, that even with an error rate of 1/2 − ǫ the correct conversation outcome is among the 1/ǫ
most likely choices for both parties. Lastly, it is easy to enhance both Algorithm 1 and similarly
Algorithm 2 to be adaptive. For this one simply adds the following three, almost obvious, rules of
an adaptive conversation:
Rules 3.7 (Rules for a Successful Adaptive Conversation).
Be fair and take turns talking and listening, unless:
1. you are sure that your conversation partner already understood you fully and correctly, in
which case you should stop talking and instead listen more to also understand him; or reversely
9
2. you are sure that you already understood your conversation partner fully and correctly, in
which case you should stop listening and instead talk more to also inform him.
Our algorithm Algorithm 3 exactly adds the formal equivalent of these rules to Algorithm 3.
A relatively simple proof that draws on the insights obtained from analyzing the optimal robust
exchange protocol then shows that this simple and natural coding scheme indeed achieves the
optimal 2/7 − ǫ tolerable error rate for adaptive unique-decoding. This means that Algorithm
3 is one intuitive and natural algorithm that simultaneously achieves the 1/4 error rate (if the
adaptivity rules are ignored), the 2/7 − ǫ error rate for adaptive protocols and the 1/2 − ǫ error
rate with optimal list size when list decoding is allowed. Of course, so far, this result comes
with the drawback of using a large (O(n)-bits) alphabet. Nonetheless, this result together with
the invariability hypothesis hold open the promise of such a "perfect" algorithm that works even
without the drastic communication overhead4.
4 Results for the Exchange Problem
In this section we study the Exchange Problem, which can be viewed as the simplest instance of a
two-way (i.e., interactive) communication problem. In the Exchange Problem, each party is given
a bit-string of n bits, that is, iA, iB ∈ {0, 1}n, and each party wants to know the bit-string of the
other party.
Recall that the 1/4 impossibility bound on tolerable error-rate for non-adaptive interactive
protocols presented by Braverman and Rao [5] is this simple setting. In Section 4.1, we show that
adding rate adaptivity to the exchange algorithms helps one break this 1/4 impossiblity bound and
tolerate an error-rate of 2/7 − ǫ, and in fact, this is done with a minimal amount of adaptivity-
based decisions regarding whether a party should transmit or listen in each round. We show in
Section 4.2 that the error-rate of 2/7 is not tolerable even for the exchange problem, even if one
is given infinite number of rounds, alphabet size, and computational power. Furthermore, the
intuition used to achieve the 2/7 − ǫ possibility result also extends to the more general simulation
problem, discussed in Section 5.
4.1 An Algorithm for the Exchange Problem under Error-Rate 2/7 − ǫ
Note that a simple solution based on error correcting codes suffices for solving exchange problem
under error-rate 1
4 − ǫ: parties use a code with relative distance 1 − ǫ. In the first half of the time,
Alice sends its encoded message and in the second half of the time, Bob sends its encoded message.
At the end, each party decodes simply by picking the codeword closest to the received string. As
discussed before, the error rate 1
4 − ǫ of this approach is the best possible if no rate adaptation
is used. In the following, we explain that a simple rate adaptation technique boosts the tolerable
error-rate to 2
7 − ǫ, which we later prove to be optimal.
Theorem 4.1. In the private randomness model with rate adaptation, there is an algorithm for the
n-bit Exchange Problem that tolerates adversarial error rate of 2/7 − ǫ ≈ 0.2857 − ǫ for any ǫ > 0.
Proof. The algorithm runs in N = 7n/ǫ rounds, which means that the budget of adversary is
(2/7 − ǫ)7n/ǫ = 2n/ǫ − 7. Throughout the algorithm, we use an error-correction code C : {0, 1}n →
4Subsequent works of the authors have already moved towards this constant size alphabet protocol by reducing
the alphabet size from O(n) bits to merely O(logǫ n) bits (see Appendix B).
10
{1, . . . , q}
concatenating κ copies of C.
ǫ that has distance n
n
ǫ − 1. Also, for simplicity, we use Cκ to denote the code formed by
The first 6N/7 rounds of the algorithm do not use any rate adaptation: Simply, Alice sends
C3(iA) in the first 3N/7 rounds and Bob sends C3(iB) in the second 3N/7 rounds. At the end of
this part, each party "estimates" the errors invested on the transmissions of the other party by
simply reading the hamming distance of the received string to the closest codeword of code C3.
If this estimate is less than N/7 = n/ǫ, the party -- say Alice -- can safely assume that the closest
codeword is the correct codeword. This is because the adversary's total budget is 2n/ǫ − 7 and
the distance between two codewords of C3(iB) is at least 3n/ǫ − 3. In this case, in the remaining
N/7 rounds of the algorithm, Alice will be sending C(iA) and never listening. On the other hand,
if Alice reads an estimated error greater than N/7 = n/ǫ, then in the remaining N/7 rounds, she
will be always listening. The algorithm for Bob is similar.
Note that because of the limit on the budget of the adversary, at most only one of the parties
will be listening in the last N/7 rounds. Suppose that there is exactly one listening party and it
is Alice. In this case, throughout the whole algorithm, Alice has listened a total of 4N/7 = 4n/ǫ
rounds where Bob has been sending C4(iB). Since the adversaries budget is less than 2n/ǫ − 7, and
because C4 has distance 4n
ǫ − 4, Alice can also decode correctly by just picking the codeword of C4
with the smallest hamming distance to the received string.
4.2
Impossibility of Tolerating Error-Rate 2/7 in the Exchange Problem
In this section, we turn our attention to impossibility results and particularly prove that the error-
rate 2/7 is not tolerable. See the formal statement in Theorem 3.2.
Note that Braverman and Rao [5] showed that it is not possible to tolerate error-rate of 1/4 with
non-adaptive algorithms. For completeness, we present a (slightly more formal) proof in the style
of distributed indistinguishably arguments, in Appendix D, which also covers random algorithms.
We first explain a simple (but informal) argument which shows that even with adaptivity, error-
rate 1/3 is not tolerable. A formal version of this proof is deferred to Appendix D. The informal
version explained next serves as a warm up for the more complex argument used for proving the
2/7 impossibility, presented formally in Theorem 4.3.
Lemma 4.2. There is no (deterministic or randomized) adaptive protocol that robustly simulates
the Exchange Protocol for an error rate of 1/3 with an o(1) failure probability even when allow-
ing computationally unbounded protocols that use an arbitrarily large number of rounds and an
unbounded alphabet.
Informal Proof. To simplify the discussion, here we only explain the reasoning about deterministic
algorithms and we also ignore the rounds in which both parties listen. Note that by the definition
of the model, in those all-listening rounds, the adversary can deliver arbitrary messages to each of
the parties at no cost. A complete proof is presented in Appendix D.
For the sake of contradiction, suppose that there is an algorithm that solves the exchange
problem under adversarial error-rate 1/3, in N rounds. We work simply with 1-bit inputs. Let
SX,Y denote the setting where Alice receives input X and Bob gets input Y . The idea is to
make either settings S0,0 and S0,1 look indistinguishable to Alice or settings S0,0 and S1,0 look
indistinguishable to Bob.
11
Alice Bob
Inputs:
0
0
Alice Bob
0
1
2R/3
R/3
Copy Bob's transmissions in
rounds when Alice is listening
alone (at most R/3).
Copy Bob's transmissions
in rounds when Alice is
listening alone.
Figure 1: The adversary's strategy for the 1/3-impossibility proof
Consider setting S0,0 and suppose that for the first 2N/3 rounds, the adversary does not inter-
fere. Without loss of generality, we can assume that in this setting, Alice listens (alone) in less than
N/3 of these 2N/3 rounds. We next explain adversary's strategy for the case that this assumption
holds. An illustration of the adversary's strategy is presented in Figure 1.
First, we explain the adversary's strategy for setting S0,1: Adversary creates a dummy person-
ality gBob0 and simulates it with Alice in setting S0,0 where adversary does not interfere. In the first
2N/3 rounds of setting S0,1, whenever Alice listens (alone), the adversary delivers transmission of
gBob0 to Alice. As a shorthand for this, we say Alice is connected to gBob0. Since Alice listens less
than N/3 of the time, the adversary will have enough budget to completely fake Bob as gBob0 (from
Alice's viewpoint). Thus, the two settings look identical to Alice for the first 2N/3 rounds. During
the last N/3 rounds of the execution in setting S0,1, the adversary lets Alice and the real Bob talk
without no interference.
Now, we explain the adversary's strategy for setting S0,0: The adversary generates another
dummy personality gBob1 by simulating Bob in setting S0,1 where alone-listening rounds of Alice
in the first 2N/3 rounds are connected to gBob0. In setting S0,0, the adversary lets Alice and Bob
adversary connects her to gBob1.
talk freely during the first 2N/3 rounds but for the last N/3 rounds, whenever Alice listens, the
To conclude, we know that in each of the settings S0,1 and S1,0, at most N/3 rounds get
corrupted by the adversary. Furthermore, the two settings look identical to Alice which means that
she can not know Bob's input. This contradiction completes the proof.
Theorem 4.3. [A rephrasing of Theorem 3.2] There is no (deterministic or randomized) adaptive
protocol that robustly simulates the Exchange Protocol for an error rate of 2/7 with an o(1) failure
probability even when allowing computationally unbounded protocols that use an arbitrarily large
number of rounds and an unbounded alphabet.
Proof. Suppose that there is an algorithm that solves the exchange problem under adversarial error-
rate 1/3, in N rounds. We study this algorithm simply with 1-bit inputs. Let SX,Y denote the
setting where Alice receives input X and Bob gets input Y . We specifically work only with settings
12
S0,0, S0,1, and S1,0. Note that if a party has an input 1, it knows in which of these three settings
we are. The idea is to present an adversarial strategy that changes the receptions of the party,
or parties, that have a 0 input so as to make that party, or parties, to not be able to distinguish
(between two of) the settings.
For simplicity, we first assume that the algorithm is deterministic and we also ignore the rounds
where both parties listen. Note that by the definition of the model for adaptive algorithms (see
Section 2), for these rounds, the adversary can deliver arbitrary messages to the parties at no cost.
For this lower bound, we need to define the party that becomes the base of indistinguishability
(whom we confuse by errors) in a more dynamic way, compared to that in Lemma 4.2 or in [5, Claim
10]. For this purpose, we first study the parties that have input 0 under a particular pattern of
received messages (regardless of the setting in which they are), without considering whether the
adversary has enough budget to create this pattern or not. Later, we argue that the adversary
indeed has enough budget to create this pattern for at least one party and make that party confused.
To determine what should be delivered to each party with input 0, the adversary cultivates
dummy personalities ]Alice0, ]Alice1, gBob0, gBob1, by simulating Alice or Bob respectively in settings
S0,0, S1,0, S0,0, and S0,1, where each of these settings is modified by adversarial interferences (to
be specified). Later, when we say that in a given round, e.g. "the adversary connects dummy
personality gBob1 to Alice", we mean that the adversary delivers the transmission of gBob1 in that
round to Alice5. For each setting, the adversary uses one method of interferences, and thus, when
we refer to a setting, we always mean the setting with the related adversarial interferences included.
We now explain the said pattern of received messages. Suppose that Alice has input 0 and
consider her in settings S0,0 and S0,1, as a priori these two settings are identical to Alice. Using
connections to dummy personalities, the adversary creates the following pattern: In the first 2N/7
rounds in which Alice listens alone, her receptions will be made to imply that Bob also has a
0. This happens with no adversarial interference in setting S0,0, but it is enforced to happen in
setting S0,1 by the adversary via connecting to Alice the dummy personality gBob0 cultivated in
setting S0,0. Thus, at the end of those 2N/7 listening-alone rounds of Alice, the two settings are
indistinguishable to Alice. In the rest of the rounds where Alice listens alone, the receptions will be
made to look as if Bob has a 1. That is, the adversary leaves those rounds of setting S0,1 intact, but
in rounds of setting S0,0 in which Alice listens alone, the adversary connects to Alice the dummy
personality gBob1 cultivated in setting S0,1 (with the adversarial behavior described above).
The adversary creates a similar pattern of receptions for Bob when he has an input 0, in settings
S0,0 and S1,0. That is, the first 2N/7 of his alone receptions are made to imply that Alice has a 0
but the later alone-receptions imply that Alice has a 1.
The described reception pattens make Alice unable to distinguish S0,0 from S0,1 and also they
make Bob unable to distinguish S0,0 from S1,0. However, the above discussions ignore the adver-
sary's budget. We now argue that the adversary indeed has enough budget to create this reception
pattern to confuse one or both of the parties.
Let xA be the total number of rounds where Alice listens, when she has input 0 and her receptions
follow the above pattern. Similarly, define xB for Bob. If xA ≤ 4N/7, then the adversary indeed
has enough budget to make the receptions of Alice in settings S0,0 and S0,1 follow the discussed
behavior, where the first 2N/7 alone-receptions of Alice are altered in S0,1 and the remaining alone-
receptions are altered in S0,0. Thus, if xA ≤ 4N/7, the adversary has a legitimate strategy to make
5This is assuming gBob1 transmits in that round, we later discuss the case where both Alice and gBob1 listen later.
13
Alice confused between S0,0 and S0,1. A similar statement is true about Bob: if xB ≤ 4N/7, the
adversary has a legitimate strategy to make Bob confused between S0,0 and S0,1.
Now suppose that xA > 4N/7 and xB > 4N/7. In this case, the number of alone-receptions of
Alice is at most xA −(xA +xB −N ) = N −xB ≤ 3N/7 and similarly, the number of alone-receptions
of Bob is at most xB − (xA + xB − N ) = N − xA ≤ 3N/7. This is because xA + xB − N is a
lower bound on the overlap of the round that the two parties listen. In this case, the adversary
has enough budget to simultaneously confuse both Alice and Bob of setting S0,0; Alice will be
confused between S0,0 and S0,1 and Bob will be confused between S0,0 and S1,0. For this purpose,
in setting S0,0, the adversary leaves the first 2N/7 alone-receptions of each party intact but alters
the remaining at most N/7 alone-receptions of each party, for a total of at most 2N/7 errors. On
the other hand, in setting S0,1, only 2N/7 errors are used on the first 2N/7 alone-receptions of
Alice and similarly, in setting S1,0, only 2N/7 errors are used on the first 2N/7 alone-receptions of
Bob. Note that these errors make the receptions of each party that has input 0 follow the pattern
explained above.
We now go back to the issue of the rounds where both parties listen. The rounds of S0,0 in
which both parties listen are treated as follows: The adversary delivers the transmission of gBob1
(cultivated in setting S0,1) to Alice and delivers the transmission of ]Alice1 (cultivated in setting
S1,0) to Bob. Recall that the adversary does not pay for these interferences. Furthermore, note
that these connections make sure that these all-listening rounds do not help Alice to distinguish
S0,0 from S0,1 and also they do not help Bob to distinguish S0,0 from S1,0.
Finally, we turn to covering the randomized algorithms. Note that for this case we only show
that the failure probability of the algorithm is not o(1) as just by guessing randomly, the two parties
can have success probability of 1/4.
First suppose that P r[xA ≤ 4N/7] ≥ 1/3. Note that the adversary can easily compute this
probability, or even simpler just get a (1 + o(1))-factor estimation of it. If P r[xA ≤ 4N/7] ≥ 1/3,
then the adversary will hedge his bets on that xA ≤ 4N/7, and thus, it will try to confuse Alice.
In particular, he gives Alice an input 0 and tosses a coin and gives Bob a 0 or a 1, accordingly.
Regarding whether Bob gets input 0 or 1, the adversary also uses the dummy personalities gBob0
and gBob1, respectively. With probability 1/3, we will have that in fact xA ≤ 4N/7, and in this
case the adversary by determining whether Alice hears from the real Bob or the dummy Bob,
the adversary makes Alice receive the messages with the pattern described above. This means
Alice would not know whether Bob has a 0 or a 1. Hence, the algorithm fails with probability
at least 1/6 (Alice can still guess in this case which is correct with probability 1/2). Similarly, if
P r[xB ≤ 4N/7] ≥ 1/3, then adversary will make Bob confused between S0,0 and S1,0. On the other
hand, if P r[xA ≤ 4N/7] < 1/3 and P r[xB ≤ 4N/7] < 1/3, then just using a union bound we know
that Pr[xA > 4N/7&xB > 4N/7] ≥ 1/3. In this case, the adversary gambles on the assumption
that it will actually happen that xA > 4N/7 and xB > 4N/7. This assumption happens with
probability at least 1/3, and in that case, the adversary makes Alice confused between S0,0 and
S0,1 and Bob confused between S0,0 and S1,0, simultaneously, using the approach described above.
Hence, in conclusion, in any of the cases regarding random variables xA and xB, the adversary can
make the algorithm fail with probability at least 1/6.
14
Figure 2: A Binary Interactive Protocol in the Canonical Form
5 Natural Interactive Coding Schemes With Large Alphabets
We start by presenting a canonical format for interactive communication and then present our
natural non-adaptive and adaptive coding schemes.
5.1
Interactive Protocols in Canonical Form
We consider the following canonical form of an n-round two party protocol over alphabet Σ: We call
the two parties Alice and Bob. To define the protocol between them, we take a rooted complete
Σ-ary tree of depths n. Each non-leaf node has Σ edges to its children, each labeled with a
distinct symbol from Σ. For each node, one of the edges towards children is preferred, and these
preferred edges determine a unique leaf or equivalently a unique path from the root to a leaf. We
say that the set X of the preferred edges at odd levels of the tree is owned by Alice and the set Y
of the preferred edges at even levels of the tree is owned by Bob. This means that at the beginning
of the protocol, Alice gets to know the preferred edges on the odd levels and Bob gets to know
the preferred edges on the even levels. The knowledge about these preferred edges is considered
as inputs X and Y given respectively to Alice and Bob. The output of the protocol is the unique
path from the root to a leaf following only preferred edges. We call this path the common path
and the edges and nodes on this path the common edges and the common nodes. The goal of the
protocol is for Alice and Bob to determine the common path. The protocol succeeds if and only
if both Alice and Bob learn the common path. Figure 2 illustrates an example: Alice's preferred
edges are indicated with blue arrows and Bob's preferred edges are indicated with red arrows, and
the common leaf is indicated by a green circle.
It is easy to see that if the channel is noiseless, Alice and Bob can determine the common path
of a canonical protocol P by performing n rounds of communication. For this Alice and Bob move
down on the tree together simply by following the path of preferred edges; they take turns and
exchange one symbol of Σ per round, where each symbol indicates the next common node. We call
this exchange the execution of the protocol P .
5.2 Natural Non-Adaptive Coding Schemes
In this section, we present a non-adaptive coding scheme which can be viewed as a straightforward
formalization of the natural high level approach presented in Section 3. This coding scheme tolerates
the optimal error rate of 1/4 − ǫ when unique decoding and simultaneously the optimal error rate of
1/2−ǫ when list decoding. The coding scheme is furthermore simple, intuitive, and computationally
15
Algorithm 2 Natural Non-Adaptive Coding Scheme at Alice's Side
B ∪ EA
Receive edge-set E ′
E ← E ′
if E is a valid edgeset then
1: X ← the set of Alice's preferred edges;
2: EA ← ∅;
3: N ← 2n
ǫ ;
4: for i = 1 to N do
5:
6:
7:
8:
9:
10:
11:
12:
13:
Add one vote to this leaf
r ← ∅
follow the common path in E
if the common path ends at a leaf then
else
⊲ EA is Alice's set of important edges. We preserve that always EA ⊆ X
B;
⊲ E ′
B is the received version of Bob's important edge-set EB
r ← {e} where e is the next edge in X continuing the common path in E (if any)
else
EA ← EA ∪ r
Send EA
14:
15:
16:
17:
18: Output the leaf with the most votes for unique decoding
19: Output the O(1/ǫ) leaves with the most votes for list decoding
Send EA
efficient, but it makes use of a large O( n
ǫ )-bit size alphabet. We note that one can also view this
algorithm as a simplified version of the Braverman-Rao algorithm [5] with larger alphabet size and
without using tree codes [10].
The algorithm, for which a pseudo code is presented in Algorithm 2, works as follows: In the
course of the algorithm, Alice and Bob respectively maintain sets EA and EB which are a subset
of their own preferred tree edges that are considered to be important. We call these important
edge-sets or sometimes simple edge-sets. Initially these edge-sets are empty and in each iteration,
Alice and Bob add one edge to their sets. In each iteration, when a party gets a turn to transmit,
it sends its edge-set to the other party. The other party receives either the correct edge-set or a
corrupted symbol which represents an edge-set made up by the adversary. In either case, the party
combines the received edge-set with its own important edge-set and follows the common path in
this set. Then, if this common path can be extended by the party's own set of preferred edges by
a new edge e, the party adds this edge e to its edge-set, and sends this new edge-set in the next
round. If, on the other hand, the common path already ends at a leaf, then the party registers this
as a vote for this leaf and simply re-sends its old edge-set. In the end, both parties simply output
the the leaf (respectively the O(1/ǫ) leaves) with the most votes for unique decoding (resp., for list
decoding).
Analysis We now prove that Algorithm 2 indeed achieves the optimal tolerable error rates for
non-adaptive unique decoding and list decoding.
Theorem 5.1. For any ǫ > 0, Algorithm 2 is a deterministic polynomial time non-adaptive simu-
lator with alphabet size of O( n
that tolerates an error-rate of 1/4 − ǫ
for unique decoding, and also tolerates an error-rate of 1/2 − ǫ for list decoding with a list of size
1
ǫ .
ǫ )-bits and round complexity 2n
ǫ
16
Proof. Clearly, both EA and EB grow by at most one edge per round. Furthermore, the edges
always attach to an already present edge and therefore, each of these edge-sets always forms a
subtree with size at most N starting at the root of the tree of the canonical form, which has depth
n. One can easily see that each such subtree can be encoded using O(N ) bits, e.g., by encoding
each edge of the breadth first search traversal of the subtree using alphabet of size 3 (indicating
"left", "right" or "up"). Hence, parties can encode their edge-sets using O( n
ǫ )-bits symbols, which
shows that the alphabet size is indeed as specified.
We now prove the correctness of the algorithm, starting with that of unique decoding. Note that
any two consecutive rounds in which Bob and Alice's transmissions are not corrupted by adversary,
one of the following two good things happens: Either the path in EA ∪ EB gets extended by at
least one edge, or both Alice and Bob receive a vote for the correct leaf.
Now suppose that the simulation runs in N = 2n/ǫ rounds which can be grouped into n/ǫ
round pairs. Given the error rate of 1/4 − ǫ at most a 1/2 − 2ǫ fraction of these round pairs can
be corrupted, which leaves a total of N/2(1/2 + 2ǫ) uncorrupted round pairs. At most n of these
round pairs grow the path while the remaining N/2(1/2 + 2ǫ) − n rounds vote for the correct leaf.
This results in at least N (1/2 + 2ǫ) − ⌈n/2⌉ = n
2ǫ + 2n − n > N/4 out of N/2 votes being correct.
For the list decoding, with error rate 1/2 − ǫ, we get that at most 1 − 2ǫ fraction of round-pairs
are corrupted, and thus at least N ǫ = 2n uncorrupted pairs exist. Hence, the correct leaf gets a
vote of at least 2n − n. Noting that the total number of the votes that one party gives to its leaves
is N/2 = n
ǫ , we get that the correct leaf has at least a ǫ fraction of all the votes. Therefore, if we
output the 1/ǫ leaves with the most votes, the list will include the correct leaf.
5.3 Natural Adaptive Coding Scheme
In this section we show that the simplest way to introduce adaptation into the natural coding
scheme presented in Algorithm 2. In particular we use the simple rules specified as 3.7 and show
that this leads to a coding scheme tolerating an error rate of 2/7 − ǫ, the optimal error rate for this
setting.
Next we explain how to incorporate the rules specified in 3.7 easily and efficiently into Algorithm
2. For this we note that for example if one party has a leaf with more than (2/7 − ǫ)N votes, since
adversary has only budget of (2/7 − ǫ)N , this leaf is the correct leaf and thus the party can follow
the second rule. Generalizing this idea, we use the rule that, if the party has a leaf v such that
only at most N
7 votes are on leaves other than v, then the party can safely assume that this is the
correct leaf. In our proof we show that this assumption is indeed safe and furthermore, at least one
party can safely decode at the end of the first 6/7 fraction of the simulation. Since both parties
know this in advance, if a party can not safely decode after 6/7 fraction of the time, it knows that
the other party has safely decoded -- which corresponds to the condition in the first rule -- and in
this case, the party only listens for the last 1/7 fraction of the protocol. The pseudo code for this
coding scheme is presented in Algorithm 3.
Theorem 5.2. Algorithm 3 is a deterministic adaptive coding scheme with alphabet size of O( n
bits, round complexity of O( n
of 2/7 − ǫ for unique decoding.
ǫ )-
ǫ ), and polynomial computational complexity that tolerates an error-rate
Proof. First, we show that if at the end of 6N
7 of
7
which are dedicated to one leaf v, then this party can safely assume that this leaf v is the correct
leaf. Proof of this part is by a simply contradiction. Note that if the party has s votes, then
rounds, one party has t votes, s ≥ t − N
17
Algorithm 3 Natural Adaptive Coding Scheme at Alice's Side
B;
ǫ );
B ∪ EA
7 N do
Receive edge-set E ′
E ← E ′
if E is a valid edgeset then
1: X ← the set of Alice's preferred edges;
2: EA ← ∅;
3: N ← Θ( n
4: for i = 1 to 6
5:
6:
7:
8:
9:
10:
11:
12:
13:
Add one vote to this leaf
r ← ∅
follow the common path in E
if the common path ends at a leaf then
else
r ← {e} where e is the next edge in X continuing the common path in E (if any)
else
7 do
Send EA
EA ← EA ∪ r
Send EA
7 then
for i = 1 to N
14:
15:
16:
17:
18: Let s be number of votes of the leaf with the most votes and t be the total number of votes
19: if s ≥ t − N
20:
21:
22: else
23:
24:
25:
26:
27:
28:
29: Output the leaf with the most votes
follow the common path in E
if the common path ends at a leaf then
Receive edge-set E ′
B; E = E ′
if E is a valid edge-set then
Add one vote to this leaf
for i = 1 to N
Send EA
B ∪ EA
7 do
7 ≥ 2N
7 − t + t − N
there are at least 3N
7 − t that either stopped the growth of the path or turned an edge-set into a
nonvalid edge-set. Furthermore, if v is not the correct leaf, then the votes v are created by errors
of adversary which means that adversary has invested s errors on turning the edge-sets sent by
the other party into other valid-looking edge-sets. Hence, in total, adversary has spent at least
3N
7 − t + s ≥ 3N
7 errors which is a contradiction.
Now that we know that the rule for safely decoding at the end of 6N
7
rounds is indeed safe, we
show that at least one party will safely decode at that point of time. Suppose that no party can
decode safely. Also assume that Alice has tA votes, rA of which are votes on the good leaf. That
means at least adversary has turned at least tA − rA edge-sets sent by Bob into other valid-looking
edge-sets. Similarly, tB − rB errors are introduced by the adversary on edge-sets sent by Alice. If
neither Alice nor Bob can decode safely, we know that tA − rA ≥ N
7 , which means
that in total, adversary has introduced at least 2N
7 errors. Since this is not possible give adversary's
budget, we conclude that at the end of 6N
7
rounds, at least one party decodes safely.
7 and tB − rB ≥ N
Now suppose that only one party, say Alice, decodes safely at the end of 6N
7
rounds. Then, in
the last N
7 rounds, Bob is listening and Alice is sending. In this case, we claim that Bob's leaf that
gets the majority of the votes at the end is the correct leaf. The reason is, suppose that Bob has
18
7 rounds and t′
tB votes from the first 6N
the correct leaf had rB votes from the first 6N
the adversary has introduced at least ( 3N
errors. Since adversaries budget is at most ( 2
clearly Bob has at most 4N
7 − tB) + ( N
B) + (tB − rB) + (t′
7 − ǫ)N , we get that rB + r′
7 votes in total, the correct leaf has the majority.
7 − t′
B votes from the last N
7 rounds and r′
7 rounds. Furthermore, suppose that
B votes from the last N
7 rounds. Then,
7 − rB + r′
B
7 . Hence, since
B − r′
B > 2N
B) = 4N
6 Coding Schemes with Small Alphabet and O(n2) Rounds
In this section, we show how to implement the natural coding schemes presented as Algorithms 2
and 3 over a channel supporting only constant size symbols at the cost of increasing the number of
rounds to O(n2).
ǫ3 ) symbols and send this encoding in O( n
ǫ2 ), using an ECC with relative distance 1 − ǫ/10, alphabet size O( 1
ǫ3 ) rounds. We call each such O( n
To emulate Algorithms 2 and 3 over a finite alphabet, we use Error Correcting Codes (ECC) and
list decoding. In particular, on the transmission side, we encode each edge-set, which will have size
at most O( n
ǫ ), and code-length
O( n
ǫ3 ) rounds related to
transmission of one edge-set a block. On the receiver side, we use a list decoder to produce a list of
O( 1
ǫ ) edge-sets such that, if the error-rate in the block is at most 1 − ǫ/3, then one of the edge-sets
in the list in indeed equal to the transmitted edge-set. If the error-rate is greater than 1 − ǫ/3, the
list provided by the list decoder does not need to provide any guarantee.
We use O( n
to O( n
each of these edge-sets to the set of important edges leads to the list decoding result:
ǫ3 ) rounds for each block as because of list decoding, now each edge-set contains up
ǫ ) edges in Algorithms 2 and 3). Adding edges corresponding to
ǫ2 ) edges (compare to O( n
Lemma 6.1. If the error rate is at most 1/2 − ǫ, then in both parties, the set of O(1/ǫ2) leaves
with the highest votes includes the correct leaf.
Proof. This is because, with error-rate 1/2 − ǫ, the adversary can corrupt at most 1/2 − ǫ/3 blocks
beyond corruption rate of 1 − ǫ/3. Hence, we are now in a regime that at most 1/2 − ǫ/3 fraction
of edge-sets are corrupted, each corrupted possibly even completely, and for every other edge-
set, the (list) decoding includes the correct transmitted edge-set. Hence, similar to the proof of
Theorem 5.1, we get that the correct leaf gets a vote of at least Ω(ǫN ). On the other hand, now
each block might give a vote of at most O(1/ǫ) to different leaves and thus, the total weight is at
most O(N/ǫ). Therefore, the correct leaf is within the top O(1/ǫ2) voted leaves.
We next explain a simple idea that allows us to extend this result to unique decoding; specifically
non-adaptive unique decoding when error-rate is at most 1/4 − ǫ, and adaptive unique decoding
when error-rate is at most 2/7 − ǫ.
The idea is to use a variation of Forney's technique for decoding concatenated codes [7]. Recall
that each received edge-set might lead to two things: (1) extending the common path, or (2)
adding a vote to a leaf. While we keep the first part as above with list decoding, we make the voting
weighted. In particular, in the receiver side, we take each edge-set leading to a leaf (when combined
with local important edge set) as a vote for the related leaf but we weight this vote according to
the hamming-distance to the received block. More precisely, if the edge-set has relative distance δ
from the received block, the related leaf gets a vote of max{1 − 2δ, 0}.
Using this weighting function, intuitively we have that if the adversary corrupts an edge-set
2 corruption rate, even though the edge-sets gets added to the set of important edges, in the
to 1
19
weighting procedure this edge-set can only add at most weight 2ǫ to any (incorrect) leaf. Hence,
for instance if the adversary wants to add a unit of weight to the votes of an incorrect leaf, it has
to almost completely corrupt the symbols of the related block.
Lemma 6.2. If the error rate is at most 1/4 − ǫ, then in both parties, the leaf with the highest
weighted votes is the correct leaf.
Proof. We show that the correct leaf ug get strictly more weighted votes compared to any other
leaf ub. For this, we use a potential Φ = W (ug) − W (ub), that is, the total weight added to the
good leaf minus that of the bad leaf, and we show this potential to be strictly positive. Let Pc
be the set of edges of the correct path. Let t be the time at which point the common path in
EA ∪ EB ∪ Pc ends in a leaf. First note that for each two consecutive blocks before time t in which
the corruption rate is at most 1 − ǫ/3, the common path in EA ∪ EB ∪ Pc gets extended by (at
least) one edge (towards the correct leaf). Hence, at most n such ("not completely corrupted")
block pairs are spent on growing the common path in EA ∪ EB ∪ Pc and also, t happens after at
most N 2(1/2 − 2ǫ)(1 + ǫ/3) < N (1 − 2ǫ) < N − 4n blocks. That is, t happens at least 4n blocks
before the end of the simulation. Before time t, we do not add any weight to W (ug) but each block
corrupted with rate x ≥ 1/2 changes Φ in the worst case as 1 − 2x ≤ 0. For each block after time
t, each block corrupted with rate x ∈ [0, 1 − ǫ/3] changes Φ in the worst case by 1 − 2x and each
block corrupted to rate x ∈ (1 − ǫ/3, 1] changes Φ by at most −1. These two cases can be covered
as a change of no worse than 1 − 2(1 + ǫ/3)x.
In total, since adversary's error rate is at most
(1/4 − ǫ), in total of before and after time t, we get that it can corrupt at most 1/2 − 2ǫ fraction
of the receptions of one party and thus, Φ ≥ 1 − 2(1/2 − 2ǫ)(1 + ǫ/3) ≥ 3ǫ > 0.
For the 2/7 − ǫ adaptive algorithm, we first present a lemma about the total weight assigned
to the leaves due to one edge-set reception.
Lemma 6.3. For each list decoded block that has corruption rate ρ, the summation of the weight
given to all the codewords in the list is at most 1 − 2ρ + 3ǫ/5.
Proof. First we show that only at most 3 codewords receive nonzero weight. The proof is by
contradiction. Suppose that there are 4 codewords x1 to x4 that each agree with the received
string x in at least ǫ/10 fraction of the symbols. Furthermore, for each xi, let Si be the set of
coordinates where xi agrees with the received string x. Let ℓ be the length of string x. Note that
∀i 6= j ∈ {1, 2, 3, 4}, Si ≥ ℓ/2 and Si ∩ Sj ≤ ǫ/10. Hence, with a simple inclusion-exclusion, we
have
ℓ ≥ ∪i Si ≥ X
Si −X
4ℓ
2
6ǫ
10
Si ∩ Sj ≥
−
> ℓ,
which is a contradiction.
i
i<j
Having that at most 3 codewords receive nonzero weight, we now conclude the proof. Let x1 to
x3 be the codewords that receive nonzero weights and assume that x1 is the closest codeword to x.
We have
∀i > 1, ∆(x, xi) ≥ ∆(x1, xi) − ∆(x, x1) ≥ 1 − ǫ/10 − 1/2 ≥ 1/2 − ǫ/10.
Thus, the weight that x2 or x3 get is at most 2ǫ/10. On the other hand,
∆(x1, x) ≥ min{ρ, 1 − ǫ/10 − ρ} =
1 − ǫ/10
2
−
1 − ǫ/10
2
− ρ.
20
Thus, the weight that x1 receives is at most ǫ/10 + 1 − ǫ/10 − 2ρ ≤ 2ǫ/10 + 1 − 2ρ. Summing up
with the weight given to x2 and x3, we get that the total weight given to all codewords is at most
3ǫ/5 + 1 − 2ρ.
The algorithm is as in Algorithm 3, now enhanced with list decoding and weighted voting. In
the end of the 6N
rounds, a party decodes safely (and switched to only transmitting after that) if
7
for the leaf u that has the most votes, the following holds: Ψ = W (u)+W∅−W (v)
> 1/7 where here,
W (v) is the weighted vote of the second leaf that has the most votes and W∅ is the total weight
for decoded codewords that are inconsistent with the local important edge set (and thus mean an
error). We call Ψ the parties confidence.
N
The following lemma serves as completing the proof of Theorem 3.1.
Lemma 6.4. If the error rate is at most 2/7 − ǫ, then in the end, in both parties, the leaf with the
highest weighted votes is the correct leaf.
Proof. We first show that at the end of the first 6N/7 rounds, at least one party has confidence at
least 1/7 + ǫ/2. For this, we show the summation of the confidence of the two parties to be at least
2/7 + ǫ. The reasoning is similar to the proof that of Lemma 6.2. Let t be the time at which point
the common path in EA ∪ EB ∪ Pc ends in a leaf and note that for each two consecutive blocks
before time t in which the corruption rate is at most 1 − ǫ/3, the common path in EA ∪ EB ∪ Pc gets
extended by (at least) one edge (towards the correct leaf). Hence, at most n such ("not completely
corrupted") block pairs are spent on growing the common path in EA ∪ EB ∪ Pc. Furthermore,
t happens after at most 2 · N (2/7 − ǫ)(1 + ǫ/3) < N (4/7 − ǫ) blocks. Specifically, t happens at
least N (2/7 + ǫ) blocks before the end of the first 6N/7 blocks. On the other hand, for each
corruption rate x on one block, only the confidence of the party receiving it gets effected and
in worst case it goes down by 1−2x(1+ǫ/3)
. Since the adversary's total budget is N (2/7 − ǫ), at
the end of the first 6N/7 rounds, the summation of the confidence of the two parties is at least
6N/7−2n−2N (2/7−ǫ)(1+ǫ/3)
N
N
≥ N (2/7 + ǫ).
Now we argue that if a party decodes at the end of 6N/7 blocks because it has confidence
at least 1/7, then the decoding of this party is indeed correct. Proof is by contradiction. Using
Lemma 6.3, each block with corruption rate x can add a weight of at most max{2x − 1 + 3ǫ/5, 0} to
the set of bad leaves or those incorrect codeword edge-sets that do not form a common path with
the local edge-sets, i.e., the weight given to W∅. Hence, knowing that the adversary has budget of
N (2/7 − ǫ), it can create a weight of at most 2N (2/7 − ǫ) − N/3(1 − 3ǫ/5) < N/7 − ǫ/5 < N/7,
which means that there can not be a confidence of more than 1/7 on an incorrect leaf.
The above shows that at least one party decides after 6N/7 blocks and if a party decides then,
it has decided on the correct leaf. If a party does not decide, it listens for the next N/7 blocks
where the other party is constantly transmitting. It is easy to see that in this case, this leaf that
has the maximum vote in this listening party is the correct party.
7 Adaptivity in the Shared Randomness Setting
In this section, we present our positive and negative results for adaptive protocols with access to
shared randomness.
In the shared randomness setting of Franklin et al. [8] Alice and Bob have access to a shared
random string that is not known to the adversary. As shown in Fanklin et al. [8] the main advantage
21
of this shared randomness is that Alice and Bob can use a larger communication alphabet for their
protocol and agree on using only random subset of it which is independent from round to round.
Since the randomness is not known to the adversary any attempt at corrupting a transmission
results with good probability on an unused symbol. This makes most corruptions detectable to
Alice and Bob and essentially results in corruptions being equivalent to erasures. Fanklin et al.
call this way of encoding the desired transmissions into a larger alphabet to detect errors a blue-
berry code. It is well known for error correcting codes that erasures are in essence half as bad as
corruption. Showing that the same holds for the use of tree codes in the algorithm of Braverman and
Rao translates this factor two improvement to the interactive setting. For non-adaptive protocols
this translates into a tolerable error rate of 1/2 − ǫ instead of a 1/4 − ǫ error rate. In what follows
we show that allowing adaptive protocols in this setting allows to further rise the tolerable error
rate to 2/3 − ǫ. We also show that no coding scheme can tolerate more than an 2/3 error rate:
Theorem 7.1. [The second part of Theorem 3.3] In the shared randomness setting there is no
(deterministic or randomized) adaptive protocol that robustly simulates the Exchange Protocol for
an error rate of 2/3 with an o(1) failure probability even when allowing computationally unbounded
protocols that use an arbitrarily large number of rounds and an unbounded alphabet.
With the intuition that in the shared randomness setting corruptions are almost as bad as
erasures we prove Theorem 7.1 directly for an strictly weaker adversary, which can only delete
transmitted symbols instead of altering them to other symbols. Formally, when the adversary
erases the transmission of a party, we say that the other party receives a special symbol ⊥ which
identifies the erasure as such.
Proof. Suppose that there is a coding scheme with N rounds that allows Alice and Bob exchange
their input bits, while the adversary erases at most 2N/3 of the transmissions. Consider each party
in the special scenario in which this party receives a ⊥ symbol whenever it listens. Let xA and xB
be the (random variable of) the number of rounds where, respectively, Alice and Bob listen when
they are in this special scenario.
Suppose this number is usually small for one party, say Alice. That is, suppose we have Pr[xA ≤
2N/3] ≥ 1/3. In this case the adversary gambles on the condition xA ≤ 2N/3 and simply erases
Bob's transmission whenever Alice listens and Bob transmits. Furthermore, if both parties listen,
the adversary delivers the symbol ⊥ to both parties at no cost. This way, with probability at least
1/3, Alice stays in the special scenario while the adversary adheres to its budget of at most 2N/3
erasures. Since, in this case, Alice only receives the erasure symbol she cannot know Bob's input.
Therefore, if the adversary chooses Bob's input to be a random bit the output of Alice will be
wrong with probability at least 1/2 leading to a total failure probability of at least 1/6.
If on the other hand Pr[xA ≤ 2N/3] ≤ 1/3 and Pr[xB ≤ 2N/3] ≤ 1/3, then even a union bound
shows that Pr[(xA ≥ 2N/3) ∧ (xA ≥ 2N/3)] ≥ 1/3. In this case, the adversary tries to erase all
transmissions of both sides. Indeed, if xA ≥ 2N/3 and xA ≥ 2N/3 this becomes possible because in
this case there must be are at least N/3 rounds in which both parties are listening simultaneously.
For these rounds, the adversary gets to deliver the erasure symbol to both sides at no cost. In
this case, which happens with probability 1/3 there remain at most 2N/3 rounds in which one
of the parties listens alone which the adversary can erase without running out of budget. With
probability Pr[(xA ≥ 2N/3) ∧ (xA ≥ 2N/3)] > 1/3 the adversary can thus prevent both parties
from learning the other party's input. Choosing the inputs randomly results in at least one party
22
being wrong in its decoding with probability 3/4 which in total leads to a failure probability of at
least 1/3 · 3/4 > 1/6.
Turning to positive results we first show how to solve the Exchange Problem robustly with an
error rate of 2/3 − ǫ. While this is a much simpler setting it already provides valuable insights
into the more complicated general case of robustly simulating arbitrary n-round protocols. For
simplicity we stay with the adversary who can only erase symbols. Our general simulation result
presented in Theorem 7.4 works of course for the regular adversary.
Lemma 7.2. Suppose ǫ > 0. In the shared hidden randomness model with rate adaptation there is
a protocol that solves the Exchange Problem in O(1/ǫ) rounds under an adversarial erasure rate of
2/3 − ǫ.
Proof. The protocol consists of 3/ǫ rounds grouped into three parts containing 1/ǫ rounds each.
In the first part Alice sends her input symbol in every round while Bob listens. In the second part
Bob sends his input symbol in each round while Alice listens.
In the last part each of the two
parties sends its input symbol if and only if they have received the other parties input and not just
erasures during the first two parts; otherwise a party listens during the last part. Note that the
adversary has a budget of 3/ǫ · (2/3 − ǫ) = 2ǫ − 1 erasures which is exactly one erasure to little
to erase all transmission during the first two parts. This results in at least one party learning the
other party's input symbol. If both parties learn each others input within the first two rounds then
the algorithm succeeded. If on the other hand one party, say without loss of generality Alice, only
received erasures then Bobs received her input symbol at least once. This results in Bob sending
his input symbol during the last part while Alice is listens. Bob therefore sends his input symbol a
total of 2/ǫ times while Alice listens. Not all these transmissions can be erased and Alice therefore
also knows Bob's input symbol in the end.
Lemma 7.2 and even more the structure of the robust Exchange Protocol presented in its proof
already give useful insights into how to achieve a general robust simulation result. We combine
these insights with the blue-berry code idea of [8] to build what we call an adaptive exchange block.
An adaptive exchange block is a simple three round protocol which will serve as a crucial building
block in Theorem 7.4. An adaptive exchange block is designed to transmit one symbol σA ∈ Σ from
Alice to Bob and one symbol σB ∈ Σ from Bob to Alice. It assumes a detection parameter δ < 1
and works over an alphabet Σ′ of size Σ′ = ⌈Σ/δ⌉ and works as follows: First, using the shared
randomness, Alice and Bob use their shared randomness to agree for each of the three rounds on
an independent random subset of of the larger communication alphabet Σ′ to be used. Then, in
the first round of the adaptive exchange block Alice sends the agreed equivalent of σA while Bob
listens. In the second round Bob sends the equivalent of σB while Alice listens. Both parties try
to translate the received symbol back to the alphabet σ and declare a (detected) corruption if this
is not possible. In the last round of the adaptive exchange block a party sends the encoding of its
σ-symbol if and only if a failure was detected; otherwise a party listens and tries again to decode
any received symbol.
The following two properties of the adaptive exchange block easily verified:
Lemma 7.3. Regardless of how many transmissions an adversary corrupted during an adaptive
exchange block, the probability that at least one of the parties decodes a wrong symbol is at most 3δ.
In addition, if an adversary corrupted at most one transmission during an adaptive exchange
block the with probability at least 1 − δ both parties received their σ symbols correctly.
23
Proof. For a party to decode to a wrong symbol it must be the case that during one of the three
rounds the adversary hit a meaningful symbol in the alphabet Σ′ during a corruption. Since
Σ/Σ′ ≤ δ this happens at most with probability δ during any specific round and at most with
probability 3δ during any of the three rounds. To prove the the second statement we note that
in order for a decoding error to happen the adversary must interfere during the first two rounds.
With probability 1 − δ such an interference is however detected leading to the corrupted party to
resend in the third round.
Next we explain how to use the adaptive exchange block together with the ideas of [8] to prove
that adaptive protocols with shared randomness can tolerate any error rate below 2/3:
Theorem 7.4. [The first part of Theorem 3.3] Suppose Π is an n-round protocol over a constant
bit-size alphabet. For any ǫ > 0, there is an adaptive protocol Π′ that uses shared randomness and
robustly simulates Π for an error rate of 2/3 − ǫ with a failure probability of 2−Ω(n).
rounds which are grouped into 2n
Proof Sketch. The protocol Π′ consists of N = 6n
ǫ adaptive
ǫ
exchange blocks. The adversary has an error budget of 4n
ǫ − 6n corruptions. Choosing the with
parameter δ in the exchange blocks to be at most ǫ/(6 · 4 + 1) and using Lemma 7.3 together with
a Chernoff bound gives that with probability 1 − 2−Ω(n) there are at most n/4 adaptive exchange
blocks in which a wrong symbol is decoded. Similarly, the number of adaptive exchange blocks
in which only one corruption occurred but not both parties received correctly is with probability
1 − 2−Ω(n) at most n/3. Lastly, the number of adaptive exchange blocks in which the adversary can
force an erasure by corrupting two transmissions is at most ( 4n
ǫ − 6n)/2. We can therefore assume
that regardless of the adversaries actions at most n corruptions and at most 2n
ǫ − 8/3n erasures
happen during any execution of Π′, at least with the required probability of 1 − 2−Ω(n).
We can now apply the arguments given in [8]. These arguments build on the result in [5]
and essentially show that detected corruptions or erasures can essentially be counted as half a
corruption. Since almost all parts of the lengthy proofs in [5] and [8] stay the same, we restrict
ourselves to a proof sketch. For this we first note that the result in [5] continues to hold if instead
of taking turns transmitting for N rounds both Alice and Bob use N/2 rounds transmitting their
next symbol simultaneously. The extensions described in [8] furthermore show that this algorithm
still performs a successful simulation if the number effectively corrupted simultaneous rounds is at
least n rounds less than half of all simultaneous rounds. Here the number of effectively corrupted
simultaneous rounds is simply the number of simultaneous rounds with an (undetected) corruption
plus half the number of simultaneous rounds suffering from an erasure (a detected corruption).
Using one adaptive exchange block to simulate one simultaneous round leads to 2n
ǫ simultaneous
rounds and an effective number of corrupted rounds of n/4 + ( 2n
ǫ )/2 − n. Putting
everything together proves that with probability at least 1 − 2−Ω(n) the protocol Π′ successfully
simulates the protocol Π, as claimed.
ǫ − 8/3n)/2 < ( 2n
This way of constructing the adaptive protocol Π′ leads to an optimal linear number of rounds
and a constant size alphabet but results in exponential time computations. We remark that using
efficiently decodable tree codes, such as the ones described in the postscript to [10] on Schulman's
webpage, one can also obtain a computationally efficient coding scheme at the cost of using a large
O(n)-bit alphabet. Lastly, applying the same ideas as in Section 6 also allows to translate this
computationally efficient coding schemes with a large alphabet into one that uses a constant size
alphabet but a quadratic number of rounds.
24
References
[1] Z. Brakerski and Y. Kalai. Efficient interactive coding against adversarial noise. In Proceedings
of the IEEE Symposium on Foundations of Computer Science (FOCS), pages 160 -- 166, 2012.
[2] Z. Brakerski and M. Naor. Fast algorithms for interactive coding. In Proceedings of the ACM-
SIAM Symposium on Discrete Algorithms (SODA), pages 443 -- 456, 2013.
[3] M. Braverman. Coding for interactive computation: progress and challenges.
In Allerton
Conference on Communication, Control, and Computing (Allerton), pages 1914 -- 1921, 2012.
[4] M. Braverman. Towards deterministic tree code constructions. In Proceedings of the Innova-
tions in Theoretical Computer Science Conference (ITCS), pages 161 -- 167, 2012.
[5] M. Braverman and A. Rao. Towards coding for maximum errors in interactive communication.
In Proceedings of the ACM Symposium on Theory of Computing (STOC), pages 159 -- 166, 2011.
[6] K.-M. Chung, R. Pass, and S. Telang. Knowledge preserving interactive coding. Proceedings
of the IEEE Symposium on Foundations of Computer Science (FOCS), 2013.
[7] G. D. Forney. Concatenated codes. PhD thesis, MIT, 1966.
[8] M. Franklin, R. Gelles, R. Ostrovsky, and L. J. Schulman. Optimal coding for streaming
In Proceedings of International Cryptology
authentication and interactive communication.
Conference (CRYPTO), pages 258 -- 276, 2013.
[9] M. Ghaffari and M. Haeupler. Optimal Error Rates for Interactive Coding II: Efficiency and
List decoding. In arXiv, 2013.
[10] L. J. Schulman. Coding for interactive communication. IEEE Transactions on Information
Theory, 42(6):1745 -- 1756, 1996.
[11] M. Sudan. Decoding of reed solomon codes beyond the error-correction bound. Journal of
Complexity, 13(1):180 -- 193, 1997.
A Adaptivity as a Natural Model for Shared-Medium Channels
In this section we briefly describe two natural interpretations of our adaptivity setting as modeling
communication over a shared medium channel (e.g., a wireless channel). We first remark that one
obvious parallel between wireless channels (or generally shared medium channels) and our model
is that full-duplex communication, that is, sending and receiving at the same time, is not possible.
Beyond this, one natural interpretation relates our model to a setting in which the signals used
by these two parties are not much stronger than the average background noise level. In this setting
having the background noise corrupt the signal a ρ fraction of the time in an undetermined way is
consistent with assuming that the variable noise level will be larger than the signal level at most
this often. It is also consistent with the impossibility of distinguishing between the presence and
absence of a signal which leads to undetermined signal decodings in the case that both parties listen.
As always, the desire to avoid making possibly unrealistic assumptions about these corruptions
25
naturally gives rise to think of any undetermined symbols as being worst case or equivalently as
being determined by an adversary.
In a second related interpretation of our model one can think of the adversary as an active
malicious entity that jams the shared medium. In this case our assumptions naturally correspond
to the jammer having an energy budget that allows to over-shout a sent signal at most a ρ fraction
of the time. In this setting it is also natural to assume that the energy required for sending a fake
signal to both parties when no signal is present is much smaller than corrupting sent signals, and
does as such not significantly reduce the jammer's energy budget.
B Further Results Supporting the Invariability Hypothesis
In this section we mention several results that further support the Invariability Hypothesis.
We first remark that the lower bound in this paper are, in all settings and all properties, already
as strong as required by the hypothesis. Our positive results, as summarized in Corollary 3.6,
furthermore show that the invariability hypothesis holds if one allows either a large alphabet or
a quadratic number of rounds. Both assumptions lead to the communication rate being O(1/n).
Next we list several results which show that the invariability hypothesis provably extends beyond
these low rate settings:
1. Subsequent results in [9] show that:
The IH is true for all eight settings when allowing randomized algorithms with exponentially
small failure probability and a round blowup of (log∗ n)O(log∗ n).
2. Subsequent results in [9] show that:
The Invariability Hypothesis is true for all eight settings if one removes points 5. and 6., that
is, when one can use randomized protocols that can generate private and public encryption
keys which the adversary cannot break.
3. A subsequent result in [9] gives an efficient randomized coding scheme for non-adaptive unique
decoding which tolerates the optimal 1/4−ǫ error rate. This improves over the coding scheme
of Brakerski-Kalai [1] which requires an error rate below 1/16. In different terms this result
shows that:
The Invariability Hypothesis is true for unique decoding if one weakens point 5.
efficient randomized algorithms with exponentially small failure probability.
to allow
4. The result of Braverman and Rao [5] shows that:
The Invariability Hypothesis is true for non-adaptive unique decoding if one removes point 4.
(which requires protocols to be computationally efficient).
5. A subsequent result of the authors show that the 1/10 − ǫ distance parameter of the tree code
construction in [4] can be improved to 1 − ǫ which shows that:
The Invariability Hypothesis is true for unique decoding if one weakens point 4.
deterministic sub-exponential time computations.
to allow
6. The improved tree code construction mentioned in point 5. can also be used together with
the analysis of Franklin at al. [8] to show that:
The Invariability Hypothesis is true for non-adaptive unique decoding with shared randomness
if one weakens point 4. to allow deterministic sub-exponential time computations.
26
7. The improved tree code construction mentioned in point 5. can also be used together with
the ideas of Theorem 7.4 to show that:
The Invariability Hypothesis is true for adaptive unique decoding with shared randomness if
one weakens point 4. to allow deterministic sub-exponential time computations.
Lastly, we remark that the 1/6 − ǫ tolerable error rate of the knowledge preserving coding
schemes given by Chung et al. [6] can be improved to the optimal 1/4 − ǫ tolerable error rate of the
unique decoding setting. The communication blowup for knowledge preserving protocols is however
inherently super constant.
C Impossibility Results for List Decodable Interactive Coding
Schemes
In this section we prove that list decodable interactive coding schemes are not possible beyond an
error rate of 1/2. This holds even if adaptivity or shared randomness is allowed (but not both). We
remark that for both results we prove a lower bound for the n-bit Exchange Problem as list decoding
with a constant list size becomes trivial for any protocol with only constantly many different inputs.
The intuitive reason why shared randomness does not help in non-adaptive protocols to go
beyond an error rate of 1/2 is because the adversary can completely corrupt the party that talks
less:
Lemma C.1. There is no (deterministic or randomized) list decodable non-adaptive protocol that
robustly simulates the n-bit Exchange Protocol for an error rate of 1/2 with an o(1) failure probability
and a list size of 2n−1 regardless of whether the protocols have access to shared randomness and
regardless of whether computationally unbounded protocols that use an arbitrarily large number of
rounds and an unbounded alphabet are allowed.
Proof. We recall that non-adaptive protocols will for every round specify a sender and receiver in
advance, that is, independent from the history of communication. We remark that the proof that
follows continues to hold if these decisions are based on the shared randomness of the protocols.
The adversary's strategy is simple: It gives both Alice and Bob random inputs, then randomly
picks one of them, and blocks all symbols sent by this party by replacing them with a fixed symbol
from the communication alphabet. With probability at least 1/2 the randomly chosen player speaks
less than half the time and so the fraction of errors introduced is at most 1/2. On the other hand
no information about the blocked player's input is revealed to the other player and so other player
can not narrow down the list of possibilities in any way. This means that even when allowed a list
size of 2n−1 there is a probability of 1/2 that the list does not include the randomly chosen input.
This results in a failure probability of at least 1/4.
A 1/2 impossibility result also holds for list decodable coding schemes that are adaptive:
Lemma C.2. There is no (deterministic or randomized) list decodable adaptive protocol that ro-
bustly simulates the n-bit Exchange Protocol for an error rate of 1/2 with an o(1) failure probability
and a list size of 2n−1 even when allowing computationally unbounded protocols that use an arbi-
trarily large number of rounds and an unbounded alphabet.
27
To show that adaptivity is not helpful, one could try to prove that the adversary can imitate
one party completely without being detected by the other party. This, however, is not possible
with an error rate of 1/2 because both parties could in principle listen for more than half of the
rounds if no error occurs and use these rounds to alert the other party if tempering is detected.
Our proof of Lemma C.2 shows that this "alert" strategy cannot work. In fact, we argue that it
is counterproductive for Alice to have such a hypothetical "alert mode" in which she sends more
than half of the rounds. The reason is that the adversary could intentionally trigger this behavior
while only corrupting less than half of the rounds (since Alice is sending alerts and not listening
during most rounds). The adversary can furthermore do this regardless of the input of Bob which
makes it impossible for Alice to decode Bob's input. This guarantees that Alice never sends for
more than half of the rounds and the adversary can therefore simply corrupt all her transmissions.
In this case Bob will not be able to learn anything about Alice's input.
Proof. Suppose there exists a protocol that robustly simulates the n-bit Exchange Pro tocol for an
error rate of 1/2 using N rounds over an alphabet Σ. We consider pairs of the form (x, ~r) where
x ∈ {0, 1}n is an input to A lice and ~r = (r1, r2, . . . , rN ) ∈ ΣN is a string over the chann el alphabet
with one symbol for each round. We now look at the following hypothetical communication between
Alice and the adversary: Alice gets input x and samples her private randomness. In each round
she then decides to send or listen. If she listens in round i she receives the symbol ri. For every
pair (x, ~r) we now define p(x, ~r) to be the probability, taken over the private randomness of Alice,
that in this communication Alice sends at least N/2 rounds (and conversely listens for at most N/2
rounds). The adversaries strategy now makes the following case distinction: If there is one (x, ~r)
for which p(x, ~r) > 1/2 then the adversary picks a random input for Bob, gives Alice input x and
during the protocol corrupts any symbol Alice listens to according to ~r. By definition of p(x, ~r)
there is a probability of at least 1/2 that the adversary can do this without using more than N/2
corruptions. In such an execution Alice has furthermore no information on Bob's input and even
when allowed a list size of 2n−1 has at most a probability of 1/2 to include Bob's input into her
decoding list. Therefore Alice will fail to list decode correctly with probability at least 1/4. If on
the other hand for every (x, ~r) it holds that p(x, ~r) < 1/2 then the adversary picks a random input
for Alice and an arbitrary input for Bob and during the protocol corrupts any symbol Alice sends
to a fixed symbol σ ∈ Σ. Furthermore in a round in which both Alice and Bob listens it chooses
to deliver the same symbol σ to Bob. By definition of p(x, ~r) there is a probability of at least 1/2
that the adversary can do this without using more than N/2 corruptions. In such an execution Bob
now has no information on Alice's input and even when allowed a list size of 2n−1 he therefore has
at most a probability of 1/2 to include Alice's input into his decoding list. This leads to a failure
probability of 1/4 as well.
D Impossibility Results for Solving the Exchange Problem Adap-
tively
In this appendix we provide the proofs we deferred in Section 4.
For completeness and to also cover randomized protocols we first reprove Claim 10 of [5] which
states that no non-adaptive uniquely decoding protocol can solve the exchange problem for an error
rate of 1/4:
28
Lemma D.1 (Claim 10 of [5]). There is no (deterministic or randomized) protocol that robustly
simulates the Exchange Protocol for an error rate of 1/4 with an o(1) failure probability even when
allowing computationally unbounded protocols that use an arbitrarily large number of rounds and
an unbounded alphabet.
Proof. Suppose that there is an algorithm A with no rate adaptation that solves the exchange
problem under adversarial error-rate 1/4, in N rounds. We work simply with 1-bit inputs. Let
SX,Y denote the setting where Alice receives input X and Bob gets input Y . For simplicity, we
first ignore the rounds in which both parties listen; note that in those rounds the adversary can
deliver arbitrary messages to each of the parties at no cost.
First we explain adversary's strategy for setting S0,1: Consider the executions of A in setting
S0,0 with no interference from the adversary. Without loss of generality, we can assume that in
this execution, with probability at least 1/2, Alice listens in at most 1/2 of the rounds. Noting
the restriction that A has no rate-adaptivity, we get that in setting S0,0, regardless of the adver-
sary's interferences, with probability at least 1/2, Alice listens alone in at most 1/2 of the rounds.
Adversary generates a dummy personality gBob0 by simulating algorithm A on Bob (and Alice) in
setting S0,0 with no interferences. This dummy personality is then used in setting S0,1. Note that
at the start, only Bob can distinguish S0,1 from S0,0. For the first N/4 times that Alice listens
alone in S0,1, the adversary connects Alice to the dummy gBob0, that is, Alice receives transmis-
sions of gBob0. Thus, up to the point that Alice has listened alone for N/4 rounds, Alice receives
inputs (with distribution) exactly as if she was in setting S0,0 with real Bob and hence, she can not
distinguish this setting from the setting S0,0 with no adversarial interference. After this point, the
adversary lets Alice talk freely with Bob with no interference.
We now explain adversary's strategy for setting S0,0: The adversary generates another dummy
personality gBob1 by simulating algorithm A on Bob (and Alice) in setting S0,1 where the first N/4
listening-alone rounds of Alice were connected to gBob0. That is, exactly the situation that will
happen to real Bob in setting S0,1. For the first N/4 rounds of setting S0,0 where Alice listens, the
adversary does not interfere in the communications. After that, for the next N/4 rounds that Alice
listens, the adversary delivers transmissions of dummy personality gBob1 to Alice.
To conclude the argument, the adversary give a random input y ∈ {0, 1} input to Bob and
gambles on that Alice will be listening alone less than 1/2 of the rounds. The adversary also uses
the dummy personality gBobi for i = 1 − y and when Alice listens alone, the adversary connects
Alice to the real Bob or this dummy personality according to the rules explained above. With
probability at least 1/2, Alice indeed listens in at most N/2 rounds. If this happens, due to the
adversarial errors, the two settings look identical to Alice (more formally, she observes the same
probability distributions for the inputs) and she can not distinguish them from each other. This
means that algorithm A has a failure probability of at least 1/4 (Alice can still guess y but the
guess would be incorrect with probability at at least 1/8).
We finally explain the adversary's rules for treating the rounds where both parties listen: For
setting S0,0, if both Alice and the real Bob are listening, Alice is connected to gBob1 at no cost.
Similarly, in setting S0,1, if both Alice and real Bob are listening, then Alice is connected to gBob0
at no cost. To make sure that this definition does not have a loop, if for a round r, both parties are
listening in both settings, then the adversary delivers a 0 to Alice in both settings. Note that in
using these rules, the behavior of the dummy personalities gBob0 and gBob1 are defined recursively
based on the round number; for example, the simulation that generates the behavior of gBob0 in
29
round r might use the behavior of gBob1 in round r − 1. Because of these rules, we get that in each
round that Alice listens (at least until Alice has had N/2 listening-alone rounds), the messages that
she receives have the same probability distribution in two settings and thus, the two settings look
indistinguishable to Alice. If the execution is such that Alice listens alone less than N/4 rounds,
which happens with probability at least 1/2, the algorithm is bound to fail with probability at least
1/2 in this case. This means algorithm A fails with probability at least 1/4.
Next we give the proof for Lemma 4.2 which shows that no adaptive protocol can robustly
simulates the Exchange Protocol for an error rate of 2/7:
Proof of Lemma 4.2. We first explain the adversaries strategy and then explain why this strategy
makes at least one of the parties unable to know the input of the other party with probability
greater than 1/2.
Suppose that there is an algorithm A that solves the exchange problem under adversarial error-
rate 1/3, in N rounds. We work simply with 1-bit inputs. Let SX,Y denote the setting where Alice
receives input X and Bob gets input Y . For simplicity, we first ignore the rounds in which both
parties listen; note that in those rounds the adversary can deliver arbitrary messages to each of the
parties at no cost.
First we explain adversary's strategy for setting S0,1: Consider setting S0,0 and suppose that
for the first 2N/3 rounds in this, the adversary does not interfere. Without loss of generality, we
can assume that with probability at least 1/2, Alice listens alone in less than N/3 of these rounds.
Adversary creates a dummy personality gBob0 and simulates it with Alice in this S0,0 setting. In
the first 2N/3 rounds of setting S0,1, whenever Alice listens alone, the adversary connects Alice
to gBob0, that is, Alice receives the transmission of gBob0. With probability at least 1/2 regarding
the randomness of Alice and gBob0, Alice listens less than N/3 of the time which means that the
adversary will have enough budget to completely fake Bob as gBob0 (from Alice's viewpoint). In
that case, the two settings look identical to Alice for the first 2N/3 rounds. During the last N/3
rounds of the execution in setting S0,1, the adversary lets Alice and the real Bob talk without no
interference.
We now explain adversary's strategy for setting S0,0: The adversary generates another dummy
personality gBob1 by simulating Bob in setting S0,1 where alone-listenings of Alice in the first 2N/3
rounds are connected to gBob0. In setting S0,0, the adversary lets Alice and Bob talk freely during
adversary connects her to gBob1.
the first 2N/3 rounds but for the last N/3 rounds, whenever Alice listens (even if not alone), the
To conclude, the adversary give a random input y ∈ {0, 1} input to Bob and gambles on that
Alice listens alone less than N/3 rounds of the first 2N/3 rounds. The adversary also uses the
dummy personality gBobi for i = 1 − y and when Alice listens alone, the adversary connects Alice
to the real Bob or this dummy personality according to the rules explained above. We know that
with probability at least 1/2, Alice listens alone less than N/3 rounds of the first 2N/3 rounds.
If that happens, with at most N/3 errors, the adversary can follow the strategy explained above.
Therefore, with probability 1/2, Alice can not know Bob's input and thus will fail with probability
at least 1/4.
Regarding the rounds where both parties are listening, the rule is similar to Lemma D.1 but
a little bit more subtle because of the rate adaptivity of algorithm A: We need to declare what
are the receptions when in the first 2N/3 rounds of setting S0,0, both Alice and Bob are listening.
30
However, at that point, it's not clear whether we will make an indistinguishability argument for
Alice or for Bob, which since it is affected by which one of them listens alone less, it can also
depend on the receptions during the all-listening rounds of the first 2N/3 rounds. The simple
remedy is to create analogous situations for Alice and Bob so that we can determine the base of
indistinguishability later at the end of 2N/3 rounds. Adversary generates dummy personalities
]Alice0 and ]Alice1, respectively in settings S0,0 and S1,0, similar to those of Bob.
In each all-
listening round of the first 2N/3 rounds, adversary makes Alice receive the transmission of gBob1
and Bob receive the transmission of ]Alice1. With these connections, whoever is more likely to listen
alone less in the first 2N/3 rounds -- which we assumed to be Alice without loss of generality in the
above discussions -- with constant probability will receive messages with the exact same probability
distributions, in each round in the two different settings. Thus she will not be able to distinguish
the two settings.
31
|
1911.03360 | 1 | 1911 | 2019-11-08T16:32:02 | Local Search for Group Closeness Maximization on Big Graphs | [
"cs.DS",
"cs.SI"
] | In network analysis and graph mining, closeness centrality is a popular measure to infer the importance of a vertex. Computing closeness efficiently for individual vertices received considerable attention. The NP-hard problem of group closeness maximization, in turn, is more challenging: the objective is to find a vertex group that is central as a whole and state-of-the-art heuristics for it do not scale to very big graphs yet.
In this paper, we present new local search heuristics for group closeness maximization. By using randomized approximation techniques and dynamic data structures, our algorithms are often able to perform locally optimal decisions efficiently. The final result is a group with high (but not optimal) closeness centrality.
We compare our algorithms to the current state-of-the-art greedy heuristic both on weighted and on unweighted real-world graphs. For graphs with hundreds of millions of edges, our local search algorithms take only around ten minutes, while greedy requires more than ten hours. Overall, our new algorithms are between one and two orders of magnitude faster, depending on the desired group size and solution quality. For example, on weighted graphs and $k = 10$, our algorithms yield solutions of $12,4\%$ higher quality, while also being $793,6\times$ faster. For unweighted graphs and $k = 10$, we achieve solutions within $99,4\%$ of the state-of-the-art quality while being $127,8\times$ faster. | cs.DS | cs | Local Search for Group Closeness Maximization
on Big Graphs
Eugenio Angriman, Alexander van der Grinten, Henning Meyerhenke
Department of Computer Science
Humboldt-Universitat zu Berlin, Germany
{angrimae, avdgrinten, meyerhenke}@hu-berlin.de
9
1
0
2
v
o
N
8
]
S
D
.
s
c
[
1
v
0
6
3
3
0
.
1
1
9
1
:
v
i
X
r
a
Abstract -- In network analysis and graph mining, closeness
centrality is a popular measure to infer the importance of a
vertex. Computing closeness efficiently for individual vertices
received considerable attention. The NP-hard problem of group
closeness maximization, in turn, is more challenging: the objective
is to find a vertex group that is central as a whole and state-of-
the-art heuristics for it do not scale to very big graphs yet.
In this paper, we present new local search heuristics for group
closeness maximization. By using randomized approximation
techniques and dynamic data structures, our algorithms are often
able to perform locally optimal decisions efficiently. The final
result is a group with high (but not optimal) closeness centrality.
We compare our algorithms to the current state-of-the-art
greedy heuristic both on weighted and on unweighted real-world
graphs. For graphs with hundreds of millions of edges, our local
search algorithms take only around ten minutes, while greedy
requires more than ten hours. Overall, our new algorithms are
between one and two orders of magnitude faster, depending on
the desired group size and solution quality. For example, on
weighted graphs and k = 10, our algorithms yield solutions
of 12.4% higher quality, while also being 793.6× faster. For
unweighted graphs and k = 10, we achieve solutions within
99.4% of the state-of-the-art quality while being 127.8× faster.
Index Terms -- centrality, group closeness, graph mining, net-
work analysis
I. INTRODUCTION
Identifying important vertices in large networks is one of
the main problems in network analysis [1]. For this purpose,
several centrality measures have been introduced over the
past decades [2]. Among them, one of the most widely-used
measures is closeness [3]. For a given vertex v, it is defined
as the reciprocal of the average shortest-path distance from
v to all other vertices. The problem of identifying the k ver-
tices with highest closeness centrality has received significant
attention [4] -- [6]. In graph mining applications, however, it is
often necessary to determine a group of vertices that is central
as a whole -- which is an orthogonal problem shown to be
NP-hard [7]. One can view group closeness as special case
(on graphs) of the well-known k-median problem for facility
location. Example applications include: (i) retailers that want
to advertise their product via social media; they could select as
promoters the group of k members with highest centrality (≈
influence over the other members) [8]; (ii) in P2P networks,
This work is partially supported by German Research Foundation (DFG)
grant ME 3619/3-2 within Priority Programme 1736 Algorithms for Big Data.
shared resources could be placed on k peers so that they are
easily accessible by others [9]; (iii) in citation networks, group
centrality measures can be employed as alternative indicators
for the influence of journals or papers within their field [10].
Everett and Borgatti [11] formalized the concept of cen-
trality for groups of vertices; the closeness of a group S
is defined as the reciprocal of the average distance from S
to all other vertices of the graph. While exact algorithms to
find a group with maximal group closeness are known (e. g.,
algorithms based on integer linear programming (ILP) [12]),
they do not scale to graphs with more than a few thousand
edges. Hence, in practice, heuristics are used to find groups
with high group closeness on large real-world data sets. For
example, Chen et al. [7] proposed a greedy algorithm that
heuristically computes a group S with high closeness C(S).
To obtain a group of size k, it performs k iterations and adds
in each iteration the vertex with highest marginal contribution
to S.1 It was shown empirically that this greedy algorithm
yields solutions of very high quality (within 97% of the
optimum) [12] -- at least for those small graphs where running
a comparison against an exact algorithm is still feasible within
a few hours. Due to greedy's promising quality, Bergamini et
al. [12] proposed techniques to speed up the algorithm by
pruning, e. g., by exploiting the supermodularity of farness,
i. e., 1/C(S). Pruning yields a significant acceleration (while
retaining the same solution); however, graphs with hundreds
of millions of edges still require several hours to complete.
Indeed, pruning is most effective when the group is already
large. When performing the first addition, however, the greedy
algorithm has to perform one (pruned) single-source shorted-
path (SSSP) computation for each vertex in the graph to
compute its marginal contribution, and this phase scales super-
linearly in the size of the graph.
Our Contribution: We present
two new local search
algorithms for group closeness: the first algorithm, Local-
Swap, requires little time per iteration but can only exchange
vertices locally. Our second algorithm, Grow-Shrink, is able
to perform non-local vertex exchanges, but updating its data
1While this greedy algorithm was claimed to have a bounded approximation
quality (e. g., in [7], [12]), the proof of this bound relied on the assumption that
C is submodular. A recent update [13] to the conference version [12] revealed
that, in fact, C is not submodular. We are not aware of any approximation
algorithm for group closeness that scales to large graphs.
© 2019 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media,
including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers
or lists, or reuse of any copyrighted component of this work in other works.
structures is more involved. The final result of both algorithms
is heuristic, i. e., no approximation guarantee is known. Yet,
each iteration of Local-Swap maximizes (in approximation) a
lower bound on the objective function, while each iteration of
Grow-Shrink locally maximizes the objective function itself.
Despite these favorable properties, the time complexity of a
single iteration of our algorithms matches the time complexity
of a single evaluation of the objective function,
i. e., for
unweighted graphs, it is linear in the size of the graph.
Our experiments show that our best algorithm, extended
Grow-Shrink, finds solutions with a closeness score greater
than 99.4% of the score of a greedy solution, while being
127.8× faster to compute (k = 10). We see this algorithm as
our main contribution for unweighted graphs. When quality
is not a primary concern, our other algorithms can further
accelerate the computation: For example, the non-extended
variant of Grow-Shrink yields solutions for groups of size 10
whose quality is 91.1% compared to the state of the art; in this
case, it is 700.2× faster. The speedup varies between 927.9×
and 43.0× for groups of sizes 5 and 100, respectively.
On weighted graphs, our algorithms even improve both the
quality and the running time performance compared to the
state of the art, returning solutions of 12.4% higher quality at a
speedup of 793.6× (k = 10). Other trade-offs between quality
and performance are possible and discussed in Section III.
II. LOCAL SEARCH FOR GROUP CLOSENESS
Let G = (V, E) be an undirected connected graph. We allow
both unweighted and positively weighted graphs G. Subsets
S ⊆ V are called groups. The farness of any given group S
is defined as:
f (S) =
dist(S, v).
(cid:88)
v∈V \S
Here, dist(S, v) refers to the minimal shortest-path distance
from any s ∈ S to v in G. Furthermore, the (group) closeness
of S is defined as C(S) = V /f (S),
i. e., C(S) is the
reciprocal of the average distance of all vertices in V \ S to
the nearest vertex in S. Recall that determining the group S∗
that maximizes C over all groups S with S ≤ k is known
to be NP-hard [7]; we are not aware of any algorithm with
a bounded approximation ratio.
We consider the problem of improving the group closeness
of a given set S via local search. More precisely, we consider
exchanges of vertices from S and V \ S. Let u be a vertex
in S and v ∈ V \ S. To simplify the presentation, we use
the notation S+v−u := (S \ {u}) ∪ {v} to denote the set that is
constructed by exchanging u and v. We also use the notation
S+v := S∪{v} and S−u := S\{u}, to denote vertex additions
and removals, respectively.
Note that, as our algorithms can only perform vertex
exchanges,
they require the construction of an initial set
S before the algorithms start. To avoid compromising our
algorithms' running times, we cannot afford a superlinear
initialization step. Thus, in all of our local search algorithms,
we simply choose the initial S uniformly at random. For
large graphs, this initialization can be expected to cover the
graph reasonably well. Exploratory experiments revealed that
other obvious initialization techniques (such as selecting the k
vertices with highest degree) did not improve the performance
of the algorithm.
A. Estimating the Quality of Vertex Exchanges
It is known that a simple greedy ascent algorithm yields
results of good quality on real-world graphs [12]. This greedy
algorithm starts with an empty set S and iteratively adds
vertices v ∈ V \ S to S that maximize f (S) − f (S+v).
Depending on the input graph and the value of k, however,
the greedy algorithm might need to evaluate the difference in
f for a substantial number of vertices -- this computation is
rather expensive for large real-world graphs.
The algorithms in this paper aim to improve upon the
running time of the greedy algorithm. We achieve this by
considering only local vertices for v, i. e., vertices that are
already "near" S. It is clear that selecting only local vertices
would decrease the quality of a greedy solution (as the greedy
algorithm does not have the ability to eventually correct
suboptimal choices). However, this is not necessarily true for
our algorithms based on vertex exchanges in Sections II-B
and II-C.
To make our notion of locality more concrete, let BS ⊆ G
be the DAG constructed by running a SSSP algorithm (i. e.,
BFS or Dijkstra's algorithm) from the vertices in S. We
remark that we work with the full SSSP DAG here (and not
a SSSP tree). Here, the vertices of S are all considered as
sources of the SSSP algorithm, i. e., they are at distance zero.
Furthermore, define
∆−(v) := f (S) − f (S+v)
(cid:88)
=
x∈V \S
dist(S, x) − dist(S+v, x).
(1)
To compute the greedy solution, it seems to be necessary to
compute ∆−(v) exactly for a substantial number of vertices
v.2 As discussed above, this seems to be impractical for large
graphs. However, a lower bound for ∆−(v) can be computed
from the shortest path DAG BS. To this end, let Dv be the
set of vertices reachable from v in BS.
Lemma 1. It holds that:
∆−(v) ≥ Dv · dist(S, v).
In the unweighted case, equality holds if v is a neighbor of S.
This lemma can be proven from the definition of BS. A
formal proof can be found in Appendix A.
The bound of Lemma 1 will be used in the two algorithms
in Sections II-B and II-C. Instead of picking vertices that
maximize ∆−, those algorithms pick vertices that maximize
the right-hand side of Lemma 1, i. e., Dv · dist(S, v). The
is more accurate for
bound is local
in the sense that
it
2The techniques of [12] can avoid some of the computations; nevertheless,
many evaluations of ∆−(v) still have to be performed.
2
Algorithm 1 Overview of Local-Swaps Algorithm
1: repeat
2:
3:
4:
5:
6: until previous iteration did not decrease f (S)
approximate Dv for all V \ S
(u, v) ← argmaxu∈S,v∈N (u)\S Dv − Λu
S ← S+v−u
run pruned BFS from v
(cid:46) to recompute f (S)
vertices near S: in particular, the reachability sets of vertices
in (V \ S) ∩ N (S) are larger in G than those in B, as B
does not contain back edges. Unfortunately, computing the
size of Dv exactly for all v still seems to be prohibitively
expensive: indeed, the fastest known algorithm to compute
the size of the transitive closure of a DAG (=(cid:80)Dv) relies
on iterated (Boolean) matrix multiplication (hence, the best
known exact algorithm has a complexity of O(n2.37) [14]).
However, it turns out that randomized algorithms can be used
to approximate the sizes of Dv for all v at the same time.
We employ the randomized algorithm of Cohen [15] for this
task. In multiple iterations, this algorithm samples a random
number for each vertex of the graph G, accumulates in each
vertex v the minimal random number of any vertex reachable
from v, and estimates Dv based on this information.
We remark that since Cohen's algorithm yields an approx-
imation, but not a lower bound for the right-hand side of
Lemma 1, the inequality of the Lemma can be violated in
our algorithms; in particular, it can happen that our algorithms
pick a vertex v such that ∆−(v) < 0. In this case, instead of
decreasing the closeness centrality of our current group, our
algorithms terminate. Nevertheless, our experiments demon-
strate that on real-world instances, a coarse approximation
of the reachability set size is enough for Lemma 1 to yield
useful candidates for vertex exchanges (see Section II-D5 and
Appendix C-B).
B. Local-Swaps Algorithm
Let us first focus on unweighted graphs. To construct a
fast local search algorithm, a straightforward idea is to allow
swaps between vertices in S and their neighbors in V \S. This
procedure can be repeated until no swap can decrease f (S).
Let u ∈ S be a vertex of the group and let v ∈ N (S)\S be one
of its neighbors outside of the group. To determine whether
swapping u and v (i. e., replacing S by S+v−u) is beneficial,
we have to check whether f (S) − f (S+v−u) > 0, i. e., whether
the farness is decreased by the swap. The challenge here is to
find a pair u, v that satisfies this inequality (without checking
all pairs u, v exhaustively) and to compute the difference
f (S) − f (S+v−u) quickly. Note that a crucial ingredient that
allows us to construct an efficient algorithm is that the distance
of S to every vertex x can only change by ±1 when doing a
swap. Hence, we only have to count the numbers of vertices
where the distance changes by −1 and the number of vertices
where it changes by +1.
Our algorithm requires a few auxiliary data structures to
compute f (S)− f (S+v−u). In particular, we store the following:
3
• the distance dist(S, x) from S to all vertices x ∈ V \ S,
• a set λx := {w ∈ S : dist(S, x) = dist(w, x)} for each
x ∈ V \ S that contains all vertices in S that realize the
shortest distance from S to x,
• the value Λw for each w ∈ S, where Λw := {x ∈ V \S :
λx = {w}} is the set of vertices for which the shortest
distance is realized exclusively by w.
Note that the sets λx consume O(kV ) memory in total.
However, since k (cid:28) V , this can be afforded even for large
real-world graphs. In our implementation, we store each λx in
only k bits.
All of those auxiliary data structures can be maintained
dynamically during the entire algorithm with little additional
overhead. More precisely, after a u-v swap is done, v is added
to all λx satisfying dist(v, x) = dist(S, x); for x ∈ V \ S that
satisfy dist(v, x) < dist(S, x), the set {v} replaces λx. u can
be removed from all λx by a linear scan through all x ∈ V \S.
Algorithm 1 states a high-level overview of the algorithm.
In the following, we discuss how to pick a good swap (line 3 of
the pseudocode) and how to update the data structures after a
swap (line 5). The running time of the algorithm is dominated
by the initialization of λx. Thus, it runs in O(kV +E) time
for each update.
1) Choosing a Good Swap: Because it would be too ex-
pensive to compute the exact difference in f for each possible
swap, we find the pair of vertices (u, v) with u ∈ S, v ∈
N (v)\S that maximizes Dv·dist(S, v)−Λu = Dv−Λu.
Note that this value is a lower bound for the decrease of f
after swapping u and v: In particular, Lemma 1 implies that
Dv is a lower bound for the decrease in farness when adding
v to S. Additionally, Λu is an upper bound for the increase
in farness when removing u from S (and hence also for the
increase in farness when removing u from S+v).3 Thus, we
can expect this strategy to yield pairs of vertices that lead to a
decrease of f. To maximize Dv−Λu, for each v ∈ V \S, we
compute the neighbor u ∈ N (v) ∩ S that minimizes Λu (in
O(V +E) time). Afterwards, we can maximize Dv−Λu
by a linear scan over all v ∈ V \ S.
2) Computing the Difference in Farness: Instead of com-
paring distances, it is sufficient to define sets of vertices whose
distance to S is increased or decreased (by 1) by the swap:
u,v := {x ∈ V : dist(S, x) < dist(S+v−u, x)},
H +
u,v := {x ∈ V : dist(S, x) > dist(S+v−u, x)}
H−
As dist(S, x) − dist(S+v−u, x) ∈ {−1, 0, +1}, it holds that:
Lemma 2. f (S) − f (S+v−u) = H−
u,v − H +
u,v
Fortunately, computing H−
u,v is straightforward: this can
be done by running a BFS rooted at v;
the BFS simply
counts those vertices x for which dist(v, x) < dist(S, x).
Hence, to check this condition, we have to store the values
of dist(S, x) for all x ∈ V . We remark that it is not necessary
to run a full BFS: indeed, we can prune the search once
dist(v, x) ≥ dist(S, x) (i. e., if the BFS is about to visit a
3 Note, however, that this bound is trivial if Dv − Λu ≤ 0.
Algorithm 2 Overview of Grow-Shrink Algorithm
1: repeat
2:
3:
4:
5:
6:
7:
8:
9: until previous iteration did not decrease f (S)
approximate Dv for all v ∈ V \ S
v ← argmaxv∈V \S Dv · d(v)
S ← S+v
run pruned BFS from v
u ← argminu∈S
x∈Ru
S ← S−u
run Dijkstra-like algo.
(cid:46) to recompute f (S), d, d(cid:48)
d(cid:48)(x) − d(x)
(cid:46) to recompute f (S), d, d(cid:48)
(cid:80)
u,v ∩ H−
u,v ⊆ Λu \ H−
vertex x satisfying this condition, the search continues without
visiting x). However, as we will see in the following, it makes
sense to sightly relax the pruning condition and only prune the
BFS if dist(v, x) > dist(S, x); this allows us to update our
auxiliary data structures on the fly.
u,v can be computed from Λu with the help of the
H +
u,v ⊆ Λu, as only
auxiliary data structures. We note that H +
vertices x where dist(S, x) is uniquely realized by u (out of all
vertices in the group) can have their distance from S increased
u,v = ∅, we can further restrict this
by the swap. As H +
u,v, but, in general, Λu \ H−
inclusion to H +
u,v
will consist of more vertices than just H +
u,v. More precisely,
u,v ∪ L0
u,v can be partitioned into Λu \ H−
Λu \ H−
u,v,
u,v = H +
u,v := {x ∈ Λu : dist(u, x) = dist(v, x)} consists
where L0
only of vertices whose distance is neither increased nor
decreased by the swap. By construction, L0
u,v are
disjoint. This proves that the following holds:
Lemma 3. H +
u,v − L0
u,v) is completely visited
We note that L0
u,v, the BFS only has to count the
by our BFS. To determine L0
vertices x that satisfy dist(v, x) = dist(S, x) and λx = {u}.
u,v, it has to count the
On the other hand, to determine Λu∩H−
vertices x satisfying dist(v, x) < dist(S, x) and λx = {u}.
C. Grow-Shrink Algorithm
u,v = Λu − Λu ∩ H−
u,v (and also Λu∩H−
u,v and H +
u,v.
The main issue with the swapping algorithm from Sec-
tion II-B is that it can only exchange a vertex u ∈ S with
one of its neighbors v ∈ N (u) \ S. Due to this behavior,
the algorithm might take a long time to converge to a local
optimum. It also makes it hard to escape a local optimum:
indeed, the algorithm will terminate if no swap with a neighbor
can improve the closeness.
Our second algorithm lifts those limitations. It also allows
G to be a weighted graph. In particular,
it allows vertex
exchanges that change the distances from S to the vertices in
V \ S by arbitrary amounts. Computing the exact differences
f (S) − S(S+v−u) for all possible pairs of u and v seems to be
impractical in this setting. Hence, we decompose the vertex
exchange of u and v into two operations: the addition of v to
S and the removal of u from S+v. In particular, we allow the
set S to grow to a size of k + 1 before we shrink the size of
S back to k. Thus, the cardinality constraint S ≤ k is tem-
porarily violated; eventually, the constraint is restored again.
Fortunately, the individual differences f (S) − f (S+v) and
f (S+v)−f (S+v−u) (or bounds for those differences) turn out to
be efficiently computable for all possible u and v, at least in
approximation. We remark, however, that while this technique
does find the vertex that maximizes f (S) − f (S+v) and the
vertex that maximizes f (S+v) − f (S+v−u), it does not neces-
sarily find the pair of vertices maximizing f (S) − f (S+v−u).
Nevertheless, our experiments in Section III demonstrate that
the solution quality of this algorithm is superior to the quality
of the local-swaps algorithm.
maintains the following data structures:
In order to perform these computations, our algorithm
• the distance d(x) of each vertex x /∈ S to S, and a
representative r(x) ∈ S that realizes this distance, i. e., it
holds that dist(S, x) = dist(r(x), x) = d(x),
• the distance d(cid:48)(x) from S\{r(x)} to x and representative
r(cid:48)(x) for this distance (satisfying the analogous equality).
Since the graph is connected, these data structures are well-
defined for all groups S of size S ≥ 1. Furthermore, the
difference between d(cid:48)(x) and d(x) yields exactly the difference
in farness when r(x) is removed from the S. Later, we will
use this fact to quickly determine differences in farness.
We remark that it can happen that d(x) = d(cid:48)(x); nev-
ertheless, r(x) and r(cid:48)(x) are always distinct. Indeed, there
can be two distinct vertices r(x) and r(cid:48)(x) in S that satisfy
dist(r(x), x) = dist(r(cid:48)(x), x) = dist(S, x). With Ru and R(cid:48)
u,
we denote the set of vertices x ∈ V \ S with r(x) = u and
r(cid:48)(x) = u, respectively.
Algorithm 2 gives a high-level overview of the algorithm. In
the following two subsections, we discuss the growing phase
(line 2-5) and the shrinking phase (line 6-8) individually. The
running time of Grow-Shrink is dominated by the Dijkstra-like
algorithm (in line 8). Therefore, it runs in O(V +E log V )
time per update (when using an appropriate priority queue).
The space complexity is O(V + E).
1) Vertex additions: When adding a vertex v to S, we
want to select v such that f (S+v) is minimized. Note that
minimizing f (S+v) is equivalent to maximizing the difference
f (S) − f (S+v) = ∆−(v). Instead of maximizing ∆−(v),
we maximize the lower bound Dv · dist(S, v). We perform
a small number of iterations of the reachability set size
approximation algorithm (see Section II-A) to select the vertex
v with (approximatively) largest Dv.
After v is selected, we perform a BFS from v to compute
∆−(v) exactly. As we only need to visit the vertices whose
distance to S+v is smaller than to S, this BFS can be pruned
once a vertex x is reached with dist(S, x) < dist(v, x). During
the BFS, the values of d, d(cid:48), r, r(cid:48) are updated to reflect the
vertex addition: the only thing that can happen here is that v
realizes either of the new distances d or d(cid:48).
2) Vertex removals: For vertex removals, we can efficiently
calculate the exact increase ∆+(u) := f (S−u) − f (S) in
farness for all vertices u ∈ S, even without relying on
approximation. In fact, ∆+(u) is given as:
(cid:88)
x∈Ru
d(cid:48)(x) − d(x).
∆+(u) =
4
d(cid:48)-boundary
d-boundary
w
u
w(cid:48)
Fig. 1: w, u and w(cid:48) are vertices in S. Vertices within the
solid regions belong to Rw, Ru and Rw(cid:48), respectively. Vertices
within the dashed regions belong to R(cid:48)
w and R(cid:48)
u, respectively.
After removing u from S, the vertices in R(cid:48)
u will have an
invalid r(cid:48) and d(cid:48).
We need to compute k such sums (i. e., ∆+(u) for each u ∈
S); but they can all be computed at the same time by a single
linear scan through all vertices x ∈ V .
It is more challenging, however, to update d, d(cid:48), r and r(cid:48)
after removing a vertex u from S. For vertices x with an
invalid d(x) (i. e., vertices in Ru), we can simply update
d(x) ← d(cid:48)(x) and r(x) ← r(cid:48)(x). This update invalidates
d(cid:48)(x) and r(cid:48)(x). In the following, we treat d(cid:48)(x) as infinite and
r(cid:48)(x) as undefined for all updated vertices x; eventually those
expressions will be restored to valid values using the algorithm
that we describe in the remainder of this section. Indeed, we
now have to handle all vertices with an invalid d(cid:48)(x) (i. e.,
those in Ru ∪ R(cid:48)
u). This computation is more involved. We
run a Dijkstra-like algorithm (even in the unweighted case) to
fix d(cid:48)(x) and r(cid:48)(x). The following definition yields the starting
points for our Dijkstra-like algorithm.
Definition 1. Let x ∈ V be any vertex and let y ∈ N (x) ∩
(Ru ∪ R(cid:48)
u) be a neighbor of x that needs to be updated.
• We call (x, y) a d-boundary pair for y iff r(x) (cid:54)= r(y).
• We call (x, y) a d(cid:48)-boundary pair for y iff r(x) = r(y)
u. In this case, we set b(x, y) := d(cid:48)(x) +
In this case, we set b(x, y) := d(x) + dist(x, y).
and x /∈ Ru ∪ R(cid:48)
dist(x, y).
In both cases, b(x, y) is called the boundary distance of (x, y).
The definition is illustrated in Figure 1. Intuitively, boundary
pairs define the boundary between regions of G that have
a valid d(cid:48)(x) (blue regions in Figure 1) and regions of the
graph that have an invalid d(cid:48)(y) (orange region in Figure 1).
The boundary distance b(x, y) corresponds to the value of d(cid:48)
that a SSSP algorithm could propagate from x to y. We need
to distinguish d-boundary pairs and d(cid:48)-boundary pairs as the
boundary distance can either be propagated on a shortest path
from S over x to y (in case of a d-boundary pair) or on a
shortest path from S−r(x) over x to y (in case of a d(cid:48)-boundary
pair).
Consider all y ∈ V \ S such that there exists at least
one (d- or d(cid:48)-)boundary pair for y. For each such y,
let
(x, y) be the boundary pair with minimal boundary distance
b(x, y). Our algorithm first determines all such y and up-
dates d(cid:48)(y) ← b(x, y). If (x, y) is a d-boundary pair, we set
r(y) ← r(x); for d(cid:48)-boundary pairs, we set r(y) ← r(cid:48)(x).
After this initial update, we run a Dijkstra-like algorithm
starting from these vertices y for which a boundary pair exists.
This algorithm treats d(cid:48) as the distance. Compared to the
usual Dijkstra algorithm, our algorithm needs the following
modifications: For each vertex x, our algorithm only visits
those neighbors y that satisfy r(y) (cid:54)= r(cid:48)(x). Furthermore,
whenever such a visit results in an update of d(cid:48)(y), we
propagate r(cid:48)(y) ← r(cid:48)(x). Note that these conditions imply
that we never update r(cid:48)(y) such that r(cid:48)(y) = r(y).
Lemma 4. After the Dijkstra-like algorithm terminates, d(cid:48) and
r(cid:48) are correct.
A proof of this lemma can be found in Appendix A.
D. Variants and Algorithmic Improvements
1) Semi-local Swaps: One weakness of the algorithm in
Section II-B is that it only performs local vertex exchanges.
In particular, the algorithm always swaps a vertex u ∈ S and
a vertex v ∈ N (u) \ S. This condition can be generalized: in
particular, it is sufficient that u ∈ S also satisfies u ∈ N (S+v).
In this situation, the distances of all vertices can still only
change by a single hop and the proofs of the correctness of
the algorithm remain valid. Note that this naturally partitions
candidates u into two sets: first, the set N (v)∩S of candidates
that the original algorithm considers, and the set N (S) ∩ S.
Candidates in the latter set can be determined independently of
v; indeed, they can be swapped with any v ∈ N (S)\S. Hence,
our swap selection strategy from Section II-B1 continues to
work with little modification.
2) Restricted Swaps: To further improve the performance
of our Local-Swap algorithm at the cost of its solution quality,
we consider the following variant: instead of selecting the pair
of vertices (u, v) that maximize Dv−Λu, we just select the
vertex v that maximizes Dv and then choose u ∈ N (v) ∩ S
such that Λu is minimized. This restricts the choices for u;
hence, we expect this Restricted Local-Swap algorithm to yield
solutions of worse quality. Due to the restriction, however, it
is also expected to converge faster.
3) Local Grow-Shrink: During exploratory experiments, it
turned out that the Grow-Shrink algorithm sometimes over-
estimates the lower bound Dv · dist(S, v) on the decrease
f (S) − f (S+v) of the farness after adding an element v.
This happens because errors in the approximation of Dv are
amplified by multiplying with a large dist(S, v). Hence, we
found that restricting the algorithm's choices for v to vertices
near S improves the solution quality of the algorithm.
It may seem that
this modification makes Grow-Shrink
vulnerable to the same weaknesses as Local-Swap. Namely,
local choices imply that large numbers of exchanges might
be required to reach local optima and it becomes hard to
escape these local optima. Fortunately, additional techniques
discussed in the next section can be used to avoid this problem.
4) Extended Grow-Shrink: Even in the case of Grow-
Shrink, the bound of Lemma 1 becomes worse for vertices
5
at
long distances from the current group. As detailed in
Section II-A, this happens as our reachability set size approx-
imation approach does not take back edges into account. This
is a problem especially on graphs with a large diameter where
we have to expect that many back edges exist. We mitigate this
problem (as well as the problems mentioned in Section II-D3)
by allowing the group to grow by more than one vertex before
we shrink it again. In particular, we allow the group to grow
to size k + h for some h ≥ 1, before we shrink it back to k.
In our experiments in Section III, we consider two strategies
to choose h. First, we consider constant values for h. However,
this is not expected to be appropriate for all graphs: specifi-
cally, we want to take the diameter of the graph into account.
Hence, a more sophisticated strategy selects h = diam(G)/kp
for a fixed p. This strategy is inspired by mesh-like graphs
(e. g., real-world road networks and some other infrastructure
networks): if we divide a quadratic two-dimensional mesh G
√
into k quadratic sub-meshes (where k is a power of 2), the
k. Hence, if we
diameter of the sub-meshes is diam(G)/
assume that each vertex of the group covers an equal amount
of vertices in the remaining graph, h = diam(G)/
k vertex
additions should be sufficient to find at least one good vertex
that will improve a size-k group. As we expect that real-world
networks deviate from ideal two-dimensional meshes to some
extend, we consider not only p = 1
2 but also other values of
p.
√
5) Engineering the reachability set size approximation al-
gorithm: Cohen's reachability set size approximation algo-
rithm [15] has multiple parameters that need to be chosen
appropriately: in particular, there is a choice of probability
distribution (exponential vs. uniform), the estimator function
(averaging vs. selection-based), the number of samples and
the width of each random number. For the estimator, we use
an averaging estimator, as this estimator can be implemented
more efficiently than a selection-based estimator (i. e., it only
requires averaging numbers instead of finding the k-smallest
number). We performed exploratory experiments to determine
a good configuration of the remaining parameters. It turns
out that, while the exponential distribution empirically offers
better accuracy than the uniform distribution, the algorithm
can be implemented much more efficiently using the uniform
distribution: in particular, for the uniform distribution, it is
sufficient to generate and store the per-vertex random numbers
as (unsigned) integers, while the exponential distribution re-
quires floating point calculations. We deal with the decrease in
accuracy by simply gathering more samples. For the uniform
distribution and real-world graphs, 16 bits per integer turns
out to yield sufficient accuracy. In this setting, we found that
16 samples are enough to accurately find the vertex with
highest reachability set size. In particular, while the theoretical
guarantee in [15] requires the number of samples to grow with
log V , we found this number to have a negligible impact
on the solution quality of our group closeness heuristic (see
Appendix C-B).
6) Memory latency in reachability set size approximation:
is well-known that
the empirical performance of graph
It
6
traversal algorithms (like BFS and Dijkstra) is often limited
by memory latency [16], [17]. Unfortunately, the reachability
set size approximation needs to perform multiple traversals of
the same graph. To mitigate this issue, we perform multiple
iterations of the approximation algorithm at the same time.
This technique increases the running time performance of the
algorithm at the cost of its memory footprint. More precisely,
during each traversal of the graph, we store 16 random integers
per vertex and we aggregate all 16 minimal values per vertex
at
the same time. This operation can be performed very
efficiently by utilizing SIMD vector operations. In particular,
we use 256-bit AVX operations of our Intel Xeon CPUs to take
the minimum of all 16 values at the same time. As mentioned
above, aggregating 16 random numbers per vertex is enough
for our use case; thus, using SIMD aggregation, we only need
to perform a single traversal of the graph.
7) Accepting swaps and stopping condition: As detailed in
Sections II-B and II-C, our algorithms stop once they cannot
find another vertex exchange that improves the closeness score
of the current group. Exchanges that worsen the score are not
accepted. To prevent vertex exchanges that change the group
closeness score only negligibly, we also set a limit on the
number of vertex exchanges. In our experiments, we choose
a conservative limit that does not impact the solution quality
measurably (see Appendix C-A).
III. EXPERIMENTS
In this section, we evaluate the performance of our al-
gorithms against
the state-of-the-art greedy algorithm of
Bergamini et al. [12].4 As mentioned in Section I, it has
been shown empirically that the solution quality yielded by
the greedy algorithm is often nearly-optimal. We evaluate two
variants, LS and LS-restrict (see Section II-D2), of our Local-
Swap algorithm, and three variants, GS, GS-local (see Sec-
tion II-D3) and GS-extended (see Section II-D4) of our Grow-
Shrink algorithm. We evaluate these algorithms for group sizes
of k ∈ {5, 10, 20, 50, 100} on the largest connected component
of the input graphs. We measure the performance in terms
of running time and closeness of the group computed by the
algorithms. Because our algorithms construct an initial group
S by selecting k vertices uniformly at random (see Section II),
we average the results of five runs, each one with a different
random seed, using the geometric mean over speedup and
relative closeness.5 Unless stated otherwise, our experiments
are based on the graphs listed in Tables II and III. They are
all undirected and have been downloaded from the public
repositories 9th DIMACS Challenge [18] and KONECT [19].
The running time of the greedy baseline varies between 10
minutes and 2 hours on those instances.
Our algorithms are implemented in the NetworKit [20] C++
framework and use PCG32 [21] to generate random numbers.
4In our experiments we do not consider the naive greedy algorithm and the
OSA heuristic of [7] because they are both dominated by [12].
5These five runs are done to average out particularly bad (or good)
selections of initial groups; as one can see from Appendix C-B, the variance
due to the randomized reachability set size algorithm is negligible.
(a) Speedup over the greedy
algorithm (geom. mean).
(b) Closeness score relative
to the score of
the group
returned by greedy (geom.
mean).
Fig. 2: Performance of the extended Grow-Shrink algorithm
for different values of h or p on unweighted graphs, k = 10.
All experiments were managed by the SimexPal software to
ensure reproducibility [22]. Experiments were executed with
sequential code on a Linux machine with an Intel Xeon Gold
6154 CPU and 1.5 TiB of memory.
A. Results for Extended Grow-Shrink
In a first experiment, we evaluate performance of our
extended Grow-Shrink algorithm and compare it to the greedy
heuristic. Because of its ability to escape local optima, we
expect this to be the best algorithm in terms of quality; hence,
it should be a good default choice among our algorithms. For
this experiment, we set k = 10.
As discussed in Section II-D4, we distinguish two strategies
to determine h: we either fix a constant h, or we fix a constant
p. For each strategy, we evaluate multiple values for h or p.
Results for both strategies are shown in Figure 2. As expected,
higher values of h (or, similarly, lower values of p) increase the
algorithm's running time; (while h > 1 allows the algorithm to
perform better choices, it does not converge h-times as fast).
Still, for all tested values of h or p, the extended Grow-Shrink
algorithm is one to two orders of magnitude faster than the
greedy baseline. Furthermore, values of p < 1 yield results
of very good quality: for p = 0.75, for example, we achieve
a quality of 99.4%. At the same time, using this setting for
p, our algorithm is 127.8× faster than the greedy algorithm.
We remark that for all but the smallest values of h (i. e., those
corresponding to the lowest quality), choosing constant p is a
better strategy than choosing constant h: for the same running
time, constant p always achieves solutions of higher quality.
B. Scalability to Large Graphs
We also analyze the running time of our extended Grow-
Shrink algorithm on large-scale networks. To this end, we
switch to graphs larger than the ones in Table II. We fix
p = 0.75, as Section III-A demonstrated that this setting
results in a favorable trade-off between solution quality and
running time. The greedy algorithm is not included in this
experiment as it requires multiple hours of running time, even
(a) R-MAT networks; 217 to
224 vertices (up to 268 million
edges).
(b) Random hyperbolic net-
works; 217 to 226 vertices (up
to 671 million edges).
Fig. 3: Running time (s) of the extended Grow-Shrink al-
gorithm on synthetic graphs (black line = linear regression);
p = 0.75, k = 10.
TABLE I: Running time of the extended Grow-Shrink algo-
rithm on large real-world networks; p = 0.75, k = 10.
Time (s)
95.3
135.6
199.9
368.0
333.1
680.1
Network
soc-LiveJournal1
livejournal-links
orkut-links
dbpedia-link
dimacs10-uk-2002
wikipedia link en
E
42,845,684
48,687,945
117,184,899
126,888,089
261,556,721
334,640,259
V
4,843,953
5,189,808
3,072,441
18,265,512
18,459,128
13,591,759
for the smallest real-world graphs that we consider in this part.
Hence, we also do not compare against its solution quality in
this experiment.
1) Results on Synthetic Data: Figure 3 shows the average
running time of our algorithm on randomly generated R-
MAT [23] graphs as well as graphs from a generator [24] for
random hyperbolic graphs. Like R-MAT, the random hyper-
bolic model yields graphs with a skewed degree distribution,
similar to the one found in real-world complex networks. In the
(log-log) plot, the straight lines represent a linear regression of
the running times. In both cases, the running time curves are
at most as steep as the regression line, i. e., the running time
behaves linearly in the number of vertices for the considered
network models and sizes.
2) Results on Large Real-World Data Sets: Table I reports
the algorithm's performance on large real-world graphs. In
contrast to the greedy algorithm (which would require hours),
our extended Grow-Shrink algorithm can handle real-world
graphs with hundreds of millions of edges in a few minutes.
For the orkut-links network, Bergamini et al. [12] report
running times for greedy of 16 hours on their machine; it
is the largest instance in their experiments.
C. Accelerating Performance on Unweighted Graphs
While the extended Grow-Shrink algorithm yields results
of very high quality, if quality is not a primary concern, even
faster algorithms might be desirable for very large graphs.
To this end, we also evaluate the performance of the non-
extended Grow-Shrink and the Local Swap algorithms. For
extended Grow-Shrink, we fix p = 0.75 again. The speedup
and the quality of our algorithms over the greedy algorithm,
for different values of the group size k, is shown in Figures 4a
7
const hconst p151050h0100200300400Speedup over greedy2.01.751.51.251.00.750.5p0.900.920.940.960.981.00Relative closeness151050h02.01.751.51.251.00.750.5p1718192021222324Scale100101102103Time (s)GS-extended17181920212223242526Scale100101102103Time (s)GS-extended Grow-Shrink on weighted graphs is shown in Figure 5. In
contrast to unweighted graphs, the quality of the non-local
Grow-Shrink algorithm is superior to the greedy baseline for
all considered group sizes. Furthermore, in contrast to the
unweighted case, the ability to perform non-local vertex ex-
changes greatly benefits the non-local Grow-Shrink algorithm
compared to local Grow-Shrink.6 Thus, Grow-Shrink clearly
dominates both the greedy algorithm and local Grow-Shrink
on the weighted graphs in our benchmark set -- both in terms
of speed and solution quality.
E. Summary of Experimental Results
On unweighted graphs, a good trade-off between solution
quality and running time is achieved by the extended Grow-
Shrink algorithm with constant p = 0.75. This strategy yields
solutions with at least 99.4% of the closeness score of a greedy
solution (greedy, in turn, was at most 3% away from the
optimum on small networks in previous work [12]). Extended
Grow-Shrink is 127.8× faster than greedy (k = 10). Thus, it
is able to handle graphs with hundreds of millions of edges in
a few minutes -- the state of the art needs multiple hours. If a
fast but inaccurate algorithm is needed for exploratory analysis
of graph data sets, we recommend to run the non-extended
Grow-Shrink algorithm, or, if only a very coarse estimate of
the group with maximal closeness is needed, restricted Local-
Swap.
On weighted graphs, we recommend to always use our
Grow-Shrink algorithm, as it outperforms the greedy state of
the art both in terms of quality (yielding solution that are on
average 12.4% better than greedy solutions) and in terms of
running time performance (with a speedup of two orders of
magnitude), at the same time.
IV. CONCLUSIONS
In this paper, we introduced two families of new local-
search algorithms for group closeness maximization in large
networks. As maximizing group closeness exactly is infea-
sible for graphs with more than a few thousand edges, our
algorithms are heuristics (just like the state-of-the-art greedy
algorithm). However, for small real-world networks, the results
are empirically known to be close to optimal solutions [12].
Compared to previous state-of-the-art heuristics, our algo-
rithms (in particular: extended Grow-Shrink) allow to find
groups with high closeness centrality in real-world networks
with hundreds of millions of edges in seconds to minutes
instead of multiple hours, while sacrificing less than 1% in
quality. In weighted graphs, Grow-Shrink (GS) even dominates
the best known heuristic: the GS solution quality is more
than 10% higher and GS is two orders of magnitude faster.
Adapting the algorithm to even larger graphs in distributed
memory is left to future work.
(a) Speedup over the greedy
algorithm (geom. mean).
(b) Closeness score relative to
the score of the group returned
by greedy (geom. mean).
Fig. 4: Performance of our local search algorithms for different
values of k; unweighted graphs.
(a) Speedup over the greedy
algorithm (geom. mean).
(b) Closeness score relative to
the score of the group returned
by greedy (geom. mean).
Fig. 5: Performance of our local search algorithms for different
values of k; weighted graphs.
and 4b, respectively. Note that the greedy algorithm scales well
to large k, so that the speedup of our algorithms decreases with
k (as mentioned in Section I, the main bottleneck of greedy is
adding the first vertex to the group). However, even for large
groups of k = 100, all of our algorithms are still at least 43.0×
faster.
After extended Grow-Shrink, our non-extended local ver-
sion of Grow-Shrink is the next best algorithm in terms of
quality. As explained in Section II-D3, this variant yields better
solutions than non-local Grow-Shrink and gives a speedup of
3.1× over extended Grow-Shrink with p = 0.75 and k = 10
(= a speedup of 365.8× over greedy); the solution quality in
this case is 92.1% of the greedy quality.
The non-restricted Local-Swap algorithm is dominated by
Grow-Shrink, both regarding running time and solution qual-
ity. Furthermore, compared to the other algorithms, the re-
stricted Local-Swap algorithm only gives a rough estimate
of the group with highest group closeness; however, it is
also significantly faster than all other algorithms and may be
employed during an exploratory analysis of graph data sets.
D. Results on Weighted Road Networks
Recall that the Local-Swaps algorithm does not work for
weighted graphs; we thus report only Grow-Shrink data in the
weighted case. The performance of Grow-Shrink and local
6For this reason, we do not include the extended Grow-Shrink algorithm in
this experiment. In fact, we expect that it improves only slightly on GS-local
(red line/bars in Fig. 5) but cannot compete with (non-local) GS: the ability to
perform non-local vertex exchanges, as done by GS (green line/bars in Fig. 5)
appears to be crucial to obtain high-quality results on weighted graphs.
8
LS-restrictLSGSGS-localGS-extended5102050100Group size050010001500Speedup over greedy0.70.80.91.0Relative closeness5102050100Group size0GSGS-local5102050100Group size02505007501000Speedup over greedy0.91.01.1Relative closeness5102050100Group size0 REFERENCES
[1] M. Newman, Networks, 2nd ed. Oxford university press, 2018.
[2] P. Boldi and S. Vigna, "Axioms for centrality," Internet Mathematics,
vol. 10, no. 3-4, pp. 222 -- 262, 2014.
[3] A. Bavelas, "A mathematical model for group structures," Human
organization, vol. 7, no. 3, p. 16, 1948.
[4] E. Bergamini, M. Borassi, P. Crescenzi, A. Marino, and H. Meyerhenke,
"Computing top-k closeness centrality faster in unweighted graphs,"
ACM Transactions on Knowledge Discovery from Data (TKDD), vol. 13,
no. 5, p. 53, 2019.
[5] K. Okamoto, W. Chen, and X.-Y. Li, "Ranking of closeness centrality
for large-scale social networks," in International Workshop on Frontiers
in Algorithmics. Springer, 2008, pp. 186 -- 195.
[6] P. W. Olsen, A. G. Labouseur, and J.-H. Hwang, "Efficient
top-k
closeness centrality search," in 2014 IEEE 30th International Conference
on Data Engineering.
IEEE, 2014, pp. 196 -- 207.
[7] C. Chen, W. Wang, and X. Wang, "Efficient maximum closeness
centrality group identification," in Australasian Database Conference.
Springer, 2016, pp. 43 -- 55.
[8] T. Zhu, B. Wang, B. Wu, and C. Zhu, "Maximizing the spread of
influence ranking in social networks," Information Sciences, vol. 278,
pp. 535 -- 544, 2014.
[9] C. Gkantsidis, M. Mihail, and A. Saberi, "Random walks in peer-to-peer
networks," in IEEE INFOCOM 2004, vol. 1.
IEEE, 2004.
[10] L. Leydesdorff, "Betweenness centrality as an indicator of the interdis-
ciplinarity of scientific journals," Journal of the American Society for
Information Science and Technology, vol. 58, no. 9, pp. 1303 -- 1319,
2007.
[11] M. G. Everett and S. P. Borgatti, "The centrality of groups and classes,"
The Journal of mathematical sociology, vol. 23, no. 3, pp. 181 -- 201,
1999.
[12] E. Bergamini, T. Gonser, and H. Meyerhenke, "Scaling up group close-
ness maximization," in 2018 Proceedings of the Twentieth Workshop on
Algorithm Engineering and Experiments (ALENEX). SIAM, 2018, pp.
209 -- 222.
[13] -- -- , "Scaling up group closeness maximization," CoRR, vol.
abs/1710.01144, 2017, updated on May 15, 2019. [Online]. Available:
http://arxiv.org/abs/1710.01144
[14] F. Le Gall, "Powers of tensors and fast matrix multiplication," in
the 39th international symposium on symbolic and
Proceedings of
algebraic computation. ACM, 2014, pp. 296 -- 303.
[15] E. Cohen, "Size-estimation framework with applications to transitive
closure and reachability," Journal of Computer and System Sciences,
vol. 55, no. 3, pp. 441 -- 453, 1997.
[16] D. A. Bader, G. Cong, and J. Feo, "On the architectural requirements for
efficient execution of graph algorithms," in Parallel Processing, 2005.
ICPP 2005. International Conference on.
IEEE, 2005, pp. 547 -- 556.
[17] A. Lumsdaine, D. Gregor, B. Hendrickson, and J. Berry, "Challenges in
parallel graph processing," Parallel Processing Letters, vol. 17, no. 01,
pp. 5 -- 20, 2007.
[18] C. Demetrescu, A. V. Goldberg, and D. S. Johnson, The Shortest
Path Problem: Ninth DIMACS Implementation Challenge. American
Mathematical Soc., 2009, vol. 74.
[19] J. Kunegis, "Konect: the Koblenz Network Collection," in Proceedings
of the 22nd International Conference on World Wide Web. ACM, 2013,
pp. 1343 -- 1350.
[20] C. L. Staudt, A. Sazonovs, and H. Meyerhenke, "Networkit: A tool
suite for large-scale complex network analysis," Network Science, vol. 4,
no. 4, pp. 508 -- 530, 2016.
[21] M. E. O'Neill, "Pcg: A family of simple fast space-efficient statistically
good algorithms for random number generation," Harvey Mudd College,
Claremont, CA, Tech. Rep. HMC-CS-2014-0905, Sep. 2014.
[22] E. Angriman, A. v. d. Grinten, M. v. Looz, H. Meyerhenke,
M. Nollenburg, M. Predari, and C. Tzovas, "Guidelines for experimental
algorithmics: A case study in network analysis," Algorithms, vol. 12,
no. 7, p. 127, 2019.
[23] D. Chakrabarti, Y. Zhan, and C. Faloutsos, "R-mat: A recursive model
the 2004 SIAM International
for graph mining," in Proceedings of
Conference on Data Mining. SIAM, 2004, pp. 442 -- 446.
[24] M. von Looz, M. S. Ozdayi, S. Laue, and H. Meyerhenke, "Generating
massive complex networks with hyperbolic geometry faster in prac-
tice," in 2016 IEEE High Performance Extreme Computing Conference
(HPEC).
IEEE, 2016, pp. 1 -- 6.
[25] R. C. Murphy, K. B. Wheeler, B. W. Barrett, and J. A. Ang, "Introducing
the graph 500," Cray Users Group (CUG), vol. 19, pp. 45 -- 74, 2010.
APPENDIX A
TECHNICAL PROOFS
Proof of Lemma 1. Let x ∈ Dv. Because of the sub path op-
timality property of shortest paths, it is clear that dist(S, x) =
dist(S, v)+dist(v, x) (as v is a predecessor of x on a shortest
path from S). On the other hand, adding v to S decreases
the length of this path (as the distance between S and v
becomes zero); in other words: dist(S+v, x) = dist(v, x).
These observations allow us to express the right-hand side
of Eq. 1 as dist(S, x) − dist(S+v, x) = dist(S, v). Sum-
ming this equation for all vertices in Dv yields the term
Dv · dist(S, v) of the lemma. For vertices x /∈ Dv, it holds
that dist(S, x) − dist(S+v, x) ≥ 0, hence the inequality.
w∈S(Z ∩ R(cid:48)
prove the lemma, it is sufficient to prove that Z =(cid:83)
Let us first prove that (cid:83)
For the statement about the unweighted case, we need to
show that the contribution of all other vertices is zero, i. e.,
dist(S, x) = dist(S+v, x) for all vertices x /∈ Dv. Note that
dist(S, x) < dist(S, v) + dist(v, x) (otherwise x would be in
Dv) and dist(S, v) = 1. Thus, dist(S, x) ≤ dist(v, x) which
completes the proof.
Proof of Lemma 4. Let Z be the set of vertices that need to
be updated by the algorithm, i. e., Z equals the set Ru ∪ R(cid:48)
u
before the Dijkstra-like algorithm runs. We have not shown yet
whether all vertices in Z are indeed updated. For the remainder
of this proof, all symbols (such as r(cid:48), d(cid:48) and R(cid:48)) refer to the
state of our data structures after the algorithm terminates. To
w∈S(Z ∩
w) (i. e., that no r(cid:48) remains undefined, or, in other words,
R(cid:48)
r(cid:48) is updated wherever necessary) and that dist(S−r(x), x) =
dist(r(cid:48)(x), x) for all x ∈ Z (i. e., that the definition of r(cid:48) is
respected).
w) = Z. Let z ∈ Z.
There exists a path from every w ∈ S to z and each such
path contained at least one boundary pair (x, y) before the
algorithm started. Indeed, there is a boundary pair for the first
vertex y on that path that is also in Z. Thus, the algorithm
sets r(cid:48)(y) = w for some w ∈ S (i. e., y ∈ Z ∩ R(cid:48)
w) and
propagates the update of r(cid:48) along the path from w to z. We
have to prove that our pruning condition does not prevent any
necessary update along this path. Hence, let (x, y) be a pair
of vertices so that the algorithm is pruned before visiting y
from x. Only the y ∈ Z case is interesting, as r(cid:48) must already
be correct otherwise. Pruning only happens if r(cid:48)(x) = r(y)
and therefore r(x) (cid:54)= r(y). But in this case, (x, y) was a d-
boundary pair and the preceding argument shows that y ∈
w∈S Z ∩ R(cid:48)
w.
Now consider the second part of the proof. Let w ∈ S
be any group vertex and let y ∈ Z ∩ Rw be any vertex that
is updated by the algorithm with r(y) = w. The algorithm
guarantees that dist(S−w, y) ≤ dist(r(cid:48)(y), y) ≤ d(cid:48)(y) as
r(cid:48)(y) ∈ S−w and (by construction of the algorithm) d(cid:48) is the
length of a (not yet proven to be shortest) path from r(cid:48)(y)
to y. It is sufficient to show that this path is a shortest one,
i. e., dist(S−w, y) = d(cid:48)(y). We prove this statement for all
(cid:83)
9
y ∈ Rw by an inductive argument using dist(S−w, y). We
distinguish two cases depending on whether there exists a
neighbor of y in Rw that is on a shortest path from S−w
to y. First, we handle the case that no such neighbor exists.
In this case, r(x) (cid:54)= w holds for all x ∈ N (y) on shortest
paths from S−w to y. As r(x) did not change during the
algorithm, all such x correspond to d-boundary pairs for
y and dist(S−w, y) is the minimal boundary distance over
these pairs (x, y). Hence, d(cid:48)(y) was updated correctly
all
before the Dijkstra-like algorithm ran. On the other hand, let
x ∈ N (y) ∩ Rw be a neighbor of y that is on a shortest
path from S−w to y. x ∈ Rw implies r(cid:48)(x) (cid:54)= w; thus, the
algorithm cannot be pruned when visiting y from x. In this
case, however, the algorithm sets d(cid:48)(y) = d(cid:48)(x) + dist(x, y).
As dist(S−w, x) < dist(S−w, y), the induction yields that
d(cid:48)(x) is already correct, i. e., d(cid:48)(x) = dist(S−w, x). Since
x is on a shortest path from S−w to y, d(cid:48)(y) is also updated
correctly.
APPENDIX B
DETAILS OF THE EXPERIMENTAL SETUP
TABLE II: Unweighted real-world networks used in the ex-
periments.
Network
dimacs9-NY
dimacs9-BAY
web-Stanford
hyves
youtube-links
com-youtube
web-Google
trec-wt10g
dimacs10-eu-2005
soc-pokec-relationships
wikipedia link ca
E
365,050
397,415
1,941,926
2,777,419
2,987,468
2,987,624
4,291,352
6,225,033
16,138,468
22,301,964
27,133,794
V
264,346
321,270
255,265
1,402,673
1,134,885
1,134,890
855,802
1,458,316
862,664
1,632,803
926,588
Category
Road
Road
Hyperlink
Social
Social
Social
Hyperlink
Hyperlink
Road
Social
Hyperlink
TABLE III: Weighted networks used in the experiments. All
networks are road networks of different states of the US.
State
DC
HI
AK
DE
RI
CT
ME
V
9,522
21,774
48,560
48,812
51,642
152,036
187,315
E
14,807
26,007
55,014
59,502
66,650
184,393
206,176
State
ND
SD
WY
ID
MD
WV
NE
V
203,583
206,998
243,545
265,552
264,378
292,557
304,335
E
249,809
249,828
293,825
310,684
312,977
320,708
380,004
Tables II and III show details about our real-world in-
stances. To generate the synthetic graphs in Figure 3, we
use the same parameter setting as in the Graph 500's bench-
mark [25] (i. e., edge factor 16, a = 0.57, b = 0.19, c = 0.19,
and d = 0.05) for the R-MAT generator. For the random
hyperbolic generator, we set the average degree to 20, and
the exponent of the power-law distribution to 3.
10
(a) Speedup over the greedy
algorithm (geom. mean).
(b) Closeness score relative to
the score of the group returned
by greedy (geom. mean).
Fig. 6: Performance of the Grow-Shrink algorithm for different
numbers of samples to estimate reachability set size; k = 10.
(a) Unweighted graphs.
(b) Weighted graphs.
Fig. 7: Behavior of the relative closeness score (compared to
the group returned by greedy, geom. mean) over the execution
of the algorithms (in terms of vertex exchanges); k = 10.
APPENDIX C
ADDITIONAL EXPERIMENTS
A. Impact of the number of vertex exchanges
Figures 7a and 7b depict the relative closeness (compared to
the closeness of the group returned by the greedy algorithm),
depending on the progress of the algorithm in terms of vertex
exchanges. For extended Grow-Shrink, we fix p = 0.75. All
of the local search algorithm quickly converge to a value
near their final result; additional vertex exchanges improve
the group closeness score by small amounts. In order to avoid
an excessive amount of iterations, it seems reasonable to set a
limit on the number of vertex exchanges. In our experiments
we set a conservative limit of 100 exchanges.
B. Impact of reachability set size approximation
As mentioned in Section II-D3, the errors in the approxima-
tion of Dv are amplified by the multiplication with dist(S, v),
and this results in GS-local computing higher quality solutions
than GS. We study how increasing the accuracy of the reach-
ability set size approximation by incrementing the number of
samples impacts the performances of both GS and GS-local.
Figure 6a shows that GS needs at least 64 samples to converge
to a better local optimum than GS-local. However, in both
cases increasing the number of samples degrades the speedup
without yielding a significant quality improvement (Figure 6b).
GSGS-local163264128256Number of samples0200400600800Speedup over greedy0.900.92Relative closeness163264128256Number of samples0LS-restrictLSGSGS-localGS-extended0.60.81.0Relative closeness02468101214Number of exchanges00.81.0Relative closeness01234567891011Number of exchanges0 |
1708.03228 | 1 | 1708 | 2017-08-10T14:13:45 | Lower bounds for several online variants of bin packing | [
"cs.DS",
"cs.DM",
"math.CO",
"math.OC"
] | We consider several previously studied online variants of bin packing and prove new and improved lower bounds on the asymptotic competitive ratios for them. For that, we use a method of fully adaptive constructions. In particular, we improve the lower bound for the asymptotic competitive ratio of online square packing significantly, raising it from roughly 1.68 to above 1.75. | cs.DS | cs |
Lower bounds for several online variants of bin packing∗
J´anos Balogh †
J´ozsef B´ek´esi ‡
Gyorgy D´osa§
Leah Epstein¶
Asaf Levink
Abstract
We consider several previously studied online variants of bin packing and prove new and
improved lower bounds on the asymptotic competitive ratios for them. For that, we use a
method of fully adaptive constructions.
In particular, we improve the lower bound for the
asymptotic competitive ratio of online square packing significantly, raising it from roughly 1.68
to above 1.75.
1
Introduction
In bin packing problems, there is an input consisting of a set of items, and the goal is to partition
it into a minimum number of subsets called bins, under certain conditions and constraints. In the
classic variant [28, 19, 20, 31, 24], items have one-dimensional rational numbers in (0, 1], called sizes,
associated with them, and the total size of items of one bin cannot exceed 1. In online variants
items are presented as a sequence and the partition is created throughout this process in the sense
that any new item should be assigned to a bin before any information regarding the next item is
provided. The conditions on the partition or packing remain as in the offline problem where the
items are all given at once as a set. Using an algorithm A to partition the items into subsets, which
is also seen as a process of packing items into bins, the number of partitions or bins used for the
packing is defined to be the cost of A.
Algorithms for bin packing problems are normally studied using the asymptotic approximation
ratio, also called asymptotic competitive ratio for the case of online algorithms (and we will use
this last term). For an algorithm A and an input I, let A(I) denote the number of bins used by A
for I, that is, the cost of A for I. Let OP T (I) denote the number of bins that an optimal solution
uses for I, that is, the cost of an optimal (offline) algorithm OP T for I. Consider the set of inputs
JQ of all inputs for which the number of bins used by OP T is Q. For the problems studied here
(and non-empty inputs for them), Q will be a positive integer. Let c(Q) = maxI∈JQ A(I) (where for
c(Q)
Q . The absolute competitive
reasonable algorithms this value is finite), and let RA = lim supQ→∞
A(I)
ratio of A is defined by supI
OP T (I) , that is, this is the supremum ratio between the cost of A and
the optimal cost, over all inputs, and the asymptotic competitive ratio is the superior limit of the
∗Gy. D´osa was supported by VKSZ 12-1-2013-0088 "Development of cloud based smart IT solutions by IBM
Hungary in cooperation with the University of Pannonia" and by National Research, Development and Innovation
Office -- NKFIH under the grant SNN 116095. L. Epstein and A. Levin were partially supported by a grant from
GIF - the German-Israeli Foundation for Scientific Research and Development (grant number I-1366-407.6/2016).
†Department of Applied Informatics, Gyula Juh´asz Faculty of Education, University of Szeged, Hungary.
[email protected]
‡Department of Applied Informatics, Gyula Juh´asz Faculty of Education, University of Szeged, Hungary.
[email protected]
§Department of Mathematics, University of Pannonia, Veszprem, Hungary, [email protected].
¶Department of Mathematics, University of Haifa, Haifa, Israel. [email protected].
kFaculty of Industrial Engineering and Management, The Technion, Haifa, Israel. [email protected].
1
absolute competitive ratios for fixed values of Q = OP T (I) when Q grows to infinity. Since the
standard measures for online bin packing problems (and offline bin packing problems, respectively),
are the asymptotic competitive ratio (and the asymptotic approximation ratio), we also use the
terms competitive ratio (and approximation ratio) for them, and always use the word absolute when
we discuss the absolute measures. To prove lower bounds on the (asymptotic) competitive ratio
one can use inputs where the optimal cost is arbitrarily large, and we use this method. The study
of lower bounds on the competitive ratio for a given problem characterizes the extent to which the
performance of the system deteriorates due to lack of information regarding the future input items.
Here, we study three versions of the online bin packing problem, providing new lower bounds
on the competitive ratio for them. Previous constructions used for proving such lower bounds were
often inputs where items arrive in batches, such that the items of one batch all have the exact same
size (and the input may stop after a certain batch or it can continue to another one). In the known
lower bounds for classic bin packing, it is even known what the next batches will be, if they are
indeed presented [23, 29, 6]. While it may be obvious that adaptive inputs where the properties of
the next item are based on the packing of previous items are harder for an algorithm to deal with,
it was not known until recently how to use this idea for designing lower bounds, except for special
cases [9, 2, 17]. In cardinality constrained bin packing [22, 21, 12, 2, 8], items are one-dimensional, a
fixed integer t ≥ 2 is given, and the two requirements for a packed bin are that its total size of items
is at most 1, and that it contains at most t items. The special case analyzed in the past [9, 2, 17] is
t = 2, which can also be seen as a matching problem, as every bin can contain at most two items.
In [4] we showed that the overall competitive ratio (supremum over all values of t) is 2 (an upper
bound was known prior to that work [2, 8]), and provided improved lower bounds for relatively
small values of t. For standard bin packing, the best known lower bound on the competitive ratio
is 1.5403 [29, 6] and the best upper bound is 1.57829 [5].
Another lower bound presented in [4] is for the competitive ratio of vector packing in at least
two dimensions. For an integer dimension d ≥ 2, the items have d-dimensional vectors associated
with them, whose components are rational numbers in [0, 1] (none of which are all-zero vectors),
and bins are all-one vectors of dimension d. A subset of items can be packed into a bin if taking
no component exceeds 1 in their vector sum. This generalizes cardinality constrained bin packing,
and we showed a lower bound of 2.03731129 on the competitive ratio of the online variant for any
d ≥ 2 (prior to that work, no lower bound strictly above 2 for a constant dimension was known).
Our main goal here is to exhibit how to exploit adaptive constructions with some connection
to those used in [4] in order to obtain lower bounds for other variants. We focus on the following
three variants. In all three variants of online bin packing which we study, the input consists of
rational numbers in (0, 1], however there is additional information received with the input in some
of the cases and the input is interpreted in different ways. Two of the problems are one-dimensional
and the input numbers are sizes of items. The third variant is two-dimensional, and the numbers
are side lengths of squares. In our first variant called bin packing with known optimal cost, the
cost of an optimal (offline) solution is given in advance, that is, it is known how many bins are
required for packing the input. This problem is also called K-O (known-OPT). It is currently hard
to find an appropriate way to use this additional piece of information for algorithm design, but
in all lower bounds known for standard online bin packing [29, 6] the property that the optimal
cost is different for different inputs is crucial for achieving the result. For K-O, a lower bound of
1.30556 on the competitive ratio was presented [14] and later improved to 1.32312 [3]. We show
a new lower bound of 87
62 ≈ 1.4032258 on the competitive ratio, improving the previous result
significantly. This problem is related to the field of semi-online algorithms and to the so-called
model of online algorithms with advice [10, 1], where the online algorithm is provided with some
(preferably very small) pieces of information regarding the input.
In the square packing (SP) problem, the goal is to assign an input set of squares whose sides
2
are rational numbers in (0, 1] into bins that are unit squares in a non-overlapping and axis-parallel
way, so as to minimize the number of non-empty bins. We use the standard definition of this
packing problem, where two squares do not overlap if their interiors do not overlap (but they may
have common points on the boundaries of the squares). The offline variant is well-studied [7, 15].
The history of lower bounds on the competitive ratio of online algorithms for this problem is as
follows. Several such lower bounds were proved for the online version of SP, starting with a simple
construction yielding a lower bound of 4
3 on the competitive ratio by Coppersmith and Raghavan
[11], and then there were several improvements [25, 16, 18], all showing bounds above 1.6. In 2016
a copy of the thesis of Blitz [9] from 1996 was found by the authors of [18]. This thesis contains
a number of lower bounds for bin packing problems, including a lower bound of 1.680783 on the
competitive ratio of online algorithms for SP. The result of Blitz [9] is now the previous best lower
bound on the competitive ratio for the problem (prior to our work), and it is higher than the lower
bounds of [25, 16, 18]. Here, we show a much higher lower bound, larger than 1.7515445, on the
competitive ratio of this problem.
Finally, we consider class constrained bin packing (CLCBP) [27, 26, 30, 13].
In this one-
dimensional variant every item has a size and a color, and for a given parameter t ≥ 1, any bin
can receive items of at most t different colors (of total size at most 1), while the number of items
of each color can be arbitrary. This problem generalizes standard bin packing, as for any input
of standard bin packing, defining a common color to all items results in an instance of CLCBP
for any t. It also generalizes bin packing with cardinality constraints, though here to obtain an
instance of CLCBP one should assign distinct colors to all items. We provide improved lower
bounds for t = 2, 3. For t = 2, the previous known lower bound was 1.5652 [13]. For t = 3, the
previous lower bound was 5
3 ≈ 1.6667 [26]. This last result was proved even for the special case
with equal size items. Interestingly, it has elements of adaptivity, but with respect to colors (as
all items have identical sizes), and the input moves to presenting items of a new color once the
algorithm performs a certain action. We show that the competitive ratio of any online algorithm
for CLCBP with t = 2 is at least 1.717668, and that the competitive ratio of any online algorithm
for CLCBP with t = 3 is at least 1.808142.
The drawback of previous results for all those problems is that while the exact input was not
known in advance, the set of sizes used for it was determined prior to the action of the algorithm.
We show here that our methods for proving lower bounds can be combined with a number of other
approaches to result in improved lower bounds for a variety of bin packing problems. We use the
following theorem proved in [4] (see the construction in Section 3.1 and Corollary 3).
Theorem 1 (i) Let N ≥ 1 and k ≥ 2 be large positive integers. Assume that we are given an
arbitrary deterministic online algorithm for a variant of bin packing and a condition C1 on the
possible behavior of an online algorithm for one item (on the way that the item is packed). An
adversary is able to construct a sequence of values ai (1 ≤ i ≤ N ) such that for any i, ai ∈
(cid:16)k−2N +3
k4(cid:1). For any item i1 satisfying C1 and any item i2
not satisfying C1, it holds that
> k. Specifically, there are values β and γ such that for any
item i1 satisfying C1, and any item i2 not satisfying C1, it holds that ai1 < γ < ai2 and
, k−2N +2(cid:17), and in particular ai ∈ (cid:0)0, 1
ai2
ai1
ai2
ai1
> β.
(ii) If another condition C ′ is given for stopping the input (it can be a condition on the packing
or on the constructed input), it is possible to construct a sequence ai consisting of N items such that
C ′ never holds, or a sequence of N ′ < N items, such that C ′ holds after N ′ items were introduced
(but not earlier), and where the sequence satisfies the requirements above.
Examples for the condition C1 can be the following: "the item is packed as a second item of
its bin", "the item is packed into a non-empty bin", "the item is packed into a bin an item of size
3
above 1
of non-empty bins".
2 ", etc. An example for the condition C ′ can be "the algorithm has at least a given number
The construction of such inputs is based on presenting items one by one, where there is an active
(open) interval of sizes out of which future values ai are selected. When a new item is presented,
and the algorithm packs it such that it does not satisfy C1, all future items will be smaller. If the
algorithm packs a new item such that it satisfies C1, all future items will be larger. This reduces
the length of the active interval. Thus, even though the active interval becomes shorter in every
step where a new item arrives, it always has a positive length. One can see this as a kind of binary
search on the value γ, which will always be contained in the remaining interval (as it remains
non-empty). For example, Fujiwara and Kobayashi [17] used a similar approach and in their work
the middle point of the active interval is the size of the next item, and the active interval has length
that it smaller by a factor of 2 after every step. To obtain the stronger property that items whose
sizes is at least the right endpoint of the active interval are larger by a factor of k than items no
larger than the left endpoint of the active interval, the selection of the next size is performed by a
process similar to geometrical binary search.
Note that an important feature is that the value ai is defined before it is known whether C1
holds for the ith item (the item corresponding to ai, that is, the item whose size is a function of
ai). We will use this theorem throughout the paper. We study the problems in the order they were
defined.
2 Online bin packing with known optimal cost (K-O)
Here, we consider the problem K-O, and prove a new lower bound on the competitive ratio for it.
We prove the following theorem.
Theorem 2 The competitive ratio of any online algorithm for K-O is at least 87
62 ≈ 1.4032258.
Let M be a large integer that is divisible by 4 (M will be the value of the known optimal cost).
We will create several alternative inputs, such that the optimal cost will be equal to M for each
one of them.
We use the following construction. For k = 10 and N = M , define an input built using Theorem
1 as follows applied twice on different parts of the input as explained below. The outline of our
lower bound construction is as follows. The first part of the input will consist of M items of sizes
slightly above 1
7 (such that some of them, those packed first into bins, are larger than the others).
Then, there are M items of sizes slightly above 1
3 (where items packed into new bins are larger than
7 or with another item of size roughly 1
others, while those combined with items of sizes roughly 1
3 ,
or both, are slightly smaller). Finally, the algorithm will be presented with a list of identical items
of one of the three sizes 1 (exactly), or slightly above 1
3 , such that every larger
item of size slightly above 1
3 ).
Additionally, after the first M items arrive, it is possible that instead of the input explained here
there are items of sizes slightly below 6
7 , either such that every such item can be packed with any
item out of the first M items, or such that it can only be combined with the smaller items out of
the first M items (due to the property that the size of an item will be just below 6
7 , in both cases
it can be combined with at most one item of size just above 1
3 cannot be packed together with such an item (of size slightly below 2
2 , or slightly below 2
Next, we formally define our input sequences. Throughout this section, let the condition C1 be
that the item is not packed as a first item into a bin. The first M items are defined as follows.
Using Theorem 1, we create M items such that the size of item i is 1
7 + ai. These items are called
S-items. The sizes of such items are in ( 1
7 , 0.143), and there is a value γ1 such that any item whose
packing satisfies condition C1 has size below 1
7 + γ1 and any item whose packing does not satisfy
7 ).
4
C1 has size above 1
items are called large S-items.
7 + γ1. The first kind of items are called small S-items, and the second kind of
Let Y7 denote the current number of bins used by the algorithm (after all S-items have arrived),
and this is also the number of large S-items. Two possible continuations at this point are M items
of sizes equal to 4
7 − γ1 (the second option).
5 (the first option), and M − ⌈ Y7
6 ⌉ items of sizes equal to 6
Lemma 3 In both options, an optimal solution has cost M .
Proof. In the first option, an optimal solution has one item of size 4
than 0.143 in every bin. It is optimal as every item of size above 1
it can be possibly packed with smaller items).
5 and one item of size no larger
2 requires a separate bin (where
In the second option, an optimal solution uses ⌈ Y7
7 − γ1 > 0.857, and M − Y7 of them also have one item (each) of size below 1
6 ⌉ bins to pack the large S-items: Every bin
can contain at most six such items, as their sizes are in ( 1
6 ), each remaining bin has one item of
size 6
7 + γ1. This is
an optimal solution as the two larger kinds of items (those of sizes above 1
2 and the large S-items)
cannot be combined into the same bins, and the packing for each of these two kinds of items is
optimal.
7 , 1
In the first case, the algorithm can use bins containing exactly one item to pack (also) an item
5 , but it cannot use any other bin again. In the second case, as every bin has exactly one
7 + γ1, the algorithm uses an empty bin for every item of size 6
7 − γ1.
of size 4
item of size above 1
We explain the continuation of the input in the case where none of the two continuations already
defined is used. The next M items are defined using Theorem 1, and we create M items such that
the size of the ith item of the current subsequence of M items is 1
3 +ai (the values ai are constructed
here again, and they are different from the values ai constructed earlier). We call these items T -
items. The sizes of T -items are in ( 1
3 , 0.33344), and there is a value γ2 such that any item whose
packing satisfies condition C1 (defined in this section) has size below 1
3 + γ2 and for any item whose
packing does not satisfy C1, it has size above 1
3 + γ2. The first kind of items are called small
T -items, and the second type items are called large T -items.
Here, there are three possible continuations. The first one is M
2 items, all of size 1. The second
one is M items, each of size 0.52. Let Y3 denote the number of new bins created for the T -items,
which is also the number of large T -items (so after the T -items are packed the algorithm uses
Y7 + Y3 bins). If Y3 ≤ M
3 − γ2 (where
3 − γ2 > 0.66656). Otherwise (Y3 > M
2
2 ⌉ items, each of
size 2
3 − γ2) in both
cases, and the number of items is M − max{ M
3 − γ2. Thus, in the third continuation, the sizes of items are the same (i.e., 2
2 ), the third continuation is with M − ⌈ Y3
2 , the third continuation is with 3M
items, each of size 2
4
4 , ⌈ Y3
2 ⌉}.
Lemma 4 The optimal cost in all cases (i.e., after the packing of the items of each possible con-
tinuation has been completed) is exactly M .
Note that it is sufficient to show that the optimal cost is at most M , as in the case where it is
strictly smaller than M , it is possible to present items of size 1 until the optimal cost is exactly M ,
while the cost of the algorithm does not decrease. We prove that the value is exactly M to stress
the property that one cannot prove a better lower bound using the same kind of input.
Proof. For the first continuation, an optimal solution packs M
2 bins, each with two S-items and
two T -items, and another M
2 bins, each with one item of size 1. This solution is optimal as every
item of size 1 has to be packed alone into a bin, and no bin can contain more than two items of
sizes above 1
3 .
For the second continuation, an optimal solution packs M bins, each with one item of size 0.52,
one T -item and one S-item. This solution is optimal as no bin can contain more than one item of
size above 1
2 .
5
2 , there are M
For the third continuation, the two options for optimal solutions are as follows.
Y3 ≤ M
packed into these bins (which is possible as there are M
There are also M
one item of size 2
total size will be below 0.143 · 2 + 2
(this is possible as the size of each small T -item is below 1
In the case
4 bins, each with two T -items and two S-items. All large T -items will be
2 ≥ Y3 T -items packed into those bins).
2 − Y3 small T -items packed into these bins. Each of the remaining bins contains
3 − γ2, where M
4 of those bins also contain two S-items (which is possible as the
2 bins has one small T -item
3 < 1), and each of the remaining M
2 , there are ⌈ Y3
In the case Y3 > M
2 ⌉ bins with two S-items and two large T -items (at most one
bin may contain a smaller number of large T -items). All large T -items are packed into these bins,
and no small T -items are packed into these bins. The remaining bins all contain one item (each) of
size 2
2 ⌉ of those bins also contain two S-items, and M − Y3 of those bins (not
containing S-items) also contain one small T -item (this is possible as ⌈ Y3
2 ⌉+M −Y3 ≤ M ).
2 ) is optimal as separate bins are needed
The solution for the second case (i.e., for the case Y3 > M
3 − γ2, where M
2 − ⌈ Y3
2 −⌈ Y3
2 ⌉+ M
3 + γ2).
for items of size 2
3 − γ2 and large T -items, and the solution obtained for each kind is optimal.
Thus, it remains to prove that in the first case (i.e., in the case Y3 ≤ M
2 ), the optimal cost is M .
Observe that we showed a feasible solution of cost M , so we need to show that the optimal cost is
at least M . In this case every bin with an item of size 2
3 − γ2 can receive either two S-items or one
small T -item. Consider an optimal solution and let ∆ ≥ 0 be the number of items of size 2
3 − γ2
packed with a T -item. The remaining (M − ∆) T -items are packed at most two in each bin, so if
∆ ≤ M
2 + 1, at most
2( 3M
2 S-items
remain to be packed with (M − ∆) T -items. Even replacing each T -item with two items of size in
( 1
7 , 1
items where a
bin can contain at most six items, so at least M
4 = M bins.
6 ] (virtually, for the sake of proof), we have to pack 2(M − ∆) + 2∆ − M
2 = 3M
4 bins are needed, for a total of 3M
2 ≥ M bins. Otherwise, ∆ ≥ M
3 − γ2, and M − 2( 3M
4 − ∆) S-items are packed with items of size 2
2 , we are done as there are at least 3M
4 − ∆) = 2∆ − M
4 + M −∆
4 + M
2
This completes the description of the input where we showed that in each case the optimal cost
is exactly M . Next, we consider the behavior of the algorithm. Consider the kinds of bins the
algorithm may have after all T -items have arrived. The T -items do not necessarily arrive, but we
will deduce the numbers of different kinds of bins the algorithm has after the S-items have arrived
from the numbers of bins assuming that the T -items have arrived. This is an approach similar to
that used in [29], where numbers of bins packed according to certain patterns (subsets of items
that can be packed into one bin) at the end of the input are considered, and based on them, the
number of bins already opened at each step of the input are counted. More precisely, if the input
consists of batches of identical (or similar) items, given the contents of a bin it is clear when it is
opened and at what times (after arrival of sub-inputs) it should be counted towards the cost of the
algorithm.
A bin with no T -items can receive an item of size 0.52 if it has at most three S-items and it can
receive an item of size 2
3 − γ2 if it has at most two S-items. The only case where a bin with at least
one S-item and at least one T -item can receive another item (out of a continuation of the input) is
the case that a bin has one of each of these types of items, and it will receive an item of size 0.52.
Let X60 denote the number of bins with four or five or six S-items and no T -items. Such a bin
cannot receive any further items in addition to its S-items. Let X30 denote the number of bins with
three S-items and no T -items. Such a bin can receive an item of size 0.52 (but not a larger item).
Let X20 and X10 denote the number of bins with two S-items and one S-item, respectively, and no
T -items. Out of possible input items, such a bin can receive an item of size 0.52 or an item of size
2
3 − γ2. We distinguish these two kinds of bins due to the possible other continuations after T -items
have arrived. Let X41 denote the number of bins with two or three or four S-items and one T -item.
Such bins cannot receive any further items out of our inputs. Let X11 denote the number of bins
6
with one S-item and one T -item. Let X12 and X22 denote the numbers of bins with two T -items
and one and two S-items, respectively. Obviously, there can be bins without S-items containing
one or two T -items, and we denote their numbers by X01 (one T -item) and X02 (two T -items).
We have five scenarios based on the different options and continuations described above, and
we use ALGi to denote the cost of a given algorithm for each one of them, in the order they were
ALGi
M , which is a
presented. Let R be the (asymptotic) competitive ratio. Let Ai = lim supM →∞
lower bound on the competitive ratio R since the optimal cost is always M (by Lemmas 3 and 4),
so for i = 1, 2, 3, 4, 5 we have the constraint Ai ≤ R. The Ai (for i = 1, 2, 3, 4, 5) will not appear
explicitly as variables in the forthcoming linear program. Instead, we will compute each Ai based
on the other variables in the program and substitute the resulting expression in the constraint
Ai ≤ R. We use yi = Yi
M for those values of i and j such that Yi and Xij are defined.
For all thirteen variables there is a non-negativity constraint. In addition, the number of items
should satisfy Pi,j j · Xij = M and Pi,j i · Xij ≥ M (the second constraint is not an equality as in
some cases Xij counts bins with at most (i) S-items). Using the definitions of Y7 and Y3 we have
Y7 = X60 + X30 + X20 + X10 + X41 + X11 + X12 + X22 and Y3 = X01 + X02.
M and xij = Xij
We get the following four constraints:
x41 + x11 + 2x12 + 2x22 + x01 + 2x02 = 1
6x60 + 3x30 + 2x20 + x10 + 4x41 + x11 + x12 + 2x22 ≥ 1
y7 − x60 − x30 − x20 − x10 − x41 − x11 − x12 − x22 = 0
y3 − x01 − x02 = 0
(1)
(2)
(3)
(4)
ALG2 = M − ⌈ Y7
The costs of the algorithm are as follows. We have ALG1 = M + X60 + X30 + X20 + X41 + X22,
2 , and ALG4 = X60 + X41 + X22 + X12 + X02 + M .
2 , we have ALG5 =
2 , we have ALG5 = Y7 + Y3 − X20 − X10 + 3M
6 ⌉ + Y7, ALG3 = Y7 + Y3 + M
4 , and if Y3 > M
If Y3 ≤ M
Y7 + Y3 − X20 − X10 + M − ⌈ Y3
2 ⌉.
The four first costs of the algorithm (for the first four scenarios) gives the constraints
R − x60 − x30 − x20 − x41 − x22 ≥ 1
6R − 5y7 ≥ 6
2R − 2y7 − 2y3 ≥ 1
R − x60 − x41 − x22 − x12 − x02 ≥ 1
(5)
(6)
(7)
(8)
The two final constraints form two cases (according to the value of y3), and therefore our list of
constraints results in two linear programs (with all previous constraints and two additional ones).
The inputs for the two cases are different, and therefore they are considered separately (due to the
different inputs, there is one other different constraint except for the constraint on the value of y3).
For each one of the linear programs, the objective is to minimize the value of R.
One pair of constraints is y3 ≤ 1
2 and 2R − 2y7 − y3 + 2x20 + 2x10 ≥ 2 (observe that the constraint y3 ≥ 1
2 and 4R − 4y7 − 4y3 + 4x20 + 4x10 ≥ 3, and the alternative pair
2 is a relaxation
is y3 ≥ 1
of the valid constraint y3 > 1
2 , and thus the weaker constraint y3 ≥ 1
2 is valid in this case).
Multiplying the first five constraints by the values 2, 1, 3, 2, 1, respectively, and taking the sum
gives:
2x60 + 2x41 + 2x12 + 2x02 + 2x22 − 2x10 − x30 − 2x20 + 3y7 + 2y3 + R ≥ 4 .
(9)
For the first case, we take the sum of the sixth, eighth, and tenth constraints multiplied by the
values 2, 20, 5, respectively, and get:
52R − 30y7 − 20y3 − 20x60 − 20x41 − 20x22 − 20x12 − 20x02 + 20x20 + 20x10 ≥ 47 .
7
Summing this with ten times (9) we get 62R − 10x30 ≥ 87, and by x30 ≥ 0 we get R ≥ 87
1.4032258.
62 ≈
For the second case, we take the sum of the seventh, eighth, and tenth constraints multiplied
by the values 1, 4, 2, respectively, and get:
10R − 6y7 − 4y3 − 4x60 − 4x41 − 4x22 − 4x12 − 4x02 + 4x20 + 4x10 ≥ 9 .
Summing this with twice (9) we get 12R − 2x30 ≥ 17, and as x30 ≥ 0, we have R ≥ 17
Thus, we have proved R ≥ 1.4032258.
12 ≈ 1.41666.
3 Online Square packing (SP)
We continue with the online square packing (SP) problem. We prove the following theorem.
Theorem 5 The competitive ratio of any online algorithm for SP is at least 1.7515445.
Here, in the description of the input, when we refer to the size of an item, this means the length
of the side of the square (and not its area). Consider the following input. For a large positive even
integer M and k = 10, we define an input based on using Theorem 1 twice. The construction is
similar to that of the previous section, though here we are not committed to a specific optimal cost,
and we take into account the multidimensionality. Moreover, for one of the item types the number
of such items is also determined by the action of the algorithm (which was difficult to implement
in the previous section when the cost of an optimal packing is fixed in advance, and we did not
use such an approach there as extensively as in the current section). Here, we only compute upper
bounds on the optimal cost for each case.
The outline of the construction is as follows. The first part of the input will consist of M items
of sizes slightly above 1
4 (such that some of them, those packed first into bins, are larger than the
others), then, there are items of sizes slightly above 1
3 (where such items that are packed into bins
containing relatively few items, where the exact condition is defined below, will be larger than other
items of this last kind). Finally, there will be items of one of the sizes: 3
5 , and slightly below 2
3
(all of them will have exactly the same size), such that every larger item of size slightly above 1
3
cannot be packed together with such an item of size slightly smaller than 2
3 . Additionally, after the
first M items arrive, it is possible that instead of the input explained here there are items of sizes
slightly below 3
4 , such that it can be only be combined with the smaller items out of the first M
items (any bin with an item of size slightly below 3
4 may have at most five smaller items out of the
first M items in a common bin).
Next, we formally define the construction. Let the condition C11 be that the item is not packed
4 . For
3 , let the condition C12 be that the item is either packed in a bin
3 , or that it contains at least five items whose sizes are in
as a first item into a bin. This is the condition we will use for items of sizes slightly above 1
items of sizes slightly above 1
already containing an item of size above 1
4 , 1
( 1
3 ].
The first M items are defined as follows. Using Theorem 1, we create M items such that the
size of item i is 1
4 + ai. These items are called F -items. The sizes of items are in (0.25, 0.2501), and
there is a value γ1 such that any item whose packing satisfies condition C11 has size below 1
4 + γ1
and for any item whose packing does not satisfy C11, it has size above 1
4 + γ1. The first kind of
items are called small F -items, and the second type items are called large F -items. No matter how
the input continues, as any packing of the first M items requires at least M
9 bins, the cost of an
optimal solution is Ω(M ).
Let Y4 denote the current number of bins used by the algorithm, and this is also the number
⌉ items of (identical) sizes equal
of large F -items. A possible continuation at this point is ⌈ M −Y4
5
8
4 − γ1. Note that such an item cannot be packed into a bin with an item of size above 1
to 3
4 + γ1,
as it cannot be packed next to it or below (or above) it, and the remaining space (not next to it
or below it or above it) is too small (the sum of the diagonals of these two items is too large to be
packed into a unit square bin).
Lemma 6 There exists a packing of the items of the presented sequence (in this case) of cost at
most M
5 − 4Y4
45 + 2.
5
Proof. A possible packing of the items of sizes 3
4 − γ1 together with the (M ) F -items is to use
⌈ M −Y4
⌉ bins for the new items and combine five small F -items into these bins (one such bin may
have a smaller number of F -items). This packing is feasible as the large item can be packed in
one corner of a unit square bin, leaving an L shaped area of width 1
4 + γ1, the opposite corner will
contain an F -item, and there are two additional such items next to it on each side of the L shaped
area. The remaining large F -items are packed into bins containing nine items each (possibly except
for one bin), such that the number of such bins is ⌈ Y4
9 ⌉. The total number of bins in this packing
is at most M
5 − 4Y4
45 + 2.
The algorithm has one large F -item in each of the first Y4 bins and therefore it uses a new bin
4 − γ1. Thus, the total number of bins in the packing of the algorithm (in
for every item of size 3
this case) is exactly Y4 + ⌈ M −Y4
⌉.
5
3 + ai, and the sizes of items are in ( 1
We explain the continuation of the input in the case where the continuation defined above is
not used. Here, for the construction, we state an upper bound on the number of items as the exact
number of items is not known in advance and it will be determined during the presentation of the
input. There will be at most 1.5M items of sizes slightly above 1
3 . We will use the variables S3
and L3 to denote the numbers of items for which condition C12 was satisfied and was not satisfied,
respectively, in the current construction.
Initialize S3 = L3 = 0, and increase the value of the
suitable variable by 1 when a new item is presented. The ith item of the current construction
has size 1
3 , 0.33344). These items are called T -items. There
is a value γ2 such that any item whose packing satisfies condition C12 has size below 1
3 + γ2 and
any item whose packing does not satisfy C12 has size above 1
3 + γ2. The first kind of items are
called smaller T -items and the second type items are called larger T -items. Present items until
8S3 + 15L3 ≥ 12M holds (this does not hold initially, so at least one item is presented, and this
is defined to be condition C ′). We show that indeed at most 1.5M items are presented. If 1.5M
items were already presented, 8S3 + 15L3 ≥ 8 · (1.5M ) = 12M , and therefore the construction is
stopped. In what follows, let S3 and L3 denote the final values of these variables. Before the last
item of this part of the input was presented, it either was the case that 8(S3 − 1) + 15L3 < 12M
or 8S3 + 15(L3 − 1) < 12M (as exactly one of S3 and L3 was increased by 1 when the last item
was presented), so 8S3 + 15L3 − 15 < 12M , or alternatively, 8S3 + 15L3 ≤ 12M + 15. Moreover,
S3 + L3 ≥ 4M
5 as 12M ≤ 8S3 + 15L3 ≤ 15(S3 + L3). Let M ′ = S3 + L3 (and we have M ′ = Θ(M )).
3 ⌋) identical items, each of size
Here, there are two possible continuations. The first one is (⌊ M ′
3 ⌋ identical items, each of size 2
exactly 0.6, and the second one is ⌊ S3
3 − γ2.
Lemma 7 The optimal cost in the first continuation is at most M
9 + 7S3
27 + 7L3
27 + 3.
Proof. A possible packing for this case consists of ⌊ M ′
3 ⌋ bins with one item of size 0.6, three T -
items, and two F -items (placing the item of size 0.6 in a corner leaves an L shaped area of width 0.4,
so we place one T -item in each of the other corners and in the remaining space between each pair of
adjacent T -items we pack an F -item). As M ′ ≤ 3M
3 ⌋ ≥ 0 unpacked F -items,
27 +2 bins, where each bin has nine items (the
2 , there are M − 2⌊ M ′
and they are packed into exactly ⌈
9 − 2M ′
M −2⌊ M ′
3 ⌋
⌉ ≤ M
9
9
last bin may have less items). In addition, there are at most two unpacked T -items, and they are
packed into a bin together. The total number of bins is at most M
27 + 3.
27 + 3 = M
9 + 7M ′
27 + 7L3
9 + 7S3
Lemma 8 The optimal cost in the second continuation is at most S3
3 + L3
4 + 2.
3 ⌋ bins with one item of size 2
Proof. A possible packing for this case consists of ⌊ S3
3 − γ2, three
small T -items, and two F -items (placing the item of size 2
3 − γ2 in a corner of a unit square bin
leaves an L shaped area of width 1
3 + γ2 where the remaining items are packed). There are at least
(S3 − 2) T -items that were packed and at least (2 S3−2
) F -items are packed. There are also ⌈ L3+2
4 ⌉
bins, each with at most four T -items and at most five F -items (there is a square with four larger
items in a corner and the smaller items are packed around them, in the L-shaped area of the bin).
This allows to pack the remaining T -items as there is space for at least S3 + L3 such items, and to
pack all F -items as there is a place for at least 2 S3−2
4 ≥ M such items, where
the last inequality holds by the condition 8S3 + 15L3 ≥ 12M . The total number of bins is at most
3 + L3
S3
3 + 5 L3+2
4 ≥ 2S3
3 + 5 L3
4 + 2.
3
Let Y3 denote the number of new bins created for the T -items (where these bins were empty prior
to the arrival of T -items). Here, there may be previously existing bins containing larger T -items
(with at most four F -items), and Y3 ≤ L3. Consider the kinds of bins the algorithm may have after
all T -items have arrived. Once again, T -items do not necessarily arrive, but we will deduce the
numbers of different kinds of bins the algorithm has after all F -items have arrived based on number
of bins existing after the arrival of T -items. After all T -items have arrived, a non-empty bin can
receive an item of size 0.6 if it has at most five items, out of which at most three are T -items. The
construction is such that any non-empty bin except for bins with at most five F -items has either
at least six items in total (each of size above 1
3 + γ2 (or both
options may occur simultaneously), and therefore it cannot receive an item of size above 2
4 ) or it has an item of size above 1
3 − γ2.
2 that may arrive (but not an item of size 3
Consider a given online algorithm for SP after the T -items were presented. Let X90 denote
the number of bins with six, seven, eight, or nine F -items and no T -items. Such a bin cannot
receive any further items in addition to its F -items in any of our continuations. Let X50 denote
the number of bins with at least one and at most five F -items and no T -items. Such a bin can
receive any item of size larger than 1
4 − γ1). Let X81
denote the number of bins with five, six, seven, or eight F -items and one (small) T -item. Let X41
denote the number of bins with at least one and at most four F -items and one (large) T -item. Let
X72 denote the number of bins with five, six, or seven F -items and two (small) T -items. Let X42
denote the number of bins with four F -items and two T -items (out of which one is small and one
is large). Let X32 be the number of bins with at least one and at most three F -items and two
T -items (out of which one is small and one is large). Let X63 denote the number of bins with five
or six F -items and three T -items (all of which are small). Let X43 denote the number of bins with
three or four F -items and three T -items (out of which two are small and one is large). Let X23
denote the number of bins with one or two F -items and three T -items (out of which two are small
and one is large). Let X54 denote the number of bins with five F -items and four T -items (all of
which are small). Let X44 denote the number of bins with two or three or four F -items and four
T -items (out of which three are small and one is large). Let X14 denote the number of bins with
one F -item and four T -items (out of which three are small and one is large).
Let X03 be the number of bins with no F -items and at least one and at most three T -items,
one of which is a large T -item, while the others (at most two) are small. Let X04 be the number of
bins with no F items and four T -items, one of which is large, while three are small.
We have three scenarios, and we use ALGi to denote the cost of the algorithm for each one
ALGi
M . The optimal cost is
of them, in the order they were presented. Let Ai = lim supM →∞
10
M and s3 = S3
M and xij = Xij
M , and the ratio Ai
OP Ti
Oi
M , so 12 ≤ 8s3 + 15ℓ3 ≤ 12 + 15
always in Θ(M ), and we let OP Ti denote our upper bounds on the optimal cost of the ith scenario,
is lower bound on the competitive ratio R. We use the
Oi = lim inf M →∞
notation yi = Yi
M for those values of i and j such that Yi and Xij are defined. Let
ℓ3 = L3
M , and for M growing to infinity, 8s3 + 15ℓ3 = 12.
Let R be the (asymptotic) competitive ratio. For all twenty variables there is a non-negativity
constraint. In addition, the number of items should satisfy Pi,j j·Xij ≥ S3+L3 and Pi,j i·Xij ≥ M
(once again, the first constraint is inequality and not equality as X03 counts also bins with less than
three T -items, and the second constraint is not an equality as in some cases Xij counts bins with
fewer than i F -items). Using the definitions of Y4 and Y3 we have Y4 = X90 + X50 + X81 + X41 +
X72 + X42 + X32 + X63 + X43 + X23 + X54 + X44 + X14 and Y3 = X03 + X04.
We also have ALG1 = Y4 + ⌈ M −Y4
⌉ while OP T1 ≤ M
5
5 − 4Y4
45 + 2, so
R ≥
A1
O1
≥
1/5 + 4y4/5
1/5 − 4y4/45
=
9 + 36y4
9 − 4y4
.
Additionally, ALG2 = Y4 + Y3 − X50 − X41 − X32 − X23 − X03 + ⌊ M ′
X32 − X23 − X03 + S3+L3
Y4 + Y3 − X50 + S3
3 − 2 while OP T2 ≤ M
3 + L3
3 − 1 while OP T3 ≤ S3
9 + 7S3
4 + 2, so
27 + 7L3
27 + 3, and ALG3 = Y4 + Y3 − X50 + ⌊ S3
3 ⌋ ≥ Y4 + Y3 − X50 − X41 −
3 ⌋ ≥
R ≥
A2
O2
≥
y4 + y3 − x50 − x41 − x32 − x23 − x03 + s3/3 + ℓ3/3
7s3/27 + 7ℓ3/27 + 1/9
and R ≥ A3
O3
≥ y4+y3−x50+s3/3
s3/3+ℓ3/4
.
We get the following set of constraints:
y4 = x90 + x50 + x81 + x41 + x72 + x42 + x32 + x63 + x43 + x23 + x54 + x44 + x14
8s3 + 15ℓ3 = 12
y3 = x03 + x04
x81 + x41 + 2x72 + 2x42 + 2x32 + 3x63 + 3x43 + 3x23 + 4x44 + 4x54 + 4x14 + 3x03 + 4x04
≥ ℓ3 + s3
x41 + x42 + x32 + x43 + x23 + x44 + x14 + x03 + x04 = ℓ3
9x90 + 5x50 + 8x81 + 4x41 + 7x72 + 4x42 + 3x32 + 6x63 + 4x43 + 2x23 + 5x54 + 4x44 + x14
≥ 1
9 + 36y4 ≤ R(9 − 4y4)
(y4 + y3 − x50 − x41 − x32 − x23 − x03 + s3/3 + ℓ3/3) ≤ R(7s3/27 + 7ℓ3/27 + 1/9)
y4 + y3 − x50 + s3/3 ≤ R(s3/3 + ℓ3/4)
(10)
(11)
(12)
(13)
(14)
(15)
(16)
(17)
(18)
The optimal objective function value of the mathematical program of minimizing R subject
to all these constraints is approximately 1.751544578513 (and it is not smaller than this number).
Thus, we have proved R ≥ 1.751544578513.
4 Online class constrained bin packing (CLCBP)
In this section we exhibit our approach to proving lower bounds for the last variant of the bin
packing problem which we study here, by improving the known lower bounds for the cases t = 2
and t = 3 of CLCBP. We will prove the following theorem.
11
Theorem 9 The competitive ratios of online algorithms for CLCBP with t = 2 and t = 3 are at
least 1.717668486 and at least 1.80814287, respectively.
The constructions for t = 2 and t = 3 have clear differences, but the general idea is similar. The
outline of the constructions is as follows. Start with a large number of tiny items, all of distinct
colors, so every bin of any algorithm will contain at most t tiny items. Here, the construction is
such that the items packed first into their bins are much larger than other items (large tiny items
will be larger by at least a constant multiplicative factor than small tiny items, but they are still
very small). One option at this point is to continue with huge items of sizes close to 1, all of distinct
colors out of the colors of small tiny items, such that every item of size almost 1 can be packed into
a bin with t small tiny items in an offline solution, one of which has the same color as the huge item
packed with it. Note that no large tiny item can be combined with a huge item, so those items will
be packed separately, t items per bin. The number of huge items is chosen in a way such that the
optimal cost is not increased. Another option to continue the construction (instead of introducing
the huge items) is with items of sizes slightly above 1
3 , where an item packed into a bin already
containing an item of size above 1
3 is smaller than an item packed into a bin with no such item
(but it could possibly be packed with tiny items). It is ensured that bins of the algorithm already
containing t (tiny) items will not be used again by the algorithm by never introducing items of
their colors again. The sizes will be 1
3 plus small values, where these small values are much larger
than sizes of tiny items (including sizes of large tiny items). An interesting feature is that there
will be exactly two items of sizes slightly above 1
3 with each color which is used for such items,
where the idea is to reuse (as much as possible) colors of tiny items packed by the algorithm into
bins with at most t − 1 tiny items (where those tiny items can be large or small), and never reuse
colors of tiny items packed in bins of t items. In some cases (if there are too few such colors which
can be reused), new colors are used as well for items of sizes slightly above 1
3 (but there are still
two items of sizes just above 1
3 for each color). After these last items are presented, the final list of
3 , 1
items will be items of sizes above 1
2 ]
with the goal of packing such pairs of one color together into bins of offline solutions. There are
two options for the final items. There are either such items not much larger than 1
2 , or there are
3 , such that such an item having a color of an item of size slightly above 1
items of sizes close to 2
3
can be combined into a bin with that item and with at most t tiny items coming from bins of the
algorithm with at most t − 1 items (no matter whether they are small or large, but one of them
has to be of the same color). However, in the case of items of sizes almost 2
3 , only small items of
sizes just above 1
3 will be combined with them in good offline solutions while others are packed in
pairs (of the same color whenever possible, and of different colors otherwise, combining tiny items
where possible).
2 whose colors will match exactly those of items of sizes in ( 1
First, we present the parts of the constructions that are identical for t = 2 and t = 3. The
condition C1 will be that the current item is not the first item of its type packed into its bin,
where a type consists of all items of similar size (the two relevant types are tiny items and items of
sizes slightly above 1
3 ). Let M > 1 be a large integer divisible by 6. The construction starts with
the first type of items, where these items are called E-items or tiny items, consisting of M items
constructed using Theorem 1. Let the value of k be 20, and the resulting values ai are smaller than
20−22M +2
. The number of tiny items presented is always exactly M (so the stopping condition is
that there are M items), and the size of the ith item is simply ai. Every E-item has its own color
that may be reused in future parts of the construction but not for E-items. Let ε1 and γ1 be such
that the size of any E-item satisfying C1 (which we call a small E-item) is below 2ε1
t and the
size of any E-item not satisfying C1 (which we call a large E-item) is above 2ε1 (but smaller than
20−22M +2
). Let Xj (for 1 ≤ j ≤ t) be the number of bins of the algorithm with j E-items. Let X
denote the total number of bins of E-items, i.e., X = Pt
20 < ε1
j=1 Xj.
12
If huge items arrive now, their number is ⌊ M −X
⌋ and their colors are distinct colors out of
colors of small E-items. The size of every huge item is 1 − ε1. If Xt ≤ M
2t , there are no other
continuations. In all other cases, there are two possible continuations except for the one with huge
items, which was just discussed.
t
22
1022M +3
=
2M +2
2M +3
2M +2
2M +2
, 10−22M +2
). We have (by M ≥ 1) 10−2
20−22M +2 = 102
In all other continuations, items of a second type are presented such that their number is at
most 2M , and they will be called T -items. They are constructed using Theorem 1 with k = 10, so
their values of ai are in (10−22M +3
22
104 > 6 > t. The size of the ith T -item is 1
3 + ai, and here condition C1 means that the
T -item is packed by the algorithm as the second T -item of its bin. Let ε2 and γ2 be such that a
T -item satisfying C1 (which we call a small T -item) has size smaller than 1
10 and a T -item not
satisfying C1 (which we call a large T -item) has size larger than 1
3 + ε2. The number of T -items is
even, and their colors are such that there are two T -items for each color. These colors are colors of
E-items that are not packed in bins of t E-items by the algorithm. As the number of such E-items
is M − t · Xt, if the number of T -items is larger than 2(M − t · Xt), new colors (which were not used
for any earlier item) are used (and for the new colors there are also two T -items for each color).
The variables Z1 and Z2 denote the numbers of bins with at least one T -item and with exactly two
T -items, respectively, used by the algorithm (so Z2 ≤ Z1). The algorithm may use bins with at
most (t − 1) E-items to pack T items (but not bins with (t) E-items, as no additional items have
colors as those items).
3 + ε2
For t = 2, the number of T -items is max{2X1, 2X2}. Since 2X2 ≤ M and 2X1 ≤ 2M , the
number of T -items does not exceed 2M . For t = 3, the stopping condition is defined as follows.
First, present items until at least one of Z1 + Z2 + 6X3 ≥ 2M − 1, 3Z1 + 4Z2 ≥ 2M − 7 holds.
Then, if the second condition holds, stop presenting items. If the first condition holds (and the
second one does not hold), continue presenting items until 2Z1 + 3Z2 ≥ 6X3 − 5 holds and stop.
At this time, if the current number of T -items is odd, one additional item is presented. Thus, we
guarantee that the value of Z1 + Z2 is an even number. Since the value X3 is already fixed when
T -items are presented, we analyze the increase in the value of each expression when a new T -item
is presented. If a new item is packed into a bin with no T -item (and it is large), then the value
of Z1 increases by 1 while the value of Z2 is unchanged. Otherwise (it is small), the value of Z2
increases by 1 while the value of Z1 is unchanged. Thus, the value of Z1 + Z2 can increase by at
most 1, while that of 3Z1 + 4Z2 can increase by at most 4, and that of 2Z1 + 3Z2 can increase by
at most 3. Thus, there are two cases. If the first condition that holds is 3Z1 + 4Z2 ≥ 2M − 7,
when it started to hold, the value of the left hand side was increased by at most 4. If another
item is presented to make the number of items even, it could increase by at most 4 again, so
3Z1 + 4Z2 ≤ 2M .
If Z1 + Z2 + 6X3 ≥ 2M − 1 holds first (note that the two conditions could
potentially start holding at the same time), then still Z1 + Z2 + 6X3 ≤ 2M . If in the current step
it holds that Z1 + Z2 + 6X3 ≥ 2M − 1 and 3Z1 + 4Z2 ≤ 2M − 8, at that time, 2Z1 + 3Z2 ≤ 6X3 − 6
holds (as otherwise, taking the sum of Z1 + Z2 + 6X3 ≥ 2M − 1 and 2Z1 + 3Z2 ≥ 6X3 − 5 gives
3Z1 +4Z2 ≥ 2M −6 > 2M −7). Therefore in the case the first condition holds first while the second
one does not, additional items are presented and finally 2Z1 + 3Z2 ≤ 6X3 (counting the last two
items). Thus, after all T -items have arrived, it is either the case that Z1 + Z2 ≤ 3Z1 + 4Z2 ≤ 2M or
that Z1 + Z2 ≤ 2Z1 + 3Z2 ≤ 6X3 ≤ 2M (as 3X3 ≤ M ), so there are indeed at most (2M ) T -items.
2 with the same color. There are two
continuations as follows. In the first one, there are items of sizes 0.6, such that there is a matching
item for every T -item (a different matching item for every item, i.e., Z1 + Z2 items of size 0.6
in total). In the second one, there are items of sizes 2
5 , such that every small T -item has a
matching item (once again, a different matching item for every item, i.e., Z2 items in total). This
concludes the description of our lower bounds constructions for the two cases of t = 2 and t = 3.
A matching item for a T -item is an item of size above 1
3 − ε2
13
4.1 The analysis
Let ALGi and OP Ti respectively denote the costs of the algorithm and of an optimal solution for
the ith continuation. We use algi = ALGi
M . This auxiliary notation will assist
us as we would like to find the bounds for M growing to infinity. The competitive ratio satisfies
. We will also use xi = Xi
ALGi
M , for values of i
R ≥ lim supM →∞
OP Ti
that these variables are defined, and x = X
M .
M and opti = OP Ti
M and zi = Zi
= lim supM →∞
algi
opti
Consider a given online algorithm and an offline solution after the huge items are presented.
Lemma 10 We have ALG1 = X + ⌊ M −X
t
⌋ and OP T1 ≤ M
t .
Proof. Every huge item can be packed with t small E-items, if one of them has the same color as
the huge item. No huge item can be packed with a large E-item in one bin. Thus, the algorithm
has ⌊ M −X
⌋ bins with huge items (one huge item packed into each such bin), and all of them contain
no other items (as every bin of the algorithm with E-items has a large E-item). A possible offline
solution has ⌊ M −X
⌋ bins with a huge item and a small E-item of the same color (as the color of
t
t
the huge item) and t − 1 other small E-items, and there are
⌋ bins with
t E-items not packed in the previous set of bins. All E-items are packed, and the total number of
bins is M
t .
t
t
= M
t − ⌊ M −X
t
M −t(⌊ M −X
⌋)
Lemma 11 We have R ≥ tx + (1 − x). If xt ≤ 1
2t , then the competitive ratio is at least 2 − 1
2t .
Proof. In this case we consider the input without continuations. By Lemma 10 and by letting M
grow to infinity, we have alg1 = t−1
2t , at least
M − tXt ≥ M
t−1 and
x ≥ xt + 1−txt
2 items are packed in bins containing at most t − 1 items, and thus x − xt ≥ 1−txt
t , and R ≥ (t − 1)x + 1. As Xt ≤ M
2t(t−1) . We get R ≥ 4t−1
t−1 ≥ 1−1/(2t)
t−1 = 2t−1
2t = 2 − 1
2t .
t x + 1
t , opt1 ≤ 1
t−1 = 1−xt
Using the first part of the last lemma, we get R ≥ 2x + (1 − x) = x + 1 = x1 + x2 + 1 for t = 2,
and R ≥ 2x + 1 for t = 3. As we prove lower bounds that are lower than 1.75 for t = 2 and lower
than 1.8333 for t = 3, by the last lemma, it is left to deal with the case xt ≥ 1
2t . Note that the
continuation of huge items is still possible for those cases. The remaining part of the analysis is
performed separately for the two cases.
2 < 2x2. As the number of
The case t = 2.
T -items is 2 max{X1, X2}, there are two T -items of any color of an E-item packed alone in a bin
by the algorithm just after the E-items have arrived.
In this case we assume x2 > 1
4 and therefore x1 < 1
Lemma 12 We have alg2 ≥ x2 + z1 + 2z2, alg3 ≥ x2 + z1 + z2, opt2 ≤ z1 + z2, and opt3 ≤
z1+2z2+2x2−max{x1,x2}
2
.
Proof. The algorithm never reuses bins with two E-items as no further item has color of any of
their colors. If the final items have sizes of 0.6, the bins with one T -item can possibly be reused
(but not those with two such items). The number of final items is the same as the T -items, that
is, Z1 + Z2. If the final items have sizes of 2
5 , as any bin with at least one T -item has a large
T -item, no bins with T -items can be reused by the algorithm. The number of final items is Z2 in
this case. The lower bounds on the costs of the algorithm follow from the numbers of items of sizes
above 1
2 in the final part of the input, and from the property that they cannot be added to bins
with two tiny items, to bins with two T -items, and in the case of items of sizes 2
5 they cannot
be added to any bin with a large T -item (in this case they cannot be added to any bin with at least
one T -item).
3 − ε2
3 − ε2
14
Consider the following offline solutions. If the final items have sizes of 0.6, every bin contains a
T -item and its matching item of size 0.6. It also contains an E-item of the same color, if it exists
(it is also possible that it exists but it is packed in another bin with a T -item of the same color),
and at most one E-item of another color. As Z1 + Z2 ≥ 2X1 and Z1 + Z2 ≥ 2X2, every E-item
packed alone in the algorithm (after all the E-items arrive) has a T -item of its color (there are two
items with this color, and it can be packed with one of them). Given the number of bins of this
solution, it is possible to add (at most) one E-item, which is packed in bins of two E-items by the
algorithm, to each bin containing an item of size 0.6 (as the number of such E-items is 2X2 and
the number of bins is the number of T -items, that is, at least 2X2). The total size of items in every
bin is below 0.94. Thus, those Z1 + Z2 bins are packed in a valid way and contain all items.
3 − ε2
If the final items have sizes of 2
3 − ε2
5 , as E-items have sizes no larger than ε2
60 , it is possible
to pack one small T -item with its matching item of size 2
5 , and at most two E-items, one
of which has the same color as the small T -item. As there are Z2 small T -items, there are at
least Z1 − Z2 large T -items such that the other T -item of the same color is large, and therefore
there are at least Z1−Z2
pairs of large T -items with common colors (as Z1 + Z2 is even, Z1 − Z2
is even too). There are Z1 − Z2 large T -items that are packed in pairs, such that Z1−Z2
pairs of
two large T -items of the same color are packed together with one E-item of their color and one
E-item of another color (because it cannot contain items of an additional color). Note that even
if there is a larger number of pairs of large T -items with common colors, exactly Z1−Z2
pairs are
packed in this way. The other large T -items and unpacked E-items are simply packed in pairs.
Note that there are (2X2) E-items with unique colors (where no other item has the same color).
We have packed Z2 + Z1−Z2
= max{X1, X2} < 2X2 items (recall that the number of
T -items is 2 · max{X1, X2} and it is also Z1 + Z2, while there are (2X2) E-items of unique colors
and the number of other E-items is X1 ≤ max{X1, X2}, while the number of colors of T -items
is max{X1, X2}) that are E-items with unique colors, so there are still such items to be packed.
There are Z1 large T -items, and therefore Z2 such items remain. Therefore, as the number of
⌉ = X2 − ⌊ Z1−Z2
unpacked E-items of unique colors is 2X2 − Z1+Z2
⌋
bins are used for the pairs. Thus, the number of bins is at most Z2 + Z1−Z2
4 + 1 =
X2 + Z1+3Z2
2 , and in the case
Z1 + Z2 = 2X1, we have X2 + Z1+3Z2
2 . In both cases the number of bins is
at most Z1
4
. The other E-items are packed with T -items of their colors.
4 + 1. In the case Z1 + Z2 = 2X2, we have X2 + Z1+3Z2
2 +Z2
2
2 + X2 − Z1−Z2
= Z1
2 + Z2 + X2 − X1
= Z1
2 + Z2 + X2
= Z1+Z2
2
2
2 + Z2 + X2 − max{X1,X2}
2
, an additional ⌈
2
2
2
2
2X2− Z1+Z2
4
4
Here we solve two mathematical programs, both minimizing R under constraints including non-
negativity constraints for all variables, and the properties x1 ≤ 2x2, x1 + x2 + 1 ≤ R, x2 + z1 + 2z2 ≤
R(z1 + z2), z2 ≤ z1, and x1 + 2x2 = 1.
The first program is for the case x2 ≥ x1, which is one of the constraints (where z1 + z2 = 2x2).
The other constraints are z2 + z1 − 2x2 = 0, x2 + z1 + z2 ≤ R(z1/2 + z2 + x2/2). Solving the program
shows that R ≥ 1.7320507 in this case.
The second program is for the case x1 ≥ x2, which is one of the constraints (and here z1 + z2 =
2x1). The other constraints are z2 + z1 − 2x1 = 0, x2 + z1 + z2 ≤ R(z1/2 + z2 + x2 − x1/2). Solving
the program shows that R ≥ 1.717668486 in this case.
The case t = 3.
In this case we assume x3 > 1
6 .
Lemma 13 We have alg2 ≥ x3 + z1 + 2z2, alg3 ≥ x3 + z1 + z2, opt2 ≤ z1 + z2, and opt3 ≤ z1+2z2
2
.
Proof. The algorithm never reuses bins with three E-items as no further item has any color of
their colors. Other than that, the arguments for the costs of the algorithm are the same as in the
case t = 2.
15
Next, we analyze offline solutions. In both cases of final items, the difference with the case t = 2
is that every bin can contain two E-items whose colors are unique (either because they come from
bins with three E-items of the algorithm or because the number of colors of T -items is smaller than
the number of items coming from bins of the algorithm with less than three items). It is possible
to add such items to the bins as the size of three E-items is still below ε2
10 .
We first calculate the number of E-items of unique colors (that is, E-items of colors that appear
In the case where Z1 + Z2 + 6X3 ≥ 2M − 1 we have in fact
only once for the entire input).
Z1 + Z2 + 6X3 ≥ 2M as the value Z1 + Z2 is even.
In this case every E-item packed in a bin
with less than three E-items by the algorithm has two T -items of its color, and it can always be
packed with one of them. In this case the number of E-items of unique colors is 3X3. Otherwise,
the number of E-items of unique colors is M − Z1+Z2
, as there are (Z1 + Z2) T -items, and there
are two T -items of each color.
2
3 − ε2
We claim that in the case of final items of sizes 2
5 , it is possible to pack all E-items of
unique colors, possibly except for a constant number of items which can be packed separately into
a constant number of bins. We claim that there is always space for at least (Z1 + 1.5Z2 − 1) E-items
of unique colors. The difference with the case t = 2 is that the bins with the final items can receive
two E-items of unique colors and not only one (and there are Z2 such bins). The bins with pairs
of large T -items of one color can receive two E-items of unique colors (and there are Z1−Z2
such
bins), and the remaining bins, with two large T -items of distinct colors can receive one such E-item
(and there are ⌈ Z2
2 ⌉) E-items
If their number if 3X3, we also have Z1 + 1.5Z2 ≥ 3X3 − 2.5, so excluding a
of unique colors.
constant number of such items, all of them are packed. If their number is M − Z1+Z2
, we also have
3Z1 + 4Z2 ≥ 2M − 7, so M − Z1+Z2
2 ≤ Z1 + 1.5Z2 + 3.5. Thus, we find opt2 ≤ z1+2z2
. In the case
where the final items have sizes of 0.6, it is possible to pack (2Z1 + 2Z2) E-items of unique colors
in those bins, and opt2 ≤ z1 + z2.
2 ⌉ such bins). Thus, it is possible to pack at least (2Z2 + 2 Z1−Z2
2 + ⌈ Z2
2
2
2
Here we also solve two mathematical programs, both minimizing R under constraints including
non-negativity constraints for all variables. Other constraints are x1+2x2+3x3 = 1, x = x1+x2+x3,
1 + 2x ≤ R, z2 ≤ z1, x3 + z1 + z2 ≤ R(z1 + 2z2)/2, and x3 + z1 + 2z2 ≤ R(z1 + z2).
The first program is for the case where Z1 + Z2 + 6X3 ≥ 2M − 1 and −5 ≤ 2Z1+ 3Z2− 6X3 ≤ 0.
These properties result in the constraints z1 + z2 + 6x3 ≥ 2 and 2z1 + 3z2 − 6x3 = 0. Solving the
program shows that R ≥ 1.902018 in this case.
The second program is for the case where Z1+Z2+6X3 ≤ 2M +12 and 2M −7 ≤ 3Z1+4Z2 ≤ 2M
hold. Note that if we stop presenting T -items due to the second case where 2M − 7 ≤ 3Z1 + 4Z2
it means that in the previous (even-indexed) step the first condition Z1 + Z2 + 6X3 ≥ 2M − 1 did
not hold. Therefore, at that time Z1 + Z2 + 6X3 ≤ 2M − 2 holds, and the value of the left hand
side may increase by at most 7 in one step (and thus by at most 14 in the last two steps). Those
properties result in the constraints z1 + z2 + 6x3 ≤ 2 and 3z1 + 4z2 = 2. Solving the program shows
that R ≥ 1.80814287 in this case.
5 Summary
We showed that the method of designing fully adaptive instances, previously used for cardinality
constrained bin packing and vector packing [4] (see also [9, 2, 17]) can be used to improve the
known lower bounds for several additional bin packing problems. We analyzed its effect (together
with many additional ideas) for several variants, and expect that it could be useful for a number
of other variants as well.
16
References
[1] S. Angelopoulos, C. Durr, S. Kamali, M. P. Renault, A. Ros´en. Online bin packing with advice
of small size. In Proc. of The 14th International Symposium Algorithms and Data Structures
(WADS'15), 40 -- 53, 2015.
[2] L. Babel, B. Chen, H. Kellerer, and V. Kotov. Algorithms for on-line bin-packing problems
with cardinality constraints. Discrete Applied Mathematics, 143(1-3):238 -- 251, 2004.
[3] J. Balogh and J. B´ek´esi. Semi-on-line bin packing: a short overview and a new lower bound.
Central European Journal of Operations Research, 21(4):685 -- 698, 2013.
[4] J. Balogh, J. B´ek´esi, Gy. D´osa, L. Epstein, and A. Levin, Online bin packing with cardinal-
ity constraints resolved. The Computing Res. Rep. (CoRR), http://arxiv.org/abs/1608.06415,
2016. Also in ESA'17, to appear.
[5] J. Balogh, J. B´ek´esi, Gy. D´osa, L. Epstein, and A. Levin, A new and improved algorithm
for online bin packing. The Computing Res. Rep. (CoRR), http://arxiv.org/abs/1707.01728,
2017.
[6] J. Balogh, J. B´ek´esi, and G. Galambos. New lower bounds for certain classes of bin packing
algorithms. Theoretical Computer Science, 440-441:1 -- 13, 2012.
[7] N. Bansal, J. Correa, C. Kenyon, and M. Sviridenko. Bin packing in multiple dimensions:
Inapproximability results and approximation schemes. Mathematics of Operations Research,
31(1):31 -- 49, 2006.
[8] J. B´ek´esi, Gy. D´osa, and L. Epstein. Bounds for online bin packing with cardinality constraints.
Information and Computation 249:190 -- 204, 2016.
[9] D. Blitz. Lower bounds on the asymptotic worst-case ratios of on-line bin packing algorithms.
M.Sc. thesis, University of Rotterdam, number 114682, 1996.
[10] J. Boyar, S. Kamali, K. S. Larsen, and A. L´opez-Ortiz. Online bin packing with advice. Algo-
rithmica 74(1):507 -- 527, 2016.
[11] D. Coppersmith and P. Raghavan. Multidimensional online bin packing: Algorithms and worst
case analysis. Operations Research Letters, 8(1):17 -- 20, 1989.
[12] L. Epstein. Online bin packing with cardinality constraints. SIAM Journal on Discrete Math-
ematics, 20(4):1015 -- 1030, 2006.
[13] L. Epstein, Cs. Imreh, and A. Levin. Class constrained bin packing revisited. Theoretical
Computer Science, 411(34-36):3073 -- 3089, 2010.
[14] L. Epstein and A. Levin. On bin packing with conflicts. SIAM J. on Optimization, 19(3):1270 --
1298, 2008.
[15] L. Epstein and A. Levin. Robust approximation schemes for cube packing. SIAM Journal on
Optimization, 23(2):1310 -- 1343, 2013.
[16] L. Epstein and R. van Stee. Online square and cube packing. Acta Informatica, 41(9):595 -- 606,
2005.
17
[17] H. Fujiwara and K. Kobayashi. Improved lower bounds for the online bin packing problem
with cardinality constraints. Journal of Combinatorial Optimization, 29(1):67 -- 87, 2015.
[18] S. Heydrich, R. van Stee, Improved Lower Bounds for Online Hypercube Packing. The Com-
puting Res. Rep. (CoRR), http://arxiv.org/abs/1607.01229, 2016.
[19] D. S. Johnson. Fast algorithms for bin packing. Journal of Computer and System Sciences,
8:272 -- 314, 1974.
[20] D. S. Johnson, A. Demers, J. D. Ullman, M. R. Garey, and R. L. Graham. Worst-case perfor-
mance bounds for simple one-dimensional packing algorithms. SIAM Journal on Computing,
3:256 -- 278, 1974.
[21] H. Kellerer and U. Pferschy. Cardinality constrained bin-packing problems. Annals of Opera-
tions Research, 92:335 -- 348, 1999.
[22] K. L. Krause, V. Y. Shen, and H. D. Schwetman. Analysis of several task-scheduling algorithms
for a model of multiprogramming computer systems. Journal of the ACM, 22(4):522 -- 550, 1975.
[23] F. M. Liang. A lower bound for on-line bin packing. Information Processing Letters, 10(2):76 --
79, 1980.
[24] S. S. Seiden. On the online bin packing problem. Journal of the ACM, 49(5):640 -- 671, 2002.
[25] S. S. Seiden and R. van Stee. New bounds for multi-dimensional packing. Algorithmica,
36(3):261 -- 293, 2003.
[26] H. Shachnai and T. Tamir. Tight bounds for online class-constrained packing. Theoretical
Computer Science, 321(1):103 -- 123, 2004.
[27] H. Shachnai and T. Tamir. Polynomial time approximation schemes for class-constrained
packing problems. Journal of Scheduling, 4(6):313 -- 338, 2001.
[28] J. D. Ullman. The performance of a memory allocation algorithm. Technical Report 100,
Princeton University, Princeton, NJ, 1971.
[29] A. van Vliet. An improved lower bound for online bin packing algorithms.
Information
Processing Letters, 43(5):277 -- 284, 1992.
[30] E. C. Xavier and F. K. Miyazawa. The class constrained bin packing problem with applications
to video-on-demand. Theoretical Computer Science, 393(1-3):240 -- 259, 2008.
[31] A. C. C. Yao. New algorithms for bin packing. Journal of the ACM, 27:207 -- 227, 1980.
18
|
1802.09709 | 1 | 1802 | 2018-02-27T03:56:48 | Fully Dynamic Maximal Independent Set with Sublinear Update Time | [
"cs.DS"
] | A maximal independent set (MIS) can be maintained in an evolving $m$-edge graph by simply recomputing it from scratch in $O(m)$ time after each update. But can it be maintained in time sublinear in $m$ in fully dynamic graphs?
We answer this fundamental open question in the affirmative. We present a deterministic algorithm with amortized update time $O(\min\{\Delta,m^{3/4}\})$, where $\Delta$ is a fixed bound on the maximum degree in the graph and $m$ is the (dynamically changing) number of edges.
We further present a distributed implementation of our algorithm with $O(\min\{\Delta,m^{3/4}\})$ amortized message complexity, and $O(1)$ amortized round complexity and adjustment complexity (the number of vertices that change their output after each update). This strengthens a similar result by Censor-Hillel, Haramaty, and Karnin (PODC'16) that required an assumption of a non-adaptive oblivious adversary. | cs.DS | cs | Fully Dynamic Maximal Independent Set
with Sublinear Update Time
Sepehr Assadi∗
University of Pennsylvania
Krzysztof Onak†
IBM Research
Baruch Schieber‡
IBM Research
Shay Solomon§
IBM Research
Abstract
A maximal independent set (MIS) can be maintained in an evolving m-edge graph by simply
recomputing it from scratch in O(m) time after each update. But can it be maintained in time
sublinear in m in fully dynamic graphs?
We answer this fundamental open question in the affirmative. We present a deterministic
algorithm with amortized update time O(min{∆, m3/4}), where ∆ is a fixed bound on the
maximum degree in the graph and m is the (dynamically changing) number of edges.
We further present a distributed implementation of our algorithm with O(min{∆, m3/4})
amortized message complexity, and O(1) amortized round complexity and adjustment complex-
ity (the number of vertices that change their output after each update). This strengthens a
similar result by Censor-Hillel, Haramaty, and Karnin (PODC'16) that required an assumption
of a non-adaptive oblivious adversary.
8
1
0
2
b
e
F
7
2
]
S
D
.
s
c
[
1
v
9
0
7
9
0
.
2
0
8
1
:
v
i
X
r
a
∗[email protected] Supported in part by NSF grant CCF-1617851.
†[email protected]
‡[email protected]
§[email protected]
1
Introduction
Dynamic graph algorithms constitute an active area of research in theoretical computer science.
Their objective is to maintain a solution to a combinatorial problem in an input graph-for exam-
ple, a minimum spanning tree or maximal matching-under insertion and deletion of edges. The
research on dynamic graph algorithms addresses the natural question of whether one essentially
needs to recompute the solution from scratch after every update.
This question has been asked over the years for a wide range of problems such as connectivity [30,
32], minimum spanning tree [20,22,29,32,45], maximal matching [7,31,37,43], approximate matching
and vertex cover [10–12,14–16,25,31,37,38], shortest paths [1,8,9,19,21,26–28,34,42,44], and graph
coloring [3, 6, 13] (this is by no means a comprehensive summary of previous results).
Surprisingly however, almost no work has been done for the prominent problem of maintaining a
maximal independent set (MIS) in dynamic graphs. Indeed, the only previous result for this problem
that we are aware of is due to Censor-Hillel et al. [17], who developed a randomized algorithm for
this problem in distributed dynamic networks and left the sequential case (the main focus of this
paper) as a major open question. We note that implementing their distributed algorithm in the
sequential setting requires Ω(∆)1 update time in expectation, where ∆ is a fixed upper bound on
the degree of vertices in the graph and can be as large as Θ(m) in sparse graphs.
The maximal independent set problem is of fundamental importance in graph theory with
natural connections to a plethora of other basic problems, such as vertex cover, matching, vertex
coloring, and edge coloring (in fact, all these problems can be solved approximately by finding an
MIS, see, e.g., the paper of Linial [35]). As a result, this problem has been studied extensively in
different settings, in particular in parallel and distributed algorithms [2, 4, 5, 18, 23, 33, 35, 36, 39].
(We refer the interested reader to the papers of Barenboim et al. [5] and Ghaffari [23] for the story
of this problem in these settings and a comprehensive summary of previous work.)
In this paper, we concentrate on sequential algorithms for maintaining a maximal independent
set in a dynamic graph. Our results are also applicable to the dynamic distributed setting and
improve upon the previous work of Censor-Hillel et al. [17].
1.1 Problem Statement and Our Results
Recall that a maximal independent set (MIS) of an undirected graph, is a maximal collection of
vertices subject to the restriction that no pair of vertices in the collection are adjacent.
In the
maximal independent set problem, the goal is to compute an MIS of the input graph.
We study the fully dynamic variant of the maximal independent set problem in which the goal
is to maintain an MIS of a dynamic graph G, denoted by M := M(G), subject to a sequence of
edge insertions and deletions. When an edge change occurs, the goal is to maintain M in time
significantly faster than simply recomputing it from scratch. Our main result is the following:
Theorem 1. Starting from an empty graph on n fixed vertices, a maximal independent set can
be maintained deterministically over any sequence of edge insertions and deletions in O(m3/4)
amortized update time, where m denotes the dynamic number of edges.
As a warm-up to our main result in Theorem 1, we also present an extremely simple deterministic
algorithm for maintaining an MIS with O(∆) amortized update time, where ∆ is a fixed upper
bound on the maximum degree of the graph. Our algorithms can be combined together to achieve
a deterministic O(min (cid:8)∆, m3/4(cid:9)) amortized update time algorithm for maintaining an MIS in
dynamic graphs. This constitutes the first improvement on the update time required for this
1 It is not clear whether O(∆) time is also sufficient for this algorithm or not; see Section 6 of their paper.
1
problem in fully dynamic graphs over the naıve O(m) bound for all possible values of m. We now
elaborate more on the details of our algorithm in Theorem 1.
Deterministic Algorithm. An important feature of our algorithm in Theorem 1 is that it is
deterministic. The distinction between deterministic and randomized algorithms is particularly im-
portant in the dynamic setting as almost all existing randomized algorithms require the assumption
of a non-adaptive oblivious adversary who is not allowed to learn anything about the algorithm's
random bits. Alternately, this setting can be viewed as the requirement that the entire sequence of
updates be fixed in advance, in which case the adversary cannot use the solution maintained by the
algorithm in order to break its guarantees. While these assumptions can be naturally justified in
many settings, they can render randomized algorithms entirely unusable in certain scenarios (see,
e.g., [8, 10, 15] for more details).
As a result of this assumption, obtaining a deterministic algorithm for most dynamic problems
is considered a distinctively harder task compared to finding a randomized one. This is evident by
the polynomial gap between the update time of best known deterministic algorithms compared to
randomized ones for many dynamic problems. For example, a maximal matching can be maintained
in a fully dynamic graph with O(1) update time via a randomized algorithm [43], assuming a non-
adaptive oblivious adversary, while the best known deterministic algorithm for this problem requires
Θ(√m) update time [37] (see [13] for a similar situation for (∆ + 1)-coloring of vertices of a graph).
O(1)-Amortized Adjustment Complexity. An important performance measure of a dynamic
algorithm is its adjustment complexity (sometimes called recourse) that counts the number of
vertices (or edges) that need to be inserted to or deleted from the maintained solution after each
update (see, e.g. [3, 13, 17, 24]). For many natural graph problems such as maintaining a maximal
matching, constant worst-case adjustment complexity can be trivially achieved since one edge
update cannot ever necessitate more than a constant number of changes in the maintained solution.
This is, however, not the case for the MIS problem: by inserting an edge between two vertices
already in M, the adversary can force the algorithm to delete at least one end point of this edge
from M, which in turn forces the algorithm to pick all neighbors of this deleted vertex to ensure
maximality (this phenomena also highlights a major challenge in the treatment of this problem
compared to the maximal matching problem which we discuss further below).
Nevertheless, we prove that the adjustment complexity of our algorithm in Theorem 1 is O(1) on
average which is clearly optimal. Can we further improve our results to achieve an O(1) worst-case
adjustment complexity? We claim that this is indeed not possible by showing that the worst-case
adjustment complexity of any algorithm for maintaining an MIS is Ω(n), using a simple adaption
of an example proposed originally by [17] for proving a similar result in distributed settings when
vertex deletions are also allowed by the adversary (this follows seamlessly from the result in [17]
and is provided in Appendix A only for completeness).
Distributed Implementation. Finding a maximal independent set is one of the most studied
problems in distributed computing.
In the distributed computing model, there is a processor
on each vertex of the graph. Computation proceeds in synchronous rounds during which every
processor can communicate messages of size O(log n) with its neighbors (this corresponds to the
CONGEST model of distributed computation; see Section 5 for further details). In the dynamic
setting, both edges and vertices can be inserted to or deleted from the graph and the goal is to
update the solution in a small number of rounds of communication, with small communication cost
and adjustment complexity.
Our results in the sequential setting also imply a deterministic distributed algorithm for main-
taining an MIS in a dynamic network with O(1) amortized round complexity, O(1) amortized ad-
justment complexity, and O(min (cid:8)∆, m3/4(cid:9)) amortized message complexity per each update. This
2
result achieves an improved message complexity compared to the distributed algorithm of [17] with
asymptotically the same round and adjustment complexity (albeit in amortized sense as opposed
to in expectation; see Section 5). More importantly, our result is achieved via a deterministic
algorithm and does not require the assumption of a non-adaptive oblivious adversary. Similar
to [17], our algorithm can also be implemented in the asynchronous model, where there is no global
synchronization of communication between nodes. We elaborate more on this result in Section 5.
Maximal Independent Set vs. Maximal Matching. We conclude this section by comparing
the maximal independent set problem to the closely related problem of maintaining a maximal
matching2 in dynamic graphs. We discuss additional challenges that one encounters for the maximal
independent set problem.
In sharp contrast to the maximal independent set problem, maintaining maximal matchings in
dynamic graphs has been studied extensively, culminating in an O(√m) worst-case update time
deterministic algorithms [37] and O(1) expected update time randomized algorithm [43] (assuming
a non-adaptive oblivious adversary).
Maintaining an MIS in a dynamic graph seems inherently more complicated than maintaining
a maximal matching. One simple reason is that as argued before, a single update can only change
the status of O(1) edges/vertices in the maximal matching, while any algorithm can be forced to
make Ω(n) changes to the MIS for a single edge update in the worst case. As a result, a maximal
matching can be maintained with an O(∆) worst-case update time via a straightforward algorithm
(see, e.g. [31, 38]), while the analogous approach for MIS only results in O(m) update time.
Another, perhaps more fundamental difference between the two problems lies in their different
level of "locality." To adjust a maximal matching after an update, we only need to consider the
neighbors of the recently unmatched vertices (to find another unmatched vertex to match with),
while to fix an MIS, we need to consider the two-hop neighborhood of a recently removed vertex
from the MIS (to add to the MIS the neighbors of this vertex which themselves do not have another
neighbor in the MIS). We note that this difficulty is similar-in-spirit to the barrier for maintaining
a better than 2 approximate matching via excluding length-3 augmenting paths in dynamic graphs.
Currently, the best known algorithm for achieving a better than 2-approximation to matching in
dynamic graphs requires O(m1/4) update time [10, 11]. Achieving sub-polynomial in m update
time-even using randomness and assuming a non-adaptive oblivious adversary-remains a major
open problem in this area (we refer the interested reader to [15] for more details).
We emphasize that even for the seemingly easier problem of maximal matching, the best upper
bound on update time using a deterministic algorithm (the focus of our paper) is only O(√m) [37].
1.2 Overview of Our Techniques
O(∆)-Amortized Update Time. Consider the following simple algorithm for maintaining an
MIS M of a dynamic graph:
for each vertex, maintain the number of its neighbors in M in a
counter, and for each update in the graph or M, spend O(∆) time to update this counter for
the neighbors of the updated vertex. What is the complexity of this algorithm? Unfortunately,
as argued before, an update to the graph may inevitably result in an update of size Ω(n) to M.
Processing it may take Ω(n · ∆) time as we have to update all neighbors of every updated vertex.
However, all we need to handle this case is the following basic observation: while a single update can
force the algorithm to insert up to Ω(n) vertices to M, it can never force the algorithm to remove
more than one vertex from M. We therefore charge the O(∆) time needed to insert a vertex into
M (and there can be many such vertices per one update) to the time spent in a previous update
in which the same vertex was (the only one) removed from M. This allows us to argue that on
2A maximal matching in a graph G can be obtained by computing an MIS of the line graph of G.
3
average, we only spend O(∆) time per update.
O(m3/4)-Amortized Update Time. Achieving an o(∆) amortized update time however is dis-
tinctly more challenging. On the one hand, we cannot afford to update all neighbors of a vertex
after every change in the graph. On the other hand, we do not have enough time to iterate over all
neighbors of an updated vertex to even check whether or not they should be added to M and hence
need to maintain this information, which is a function of vertices in the two-hop neighborhood of
a vertex, explicitly for every vertex.
To bypass these challenges, we relax our requirement for knowing the status of all vertices in
the neighborhood of a vertex, and instead maintain the status of some vertices that are in the
two-hop neighborhood of a vertex. More concretely, we allow "high" degree vertices to not update
their "low" degree neighbors about their status (as the number of low degree neighbors can be very
large), while making every "low" degree vertex update not only all its neighbors but even some of
its neighbor's neighbors, using the extra time available to this vertex (as its degree is small). This
approach allows us to maintain a "noisy" version of the information described above. Note that this
information is not completely accurate as the status of some vertices in M would be unknown to
their neighbors and their neighbor's neighbors (in the actual algorithm, we use a more fine-grained
partition of vertices based on their degree into more than two classes, not only "high" and "low").
We now need to address a new challenge introduced by working with this "noisy" information:
we may decide that a vertex is ready to join M based on the information stored in the algorithm
and insert this vertex to M, only to find out that there are already some vertices in M adjacent
to this vertex. To handle this, we also relax the property of the basic algorithm above that only
allowed for deleting one vertex from M per each update. This allows us to insert multiple vertices
to M as long as a large portion (but not all) of their neighbors are known to be not in M. Then
we go back and delete a small number of "violating" vertices from M to make sure it is indeed an
independent set. Note that deleting those vertices may now require inserting a new set vertices in
their neighborhood to M to ensure maximality.
In order to be able to perform all those operations and recursively treat the newly deleted
vertices in a timely manner, we maintain the invariant that whenever we need to remove more
than one vertex from M, the number of inserted vertices leading to this case is much larger than
the number of removed vertices. This allows us to extend the simple charging scheme used in the
analysis of the basic algorithm above to this new algorithm and prove our upper bound on the
amortized update time of the algorithm.
We point out that despite the multiple challenges along the way that are described above, our
algorithm turned to be quite simple in hindsight. The main delicate matters are in the choice of
parameters and in the analysis. This in turn makes the implementation of our results in sequential
and distributed settings quite practical.
Organization. We introduce our notation and preliminaries in Section 2. We then present a
simple proof of the O(∆)-amortized update time algorithm in Section 3 as a warm-up to our
main result. Section 4 contains the proof of our main result in Theorem 1. The distributed
implementation of our result and a detailed comparison of our results with that of Censor-Hillel et
al. [17] appear in Section 5.
2 Preliminaries
Notation. We denote the static vertex set of the input graph by V . Let G = hG0, G1, . . .i be the
sequence of graphs that are given to the algorithm: initial graph G0 is empty and each graph Gt
is obtained from the previous graph Gt−1 by either inserting or deleting a single edge et = (ut, vt).
We use Gt(V, Et) to denote the graph at step t and define mt := Et. Finally, throughout the
4
paper, M denotes the maximal independent set maintained by the algorithm at every time step.
Greedy MIS Algorithm. Consider the following algorithm for computing an MIS of a given
graph: Fix an arbitrary ordering of the vertices in the graph, add the first vertex to the MIS,
remove all its neighbors from the list, and continue. This algorithm clearly computes an MIS of
the input graph. In the rest of the paper, we refer to this algorithm as the greedy MIS algorithm.
Fact 2.1. For an n-vertex graph G with maximum degree ∆, the greedy MIS algorithm computes
an MIS of size at least n/(∆ + 1).
3 Warm-Up: A Simple O(∆)-Update-Time Dynamic Algorithm
As a warm-up to our main result, we describe a straightforward algorithm for maintaining an MIS
M in a dynamic graph with O(∆) amortized update time, where ∆ is a fixed upper bound on
the maximum degree in the graph. For every vertex v in the graph, we simply maintain a counter
MISCounter[v], counting number of its neighbors in M. In the following, we consider updating M
and this counter after each edge update.
Let et = (ut, vt) be the updated edge. Suppose first that we delete this edge. In this case, ut
and vt cannot both be in M by definition of an independent set. Also, if none of them belong to
M, there is nothing to do. The interesting case is thus when exactly one of ut or vt belongs to M;
without loss of generality, we assume this vertex is ut. We first subtract one from MISCounter[vt]
(as it is no longer adjacent to ut). If MISCounter[vt] > 0 still, it means that vt is adjacent to some
vertex in M and hence we are done. Otherwise, we add vt to M and update the counter of all
its neighbors in O(∆) time. Clearly, this step takes O(∆) time in the worst case, after that M is
indeed an MIS.
Now suppose et was inserted to the graph. The only interesting case here is when both ut
and vt belong to M (we do not need to do anything in the remaining cases, other than perhaps
updating the neighbor list of ut and vt in O(1) time). To ensure that M remains an independent
set, we need to remove one of these vertices, say ut, from M. After this, to ensure the maximality,
we have to insert to M any neighbor of ut that can now join M. To do this, we first update the
MISCounter[·] of all neighbors of ut in O(∆) time. Next (using the updated counter), we iterate
over all neighbors of ut and for each one check if they can be inserted to M now or not. If so, we
add this new vertex to M and inform all its neighbors in O(∆) time to update their MISCounter[·].
It is easy to see that in this case, we spend O(k · ∆) time in the worst case, where k is the number
of vertices added to M.
The correctness of this algorithm is straightforward to verify. We now prove that the amortized
running time of the algorithm is O(∆). The crucial observation is that whenever we change M, we
may increase its size without any restriction, but we never decrease its size by more than one. We
use the following straightforward charging scheme.
Initially, we start with all vertices being in M as the original graph is empty. Whenever we
delete one vertex from M, we spend O(∆) time to handle this vertex (including updating its
neighbors and checking which ones can join M), and place O(∆) "extra budget" on this vertex
to be spent later. We use this budget when this vertex is being inserted to M again. Whenever
we want to bring this vertex back to M, we only need to spend this extra budget and hence the
O(∆) time spent for inserting this vertex back to M can be charged to the time spent for this
vertex when we removed it from M. This implies that the update time is O(∆) in average. We
can therefore conclude the following lemma.
Lemma 3.1. Starting from an empty graph on n vertices, a maximal independent set can be
maintained deterministically over any sequence of K edge insertions and deletions in O(K · ∆)
time where ∆ is a fixed bound on the maximum degree in the graph.
5
4 An O(m3/4)-Update-Time Dynamic Algorithm
We present our fully dynamic algorithm for maintaining a maximal independent set in this section
and prove Theorem 1. The following lemma is a somewhat weaker looking version of Theorem 1.
However, we prove next that this lemma is all we need to prove Theorem 1.
Lemma 4.1. Starting with any arbitrary graph on n vertices and m edges, a maximal independent
set M can be maintained deterministically over any sequence of K = Ω(m) edge insertions and
deletions in O(K · m3/4) time, as long as the number of edges remains within a factor 2 of m.
We first show that this lemma implies Theorem 1.
Proof of Theorem 1. For simplicity, we define m = 1 in case of empty graphs. We start from the
empty graph and run the algorithm in Lemma 4.1 until the number of edges mt in the graph differs
from m by a factor more than 2. This crucially implies that the total number of updates before
terminating the algorithm (the parameter K in Lemma 4.1), is Ω(m). As such, we can invoke
Lemma 4.1 to obtain an upper bound of O(m3/4) on the amortized update time of the algorithm
throughout these updates. We then update m = mt and start running the algorithm in Lemma 4.1
on the current graph using the new choice of m. Clearly, this results in an amortized update time
of O(m3/4) where m now denotes the number of dynamic edges in the graph. As the algorithm in
Lemma 4.1 always maintain an MIS of the underlying graph, we obtain the final result.
The rest of this section is devoted to the proof of Lemma 4.1. In the following, we first describe
the data structure maintained in the algorithm for storing the required information and its main
properties and then present our update algorithm.
4.1 The Data Structure
For every vertex v, we maintain the following information:
• neighbors[v]: a list of current neighbors of v in the graph.
• degree[v]: an estimate of the degree of v to within a factor of two.
• neighbors-degree[v]: a list containing degree[u] for every vertex u in neighbors[v].
• MIS-flag[v]: a boolean entry indicating whether or not v belongs to M.
• MIS-neighbors[v]: a counter denoting the size of a suitable subset of current neighbors of
v in M. Any vertex counted in MIS-neighbors[v] belongs to M but not all neighbors of v
in M are (necessarily) counted in MIS-neighbors[v] (see Invariant 1 for more detail).
• MIS-2hop-neighbors[v]: a list, containing for every vertex w in neighbors[v], a counter
that counts the size of a suitable subset of current neighbors of w in M. Any vertex counted
in MIS-2hop-neighbors[v][w] is also counted in MIS-neighbors[w] but not vice versa (see
Invariant 2 below for more detail).
Additionally, we maintain a partition (VHigh, VMed-High, VMed-Low, VLow) of the vertices into four
sets based on their current approximate degree, namely, degree[v]. In particular, v belongs to VHigh
iff degree[v] ≥ m3/4, to VMed-High iff m3/4 > degree[v] ≥ m1/2, to VMed-Low iff m1/2 > degree[v] ≥
m1/4, and to VLow iff degree[v] < m1/4. We refer to the vertices of VLow as the low-degree vertices.
Throughout, we assume that in any of the lists maintained for a vertex by the algorithm, we can
directly iterate over vertices of a particular subset in (VHigh, VMed-High, VMed-Low, VLow). (This can be
done, for example, by storing these lists as four separate linked lists, one per each such subset.)
6
The following invariant is concerned with the information we need from MIS-neighbors[v].
Invariant 1. For any vertex v ∈ V \VLow, MIS-neighbors[v] counts the number of all neighbors
of v in M. For any vertex v ∈ VLow, MIS-neighbors[v] counts the number of neighbors of v
that are in M but not in VHigh, i.e., are in M \ VHigh.
By Invariant 1, any vertex either knows the number of all its neighbors in M or is a low-degree
vertex and can iterate over all its neighbors in O(m1/4) time to count this number. Moreover, even
a low-degree vertex knows the number of its neighbors in M\VHigh. This is crucial for our algorithm
as in some cases, we need to iterate over many vertices that belong to VLow and decide if they can
join M and hence cannot spend O(m1/4) time per each vertex to determine this information. Note
that the information we obtain in this way is "noisy", as we ignore some neighbors of vertices in
VLow that are potentially in M. We shall address this problem using a post-processing step that
exploits the fact that the total number of ignored vertices, i.e., vertices in VHigh, is small.
The following invariant is concerned with the information we need from MIS-2hop-neighbors[v].
Invariant 2. For any v ∈ V and and every u ∈ neighbors[v]∩VLow, MIS-2hop-neighbors[v][u]
counts the number of vertices in VMed-Low ∪ VLow that belong to M and are neighbors of u (the
entry in MIS-2hop-neighbors[v][u] for any vertex u /∈ VLow is ⊥).
Invariant 2 allows us to infer some nontrivial information about the two-hop neighborhood of
any vertex. We use Invariant 2 to quickly determine which neighbors of a vertex v can be added to
M in case v is deleted from it. Similar to the one-hop information we obtain through maintaining
Invariant 1, the information we obtain in this way is also "noisy".
We show how to update the information per each vertex after a change in the topology or M.
Maintaining neighbors[v] under edge updates is straightforward. To maintain degree[v], each
vertex simply keeps a 2-approximation of its degree in degree[v]. Whenever the current actual
degree of v differs from degree[v] by more than a factor of two, v updates degree[v] to its actual
degree and informs all its neighbors u ∈ neighbors[v] to update neighbors-degree[u]. This
requires only O(1) amortized time.
The above information is a function of the underlying graph and not M. We also need to
update the information per each vertex that are functions of M whenever M changes. Maintaining
MIS-flag[v] is trivial for any vertex v, hence in the following we focus on the remaining two parts.
Once a vertex u changes its status in M, we apply the following algorithm to update the value
of MIS-neighbors[v] for every vertex v (we only need to update this for v ∈ neighbors[u]).
Algorithm UpdateNeighbors(u). An algorithm called whenever a vertex u enters or exists M
to update MIS-neighbors[v] for neighbors of u.
1. If u ∈ VHigh, update MIS-neighbors[v] for any vertex v ∈ neighbors[u] not in VLow accord-
ingly (i.e., add or subtract one depending on whether u joined or left M).
2. If u /∈ VHigh, update MIS-neighbors[v] for every vertex v ∈ neighbors[u].
It is immediate to see that by running UpdateNeighbors(u) in our main algorithm whenever
a vertex u is updated in M, we can maintain Invariant 1. Also each call to UpdateNeighbors(u)
takes O(m3/4) time in worst-case since in both cases of the algorithm, we only need to update
O(m3/4) vertices: (i) if u ∈ VHigh, the algorithm only updates the vertices in V \ VLow whose size
7
is O(m3/4), and (ii) if u /∈ VHigh, u only has O(m3/4) neighbors to update. We also point out that
MIS-neighbors[v] can be updated easily whenever an edge incident on (u, v) is inserted or deleted
in O(1) time by simply visiting MIS-flag[u] and updating MIS-neighbors[v] accordingly.
Now consider updating MIS-2hop-neighbors[·]. We use the following algorithm on a vertex u
that has changed its status in M to update MIS-2hop-neighbors[v] for every vertex v in the graph
(we only need to update this information for the two-hop neighborhood of u).
Algorithm UpdateTwoHopNeighbors(u). An algorithm called when a vertex u enters or exists
M to update MIS-2hop-neighbors[v] for the two-hop neighborhood of u.
1. If u ∈ VMed-Low ∪ VLow, for any vertex w ∈ neighbors[u]:
(a) If w belongs to VLow, iterate over all vertices v ∈ neighbors[w].
(b) For any such v, update MIS-2hop-neighbors[v][w] accordingly (i.e., add or subtract
one depending on whether u joined or left M).
Each call to UpdateTwoHopNeighbors(u) takes O(m3/4) time in the worst case. This is because
u only updates its neighbors if it has O(m1/2) neighbors as u should be in VMed-Low ∪ VLow and when
it updates its neighbors, it changes the counter of O(m1/4) vertices (as w should be in VLow). This
ensures that the running time of the algorithm is O(m1/2 · m1/4) = O(m3/4). Whenever an edge
(u, v) is updated in the graph, we can run a similar algorithm to update the two-hop neighborhood
of u and v in the same way in O(m3/4) time; we omit the details. It is also straightforward to verify
that by running UpdateTwoHopNeighbors(u) in our main algorithm whenever a vertex u is updated
in M, we preserve Invariant 2.
Finally, recall that we also need a preprocessing step that given a graph G initializes this
data structure. We can implement this step by first initializing all non-MIS-related information
in this data structure in O(m) time (we do not need to handle isolated vertices at this point).
Next, we run the greedy MIS algorithm to compute an MIS M of this graph in O(m) time (again
only on non-isolated vertices). Finally, we update the information for every vertex in M using
the two procedures above which takes O(m · m3/4) time in total. (We note that a more efficient
implementation for this initial stage is possible.) As K = Ω(m) in Lemma 4.1, this (one time only)
initialization cost is within the bounds stated in the lemma statement.
We summarize the results in this section in the following two lemmas.
Lemma 4.2. After updating any single edge, the data structure stored in the algorithm can be
updated in O(m3/4) amortized time. Moreover, Invariants 1 and 2 hold after this update.
Lemma 4.3. After updating any single vertex in M, the data structure stored in the algorithm can
be updated in O(m3/4) worst case time. Moreover, Invariants 1 and 2 hold after this update.
4.2 The Update Algorithm
The update algorithm is applied following edge insertions and deletions to and from the graph.
After any edge update, the algorithm updates the data structure and M.
In order to do the
latter task, the algorithm may need to remove and/or insert multiple vertices from and to M.
Since we already argued that maintaining the data structure requires O(m3/4) amortized time (by
Lemma 4.2), from now on, without loss of generality, we only measure the time needed to fix M
after any edge update and ignore the additive term needed to update the data structure. The
following is the core invariant that we aim to maintain in our algorithm.
8
Invariant 3 (Core Invariant). Following every edge update, the set M maintained by the
algorithm is an MIS of the input graph. Moreover,
(i) if only a single vertex leaves M, then there is no restriction on the number of vertices
joining M (which could be zero).
(ii) if at least two vertices leave M, then at least twice as many vertices join M.
In either case, the total time spent by the algorithm to fix M for an edge update is at most an
O(m3/4) factor larger than the total number of vertices leaving and joining M.
Before showing how to maintain Invariant 3, we present the proof of Lemma 4.1 using this invariant.
Proof. The main idea behind the proof is as follows. By Invariant 3, after each step, the size of M
either decreases by at most one, or it will increase. At the same time, M cannot grow more than
n, the number of vertices in the graph. It then follows that the average number of changes to M
per each update is O(1). As we only spend O(m3/4) per each update, we obtain the final result.
We now present the formal proof using the following charging scheme.
Recall that we compute an MIS M of the initial graph in the preprocessing step and that the
initialization phase takes O(m · m3/4) time in total. We place O(m3/4) "extra budget" on vertices
in the initial graph that do not belong to M to be spent later when these vertices are inserted to
M. As the number of such vertices is O(m), this extra budget can be charged to the time spent in
the initialization phase. Note that at this point, an extra budget is allocated to any vertex not in
M and we maintain this throughout the algorithm.
Whenever an update results in only a single vertex leaving M (corresponding to Part (i) of
Invariant 3), we spend O(m3/4) time to handle this vertex and additionally place O(m3/4) budget
on this vertex and then for the vertices inserted to M, we simply use the extra budget allocated
to these vertices before to charge for the O(m3/4) time needed to handle each. If an update results
in removing k > 1 vertices from M, we know that at least 2 · k vertices would be added to M
after this update (corresponding to Part (ii) of Invariant 3). In this case, we use the O(m3/4) extra
budget on these (at least) 2 · k vertices that are joining M to charge for the time needed to insert
these vertices to M, remove the k initial vertices from M, and place O(m3/4) extra budget on
every removed vertex. As a result, this type of updates can be handled free of charge. Finally, if
an update only involves inserting some vertices to M, we simply use the budgets on these vertices
to handle them free of charge. This finalizes the proof of Lemma 4.1.
We point out that using the above charging scheme, we can also argue that the average number
of changes to M is O(1) in each update.
Fix a time step t and suppose the invariant holds up until this time step. Let et = (ut, vt) be
the edge updated at this time step. In the remainder of this section, we describe one round of the
update algorithm to handle this single edge update and preserve Invariant 3.
4.2.1 Edge Deletions
We start with the easier case of deleting an edge et = (ut, vt).
Case 1: Neither ut nor vt belong to M. In this case, there is nothing to do.
Case 2: ut belongs to M but not vt (or vice versa). After deleting the edge et, it is possible that vt
may need to join M as well. We first check whether MIS-neighbors[vt] = 0. If not, there is nothing
else to do as vt is still adjacent to some vertex in M. Otherwise, we need to ensure that vt does not
have any neighbor in M (outside those vertices counted in MIS-neighbors[vt]). If vt ∈ V \ VLow,
9
by Invariant 1, MIS-neighbors[vt] counts all neighbors of vt and hence there is nothing more to
check. If vt ∈ VLow, we can go over all the O(m1/4) vertices in the neighborhood of vt and check
whether vt has a neighbor in M or not. This only takes O(m1/4) time in the worst case. Again, if
we find a neighbor in M there is nothing else to do. Otherwise, we add vt to M and update the
data structure which takes O(m3/4) time in the worst case by Lemma 4.3. After this step, M is
again a valid MIS and hence Invariant 3 is preserved as we only spent O(m3/4) time and inserted
at most one vertex to M without deleting any vertex from it.
Case 3: Both ut and vt belong to M. This case is not possible in the first place by Invariant 3 as
otherwise M maintained by the algorithm before this edge update was not an MIS.
4.2.2 Edge Insertions
We now consider the by far more challenging case of edge insertions where we concentrate bulk of
our efforts. It is immediate to see that the only time we need to handle an edge insertion is when
the inserted edge et = (ut, vt) connects two vertices already in M (there is nothing to do in the
remaining cases). Hence, in the following, we assume both ut and vt belong M.
To ensure that M is an independent set, we first need to remove one of ut or vt from it and
then potentially insert some of the neighbors of the deleted vertex to M to ensure its maximality.
Let ut be the deleted vertex (the choice of which vertex to delete is arbitrary). After deleting ut,
we update the algorithm's data structure in O(m3/4) time by Lemma 4.3.
Let L := neighbors[ut] ∩ VLow denote the set of low degree neighbors of ut. We first show that
one can easily handle all neighbors of ut which are not in L. To do so, we can iterate over these
vertices as there are O(m3/4) of them and for any vertex w, by Invariant 1, we know whether w
can be added M or not by simply checking MIS-neighbors[w]. Hence, we can add the necessary
vertices to M and spend O(m3/4) time for each inserted one using Lemma 4.3. As such, we spend
O(m3/4) time for iterating the vertices which did not join M and k· O(m3/4) time for the k vertices
that joined M. Hence, Invariant 3 is preserved after this step.
We now consider the challenging case of updating the neighbors of ut that belong to L. As
the number of such vertices is potentially very large, we cannot iterate over all of them anymore.
Define the following subsets of L:
• LMIS ⊆ L: the set of vertices in L that do not have any neighbor in M.
• L1-hop ⊇ LMIS: all vertices w ∈ L where MIS-neighbors[w] = 0 i.e., our algorithm did not count
any neighbor for them in M. Recall that MIS-neighbors[w] does not count all neighbors of w
in M; it is missing the vertices in VHigh by Invariant 1.
• L2-hop ⊇ L1-hop ⊇ LMIS: all vertices w ∈ L, where MIS-2hop-neighbors[ut][w] = 0. Again, recall
that MIS-2hop-neighbors[ut][wt] does not count all neighbors of w ∈ MIS-neighbors[w] (and
consequently in M); it misses the vertices in VMed-High in MIS-neighbors[w] (and additionally
VHigh in M) by Invariant 2.
Let ℓMIS := LMIS, ℓ1-hop := L1-hop and ℓ2-hop := L2-hop, where ℓMIS ≤ ℓ1-hop ≤ ℓ2-hop.
Our algorithm does not know the sets LMIS and L1-hop or even their sizes. However, the up-
date algorithm knows the value of ℓ2-hop and has access to vertices in L2-hop through the list
MIS-2hop-neighbors[ut] and can iterate over them in O(1) time per each vertex in L2-hop (notice
that even this can be potentially too time consuming as size of this list can be too large). We
consider different cases based on the value of these parameters.
Case 1: when ℓ2-hop is small, i.e., ℓ2-hop ≤ 4· m3/4. In this case, we iterate over vertices w ∈ L2-hop
in O(ℓ2-hop) = O(m3/4) time and check whether MIS-neighbors[w] = 0 or not. This allows us to
compute the set L1-hop and ℓ1-hop as well. We further distinguish between two cases.
10
Case 1-a: when ℓ1-hop is very small, i.e., ℓ1-hop ≤ 4 · m1/2. We iterate over vertices w ∈ L1-hop
and for each vertex, spend O(m1/4) time to go over all its neighbors and decide whether w has any
neighbor in M or not (degree of w is O(m1/4) since it belongs to VLow). Hence, in this case, we
can obtain the set LMIS fully in O(m3/4) time in total.
We then iterate over vertices in LMIS, insert each one greedily to M, and update the data
structure in O(m3/4) time using Lemma 4.3. It is possible that some vertices in LMIS are adjacent
to each other and hence before inserting any vertex w, we first need to check MIS-neighbors[w] to
make sure it is zero still (by Invariant 1 and since all vertices in LMIS belong to VLow, any vertex
added to M here would update MIS-neighbors[w] for any neighbor w). Hence, in this case, we
spend O(m3/4) time for each vertex inserted to M and did not delete any vertex from it. Therefore,
Invariant 3 is preserved after the edge update in this case.
Case 1-b: when ℓ1-hop is not very small, i.e., ℓ1-hop > 4 · m1/2. In this case, we cannot afford
to compute LMIS explicitly. Rather, we simply add the vertices in L1-hop to M directly, without
considering whether they are adjacent to vertices already in M or not at all (although we check
that they are not adjacent to the previously inserted vertices from L1-hop). As a result, it is possible
that after this process, M is not an independent set of the graph anymore. To fix this, we perform a
post processing step in which we delete some vertices from M to ensure that the remaining vertices
indeed form an MIS of the original graph.
Concretely, we go over vertices in L1-hop and insert each to M if none of its neighbors have been
added to M in this step, and then invoke Lemma 4.3 to update the algorithm's data structure.
Since in this step, we are only adding vertices that are in VLow, we can check in O(1) time whether
a vertex has a neighbor in M (that has been added in this step) or not by Invariant 1. This step
clearly takes O(m3/4) time per each vertex inserted to the MIS.
At this point, it is possible that there are some vertices in M which are adjacent to the newly
inserted vertices. By Invariant 1, we know that these vertices can only belong to VHigh and hence
there are at most m1/4 of them. We iterate over all vertices in VHigh and check whether they
have a neighbor in M (by Invariant 1, we stored this information for these vertices) and mark all
such vertices. Next, we remove all these marked vertices from M simultaneously and update the
algorithm's state by Lemma 4.3. We are not done yet though because after removing these vertices,
it is possible that we may need to bring some of their neighbors back to M. We solve this problem
recursively using the same update algorithm by treating these marked vertices the same as ut.
We argue that Invariant 3 is preserved. As the degree of vertices in L1-hop is bounded by m1/4,
(m1/4+1) ≥ 2 · m1/4 (by Fact 2.1 and
the number of vertices added to M in this part is at least
the assumption on ℓ1-hop in this case). On the other hand, the number of vertices removed from
M is at most equal to size of VHigh which is m1/4. As a result, in this specific step, the number of
vertices inserted to M is at least twice as many as the vertices removed from it. For any vertex
inserted or deleted from M also, we spent O(m3/4) time. As we are performing the recursive step
using the same algorithm, we can argue inductively that for any vertex deleted in those recursive
calls, at least twice as many vertices would be added to M and that the total running time would
be proportional to the number of vertices added or removed from M times O(m3/4). We point out
that any recursive call that leads to another one in this algorithm necessarily increase the number
of vertices in M and hence the algorithm does indeed terminate (see also case 2 ).
Case 2: when ℓ2-hop is not small, i.e., ℓ2-hop > 4· m3/4. We use a similar strategy as case 1-b here
as well. We iterate over all vertices in L2-hop, greedily add each vertex to M as long as this vertex is
not adjacent to any of the newly added vertices (which can be checked in O(1) time by Invariant 1),
and update the data structure using Lemma 4.3. As the maximum degree of vertices in L2-hop is at
most m1/4, we add at least
m1/4+1 > 2· m1/2 vertices to M by Fact 2.1. By Invariant 2, if a vertex
ℓ2-hop
ℓ1-hop
11
belongs to L2-hop, the only neighbors of this vertex in M belong to VHigh or VMed-High and hence has
degree at least m1/2. We go over these vertices next and mark them. Then, we remove all of them
from M simultaneously and update the algorithm by Lemma 4.3. Similar to case 1-b, we now also
have to consider bringing some of the neighbors of these vertices to M which is handled recursively
exactly the same way as in case 1-b.
We first analyze the time complexity of this step. Iterating over L2-hop takes O(ℓ2-hop) = O(m)
time and since we are inserting at least m1/2 vertices from L2-hop to M, we can charge the time
needed for this step to the time allowed for inserting these vertices to M. Moreover, we inserted
at least 2 · m1/2 vertices to M and would remove at most m1/2 vertices after considering violating
vertices in VHigh and VMed-High. Hence, number of inserted vertices is at least twice the number of
removed ones at this step. We can also argue inductively that this property hold for each recursive
call similar to the case 1-b. This finalizes the proof of this case.
To conclude, we proved that Invariant 3 is preserved after any edge insertion or deletion in the
algorithm, which finalizes the proof of Lemma 4.1.
5 Maximal Independent Set in Dynamic Distributed Networks
We consider the CONGEST model of distributed computation (cf. [41]) which captures the essence
of both spatial locality and congestion. The network is modeled by an undirected graph G(V, E)
where the vertex-set is V , and E corresponds to both the edge-set in the current graph and also
the vertex pairs that can directly communicate with each other. We assume a synchronous com-
munication model, where time is divided into rounds and in each round, each vertex can send a
message of size O(log n) bits to any of its neighbors, where n = V . The goal is to maintain an
MIS M in G in a way that each vertex is able to output whether or not it belongs to M.
We focus on dynamically changing networks where both edges and vertices can be inserted
to or deleted from the network. For deletions, we consider graceful deletions where the deleted
vertex/edge may be used for passing messages between its neighbors (endpoints), and is only
deleted completely once the network is stable again. After each change, the vertices communicate
with each other to adjust their outputs, namely make the network stable again. We make the
standard assumption that the changes occur in large enough time gaps, and hence the network
is always stable before the next change occurs (see, e.g., [17, 40]). We further assume that each
change in the network is indexed and vertices affected by this change know how many updates have
happened before3.
There are three complexity measures for the algorithms in this model. The first is the so-called
adjustment complexity, which measures the number of vertices that change their output as a result
of a recent topology change. The second is the round complexity, the number of rounds required
for the network to become stable again after each update. The third is the message complexity,
measuring the total number of O(log n)-length messages communicated by the algorithm.
Our main result in this section is an implementation of Theorem 1 in this distributed setting
for maintaining an MIS in a dynamically changing network.
Theorem 2. Starting from an empty distributed network on n vertices, a maximal independent set
can be maintained deterministically in a distributed fashion (under the CONGEST communication
model) over any sequence of vertex/edge insertions and (graceful) deletions with (i) O(1) amor-
tized adjustment complexity, (ii) O(1) amortized round complexity, and (iii) O(m3/4) amortized
message complexity. Here, m denotes the number of dynamic edges.
3This is only needed by our algorithm in Theorem 2 to have an approximation of the number of edges in the
graph, which is a global quantity and cannot be maintained by each vertex locally
12
The algorithm in Lemma 3.1 can also be trivially implemented in this distributed setting,
resulting in an extremely simple deterministic distributed algorithm for maintaining an MIS of a
dynamically changing graph in O(1) amortized adjustment complexity and round complexity, and
O(∆) amortized message complexity. As argued before, this simple algorithm already strengthens
the previous randomized algorithm of Censor-Hillel et al. [17] by virtue of being deterministic and
not requiring an assumption of a non-adaptive oblivious adversary. In the following, we compare
our results in the distributed setting with those of [17].
Amortized vs in Expectation Guarantee. The guarantees on the complexity measures pro-
vided by our deterministic algorithms in this setting are amortized, while the randomized algorithm
in [17] achieves its bound in expectation which may be considered somewhat stronger than our guar-
antee. To achieve this guarantee however, the algorithm in [17], besides using randomization, also
assumes a non-adaptive oblivious adversary. An adaptive adversary (the assumption supported by
all our algorithms in this paper) can force the algorithm in [17] to adjust the MIS by Ω(n) vertices
in every round, which in turn blows up all the complexity measures in [17] by a factor of Ω(n).
It is also worth mentioning that the guarantee achieved by [17] only holds in expectation and not
with high probability and for a fundamental reason: It was shown in [17] that for every value of
k, there exists an instance for which at least Ω(k) adjustments are needed for any algorithm with
probability at least 1/k (see Section 1.1 of their paper).
Broadcast vs Unicast. The communication in algorithm of [17] in each round is O(1) broadcast
messages in expectation that requires only O(1) bits on every edge (i.e., each vertex communicates
the same O(1) bits to every one of its neighbors). As such, the total communication at every round
of this algorithm is O(∆) bits in expectation. Our amortized O(∆)-message complexity algorithm
(distributed implementation of Lemma 3.1) also works with the same guarantee:
indeed, every
vertex simply needs to send O(1) bits to all its neighbors in a broadcast manner so that their
neighbors know whether to add or subtract the contribution of this vertex to or from their counter.
This is however not the case for our main algorithm in Theorem 2 which requires a processor to
communicate differently to its neighbor over each edge (in general, one cannot hope to achieve
o(∆) communication with only broadcast messages). Additionally, this algorithm now requires to
communicate O(log n) bits (as opposed to O(1) in the previous two algorithms) over every edge.
This is mainly due to the fact that in this new algorithm we need to communicate with vertices
which are at distance 2 of the current vertex and hence we need to carry the ID of original senders
in the messages also.
Graceful vs Abrupt Deletions. A stronger notion of deletion in the dynamic setting is abrupt
deletion in which the neighbors of the deleted vertex/edge simply discover that this vertex/edge is
being deleted and the deleted vertex/edge cannot be used for communication anymore right after
the deletion happens. Censor-Hillel et al. [17] also extend their result to this more general setting
and achieved the same guarantees except for message complexity of abrupt deletion of a node which
is now O(min{log n, ∆}) broadcasts as opposed to O(1). We do not consider this model explicitly.
However, it is straightforward to verify that our amortized O(∆)-message complexity algorithm
(distributed implementation of Lemma 3.1) works in this more general setting with virtually no
change and even still achieves amortized O(1) broadcast per abrupt deletion of a vertex as well.
We believe that our main algorithm in Theorem 2 should also work in this more general setting
with proper modifications but we did not prove this formally.
Synchronous vs Asynchronous Communication. We focused only on the synchronous com-
munication in this paper. Censor-Hillel [17] also considered the asynchronous model of communi-
cation and showed that their algorithm holds in this model as well, albeit with a weaker guarantee
13
on its message complexity. Our algorithms can be modified to work in an asynchronous model as
well, as at each stage of the algorithm we can identify a (different) local "coordinator" that can be
used to synchronize the operations with an added overhead that is within a constant multiplicative
of the synchronous complexity (as per each update only vertices within two-hop neighborhood of
a vertex need to communicate with each other in our algorithm); we omit the details but refer the
reader to Section 5.2.2 for more information on the use of a local coordinator in our algorithms.
We now turn to proving Theorem 2, using the following lemma the same way we used Lemma 4.1
in the proof of Theorem 1.
Lemma 5.1. Starting with any arbitrary graph on n vertices and m edges, a maximal independent
set M can be maintained deterministically in a distributed fashion (under the CONGEST commu-
nication model) over any sequence of K = Ω(m) vertex/edge insertions and (graceful) deletions as
long as the number of edges in the graph remains within a factor 2 of m. The algorithm:
(i) makes O(K) adjustment to M in total, i.e., has O(1) amortized adjustment complexity,
(ii) requires O(K) rounds in total, i.e., has O(1) amortized round complexity, and
(iii) communicates O(K·m3/4) messages in total, i.e., has O(m3/4) amortized message complexity.
The algorithm in Lemma 5.1 is a simple implementation of our sequential dynamic algorithm
in Lemma 4.1. In the following, we first adapt the data structures introduced in Section 4.1 to the
distributed setting. We then show that with proper adjustments, the (sequential) update algorithm
in Section 4.2 can also be used in the CONGEST model and prove Theorem 2.
5.1 The Data Structure
We store the same exact information in Section 4.1 per each vertex here as well and maintain Invari-
ants 1 and 2. We first prove that the two procedures UpdateNeighbors and UpdateTwoHopNeighbors
can both be implemented in constant rounds and O(m3/4) messages in total. In particular,
Lemma 5.2. For any vertex u ∈ V ,
(i) UpdateNeighbors(u) operation requires spending 1 round and m3/4 messages in total, and
(ii) UpdateTwoHopNeighbors(u) operation requires 2 rounds and 2 · m3/4 messages in total.
Proof. Part (i). If u ∈ VHigh, it only needs to send a message to its neighbors in V \ VLow and
inform them on the status of u (whether it is inserted to or deleted from M), which requires only
1 round (as they are all neighbors to u) and m3/4 messages as V \ VLow ≤ m/m1/4 = m3/4. If
u /∈ VHigh, it would update all its neighbors again in 1 round and m3/4 messages as the latter is an
upper bound on number of its neighbors.
Part (ii). If u /∈ VLow ∪ VMed-Low there is nothing to do. Otherwise, u needs to send a message
to all its (at most m1/2) neighbors that belong to VLow and ask them to relay this information to
their neighbors. These vertices can then spend another round to inform all their (at most m1/4)
neighbors about the status of u. This takes 2 rounds and m1/2 + m1/2 · m1/4 < 2 · m3/4 messages
in total.
Lemma 5.2 ensures that Invariants 1 and 2 (the only MIS-related information stored for u beside
MIS-flag[u] that can be trivially updated) are preserved after any change in M within a constant
number of rounds and O(m3/4) messages.
In the following, we briefly describe how to update the information stored for vertices per each
topology change in the graph.
14
Vertex Updates. Let u be the updated vertex. In case of vertex insertion, we simply initial-
ize the data structures at u and we are almost done as the neighbors of u are already informed
about u being inserted to the graph and hence can update their information locally. We only need
to send degree[u] to all the neighbors (the time needed for this can be charged to the initializa-
tion cost of this algorithm). Now suppose u is being deleted. We can update neighbors[v] and
neighbors-degree[v] for any neighbor v of u without any communication as they are informed
that u is deleted. We can also run UpdateNeighbors(u) (virtually) with no communication as this
procedure only informs the neighbors of u that this vertex is being deleted from M and by knowing
that u has left the graph, any vertex v in the neighborhood of u can update MIS-neighbors[v]
accordingly. Finally, we can also run UpdateTwoHopNeighbors(u) with only 1 round of communi-
cation and m3/4 messages (see Lemma 5.2) by relaying the information from the neighbors of u
(which are informed about u leaving the graph) to their neighbors.
Edge Updates. These updates are handled exactly the same as in our sequential algorithm.
Let (u, v) be the updated edge. Nertices u and v can update all information except for updat-
ing MIS-2hop-neighbors[·] (in particular neighbors-degree[·] can be updated by the procedure
described in Section 4.1 with O(1) worst-case round complexity and O(1) amortized message com-
plexity). To do the latter task, vertex u (resp. v) can simulate UpdateTwoHopNeighbors(v) (resp.
UpdateTwoHopNeighbors(u)) as described above, which takes m3/4 messages and 1 round.
We hence showed that after each change in the topology, all the information stored for vertices
can be updated in O(1) rounds and O(m3/4) amortized messages.
5.2 The Distributed Algorithm
We design a distributed algorithm for updating M in the network in the spirit of our update
algorithm in Section 4.2. The algorithm is a simple adaption of our sequential algorithm to this
dynamic model. For every update, we first perform the steps in the previous section to update the
information on every vertex in the graph and then make the network stable again by adjusting M.
Throughout, we aim to maintain the following invariant which is the direct analogue of Invari-
ant 3 in the dynamic setting.
Invariant 4. Following every vertex/edge update, the set M maintained by the algorithm is
an MIS of the input graph. Moreover,
(i) if only a single vertex leaves M then there is no restriction on the number of vertices
joining M (which could be zero).
(ii) if at least two vertices leave M, then at least twice as many vertices are added to M.
In either case, the worst case number of rounds and messages spent by the algorithm for any
update is within, respectively, an O(1) and an O(m3/4) factor of the total number of vertices
leaving and joining M.
Using the same exact argument as in the proof of Lemma 4.1, maintaining Invariant 4 ensures
that the amortized adjustment complexity and amortized round complexity of the algorithm is O(1)
and its amortized message complexity is O(m3/4). Hence, to prove Lemma 5.1, it suffices to prove
that Invariant 4 is preserved after every update. We consider different cases based on insertion and
deletion of edges and vertices.
5.2.1 Edge Deletions
Suppose we delete the edge (u, v). We only consider the case that u belongs to M and v is not; the
remaining cases are either symmetric to this one or need no update in M (see Section 4.2.1). If v
15
is not in VLow, by Invariant 1, it knows all its neighbors in M and can decide whether to join M or
not to locally; if it enters M, it can update the network in O(1) rounds and O(m3/4) messages by
Lemma 5.2. If v is in VLow, it first sends a message to all its O(m1/4) neighbors and ask for their
status to which its neighbors reply whether they belong to M or not. This only takes 2 rounds
and O(m1/4) communication and then v can decide again whether to join M or not to. Note that
this part of the result holds even with abrupt deletions.
5.2.2 Edge Insertions
Suppose we insert the edge (u, v). We only consider the case when both u and v belong to M;
the remaining cases need no update in M (see Section 4.2.2). Remember that in Section 4.2.2, we
needed to handle these updates in three separate cases. While the algorithm and analysis in each
case is different, the procedures needed to carry the information around the network are essentially
the same among these cases and hence in the following, for simplicity, we only consider one of the
main cases, namely case 1-b (see Section 4.2.2 for definition of this case). The algorithm in the
remaining cases can be adapted to this setting in the same exact way.
Recall that in this case, the vertex u is deleted from M and moreover u knows the set L1-hop
entirely, which is of size O(m3/4). The general approach is to make u a "coordinator" for running
the update algorithm in Section 4.2.2 by communicating with its two-hop neighborhood and gather
the necessary information to run the sequential update algorithm.
Vertex u first sends a message to all its neighbors in L1-hop and asks for their status to which
they respond whether or not they belong to M. This takes 2 rounds and O(m3/4) messages. Next,
u informs one of its neighbors that it can join M and this new vertex updates its status and the
information in the graph which takes O(1) rounds and O(m3/4) messages by Lemma 5.2. After
this, u again sends a message to all its neighbors in L1-hop and asks for their status in M to which
they respond whether they belong to M or whether one of their neighbors in L1-hop has been added
to M in this step. Then, again, u informs one of its neighbors (if such exists) that it can join
M, and continues. This way, we only spend O(1) rounds and O(m3/4) communication per each
vertex entering M in addition to O(1) rounds and O(m3/4) communication for communicating with
neighbors of u that would not join M eventually.
After processing the list L1-hop, we also need to delete from M, the set of vertices in VHigh that are
now incident to vertices in L1-hop that just joined M. Note that such vertices are necessarily in the
two-hop neighborhood of u and hence u can communicate with them (which are only O(m1/4) many)
in O(1) rounds and use the above idea to implement the same update algorithm in Section 4.2.2 in
this model. This allows us to preserve Invariant 4 by the same exact analysis in Section 4.2.2.
5.2.3 Vertex Deletions
This case is essentially equivalent to the edge insertion case discussed above. Since we have a
graceful deletion, we can treat the deleted vertex the same way as in Section 5.2.2 by deleting it
from M (if it belonged to it) and using it as the "coordinator" to implement the process described
in Section 5.2.2.
5.2.4 Vertex Insertions
The only thing we need to do in this case is to check whether we need to add this new vertex to
M or not. If this vertex is not in VLow, it already knows this information and hence can decide
whether or not to join M; after that we are done. Otherwise, if the vertex belongs to VLow, it sends
a message to all its O(m1/4) neighbors and ask for their status in M, and use that to decide about
joining M. In either case, we only need O(1) rounds and O(m1/4) total communication. After
this, we update the neighbors using first part of Lemma 5.2 in O(m3/4) communication and O(1)
16
rounds.
To conclude, we showed that Invariant 4 is preserved after any edge or vertex insertion or
deletion by the distributed algorithm, hence proving Lemma 5.1. We are now ready to prove
Theorem 2.
Proof of Theorem 2. The proof is identical to the proof of Theorem 1. The only difference is that
in this distributed setting, we are not able to maintain the exact number of edges in the graph
in a distributed fashion across all vertices. However, recall that we assumed vertices affected by
an update in the topology know the index of this update, i.e., how many updates have happened
before this one. Hence, whenever the number of updates reaches Ω(m), any vertex that knows
this information sends a message to all its neighbors to terminate the process which would then be
broadcast across the whole graph. This takes O(m) rounds and O(m) communication and can be
charged to the total number of updates, i.e., Ω(m) in this step. Hence, the vertices can initialize
their data structure using the new choice of m and continue the distributed algorithm in Lemma 5.1.
References
[1] I. Abraham, S. Chechik, and S. Krinninger. Fully dynamic all-pairs shortest paths with worst-
In Proceedings of the 28th Annual ACM-SIAM Symposium on
case update-time revisited.
Discrete Algorithms, SODA 2017, Barcelona, Spain, January 16-19, 2017, pages 440–452,
2017.
[2] N. Alon, L. Babai, and A. Itai. A fast and simple randomized parallel algorithm for the
maximal independent set problem. J. Algorithms, 7(4):567–583, 1986.
[3] L. Barba, J. Cardinal, M. Korman, S. Langerman, A. van Renssen, M. Roeloffzen, and S. Ver-
donschot. Dynamic graph coloring. In Proceedings of the 15th International Symposium on
Algorithms and Data Structures, WADS 2017, St. John's, NL, Canada, July 31 - August 2,
2017, pages 97–108, 2017.
[4] L. Barenboim, M. Elkin, and F. Kuhn. Distributed (∆ + 1)-coloring in linear (in ∆) time.
SIAM J. Comput., 43(1):72–95, 2014.
[5] L. Barenboim, M. Elkin, S. Pettie, and J. Schneider. The locality of distributed symmetry
breaking. J. ACM, 63(3):20:1–20:45, 2016.
[6] L. Barenboim and T. Maimon. Fully-dynamic graph algorithms with sublinear time inspired
by distributed computing. In Proceedings of the International Conference on Computational
Science, ICCS 2017, Zurich, Switzerland, June 12-14, 2017, pages 89–98, 2017.
[7] S. Baswana, M. Gupta, and S. Sen. Fully dynamic maximal matching in O(log n) update
time.
In Proceedings of the 52nd IEEE Annual Symposium on Foundations of Computer
Science, FOCS 2011, Palm Springs, CA, October 23-25, 2011, pages 383–392, 2011 (see also
SICOMP'15 version, and subsequent erratum).
[8] A. Bernstein and S. Chechik. Deterministic decremental single source shortest paths: beyond
the O(mn) bound. In Proceedings of the 48th Annual ACM SIGACT Symposium on Theory
of Computing, STOC 2016, Cambridge, MA, USA, June 18-21, 2016, pages 389–397, 2016.
17
[9] A. Bernstein and L. Roditty.
Improved dynamic algorithms for maintaining approximate
shortest paths under deletions. In Proceedings of the 22nd Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA 2011, San Francisco, CA, USA, January 23-25, 2011, pages
1355–1365, 2011.
[10] A. Bernstein and C. Stein. Fully dynamic matching in bipartite graphs. In Proc. 42nd ICALP,
pages 167–179, 2015.
[11] A. Bernstein and C. Stein. Faster fully dynamic matchings with small approximation ratios. In
Proceedings of the 27th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2016,
Arlington, VA, USA, January 10-12, 2016, pages 692–711, 2016.
[12] S. Bhattacharya, D. Chakrabarty, and M. Henzinger. Deterministic fully dynamic approximate
vertex cover and fractional matching in O(1) amortized update time. In Proceedings of the 19th
International Conference on Integer Programming and Combinatorial Optimization, IPCO
2017, Waterloo, ON, Canada, June 26-28, 2017, pages 86–98, 2017.
[13] S. Bhattacharya, D. Chakrabarty, M. Henzinger, and D. Nanongkai. Dynamic algorithms
In Proceedings of the 29th Annual ACM-SIAM Symposium on Discrete
for graph coloring.
Algorithms, SODA 2018, New Orleans, LA, USA, January 7-10, 2018, pages 1–20, 2018.
[14] S. Bhattacharya, M. Henzinger, and G. F. Italiano. Deterministic fully dynamic data structures
for vertex cover and matching. In Proceedings of the 26th Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015, pages 785–804,
2015.
[15] S. Bhattacharya, M. Henzinger, and D. Nanongkai. New deterministic approximation algo-
rithms for fully dynamic matching. In Proceedings of the 48th Annual ACM SIGACT Sympo-
sium on Theory of Computing, STOC 2016, Cambridge, MA, USA, June 18-21, 2016, pages
398–411, 2016.
[16] S. Bhattacharya, M. Henzinger, and D. Nanongkai. Fully dynamic maximum matching and
vertex cover in O(log3 n) worst case update time. In Proceedings of the 28th Annual ACM-
SIAM Symposium on Discrete Algorithms, SODA 2017, Barcelona, Spain, January 16-19,
2017, pages 470–489, 2017.
[17] K. Censor-Hillel, E. Haramaty, and Z. S. Karnin. Optimal dynamic distributed MIS.
In
Proceedings of the 2016 ACM Symposium on Principles of Distributed Computing, PODC
2016, Chicago, IL, USA, July 25-28, 2016, pages 217–226, 2016.
[18] S. A. Cook. An overview of computational complexity. Commun. ACM, 26(6):400–408, 1983.
[19] C. Demetrescu and G. F. Italiano. A new approach to dynamic all pairs shortest paths. J.
ACM, 51(6):968–992, 2004.
[20] D. Eppstein, Z. Galil, G. F. Italiano, and A. Nissenzweig. Sparsification - a technique for
speeding up dynamic graph algorithms. J. ACM, 44(5):669–696, 1997.
[21] S. Even and Y. Shiloach. An on-line edge-deletion problem. J. ACM, 28(1):1–4, 1981.
[22] G. N. Frederickson. Data structures for on-line updating of minimum spanning trees, with
applications. SIAM J. Comput., 14(4):781–798, 1985.
18
[23] M. Ghaffari. An improved distributed algorithm for maximal independent set. In Proceedings
of the 27th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2016, Arlington,
VA, USA, January 10-12, 2016, pages 270–277, 2016.
[24] A. Gupta, R. Krishnaswamy, A. Kumar, and D. Panigrahi. Online and dynamic algorithms
In Proceedings of the 49th Annual ACM SIGACT Symposium on Theory of
for set cover.
Computing, STOC 2017, Montreal, QC, Canada, June 19-23, 2017, pages 537–550, 2017.
[25] M. Gupta and R. Peng. Fully dynamic (1 + ǫ)-approximate matchings. In Proceedings of the
54th IEEE Annual Symposium on Foundations of Computer Science, FOCS 2013, Berkeley,
CA, USA, October 26-29, 2013, pages 548–557, 2013.
[26] M. Henzinger, S. Krinninger, and D. Nanongkai. Decremental single-source shortest paths
on undirected graphs in near-linear total update time. In Proceedings of the 55th IEEE An-
nual Symposium on Foundations of Computer Science, FOCS 2014, Philadelphia, PA, USA,
October 18-21, 2014, pages 146–155, 2014.
[27] M. Henzinger, S. Krinninger, and D. Nanongkai. Sublinear-time decremental algorithms for
single-source reachability and shortest paths on directed graphs. In Proceedings of the 46th
Annual ACM on Symposium on Theory of Computing, STOC 2014, New York, NY, USA,
May 31 - June 03, 2014, pages 674–683, 2014.
[28] M. Henzinger, S. Krinninger, D. Nanongkai, and T. Saranurak. Unifying and strengthening
hardness for dynamic problems via the online matrix-vector multiplication conjecture.
In
Proceedings of the 47th Annual ACM on Symposium on Theory of Computing, STOC 2015,
Portland, OR, USA, June 14-17, 2015, pages 21–30, 2015.
[29] M. R. Henzinger and V. King. Maintaining minimum spanning trees in dynamic graphs. In
Proceedings of the 24th International Colloquium on Automata, Languages, and Programming,
ICALP 1997, Bologna, Italy, July 7-11, 1997, pages 594–604, 1997.
[30] M. R. Henzinger and V. King. Randomized fully dynamic graph algorithms with polylogarith-
mic time per operation. J. ACM, 46(4):502–516, 1999.
[31] Z. Ivkovi´c and E. L. Lloyd. Fully dynamic maintenance of vertex cover. In Proceedings of the
19th International Workshop on Graph-Theoretic Concepts in Computer Science, WG 1993,
Utrecht, The Netherlands, June 16-18, 1993, pages 99–111, 1993.
[32] M. T. J. Holm, K. de. Lichtenberg. Poly-logarithmic deterministic fully-dynamic algorithms
for connectivity, minimum spanning tree, 2-edge, and biconnectivity. J. ACM, 48(4):723–760,
2001.
[33] R. M. Karp and A. Wigderson. A fast parallel algorithm for the maximal independent set
problem. J. ACM, 32(4):762–773, 1985.
[34] V. King. Fully dynamic algorithms for maintaining all-pairs shortest paths and transitive
closure in digraphs. In Proceedings of the 40th Annual Symposium on Foundations of Computer
Science, FOCS 1999, New York, NY, USA, October 17-18, 1999, pages 81–91, 1999.
[35] N. Linial. Distributive graph algorithms-global solutions from local data. In Proceedings of
the 28th IEEE Annual Symposium on Foundations of Computer Science, FOCS 1987, Los
Angeles, CA, USA, October 27-29, 1987, pages 331–335, 1987.
19
[36] M. Luby. A simple parallel algorithm for the maximal independent set problem. SIAM J.
Comput., 15(4):1036–1053, 1986.
[37] O. Neiman and S. Solomon. Simple deterministic algorithms for fully dynamic maximal match-
ing. In Proceedings of the 45th Annual ACM SIGACT Symposium on Theory of Computing,
STOC 2013, Palo Alto, CA, USA, June 1-4, 2013, pages 745–754, 2013.
[38] K. Onak and R. Rubinfeld. Maintaining a large matching and a small vertex cover.
In
Proceedings of the 42nd Annual ACM SIGACT Symposium on Theory of Computing, STOC
2010, Cambridge, MA, USA, June 6-8, 2010, pages 457–464, 2010.
[39] A. Panconesi and A. Srinivasan. On the complexity of distributed network decomposition. J.
Algorithms, 20(2):356–374, 1996.
[40] M. Parter, D. Peleg, and S. Solomon. Local-on-average distributed tasks. In Proceedings of
the 27th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2016, Arlington, VA,
USA, January 10-12, 2016, pages 220–239, 2016.
[41] D. Peleg. Distributed Computing: A Locality-Sensitive Approach. SIAM, 2000.
[42] L. Roditty and U. Zwick. On dynamic shortest paths problems. Algorithmica, 61(2):389–401,
2011.
[43] S. Solomon. Fully dynamic maximal matching in constant update time.
In Proceedings of
the 57th IEEE Annual Symposium on Foundations of Computer Science, FOCS 2016, New
Brunswick, NJ, USA, October 9-11, 2016, pages 325–334, 2016.
[44] M. Thorup. Worst-case update times for fully-dynamic all-pairs shortest paths. In Proceed-
ings of the 37th Annual ACM SIGACT Symposium on Theory of Computing, STOC 2005,
Baltimore, MD, USA, May 21-24, 2005, pages 112–119, 2005.
[45] C. Wulff-Nilsen. Fully-dynamic minimum spanning forest with improved worst-case update
time. In Proceedings of the 49th Annual ACM SIGACT Symposium on Theory of Computing,
STOC 2017, Montreal, QC, Canada, June 19-23, 2017, pages 1130–1143, 2017.
20
A An Ω(n) Lower Bound on Worst-Case Adjustment Complexity
By a straightforward modification of the lower bound example in [17], we can show that adjustment-
complexity of any deterministic algorithm is Ω(n) in the worst-case.
Consider the following example: Let G1 be a complete bipartite graph between two sets of
vertices L1 and R1, each of size n/4. We create an identical copy of G1 named G2 with bipartition
L2, R2 on the remaining vertices and let G be the union of these graphs. Consider any deterministic
algorithm A for maintaining an MIS on G. Without loss of generality, assume L1 ∪ L2 is the MIS
chosen by A (in any MIS of G either L1 is entirely in the MIS or R1 and similarly for L2 and
R2). Let u1 and u2 be two arbitrary vertices in L1 and L2. The adversary starts deleting all edges
incident to all vertices in L1 \ {u1} and L2 \ {u2}. Finally, it adds an edge between u1 and u2. We
argue that at some point during these updates, A adjusted Ω(n) vertices in the maintained MIS.
There are two cases to consider. For simplicity of exposition, we assume that at each time step, all
edges incident to a vertex are deleted at once.
Suppose at some point before inserting the last edge, A decides to add a vertex in R1 to the
MIS for the first time (the argument is symmetric for R2 as well). Since u1 is incident to all
vertices in R1, this means that u1 needs to leave the MIS. Also, since a vertex in R1 has joined
the MIS, we know that there cannot be any edge from vertices in the MIS to any vertex in R1
(as we start with a complete bipartite graph and assumed that all edges incident to a vertex are
deleted simultaneously). This means that after this step, all vertices in R1 should join the MIS to
ensure maximality. Therefore, at this step, Ω(n) vertices are inserted to the MIS at once, proving
the claim in this case.
Now suppose that before inserting the last edge, no vertex in R1 and R2 belong to the MIS and
hence both u1 and u2 should be inside it. By adding an edge between u1 and u2, A is forced to
remove at least one of them, say u1, from the MIS, which in turn forces all R1 to join MIS to keep
the maximality. Hence, again, Ω(n) vertices are inserted to the MIS at once, finalizing the proof.
Remark A.1. We remark that this simple example explains why we obtain our results in amortized
bounds rather than worst-case bounds.
21
|
1604.02188 | 1 | 1604 | 2016-04-07T22:31:38 | Simultaneous Nearest Neighbor Search | [
"cs.DS",
"cs.CG"
] | Motivated by applications in computer vision and databases, we introduce and study the Simultaneous Nearest Neighbor Search (SNN) problem. Given a set of data points, the goal of SNN is to design a data structure that, given a collection of queries, finds a collection of close points that are compatible with each other. Formally, we are given $k$ query points $Q=q_1,\cdots,q_k$, and a compatibility graph $G$ with vertices in $Q$, and the goal is to return data points $p_1,\cdots,p_k$ that minimize (i) the weighted sum of the distances from $q_i$ to $p_i$ and (ii) the weighted sum, over all edges $(i,j)$ in the compatibility graph $G$, of the distances between $p_i$ and $p_j$. The problem has several applications, where one wants to return a set of consistent answers to multiple related queries. This generalizes well-studied computational problems, including NN, Aggregate NN and the 0-extension problem.
In this paper we propose and analyze the following general two-step method for designing efficient data structures for SNN. In the first step, for each query point $q_i$ we find its (approximate) nearest neighbor point $\hat{p}_i$; this can be done efficiently using existing approximate nearest neighbor structures. In the second step, we solve an off-line optimization problem over sets $q_1,\cdots,q_k$ and $\hat{p}_1,\cdots,\hat{p}_k$; this can be done efficiently given that $k$ is much smaller than $n$. Even though $\hat{p}_1,\cdots,\hat{p}_k$ might not constitute the optimal answers to queries $q_1,\cdots,q_k$, we show that, for the unweighted case, the resulting algorithm is $O(\log k/\log \log k)$-approximation. Also, we show that the approximation factor can be in fact reduced to a constant for compatibility graphs frequently occurring in practice.
Finally, we show that the "empirical approximation factor" provided by the above approach is very close to 1. | cs.DS | cs |
Simultaneous Nearest Neighbor Search∗
Piotr Indyk
MIT
[email protected]
Robert Kleinberg
Cornell and MSR
[email protected]
Sepideh Mahabadi
MIT
[email protected]
Yang Yuan
Cornell University
[email protected]
Abstract
Motivated by applications in computer vision and databases, we introduce and study the Simultane-
ous Nearest Neighbor Search (SNN) problem. Given a set of data points, the goal of SNN is to design a
data structure that, given a collection of queries, finds a collection of close points that are "compatible"
with each other. Formally, we are given k query points Q = q1,··· , qk, and a compatibility graph G
with vertices in Q, and the goal is to return data points p1,··· , pk that minimize (i) the weighted sum
of the distances from qi to pi and (ii) the weighted sum, over all edges (i, j) in the compatibility graph
G, of the distances between pi and pj. The problem has several applications in computer vision and
databases, where one wants to return a set of consistent answers to multiple related queries. Further-
more, it generalizes several well-studied computational problems, including Nearest Neighbor Search,
Aggregate Nearest Neighbor Search and the 0-extension problem.
In this paper we propose and analyze the following general two-step method for designing efficient
data structures for SNN. In the first step, for each query point qi we find its (approximate) nearest
neighbor point pi; this can be done efficiently using existing approximate nearest neighbor structures. In
the second step, we solve an off-line optimization problem over sets q1,··· , qk and p1,··· , pk; this can
be done efficiently given that k is much smaller than n. Even though p1,··· , pk might not constitute the
optimal answers to queries q1,··· , qk, we show that, for the unweighted case, the resulting algorithm
satisfies a O(log k/ log log k)-approximation guarantee. Furthermore, we show that the approximation
factor can be in fact reduced to a constant for compatibility graphs frequently occurring in practice, e.g.,
2D grids, 3D grids or planar graphs.
Finally, we validate our theoretical results by preliminary experiments. In particular, we show that
the "empirical approximation factor" provided by the above approach is very close to 1.
1 Introduction
The nearest neighbor search (NN) problem is defined as follows: given a collection P of n points, build
a data structure that, given any query point from some set Q, reports the data point closest to the query.
The problem is of key importance in many applied areas, including computer vision, databases, information
retrieval, data mining, machine learning, and signal processing. The nearest neighbor search problem, as
well as its approximate variants, have been a subject of extensive studies over the last few decades, see,
e.g., [2, 3, 4, 5, 6, 7] and the references therein.
Despite their success, however, the current algorithms suffer from significant theoretical and practical
limitations. One of their major drawbacks is their inability to support and exploit structure in query sets
that is often present in applications. Specifically, in many applications (notably in computer vision), queries
∗This work was in part supported by NSF grant CCF 1447476 [1] and the Simons Foundation.
1
issued to the data structure are not unrelated but instead correspond to samples taken from the same object.
For example, queries can correspond to pixels or small patches taken from the same image. To ensure
consistency, one needs to impose "compatibility constraints" that ensure that related queries return similar
answers. Unfortunately, standard nearest neighbor data structures do not provide a clear way to enforce such
constraints, as all queries are processed independently of each other.
To address this issue, we introduce the Simultaneous Nearest Neighbor Search (SNN) problem. Given
k simultaneous query points q1, q2,··· , qk, the goal of a SNN data structure is to find k points (also called
labels) p1, p2,··· , pk in P such that (i) pi is close to qi, and (ii) p1,··· , pk are "compatible". Formally, the
compatibility is defined by a graph G = (Q, E) with k vertices which is given to the data structure, along
with the query points Q = q1,··· , qk. Furthermore, we assume that the data set P is a subset of some
space X equipped with a distance function distX, and that we are given another metric distY defined over
P ∪ Q. Given the graph G and the queries q1,··· , qk, the goal of the SNN data structure is to return points
p1,··· , pk from P that minimize the following function:
κidistY (pi, qi) +
λi,jdistX (pi, pj)
(1)
k(cid:88)
i=1
(cid:88)
(i,j)∈E
where κi and λi,j are parameters defined in advance.
The above formulation captures a wide variety of applications that are not well modeled by traditional
NN search. For example, many applications in computer vision involve computing nearest neighbors of
pixels or image patches from the same image [8, 9, 10]. In particular, algorithms for tasks such as de-
noising (removing noise from an image), restoration (replacing a deleted or occluded part of an image) or
super-resolution (enhancing the resolution of an image) involve assigning "labels" to each image patch1.
The labels could correspond to the pixel color, the enhanced image patch, etc. The label assignment should
have the property that the labels are similar to the image patches they are assigned to, while at the same
time the labels assigned to nearby image patches should be similar to each other. The objective function in
Equation 1 directly captures these constraints.
find a data point p that minimizes the sum2(cid:80)
From a theoretical perspective, Simultaneous Nearest Neighbor Search generalizes several well-studied
computational problems, notably the Aggregate Nearest Neighbor problem [12, 13, 14, 15, 16] and the 0-
extension problem [17, 18, 19, 20]. The first problem is quite similar to the basic nearest neighbor search
problem over a metric dist, except that the data structure is given k queries q1 ··· qk, and the goal is to
i dist(qi, p). This objective can be easily simulated in SNN
by setting distY = dist and distX = L · uniform, where L is a very large number and uniform(p, q) is
the uniform metric. The 0-extension problem is a combinatorial optimization problem where the goal is to
minimize an objective function quite similar to that in Equation 1. The exact definition of 0-extension as
well as its connections to SNN are discussed in detail in Section 2.1.
1.1 Our results
In this paper we consider the basic case where distX = distY and λi,j = κi = 1; we refer to this variant
as the unweighted case. Our main contribution is a general reduction that enables us to design and analyze
efficient data structures for unweighted SNN. The algorithm (called Independent Nearest Neighbors or INN)
1This problem has been formalized in the algorithms literature as the metric labeling problem [11]. The problem considered in
this paper can thus be viewed as a variant of metric labeling with a very large number of labels.
2Other aggregate functions, such as the maximum, are considered as well.
2
consists of two steps. In the first (pruning) step, for each query point qi we find its nearest neighbor3 point
pi ; this can be done efficiently using existing nearest neighbor search data structures. In the second (opti-
mization) step, we run an appropriate (approximation) algorithm for the SNN problem over sets q1,··· , qk
and p1,··· , pk; this can be done efficiently given that k is much smaller than n. We show that the resulting
algorithm satisfies a O(b log k/ log log k)-approximation guarantee, where b is the approximation factor of
the algorithm used in the second step. This can be further improved to O(bδ), if the metric space dist admits
a δ-padding decomposition (see Preliminaries for more detail). The running time incurred by this algorithm
is bounded by the cost of k nearest neighbor search queries in a data set of size n plus the cost of the ap-
proximation algorithm for the 0-extension problem over an input of size k. By plugging in the best nearest
neighbor algorithms for dist we obtain significant running time savings if k (cid:28) n.
We note that INN is somewhat similar to the belief propagation algorithm for super-resolution described
in [8]. Specifically, that algorithm selects 16 closest labels for each qi, and then chooses one of them by
running a belief propagation algorithm that optimizes an objective function similar to Equation 1. However,
we note that the algorithm in [8] is heuristic and is not supported by approximation guarantees.
We complement our upper bound by showing that the aforementioned reduction inherently yields super-
constant approximation guarantee. Specifically, we show that, for an appropriate distance function dist,
queries q1,··· , qk, and a label set P , the best solution to SNN with the label set restricted to p1,··· , pk
√
can be Θ(
log k) times larger than the best solution with label set equal to P . This means that even if the
second step problem is solved to optimality, reducing the set of labels from P to P inherently increases the
cost by a super-constant factor.
However, we further show that the aforementioned limitation can be overcome if the compatibility graph
G has pseudoarboricity r (which means that each edge can be mapped to one of its endpoint vertices such
that at most r edges are mapped to each vertex). Specifically, we show that if G has pseudoarboricity r, then
the gap between the best solution using labels in P , and the best solution using labels in P , is at most O(r).
Since many graphs used in practice do in fact satisfy r = O(1) (e.g., 2D grids, 3D grids or planar graphs),
this means that the gap is indeed constant for a wide collection of common compatibility graphs.
In Appendix 6 we also present an alternative algorithm for the r-pseudoarboricity case. Similarly to
INN, the algorithm computes the nearest label to each query qi. However, the distance function used to
compute the nearest neighbor involves not only the distance between qi and a label p, but also the distances
between the neighbors of qi in G and p. This nearest neighbor operation can be implemented using any
data structure for the Aggregate Nearest Neighbor problem [12, 13, 14, 15, 16]. Although this results in a
more expensive query time, the labeling computed by this algorithm is final, i.e., there is no need for any
additional postprocessing. Furthermore, the pruning gap (and therefore the final approximation ratio) of the
algorithm is only 2r + 1, which is better than our bound for INN.
Finally, we validate our theoretical results by preliminary experiments comparing our SNN data structure
with an alternative (less efficient) algorithm that solves the same optimization problem using the full label set
P . In our experiments we apply both algorithms to an image denoising task and measure their performance
using the objective function (1).
In particular, we show that the "empirical gap" incurred by the above
approach, i.e, the ratio of objective function values observed in our experiments, is very close to 1.
3Our analysis immediately extends to the case where the we compute approximate, not exact, nearest neighbors. For simplicity
we focus only on the exact case in the following discussion.
3
1.2 Our techniques
We start by pointing out that SNN can be reduced to 0-extension in a "black-box" manner. Unfortunately,
this reduction yields an SNN algorithm whose running time depends on the size of labels n, which could
be very large; essentially this approach defeats the goal of having a data structure solving the problem. The
INN algorithm overcomes this issue by reducing the number of labels from n to k. However the pruning step
can increase the cost of the best solution. The ratio between the optimum cost after pruning to the optimum
cost before pruning is called the pruning gap.
To bound the pruning gap, we again resort to existing 0-extension algorithms, albeit in a "grey box"
manner. Specifically, we observe that many algorithms, such as those in [19, 20, 18, 21], proceed by first
creating a label assignment in an "extended" metric space (using a LP relaxation of 0-extension), and then
apply a rounding algorithm to find an actual solution. The key observation is that the correctness of the
rounding step does not rely on the fact that the initial label assignment is optimal, but instead it works for
any label assignment. We use this fact to translate the known upper bounds for the integrality gap of linear
programming relaxations of 0-extension into upper bounds for the pruning gap. On the flip side, we show
a lower bound for the pruning gap by mimicking the arguments used in [19] to lower bound the integrality
gap of a 0-extension relaxation.
To overcome the lower bound, we consider the case where the compatibility graph G has pseudoarboric-
ity r. Many graphs used in applications, such as 2D grids, 3D grids or planar graphs, have pseudoarboricity
r for some constant r. We show that for such graphs the pruning gap is only O(r). The proof proceeds by
directly assigning labels in P to the nodes in Q and bounding the resulting cost increase. It is worth not-
ing that the "grey box" approach outlined in the preceding paragraph, combined with Theorem 11 of [19],
yields an O(r3) pruning gap for the class of Kr,r-minor-free graphs, whose pseudoarboricity is O(r). Our
O(r) pruning gap not only improves this O(r3) bound in a quantitative sense, but it also applies to a much
broader class of graphs. For example, three-dimensional grid graphs have pseudoarboricity 6, but the class
of three-dimensional grid graphs includes graphs with Kr,r minors for every positive integer r.
Finally, we validate our theoretical results by experiments. We focus on a simple de-noising scenario
where X is the pixel color space, i.e., the discrete three-dimensional space space {0 . . . 255}3. Each pixel in
this space is parametrized by the intensity of the red, green and blue colors. We use the Euclidean norm to
measure the distance between two pixels. We also let P = X. We consider three test images: a cartoon with
an MIT logo and two natural images. For each image we add some noise and then solve the SNN problems
for both the full color space P and the pruned color space P . Note that since P = X, the set of pruned
labels P simply contains all pixels present in the image.
Unfortunately, we cannot solve the problems optimally, since the best known exact algorithm takes
exponential time. Instead, we run the same approximation algorithm on both instances and compare the
solutions. We find that the values of the objective function for the solutions obtained using pruned labels
and the full label space are equal up to a small multiplicative factor. This suggests that the empirical value
of the pruning gap is very small, at least for the simple data sets that we considered.
2 Definitions and Preliminaries
We define the Unweighted Simultaneous Nearest Neighbor problem as follows. Let (X, dist) be a metric
space and let P ⊆ X be a set of n points from the space.
Definition 2.1. In the Unweighted Simultaneous Nearest Neighbor problem, the goal is to build a data
structure over a given point set P that supports the following operation. Given a set of k points Q =
4
{q1,··· , qk} in the metric space X, along with a graph G = (Q, E) of k nodes, the goal is to report k (not
necessarily unique) points from the database p1,··· , pk ∈ P which minimize the following cost function:
k(cid:88)
i=1
(cid:88)
(qi,qj )∈E
dist(pi, qi) +
dist(pi, pj)
(2)
We refer to the first term in sum as the nearest neighbor (NN) cost, and to the second sum as the pairwise
(PW) cost. We denote the cost of the optimal assignment from the point set P by Cost(Q, G, P ).
In the rest of this paper, simultaneous nearest neighbor (SNN) refers to the unweighted version of the
problem (unless stated otherwise). Next, we define the pseudoarboricity of a graph and r-sparse graphs.
Definition 2.2. Pseudoarboricity of a graph G is defined to be the minimum number r, such that the edges
of the graph can be oriented to form a directed graph with out-degree at most r. In this paper, we call such
graphs as r-sparse.
Note that given an r-sparse graph, one can map the edges to one of its endpoint vertices such that there are
at most r edges mapped to each vertex. The doubling dimension of a metric space is defined as follows.
Definition 2.3. The doubling dimension of a metric space (X, dist) is defined to be the smallest δ such that
every ball in X can be covered by 2δ balls of half the radius.
It is known that the doubling dimension of any finite metric space is O(log X). We then define padding
decompositions.
Definition 2.4. A metric space (X, dist) is δ-padded decomposable if for every r, there is a randomized
partitioning of X into clusters C = {Ci} such that, each Ci has diameter at most r, and that for every
x1, x2 ∈ X, the probability that x1 and x2 are in different clusters is at most δdist(x1, x2)/r.
It is known that any finite metric with doubling dimension δ admits an O(δ)-padding decomposition [22].
2.1 0-Extension Problem
The 0-extension problem, first defined by Karzanov [17] is closely related to the Simultaneous Nearest
Neighbor problem. In the 0-extension problem, the input is a graph G(V, E) with a weight function w(e),
and a set of terminals T ⊆ V with a metric d defined on T . The goal is to find a mapping from the vertices
to the terminals f : V → T such that each terminal is mapped to itself and that the following cost function
is minimized:
(cid:88)
(u,v)∈E
w(u, v) · d(f (u), f (v))
It can be seen that this is a special case of the metric labeling problem [11] and thus a special case of the
general version of the SNN problem defined by Equation 1. To see this, it is enough to let Q = V and
P = T , and let κi = ∞ for qi ∈ T , κi = 0 for qi (cid:54)∈ T , and λi,j = w(i, j) in Equation 1.
Calinescu et al. [19] considered the semimetric relaxation of the LP for the 0-extension problem and
gave an O(log T) algorithm using randomized rounding of the LP solution. They also proved an integrality
Later Fakcharoenphol et al. [18] improved the upper-bound to O(log T/ log log T), and Lee and Naor
[21] proved that if the metric d admits a δ-padded decomposition, then there is an O(δ)-approximation
ratio of O((cid:112)log T) for the semimetric LP relaxation.
5
algorithm for the 0-extension problem. For the finite metric spaces, this gives an O(δ) algorithm where
δ is the doubling dimension of the metric space. Furthermore, the same results can be achieved using
another metric relaxation (earth-mover relaxation), see [20]. Later Karloff et al. [23] proved that there is
no polynomial time algorithm for 0-extension problem with approximation factor O((log n)1/4−) unless
N P ⊆ DT IM E(npoly(log n)).
SNN can be reduced to 0-extension in a "black-box" manner via the following lemma.
Lemma 2.5. Any b-approximate algorithm for the 0-extension problem yields an O(b)-approximate algo-
rithm for the SNN problem.
Proof. Given an instance of the SNN problem (Q, G(cid:48), P ), we build an instance of the 0-extension problem
(V, T, G) as follows. Let T = P and V = T ∪ Q. The metric d is the same as dist. However the graph G
of the 0-extension problem requires some modification. Let G(cid:48) = (Q, EG(cid:48)), then G = (V, E) is defined as
follows. For each qi, qj ∈ Q, we have the edge (qi, qj) ∈ E iff (qi, qj) ∈ EG(cid:48). We also include another type
of edges in the graph: for each qi ∈ Q, we add an edge (qi, pi) ∈ E where pi ∈ P is the nearest neighbor of
qi. Note that we consider the graph G to be unweighted.
Using the b-approximation algorithm for this problem, we get an assignment µ that maps the non-
terminal vertices q1,··· , qk to the terminal vertices. Suppose qi is mapped to the terminal vertex pi in this
assignment. Let p∗
k be the optimal SNN assignment. Next, we show that the same mapping µ for
the SNN problem, gives us an O(b) approximate solution. The SNN cost of the mapping µ is denoted as
follows:
CostSNN(µ) =
dist(qi, pi) +
dist(pi, pj)
≤ Cost(Q, G(cid:48), P ) + b · [
dist(qi, p∗
i ) +
dist(p∗
i , p∗
j )]
where we have used triangle inequality and the following facts in the above. First, pi is the closest point
in P to qi and thus dist(qi, pi) ≤ dist(qi, p∗
i ). Second, by definition we have that Cost(Q, G(cid:48), P ) =
j ). Finally, since µ is a b approximate solution for the 0-extension
(qi,qj )∈EG(cid:48) dist(pi, pj) is smaller than b times the 0-extension
(cid:80)k
i )+(cid:80)
i=1 dist(pi, pi) +(cid:80)
problem, we have that(cid:80)k
i=1 dist(qi, p∗
cost of any other assignment, and in particular(cid:80)k
(qi,qj )∈EG(cid:48) dist(p∗
i ) +(cid:80)
i , p∗
j ).
By plugging in the known 0-extension algorithms cited earlier we obtain the following:
(qi,qj )∈EG(cid:48) dist(p∗
i=1 dist(pi, p∗
i , p∗
6
1,··· , p∗
k(cid:88)
≤ k(cid:88)
≤ k(cid:88)
i=1
i=1
i=1
dist(qi, p∗
(cid:88)
k(cid:88)
(qi,qj )∈EG(cid:48)
i=1
i=1
k(cid:88)
i ) + b · [
k(cid:88)
k(cid:88)
i=1
≤ Cost(Q, G(cid:48), P ) + b · [
≤ Cost(Q, G(cid:48), P )(2b + 1)
i=1
dist(qi, pi) +
dist(pi, pi) +
dist(pi, pj)
(cid:88)
(qi,qj )∈EG(cid:48)
(cid:88)
k(cid:88)
(qi,qj )∈EG(cid:48)
dist(pi, p∗
i ) +
dist(pi, qi) +
i=1
dist(pi, qi) + Cost(Q, G(cid:48), P )]
dist(p∗
i , p∗
j )]
(cid:88)
(qi,qj )∈EG(cid:48)
Corollary 2.6. There exists an O(log n/ log log n) approximation algorithm for the SNN problem with
running time nO(1), where n is the size of the label set.
Corollary 2.7. If the metric space (X, dist) is δ-padded decomposable, then there exists an O(δ) approxi-
mation algorithm for the SNN problem with running time nO(1). For finite metric spaces X, δ could represent
the doubling dimension of the metric space (or equivalently the doubling dimension of P ∪ Q).
Unfortunately, this reduction yields a SNN algorithm with running time depending on the size of labels
n, which could be very large. In the next section we show how to improve the running time by reducing the
labels set size from n to k. However, unlike the reduction in this section, our new reduction will no longer be
"black-box". Instead, its analysis will use particular properties of the 0-extension algorithms. Fortunately
those properties are satisfied by the known approximation algorithms for this problem.
3 Independent Nearest Neighbors Algorithm
In this section, we consider a natural and general algorithm for the SNN problem, which we call Independent
Nearest Neighbors (INN). The algorithm proceeds as follows. Given the query points Q = {q1,··· , qk},
for each qi the algorithm picks its (approximate) nearest neighbor pi. Then it solves the problem over the
set P = { p1,··· , pk} instead of P . This simple approach reduces the size of search space from n down to
k.
The details of the algorithm are shown in Algorithm 1.
Algorithm 1 Independent Nearest Neighbors (INN) Algorithm
Input Q = {q1,··· , qk}, and input graph G = (Q, E)
1: for i = 1 to k do
2:
3: end for
4: Find the optimal (or approximately optimal) solution among the set P = { p1,··· , pk}.
Query the NN data structure to extract a nearest neighbor (or approximate nearest neighbor) pi for qi
In the rest of the section we analyze the quality of this pruning step. More specifically, we define the
pruning gap of the algorithm as the ratio of the optimal cost function using the points in P over its value
using the original point set P .
Definition 3.1. The pruning gap of an instance of SNN is defined as α(Q, G, P ) = Cost(Q,G, P )
the pruning gap of the INN algorithm, α, as the largest value of α(Q, G, P ) over all instances.
Cost(Q,G,P ). We define
First, in Section 3.1, by proving a reduction from algorithms for rounding the LP solution of the 0-
extension problem, we show that for arbitrary graphs G, we have α = O(log k/ log log k), and if the metric
(X, dist) is δ-padded decomposable, we have α = O(δ) (for example, for finite metric spaces X, δ can
represent the doubling dimension of the metric space). Then, in Section 3.2, we prove that α = O(r)
where r is the pseudoarboricity of the graph G. This would show that for the sparse graphs, the pruning gap
√
remains constant. Finally, in Section 4, we present a lower bound showing that the pruning gap could be as
large as Ω(
log k) and as large as Ω(r) for (r ≤ √
log k). Therefore, we get the following theorem.
Theorem 3.2. The following bounds hold for the pruning gap of the INN algorithm. First we have α =
O( log k
log log k ), and that if metric (X, dist) is δ-padded decomposable, we have α = O(δ). Second, α = O(r)
7
√
where r is the pseudoarboricity of the graph G. Finally, we have that α = Ω(
r ≤ √
log k.
log k) and α = Ω(r) for
Note that the above theorem results in an O(b · α) time algorithm for the SNN problem where b is the
approximation factor of the algorithm used to solve the metric labeling problem for the set P , as noted in
line 4 of the INN algorithm. For example in a general graph b would be O(log k/ log log k) that is added on
top of O(α) approximation of the pruning step.
3.1 Bounding the pruning gap using 0-extension
In this section we show upper bounds for the pruning gap (α) of the INN algorithm. The proofs use specific
properties of existing algorithms for the 0-extension problem.
Definition 3.3. We say an algorithm A for the 0-extension problem is a β-natural rounding algorithm if,
given a graph G = (V, E), a set of terminals T ⊆ V , a metric space (X, dX ), and a mapping µ : V → X,
it outputs another mapping ν : V → X with the following properties:
• ∀t ∈ T : ν(t) = µ(t)
• ∀v ∈ V : ∃t ∈ T s.t. ν(v) = µ(t)
• Cost(ν) ≤ βCost(µ), i.e.,(cid:80)
(u,v)∈E dX (ν(u), ν(v)) ≤ β ·(cid:80)
(u,v)∈E dX (µ(u), µ(v))
Many previous algorithms for the 0-extension problem, such as [19, 20, 18, 21], first create the mapping µ
using some LP relaxation of 0-extension (such as semimetric relaxation or earth-mover relaxation), and then
apply a β-natural rounding algorithm for the 0-extension to find the mapping ν which yields the solution
to the 0-extension problem. Below we give a formal connection between guarantees of these rounding
algorithms, and the quality of the output of the INN algorithm (the pruning gap of INN).
Lemma 3.4. Let A be a β-natural rounding algorithm for the 0-extension problem. Then we can infer that
the pruning gap of the INN algorithm is O(β), that is, α = O(β).
Proof. Fix any SNN instance (Q, GS, P ), where GS = (Q, EP W ), and its corresponding INN invocation.
We construct the inputs to the algorithm A from the INN instance as follows. Let the metric space of A
be the same as (X, dist) defined in the SNN instance. Also, let V be a set of 2k vertices corresponding to
P ∪ P ∗ with T corresponding to P . Here P ∗ = {p∗
k} is the set of the optimal solutions of SNN,
and P is the set of nearest neighbors as defined by INN. The mapping µ simply maps each vertex from
V = P ∪ P ∗ to itself in the metric X defined in SNN. Moreover, the graph G = (V, E) is defined such that
E = {(pi, p∗
i , p∗
Cost(Q, GS, P ) = Cost(Q, GS, P ∗))
First we claim the following (note that Cost(µ) is defined in Definition 3.3, and that by definition
i )1 ≤ i ≤ k} ∪ {(p∗
j )(qi, qj) ∈ EP W}.
1,··· , p∗
Cost(µ) ≤ 2Cost(Q, GS, P ∗) = 2Cost(Q, GS, P )
We know that Cost(Q, GS, P ∗) can be split into NN cost and PW cost. We can also split Cost(µ) into
i )1 ≤ i ≤ k}) and PW cost (corresponding to edge set
NN cost (corresponding to edge set {(pi, p∗
j )(qi, qj) ∈ EP W}). By definition we know the PW costs of Cost(Q, GS, P ) and Cost(µ) are equal.
{(p∗
i ) ≤ dist(pi, qi) + dist(qi, p∗
i ) ≤ 2 · dist(qi, p∗
For NN cost, by triangle inequality, we know dist(pi, p∗
i ).
Here we use the fact that pi is the nearest database point of qi. Thus, the claim follows.
i , p∗
8
We then apply algorithm A to get the mapping ν. By the assumption on A, we know that Cost(ν) ≤
βCost(µ). Given the mapping ν by the algorithm A, consider the assignment in the SNN instance where
i ) ∈ T , this would map all points qi to points in
each query qi is mapped to ν(p∗
P . Thus, by definition, we have that
Cost(Q, GS, P ) ≤ k(cid:88)
≤ k(cid:88)
≤ k(cid:88)
i=1
i=1
dist(qi, pi) +
i=1
dist(qi, pi) + Cost(ν)
i ), and note that since ν(p∗
(cid:88)
dist(qi, ν(p∗
i )) +
dist(ν(p∗
k(cid:88)
(qi,qj )∈EP W
dist(pi, ν(p∗
i )) +
i ), ν(p∗
j ))
(cid:88)
(qi,qj )∈EP W
dist(ν(p∗
i ), ν(p∗
j ))
i=1
≤ Cost(Q, GS, P ) + βCost(µ)
≤ (2β + 1)Cost(Q, GS, P )
where we have used the triangle inequality. Therefore, we have that the pruning gap α of the INN algorithm
is O(β), as claimed.
Using the previously cited results, and noting that in the above instance V = O(k), we get the follow-
ing corollaries.
Corollary 3.5. The INN algorithm has pruning gap α = O(log k/ log log k).
Corollary 3.6. If the metric space (X, dist) admits a δ-padding decomposition, then the INN algorithm has
pruning gap α = O(δ). For finite metric spaces (X, dist), δ is at most the doubling dimension of the metric
space.
3.2 Sparse Graphs
In this section, we prove that the INN algorithm performs well on sparse graphs. More specifically, here we
prove that when the graph G is r-sparse, then α(Q, G, P ) = O(r). To this end, we show that there exists an
assignment using the points in P whose cost function is within O(r) of the optimal solution using the points
in the original data set P .
Let p∗
Given a graph G of pseudoarboricity r, we know that we can map each edge to one of its end points
such that the number of edges mapped to each vertex is at most r. For each edge e, we call the vertex that e
is mapped to as the corresponding vertex of e. This would mean that each vertex is the corresponding vertex
of at most r edges.
k ∈ P denote the optimal solution of SNN. Algorithm 2 shows how to find an assignment
1,··· , p∗
p1,··· , pk ∈ P . We show that the cost of this assignment is within a factor O(r) from the optimum.
Lemma 3.7. The assignment defined by Algorithm 2, has O(r) approximation factor.
Proof. For each qi ∈ Q, let yi = dist(p∗
i , qi) and for each edge e = (qi, qj) ∈ E let xe = dist(p∗
i , p∗
j ).
e∈E xe. Note that Y is the NN cost and X is the PW cost of the optimal
e, Y (cid:48) , X(cid:48) in the same way
assignment and that OP T = Cost(Q, G, P ) = X + Y . Define the variables y(cid:48)
i=1 yi and X =(cid:80)
Also let Y =(cid:80)k
i, x(cid:48)
9
1,··· , p∗
Algorithm 2 r-Sparse Graph Assignment Algorithm
Input Query points q1,··· , qk, Optimal assignment p∗
input graph G = (Q, E)
Output An Assignment p1,··· , pk ∈ P
1: for i = 1 to k do
Let j0 = i and let qj1,··· , qjt be all the neighbors of qi in the graph G
2:
3: m ← arg mint
Assign pi ← pjm
4:
5: end for
(cid:96)=0 dist(p∗
) + dist(p∗
i , p∗
, qj(cid:96))
j(cid:96)
j(cid:96)
k, Nearest Neighbors p1,··· , pk, and the
i as follows.
but for the assignment p1,··· , pk produced by the algorithm. That is, for each qi ∈ Q, y(cid:48)
i = dist(pi, qi),
and for each edge e = (qi, qj) ∈ E, x(cid:48)
e = dist(pi, pj). Moreover, for a vertex qi, we define the designated
neighbor of qi to be qjm for the value of m defined in the line 3 of Algorithm 2 (note that the designated
neighbor might be the vertex itself). Fix a vertex qi and let qc be the designated neighbor of qi. We can
bound the value of y(cid:48)
y(cid:48)
i = dist(qi, pi) = dist(qi, pc)
≤ dist(qi, p∗
i ) + dist(p∗
≤ yi + dist(p∗
c) + 2dist(p∗
i , p∗
≤ yi + 2[dist(p∗
i , p∗
c) + dist(p∗
≤ 3yi
(by definition of designated neighbor and the value m in line 3 of Algorithm 2)
Thus summing over all vertices, we get that Y (cid:48) ≤ 3Y . Now for any fixed edge e = (qi, qs) (with qi being
its corresponding vertex), let qc be the designated neighbor of qi, and qz be the designated neighbor of qs.
Then we bound the value of x(cid:48)
c, qc) + dist(qc, pc)
(since pc is the nearest neighbor of qc)
(by triangle inequality)
c) + dist(p∗
c, qc)
c, qc)]
i , p∗
e as follows.
x(cid:48)
e = dist(pi, ps) = dist(pc, pz)
≤ dist(pc, qc) + dist(qc, p∗
z) + dist(p∗
+ dist(p∗
s, p∗
≤ 2dist(qc, p∗
c) + dist(p∗
z) + 2dist(p∗
+ dist(p∗
s, p∗
≤ 2[dist(qc, p∗
c) + dist(p∗
≤ 2yi + xe + 2[xe + yi]
≤ 4(xe + yi)
(by definition of designated neighbor and line 4 of Algorithm 2)
i , p∗
s)
i ) + dist(p∗
i ) + dist(p∗
c) + dist(p∗
(by triangle inequality)
c, p∗
z, qz) + dist(qz, pz)
c, p∗
i , p∗
s)
z, qz)
i )] + dist(p∗
c, p∗
(since qc(qz respectively) is designated neighbor of qi(qs respectively))
(since pc(pz respectively) is a NN of qc(qz respectively))
s) + 2[dist(p∗
z) + dist(p∗
s, p∗
i , p∗
z, qz)]
Hence, summing over all the edges, since each vertex qi is the corresponding vertex of at most r edges, we
get that X(cid:48) ≤ 4X + 4rY . Therefore we have the following.
Cost(Q, G, P ) ≤ X(cid:48) + Y (cid:48) ≤ 3Y + 4X + 4rY ≤ (4r + 3) · Cost(Q, G, P )
and thus α(Q, G, P ) = O(r).
10
4 Lower bound
√
log k) for the approximation factor of the INN algorithm.
In this section we prove a lower bound of Ω(
√
Furthermore, the lower bound example presented in this section is a graph (in fact a multi-graph) that has
pseudoarboricity equal to O(
log k), showing that in a way, the upper bound of α = O(r) for the r-sparse
log k, we have α = Ω(r). We note that the lower
bound construction presented in this paper is similar to the approach of [19] for proving a lower bound for
the integrality ratio of the LP relaxation for the 0-extension problem.
graphs is tight. More specifically, we show that for r ≤ √
Lemma 4.1. For any value of k, there exists a set of points P of size O(k) in a metric space X, and a query
(Q, G) such that Q = k and the pruning step induces an approximation factor of at least α(Q, G, P ) =
√
Ω(
log k).
√
Proof. In what follows, we describe the construction of the lower bound example.
Let H = (V, E) be an expander graph with k vertices V = {v1,··· , vk} such that each vertex has
constant degree d and the vertex expansion of the graph is also a constant c. Let H(cid:48) = (V (cid:48), E(cid:48), W (cid:48)) be
a weighted graph constructed from H by adding k vertices {u1,··· , uk} such that each new vertex ui is
log k. All the other edges between {v1,··· , vk} (which
a leaf connected to vi with an edge of weight
were present in H) have weight 1. This graph H(cid:48) defines the metric space (X, dist) such that X is the set
of nodes V (cid:48) and dist is the weight of the shortest path between the nodes in the graph H(cid:48). Moreover, let
P = V (cid:48) be all the vertices in the graph H(cid:48).
Let the set of k queries be Q = V (cid:48) \ V = {u1, . . . , uk}. Then, while running the INN algorithm, the set
of candidates P would be the queries themselves, i.e., P = Q = {u1,··· , uk}. Also, let the input graph
G = (Q, EG) be a multi-graph which is obtained from H by replacing each edge (vi, vj) in H with
log k
copies of the edge (ui, uj) in G. This is the input graph given along with the k queries to the algorithm.
Consider the solution P ∗ = {p∗
√
1,··· , p∗
k} where p∗
(cid:88)
i = vi. The cost of this solution is
dist(vi, vj) = k(cid:112)log k + kd(cid:112)log k/2
k(cid:88)
i=1
dist(qi, p∗
i ) +
(ui,uj )∈EG
Therefore, the cost of the optimal solution OP T = Cost(Q, G, P ) is at most O(k
the optimal labeling P ∗ = { p∗
1,··· , p∗
of the following forms.
Case 1: For all 1 ≤ i ≤ k, we have p∗
log k). Next, consider
k} ⊆ P using only the points in P . This optimal assignment has one
i = ui. The cost of P ∗ in this case would be
√
k(cid:88)
i=1
(cid:88)
dist(ui, uj) ≥ 0 + EG · 2(cid:112)log k ≥ dk
log k
2
Cost(Q, G, P ) =
dist(qi, ui) +
Thus the cost in this case would be Ω(OP T
Case 2: All the p∗
would be:
(ui,uj )∈EG
√
log k).
i 's are equal. Without loss of generality suppose they are all equal to u1. Then the cost
Cost(Q, G, P ) =
dist(qi, u1) +
dist(u1, u1) ≥ Ω(k log k) + 0
k(cid:88)
i=1
(cid:88)
(ui,uj )∈EG
11
2
of any vertex is at most 1 + d + ··· , d
2 log d = Θ(log k). Thus, again the cost of the assignment P in this case would be Ω(OP T
This is true because in an expander graph with constant degree, the number of vertices at distance less
√
than logd k
k. Thus Θ(k) vertices are farther than
logd k
2 = log k
log k).
Case 3: Let S = {S1,··· , St} be a partition of [k] such that each part corresponds to all the indices i having
their p∗
i(cid:48). Now, two cases are possible. First
if all the parts Sj have size at most k/2. In this case, since the graph H has expansion c, the total number of
edges between different parts would be at least
i equal. That is, for each 1 ≤ j ≤ t, we have ∀i, i(cid:48) ∈ Sj : p∗
2 ≤ 2
i = p∗
√
logd k
(cid:12)(cid:12)(cid:12){(ui, uj) ∈ EG p∗
i (cid:54)= p∗
j}(cid:12)(cid:12)(cid:12) ≥ 1
2
t(cid:88)
j=1
cSj(cid:112)log k ≥ kc(cid:112)log k/2
log k/2 · √
√
Therefore similar to Case 1 above, the PW cost would be at least kc
log k = Ω(k log k).
Otherwise, at least one of the parts such as Sj has size at least k/2. In this case, similar to Case 2 above,
the NN cost would be at least Ω(k log k). Therefore, in both cases the cost of the assignment P ∗ would be
at least Ω(OP T
√
log k). Hence, the pruning gap of the INN algorithm on this graph is Ω(
log k).
√
√
Since the degree of all the vertices in the above graph is d
√
also Θ(
above proof, the same arguments hold and we get the following corollary.
log k). It is easy to check that if we repeat each edge r times instead of
log k, the pseudoarboricity of the graph is
log k times in EG in the
√
Corollary 4.2. For any value of r ≤ √
log k, there exists an instance of SNN(Q,G,P) such that the input
graph G has arboricity O(r) and that the pruning gap of the INN algorithm is α(Q, G, P ) = Ω(r).
5 Experiments
We consider image denoising as an application of our algorithm. A popular approach to denoising (see e.g.
[24]) is to minimize the following objective function:
(cid:88)
i∈V
(cid:88)
(i,j)∈E
κid(qi, pi) +
λi,jd(pi, pj)
Here qi is the color of pixel i in the noisy image, and pi is the color of pixel i in the output. We use the
standard 4-connected neighborhood system for the edge set E, and use Euclidean distance as the distance
function d(·,·). We also set all weights κi and λi,j to 1.
When the image is in grey scale, this objective function can be optimized approximately and efficiently
using message passing algorithm, see e.g. [25]. However, when the image pixels are points in RGB color
space, the label set becomes huge (n = 2563 = 16, 777, 216), and most techniques for metric labeling are
not feasible.
Recall that our algorithm proceeds by considering only the nearest neighbor labels of the query points,
i.e., only the colors that appeared in the image. In what follows we refer to this reduced set of labels as the
image color space, as opposed to the full color space where no pruning is performed.
In order to optimize the objective function efficiently, we use the technique of [24]. We first embed
the original (color) metric space into a tree metric (with O(log n) distortion), and then apply a top-down
divide and conquer algorithm on the tree metric, by calling the alpha-beta swap subroutine [26]. We use
the random-split kd-tree for both the full color space and the image color space. When constructing the kd-
tree, split each interval [a, b] by selecting a random number chosen uniformly at random from the interval
[0.6a + 0.4b, 0.4a + 0.6b].
12
Avg cost for full color Avg cost for image color Empirical pruning gap
MIT
Snow
Surf
341878 ± 3.1%
9338604 ± 4.5%
8304184 ± 6.6%
340477 ± 1.1%
9564288 ± 6.2%
7588244 ± 5.1%
0.996
1.024
0.914
Table 1: The empirical values of objective functions for the respective images and algorithms.
To evaluate the performance of the two algorithms, we use one cartoon image with MIT logo and two
images from the Berkeley segmentation dataset [27] which was previously used in other computer vision
papers [24]. We use Matlab imnoise function to create noisy images from the original images. We run each
instance 20 times, and compute both the average and the variance of the objective function (the variance is
due to the random generating process of kd tree).
Table 2: MIT logo (first column, size 45 ∗ 124), and two images from the Berkeley segmentation dataset
[27] (second & third columns, size 321∗ 481). The first row shows the original image; the second row shows
the noisy image; the third row shows the denoised image using full color space; the fourth row shows the
denoised image using image space (our algorithm).
13
The results are presented in Figure 2 and Table 1. In Figure 2, one can see that the images produced
by the two algorithms are comparable. The full color version seems to preserve a few more details than
the image color version, but it also "hallucinates" non-existing colors to minimize the value of the objective
function. The visual quality of the de-noised images can be improved by fine-tuning various parameters of
the algorithms. We do not report these results here, as our goal was to compare the values of the objective
function produced by the two algorithms, as opposed to developing the state of the art de-noising system.
Note that, as per Table 1, for some images the value of the objective function is sometimes lower for
the image color space compared to the full color space. This is because we cannot solve the optimization
problem exactly. In particular, using the kd tree to embed the original metric space into a tree metric is an
approximate process.
5.1 De-noising with patches
To improve the quality of the de-noised images, we run the experiment for patches of the image, instead
of pixels. Moreover, we use Algorithm 3 which implements not only a pruning step, but also computes
the solution directly. In this experiment (see Figure 3 for a sample of the results), each patch (a grid of
pixels) from the noisy image is a query point, and the dataset consists of available patches which we use as
a substitute for a noisy patch.
In our experiment, to build the dataset, we take one image from the Berkeley segmentation data set, then
add noise to the right half of the image, and try to use the patches from the left half to denoise the right half.
Each patch is of size 5 × 5 pixels. We obtain 317 × 236 patches from the left half of the image and use it as
the patch database. Then we apply Algorithm 3 to denoise the image. In particular, for each noisy patch qn
(out of 317 × 237 patches) in the right half of the image, we perform a linear scan to find the closest patch
pi from the patch database, based on the following cost function:
(cid:88)
dist(pj, pi)
5
dist(qn, pi) +
pj∈neighbor(qn)
where dist(p, q) is defined to be the sum of squares of the l2 distances between the colors of corresponding
pixels in the two patches.
After that, for each noisy patch we retrieve the closest patch from the patch database. Then for each
noisy pixel x, we first identify all the noisy patches (there are at most 25 of them) that cover it. The denoised
color of this pixel x is simply the average of all the corresponding pixels in those noisy patches which cover
x.
Since the nearest neighbor algorithm is implemented using a linear scan, it takes around 1 hour to
denoise one image. One could also apply some more advanced techniques like locality sensitive hashing to
find the closest patches with much faster running time.
Acknowledgements The authors would like to thank Pedro Felzenszwalb for formulating the Simul-
taneous Nearest Neighbor problem, as well as many helpful discussions about the experimental setup.
6 2r + 1 approximation
Motivated by the importance of the r-sparse graphs in applications, in this section we focus on them and
present another algorithm (besides INN) which solves the SNN problem for these graphs. We note that
unlike INN, the algorithm presented in this section is not just a pruning step, but it solves the whole SNN
problem.
14
Table 3: Two images from the Berkeley segmentation dataset [27] (size 321 ∗ 481). The first column shows
the original image; the second column shows the half noisy image; the third column shows the de-noised
image using our algorithm for the patches.
For a graph G = (Q, E) of pseudoarboricity r, let the mapping function be f : E → Q, such that for
every e = (qi, qj), f (e) = qi or f (e) = qj, and that for each qi ∈ Q, C(qi) ≤ r, where C(qi) is defined as
{ef (e) = qi}.
Once we have the mapping function f, we can run Algorithm 3 to get an approximate solution. Although
the naive implementation of this algorithm needs O(rkn) running time, by using the aggregate nearest
neighbor algorithm, it can be done much more efficiently. We have the following lemma on the performance
of this algorithm.
Algorithm 3 Algorithm for graph with pseudoarboricity r
Input Query points q1,··· , qk, the input graph G = (Q, E) with pseudoarboricity r
Output An Assignment p1,··· , pk ∈ P
1: for i = 1 to k do
2:
3: end for
Assign pi ← minp∈P dist(qi, p) +(cid:80)
j:(qi,qj )∈C(qj )
dist(p,qj )
r+1
Lemma 6.1. If G has pseudoarboricity r, the solution of Algorithm 3 gives 2r + 1 approximation to the
optimal solution.
Proof. Denote the optimal solution as P ∗ = {p∗
1,··· , p∗
Cost(Q, G, P ∗) =
dist(qi, p∗
i ) +
dist(p∗
i , p∗
(cid:88)
i
(cid:88)
(qi,qj )∈E
k}. We know the optimal cost is
(cid:88)
dist(p∗
(cid:88)
i , qi) +
j ) =
dist(p∗
i , p∗
j )
j:(qi,qj )∈C(qj )
i
15
Let Sol be the solution reported by Algorithm 3. Then we have
Cost(Sol) =
dist(pi, pj)
(cid:88)
j:(qi,qj )∈C(qj )
dist(pi, qj) +
dist(pi, qj)
≤(cid:88)
≤(cid:88)
i
i
≤(r + 1)
≤(r + 1)
dist(qi, pi) +
(cid:88)
(cid:88)
j:(qi,qj )∈C(qj )
j:(qi,qj )∈C(qj )
i
(cid:88)
dist(qi, pi) +
dist(qi, pi) +
(cid:88)
dist(qi, p∗
(cid:88)
dist(qi, p∗
(cid:88)
i
i
i ) +
i ) +
(qi,qj )∈C(qj )
(cid:88)
(cid:88)
(cid:88)
(cid:88)
j:(qi,qj )∈C(qj )
j:(qi,qj )∈C(qj )
(cid:88)
(cid:88)
i
=(r + 1)
dist(qi, pi) +
dist(pi, qj)
j:(qi,qj )∈C(qj )
+ r
(by definition of pseudoarboricity)
j
dist(qj, pj)
dist(qj, pj)
(by triangle inequality)
(cid:88)
(cid:88)
(by the optimality of pi in the algorithm)
(by triangle inequality)
j ) + dist(p∗
r + 1
j , qj)
i , qj)
dist(p∗
r + 1
dist(p∗
i , p∗
i
≤(r + 1)Cost(Q, G, P ∗) +
≤(r + 1)Cost(Q, G, P ∗) + r
dist(p∗
j , qj)
(by definition of pseudoarboricity)
j:(qi,qj )∈C(qj )
dist(p∗
j , qj)
= (2r + 1) Cost(Q, G, P ∗)
j
References
[1] Pedro Felzenszwalb, William Freeman, Piotr Indyk, Robert Kleinberg, and Ramin Zabih. Big-
data: F: Dka: Collaborative research: Structured nearest neighbor search in high dimensions.
http://cs.brown.edu/pff/SNN/, 2015.
[2] Jon Louis Bentley. Multidimensional binary search trees used for associative searching. Communica-
tions of the ACM, 18(9):509 -- 517, 1975.
[3] Sunil Arya, David M Mount, Nathan S Netanyahu, Ruth Silverman, and Angela Y Wu. An optimal
algorithm for approximate nearest neighbor searching fixed dimensions. Journal of the ACM (JACM),
45(6):891 -- 923, 1998.
[4] Piotr Indyk and Rajeev Motwani. Approximate nearest neighbors: towards removing the curse of
dimensionality. In Proceedings of the thirtieth annual ACM symposium on Theory of computing, pages
604 -- 613. ACM, 1998.
[5] Eyal Kushilevitz, Rafail Ostrovsky, and Yuval Rabani. Efficient search for approximate nearest neigh-
bor in high dimensional spaces. SIAM Journal on Computing, 30(2):457 -- 474, 2000.
16
[6] Robert Krauthgamer and James R Lee. Navigating nets: simple algorithms for proximity search. In
Proceedings of the fifteenth annual ACM-SIAM symposium on Discrete algorithms, pages 798 -- 807.
Society for Industrial and Applied Mathematics, 2004.
[7] Alexandr Andoni, Piotr Indyk, Huy L Nguyen, and Ilya Razenshteyn. Beyond locality-sensitive hash-
ing. In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, pages
1018 -- 1028. SIAM, 2014.
[8] William T Freeman, Thouis R Jones, and Egon C Pasztor. Example-based super-resolution. Computer
Graphics and Applications, IEEE, 22(2):56 -- 65, 2002.
[9] Yuri Boykov, Olga Veksler, and Ramin Zabih. Fast approximate energy minimization via graph cuts.
Pattern Analysis and Machine Intelligence, IEEE Transactions on, 23(11):1222 -- 1239, 2001.
[10] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan Goldman. Patchmatch: A randomized
correspondence algorithm for structural image editing. ACM Transactions on Graphics-TOG, 28(3):24,
2009.
[11] Jon Kleinberg and Eva Tardos. Approximation algorithms for classification problems with pairwise
relationships: Metric labeling and markov random fields. Journal of the ACM (JACM), 49(5):616 -- 639,
2002.
[12] Man Lung Yiu, Nikos Mamoulis, and Dimitris Papadias. Aggregate nearest neighbor queries in road
networks. Knowledge and Data Engineering, IEEE Transactions on, 17(6):820 -- 833, 2005.
[13] Yang Li, Feifei Li, Ke Yi, Bin Yao, and Min Wang. Flexible aggregate similarity search. In Proceedings
of the 2011 ACM SIGMOD international conference on management of data, pages 1009 -- 1020. ACM,
2011.
[14] Feifei Li, Bin Yao, and Piyush Kumar. Group enclosing queries. Knowledge and Data Engineering,
IEEE Transactions on, 23(10):1526 -- 1540, 2011.
[15] Pankaj K Agarwal, Alon Efrat, and Wuzhou Zhang. Nearest-neighbor searching under uncertainty. In
Proceedings of the 32nd symposium on Principles of database systems. ACM, 2012.
[16] Tsvi Kopelowitz and Robert Krauthgamer. Faster clustering via preprocessing.
arXiv preprint
arXiv:1208.5247, 2012.
[17] Alexander V Karzanov. Minimum 0-extensions of graph metrics. European Journal of Combinatorics,
19(1):71 -- 101, 1998.
[18] Jittat Fakcharoenphol, Chris Harrelson, Satish Rao, and Kunal Talwar. An improved approximation
algorithm for the 0-extension problem. In Proceedings of the fourteenth annual ACM-SIAM symposium
on Discrete algorithms, pages 257 -- 265. Society for Industrial and Applied Mathematics, 2003.
[19] Gruia Calinescu, Howard Karloff, and Yuval Rabani. Approximation algorithms for the 0-extension
problem. SIAM Journal on Computing, 34(2):358 -- 372, 2005.
17
[20] Aaron Archer, Jittat Fakcharoenphol, Chris Harrelson, Robert Krauthgamer, Kunal Talwar, and ´Eva
In Proceedings of the fifteenth annual
Tardos. Approximate classification via earthmover metrics.
ACM-SIAM symposium on Discrete algorithms, pages 1079 -- 1087. Society for Industrial and Applied
Mathematics, 2004.
[21] James R Lee and Assaf Naor. Metric decomposition, smooth measures, and clustering. Preprint, 2004.
[22] Anupam Gupta, Robert Krauthgamer, and James R Lee. Bounded geometries, fractals, and low-
distortion embeddings. In Foundations of Computer Science, 2003. Proceedings. 44th Annual IEEE
Symposium on, pages 534 -- 543. IEEE, 2003.
[23] Howard Karloff, Subhash Khot, Aranyak Mehta, and Yuval Rabani. On earthmover distance, metric
labeling, and 0-extension. SIAM Journal on Computing, 39(2):371 -- 387, 2009.
[24] Pedro F Felzenszwalb, Gyula Pap, Eva Tardos, and Ramin Zabih. Globally optimal pixel labeling algo-
rithms for tree metrics. In Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference
on, pages 3153 -- 3160. IEEE, 2010.
[25] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient belief propagation for early vision. Inter-
national journal of computer vision, 70(1):41 -- 54, 2006.
[26] Yuri Boykov and Vladimir Kolmogorov. An experimental comparison of min-cut/max-flow algorithms
for energy minimization in vision. Pattern Analysis and Machine Intelligence, IEEE Transactions on,
26(9):1124 -- 1137, 2004.
[27] David R Martin, Charless C Fowlkes, and Jitendra Malik. Learning to detect natural image boundaries
using local brightness, color, and texture cues. Pattern Analysis and Machine Intelligence, IEEE
Transactions on, 26(5):530 -- 549, 2004.
18
|
1204.2933 | 1 | 1204 | 2012-04-13T09:43:27 | Improved Randomized Online Scheduling of Intervals and Jobs | [
"cs.DS"
] | We study the online preemptive scheduling of intervals and jobs (with restarts). Each interval or job has an arrival time, a deadline, a length and a weight. The objective is to maximize the total weight of completed intervals or jobs. While the deterministic case for intervals was settled a long time ago, the randomized case remains open. In this paper we first give a 2-competitive randomized algorithm for the case of equal length intervals. The algorithm is barely random in the sense that it randomly chooses between two deterministic algorithms at the beginning and then sticks with it thereafter. Then we extend the algorithm to cover several other cases of interval scheduling including monotone instances, C-benevolent instances and D-benevolent instances, giving the same competitive ratio. These algorithms are surprisingly simple but have the best competitive ratio against all previous (fully or barely) randomized algorithms. Next we extend the idea to give a 3-competitive algorithm for equal length jobs. Finally, we prove a lower bound of 2 on the competitive ratio of all barely random algorithms that choose between two deterministic algorithms for scheduling equal length intervals (and hence jobs). | cs.DS | cs |
Improved Randomized Online Scheduling of Intervals
and Jobs∗
Stanley P. Y. Fung†
Chung Keung Poon‡
Feifeng Zheng§
November 20, 2018
Abstract
We study the online preemptive scheduling of intervals and jobs (with restarts). Each inter-
val or job has an arrival time, a deadline, a length and a weight. The objective is to maximize the
total weight of completed intervals or jobs. While the deterministic case for intervals was settled
a long time ago, the randomized case remains open. In this paper we first give a 2-competitive
randomized algorithm for the case of equal length intervals. The algorithm is barely random
in the sense that it randomly chooses between two deterministic algorithms at the beginning
and then sticks with it thereafter. Then we extend the algorithm to cover several other cases
of interval scheduling including monotone instances, C-benevolent instances and D-benevolent
instances, giving the same competitive ratio. These algorithms are surprisingly simple but have
the best competitive ratio against all previous (fully or barely) randomized algorithms. Next
we extend the idea to give a 3-competitive algorithm for equal length jobs. Finally, we prove a
lower bound of 2 on the competitive ratio of all barely random algorithms that choose between
two deterministic algorithms for scheduling equal length intervals (and hence jobs).
keywords: interval and job scheduling; preemption with restart; online algorithms; randomized;
lower bound
1
Introduction
In this paper, we study two online preemptive scheduling problems.
In the interval scheduling
problem, we are to schedule a set of weighted intervals which arrive online (in the order of their
left endpoints) so that at any moment, at most one interval is being processed. We can abort the
interval currently being processed in order to start a new one. The goal is to maximize the sum
of the weights of completed intervals. The problem can be viewed as a job scheduling problem in
which each job has, besides its weight, an arrival time, a length and a deadline. Moreover, the
deadline is always tight, i.e., deadline always equals arrival time plus length. Thus, if one does
not start an interval immediately upon its arrival, or if one aborts it before its completion, that
∗The work described in this paper was fully supported by grants from the Research Grant Council of the Hong
Kong SAR, China [CityU 119307] and NSFC Grant No. 60736027 and 70702030.
†Department of Computer Science, University of Leicester, Leicester, United Kingdom.
Email:
[email protected]
‡Department of Computer Science, City University of Hong Kong, Hong Kong. Email: [email protected]
§School of Management, Xi'an Jiaotong University, Xi'an, China. Email: [email protected]
1
interval will never be completed. The problem is fundamental in scheduling and is clearly relevant
to a number of online problems such as call control and bandwidth allocation (see e.g., [2, 6, 19]).
We also study the more general problem of job scheduling with restart. Here, the deadline of
a job needs not be tight and we can abort a job and restart it from the beginning some time
later. Both problems are in fact special cases of the broadcast scheduling problem which gains
much attention recently due to its application in video-on-demand, stock market quotation, etc
(see e.g., [13, 18, 20]). In that problem, a server holding a number of pages receives requests from
its clients and schedules the broadcasting of its pages. A request is satisfied if the requested page is
broadcasted in its entirety after the arrival time and before the deadline of the request. The page
currently being broadcasted can be aborted in order to start a new one, and the aborted page can
be re-broadcasted from the beginning later. Interval and job scheduling with restart can be seen
as a special case in which each request asks for a different page.
Our results concern barely random algorithms, i.e., randomized algorithms that randomly choose
from a very small (constant) number of deterministic algorithms at the beginning and then stick
with it thereafter. Quite some previous work in online scheduling considered the use of barely
random algorithms (see e.g. [1, 9, 17]); it is interesting to consider how the competitiveness improves
(upon their deterministic counterparts) by combining just a few deterministic algorithms. From
now on, whenever we refer to "barely random algorithms", we mean algorithms that choose between
two deterministic algorithms but possibly with unequal probability.
Types of instances.
In this paper, we consider the following special types of intervals or jobs:
1. equal length instances where all intervals or jobs have the same length,
2. monotone instances where intervals arriving earlier also have earlier deadlines, and
3. C- and D-benevolent instances where the weight of an interval is given by some 'nice' function
of its length (convex increasing for C-benevolent, and decreasing for D-benevolent).
The models will be defined precisely in the next section. These cases are already highly non-trivial,
as we will see shortly, and many previous works on these problems put further restrictions on the
inputs (such as requiring jobs to be unweighted or arrival times to be integral, in addition to being
equal-length). The power of randomization for these problems is especially unclear.
1.1 Previous work
The general case where intervals can have arbitrary lengths and weights does not admit constant
competitive algorithms [19], even with randomization [6]. Therefore, some special types of instances
have been studied in the literature.
We first mention results for equal length interval scheduling. The deterministic case was settled
in [19] where a 4-competitive algorithm and a matching lower bound were given. Miyazawa and
Erlebach [16] were the first to give a better randomized algorithm: its competitive ratio is 3 but it
only works for a special case where the weights of the intervals form a non-decreasing sequence. They
also gave the first randomized lower bound of 5/4. The first randomized algorithm for arbitrary
weight that has competitive ratio better than 4 (the bound for deterministic algorithms) was devised
in [12]. It is 3.618-competitive and is barely random, choosing between two deterministic algorithms
with equal probability. In the same paper, a lower bound of 2 for such barely random algorithms and
2
equal length
upper bound
2.455 [11]
3.227 (barely random) [11]
lower bound
1.693 [11]
2 (barely random) [this paper]
2 (barely random) [this paper]
monotone
C-benevolent
same as above
3.732 [17]
2 (barely random) [this paper]
same as above
1.693 [11]
D-benevolent
2.455 [11]
1.5 [11] (with a surjective condition)
3.227 (barely random) [11]
2 (barely random) [this paper]
Table 1: Best previous and new results for randomized interval scheduling
a lower bound of 4/3 for general randomized algorithms were also proved. Recently, Epstein and
Levin [11] gave a 2.455-competitive randomized algorithm and a 3.227-competitive barely random
algorithm. They also gave a 1.693 lower bound on the randomized competitive ratio.
The class of monotone instances (also called similarly ordered [9] or agreeable [15] instances
in the literature) is a generalization of the class of equal length instances. Therefore, the former
class inherits all the lower bounds for the latter class. In the offline case, the class of monotone
instances is actually equivalent to that of equal length instances because of the result (see e.g. [4])
that the class of proper interval graphs (intersection graphs of intervals where no interval is strictly
contained in another) is equal to the class of unit interval graphs. In the online case however, it is
not completely clear that such an equivalence holds although some of the algorithms for the equal
length case also work for the monotone case (e.g. [16, 12, 11]).
Some of the aforementioned results for equal length instances also work for C- and D-benevolent
instances, including Woeginger's 4-competitive deterministic algorithm, the lower bound of 4/3 in
[12]1, the upper bounds in [11] (for D-benevolent instances only) and the lower bound in [11] (for C-
benevolent instances only; they gave another slightly weaker lower bound of 3/2 for D-benevolent
instances). A 3.732-competitive barely random algorithm for C-benevolent instances was given
by Seiden [17]. Table 1 summarizes the various upper and lower bounds for randomized interval
scheduling.
Next we consider the problem of job scheduling with restarts. Zheng et al.
[20] gave a 4.56-
competitive deterministic algorithm. The algorithm was for the more general problem of scheduling
broadcasts but it works for jobs scheduling with restarts too. We are not aware of previous results
in the randomized case. Nevertheless, Chrobak et al. [9] considered a special case where the jobs
have no weights and the objective is to maximize the number of completed jobs. For the randomized
nonpreemptive case they gave a 5/3-competitive barely random algorithm and a lower bound of 3/2
for barely random algorithms that choose between two deterministic algorithms. They also gave
an optimal 3/2-competitive algorithm for the deterministic preemptive (with restart) case, and a
lower bound of 6/5 for the randomized preemptive case.
We can also assume that the time is discretized into unit length slots and all (unit) jobs can
only start at the beginning of each slot. Being a special case of the problem we consider in this
paper, this version of unit job scheduling has been widely studied and has applications in buffer
1 This and most other lower bounds for D-benevolent instances only work for a subclass of functions that satisfy
a surjective condition.
3
management of QoS switches. For this problem, a e/(e − 1)-competitive randomized algorithm was
given in [7], and a randomized lower bound of 1.25 was given in [8]. The current best deterministic
algorithm is 1.828-competitive [10].
An alternative preemption model is to allow the partially-executed job to resume its execution
from the point that it is preempted. This was studied, for example, in [3, 14].
1.2 Our results
In this paper we give new randomized algorithms for the different versions of the online interval
scheduling problem. They are all barely random and have a competitive ratio of 2. Thus they
substantially improve previous results. See Table 1. It should be noted that although the algorithms
are fairly simple, they were not discovered in several previous attempts by other researchers and
ourselves [11, 12, 16]. Moreover the algorithms for all these versions of the problem are based on
the same idea, which gives a unified way of analyzing these algorithms that were not present in
previous works.
Next we extend the algorithm to the case of job scheduling (with restarts), and prove that it
is 3-competitive. This is the first randomized algorithm we are aware of for this problem. The
extension of the algorithm is very natural but the proof is considerably more involved.
Finally we prove a lower bound of 2 for barely random algorithms for scheduling equal length
intervals (and jobs) that choose between two deterministic algorithms, not necessarily with equal
probability. Thus it matches the upper bound of 2 for this class of barely random algorithms.
Although this lower bound does not cover more general classes of barely random or randomized
algorithms, we believe that this is still of interest. For example, a result of this type appeared in
[9]. Also, no barely random algorithm using three or more deterministic algorithms with a better
performance is known. The proof is also much more complicated than the one in [12] with equal
probability assumption.
2 Preliminaries
A job J is specified by its arrival time r(J), its deadline d(J), its length (or processing time) p(J)
and its weight w(J). All r(J), d(J), p(J) and w(J) are nonnegative real numbers. An interval is a
job with tight deadline, i.e. d(J) = r(J) + p(J). We further introduce the following concepts for
intervals: for intervals I and J with r(I) < r(J), I contains J if d(I) ≥ d(J); if r(J) < d(I) < d(J),
the two intervals overlap; and if d(I) ≤ r(J) < d(J), the intervals are disjoint.
Next we define the types of instances that we consider in this paper. The equal length case is
where p(J) is the same for all J; without loss of generality we can assume p(J) = 1. The remaining
notions apply to intervals only. An instance is called monotone if for any two intervals I and J,
if r(I) < r(J) then d(I) ≤ d(J). An instance is called C-benevolent if the weights of intervals are
given by a function f of their lengths, where the function f satisfies the following three properties:
(i) f (0) = 0 and f (p) > 0 for all p > 0,
(ii) f is strictly increasing, and
(iii) f is convex, i.e. f (p1) + f (p2) ≤ f (p1 − ǫ) + f (p2 + ǫ) for 0 < ǫ ≤ p1 ≤ p2.
Finally, an instance is called D-benevolent if the weights of intervals are given by a function f of
their lengths where
4
(i) f (0) = 0 and f (p) > 0 for any p > 0, and
(ii) f is decreasing in (0, ∞).
In our analysis, we partition the time axis into segments called slots, s1, s2, . . ., such that each
time instant belongs to exactly one slot and the union of all slots cover the entire time axis. The
precise way of defining the slots depends on the case being studied (equal-length, monotone, C- or
D-benevolent instances). Slot si is an odd slot if i is odd, and is an even slot otherwise.
The following is an important, though perhaps unusual, definition used throughout the paper.
We say that a job (or an interval) is accepted by an algorithm A in a slot s if it is started by A
within the duration of slot s and is then completed without interruption. Note that the completion
time may well be after slot s. A may start more than one job in a slot, but it will become clear that
for all online algorithms that we consider, at most one job will be accepted in a slot; all other jobs
that were started will be aborted. For OP T we can assume that it always completes each interval
or job it starts.
The value of a schedule is the total weight of the jobs that are completed in the schedule. The
performance of online algorithms is measured using competitive analysis [5]. An online randomized
algorithm A is c-competitive if the expected value obtained by A is at least 1/c the value obtained
by the optimal offline algorithm, for any input instance. The infimum of all such c is called the
competitive ratio of A. We use OP T to denote the optimal algorithm (and its schedule).
3 Algorithms for Scheduling Intervals
3.1 Equal Length Instances
In this section we describe and analyse a very simple algorithm RAN for the case of equal length
intervals. RAN is barely random and consists of two deterministic algorithms A and B, described
as follows. The time axis is divided into unit length slots, s1, s2, . . ., where slot si covers time
[i − 1, i) for i = 1, 2, . . ..
Intuitively, A takes care of odd slots and B takes care of even slots.
Within each odd slot si, A starts the interval arriving first. If a new interval arrives in this slot
while an interval is being processed, A will abort and start the new interval if its weight is larger
than the current interval; otherwise the new interval is discarded. At the end of this slot, A is
running (or about to complete) an interval with the largest weight among those that arrive within
si; let Ii denote this interval. A then runs Ii to completion without abortion during the next (even)
slot si+1. (Thus, Ii is the only interval accepted by A in slot si.) Algorithm A then stays idle until
the beginning of the next odd slot. B runs similarly on even slots. RAN chooses one of A and B
with equal probability 1/2 at the beginning.
Theorem 3.1 RAN is 2-competitive for online interval scheduling on equal length instances.
Proof. Each Ii is accepted by either A or B. Therefore, RAN completes each Ii with probability
1/2. On the other hand, OP T can accept at most one interval in each slot si, with weight at most
w(Ii). It follows that the total value of OP T is at most 2 times the expected value of RAN .
✷
Trivial examples can show that RAN is not better than 2-competitive (e.g. a single interval).
In fact we will show in Section 5 that no barely random algorithm that chooses between two
deterministic algorithms is better than 2-competitive. But first we consider how this result can be
generalized to other types of instances.
5
3.2 Monotone Instances
Algorithm RAN -M . We adapt the idea of RAN to the case of monotone instances and call the
algorithm RAN -M . Similar to RAN , RAN -M consists of two deterministic algorithms A and B,
each chosen to execute with probability 1/2 at the beginning. The difference is that we cannot use
the idea of unit length slots but we must define the lengths of the slots in an online manner.
The execution of the algorithm is divided into phases and we name the slots in each phase
locally as s1, s2, . . . independent of other phases. After the end of a phase and before the beginning
of the next phase, the algorithm (both A and B) is idle with no pending intervals. A new phase
starts when the first interval arrives while the algorithm is idle. Among all intervals that arrive at
this time instant, let I0 be the one with the earliest deadline (ties broken arbitrarily). Then slot
s1 is defined as [r(I0), d(I0)). A aims to accept the heaviest interval among those with arrival time
falling within slot s1. To do this, A simply starts the first interval arriving in s1, and then whenever
a new interval arrives that is heavier than the interval that A is currently executing, A aborts the
current one and starts the new heavier interval. This is repeated until the time d(I0) is reached.
By the property of monotone instances and the choice of I0, these intervals all have finishing time
on or after d(I0). Let I1 denote the interval that A is executing (or about to complete) at the end
of slot s1, i.e., time d(I0). B remains idle during the whole slot. If A just finishes I1 at time d(I0),
then it will become idle again and this phase ends. Otherwise, d(I1) > d(I0) and slot s2 is now
defined as [d(I0), d(I1)).
In slot s2, A continues to execute I1 to completion without any interruption. (Thus, I1 is the
only interval accepted by A in slot s1.) B accepts the heaviest interval among those with arrival
time falling within slot s2, in the same manner A did in the previous slot. This interval is denoted
by I2 and B will run it to completion during slot s3 (if its deadline is after the end of slot s2).
In general, slot si (where i > 1) is defined as [d(Ii−2), d(Ii−1)). If i is odd, then at the beginning
of slot si, B is executing Ii−1 (the interval accepted by B in slot si−1) and A is idle. B will run
Ii−1 to completion while A will accept the heaviest interval among those arriving during this slot.
If i is even, the actions are the same except that the roles of A and B are reversed.
Theorem 3.2 RAN-M is 2-competitive for online interval scheduling on monotone instances.
Proof. No interval will arrive during the idle time between phases (since otherwise RAN -M would
have started a new phase), so each phase can be analyzed separately. Each interval completed by
OP T will be analyzed according to the slot its arrival time falls into.
In each slot si, OP T can accept at most one interval: This is true for s1 by the way s1 is
chosen. For i > 1, consider the first interval I ′ accepted by OP T in slot si = [d(Ii−2), d(Ii−1)).
(Recall that accepting a job means starting the job and then executing it to completion without
interruption.) Since the start of slot si is after r(Ii−1), we have r(I ′) > r(Ii−1). By the monotone
property, d(I ′) ≥ d(Ii−1). So, OP T cannot accept another interval in slot si. The rest of the proof
is the same as the equal length case, namely, that the interval accepted by OP T in each slot has
weight at most that of the interval accepted by A or B in the same slot. It follows that RAN -M
is 2-competitive.
✷
3.3 C-benevolent Instances
Algorithm RAN -C. Once again, the algorithm for C-benevolent instances RAN -C consists of
two deterministic algorithms A and B, each with probability 1/2 of being executed. The execution
6
of the algorithm is divided into phases as in the monotone case.
When a new phase begins, the earliest arriving interval, denoted by I0, defines the first slot s1,
i.e., s1 = [r(I0), d(I0)). (If there are several intervals arriving at the same time, let I0 be the one
with the longest length.) We first describe the processing of intervals in slot s1, which is slightly
different from the other slots. First, B starts and completes I0. During s1, A accepts the longest
interval among those with arrival time during (r(I0), d(I0)) and finishing time after d(I0). Denote
this interval by I1. (Note that there may be other intervals that arrive and end before I1 arrives.
Naturally, A could finish them in order to gain more value. However, to simplify our analysis, we
assume that A will not process them.) If there is no such I1, i.e., no interval arrives within s1 and
ends after d(I0), the phase ends at the end of s1.
Suppose I1 exists. Then define slot s2 as [d(I0), d(I1)). A uses the entire slot s2 to complete I1
without interruption. After completing I0 at time d(I0), B accepts the longest interval (denoted
I2) among those arriving within slot s2 and finishing after d(I1), in a way similar to the action of A
in the previous slot. Again, if such an I2 does not exist, the phase ends at the end of s2. Otherwise,
slot s3 is defined as [d(I1), d(I2)) and B will complete I2 that ends after d(I1). Similarly, after
A finishes I1 in time d(I1), it starts the longest interval (denoted by I3) arriving during s3 and
finishing after d(I2), and so on.
In general, slot si (for i > 1) is defined as [d(Ii−2), d(Ii−1)). If i is odd, then B takes the entire
slot to complete the interval Ii−1 without interruption while A accepts the longest interval Ii that
arrives during slot si and ends after d(Ii−1). If i is even then the roles of A and B are reversed.
Competitive Analysis. We first state the following useful lemma which holds for any C-benevolent
function f .
Lemma 3.1 For any C-benevolent function f , given any k + 1 positive real numbers pi (1 ≤ i ≤ k)
and P , if P ≥ Pk
Proof. f (P ) ≥ f (Pk
i=1 pi, then f (P ) ≥ Pk
i=1 pi) ≥ f (p1) + f (Pk
i=1 f (pi). ✷
i=1 f (pi).
i=2 pi) ≥ P2
i=1 f (pi) + f (Pk
i=3 pi) ≥ . . . ≥ Pk
Theorem 3.3 RAN -C is 2-competitive for online interval scheduling on C-benevolent instances.
Proof. As a first step to the proof we simplify the OP T schedule. Within each slot si in a
phase, i ≥ 1, OP T starts a sequence of disjoint intervals (in increasing order of starting times)
Oi = {oi,1, oi,2, . . . , oi,ki}. Only the last interval, oi,ki, may end later than d(Ii−1) (the ending time
of si). If it does, then we merge oi,1, oi,2, . . . , oi,ki−1 into one interval prei such that r(prei) = r(oi,1)
and d(prei) = r(oi,ki), and thus p(prei) = r(oi,ki) − r(oi,1) ≥ Pki−1
j=1 p(oi,j). By Lemma 3.1,
f (p(prei)) ≥ Pki−1
j=1 f (p(oi,j)). Otherwise, (i.e. oi,ki ends before d(Ii−1)), we merge all the intervals
in Oi into one interval prei such that r(prei) = r(oi,1) and p(prei) = d(oi,ki)−r(oi,1) ≥ Pki
j=1 p(oi,j).
Thus, in both cases, such merging can only make OP T 's value larger. So we can assume that OP T
starts at most two intervals prei and oi,ki in slot si. After understanding the notations, we simply
denote the two intervals prei and oi,ki by oi,1 and oi,2, respectively.
The interval oi,1 (if exist) is contained in Ii−1 and so p(oi,1) ≤ p(Ii−1). The interval oi,2 (if exist)
will end after d(Ii−1), and p(oi,2) ≤ p(Ii) since Ii is defined to be the longest interval that arrives
during slot si and ends after d(Ii−1). Note that oi,2 may also end after d(Ii+l) for some l ≥ 0. In
this case, neither ov,1 nor ov,2 exist for i ≤ v ≤ i + l. If any oi,1 or oi,2 does not exist, we set its
length to zero.
7
We now analyze the competitive ratio of RAN -C. As in the monotone case, each phase can be
analyzed separately. Consider an arbitrary schedule S = {I0, I1, . . . , In−1} produced by RAN -C in
a phase with n ≥ 1 slots, where Ii overlaps Ii+1 (0 ≤ i < n − 1), and the corresponding schedule
S∗ = {o1,1, o1,2, o2,1, . . . , on,1} produced by OP T as RAN -C produces S. (Note that on,2 cannot
exist since otherwise this means there are some intervals that arrive within [d(In−2), d(In−1)) and
end after d(In−1), and hence the phase will not end and RAN -C will start an In.)
For each slot i, OP T starts two intervals oi,1 and oi,2 while RAN -C accepts Ii−1. For presenta-
tion convenience, let xi,1 = p(oi,1), xi,2 = p(oi,2) and yi = p(Ii). We already have that xi,1 ≤ yi−1
for 1 ≤ i ≤ n and xi,2 ≤ yi for 1 ≤ i < n. We will show that
n
X
i=1
f (xi,1) +
n−1
X
i=1
f (xi,2) ≤
n−1
X
i=0
f (yi).
(1)
The left hand side of (1) represents the total weight of intervals in S∗ (note that on,2 does
not exist) while the right hand side represents the total weight of intervals in S. Since RAN -C
completes each interval in S with probability 1/2, its expected value is half of the right hand side
of (1). Thus by proving (1) we show the 2-competitiveness of RAN -C.
i=1 f (xi,1) + Pk−2
i=1 f (xi,2) ≤ Pk−2
i=1 f (xi,2) ≤ Pk−1
i=0 f (yi) and thus the claim holds for n = k.
We prove (1) by induction on n. When n = 1, (1) reduces to f (x1,1) ≤ f (y0) which is true since
x1,1 ≤ y0. Assume the claim holds for n = k − 1, i.e., Pk−1
i=1 f (yi).
Consider Ik, ok−1,2 and ok,1. We have xk−1,2 ≤ yk−1 and xk,1 ≤ yk−1. If xk−1,2 + xk,1 ≤ yk−1, then
f (xk−1,2) + f (xk,1) ≤ f (xk−1,2 + xk,1) ≤ f (yk−1). Adding this to the induction hypothesis gives
Pk
i=1 f (xi,1) + Pk−1
Otherwise, if xk−1,2 + xk,1 > yk−1, we first change the schedule S∗ as follows: we increase the
length of xk,1 to yk−1 and decrease the length of xk−1,2 by the same amount. The corresponding
r(ok−1,2) and d(ok,1) are fixed while both d(ok−1,2) and r(ok,1) decrease by an amount of yk−1 −xk,1.
OP T will only get better since f (xk,1) + f (xk−1,2) ≤ f (yk−1) + f (xk−1,2 − (yk−1 − xk,1)) by the
properties of C-benevolent functions. After this change, Ik−1 and ok,1 have the same length. The
new ok−1,2 now ends on or before d(Ik−2). We merge the new ok−1,2 into ok−1,1 so that the new
ok−1,1 extends its length to xk−1,2 + xk−1,1 and keeps its start time r(ok−1,1) unchanged. In the case
that xk−1,1 = 0 before merging ok−1,2, we set r(ok−1,1) = r(ok−1,2). The new ok−1,1 is still contained
by Ik−2 and thus xk−1,1 ≤ yk−2 still holds. After merging, xk−1,2 = 0 and xk,1 = yk−1. Therefore
Pk
i=0 f (yi) + f (yk−1) =
Pk−1
✷
i=1 f (xi,2) + f (xk,1) ≤ Pk−2
i=1 f (xi,1) + Pk−1
i=0 f (yi). Thus the claim is true for n = k.
i=1 f (xi,2) = Pk−1
i=1 f (xi,1) + Pk−2
3.4 D-benevolent Instances
Algorithm RAN -D. The basic idea of RAN -D is same as RAN : two algorithms A or B are
executed each with probability 1/2. Intuitively, in an odd slot (where slots will be defined precisely
in the following paragraphs), A accepts the largest-weight interval arriving during that slot, by
starting an interval and preempting if a new one arrives with a larger weight. We call the interval
being executed by A the main interval, denoted by IM . Meanwhile, B continues to run to com-
pletion the interval started in the previous slot; we call this the residual interval, denoted by IR.
This residual interval must be completed (as in the equal length case) because this is the interval
accepted in the previous slot. However in the D-benevolent case, if a shorter (and therefore larger
weight) interval arrives, the residual interval can actually be preempted and replaced by this new
8
interval. For even slots the roles of A and B are reversed (and the interval started by B is the main
interval and the one completed by A the residual interval).
Unlike RAN -M or RAN -C, here when slot si−1 finishes, the next slot si is not completely
determined: slot si begins where si−1 ends, but the ending time of slot si will only get a provisional
value, which may become smaller (but not larger) later on. This is called the provisional ending
time of the slot, denoted by ei. Slots will also be grouped into phases as in the other types of
instances.
Note that IM , IR and ei change during the execution of the algorithm, even within the same
slot. But RAN -D always maintains the following invariant:
Invariant: Suppose IR and IM are the residual and main interval respectively during
execution in a slot si. Then ei = d(IR) ≤ d(IM ) (if the intervals exist). Moreover ei
can only be decreased, not increased.
We describe the processing of intervals in a slot si (i ≥ 1). Consider an odd slot si (the case of
even slots is the same with the roles of A and B reversed). At the beginning of si, A is idle and B
is continuing the execution of a residual interval IR. At this point ei is provisionally set to d(IR).
In the case of the first slot, there is no residual interval left over from the previous slot, so we set
ei to be the deadline of the first interval that arrives. If more than one interval arrive at the same
instant, choose anyone.
Consider a time during si when an interval I arrives while A and B are respectively executing
some intervals IM and IR. If more than one interval arrive at the same instant, process them in
any order. If A or B is idle, assume IM or IR to have weight 0. Then A and B react according to
the following three cases:
1. If d(I) ≥ ei and w(I) > w(IM ), then I preempts IM , and this I becomes the new IM . In this
case, ei remains unchanged.
2. If d(I) < ei (which implies w(I) ≥ w(IM ) and w(I) ≥ w(IR) because by the invariant,
d(IM ) ≥ d(IR) = ei > d(I), and I arrives no earlier than either IM or IR, and thus I is
shorter), then I preempts both IM in A and IR in B. Here I becomes the new IM and IR,
and ei is then set to d(I).
3. Otherwise, w(I) ≤ w(IM ) and I is discarded.
Observe that the invariant is always maintained when we change any of IM , IR or ei.
This process repeats until time ei is reached and slot si ends. If d(IM ) > ei at the end of slot
si, then a new slot si+1 begins where slot si ends. A has not finished execution of IM yet, so it
now becomes the IR of slot si+1, and ei+1 is provisionally set to d(IR). Otherwise, d(IM ) = ei and
A just finishes execution of IM , then the phase ends. In this case we wait until the next interval
arrival, then a new phase starts.
Note that RAN -D needs to simulate the execution of both A and B (to determine when slots
end) but the actual execution follows only one of them.
Theorem 3.4 RAN -D is 2-competitive for online interval scheduling on D-benevolent instances.
9
Proof. Consider each slot si = [ei−1, ei). We claim that OP T can start at most one interval in
si and that this interval cannot finish strictly before ei. The first part of the claim follows from
the second since if OP T starts two or more intervals within si, then the first such interval must
end strictly before ei. Assume to the contrary that OP T starts an interval I that finishes strictly
before ei. Then I also finishes strictly before the provisional value of ei at the moment I arrives,
since the provisional ending time only decreases. By the design of the algorithm, at that point ei
will be reduced to d(I). ei may be reduced further subsequently, but in any case this contradicts
the fact that d(I) < ei. Hence the claim follows.
Now suppose OP T starts an interval I in an odd slot si and eventually completes it. We will
show that if si is not the last slot in the phase, A will complete an interval of weight no less than
w(I) in slot si+1; if si is the last slot, then A will complete an interval of weight no less than w(I)
in slot si.
Consider the moment when I arrives in si. If I has larger weight than the current IM , A will
preempt it and start I. Thus, by the end of si, A should have started a main interval IM of weight
at least w(I). If this is the last slot, then A completes IM at the end of si. Otherwise, IM becomes
the residual interval in slot si+1 and A will execute it to completion (as an residual interval) in si+1
unless another interval I ′ arrives in si+1 such that d(I ′) < ei+1 (and hence w(I ′) ≥ w(IM )). Note
that ei+1 will then be reduced to d(I ′). This I ′ may still be preempted by intervals of even larger
weight and earlier deadline. In any case, at exactly the end of the next slot si+1, A would have
completed the residual interval.
We can make a similar claim for even slots. Therefore it follows that, for every interval started
by OP T , either A or B will complete an interval of at least the same weight in the same or the
next slot. Thus the total value of A and B is no less than that of OP T . The 2-competitiveness
then follows since each of A/B is executed with 1/2 probability.
✷
4 Algorithms for Equal Length Jobs
Algorithm RAN -J.
In this section we extend RAN to the online scheduling of equal length jobs
with restarts. The algorithm remains very simple but the analysis is more involved. Again RAN -J
chooses between two deterministic algorithms A and B, each with probability 1/2, and again A
takes care of odd slots and B takes care of even slots, where the slots are defined as in the equal
length interval case (i.e. they all have unit length). At the beginning of each odd slot, A considers
all pending jobs that can still be completed, and starts the one with the largest weight. (If there
are multiple jobs with the same maximum weight, start an arbitrary one.) If another job of a larger
weight arrives within the slot, A aborts the current job and starts the new one instead. At the end
of this odd slot, the job that is being executed will run to completion (into the following even slot)
without abortion. A will then stay idle until the beginning of the next odd slot. Even slots are
handled by B similarly.
The following simple example (see Figure 1(a)) illustrates the algorithm, and shows that RAN -
J is not better than 3-competitive. Consider three jobs X, Y, Z, where r(X) = 0, d(X) = 3, w(X) =
1 + ǫ for arbitrary small ǫ > 0; r(Y ) = 0, d(Y ) = 1, w(Y ) = 1; and r(Z) = 1, d(Z) = 2, w(Z) = 1.
Both A and B will complete X only, but OP T can complete all three.
Notations. We define some additional notations that will be used in the rest of this section
to make our discussion clearer. The notation [s1..s2] denotes a range of slots from slot s1 to s2
10
X
Z
X
Y
X
Y
X
(a)
OPT
A
B
I
I
X
Z
Z
OPT
A
B
Z
J
J
Y
(b)
K
K
Y
Y
X
X
Y
Figure 1: (a) An example showing RAN -J is not better than 3-competitive.
(b) An example
showing the charges and a bad slot. The weight of the jobs are (for small ǫ > 0): w(X) =
1 + ǫ, w(Y ) = 1 + 2ǫ, w(Z) = 1 + 3ǫ; w(I) = w(J) = w(K) = 1. Slot 4 is a bad slot.
inclusive, where s1 is before s2. Arithmetic operators on slots carry the natural meaning, so s + 1
is the slot immediately after s, s − 1 is the slot immediately before s, s1 < s2 means s1 is before s2,
etc. The job accepted by an algorithm A in slot s is denoted by A(s). (Any algorithm can accept
at most one job in each slot since the slot has the same length as a job.) We define the inverse
A−1(x) to be the slot s with A(s) = x, if it exists; otherwise it is undefined.
Charging scheme. Our approach to the proof is to map (or charge) the weights of jobs accepted
by OP T to slots where A or B have accepted 'sufficiently heavy' jobs; namely, that each slot s
receives a charge at most 1.5 times of w(A(s)) or w(B(s)). In some cases this is not possible and
we pair up slots with large charges with slots with small charges so that the overall ratio is still at
most 1.5. Since each job in A or B is completed with probability 1/2 only, the expected value of
the online algorithm is half the total value of A and B. This gives a competitiveness of 3.
The charging scheme is defined as follows. Consider a slot s where OP T accepts the job OP T (s).
Suppose s is odd (so A is choosing the heaviest job to start). If w(A(s)) ≥ w(OP T (s)), charge the
weight of OP T (s) to s. We call this a downward charge. Otherwise, A must have accepted OP T (s)
at some earlier slot s′. Charge half the weight of OP T (s) to this slot s′. This is called a self charge.
For B, either it has accepted the job OP T (s) before s, in which case we charge the remaining half
to that slot (this is also a self charge); or OP T (s) is still pending at slot s − 1, which means at
slot s − 1, B accepts a job with weight at least w(OP T (s)). Charge the remaining half to the slot
s − 1. This is called a backward charge. When s is an even slot the charges are similarly defined.
Clearly, all job weights in OP T are charged to some slots. Observe that for each charge from
OP T to a slot, the weight of the job generating the charge is no more than that of the job accepted
in the slot receiving the charge. We define each downward charge to be of one unit, and each self
or backward charge to be of 0.5 unit. With this definition, if every slot receives at most 1.5 units
of charge, then we are done. Unfortunately, slots can receive up to 2 units of charges because a
slot can receive at most one charge of each type. Slots receiving 2 units of charges are called bad;
they must receive a backward charge. Slots with at most 1 unit charge are called good. Each bad
slot s can be characterized by a pair (X, Y ) where X is the job A(s) or B(s), and Y is the job
OP T (s + 1) generating the backward charge. The example in Figure 1(b) illustrates the charges
11
and the existence of bad slots.
Competitive Analysis. The key part of the proof is to deal with bad slots. For each bad slot,
we pair it up with a good slot so that the 'overall' charge is still under a ratio of 1.5. The proof of
the following lemma will show how this is done.
Lemma 4.1 For each bad slot s = (X, Y ), there is a good slot s′ such that the weight of A(s′) or
B(s′) is at least w(Y ). Moreover, any two bad slots are paired with different good slots.
If Lemma 4.1 is true, then we have
Lemma 4.2 Slots s and s′ as defined in Lemma 4.1 together receive a charge at most 1.5 times
the total weight of the jobs in A/B in the two slots.
Proof. Let ws and ws′ be the weight of jobs accepted by A/B in s and s′ respectively. The charges
to s is at most 1.5ws + 0.5w(Y ) while the charges to s′ is at most ws′. The overall ratio is therefore
(1.5ws + 0.5w(Y ) + ws′)/(ws + ws′) ≤ 1.5 since w(Y ) ≤ ws′.
✷
Theorem 4.1 RAN -J is 3-competitive for the online scheduling of equal length jobs with restarts.
Proof. All the weights of jobs accepted by OP T are charged to slots in A or B. Each slot in A
and B receives charges at most 1.5 times the weight of the job in the slot, either as a single slot
or as a pair of slots as defined in Lemma 4.1. Since each job in A or B is only completed with
probability 1/2, the expected value of the online algorithm is half the total weight of jobs in A and
B. It follows that the competitive ratio is 3.
✷
Before proving Lemma 4.1 we first show some properties of bad slots in the following lemma.
(Although the lemma is stated in terms of odd slots, the case of even slots is similar.)
Lemma 4.3 For each bad slot s = (X, Y ), where s is an odd slot,
(i) both X and Y are accepted by B in some slots before s (call the slots s0 and s1, where s0 < s1);
and
(ii) for each odd slot s in [s0 + 1..s1 − 1], w(A(s)) ≥ w(B(s0)) ≥ w(Y ), and for each odd slot s in
[s1 + 1..s − 1], w(A(s)) ≥ w(B(s1)) ≥ w(Y ).
Proof.
(i) Since slot s + 1 makes a backward charge instead of a downward charge, we have
w(B(s + 1)) < w(Y ). Hence B must have accepted Y before s, or else Y could have been a
candidate for B(s + 1). Furthermore, w(X) ≥ w(Y ) > w(B(s + 1)). By the same reasoning, B
must have accepted X before s.
(ii) If B(s0) = Y , then Y has already arrived before the end of slot s0 but is not accepted by
A at/before s. Hence A must have accepted jobs with weights at least w(Y ) in all odd slots in
[s0 + 1..s1 − 1]. If B(s0) = X then the same reasoning implies that A accepted jobs with weights
at least w(X), which is at least w(Y ), in these slots. The same argument holds for A[s1 + 1..s]. ✷
We now prove Lemma 4.1. We give a step-by-step procedure for identifying a good slot (in
which A or B has accepted a job of sufficient weight) for every bad slot. Consider an odd bad slot
12
s = (X, Y ). (The case for even slots is similar.) Roughly speaking, the procedure initially identifies
the two slots s0 and s1 defined in Lemma 4.3 and designates s1 as a special slot, denoted by s∗.
Then it checks if s∗ or s∗ − 1 is a good slot. If a good slot is found, the procedure stops. Otherwise,
it will identify a new slot not found before, pick a new special slot s∗ from among the identified
slots; and then move to the next step (which checks on s∗, s∗ − 1 and so on).
In more detail, at the beginning of step i (i ≥ 1), a collection of i + 1 slots, s0 < s1 < · · · < si,
have been identified. They are all even slots before the bad slot s and one of them is designated as
the special slot s∗. Denote by Yj the job B(sj) for all j ∈ {0, . . . , i} and for convenience, let si+1
denote s. Step i proceeds as follows:
Step i.1. Consider the job Y ∗ = B(s∗) in slot s∗. By Lemma 4.4(i) below, Y ∗ has weight at least
w(Y ). So, if the slot s∗ receives at most 1 unit of charge, then we have identified a good slot
of sufficient weight and we stop.
Step i.2. Otherwise, s∗ has at least 1.5 unit of charge and must therefore have a downward charge.
Denote by Z the job A(s∗ − 1). By Lemma 4.4(ii) below, w(Z) ≥ w(Y ). Since slot s∗ must
have a downward charge, slot s∗ − 1 cannot receive a backward charge. If slot s∗ − 1 receives
no self charge as well, then it is a good slot and we are done.
Step i.3. Otherwise s∗−1 receives a self charge and hence Z is accepted by OP T in some slot s′ after
s∗. In Lemma 4.5, we will show that B must also accept Z at a slot s′′ where s′′ < min{s, s′}.
Note that Z is not in {Y0, Y1, . . . , Yi}. (A job in {Y0, Y1, . . . , Yi} is either the job Y , which is
not accepted by A before slot s, or a job accepted by A in a slot other than s∗ − 1.) Therefore,
s′′ is a different slot than s0, s1, . . . , si.
Mark slot s0 or s′′, whichever is later, as the new special slot s∗. Re-index s0, . . . , si and s′′
as s0 < s1 < · · · < si+1 and move on to Step (i + 1).
We need to show that (i) the procedure always terminates, (ii) the claims made in the above
procedure are correct, and (iii) any two bad slots are paired with different good slots following this
procedure. The first is easy: note that in each step, if a good slot is not found, a new slot, s′′, which
is before s, is identified instead. But there are only a finite number of slots before s. Therefore,
the procedure must eventually terminate and return a good slot.
The claim in Step i.1 and i.2 is proved in the lemma below, which is basically a generalization
of Lemma 4.3.
Lemma 4.4 For any step i (i ≥ 1) and any j ∈ {0, . . . , i},
(i) w(Yj) ≥ w(Y ) and
(ii) for all odd slots s in [sj + 1 . . . sj+1 − 1], w(A(s)) ≥ w(Yj).
Proof. The proof is by induction on i. Clearly, (i) and (ii) are true for i = 1 as proved by Lemma
4.3. Suppose (i) and (ii) are true at the beginning of some step i. We will show that they are
maintained at the beginning of step i + 1.
Recall that Z is accepted by A in slot s∗ − 1 and by B in slot s′′. By (ii), w(Z) ≥ w(Y ). Thus,
w(B(s′′)) ≥ w(Y ) and hence (i) is maintained in the next step.
To show that (ii) is also maintained in the next step, it suffices to show that for any odd slot
s in [s′′ + 1 . . . sq − 1] where sq is the closest slot among s0, s1, . . . , si after s′′ (or sq = s if s′′ lies
after si), w(A(s)) ≥ w(Z). We consider two cases:
13
If s′′ < s∗ − 1, then Z is available before the end of slot s′′ and yet is not accepted by A until
s∗ − 1. See Figure 2(a). Therefore, w(A(s)) ≥ w(Z) for every s ∈ [s′′ + 1..s∗ − 1].
A
B
A
B
Z
Y*
s*−1
s*
(a)
Z
s''
Z
Y*
s*−1
s*
Yp
sp
(b)
Z
s''
Yq
sq
Figure 2: Positions of Yi+1 and Yp in A and B.
If s′′ > s∗ − 1, then let sp be the closest slot among s0, s1, . . . , si before s′′. See Figure 2(b).
Such slot must exist because s∗ is one such candidate. Then w(Yp) ≥ w(Z) or else Z would have
been accepted in slot sp. Therefore, w(A(s)) ≥ w(Yp) ≥ w(Z) for every odd slot s ∈ [s′′ + 1..sq − 1].
✷
Lemma 4.5 In Step i.3, B accepts Z in a slot s′′ where s′′ < min(s, s′).
Proof. First notice that w(B(s + 1)) < w(Y ) ≤ w(Yi+1). Therefore, s′′ < s or else Z would have
been a candidate for slot s + 1.
Now we assume that s′ < s and show that s′′ < s′. We distinguish two cases.
Case i. s′ is an odd slot. Let U = A(s′). Let sp be the slot in s0, . . . , si that is closest to and
before s′ (which must exist because s∗ itself is a candidate). Note that s∗ − 1 must be before sp
since it must be before s′ and immediately before one of the sj's (in this case s∗), and sp is the
latest such sj's before s′.
OPT
A
B
Z
Y*
s*−1
s*
Yp
sp
Z
U
s'
We have w(Z) > w(U ) since a self charge is made instead of a downward charge. By Lemma
4.4(ii), w(U ) ≥ w(Yp). Therefore w(Z) > w(Yp). Hence Z must be accepted before Yp in B or else
it can take Yp's place in B. By definition of sp, sp < s′. Hence s′′ < s′.
Case ii. s′ is an even slot. Let U = B(s′). Then w(Z) > w(U ) due to no downward charge in
slot s′. Thus B must have accepted Z before s′, i.e., s′′ < s′.
So in all cases B accepts Z at some slot s′′ < min(s, s′).
✷
Finally, the lemma below shows that two bad slots are paired with different good slots.
14
Lemma 4.6 All bad slots are paired with different good slots.
Proof. There are two possible places in our procedure where good slots can be identified: in Step
j.1 or in Step k.2 for some j and k. Call them substeps 1 and 2. Note that, for an odd bad slot,
good slots identified in substep 1 are always when B is accepting jobs, and good slot identified in
substep 2 are always when A is accepting jobs, and vice versa for even bad slots.
Consider two distinct bad slots s = (X, Y ) and s′ = (X ′, Y ′). First, consider the case when s
and s′ has different parity (odd or even slots). Then they can match with the same good slot only
if one of them identifies it in substep 1 and the other in substep 2. However, a good slot in substep
1 must receive self-charge (this is how the Yj's are identified) while a good slot in substep 2 cannot
receive a self charge (otherwise we would have moved on to some Step i.3 in the procedure). Thus
it is impossible that a substep 1 good slot is also a substep 2 good slot.
Next, consider the case when s and s′ are of the same parity. Without loss of generality assume
that they are both odd slots. To facilitate our discussion, we re-index the Yj's in the order they are
identified. So we let Y0, Y1, Y2, . . . be the chain of Yj's associated with s where {Y0, Y1} = {X, Y }
and for j ≥ 2, Yj is the job identified in step (j − 1).3. Similarly, we let Y ′
2 . . . be the chain
associated with s′. We will show that no job appears in both chains. This proves the claim because
for two bad slots of the same parity to be matched to the same good slot, they must both be
identified in substep 1 or both in substep 2. But if the chains of Yi's associated with them are
different, this is not possible.
1, Y ′
0, Y ′
0, Y ′
0 and Y ′
To show the chains are distinct, we first show that Y0, Y1, Y ′
1 are all distinct. Recall
1}. Clearly X 6= Y, X ′ 6= Y ′, X 6= X ′ and Y 6= Y ′.
that {X, Y } = {Y0, Y1} and {X ′, Y ′} = {Y ′
Thus we only need to show that X 6= Y ′ and X ′
If this is not true, then either: (1)
Y is accepted in A in a bad slot; or (2) X in OP T generates a backward charge. For (1), if
A−1(Y ) < A−1(X)(= s), then OP T (s + 1)(= Y ) would not make a backward charge to s; while if
A−1(Y ) > s, then A−1(Y ) cannot get a self charge and hence receives at most 1.5 units of charge.
For (2), OP T −1(X) > OP T −1(Y ) due to the self charge to slot s, and by Lemma 4.3(i), X must
also be accepted in B before s. Hence X cannot generate a backward charge. Thus neither (1) nor
(2) can be true.
6= Y .
j 6= Y ′
We have now established that Y0, Y1, Y ′
0 and Y ′
k for any j and k. Note that, if Yj = Y ′
1 are all different. It is also clear that Yj 6= Yk
and Y ′
k, for some j > 1 and k > 1, then there must be
some j′ < j and k′ < k such that Yj′ = Y ′
k′, because they are uniquely defined in such a way (in
some substep 2). So the only remaining case to consider is Y0 or Y1 being the same as Y ′
j for some
j > 1. Recall Y0 and Y1 are the X and Y of s. Y ′
j cannot be Y because by Lemma 4.5, both A and
B accept Y ′
j cannot be X because this would
mean the job B(s + 1) is Y ′
k for some k < j, so slot s + 1 should receive a downward charge but
this contradicts that OP T (s + 1) makes a backward charge to s instead of a downward charge. ✷
j before OP T does but A accepts Y after OP T . Y ′
5 Lower Bound for Equal Length Intervals
In this section, we show a lower bound of 2 for barely random algorithms for scheduling equal length
intervals that choose between two deterministic algorithms, possibly with unequal probability.
Theorem 5.1 No barely random algorithm choosing between two deterministic algorithms for the
online scheduling of equal length intervals has a competitive ratio better than 2.
15
w
w
ε'ε'
v+ 3
v+ 2
ε'
v+
v
v
Figure 3: SET (v, w, ǫ). On the left is the actual set of intervals; the vertical arrow on the right is
the notation we use to denote such a set.
Let ALG be a barely random algorithm that chooses between two deterministic algorithms A
and B with probability p and q respectively such that p + q = 1 and 0 < p ≤ q < 1. Let δ be an
arbitrarily small positive constant less than 1. We will show that there is an input on which OP T
gains at least 2 − δ times of what ALG gains.
We will be using sets of intervals similar to that in Woeginger [19]. More formally, let ǫ be an
arbitrary positive real number and let v, w be any pair of real numbers such that 0 ≤ v ≤ w. We
define SET (v, w, ǫ) as a set of intervals of weight v, v + ǫ′, v + 2ǫ′, . . . , w (where ǫ′ is the largest
number such that ǫ′ ≤ ǫ and w − v is a multiple of ǫ′) and their relative arrival times are such that
intervals of smaller weight come earlier and the last interval (i.e., the one that arrives last and has
weight w) arrives before the first interval finishes. Thus, there is overlapping between any pair of
intervals in the set. See Figure 3. This presents a difficulty for the online algorithm as it has to
choose the right interval to process without knowledge of the future.
To facilitate our discussion, we assume that all intervals have weight at least 1 throughout this
section, except Section 5.3.
If I is an interval in SET (v, w, ǫ) and w(I) > v (i.e., I is not the
earliest interval in the set), then I − denotes the interval that arrives just before I in SET (v, w, ǫ).
So, w(I −) ≥ w(I) − ǫ.
5.1 A Few Simple Cases
We first present a few simple situations in which OP T can gain a lot compared with what ALG
can gain. The first lemma shows that an algorithm should not start an interval that is lighter than
the current interval being processed by the other algorithm. The second lemma shows that it is not
good to have A processing an interval of equal or heavier weight than the interval currently being
processed by B. Moreover, the two algorithms should avoid processing almost non-overlapping
intervals as shown in the third lemma.
Lemma 5.1 Suppose at some moment, one of the algorithms (say B) is processing an interval J
from a set S = SET (v, w, ǫ) while the other algorithm (A) is processing another interval I, where
w(J) > v, w(I) ≤ w(J) and r(I) > r(J). (Note that here the interval I cannot come from S.)
Then OP T gains at least 2 − ǫ of ALG's gain on an input consisting of S, I and some subsequently
arrived intervals.
Proof. We illustrate the scenario in Figure 4. (There, the vertical line represents the set S and the
horizontal line labelled J is one of the intervals in S. The horizontal line labelled I arrives later
than J and has a smaller weight.) To defeat ALG, an interval J ′ with the same weight as J is
released between d(J −) and d(J) and no more intervals are released. (See Figure 4.) Then OP T
16
w
J
v
J'
I
Figure 4: (Lemma 5.1) Starting a lighter interval later
J'
w'
J
v'
w
I
v
Figure 5: (Lemma 5.2) A processing a heavier interval (J)
completes J − (which finishes just before J ′ starts) and J ′, gaining w(J −) + w(J ′) ≥ 2w(J) − ǫ
while ALG gains at most (p + q)w(J) = w(J) even if it aborts I to start J ′. Since w(J) ≥ 1,
2w(J)−ǫ
✷
w(J) ≥ 2 − ǫ.
Lemma 5.2 Suppose at some moment, algorithm B is processing an interval I from a set S =
SET (v, w, ǫ) while A is processing an interval J from a set S′ = SET (v′, w′, ǫ′), where w(I) > v,
w(J) > v′, w(I) ≤ w(J) and r(I) ≤ r(J). (Note that S and S′ can be the same set; I and J can
even be the same interval.) Then OP T gains at least 2 − max{ǫ, ǫ′} of ALG's gain on an input
consisting of S, S′ and some subsequently arrived intervals.
Proof. An interval J ′ with the same weight as J is released between d(I −) and d(I). See Figure
5. Clearly there is no point in algorithm A aborting J to start J ′. If algorithm B continues with
I, then no more intervals arrived. OP T gains w(I −) + w(J ′) ≥ w(I) + w(J) − ǫ while ALG gains
qw(I) + pw(J) ≤ (w(I) + w(J))/2. Since (w(I) + w(J))/2 ≥ 1, we have w(I)+w(J)−ǫ
(w(I)+w(J))/2 ≥ 2 − ǫ.
If B aborts I and starts J ′, then using Lemma 5.1, we can see that OP T gains at least 2 − ǫ′ of
what ALG gains on an input consisting of S, S′ and the subsequently arrived intervals specified in
Lemma 5.1.
✷
Lemma 5.3 Suppose at some moment, algorithm A is processing an interval I from a set S =
SET (v, w, ǫ) while algorithm B is processing an interval J from another set S′ = SET (v′, w′, ǫ′),
where w(I) > v, w(J) > v′, w(I) ≤ w(J) and the intervals of S′ arrive between d(I −) and d(I).
Then OP T gains at least 2 − (ǫ + ǫ′) of ALG's gain on an input consisting of S, S′ and some
subsequently arrived intervals. in the worst case.
Proof. An interval J ′ with the same weight as J is released between d(J −) and d(J) and no
more intervals are released. See Figure 6. OPT completes I − in S, J − in S′ and J ′. So it gains
w(I −) + w(J −) + w(J ′) ≥ w(I) + 2w(J) − (ǫ + ǫ′). On the other hand, A completes I and then
J ′ while B completes J. Thus ALG gains at most p(w(I) + w(J)) + qw(J) = pw(I) + w(J)
≤ (1/2)w(I) + w(J). Since w(I)/2 + w(J) ≥ 1, we have w(I)+2w(J)−(ǫ+ǫ′)
✷
≥ 2 − (ǫ + ǫ′).
w(I)/2+w(J)
17
w
I
v
w'
J
v'
J'
Figure 6: (Lemma 5.3) A and B processing almost non-overlapping intervals
5.2 Constructing the Sequence of Intervals
Our lower bound proof takes a number of steps. In each step, the adversary will release some set of
intervals SET (·, ·, ·) adaptively according to how ALG reacts in the previous steps. In each step,
the adversary forces ALG not to finish any interval (and hence gain no value) while OP T will gain
some. Eventually, OP T will accumulate at least 2 − δ times of what ALG can gain no matter what
ALG does in the last step.
5.2.1 Step 1
Let c = 2 − δ/2. The adversary releases S1 = SET (v1, w1, ǫ1) where v1 is some positive real number
at least one, w1 = c(q/p)(4/δ)v1 and ǫ1 = δ/8. Denote by I1 and J1, where w(I1) ≤ w(J1), the
intervals chosen by ALG. We claim that
Lemma 5.4 Both algorithms A and B do not process the smallest-weight interval in S1, i.e.,
(i) w(I1) > v1 and
(ii) w(J1) > v1.
Hence both I −
1 and J −
1 exist.
Proof. We first prove part (ii). By Lemma 5.2, we assume that I1 is processed by A and J1 is
processed by B. So, the expected gain by ALG is pw(I1) + qw(J1) ≤ w(J1). Then we deduce that
w(J1) > w1/c or else the adversary stops, OP T schedules the heaviest interval in S1 (of weight w1)
so that it gains at least c > 2 − δ times the expected gain by ALG. Since v1 = (p/q)(δ/4)(w1/c)
< w1/c, we have w(J1) > v1.
1) ≥ 2w(J1)−ǫ1 by executing J −
1 in S1 and then J ′
1 )+w(J ′
To prove part (i), we assume to the contrary that w(I1) = v1. Then an interval J ′
1 with
the same weight as J1 is released between d(J −
1 ) and d(J1). See Figure 7(a). OP T can gain
w(J −
1. Upon finishing I1, algorithm A can
go on to finish J ′
1))+qw(J1) = pw(I1)+w(J1).
Note that v1 = (p/q)(δ/4)(w1/c) ≤ (δ/4)w(J1) and pw(I1) = pv1 < (δ/4)w(J1). Thus ALG's gain
is at most pw(I1) + w(J1) ≤ (1 + δ/4)w(J1). So OP T 's gain is more than 2 − δ times that of ALG's.
✷
1. The expected gain of ALG is at most p(w(I1)+w(J ′
Lemma 5.5 w(J1) < 2w(I1).
We defer this to Section 5.3, where we prove that if w(J1) ≥ 2w(I1) then the adversary can
force the competitive ratio to be at least 2 − δ.
18
w1
J1
I1
v1
J'
(a)
w2
v2
(b)
w1
J1
I1
v1
Figure 7: Step 1. (left) Lemma 5.4, (right) Lemma 5.6
The adversary then releases a new set of intervals S2 = SET (v2, w2, ǫ2) such that all these
1 ) and d(I1), where v2 = w(I1), w2 = max{c(pw(I1) + qw(J1)) −
intervals arrive between d(I −
w(I1), v2} and ǫ2 = ǫ1/2. See Figure 7(b).
Lemma 5.6 Upon the release of S2, both A and B must abort their current intervals in S1 and
start some intervals I2 and J2 respectively in S2. Moreover, v2 < w(I2) < w(J2).
Proof.
is pw(I1) + qw(J1) while OP T can complete I −
≥ c(pw(I1) + qw(J1)) − ǫ1 = (2 − δ
If ALG ignores S2 and continues with both I1 and J1, then the expected gain of ALG
1 ) + w2
1 and the last interval in S2, gaining w(I −
Suppose algorithm B aborts J1 and starts some J2 in S2 while algorithm A continues to process
I1. Then by Lemma 5.3, ALG loses. Suppose algorithm B continues with J1 but A aborts I1 to
start some I2 in S2. By Lemma 5.2, it must be the case that w(I2) < w(J1). But then by Lemma
5.1, ALG loses too.
2 )(pw(I1) + qw(J1)) − δ
8 ≥ (2 − δ)(pw(I1) + qw(J1)).
Based on the above discussion, the only remaining sensible response for ALG is to abort both I1
and J1 and start some I2 and J2 in S2. By Lemma 5.2, we can further assume that w(I2) < w(J2).
Moreover, we claim that w(I2) > v2. Otherwise, ALG is effectively aborting only J1 but not I1.
Then the construction of inputs given in Lemma 5.3 can be used to defeat ALG.
✷
This finishes our discussion on Step 1 and we now proceed to Step 2.
5.2.2 Step i
1 )+w(I −
2 )+· · · w(I −
In general, at the beginning of Step i ≥ 2, we have the following situation: OP T has gained
w(I −
i−1) while ALG has not gained anything yet. Moreover, A and B of ALG are
respectively executing Ii and Ji in Si = SET (vi, wi, ǫi) where vi = w(Ii−1), wi = max{c(pw(Ii−1) +
qw(Ji−1)) − Pi−1
j=1 w(Ij), vi}, ǫi = ǫ1/2i−1 and vi < w(Ii) < w(Ji). We go through a similar analysis
to that in Step 1.
First, as in Lemma 5.5, we have w(Ji) < 2w(Ii) (the case w(Ji) ≥ 2w(Ii) is handled in the next
subsection). Next, the adversary releases Si+1 = SET (vi+1, wi+1, ǫi+1) in the period between d(I −
i )
and d(Ii) where vi+1 = w(Ii), wi+1 = max{c(pw(Ii) + qw(Ji)) −Pi
j=1 w(Ij), vi+1} and ǫi+1 = ǫ1/2i.
See Figure 8.
Similar to Lemma 5.6, we can prove that
Lemma 5.7 Upon the release of Si+1, both A and B must abort their current intervals in Si and
start some intervals Ii+1 and Ji+1 respectively in Si+1. Moreover, vi+1 < w(Ii+1) < w(Ji+1).
19
w i +1
w
i
J
i
I
i
I
i
−1
v
i
v
i
+1
Figure 8: Step i
...
I
1
Proof. ALG cannot continue with both Ii and Ji. Otherwise, OP T schedules (after finishing
I −
1 , . . . , I −
i and then the last interval of Si+1, thus gaining at least
i−1) I −
i
X
j=1
w(I −
j ) + c(pw(Ii) + qw(Ji)) −
i
X
j=1
w(Ij )
≥ c(pw(Ii) + qw(Ji)) −
i
X
j=1
ǫj
> (2 −
δ
2
)(pw(Ii) + qw(Ji)) − 2ǫ1
≥ (2 − δ)(pw(Ii) + qw(Ji)).
Suppose B aborts Ji in order to start some Ji+1 in Si+1 while A continues with Ii. Then by
Lemma 5.3, ALG loses. Suppose B continues with Ji while A aborts Ii to start Ii+1. By Lemma 5.2,
we have that w(Ii+1) < w(Ji). Then by Lemma 5.1, ALG loses too.
Based on the above reasoning, we conclude that ALG has to abort both Ii and Ji and start
some Ii+1 and Ji+1 in Si+1. By Lemma 5.2, we can assume that w(Ii+1) < w(Ji+1). We can also
argue that w(Ii+1) > vi+1 in the same way as proving w(I2) > v2 in Step 1.
✷
We now proceed to Step i + 1. Note that OP T has already gained w(I −
1 ) + · · · + w(I −
i ) while
ALG still has not gained anything. We will make use of Lemma 4.3 of Woeginger [19]:
Lemma 5.8 (Woeginger [19]). For 2 < d < 4, any strictly increasing sequence of positive numbers
ha1, a2, . . .i fulfilling the inequality
ai+1 ≤ dai −
i
X
j=1
aj
for every i ≥ 1 must be finite.
j=1 w(Ij ) or vi+1. If c(pw(Ii) + qw(Ji)) − Pi
Consider the sequence hw(I1), w(I2), . . .i. It is strictly increasing since w(Ii+1) > vi+1 = w(Ii)
for all i. Moreover, recall that wi+1 is set to be the maximum of either c(pw(Ii) + qw(Ji)) −
Pi
j=1 w(Ij) > vi+1 for all i ≥ 1, then we have w(Ii+1) ≤
wi+1 = max{c(pw(Ii) + qw(Ji)) − Pi
j=1 w(Ij ) ≤ cw(Ji) −
Pi
j=1 w(Ij) (since w(Ji) < 2w(Ii)) for all i. The existence of such an
infinite sequence contradicts Lemma 5.8. So, eventually, there is a finite k such that c(pw(Ik) +
j=1 w(Ij ), vi+1} = c(pw(Ii) + qw(Ji)) − Pi
j=1 w(Ij ) < (4 − δ)w(Ii) − Pi
20
qw(Jk)) − Pk
consists of a single interval of weight wk+1 (= w(Ik)).
j=1 w(Ij) ≤ vk+1 and hence we set wk+1 = vk+1 and the next (and final) set Sk+1
In that situation, it makes no difference whether A or B of ALG aborts Ik or Jk to start the
interval in Sk+1 since it has weight equal to w(Ik). Its expected gain is still at most pw(Ik)+qw(Jk).
On the other hand, OP T schedules the interval of Sk+1 and gains in total w(I −
k )+wk+1
≥ c(pw(Ik) + qw(Jk)) − 2ǫ1 which is at least 2 − δ times of ALG's gain.
1 )+· · ·+w(I −
5.3 The Case of w(Ji) ≥ 2w(Ii)
We now consider the case where in some Step i ≥ 1, w(Ji) ≥ 2w(Ii). We will show how the
adversary forces ALG to lose the game, i.e., OP T will gain at least 2 − δ of what ALG can on Si
and a set of subsequently arrived intervals. For simplicity, we drop the subscript i in Ii, Ji, and ǫi
in the following discussion.
Intuitively, when w(J) is relatively large compared with w(I), we can afford to let algo-
rithm A finish the interval I and gain pw(I), which is relatively small. Therefore, a set Su =
SET (0, uw(J), ǫ) is released between d(J −) and d(J), where u ≥ 1 is some parameter to be de-
termined as a function of p. This allows algorithm A to finish I and then start some job in this
new set Su. On the other hand, algorithm B has to decide whether to abort or continue with the
current interval J. We will show that there is a choice of u such that no matter what B does, OP T
can gain at least 2 − δ of what ALG gains.
Case 1: Algorithm B continues with J. Then ALG gains at most p(w(I) + uw(J)) + qw(J)
while OP T can gain w(J −) + uw(J). Therefore, the ratio of the gain by OP T to that of ALG on
S and Su is at least
w(J −) + uw(J)
p(w(I) + uw(J)) + qw(J)
≥
≥
(1 + u)w(J) − ǫ
p(w(J)/2 + uw(J)) + (1 − p)w(J)
1 + u − ǫ/w(J)
p(1/2 + u) + 1 − p
where the first inequality makes use of the condition that w(I) ≤ w(J)/2. This ratio is at least
2 − δ provided
1 + u −
ǫ
w(J)
≥ (2 − δ)(1 −
+ pu)
p
2
pδ
2
= (2 − δ − p +
) + (2p − pδ)u.
After simplifying using ǫ/w(J) ≤ ǫ < δ, this condition is satisfied by having
u ≥
1 − p + pδ/2
1 − 2p + pδ
.
(2)
Case 2: Algorithm B aborts J. Then algorithms A and B can start some intervals I ′ and
J ′ in Su = SET (0, uw(J), ǫ). By Lemma 5.2, we can assume that w(I ′) ≤ w(J ′). Then another
interval J ′′ with the same weight as J ′ is released between d((I ′)−) and d(I ′).
21
If ALG aborts I ′ to start J ′′, by Lemma 5.1, OP T gains at least 2 − ǫ of ALG's gain on Su and
a set of subsequently arrived intervals. Also, on the set of intervals S = SET (v, w, ǫ), ALG gains
pw(I) ≤ w(I)/2 while OP T gains at least w(J) − ǫ ≥ w(I) − ǫ. So ALG loses.
On the other hand, if ALG does not abort I ′, then its expected gain is at most p(w(I)+w(I ′))+
qw(J ′) ≤ p(w(J)/2 + w(I ′)) + (1 − p)w(J ′). OP T will complete J − in S, (I ′)− in Su and then J ′′.
(Note that if I ′ is the first interval with weight 0 in Su, then we set (I ′)− to be of weight 0 too.)
The ratio of the gain by OP T to that of ALG on S, Su and J ′′ is at least
(w(J) − ǫ) + (w(I ′) − ǫ) + w(J ′′)
p(w(J)/2 + w(I ′)) + (1 − p)w(J ′)
=
w(I ′) + w(J) + w(J ′) − 2ǫ
pw(I ′) + pw(J)/2 + (1 − p)w(J ′)
.
Since 1/p ≥ 2 > 2 − δ, it suffices to show that
f =
w(J) + w(J ′) − 2ǫ
pw(J)/2 + (1 − p)w(J ′)
is at least 2 − δ. Furthermore, 1/(p/2) ≥ 2 and 1−2ǫ/w(J ′)
a function of w(J ′), is minimized when w(J ′) is maximized. That means,
< 1
1−p
1−p ≤ 2. Therefore, the fraction f , as
f ≥
w(J) + uw(J) − 2ǫ
pw(J)/2 + (1 − p)uw(J)
≥
1 + u − 2ǫ
p/2 − pu + u
.
2 − pu + u ≥ p
Observe that p
because ǫ ≤ δ/8. Hence to show that f ≥ 2 − δ, it suffices to show that
condition is satisfied provided
2 − p + 1 ≥ 1 − p
2 > 1
2 using u ≥ 1 and p ≤ 1/2. Thus
1+u
p/2−pu+u ≥ 2 − δ
2ǫ
p/2−pu+u ≤ δ
2
2 . This
1 + u ≥ (2 −
= (p −
)(
δ
2
pδ
4
p
2
− pu + u)
) + (2 − 2p +
pδ
2
−
δ
2
)u.
or
This is satisfied if
i.e.,
By setting
(1 − 2p +
pδ
2
−
δ
2
)u ≤ 1 − p +
pδ
4
.
(1 − 2p +
pδ
2
)u ≤ 1 − p +
pδ
4
,
u ≤
1 − p + pδ/4
1 − 2p + pδ/2
.
u =
1 − p + pδ/4
1 − 2p + pδ/2
,
(3)
both ineq. (1) and ineq. (2) are satsified. This completes the proof for the case of w(Ji) ≥ 2w(Ii).
22
6 Conclusion
In this paper, we designed 2-competitive barely random algorithms for various versions of preemp-
tive scheduling of intervals. They are surprisingly simple and yet improved upon previous best
results. Based on the same approach, we designed a 3-competitive algorithm for the preemptive
(with restart) scheduling of equal length jobs. This is the first randomized algorithm for this prob-
lem. Finally, we gave a 2 lower bound for barely random algorithms that choose between two
deterministic algorithms, possibly with unequal probability.
An obvious open problem is to close the gap between the upper and lower bounds for randomized
preemptive scheduling of intervals in the various cases. We conjecture that the true competitive
ratio is 2. Also, it is interesting to prove a randomized lower bound for the related problem of job
scheduling with restart.
References
[1] S. Albers. On randomized online scheduling. In Proc. 34th ACM Symposium on Theory of
Computing, 134 -- 143, 2002.
[2] B. Awerbuch, Y. Bartal, A. Fiat, and A. Rosen. Competitive non-preemptive call control. In
Proc. 5th ACM-SIAM Symposium on Discrete Algorithms, 312 -- 320, 1994.
[3] S. Baruah, G. Koren, D. Mao, B. Mishra, A. Raghunathan, L. Rosier, D. Shasha, and F. Wang.
On the competitiveness of on-line real-time task scheduling. Real-Time Systems, 4:125 -- 144,
1992.
[4] K. P. Bogart and D. B. West. A short proof that 'Proper = Unit'. Discrete Mathematics,
201:21 -- 23, 1999.
[5] A. Borodin and R. El-Yaniv. Online Computation and Competitive Analysis. Cambridge
University Press, New York, 1998.
[6] R. Canetti and S. Irani. Bounding the power of preemption in randomized scheduling. SIAM
Journal on Computing, 27(4):993 -- 1015, 1998.
[7] F. Y. L. Chin, M. Chrobak, S. P. Y. Fung, W. Jawor, J. Sgall and T. Tich´y. Online competitive
algorithms for maximizing weighted throughput of unit jobs. Journal of Discrete Algorithms,
4(2):255 -- 276, 2006.
[8] F. Y. L. Chin and S. P. Y. Fung. Online scheduling with partial job values: does timesharing
or randomization help? Algorithmica, 37(3):149 -- 164, 2003.
[9] M. Chrobak, W. Jawor, J. Sgall and T. Tich´y. Online scheduling of equal-length jobs: ran-
domization and restarts help. SIAM Journal on Computing 36(6):1709 -- 1728, 2007.
[10] M. Englert and M. Westermann. Considering suppressed packets improves buffer management
in QoS switches. In Proc. 18th ACM-SIAM Symposium on Discrete Algorithms, 209 -- 218, 2007.
[11] L. Epstein and A. Levin.
Improved randomized results for the interval selection problem.
Theoretical Computer Science, 411(34-36):3129 -- 3135, 2010.
23
[12] S. P. Y. Fung, C. K. Poon and F. Zheng, Online interval scheduling: randomized and multi-
processor cases. In Proc. 13th International Computing and Combinatorics Conference, LNCS
4598, 176 -- 186, 2007.
[13] J.-H. Kim and K.-Y. Chwa. Scheduling broadcasts with deadlines. Theoretical Computer
Science, 325(3):479 -- 488, 2004.
[14] G. Koren and D. Shasha. Dover: An optimal on-line scheduling algorithm for overloaded
uniprocessor real-time systems. SIAM Journal on Computing, 24:318 -- 339, 1995.
[15] F. Li, J. Sethuraman and C. Stein. An optimal online algorithm for packet scheduling with
agreeable deadlines. In Proceedings of 16th ACM-SIAM Symposium on Discrete Algorithms,
pages 801 -- 802, 2005.
[16] H. Miyazawa and T. Erlebach. An improved randomized on-line algorithm for a weighted
interval selection problem. Journal of Scheduling, 7(4):293 -- 311, 2004.
[17] S. S. Seiden. Randomized online interval scheduling. Operations Research Letters, 22(4 -- 5):171 --
177, 1998.
[18] H.-F. Ting. A near optimal scheduler for on-demand data broadcasts. In Proc. 6th Italian
Conference on Algorithms and Complexity, LNCS 3998, 163 -- 174, 2006.
[19] G. J. Woeginger. On-line scheduling of jobs with fixed start and end times. Theoretical
Computer Science, 130(1):5 -- 16, 1994.
[20] F. Zheng, S. P. Y. Fung, W.-T. Chan, F. Y. L. Chin, C. K. Poon, and P. W. H. Wong.
Improved on-line broadcast scheduling with deadlines. In Proc. 12th International Computing
and Combinatorics Conference, LNCS 4112, 320 -- 329, 2006.
24
|
1811.00950 | 1 | 1811 | 2018-11-02T15:59:32 | Near-Linear Time Algorithm for n-fold ILPs via Color Coding | [
"cs.DS"
] | We study an important case of ILPs $\max\{c^Tx \ \vert\ \mathcal Ax = b, l \leq x \leq u,\, x \in \mathbb{Z}^{n t} \} $ with $n\cdot t$ variables and lower and upper bounds $\ell, u\in\mathbb Z^{nt}$. In $n$-fold ILPs non-zero entries only appear in the first $r$ rows of the matrix $\mathcal A$ and in small blocks of size $s\times t$ along the diagonal underneath. Despite this restriction many optimization problems can be expressed in this form. It is known that $n$-fold ILPs can be solved in FPT time regarding the parameters $s, r,$ and $\Delta$, where $\Delta$ is the greatest absolute value of an entry in $\mathcal A$. The state-of-the-art technique is a local search algorithm that subsequently moves in an improving direction. Both, the number of iterations and the search for such an improving direction take time $\Omega(n)$, leading to a quadratic running time in $n$. We introduce a technique based on Color Coding, which allows us to compute these improving directions in logarithmic time after a single initialization step. This leads to the first algorithm for $n$-fold ILPs with a running time that is near-linear in the number $nt$ of variables, namely $(rs\Delta)^{O(r^2s + s^2)} L^2 \cdot nt \log^{O(1)}(nt)$, where $L$ is the encoding length of the largest integer in the input. In contrast to the algorithms in recent literature, we do not need to solve the LP relaxation in order to handle unbounded variables. Instead, we give a structural lemma to introduce appropriate bounds. If, on the other hand, we are given such an LP solution, the running time can be decreased by a factor of $L$. | cs.DS | cs |
Near-Linear Time Algorithm for n-fold ILPs via
Color Coding∗
Klaus Jansen
Department of Computer Science, Kiel University, Kiel, Germany
[email protected]
Alexandra Lassota
Department of Computer Science, Kiel University, Kiel, Germany
[email protected]
Lars Rohwedder
Department of Computer Science, Kiel University, Kiel, Germany
[email protected]
Abstract
We study an important case of ILPs max{cT x Ax = b, l ≤ x ≤ u, x ∈ Znt} with n · t variables
and lower and upper bounds ℓ, u ∈ Znt.
In n-fold ILPs non-zero entries only appear in the
first r rows of the matrix A and in small blocks of size s × t along the diagonal underneath.
Despite this restriction many optimization problems can be expressed in this form. It is known
that n-fold ILPs can be solved in FPT time regarding the parameters s, r, and ∆, where ∆ is
the greatest absolute value of an entry in A. The state-of-the-art technique is a local search
algorithm that subsequently moves in an improving direction. Both, the number of iterations
and the search for such an improving direction take time Ω(n), leading to a quadratic running
time in n. We introduce a technique based on Color Coding, which allows us to compute these
improving directions in logarithmic time after a single initialization step. This leads to the first
algorithm for n-fold ILPs with a running time that is near-linear in the number nt of variables,
namely (rs∆)O(r2s+s2)L2 · nt logO(1)(nt), where L is the encoding length of the largest integer
in the input.
In contrast to the algorithms in recent literature, we do not need to solve the
LP relaxation in order to handle unbounded variables. Instead, we give a structural lemma to
introduce appropriate bounds.
If, on the other hand, we are given such an LP solution, the
running time can be decreased by a factor of L.
2012 ACM Subject Classification Mathematics of computing → Integer programming
Keywords and phrases Near-Linear Time Algorithm, n-fold, Color Coding
1
Introduction
Solving integer linear programs of the form max {cT x Ax = b, x ∈ Z≥0} is one of the most
fundamental tasks in optimization. This problem is very general and broadly applicable,
but unfortunately also very hard. In this paper we consider n-fold ILPs, a class of integer
linear programs with a specific block structure. This is, when non-zero entries appear only
in the first r rows of A and in blocks of size s × t along the diagonal underneath. More
∗ This work was partially supported by DFG Project "Strukturaussagen und deren Anwendung in
Scheduling- und Packungsprobleme", JA 612/20-1
XX:2
Near-Linear Time Algorithm for n-fold ILPs via Color Coding
precisely, an n-fold matrix has the form
A =
A1 A2
B1
0
0 B2
...
...
0
0
. . . An
0
. . .
0
. . .
...
. . .
. . . Bn
,
where A1, . . . , An are r × t matrices and B1, . . . , Bn are s × t matrices. In n-fold ILPs we
also allow upper and lower bounds on the variables. Throughout the paper we subdivide a
solution x into bricks of length t and denote by x(i) the i-th one. The corresponding columns
in A will be called blocks.
Lately, n-fold ILPs received great attention [2, 7, 12, 14, 16] and were studied intensively
due to two reasons. Firstly, many optimization problems are expressible as n-fold ILPs [5,
10, 12, 14]. Secondly, n-fold ILPs indeed can be solved much more efficiently than arbitrary
ILPs [7, 10, 16]. The previously best algorithm has a running time of (rs∆)O(r2s+rs2)L ·
(nt)2 log2(n·t)+ LP and is due to Eisenbrand et al. [7]. Here LP is the running time required
for solving the corresponding LP relaxation. This augmentation algorithm is the last one
in a line of research, where local improvement/augmenting steps are used to converge to
an optimal solution. Clever insights about the structure of the improving directions allow
them to be computed fast. Nevertheless, the dependence on n in the algorithm above is
still high.
Indeed, in practice a quadratic running time is simply not suitable for large
data sets [3, 6, 13]. For example when analyzing big data, large real world graphs as in
telecommunication networks or DNA strings in biology, the duration of the computation
would go far beyond the scope of an acceptable running time [3, 6, 13]. For this reason
even problems which have an algorithm of quadratic running time are still studied from the
viewpoint of approximation algorithms with the objective to obtain results in subquadratic
time, even for the cost of a worse quality [3, 6, 13]. Hence, it is an intriguing question,
whether the quadratic dependency on the number nt of variables can be eliminated. In this
paper, we answer this question affirmatively. The technical novelty comes from a surprising
area: We use a combinatorial structure called splitter, which has been used to derandomize
Color Coding algorithms. It allows us to build a powerful data structure that is maintained
during the local search and from which we can derive an improving direction in logarithmic
time. Handling unbounded variables in an n-fold is a non-trivial issue in the previous
algorithms from literature. They had to solve the corresponding LP relaxation and use
proximity results. Unfortunately, it is not known whether linear programming can be solved
in near-linear time in the number of variables. Hence, it is an obstacle for obtaining a
near-linear running time. We manage to circumvent the necessity of solving the LP by
introducing artificial bounds as a function of the finite upper bounds and the right-hand
side of the n-fold.
Summary of Results
We present an algorithm, which solves n-fold ILPs in time (rs∆)O(r2s+s2)L · nt log4(nt) +
LP, where LP is the time to solve the LP relaxation of the n-fold. This is the first
algorithm with a near-linear dependence on the number of variables. The crucial step is
to speed up the computation of the improving directions.
We circumvent the need for solving the LP relaxation. This leads to a purely combina-
torial algorithm with running time (rs∆)O(r2s+s2)L2 · nt log6(nt).
K. Jansen, A. Lassota, L. Rohwedder
XX:3
In the running times above the dependence on the parameters, i.e., (rs∆)O(r2s+s2), im-
proves on the function (rs∆)O(r2s+rs2) in the previous best algorithms.
Outline of New Techniques
We will briefly elaborate the main technical novelty in this paper. Let x be some feasible,
It is clear that when y∗ is an optimal solution for
non-optimal solution for the n-fold.
max{cT y Ay = 0, ℓ − x ≤ y ≤ u − x, y ∈ Znt}, then x + y∗ is optimal for the initial n-fold.
In other words, y∗ is a particularly good improving step. A sensible approximation of y∗ is
to consider directions y of small size and multiplying them by some step length, i.e., find
some λ · y with kyk1 ≤ k for a value k depending only on ∆, r, and s. This implies that at
most k of the n blocks are used for y. If we randomly color the blocks into k2 colors, then
with high probability at most one block of every color is used. This reduces the problem to
choosing a solution of a single brick for every color and to aggregate them. We add data
structures for every color to implement this efficiently. There is of course a chance that the
colors do not split y perfectly. We handle this by using a deterministic structure of multiple
colorings (instead of one) such that it is guaranteed that at least one of them has the desired
property.
Related Work
The first XP-time algorithm for solving n-fold integer programs is due to De Loera et al. [5]
with a running time of ng(A)L. Here g(A) denotes a so-called Graver complexity of the
constraint matrix A and L is the encoding length of the largest number in the input. This
algorithm already uses the idea of iterative converging to the optimal solution by finding
improving directions. Nevertheless, the Graver complexity appears to be huge even for
small n-fold integer linear programs and thus this algorithm was of no practical use [10].
The exponent of this algorithm was then greatly improved by Hemmecke et al. in [10] to a
constant factor yielding the first cubic time algorithm for solving n-fold ILPs. More precisely,
the running time of their algorithm is ∆O(t(rs+st))L · (nt)3, i.e., FPT-time parameterized
over ∆, r, s, and t. Lately, two more breakthroughs were obtained. One of the results is
due to Koutecký et al. [16], who gave a strongly polynomial algorithm with running time
∆O(r2s+rs2)(nt)6 · log(nt) + LP . Here LP is the running time for solving the corresponding
LP relaxation, which is possible in strongly polynomial time, since the entries of the matrix
are bounded. Simultaneously, Eisenbrand et al. reduced in [7] the running time from a cubic
factor to a quadratic one by introducing new proximity and sensitivity results. This leads
to an algorithm with running time (∆rs)O(r2s+rs2)L · (nt)2 log2(nt) + LP . Note that both
results require only polynomial dependency on t.
As for applications, n-fold ILPs are broadly used to model various problems. We refer
to the works [5, 10, 11, 12, 15, 18] and the references therein for an overview.
Structure of the Document
In Section 2 we introduce the necessary preliminaries. Section 3 gives the algorithm for
efficiently computing the augmenting steps. This is then integrated into an algorithm for
n-fold ILPs in Section 4. At first we require finite variable bounds and then discuss how
to eliminate this requirement using the solution of the LP relaxation. Finally, in Section 5
we discuss how to handle infinite variable bounds without the LP relaxation and give new
structural results.
XX:4
Near-Linear Time Algorithm for n-fold ILPs via Color Coding
2
Preliminaries
In the following we introduce n-folds formally and state the main results regarding them.
Further we familiarize splitters, a technique known from Color Coding.
◮ Definition 1. Let n, r, s, t ∈ N. Furthermore let A1, . . . , An be r × t integer matrices and
B1, . . . , Bn be s × t integer matrices. Then an n-fold A is of following form:
A =
A1 A2
B1
0
0 B2
...
...
0
0
. . . An
0
. . .
. . .
0
...
. . .
. . . Bn
.
The matrix A is of dimension (r +n·s)×n·t. We will divide A into blocks of size (r +n·s)×t.
Similarly, the variables of a solution x are partitioned into bricks of length t. This means
each brick x(i) corresponds to the columns of one submatrix Ai and therefore also Bi. Given
c, ℓ, u ∈ Zn·t and b ∈ Zr+n·s, the corresponding n-fold Integer Linear Programming problem
is defined by:
max {cT x Ax = b, ℓ ≤ x ≤ u, x ∈ Zn·t}.
The main idea for the state-of-the-art algorithms relies on some insight about the Graver
basis of n-folds, which are special elements of the kern of A. More formally, we introduce
the following definitions:
◮ Definition 2. The kern of a matrix A is defined as the set of integral vectors x with
Ax = 0. We write kern(A) for them.
◮ Definition 3. A Graver basis element g is a minimal element of kern(A). An element is
minimal, if it is not the sum of two sign-compatible elements u, v ∈ kern(A).
Here, sign-compatible means that ui < 0 if and only if vi < 0 for every i.
◮ Theorem 4 ([4]). Let A ∈ Zn×m and let x ∈ kern(A). Then there exist 2n − 1 Graver
basis elements g1, . . . , g2n−1, which are sign-compatible with x such that
x = X2n−1
i=1
λigi
for some λ1, . . . , λ2n−1 ∈ Z≥0.
Many results for n-fold ILPs rely on the fact that the ℓ1-norm of Graver basis elements for
n-fold matrices are small. The best bound known for the ℓ1-norm is due to [7].
◮ Theorem 5 ([7]). The ℓ1-norm of the Graver basis elements of an n-fold matrix A is
bounded by O(rs∆)rs.
Next, we will introduce a technique called splitters (see e.g. [17]), which has its origins in
the FPT community and was used to derandomize the Color Coding technique [1]. So far
it has not been used with n-fold ILPs. We refer the reader to the outline of techniques in
the introduction for the idea on how we apply the splitters.
◮ Definition 6. An (n, k, ℓ) splitter is a family of hash functions F from {1, . . . , n} to
{1, . . . , ℓ} such that for every S ⊆ {1, . . . , n} with S = k, there exists a function f ∈ F
that splits S evenly, that is, for every j, j′ ≤ ℓ we have f −1(j) ∩ S and f −1(j′) ∩ S differ
by at most 1.
K. Jansen, A. Lassota, L. Rohwedder
XX:5
If ℓ ≥ k, the above means that there is some hash function that has no collisions when
restricted to S. Interestingly, there exist splitters of very small size.
◮ Theorem 7 ([1]). There exists an (n, k, k2) splitter of size kO(1) log(n) which is computable
in time kO(1) · n log(n).
We note that an alternative approach to the result above is to use FKS hashing. Although
it has an extra factor of log(n), it is particularly easy to implement.
◮ Theorem 8 (Corollary 2 and Lemma 2 in [9]). Define for every prim q < k2 log(n) and
prim p < q the hash function x 7→ 1 + (p · (x mod q) mod k2). This is an (n, k, k2) splitter
of size O(k4 log2(n)).
3
Efficient Computation of Improving Directions
The backbone of our algorithm is the efficient computation of augmenting steps. The impor-
tant aspect is the fact that we can update the augmenting steps very efficiently if the input
changes only slightly. In other words, whenever we change the current solution by applying
an augmenting step, we do not have to recompute the next augmenting step from scratch.
The augmenting steps depend on a partition of the bricks. In the following we define the
notion of a best step based on a fixed partition. Later, we will independently find steps for
a number of partitions and take the best among them.
◮ Definition 9. Let P be a partition of the n bricks into k2 disjoint sets P1, P2, . . . , Pk2 . Let
u ∈ Znt
≤0 be some upper and lower bounds on the variables (not necessarily the
same as in the n-fold). A (P, k)-best step is an optimal solution of
≥0 and ℓ ∈ Znt
max cT x
Ax = 0
xi ≤ k
Xi∈Bj
xi = 0
ℓ ≤ x ≤ u
x ∈ Znt
Bj ∈ Pj
∀j ∈ {1, . . . , k2}
∀j ∈ {1, . . . , k2}, Bj ′ ∈ Pj \ {Bj}, i ∈ Bj ′
∀j ∈ {1, . . . , k2}
This means a (P, k)-best step is an element of kern(A), which uses only one brick of every
Pj ∈ P . Within that brick the norm of the solution must be at most k.
◮ Theorem 10. Consider the problem of finding a (P, k)-best step in an n-fold where the
lower and upper bounds u, ℓ can change. This problem can be solved initially in time kO(r) ·
∆O(r2+s2) · nt and then in kO(r) · ∆O(r2+s2) · log(nt) update time whenever the bounds of a
single variable change.
Proof. Let P be a partition of the bricks from matrix A into k2 disjoint sets P1, P2, . . . , Pk2 .
Solving the (P, k)-best step problem requires that from each set Pj ∈ P we choose at most
one brick and set this brick's variables. All variables in other bricks of Pj must be 0.
Let x be a (P, k)-best step and let x(j) have the values of x in variables of Pj and 0 in
all other variables. Then by definition, kx(j)k1 ≤ k. This implies that the right-hand side
XX:6
Near-Linear Time Algorithm for n-fold ILPs via Color Coding
regarding x(j), that is to say, Ax(j), is also small. Since the absolute value of an entry in
A is at most ∆, we have that kAx(j)k∞ ≤ k∆. Let ai be the i-th row of A. If i > r, then
aix(j) = 0. This is because Ax = 0 and ai has all its support either completely inside Pj or
completely outside Pj. Meaning, the value of Ax(j) is one of the (2k∆ + 1)r many values
we get by enumerating all possibilities for the first r rows. Furthermore, since P has only
k2 sets, the partial sum A(x(1) + · · · + x(j)) is always one of (2k3∆ + 1)r = (k∆)O(r) many
candidates.
Hence to find a (P, k)-best step we can restrict our search to solutions whose partial sums
stay in this range. To do so, we set up a graph containing k2 + 2 layers L0, L1, . . . Lk2 , Lk2+1.
An example is given in figure 1. The first layer L0 will consist of just one node marking the
starting point with partial sum zero. Similarly, the last layer Lk2+1 will just contain the
target point also having partial sum zero, since a (P, k)-best step is an element of kern(A).
Each layer Lj with 1 ≤ j ≤ k2 will contain (2k3∆ + 1)r many nodes, each representing one
possible value of A(x(1) + · · · x(j)). Two points v, w from adjacent layers Lj−1, Lj will be
connected if the difference of the corresponding partial sums, namely w − v, can be obtained
by a solution y of variables from only one brick of Pj (with kyk1 ≤ k). The weight of the
edge will be the largest gain for the objective function cT y over all possible bricks. Hence, it
could be necessary to compute and compare up to n values for each Pj and each difference
in the partial sums to insert one edge into the graph. Finally, we just have to find the
longest path in this graph as it corresponds to a (P, k)-best step. The out-degree of each
node is bounded by (2k3∆ + 1)r since at most this many nodes are reachable in the next
layer. Therefore the overall number of edges is bounded by
(k2 + 2) · (2k3∆ + 1)r · (2k3∆ + 1)r = (k∆)O(r).
Using the Bellman-Ford algorithm we can solve the Longest Path problem for a graph with
N vertices and M edges in time N · M as the graph does not contain any circles. This gives
a running time of (k∆)O(r) · (k∆)O(r) = (k∆)O(r) for solving the problem. Constructing the
graph, however, requires solving a number of IPs of the form
max c′T x
Bj(cid:19) x = (cid:18)b′
(cid:18)Aj
0(cid:19)
kxk1 ≤ k
ℓ′ ≤ x ≤ u′
x ∈ Zt,
where b′ ∈ Zr is the corresponding right-hand side of the top rows and ℓ′, u′, c′ are the upper
and lower bounds, and the objective of the block. This is an IP with r + s constraints,
t variables, lower and upper bounds, and entries of the matrix bounded by ∆ in absolute
value. Using the algorithm by Eisenbrand and Weismantel [8], solving one of them requires
time
t · O(r + s + 1)r+s+4 · O(∆)(r+s+1)(r+s+4) · log2((r + s + 1)∆) = t · ∆O(r2+s2).
In fact, a little thought allows us to reduce the dependency on t to a logarithmic one: Since
the number of constraints in the ILP above is very small, there are only ∆O(r+s) many
different columns. Because of the cardinality constraint kxk1 ≤ k, we only have to consider
2k many variables of each type of column, namely:
K. Jansen, A. Lassota, L. Rohwedder
XX:7
Layer L0
Layer L1
Layer L2
Layer Lk2
Layer Lk2+1
...
...
. . .
. . .
. . .
...
Figure 1 This figure shows an example for a layered graph obtained while solving the (P, k)-best
step problem. There are k + 2 layers, visually separated by gray dashed lines. This includes one
source layer L0, one target layer Lk2+1 both with just a single node representing the zero sum.
Further there are k2 layers with (2k3∆ + 1)r nodes each, where in one layer the nodes stand for
all reachable partial sums. Two points v, w from adjacent layers Lj−1, Lj will be connected if the
difference of the corresponding partial sums, namely w − v, can be obtained by a solution y of
variables from only one brick of Pj (with kyk1 ≤ k). The weight of the edge will be the largest gain
for the objective function cT y over all possible bricks. For the sake of clarity both the values of the
nodes and the edges are not illustrated.
The k many with u′
the k many with ℓ′
i > 0 and maximal c′
i < 0 and minimal c′
i.
i and
If some solution uses a variable not in this set, then by pigeonhole principle there is a variable
with the same column values and a superior objective value and which can be increased/de-
creased. We can reduce the variable outside this set and increase the corresponding variable
inside this set until all variables outside the set are 0. We can use an appropriate data struc-
ture (e.g. AVL trees) to maintain a set of all variables with u′
i < 0) such that we
can find the k best among them in time O(k log(t)). Whenever the bounds of some variable
change, we might have to add or remove entries, which also takes only logarithmic time.
After initialization in time O(nt) (in total for all bricks) solving such an IP can therefore be
implemented in time
i > 0 (ℓ′
k log(t) + 2k∆O(r+s)∆O(r2+s2) ≤ k log(t)∆O(r2+s2).
The number of IPs to solve is at most n times the number of edges, since we have to compare
the values of up to n bricks. This gives a running time of
O(nt) + n · (k∆)O(r) · log(t) · ∆O(r2+s2) ≤ nt · kO(r) · ∆O(r2+s2)
for constructing the graph. To obtain the update time from the premise of the theorem,
it is perfectly fine to solve the Longest Path problem again, but we cannot construct the
graph from scratch. However, in order to construct the graph we still have to find the best
value over all bricks for each edge. Fortunately, if only a few bricks are updated (in their
lower and upper bounds) it is not necessary to recompute all values. Each edge corresponds
to a particular Pj ∈ P and a fixed right-hand side (a possible value of Ax(j)). We require
XX:8
Near-Linear Time Algorithm for n-fold ILPs via Color Coding
an appropriate data structure De for every edge e, which supports fast computation of the
operations FindMax, Insert, and Delete. Again, an AVL tree computes each of these
operations in time O(log(N )), where N is the number of elements. In De we store pairs
(v, i) where i is a brick in Pj and v is the maximum gain of brick i for the right-hand side of
e. The pairs are stored in lexicographical order. Since there are at most n bricks in Pj, the
data structure will have at most n elements. Initially, we can build De in time nt · ∆O(r2+s2)
(this is replicated for each edge). Now consider a change to the instance. Recall that we
are looking at changes that affect only a single brick, namely the upper and lower bounds
within that brick change. We are going to update the data structure De (for each edge) to
reflect the changes and we are going to recompute the edge value of each edge e using De.
Then we simply solve the Longest Path problem again. Let Pj ∈ P be the set that contains
the brick i that has changed in some variable. We only have to consider edges from Lj−1
to Lj, since none of the other edges are affected by the change. For a relevant edge e we
compute the previous value v and current value v′ that the brick i would produce (before
and after the bounds have changed). In De we have to remove (v, i) and insert (v′, i). Both
operations need only O(log(n)) time. Then the running time to update De for one edge is
k log(t) · ∆O(r2+s2) + O(log(n)) ≤ k log(nt) · ∆O(r2+s2).
In order to update the edge value of e using De, we simply have to find the maximum
element in De, which again takes time O(log(n)). To summarize, the total time to update
the (P, k)-best step after a change to a single brick consists of (1) updating each De, (2)
finding the maximum in each De, and (3) solving the Longest Path problem. We conclude
that the update time is
k log(nt) · ∆O(r2+s2) · (k∆)O(r) + log(n) · (k∆)O(r) + (k∆)O(r)
≤ kO(r)∆O(r2+s2) · log(nt). ◭
4
The Augmenting Step Algorithm
In this section we will assume that all lower and upper bounds are finite and give a complete
algorithm for this case. Later, we will explain how to cope with infinite bounds. We start
by showing how to converge to an optimal solution when an initial feasible solution is given.
To compute the initial solution, we also apply this algorithm on a slightly modified instance.
The approach resembles the procedure in previous literature, although we apply the results
from the previous section to speed up the computation of augmenting steps.
Let x be a feasible solution for the n-fold, in particular Ax = b. Let x∗ be an optimal
one. Theorem 4 states that we can decompose the difference vector x′ = x∗ − x into at most
2nt weighted Graver basis elements, that is
x′ = x∗ − x = X2nt−1
j=1
λjgj.
For intuition, consider the following simple approach (this is similar to the algorithm in [10]).
Suppose we are able to guess the best vector λigi = argmaxj{cT (λjgj)} regarding the gain
for the objective function. This pair of step length λi and Graver element gi is called the
Graver best step. Then we can augment the current solution x by adding λigi to it, i.e.,
we set x ← x + λigi. Feasibility follows because all gj are sign-compatible. This procedure
is repeated until no improving step is possible and therefore x must be optimal. In each
iteration this decreases the gap to the optimal solution by a factor of at least 1 − 1/(2nt) by
K. Jansen, A. Lassota, L. Rohwedder
XX:9
the pigeonhole principle. It may be costly to guess the precise Graver best step, but for our
purposes it will suffice to find an augmenting step that is approximately as good.
We will now describe how to guess λi. Since x + λigi is feasible, we have that λigi ≤
u − x ≤ u − ℓ and λigi ≥ ℓ − x ≥ ℓ − u. Let (gi)j ∈ supp(gi) be some non-zero variable.
If (gi)j > 0, then λi ≤ (λigi)j ≤ uj − ℓj. Otherwise, (gi)j < 0 and λi ≤ −(λigi)j ≤
−(ℓj − uj) = uj − ℓj. Hence, it suffices to check all values in the range {1, . . . , Γ}, where
Γ = maxj{uj − ℓj}. Proceeding like in [7], we lower the time a bit further by not taking
Instead, we look at guesses of the form λ′ = 2k for k ∈
every value into consideration.
{0, . . . , ⌊log(Γ)⌋}. Doing so we lose a factor of at most 2 regarding the improvement of the
objective function, since cT (λ′gi) > 0.5 · cT (λigi) when taking λ′ = 2⌊log(λi)⌋ > λi/2. Fix λ′
to the value above. Next we describe how to compute an augmenting step that is at least
as good as λ′gi. Note that gi is a solution of
Ay = 0
kyk1 ≤ k
⌈
ℓ − x
λ′ ⌉ ≤ y ≤ ⌊
u − x
λ′
⌋,
λ′ ⌉ and u = ⌊ u−x
where k = O(rs∆)rs is the bound on the norm of Graver elements from Theorem 5. Suppose
we have guessed some partition P = {P1, . . . , Pk2 } of the bricks such that of each Pj only
a single brick has non-zero variables in gi. Clearly, the augmenting step λ′y∗, where y∗ is
a (P, k)-best step with bounds ℓ = ⌈ ℓ−x
λ′ ⌋ would be at least as good as λ′gi.
Indeed Theorem 10 explains how to compute such a (P, k)-best step dynamically and when
we add λ′gj to x we only change the bounds of at most k3 many variables. Hence, it is
very efficient to recompute (P, k)-best steps until we have converged to the optimal solution.
However, valid choices of λ′ and P might be different in every iteration. Regarding λ′,
we simply compute (P, k)-best steps for every of the O(log(Γ)) many guesses and take the
best among them. We proceed similarly for P . We guess a small number of partitions and
guarantee that always at least one of them is valid. For this purpose we employ splitters.
More precisely, we compute a (n, k, k2) splitter of the n bricks. Since gj has a norm bounded
by k, it can also only use at most k bricks. Therefore the splitter always contains a partition
P = {P1, . . . , Pk2 } where gj only uses a single brick in every Pj.
To recap, in every iteration we solve a (P, k)-best step problem for every guess λ′ and
every partition P in the splitter and take the overall best solution as an improving direction
λ′y∗. Then we update our solution x by adding λ′y∗ onto it. At most k2 many bricks
change (and within each brick only k variables can change) and therefore we can efficiently
recompute the (P, k)-best steps for every guess for the next iteration. This way we guarantee
that we improve the solution by a factor of at least 1−1/(4nt) in every iteration. The explicit
running time of these steps will be analyzed in the next theorem.
Initial Solution
Recall that we still have to find an initial solution. This solution indeed can be computed
by using the augmenting step algorithm described above. We construct a new n-fold ILP
which has a trivial feasible solution and whose optimal solution corresponds to a feasible
solution of the original problem.
First we extend our n-fold A by adding (r + s)n new columns as follows: After the first
block (A1, B1, 0, . . . , 0)T add r + s columns. The first r ones will contain an r × r identity
matrix we call Ir. This matrix Ir has all ones in the diagonal. All other entries are zero.
The next s columns will contain an s × s identity matrix Is. This submatrix will start at
XX:10 Near-Linear Time Algorithm for n-fold ILPs via Color Coding
row r + 1. Again all other entries are zeros in these columns. After the next block we again
introduce r + s new columns, the first r ones containing just zeros, the next an Is matrix
at the height of B2. We repeat this procedure of adding r + s columns after each block, the
first r having solely zero entries and the next s containing Is at the height of Bi until our
resulting matrix Ainit for finding the initial solution looks like the following:
Ainit =
A1
B1
0
0
...
0
Ir
0
0
0
...
0
0 A2
Is
0
0 B2
0
...
0
0
...
0
0
0
0
0
...
0
0
0
Is
0
...
0
. . . An
0
. . .
. . .
. . .
0
0
...
Bn
0
0
0
0
...
0
0
0
0
0
...
Is
.
Due to our careful extension Ainit has again n-fold structure. For clarity the relevant sub-
matrices are framed in the matrix above. Remark that zero entries inside of a block do not
harm as solely the zeros outside of the blocks are necessary for an n-fold structure. At first
glance, it seems that for the right-hand side b we now have a trivial solution consisting only
of the new columns. Keep in mind, however, that the old variables have upper and lower
bounds and that 0 might be outside these bounds. In order to handle this case we subtract ℓ,
the lower bound, from all upper and lower bounds and set the right-hand side to b′ = b − Aℓ.
We get an equivalent n-fold where every solution is shifted by ℓ. Now we can find a feasible
solution (for b′) using solely the new variables by defining
y′ = (0, . . . 0, b′
1, b′
2, . . . b′
r+s, 0, . . . 0, b′
r+s+1, . . . b′
r+2s, 0, . . . 0, b′
r+ns−s, b′
r+ns)T
where each non-zero entry corresponds to the columns containing the submatrices Ir and
Is respectively with a multiplicity of the remaining right-hand side b′. Next we introduce
an objective function that penalizes using the new columns by having non-zero entries c′
i
corresponding to the positions of the new variables. We set
cinit = (0, . . . 0, c′
1, . . . c′
r+s, 0, . . . 0, c′
r+ns−s, . . . , c′
r+ns),
where the zero entries correspond to old variables. The values c′
bounds for the new variables depend on the sign of the right-hand side.
i and the lower and upper
i = −1, the lower bound to 0, and the upper bound to b′
i. This way
i ≥ 0, then set c′
If b′
the variable can only be non-negative.
If b′
must be non-positive.
i < 0, set ci = 1, the lower bound to b′
i and the upper bound to 0. Hence this variable
Clearly a solution has a value of 0, if and only if none of the new columns are used and
no solution of better value is possible. Hence, if we use our augmenting step algorithm and
solve this problem optimally, we either find a solution with value 0 or one with a negative
value. In the former, we indeed have not taken any of the new columns into our solution,
therefore we can delete the new columns and obtain a solution for the original problem (after
adding ℓ to it). Otherwise, there is no feasible solution for the original problem as we solved
the problem optimal regarding the objective function.
◮ Theorem 11. The dynamic augmenting step algorithm described above computes an opti-
mal solution for the n-fold Integer Linear Program problem in time (rs∆)O(r2s+s2) · O(L2 ·
K. Jansen, A. Lassota, L. Rohwedder
XX:11
nt log4(nt)) when finite variable bounds are given for each variable. Here L is the encoding
length of the largest occurring number in the input.
Proof. Due to Theorem 4 we know that the difference vector of an optimal solution x∗ to
our current solution x, i.e. x′ = x∗ − x, can be decomposed into 2nt weighted Graver basis
elements. Hence, if we adjust our solution x with the Graver best step, we reduce the gap
between the value of an optimal solution and our current solution by a factor of at least
1 − 1/(2nt) due to the pigeonhole principle. Our algorithm finds an augmenting step that
is at least half as good as the Graver best step. Therefore, the gap to the optimal solution
is still reduced by at least a factor of 1 − 1/(4nt).
Regarding the running time we first have to compute the splitter. Theorem 7 says, that
this can be done in time kO(1) · n log(n) = (rs∆)O(rs) · n log(n). Next we have to try all
values for the weight λ. Due to our step-length we get O(log(Γ)) guesses. Recall that Γ
denotes the largest difference between an upper bound and the corresponding lower bound,
i.e., Γ = maxj{uj − ℓj}. Fixing one, we have to find the best improving direction regarding
each of the ((rs∆)O(rs))O(1) log(n) = (rs∆)O(rs) log(n) partitions.
In the first iteration
we have to set up the tables in time kO(r) · ∆O(r2+s2) · nt = (rs∆)O(r2s) · ∆O(r2+s2) · nt
by computing the gain for each possible summand for each set and setting up the data
structure. In each following iteration we update each table and search for the optimum in
time kO(r) · ∆O(r2+s2) · log(nt) = (rs∆)O(r2s) · ∆O(r2+s2) · log(nt). Now it remains to bound
the number I of iterations needed to converge to an optimal solution. To obtain such a
bound we calculate:
1 > (1 − 1/(4nt))IcT (x∗ − x).
By reordering the term, we get
I <
− log(cT (x∗ − x))
log(1 − 1/(4nt))
.
As log(1 + x) = Θ(x), we can bound log(1 − 1/(4nt)) by Θ(−1/(4nt)) and thus
I < O(
− log(cT (x∗ − x))
−1/(4nt)
) ≤ O(4nt log(cT (x∗ − x))).
As the maximal difference between the current solution x and an optimal one x∗ can be
at most the maximal value of c times the largest number in between the bounds for each
variable, we get cT (x∗ − x) ≤ nt maxi ci · Γ and thus
I < O(4nt log(cT (x∗ − x))) ≤ O(nt log(nt max
i
ci · Γ)) ≤ O(nt log(ntΓ max
i
ci)).
Let L denote the encoding length of largest integer in the input. Clearly 2L bounds the
largest absolute value in c and thus we get
I < O(nt log(ntΓ max
i
ci)) = O(nt log(ntΓ2L)) = O(nt log(ntΓ2L)).
Hence after this amount of steps by always improving the gain by a factor of at least
1 − 1/(4nt) we close the gap between the initial solution and an optimal one. Given this, we
XX:12 Near-Linear Time Algorithm for n-fold ILPs via Color Coding
can now bound the overall running time with:
(rs∆)O(rs) · n log(n)
+ (rs∆)O(rs)
· (rs∆)O(r2s) · (rs∆)O(r2+s2) · nt
+
Splitter
Partitions
First Iteration
O(nt log(ntΓ2L))
· O(log(Γ))
· (rs∆)O(rs)
· (rs∆)O(r2s) · (rs∆)O(r2+s2) · log(nt))
{z
{z
I
}
}
{z
{z
}
}
{z
}
= O((nt log(ntΓ2L)) · O(log(Γ)) · (rs∆)O(r2s+s2) · log(nt)
= (rs∆)O(r2s+s2) · O(log2(Γ + 2L) · nt log2(nt)).
λ Guesses
Partitions
Update Time
{z
}
{z
}
Here Splitter denotes the time to compute the initial set P of partitions and Partitions
denotes the cardinality of P. First Iteration is the time to solve the first iteration of the
(P, k)-best step problem. Further λ Guesses is the number of guesses we have to do to get
the right weight and lastly Update Time is the time needed to solve each following (P, k)-best
step including updating the bounds and data structures.
Note, that we still have to argue about finding the initial solution, since in the con-
struction the parameters of the n-fold slightly change. The length of a brick changes to
t′ = t + r + s. This, however, can be hidden in the O-Notation of (rs∆)O(r2s+s2). Fur-
ther, Γ′, the biggest difference in upper and lower bounds can be bounded by a function in
Γ, ∆, L, t and n. Recall, that the difference between the bounds of old variables does not
change. For the new variables, however, the difference can be as large as kb′k∞. Thus we
bound this value by
kb′k∞ = kb − Aℓk∞ ≤ kbk∞ + kAℓk∞ ≤ kbk∞ + ∆ · kℓk1 ≤ O(∆ · nt · 2L).
We conclude that the running time for finding an initial solution (and also the overall running
time) is
(rs∆)O(r2s+s2)O(log2(Γ′ +2L)nt2 log2(nt)) = (rs∆)O(r2s+s2)O(log2(∆2Lnt)nt2 log2(nt))
= (rs∆)O(r2s+s2) · L2nt log4(nt). ◭
Handling Infinite Bounds
Remark, that if no finite bounds are given for all variables, we have to introduce some
artificial bounds first. Here we can proceed as in [7], where first the LP relaxation is solved
to obtain an optimal fractional solution z∗. Using the proximity results from [7], we know
that an optimal integral solution x∗ exists such that kx∗ − z∗k1 ≤ nt(rs∆)O(rs). This allows
us to introduce artificial upper bounds for the unbounded variables. Remark that this comes
at the price of solving the corresponding relaxation of the n-fold Integer Linear Program
problem. However we also lessen the dependency from L2 to L as the finite upper and lower
bounds can also be bounded more strictly due to the same proximity result. This yields an
overall running time of (rs∆)O(r2s+s2) · L · nt log4(nt) + LP. Nevertheless, solving this LP
can be very costly, indeed it is not clear if a potential algorithm even runs in time linear in
n. Thus, it may even dominate the running time of solving the n-fold ILP with finite upper
bounds. Fortunately we can circumvent the necessity of solving the LP as we will describe
in the following section using new structural results.
◮ Theorem 12. The dynamic augmenting step algorithm described above computes an op-
timal solution for the n-fold Integer Linear Program problem in time (rs∆)O(r2s+s2) · L ·
nt log4(nt) + LP when some variables have infinite upper bounds. Here LP is the running
time to solve the corresponding relaxation of the n-fold ILP problem.
K. Jansen, A. Lassota, L. Rohwedder
XX:13
5
Bounds on ℓ1-norm
In the following, we prove that even with infinite variable bounds in an n-fold there always
exists a solution of small norm (if the n-fold has a finite optimum). Therefore, we can apply
the algorithm for finite variable bounds by replacing every infinite one with this value.
◮ Lemma 13. If the n-fold is feasible and y is some (possibly infeasible) solution satisfying
the variable bounds, then there exists a feasible solution x with kxk1 ≤ O(rs∆)rs+1 · (kyk1 +
kbk1)
Proof. We take the same construction as in the algorithm for finding a feasible solution in
Section 4. Indeed, this construction was not setup for infinite bounds, but we consider the
straight-forward adaption where infinite bounds simply stay the same. The useful property
is that an optimal solution for this n-fold is a feasible solution for the original n-fold. Recall,
the construction has a right-hand side b′ with kb′k1 ≤ kAyk1 + kbk1, the value of t becomes
t′ = t + r + s, and the objective function c′ consists only of the values {−1, 0, 1}. Moreover,
there is a feasible solution y with kyk1 = kb′k1. Let x∗ be an optimal solution for this altered
n-fold that minimizes kx∗ − yk1. We consider the decomposition into Graver elements
λigi = x∗ − y. Then c′T gi > 0 or λi = 0 for all i, since otherwise x∗ − gi
would be a better solution than x∗. It follows that c′T gi ≥ 1 by discreteness of c. Also, by
Theorem 5, kgik1 ≤ O(rs∆)rs. Recall that by construction c′T x∗ = 0 and c′T y = −kb′k1,
P2n(t+r+s)−1
which implies Pi λi ≤ kb′k1. Therefore,
i=1
λikgik1
kx∗k1 ≤ kyk1 + kXi
λigik1 ≤ kb′k1 +Xi
≤ kb′k1 + kb′k1 · O(rs∆)rs ≤ (kbk1 + kyk1) · O(rs∆)rs+1.
Here we use that kb′k1 ≤ kbk1 + kAyk ≤ kbk1 + (r + s)∆ · kyk1.
◭
◮ Lemma 14. If the n-fold is bounded and feasible, then there exists an optimal solution x
with kxk1 ≤ (rs∆)O(rs) · (kbk1 + ntζ), where ζ denotes the largest absolute value among all
finite variable bounds.
Proof. Clearly there exists a (possibly infeasible) solution y satisfying the bounds with
kyk1 ≤ ntζ. By the previous lemma we know that there is a feasible solution y with
kyk1 ≤ (rs∆)O(rs) · (kbk1 + ntζ). Let x∗ be an optimal solution of minimal norm. W.l.o.g.
assume that x∗ − y has only non-negative entries. If there is a negative entry, consider the
equivalent n-fold problem with the corresponding column inverted and its bounds inverted
and swapped.
We know that there is a decomposition of x∗ − y into weighted Graver basis elements
P2nt−1
i=1 λigi = x∗ − y. Since every gi is sign-compatible with x∗ − y, we have that all
gi are non-negative as well. Furthermore, it holds that cT gi > 0 or λi = 0 for every gi,
since otherwise x∗ − gi would be a solution of smaller norm with an objective value that is
not worse. Now suppose toward contradiction that there is some gi where all variables in
supp(gi) have infinite upper bounds. Then the n-fold is clearly unbounded, since y + α · gi is
feasible for every α > 0 and in this way we can scale the objective value beyond any bound.
Thus, every Graver basis element adds at least the value 1 to some finitely bounded variable.
This implies that Pi λi ≤ kyk1 + ntζ: If not, then by pigeonhole principle there is some
finitely bounded variable x∗
j with
x∗
j = yj + (Xi
λigi)j > yj + ζ + yj ≥ ζ.
XX:14 Near-Linear Time Algorithm for n-fold ILPs via Color Coding
Since x∗ is feasible, this cannot be the case. We conclude,
kx∗k1 ≤ kyk1 +Xi
λikgik1 ≤ kyk1 + O(rs∆)rs ·Xi
λi
≤ O(rs∆)rs · (kyk1 + ntζ) ≤ (rs∆)O(rs) · (kbk1 + ntζ). ◭
This yields an alternative approach to solving the LP relaxation, because now we can simply
replace all infinite bounds with ±(rs∆)O(rs) · nt · 2L. Then we can apply the algorithm that
works only on finite variable bounds. The new encoding length L′ of the largest integer in
the input can be bounded by
L′ ≤ log((rs∆)O(rs) · 2L · nt) ≤ O(rs · log(rs∆) · L · log(nt)).
This way we obtain the following.
◮ Corollary 15. We can compute an optimal solution for an n-fold in time (rs∆)O(r2s+s2)L2·
nt log6(nt).
In a similar way, we can derive the following bound on the sensitivity of an n-fold ILP. This
bound is not needed in our algorithm, but may be of independent interest, since it implies
small sensitivity for problems that can be expressed as n-fold.
◮ Theorem 16. Let x be an optimal solution of an n-fold with right-hand side b, in particular,
Ax = b. If the right hand side changes to b′ and the n-fold still has a finite optimum, then
there exists an optimal solution x′ for b′ (Ax′ = b′) with kx − x′k1 ≤ O(rs∆)rs · kb − b′k1.
It is notable that this bound does not depend on n. This is in contrast to the known bounds
for the distance between LP and ILP solutions of an n-fold [7].
Proof. Consider the matrix Ainit from the construction used for finding an initial solution,
that is, identity matrices are added after every block. As opposed to the proof of Lemma 13,
we leave everything except for the matrix the same. In particular, we do not change the
value in the objective function c and new columns get a value of 0. As the right-hand side of
the n-fold we use b′. For some solution x, we write xold and xnew for the vector restricted to
the old variables (with all others 0) and the variables added in the matrix Ainit, respectively.
This means x = xold + xnew.
new. Let P2n(t+r+s)−1
i=1
Let x be an optimal solution with Ainit · xnew = b′ − b and x′ one with Ainit · x′
new = 0.
Here we assume that x′ is chosen so as to minimize kx − x′k1. Those solutions naturally
correspond to solutions of the original n-fold with right-hand side b = b′ − (Ainit · xnew) and
b′ = b′ − Ainit · x′
λigi = x′ − x be the decomposition into Graver
basis elements. Suppose toward contradiction there is some gi where all of supp(gi) are old
variables. If cT gi > 0, then x is not optimal, because x + gi is feasible and has a better
objective value. If on the other hand cT gi ≤ 0, then x′ − gi is a solution of at least the same
value as x′ and thus kx − x′k1 is not minimal. Indeed, this means k(gi)newk1 ≥ 1 for all gi.
In other words, each graver element contains a non-zero new variable. Recall that Ainit is
the identity matrix when restricted to the new variables (plus some zero columns). Due to
the sign-compatibility we get
Xi
λi ≤ k(Xi
λigi)newk1 = kAinit · (Xi
λigi)newk1 = kAinit · (x′ − x)newk1 = kb − b′k1.
We conclude,
kx − x′k1 = kXi
λigik1 ≤ O(rs∆)rs ·Xi
λi ≤ O(rs∆)rs · kb − b′k1.
◭
K. Jansen, A. Lassota, L. Rohwedder
XX:15
References
1 Noga Alon, Raphael Yuster, and Uri Zwick. Color-coding. Journal of the ACM, 42(4):844 --
856, 1995. Previously appeared in STOC 1994.
2 Katerina Altmanová, Dusan Knop, and Martin Koutecký. Evaluating and Tuning n-fold
Integer Programming. In 17th International Symposium on Experimental Algorithms, vol-
ume 103 of Leibniz International Proceedings in Informatics (LIPIcs), pages 10:1 -- 10:14,
Germany, 2018. Schloss Dagstuhl -- Leibniz-Zentrum fuer Informatik.
3 Alexandr Andoni and Krzysztof Onak. Approximating edit distance in near-linear time.
SIAM Journal on Computing, 41(6):1635 -- 1648, 2012. Previously appeared in STOC 2009.
An integer analogue of
4 William Cook, Jean Fonlupt, and Alexander Schrijver.
5
Carathéodory's theorem. Journal of Combinatorial Theory, Series B, 40(1):63 -- 70, 1986.
Jesús A De Loera, Raymond Hemmecke, Shmuel Onn, and Robert Weismantel. N-fold
integer programming. Discrete Optimization, 5(2):231 -- 241, 2008.
6 Doratha E Drake and Stefan Hougardy. A linear time approximation algorithm for weighted
matchings in graphs. Journal of the ACM Transactions on Algorithms, 1(1):107 -- 122, 2005.
7
8
Friedrich Eisenbrand, Christoph Hunkenschröder, and Kim-Manuel Klein. Faster Algo-
rithms for Integer Programs with Block Structure. In 45th International Colloquium on Au-
tomata, Languages, and Programming, volume 107 of Leibniz International Proceedings in
Informatics (LIPIcs), pages 49:1 -- 49:13, Germany, 2018. Schloss Dagstuhl -- Leibniz-Zentrum
fuer Informatik.
Friedrich Eisenbrand and Robert Weismantel. Proximity results and faster algorithms for
integer programming using the steinitz lemma. In Proceedings of the Twenty-Ninth Annual
ACM-SIAM Symposium on Discrete Algorithms, pages 808 -- 816. Society for Industrial and
Applied Mathematics, 2018.
9 Michael L. Fredman, János Komlós, and Endre Szemerédi. Storing a sparse table with O(1)
worst case access time. Journal of the ACM, 31(3):538 -- 544, 1984. Previously appeared in
FOCS 1982.
10 Raymond Hemmecke, Shmuel Onn, and Lyubov Romanchuk. N-fold integer programming
in cubic time. Mathematical Programming, pages 1 -- 17, 2013.
11 Raymond Hemmecke, Shmuel Onn, and Robert Weismantel. N -fold integer programming
and nonlinear multi-transshipment. Optimization Letters, 5(1):13 -- 25, 2011.
12 Klaus Jansen, Kim-Manuel Klein, Marten Maack, and Malin Rau. Empowering the
configuration-ip - new ptas results for scheduling with setups times. In 10th Innovations
in Theoretical Computer Science (ITCS), 2019. to appear.
13
Jonathan A Kelner, Yin Tat Lee, Lorenzo Orecchia, and Aaron Sidford. An almost-linear-
time algorithm for approximate max flow in undirected graphs, and its multicommodity
generalizations. In Proceedings of the twenty-fifth annual ACM-SIAM symposium on Dis-
crete algorithms, pages 217 -- 226. SIAM, 2014.
14 Dušan Knop and Martin Kouteck`y. Scheduling meets n-fold integer programming. Journal
of Scheduling, pages 1 -- 11, 2017.
15 Dusan Knop, Martin Koutecký, and Matthias Mnich. Combinatorial n-fold integer pro-
In 25th Annual European Symposium on Algorithms, ESA
gramming and applications.
2017, September 4-6, 2017, Vienna, Austria, pages 54:1 -- 54:14, 2017.
16 Martin Koutecký, Asaf Levin, and Shmuel Onn. A Parameterized Strongly Polynomial
Algorithm for Block Structured Integer Programs. In 45th International Colloquium on Au-
tomata, Languages, and Programming, volume 107 of Leibniz International Proceedings in
Informatics (LIPIcs), pages 85:1 -- 85:14, Germany, 2018. Schloss Dagstuhl -- Leibniz-Zentrum
fuer Informatik.
XX:16 Near-Linear Time Algorithm for n-fold ILPs via Color Coding
17 Moni Naor, Leonard J. Schulman, and Aravind Srinivasan. Splitters and near-optimal
derandomization. In 36th Annual Symposium on Foundations of Computer Science, Mil-
waukee, Wisconsin, USA, 23-25 October 1995, pages 182 -- 191, 1995.
Shmuel Onn and Pauline Sarrabezolles. Huge unimodular n-fold programs. SIAM J. Dis-
crete Math., 29(4):2277 -- 2283, 2015.
18
|
1104.3806 | 1 | 1104 | 2011-04-19T17:16:59 | How to Play Unique Games against a Semi-Random Adversary | [
"cs.DS",
"cs.CC"
] | In this paper, we study the average case complexity of the Unique Games problem. We propose a natural semi-random model, in which a unique game instance is generated in several steps. First an adversary selects a completely satisfiable instance of Unique Games, then she chooses an epsilon-fraction of all edges, and finally replaces ("corrupts") the constraints corresponding to these edges with new constraints. If all steps are adversarial, the adversary can obtain any (1-epsilon) satisfiable instance, so then the problem is as hard as in the worst case. In our semi-random model, one of the steps is random, and all other steps are adversarial. We show that known algorithms for unique games (in particular, all algorithms that use the standard SDP relaxation) fail to solve semi-random instances of Unique Games.
We present an algorithm that with high probability finds a solution satisfying a (1-delta) fraction of all constraints in semi-random instances (we require that the average degree of the graph is Omega(log k). To this end, we consider a new non-standard SDP program for Unique Games, which is not a relaxation for the problem, and show how to analyze it. We present a new rounding scheme that simultaneously uses SDP and LP solutions, which we believe is of independent interest.
Our result holds only for epsilon less than some absolute constant. We prove that if epsilon > 1/2, then the problem is hard in one of the models, the result assumes the 2-to-2 conjecture.
Finally, we study semi-random instances of Unique Games that are at most (1-epsilon) satisfiable. We present an algorithm that with high probability, distinguishes between the case when the instance is a semi-random instance and the case when the instance is an (arbitrary) (1-delta) satisfiable instance if epsilon > c delta. | cs.DS | cs | How to Play Unique Games against a Semi-Random Adversary
Study of Semi-Random Models of Unique Games
Alexandra Kolla
Microsoft Research
Konstantin Makarychev
Yury Makarychev
IBM Research
TTIC
Abstract
In this paper, we study the average case complexity of the Unique Games problem. We propose
a natural semi-random model, in which a unique game instance is generated in several steps. First an
adversary selects a completely satisfiable instance of Unique Games, then she chooses an ε -- fraction
of all edges, and finally replaces ("corrupts") the constraints corresponding to these edges with new
constraints. If all steps are adversarial, the adversary can obtain any (1 − ε) satisfiable instance, so then
the problem is as hard as in the worst case. In our semi-random model, one of the steps is random,
and all other steps are adversarial. We show that known algorithms for unique games (in particular, all
algorithms that use the standard SDP relaxation) fail to solve semi-random instances of Unique Games.
We present an algorithm that with high probability finds a solution satisfying a (1 − δ) fraction of all
constraints in semi-random instances (we require that the average degree of the graph is Ω(log k)). To
this end, we consider a new non-standard SDP program for Unique Games, which is not a relaxation for
the problem, and show how to analyze it. We present a new rounding scheme that simultaneously uses
SDP and LP solutions, which we believe is of independent interest.
Our result holds only for ε less than some absolute constant. We prove that if ε ≥ 1/2, then the
problem is hard in one of the models, that is, no polynomial -- time algorithm can distinguish between the
following two cases: (i) the instance is a (1 − ε) satisfiable semi -- random instance and (ii) the instance
is at most δ satisfiable (for every δ > 0); the result assumes the 2 -- to -- 2 conjecture.
Finally, we study semi-random instances of Unique Games that are at most (1 − ε) satisfiable. We
present an algorithm that with high probability, distinguishes between the case when the instance is a
semi-random instance and the case when the instance is an (arbitrary) (1 − δ) satisfiable instance if
ε > cδ (for some absolute constant c).
1
1
0
2
r
p
A
9
1
]
S
D
.
s
c
[
1
v
6
0
8
3
.
4
0
1
1
:
v
i
X
r
a
1 Introduction
In this paper, we study the average case complexity of the Unique Games problem in a semi-random model.
In the Unique Games problem, we are given a graph G = (V, E) (denote n = V ), a set of labels [k] =
{0, . . . , k − 1} and a set of permutations πuv on [k], one permutation for every edge (u, v). Our goal is to
assign a label (or state) xu ∈ [k] to every vertex u so as to maximize the number of satisfied constraints of
the form xv = πuv(xu). The value of the solution is the number of satisfied constraints.
The problem is conjectured to be very hard in the worst case. The Unique Games Conjecture (UGC) of
Khot [17] states that for every positive ε, δ and sufficiently large k, it is NP-hard to distinguish between the
case where at least a 1 − ε fraction of constraints is satisfiable, and the case where at most a δ fraction of all
constraints is satisfiable.
One reason which makes UGC particularly intriguing is its numerous implications. The conjecture, if
true, implies that the currently best known approximation algorithms for many important computational
problems have optimal approximation ratios. Indeed, since its origin, UGC has been successfully used to
prove often optimal hardness of approximation results for several important NP-hard problems such as MAX
CUT [18], Vertex Cover [19], Maximum Acyclic Subgraph [14], Max k-CSP [23, 15, 25], which are not
known to follow from standard complexity assumptions.
Arguably, a seemingly strong reason for belief in UGC is the failure of several attempts to design ef-
ficient algorithms for Unique Games using current state-of-the-art techniques, even though a large amount
of research activity in recent years has focused on the design of such algorithms. One direction of research
has concentrated on developing polynomial-time approximation algorithms for arbitrary instances of unique
games. The first algorithm was presented by Khot in his original paper on the Unique Games Conjec-
ture [17], and then several algorithms were developed in papers by Trevisan [26], Gupta and Talwar [10],
Charikar, Makarychev and Makarychev [8], Chlamtac, Makarychev, and Makarychev [9]. Another direc-
tion of research has been to study subexponential approximation algorithms for Unique Games. The work
was initiated by Kolla [21] and Arora, Impagliazzo, Matthews and Steurer [4] who proposed subexponen-
tial algorithms for certain families of graphs. Then, in a recent paper, Arora, Barak and Steurer [3] gave a
subexponential algorithm for arbitrary instances of Unique Games.
These papers, however, do not disprove the Unique Games Conjecture. Moreover, Khot and Vishnoi [20]
showed that it is impossible to disprove the Conjecture by using the standard semidefinite programming re-
laxation for Unique Games, the technique used in the best currently known polynomial-time approximation
algorithms for general instances of Unique Games. Additionally, Khot, Kindler, Mossel, and O'Donnell [18]
proved that the approximation guarantees obtained in [8] cannot be improved if UGC is true (except possibly
for lower order terms).
All that suggests that Unique Games is a very hard problem. Unlike many other problems, however,
we do not know any specific families of hard instances of Unique Games. In contrast, we do know many
specific hard instances of other problems. Many such instances come from cryptography; for example, it is
hard to invert a one-way function f on a random input, it is hard to factor the product z = xy of two large
prime numbers x and y. Consequently, it is hard to satisfy SAT formulas that encode statements "f (x) = y"
and "xy = z". There are even more natural families of hard instances of optimization problems; e.g.
• 3-SAT: Feige's 3-SAT Conjecture [11] states that no randomized polynomial time algorithm can dis-
tinguish random instances of 3-SAT (with a certain clause to variable ratio) from 1 − ε satisfiable
instances of 3-SAT (with non-negligible probability).
• Linear Equations in Z/2Z: Alekhnovich's Conjecture [1] implies that given a random (1 − ε)
satisfiable instance of a system of linear equations in Z/2Z, no randomized polynomial time algorithm
1
can find a solution that satisfies a 1/2 + δ fraction of equations (for certain values of parameters ε and
δ).
• Maximum Clique Problem: It is widely believed [16] that no randomized polynomial time algorithm
can find a clique of size (1 + ε) log2 n in a G(n, 1/2) graph with a planted clique of size m = n1/2−δ
(for every constant ε, δ > 0).
No such results are known or conjectured for Unique Games. In order to better understand Unique
Games, we need to identify, which instances of the problem are easy and which are potentially hard. That
motivated the study of specific families of Unique Games. Barak, Hardt, Haviv, Rao, Regev and Steurer [6]
showed that unique game instances obtained by parallel repetition are "easy" (we say that a family of 1 − ε
satisfiable instances is easy if there is a randomized polynomial-time algorithm that satisfies a constant
fraction of constraints) . Arora, Khot, Kolla, Steurer, Tulsiani, and Vishnoi [5] showed that unique games on
spectral expanders are easy (see also Makarychev and Makarychev [22], and Arora, Impagliazzo, Matthews
and Steurer [4]).
In this paper, we investigate the hardness of semi-random (semi-adversarial) instances of Unique Games.
In a semi-random model, an instance is generated in several steps; at each step, choices are either made
adversarially or randomly. Semi-random models were introduced by Blum and Spencer [7] (who considered
semi-random instances of the k-coloring problem) and then studied by Feige and Kilian [12], and Feige and
Krauthgamer [13].
In this paper, we propose and study a model, in which a 1 − ε satisfiable unique game instance is
generated as follows:
1. Graph Selection Step. Choose the constraint graph G = (V, E) with n vertices and m edges.
2. Initial Instance Selection Step. Choose a set of constraints {πuv}(u,v)∈E so that the obtained instance
is completely satisfiable.
3. Edge Selection Step. Choose a set of edges Eε of size εm = εE.
4. Edge Corruption Step. Replace the constraint for every edge in Eε with a new constraint.
Note that if an adversary performs all four steps, she can obtain an arbitrary 1 − ε satisfiable instance, so,
in this fully -- adversarial case, the problem is as hard as in the worst case. The four most challenging semi-
random cases are when choices at one out of the four steps are made randomly, and all other choices are
made adversarially. The first case -- when the graph G is random and, in particular, is an expander -- was
studied by Arora, Khot, Kolla, Steurer, Tulsiani, and Vishnoi [5], who showed that this case is easy.
of constraints (if the average degree of G is at least Ω(log k) and ε is less than some absolute constant).
We present algorithms for the other three cases that with high probability (w.h.p.) satisfy a 1− δ fraction
Theorem 1.1. For every k ≥ k0, ε0 > 0 and δ > C max(ε0, log k/√k) (where C and k0 are absolute con-
stants), there exists a randomized polynomial time algorithm that given a semi-random instance of Unique
Games with ε = ε0 (generated in one of the three models; see Section 2.3 for details) on a graph G with
average degree at least Ω(log k)δ−3, finds a solution of value at least 1 − δ with probability1 1 − o(1).
1The probability is over both the random choices that we make when we generate the semi-random instance, and the random
choices that the algorithm does. That is, the probability that the model generates a semi-random instance I, such that the algorithm
finds a solution of I of value at least 1 − δ with probability 1 − o(1), is 1 − o(1).
2
The theorem follows from Theorems 3.1, 4.1, and 6.1, in which we analyze each model separately, and
establish more precise bounds on the parameters for each model.
In our opinion, this is a very surprising result since the adversary has a lot of control over the semi-
random instance. Moreover, our results suggest that the Unique Games problem is different in nature than
several NP-hard problems like SAT, which are thought to be hard on average.
We want to point out that previously known approximation algorithms for Unique Games cannot find
good solutions of semi-random instances. Also techniques developed for analyzing semi-random instances
of other problems such as local analysis, statistical analysis, spectral gap methods, standard semidefinite
programming techniques seem to be inadequate to deal with semi-random instances of Unique Games. To
illustrate this point, consider the following example. Suppose that the set of corrupted edges is chosen at
random, and all other steps are adversarial ("the random edges, adversarial constraints case"). The adversary
generates a semi-random instance as follows. It first prepares two instances I1 and I2 of Unique Games.
The first instance I1 is the Khot -- Vishnoi instance [20] on a graph G with the label set [k] = {0, . . . , k − 1}
uv} whose SDP value is ε′ < ε/2 but which is only k−Ω(ε′) satisfiable. The second
and permutations {π1
instance I2 is a completely satisfiable instance on the same graph G with the label set {k, . . . , 2k − 1} and
permutations π2
uv = id. She combines these instances together: the combined instance is an instance on
the graph G with the label set [2k] = {0, . . . , 2k − 1}, and permutations {πuv : πuv(i) = π1
uv(i) if i ∈
[k], and πuv(i) = π2
uv(i), otherwise}. Once the adversary is given a random set of edges Eε, she randomly
changes ("corrupts") permutations {π2
uv, and then updates permutations
{πuv}(u,v)∈Eε accordingly. It turns out that the SDP value of I2 with corrupted edges is very close to ε, and
therefore, it is larger than ε′, the SDP value of I1 (if we choose parameters properly). So in this case the
SDP solution assigns positive weight only to the labels in [k] from the first instance. That means that the
SDP solution does not reveal any information about the optimal solution (the only integral solution we can
obtain from the SDP solution has value k−Ω(ε)). Similarly, algorithms that analyze the spectral gap of the
label extended graph cannot deal with this instance. Of course, in this example, we let our first instance,
I1, to be the Khot -- Vishnoi instance because it "cheats" SDP based algorithms. Similarly, we can take as
I1 another instance that cheats another type of algorithms. For instance, if UGC is true, we can let I1 to
be a 1 − ε′ satisfiable unique game that is indistinguishable in polynomial-time from a δ -- satisfiable unique
game.
Our algorithms work only for values of ε less than some absolute constants. We show that this restriction
is essential. For every ε ≥ 1/2 and δ > 0, we prove that no polynomial time algorithm satisfies a δ fraction
of constraints in the "adversarial constraints, random edges" model (only the third step is random), assuming
the 2 -- to -- 2 conjecture.
uv}(u,v)∈Eε but does not change π1
One particularly interesting family of semi-random unique games (captured by our model) are mixed
instances.
In this model, the adversary prepares a satisfiable instance, and then chooses a δ fraction of
edges and replaces them with adversarial constraints (corrupted constraints); i.e. she performs all four
steps in our model and can obtain an arbitrary 1 − δ satisfiable instance. Then the "nature" replaces every
corrupted constraint with the original constraint with probability 1 − ε. In our model, this case corresponds
to an adversary who at first prepares a list of corrupted constraints π′
uv, and then at the fourth step replaces
constraints for edges in Eε with constraints π′
uv (if an edge from Eε is not in the list, the adversary does not
modify the corresponding constraint).
Distinguishing Semi-Random At Most (1 − ε) Satisfiable Instances From Almost Satisfiable In-
stances. We also study whether semi-random instances of Unique Games that are at most (1− ε) satisfiable
can be distinguished from almost satisfiable instances of Unique Games. This question was studied for other
problems. In particular, Feige's "Random 3-SAT Conjecture" states that it is impossible to distinguish be-
3
tween random instances of 3-SAT (with high enough clause density) and 1−δ satisfiable instances of 3-SAT.
In contrast, we show that in the "adversarial edges, random constraints" case (the fourth step is random),
semi-random (1− ε)-satisfiable instances can be efficiently distinguished from (arbitrary) (1− δ)-satisfiable
instances when ε > cδ (for some absolute constant c). (This problem, however, is meaningless in the other
two cases -- when the adversary corrupts the constraints -- since then she can make the instance almost
satisfiable even if ε is large.)
Linear Unique Games We separately consider the case of Linear Unique Games (MAX Γ-LIN). In the
semi-random model for Linear Unique Games, we require that constraints chosen at the second and fourth
steps are of the form xu − xv = suv(mod k). Note that in the "random edges, adversarial constraints"
model, the condition that constraints are of the form xu − xv = suv(mod k) only restricts the adversary
(and does not change how the random edges are selected). Therefore, our algorithms for semi-random
general instances still applies to this case. However, in the "adversarial edges, random constraints" case, we
need to sample constraints from a different distribution of permutations at the fourth step: for every edge
(u, v) we now choose a random shift permutation xv = xu − suv, where suv ∈U Z/kZ. We show that
our algorithm still works in this case; the analysis however is different. We believe that it is of independent
interest. We do not consider the case where only the initial satisfying assignment is chosen at random,
since for Linear Unique Games, the initial assignment uniquely determines the constraints between edges
(specifically, suv = xu − xv(mod k)). Thus the case when only the second step is random is completely
adversarial.
It is interesting that systems of linear equations affected by noise with more than two variables per
equations are believed to be much harder. Suppose we have a consistent system of linear equations Ax = b
over Z/2Z. Then we randomly change an ε fraction of entries of b. Alekhnovich [1] conjectured that no
polynomial-time algorithm can distinguish the obtained instance from a completely random instance even
if ε ≈ n−c, for some constant c (Alekhnovich stated his conjecture both for systems with 3 variables per
equation and for systems with an arbitrary number of variables per equation).
Our results can be easily generalized to Unique Games in arbitrary Abelian groups. We omit the details
in the conference version of this paper.
1.1 Brief Overview of Techniques
In this paper, we develop new powerful algorithmic techniques for solving semi-random instances of unique
games. We use different algorithms for different models. First, we outline how we solve semi-random
unique games in the "adversarial constraints, random edges" model (see Section 3 for details). As we
explained above, we cannot use the standard SDP relaxation (or other standard techniques) to solve semi-
random instances in this model. Instead, we consider a very unusual SDP program for the problem, which
we call "Crude SDP" (C-SDP). This SDP is not even a relaxation for Unique Games and its value can
be large when the instance is satisfiable. The C-SDP assigns a unit vector ui to every vertex (u, i) of
the label -- extended graph (for a description of the label -- extended graph we refer the reader to Section 2).
We use vectors ui to define the length of edges of the label -- extended graph: the length of ((u, i), (v, j))
equals kui − vjk2. Then we find super short edges w.r.t. the C-SDP solution, those edges that have length
O(1/ log k). One may expect that there are very few short edges since for a given C-SDP most edges will be
long if we choose the unique games instance at random. We prove, however, that for every C-SDP solution
{ui}, with high probability (over the semi-random instance) either
1. there are many super short edges w.r.t. {ui} in the satisfiable layer of the semi-random game,
2. or there is another C-SDP solution of value less than the value of the solution {ui}.
4
Here, as we describe later on in section 2, the "satisfiable layer" corresponds to the representation of the
satisfying assignment in the label -- extended graph.
Note that if our instance is completely satisfiable, then in the optimal (integral) C-SDP solution all the
edges that correspond to the satisfiable layer have length zero and, therefore are super short.
Our proof shows how to combine the C-SDP solution with an integral solution so that the C-SDP value
goes down unless almost all edges in the satisfiable layer are super short. We then show that this claim holds
with high probability not only for one C-SDP solution but also for all C-SDP solutions simultaneously.
The idea behind this step is to find a family F of representative C-SDP solutions and then use the union
bound. One of the challenges is to choose a very small family F, so that we can prove our result under the
assumption that the average degree is only Ω(log k). The result implies that w.h.p. there are many super
short edges w.r.t. the optimal C-SDP solution.
Now given the set of super short edges, we need to find which of them lie in the satisfiable layer. We
write and solve an LP relaxation for Unique Games, whose objective function depends on the set of super
short edges. Then we run a rounding algorithm that rounds the C-SDP and LP solutions to a combinatorial
solution using a variant of the "orthogonal separators" technique developed in [9].
Our algorithm for the "adversarial edges, random constraints" model is quite different. First, we solve
the standard SDP relaxation for Unique Games. Now, however, we cannot claim that many edges of the
label -- extended graph are short. We instead find the "long" edges of the graph G w.r.t. the SDP solution.
We prove that most corrupted edges are long, and there are at most O(ε) long edges in total (Theorem 4.4).
We remove all long edges and obtain a highly -- satisfiable instance. Then we write and solve an SDP for this
instance, and round the SDP solution using the algorithm of [8].
We also present algorithms for two more cases: the case of "adversarial edges, random constraints"
where the constraints are of the special form MAX-Γ-LIN and the case of "random initial constraints".
In this paper, we develop several new techniques. In particular, we propose a novel C-SDP program, and
then show how to exploit the optimality of a C-SDP solution in the analysis. We develop a rounding algo-
rithm that simultaneously uses SDP and LP solutions. We demonstrate how to bound the number of different
SDP solutions using dimension reduction methods and other tricks. We believe that our techniques are of
independent interest and that they will prove useful for solving semi-random instances of other problems.
2 Notation and Preliminaries
2.1 The Label-Extended Graph
For a given instance of Unique Games on a constraint graph G = (V, E), with alphabet size k and constraints
{πuv}(u,v)∈E we define the Label-Extended graph M (V ′ = V × [k], E′) associated with that instance as
follows: M has k vertices Bv = {v0,··· , vk−1} for every vertex v ∈ V . We refer to this set of vertices as
the block corresponding to v. M has a total of V blocks, one for each vertex of G. Two vertices ui, vj ∈ V ′
are connected by an edge if (u, v) ∈ E and πuv(i) = j. We refer to a set of nodes L = {u(z)
z=1 as
a "layer" if L contains exactly one node from each block Bu(z). We note that a layer L can be seen as
an assignment of labels to each vertex of G. If a layer L consists of vertices with the same index i, i.e.
L = {u(z)
z=1, we will call L the i-th layer.
i(z)}V
i }V
5
2.2 Standard SDP for Unique Games
Our algorithms use the following standard SDP relaxation for Unique Games (see also [17, 20, 8, 9]).
min
subject to
1
k
kui − vπuv(i)k2
2E X(u,v)∈E Xi∈[k]
Xi=1
kuik2 = 1
hui, uji = 0
hui, vji ≥ 0
kui − vjk2 ≤ kui − wlk2 + kwl − vjk2
for all u ∈ V
for all u ∈ V, i, j ∈ [k](i 6= j)
for all u, v ∈ V, i, j ∈ [k](i 6= j)
for all u, v, w ∈ V, i, j, l ∈ [k].
In this relaxation, we have a vector variable ui for every vertex u and label i. In the intended solution, ui is
an indicator variable for the event "xu = i". That is, if xu = i then ui = e; otherwise, and ui = 0; where e
is a fixed unit vector. The objective function measures the fraction of unsatisfied constraints: if the unique
game has value 1 − ε, then the value of the intended SDP solution equals ε (and, therefore, the value of the
optimal SDP solution is at most ε).
finds a solution of value 1 − O(√ε log k). We will use this approximation algorithm as a subroutine (we
Given an SDP solution of value ε, the approximation algorithm of Charikar, Makarychev, and Makarychev [8]
will refer to it as CMMa). We will also use the following fact.
Lemma 2.1 (see e.g. Lemmas A.1 and A.2 in [9]). Suppose, we are given two random Gaussian variables
γ1 and γ2 with mean 0 and variance 1 (not necessarily independent), and a parameter k ≥ 2. Let α =
1/(2k2). Consider a threshold t s.t. Pr(γ1 ≥ t) = Pr(γ2 ≥ t) = α. Then Pr(γ1 ≥ t and γ2 ≥ t) ≥
α(cid:16)1 −q 1
c∗ Var(γ1 − γ2) log k(cid:17) for some absolute constant c∗.
2.3 Models
In what follows, we will use several models for generating semi-random (1 − ε) satisfiable instances of
Unique Games2:
1. "Random Edges, Adversarial Constraints" Model. The adversary selects a graph G(V, E) on n
vertices and m edges and an initial set of constraints {πuv}(u,v)∈E so that the instance is completely
satisfiable. Then she adds every edge of E to a set Eε with probability ε (the choices for different
edges are independent). Finally, the adversary replaces the constraint for every edge in Eε with a
new constraint of her choice. Note that this model also captures the case where at the last step the
constraints for every edge in Eε are replaced with a new random constraint (random adversary).
2. "Adversarial Edges, Random Constraints" Model. The adversary selects a graph G(V, E) on n
vertices and m edges and an initial set of constraints {πuv}(u,v)∈E so that the instance is completely
satisfiable. Then she chooses a set Eε of εE edges. Finally, the constraint for every edge in Eε is
randomly replaced with a new constraint. We will also consider some variations of this model, where
at all steps the constraints are MAX Γ-LIN. In particular, at the last step, choosing a random constraint
of the form MAX Γ-LIN, corresponds to choosing a random value s ∈ [Γ].
2The parameters of the models are the number of vertices n, the number of edges m, and the probability ε that an edge is
corrupted.
6
3. "Random Initial Constraints" Model. The adversary chooses the constraint graph G = (V, E) and
a "planted solution" {xu}. Then for every edge (u, v) ∈ E, she randomly chooses a permutation
(constraint) πuv such that πuv(xu) = xv (among (k − 1)! possible permutations). Then the adversary
chooses an arbitrary set Eε of edges of size at most εE and replaces constraint πuv with a constraint
π′
uv of her choice for (u, v) ∈ Eε.
Remark 2.2. Without loss of generality, we will assume, when we analyze the algorithms, that the initial
completely satisfying assignment corresponds to the "zero" layer. I.e. for every edge (u, v), πuv(0) = 0.
Note that in reality, the real satisfying assignment is hidden from us.
3 Random Edges, Adversarial Constraints
In this section, we study the "random edges, adversarial constraints" model and prove the following result.
Theorem 3.1. Let k ∈ N (k ≥ 2), ε ∈ (0, 1/3), and η ∈ (0, 1). There exists a polynomial-time approxima-
tion algorithm, that given an instance of Unique Games from the "random edges, adversarial constraints"
model on graph G with Cη−3(1/3 − ε)−4n log k(log(η−1 log k))2 edges (C is a sufficiently large absolute
constant), finds a solution of value (1 − ε − η)/(1 + ε + η) + O(1/k), with probability 1 − o(1).
Corollary 3.2. There exists a polynomial-time approximation algorithm, that given an instance of unique
games from the "random edges, adversarial constraints" model on graph G with Cn log k(log log k)2 edges
(C is a sufficiently large absolute constant) and ε ≤ 1/4, finds a solution of value 1/2, with probability
1 − o(1).
Remark 3.3. We can make the constant in the O(1/k) term in (1− ε− η)/(1 + ε + η) + O(1/k) arbitrarily
small by increasing the value of C (and decreasing the value of α in the proof). We omit the details to
simplify the exposition.
The main challenge in solving "random edges, adversarial constraints" unique games is that the stan-
dard SDP relaxation may assign zero vectors to layers corresponding to the optimal solution (as well as to
some other layers) and assign non-zero vectors to layers, where every integral solution satisfies very few
constraints. To address this issue, we introduce a new slightly modified SDP. As usual the SDP has a vector
ui for every vertex -- label pair (u, i) ∈ V × [k]. We require that vectors ui, ui′ corresponding to the same
vertex u are orthogonal: hui, ui′i = 0 for all u ∈ V and i, i′ ∈ [k], i 6= i′. We also impose triangle inequality
constraints:
1
2kui − vjk2 +
1
2kui′ − vjk2 ≥ 1,
for all (u, v) ∈ E and i, i′, j ∈ [k], i 6= i′; and require that all vectors have unit length: kuik = 1 for all
u ∈ V and i ∈ [k]. Observe, that our SDP is not a relaxation3, since the integer solution does not satisfy the
last constraint. The objective function is
min X(u,v)∈E Xi∈[k]
j=πuv(i)
kui − vjk2
2
.
Usually, this objective function measures the number of unsatisfied unique games constraints. However, in
our case it does not. In fact, it does not measure any meaningful quantity. Note, that the value of the SDP
3Unless, the unique game is from a special family like Linear Unique Games (see Section 5).
7
can be arbitrary large even if the unique games instance is satisfiable. We call this SDP -- the Crude SDP or
C-SDP. Given a C-SDP solution, we define the set of super short edges, which play the central role in our
algorithm.
Definition 3.4. We say that an edge ((u, i), (v, j)) in the label -- extended graph is η -- super short, if kui −
vjk2 ≤ c∗η2/ log k, here c∗ is an absolute constant defined in Lemma 2.1. We denote the set of all η -- super
short edges by Γη.
In Section 3.1, we prove the following surprising result (Theorem 3.5), which states that all but very few
edges in the zero level of the label-extended graph are super short.
Theorem 3.5. Let k ∈ N (k ≥ 2), c ∈ (0, 1), ε ∈ (0, 1/3), η ∈ (c, 1) and γ ∈ (ε + c, 1) and let G = (V, E)
be an arbitrary graph with at least Cη−2(γ − ε)−1(1/3− ε)−4 × n log k(log(c−1 log k))2, edges. Consider
a semi-random instance of Unique Games in the "random edges, adversarial constraints" model. Let {ui}
be the optimal solution of the C-SDP. Then with probability 1 − o(1), the set
Γ0
η = Γη ∩ {((u, 0), (v, 0)) : (u, v) ∈ E}
contains at least (1 − γ)E edges.
More concretely, we proceed as follows. First, we solve the C-SDP. Then, given the C-SDP solution,
we write and solve an LP to obtain weights x(u, i) ∈ [0, 1] for every (u, i) ∈ V × [k]. These weights are
in some sense substitutes for lengths of vectors in the standard SDP relaxation. In the LP, for every vertex
u ∈ V , we require that
x(u, i) = 1.
Xi∈[k]
max X((u,i),(v,j))∈Γη
The objective function is
min(x(u, i), x(v, j))
(note that the objective function depends on the C-SDP solution). Denote the value of the LP by LP .
The intended solution of this LP is x(u, 0) = 1 and x(u, i) = 0 for i 6= 0. Since the LP contribution
η is 1, the value of the intended solution is at least Γ0
of every edge in Γ0
η. Applying Theorem 3.5 with
γ = ε + η, we get Γ0
η ≥ (1 − γ)E = (1 − ε − η)E, so LP ≥ (1 − ε − η)E. In the next section, we
present an approximation algorithm (which rounds C-SDP an LP solutions) and its analysis. We prove the
approximation guarantee in Lemma 3.9, which implies Theorem 3.1.
3.1 Lower Bound on the Number of Super Short Edges: Proof of Theorem 3.5
We need the following lemma.
Lemma 3.6. Let G = (V, E) be an arbitrary graph on n vertices, and let ε ∈ [0, 1/3), ρ = 1/3 − ε,
ν ∈ (0, ρ). Suppose, that {Zuv}(u,v)∈E are i.i.d. Bernoulli random variables taking values 1 with probability
ε and 0 with probability (1 − ε). Define the payoff function p : {0, 1} × R → R as follows
p(z, α) =(−2α,
α,
if z = 1;
if z = 0.
8
Then, with probability at least 1 − o(1) for every set of vectors {u0}u∈V satisfying (for some significantly
large absolute constant C)
1
2 X(u,v)∈E
ku0 − v0k2 ≥ CνE + Cρ−4 log2(1/ν)n
the following inequality holds
p(Zuv,ku0 − v0k2) > 0.
X(u,v)∈E
(1)
(2)
Proof. We need the following dimension reduction lemma, which is based on the Johnson -- Lindenstrauss
Lemma and is fairly standard (for an example of using the Johnson -- Lindenstrauss Lemma in SDP rounding
see Raghavendra and Steurer [24]).
Lemma 3.7. For every positive ζ, η and ν, there exists a set N of unit vectors of size at most
such that for every set of unit vectors U there exists a randomized mapping ϕ : U → N satisfying the
following property: for every u, v ∈ U,
exp(cid:0)O(ζ −2 log(1/η) log(1/ν))(cid:1)
Pr((1 + ζ)−1ku − vk2 − η2 ≤ kϕ(u) − ϕ(v)k2 ≤ (1 + ζ)ku − vk2 + η2) ≥ 1 − ν.
(3)
Proof sketch. To prove the lemma, we consider an O(ζ −2 log(1/ν)) dimensional space L and choose a
η2/32-net N in it. The size of N satisfies the bound in the lemma. To construct the mapping ϕ : U → N,
we project all vectors from U to L using the Johnson -- Lindenstrauss lemma and then define ϕ(u) to be the
closest vector u∗ ∈ N to the projection of u.
Proof of Lemma 3.6. Set ζ = ρ/5 ≡ (1/3 − ε)/5. Define a new payoff function,
pζ(z, α) =(−2(1 + ζ)α,
(1 + ζ)−1α,
if z = 1;
if z = 0.
Note, that ζ < 1/15 and pζ(0, α)− pζ (1, α) = ((1+ ζ)−1− (−2(1+ ζ)))α ≤ 4α (for α ≥ 0). Suppose, that
for a given realization {Z ∗
uv}(u,v)∈E of {Zuv}(u,v)∈E there exists a set of unit vectors {u0}u∈V satisfying
condition (1) and violating (2). Embed vectors {u0}u∈V into a net N of size exp(cid:0)O(ρ−2 log2(1/ν))(cid:1) using
Lemma 3.7, such that for (1 − ν2/2) fraction of all edges (u, v) ∈ E, the following condition holds
(1 + ζ)−1ku0 − v0k2 − ν2/8 ≤ ku∗ − v∗k2 ≤ (1 + ζ)ku0 − v0k2 + ν2/4,
here u∗ is the image of u0; v∗ is the image of v0. Then,
pζ(Z ∗
uv,ku∗ − v∗k2) < ν2E < νρE = 5νζE.
X(u,v)∈E
and (since we choose C to be sufficiently large)
ku∗ − v∗k2 ≥ 40νE + Cρ−4 log2(1/ν)n.
X(u,v)∈E
9
(4)
(5)
Thus, the existence of vectors {u0}u∈V satisfying condition (1) and violating (2) implies the existence
of vectors {u∗}u∈V satisfying (4) and (5). We now show that for a random {Zuv} such vectors {u∗} exist
with exponentially small probability. Fix a set {u∗}u∈V ⊂ N and consider random {Zuv},
E X(u,v)∈E
pζ(Zuv,ku∗ − v∗k2) ≥ X(u,v)∈E
((1 − ε)(1 − ζ) − 2(1 + ζ)ε)ku∗ − v∗k2
ku∗ − v∗k2
= (1 − 3ε − 3εζ − ζ) X(u,v)∈E
≥ (1 − 3ε − 4ζ) X(u,v)∈E
≥ ζ X(u,v)∈E
ku∗ − v∗k2.
ku∗ − v∗k2
pζ(Zuv,ku∗ − v∗k2) < 2νρE
By Hoeffding's inequality (using that pζ(0,ku∗−v∗k2)−pζ(1,ku∗−v∗k2) ≤ 4ku∗−v∗k2,P(u,v)∈E ku∗−
v∗k2 − 10νE ≥ 3
2ζ 2(cid:16)P(u,v)∈E ku∗ − v∗k2 − 10νE(cid:17)2
Pr
X(u,v)∈E
P(u,v)∈E 16ku∗ − v∗k4
ζ 2(cid:16)P(u,v)∈E ku∗ − v∗k2(cid:17)2
P(u,v)∈E 64ku∗ − v∗k2
ku∗ − v∗k2
64 X(u,v)∈E
ρ−4 log2(1/ν)n(cid:19)
ρ−2 log2(1/ν)n(cid:19)
4P(u,v)∈E ku∗ − v∗k2 and ku∗ − v∗k2 ≤ 4),
≤ exp
−
≤ exp
−
≤ exp
−
≤ exp(cid:18)−
= exp(cid:18)−
The number of all possible subsets {u∗}u∈V ⊂ N is at most
Cζ 2
64
C
25 · 64
ζ 2
Hence, by the union bound with probability at least 1 − exp(−n) = 1 − o(1) for random {Zuv}(u,v)∈E,
there does not exist a set of vectors {u0}u∈V satisfying condition (1) and violating (2).
Nn ≤ exp(cid:0)O(nρ−2 log2(1/ν))(cid:1) .
i } be the optimal SDP solution. Pick a unit vector e orthogonal to all vectors
i for i 6= 0 (for all u ∈ V ). Note that restricted to
0 = e and uint
i = u∗
Proof of Theorem 3.5. Let {u∗
u∗
i . Define a new SDP solution uint
0 }u∈V this solution is integral. Since {u∗
{uint
i − v∗
ku∗
X(u,v)∈E Xi∈[k]
j=πuv(i)
i } is the optimal solution,
jk2 ≤ X(u,v)∈E Xi∈[k]
j=πuv(i)
kuint
i − vint
j k2.
10
Denote by Eε the set of corrupted edges. Let Zuv = 1, if (u, v) ∈ Eε and Zuv = 0, otherwise. Let
Eε = {(u, v) ∈ E : πuv(0) 6= 0}. Clearly, Eε ⊂ Eε. Write,
X(u,v)∈E Xi∈[k]
jk2 − kuint
i − vint
j k2 =
i − v∗
ku∗
j=πuv(i)
0 − v∗
ku∗
0k2 + X(u,v)∈ Eε
πuv(0)k2 + kuint
0 − vint
πvu(0) − vint
0 k2.
0 − v∗
πuv(0)k2 + ku∗
πvu(0) − v∗
0k2
ku∗
= X(u,v)∈E\ Eε
− X(u,v)∈ Eε
πuv(0)k2 = kuint
kuint
For (u, v) ∈ Eε, we have kuint
Thus,
0 − vint
πvu(0) − vint
0 k2 = 2 and ku∗
0− v∗
πuv(0)k2 ≥ 2−ku∗
0− v∗
0k2.
X(u,v)∈E Xi∈[k]
j=πuv(i)
ku∗
i − v∗
jk2 − kuint
i − vint
j k2 ≥ X(u,v)∈E\ Eε
≥ X(u,v)∈E\Eε
= X(u,v)∈E
ku∗
0 − v∗
ku∗
0 − v∗
0k2 − 2 X(u,v)∈ Eε
0k2 − 2 X(u,v)∈Eε
0 − v∗
0k2),
p(Zuv,ku∗
ku∗
0 − v∗
0k2
ku∗
0 − v∗
0k2
where p(·,·) is the function from Lemma 3.6. By Lemma 3.6, with probability 1 − o(1), for some absolute
constant C ′, ρ = (1/3 − ε), ν = c∗η2(γ−ε)
2C ′ log k ,
1
2E X(u,v)∈E
ku∗
0 − v∗
0k2 <
C ′νE + C ′ρ−4 log2(1/ν)n
E
≤ c∗η2(γ − ε) log−1 k.
The last inequality holds, because
E ≥ Cη−2(γ − ε)−1ρ−4n log k(log(c−1 log k))2 ≥
C ′
2c∗ η−2(γ − ε)−1ρ−4n log k log2(1/ν)2.
0 − v∗
0 − v∗
0k2 ≤ c∗η2/ log k.
By the Markov inequality, for all but (γ − ε)/2 fraction of edges (u, v) ∈ E, ku∗
Recall that a pair ((u, 0), (v, 0)) is an η -- super short edge in the label -- extended graph if πuv(0) = 0 (i.e.,
0k2 ≤ c∗η2/ log k. By the Chernoff Bound,
(u0, v0) is an edge of the label -- extended graph) and ku∗
Eε ≤ (ε + (γ − ε)/2)E with probability (1 − o(1)); therefore, πuv(0) 6= 0 for at most a ε + (γ − ε)/2
fraction of edges. Thus, with probability (1− o(1)), there are at least (1− (γ − ε)/2− (ε + (γ − ε)/2))m =
(1 − γ)E η -- super short edges.
3.2 SDP and LP Rounding
We now present an algorithm that given a C-SDP solution {ui} and an LP solution {x(u, i)} finds an integer
solution. We first present a procedure for sampling subsets of vertex -- label pairs, which is an analog of the
algorithm for finding orthogonal separators in CMMb [9].
11
LP Weighted Orthogonal Separators
Input: An SDP solution {ui}, an LP solution {x(u, i)}.
Output: A set S of label vertex pairs (u, i).
1. Set a parameter α = 1/(2k2), which we call the probability scale.
2. Generate a random Gaussian vector g with independent components distributed as N (0, 1).
3. Fix a threshold t s.t. Pr(ξ ≥ t) = α, where ξ ∼ N (0, 1).
4. Pick a random uniform value r in the interval (0, 1).
5. Find set
S = {(u, i) ∈ V × [k] : hui, gi ≥ t and x(u, i) ≥ r} .
6. Return S.
The rounding algorithm is given below.
LP and SDP Based Rounding Algorithm
Input: An instance of unique games.
Output: An assignment of labels to the vertices.
1. Solve the SDP.
2. Find the set of all super short edges Γη.
3. Solve the LP.
4. Mark all vertices unprocessed.
5. while (there are unprocessed vertices)
• Sample a set S of vertex -- label pairs using LP weighted orthogonal separators.
• For all unprocessed vertices u:
-- Let Su = {i : (u, i) ∈ S}
-- If Su contains exactly one element i, assign label i to u and mark u as processed.
If after nk/α iterations, there are unprocessed vertices, the algorithm assigns arbitrary labels to them and
terminates.
Lemma 3.8. Let S be an LP weighted orthogonal separator. Then, for every (u, i) ∈ V × [k],
• Pr((u, i) ∈ S) = αx(u, i).
For every ((u, i), (v, j)) ∈ Γη and (u, i′) ∈ V × [k] (i′ 6= i),
12
• Pr((u, i) ∈ S and (v, j) ∈ S) ≥ α min(x(u, i), x(v, j))(1 − η).
• Pr((u, i) ∈ S; (v, j) ∈ S; (u, i′) ∈ S) ≤ α min(x(u, i), x(v, j))/(2k2 ).
Proof. We have
Pr((u, i) ∈ S) = Pr(hui, gi ≥ t and x(u, i) ≥ r) = Pr(hui, gi ≥ t) Pr(x(u, i) ≥ r) = αx(u, i).
Then, by Lemma 2.1,
Pr((u, i) ∈ S and (v, j) ∈ S) = Pr(hui, gi ≥ t and hvj, gi ≥ t) Pr(min(x(u, i), x(v, j)) ≥ r)
≥ α(1 − η) min(x(u, i), x(v, j)).
Finally, we have (below we use that hui, gi and hui′, gi are independent random variables)
Pr((u, i) ∈ S; (v, j) ∈ S; (u, i′) ∈ S) ≤ Pr(hui, gi ≥ t) Pr(hui′ , gi ≥ t) Pr(min(x(u, i), x(v, j)) ≥ r)
= α2 min(x(u, i), x(v, j)) ≤ α min(x(u, i), x(v, j))/(2k2 ).
Lemma 3.9. Given a C-SDP solution {ui} and an LP solution {x(u, i)} of value at least LP ≥ xE, the
algorithm finds a solution to the unique games instance that satisfies (1 − η)x/(2 − (1 − η)x) − O(1/k)
fraction of all constraints in the expectation.
Proof. Consider an arbitrary edge (u, v). We estimate the probability that the algorithm assigns labels that
satisfy the constraint πuv. For simplicity of presentation, assume that πuv(i) = i (we may always assume this
by renaming the labels of v). Let δi(u, v) = min(x(u, i), x(v, i)) if ((u, i), (v, i)) ∈ Γη; and δi(u, v) = 0,
otherwise. Let δ(u, v) =Pi δi(u, v) and δ′(u, v) = δ(u, v)(1 − η − 1/k).
Consider an arbitrary iteration at which both u and v have not yet been processed. By Lemma 3.8 (item
2), if ((u, i), (v, i)) ∈ Γη, then
Pr(i ∈ Su and i ∈ Sv) ≥ α min(x(u, i), x(v, i))(1 − η).
Then, by Lemma 3.8 (3) and the union bound, the probability that Su or Sv contains more than one element
and i ∈ Su, i ∈ Sv is at most α min(x(u, i), x(v, i))/k. Hence, the algorithm assigns i to both u and v with
probability at least
α min(x(u, i), x(v, i))(1 − η − 1/k) = αδi(u, v) × (1 − η − 1/k).
The probability that the algorithm assigns the same label to u and v is at least
Xi:((u,i),(v,i))∈Γη
αδi(u, v) × (1 − η − 1/k) = αδ′(u, v).
The probability that the algorithm assigns a label to u is at most α and similarly the probability that the
algorithm assigns a label to v is at most α. Thus the probability that it assigns a label to either u or v is at
most α(2 − δ′(u, v)).
Hence, the probability that the algorithm assigns the same label to u and v at one of the iterations is
at least (note that the probability that there are unlabeled vertices when the algorithm stops after nk/α
13
iterations is exponentially small, therefore, for simplicity we may assume that the number of iterations is
infinite)
∞
Xt=0
(1 − α(2 − δ′(u, v)))tαδ′(u, v) =
αδ′(u, v)
α(2 − δ′(u, v))
=
δ′(u, v)
2 − δ′(u, v)
.
The function δ 7→ δ/(2 − δ) is convex on (0, 2) and
E X(u,v)∈E
E X(u,v)∈E
δ′(u, v) =
1
1
δ(u, v)(1 − η − 1/k) ≥ x(1 − η − 1/k),
thus, by Jensen's inequality, the expected number of satisfied constraints is at least
x(1 − η − 1/k)
2 − x(1 − η − 1/k)E.
3.3 Hardness: Semi-Random Instances for ε ≥ 1/2
In this section, we show that the problem becomes hard when ε ≥ 1/2 assuming the 2 -- to -- 2 conjecture.
The 2 -- to -- 2 conjecture follows from Khot's 2 -- to -- 1 conjecture4, which is a frequently used complexity
assumption [17]. We prove the following theorem.
Theorem 3.10. For every ε ≥ 1/2 and δ > 0, no polynomial-time algorithm can distinguish with probability
greater than o(1) between the following two cases:
1. Yes Case: the instance is a 1− ε satisfiable semi-random instance (in the "random edges, adversarial
constraints" model),
2. No Case: the instance is at most δ satisfiable.
This result holds if the 2 -- to -- 2 conjecture holds.
Before we proceed with the proof we remind the reader the definition of 2 -- to -- 2 games and 2 -- to -- 2
conjecture.
Definition 3.11. In a 2 -- to -- 2 game, we are given a graph G = (V, E), a set of labels [k] = {0, . . . , k−1} (k
is even) and set of constraints, one constraint for every edge (u, v). Each constraint is defined by a 2 -- to -- 2
predicate Πuv: for every label i there are exactly two labels j such that Πuv(i, j) = 1 (the predicate is
satisfied); similarly, for every j there are exactly two labels i such that Πuv(i, j) = 1. Our goal is to assign
a label xu ∈ [k] to every vertex u so as to maximize the number of satisfied constraints Πuv(xu, xv) = 1.
The value of the solution is the number of satisfied constraints.
Definition 3.12. The 2 -- to -- 2 conjecture states that for every δ > 0 and sufficiently large k, there is no
polynomial time algorithm that distinguishes between the following two cases (i) the instance is completely
satisfiable and (ii) the instance is at most δ satisfiable.
4 The 2 -- to -- 1 conjecture implies the 2 -- to -- 2 conjecture because every 2 -- to -- 1 game can be converted to a 2 -- to -- 2 game of the
same value as follows. Let I be an instance of a 2 -- to -- 1 game on a graph (L, R, E), s.t. every constraint has degree 2 on the left
side, and degree 1 on the right side. We create two copies a1 and a2 for every label a for vertices in L. We replace each constraint
Πlr with a constraint Π∗
lr(a2, b) = Πlr(a, b). We obtain a 2 -- to -- 2 game. It is clear that its value
equals the value of the original 2 -- to -- 1 game.
lr defined by Π∗
lr(a1, b) = Π∗
14
Proof. We construct a randomized reduction that maps every satisfiable 2 -- to -- 2 game to a semi-random
unique game, and every at most δ satisfiable 2 -- to -- 2 game to an at most δ satisfiable unique game. Suppose
we are given an instance of a 2 -- to -- 2 game on a graph G with predicates Πuv. For each predicate Πuv, we find
two permutations π0
uv(i)
for every label i as follows. We consider a bipartite graph on vertex set {(u, i) : i ∈ [k]}∪{(v, j) : j ∈ [k]},
in which two vertices (u, i) and (v, j) are connected with an edge if Πuv(i, j) = 1. Every vertex has degree
2 in this graph. Therefore, it is a union of two matchings; each of them defines a permutation. Now for
every edge (u, v), we choose one of the two permutations, π0
uv , at random. We obtain a unique game
instance.
uv(i)) = 1 and Πuv(i, π1
uv such that Πuv(i, π0
uv(i)) = 1, and π0
uv or π1
uv and π1
uv(i) 6= π1
uv(xu) or xv = π1
First of all it is clear, that the value of the unique game is at most the value of the 2 -- to -- 2 game (if
a labeling {xu} satisfies xv = π0
uv(xu) then it also satisfies Πuv(xu, xv)). Hence our
reduction maps an at most δ satisfiable 2 -- to -- 2 game to an at most δ satisfiable unique game. Now suppose
that the 2 -- to -- 2 game instance is completely satisfiable. We show how an adversary can generate semi-
random instances that have the same distribution as instances generated by our reduction. The adversary
finds a solution {x∗
u} to the 2 -- to -- 2 game that satisfies all constraints Πuv. For every edge (u, v), she
chooses ruv ∈ {0, 1} such that πruv
v. She obtains a completely satisfiable unique game on G with
constraints πruv
uv (xu) = xv. Now she performs the random step -- she chooses a random set of edges Eε;
every edge belongs to Eε with probability ε. She replaces every constraint πruv
uv (xu) = xv with the constraint
π1−ruv
(xu) = xv with probability 1/(2ε), and returns the obtained instance. Note that the constraint for the
uv
edge (u, v) is π0
uv(xu) = xv with probability 1/2. Therefore, the
distribution of instances coincides with the distribution generated by our reduction.
uv(xu) = xv with probability 1/2 and π1
u) = x∗
uv (x∗
Our positive results (for ε < 1/3) apply only to graphs with average degree greater than ε−1 log k log log k.
We want to point out that our negative results (for ε ≥ 1/2) also apply to graphs with high average degree.
Indeed, note that our reduction does not change the constraint graph. As the following lemma shows, if the
2 -- to -- 2 conjecture is true then it is also true for instances on graphs with very high average degree (as large
as n1−θ ≫ k log k).
Lemma 3.13. Suppose that the 2 -- to -- 2 conjecture holds. Then it also holds for graphs with average degree
at least n1−θ (for every θ > 0).
Proof. Suppose we are given an instance I of a 2 -- to -- 2 game on a graph G = (V, E) with constraints
Πuv(xu, xv). We construct a new instance I ′ of a 2 -- to -- 2 game as follows. For every vertex u, we create a
"cloud" of N = V 2/θ new vertices Vu. We connect all vertices in Vu with all vertices in Vv for every edge
(u, v). For every edge between vertices a ∈ Vu and b ∈ Vv, we add the constraint Πuv(xa, xb) = 1 (where
Πuv is the predicate for the edge (u, v) in I). We obtain a 2 -- to -- 2 game on a graph with average degree
2N 2E/(NV ) ≥ 2N/V ≥ (NV )1−θ.
Clearly, this is a polynomial -- time reduction. We show that the values of I and I ′ are equal. If a solution
{xu} for I has value t then the solution defined by xa = xu if a ∈ Vu for I ′ has also value t. On the
other hand, given a solution {xa} for I ′ of value t, we construct a solution for I as follows: for every u, we
choose a random vertex a from Vu and let xu = xa. Then the expected value of this solution is t. Therefore,
the value of the optimal solution for I is at least t.
4 Adversarial Edges, Random Constraints
Theorem 4.1. There exists a polynomial-time approximation algorithm that given k ∈ N (k ≥ k0), ε ∈
(0, 1), η ∈ (cpε log k/k, 1) and a semi-random instance of unique games from the "adversarial edges,
15
random constraints" model on graph G = (V, E) with at least Cη−2n log k edges (C is a sufficiently large
absolute constant) finds a solution of value 1 − O(ε + η), with probability 1 − o(1).
Our algorithm proceeds in several steps. First, it solves the standard SDP relaxation for Unique Games.
Then it removes "η -- long edges" (see below) with respect to the SDP solution, and finally it runs the
CMMa [8] algorithm to solve the unique game on the remaining graph (the CMMa algorithm will again
solve the SDP relaxation for Unique Games -- it cannot reuse our SDP solution).
Definition 4.2. We say that an edge (u, v) is η -- long with respect to an SDP solution {ui}, if
1
2 Xi∈[k]
kui − vπuv(i)k2 > η.
Otherwise, we say that the edge (u, v) is η -- short.
Now we formally present the algorithm.
Input: An instance of unique games.
Output: An assignment of labels to the vertices.
1. Solve the SDP and obtain an SDP solution {u∗
i }.
2. Remove all 1/16 -- long (with respect to {u∗
graph G∗.
i }) edges (u, v) ∈ E from the graph G. Denote the new
3. Solve the SDP on the graph G∗ and run the CMMa algorithm.
In Theorem 4.4, we show that after removing all 1/16 -- long edges from the graph G, the unique games
instance contains at most O(γ)E corrupted constraint w.h.p, where γ = η2/ log k. Since the value of the
optimal SDP is at least ε, the algorithm removes at most 16ε edges at step 2. In the remaining graph, G′, the
CMMa algorithm finds an assignment satisfying 1 − O(√γ log k) = 1 − O(η) fraction of all constraints.
This assignment satisfies at least 1 − O(ε + η) fraction of all constraints in G.
Remark 4.3. In the previous section we proved that a typical instance of unique games in the "random
edges, adversarial constraints" model contains many "super short" edges of the label-extended graph.
Then we showed how we can find an assignment satisfying many super short edges. Note, that edges in the
set Eε are not necessarily short or long. In this section, we show something very different: in the typical
instance of unique games in the "adversarial edges, random constraints" model, most edges in the set Eε
are long. However, note that the label-extended graph does not have to have any super short edges at all.
4.1 Almost All Corrupted Edges are Long
Theorem 4.4. Let k ∈ N (k ≥ k0), ε ∈ (0, 1], γ ∈ (cε/√k, 1/ log k). Consider a graph G = (V, E) with
at least Cγ−1n edges and a unique game instance on G (c, C, and k0 are absolute constants). Suppose that
all constraints for edges in Eε are chosen at random; where Eε is a set of edges of size εE. Then, the set
Eε contains less than γE 1/16 -- short edges w.r.t. every SDP solution {ui} with probability 1 − o(1).
16
Proof. Consider a semi-random instance of unique games. Let {ui} be an SDP solution. Suppose, that at
least γE edges in Eε is 1/16 -- short. Define,
fα,{ui}(u, i, v) =(1,
0,
if kui − vπuv(i)k2 < α(kuik2 + kvπuv(i)k2);
otherwise.
We shall omit the second index of f , when it is clear that we are talking about the set {ui}. For every short
1/16 -- short edge (u, v) ∈ Eε, we have
kuik2 + kvjk2
(6)
1
1
8
× (1 − f1/8(u, i, v)) ≤
kui − vjk2 ≤
1
16
,
2 Xi∈[k]
j=πuv(i)
2 Xi∈[k]
j=πuv(i)
and, hence,
1
2 Xi∈[k]
j=πuv(i)
(kuik2 + kvjk2)f1/8(u, i, v) ≥
1
2 Xi∈[k]
j=πuv(i)
(kuik2 + kvjk2) −
1
2
=
1
2
.
(7)
We get (from (7) and the assumption that there are at least γE 1/16 -- short edges in Eε)
X(u,i)∈V ×[k] Xv:(u,v)∈Eε
kuik2f1/8(u, i, v) = X(u,v)∈Eε Xi∈[k]
(kuik2 + kvjk2)f1/8(u, i, v) ≥ γE.
=
γ
2E +
j=πuv(i)
γ
8 X(u,i)∈V ×[k]
kuik2(ε−1 degEε(u) + ∆).
Here degEε(u) denotes the number of edges in Eε incident to u; and ∆ denotes the average degree in
the graph G. We used that Pi∈[k] kuik2 = 1. We now group all vertices of the label -- extended graph
(u, i) ∈ E×[k] depending on the value of log2(kuik2). Specifically, we pick a random r ∈ (1, 2) distributed
with density (x ln 2)−1 and for every t ∈ N, define
Vt,r = {(u, i) ∈ V × [k] : kuik2 ∈ [r2−(t+1), r2−t]}.
We define V∞,r = {(u, i) ∈ V × [k] : kuik2 = 0}. Let gr(u, i, v) = 1, if (u, i) and (v, πuv(i)) lie in the
same group Vt,r; and gr(u, i, v) = 0, otherwise. We claim, that if f1/8(u, i, v) = 1, then Prr(gr(u, i, v) =
1) ≥ 1/2.
Claim 4.5. If f1/8(u, i, v) = 1, then Prr(gr(u, i, v) = 1) ≥ 1/2.
Proof. Denote j = πuv(i). If f1/8(u, i, v) = 1, then ui 6= 0 and vj 6= 0. Assume w.l.o.g. that kuik2 ≥
kvjk2. Note, that gr(u, i, v) = 0, if and only if r2−t ∈ [kvjk2,kuik2] for some t ∈ N. The probability of
this event is bounded by ln(kuik2/kvjk2)/ ln 2, since the probability that an interval [x, x + ∆x] contains a
point from {r2−t : t ∈ N} is ∆x/(x ln(2)) + o(1). Using the inequality
kuik2 − kvjk2 ≤ kui − vjk2 ≤ kuik2 + kvjk2
8
,
we get kuik2/kvjk2 ≤ 9/7. Thus,
Pr
r
(gr(u, i, v) = 0) ≤ ln(9/7)/ ln 2 < 1/2.
17
Therefore, for some r ∈ [1, 2],
X(u,i)∈V ×[k] Xv:(u,v)∈Eε
kuik2f1/8(u, i, v)gr(u, i, v) ≥
Rewrite this inequality as follows
γ
4E +
γ
16 X(u,i)∈V ×[k]
kuik2(ε−1 degEε(u) + ∆).
Xt∈N X(u,i)∈Vt,r×[k] Xv:(u,v)∈Eε
kuik2f1/8(u, i, v)gr (u, i, v) ≥
2−(t+1) γ
Xt∈N
4E +Xt∈N X(u,i)∈Vt,r
kuik2(ε−1 degEε(u) + ∆).
kuik2f1/8(u, i, v)gr (u, i, v) ≥
For some t ∈ N,
X(u,i)∈Vt,r Xv:(u,v)∈Eε
We now replace each term kuik2 in the left hand side with the upper bound r2−t and each term kuik2 in the
right hand side with the lower bound r2−(t+1). Then, we divide both sides by r2−t ≤ 2−t.
γ
ε−1 degEε(u) +
32Vt,r∆.
kuik2(ε−1 degEε(u) + ∆).
f1/8(u, i, v)gr(u, i, v) ≥
16 X(u,i)∈Vt,r
2−(t+1)γE
γ
16E +
+
γ
γ
4
X(u,i)∈Vt,r Xv:(u,v)∈Eε
32 X(u,i)∈Vt,r
Define a new set of vectors u∗
1, then fα,{u∗
i }(u, i, v) = 1 since for j = πuv(i),
i = ui/kuik, if (u, i) ∈ Vt,r; u∗
i = 0, otherwise. Observe, that if fα(u, i, v)gr(u, i, v) =
ku∗
i − v∗
jk2 = 2 − 2hu∗
i , v∗
ji = 2 − 2 hui, vji
kuik kvjk
kuik kvjk
kuik kvjk ≤ 2 − (1 − α) · 2 = 2α = α(ku∗
= 2 − kuik2 + kvjk2 − kui − vjk2
i k2 + kv∗
jk2).
< 2 − (1 − α)kuik2 + kvjk2
Therefore,
f1/8,{u∗
i }(u, i, v) ≥
X(u,i)∈Vt,r Xv:(u,v)∈Eε
We now embed vectors {u∗
i } in a net N of size exp(O(log k)) using a randomized mapping ϕ (see Lemma 3.7,
a variant of the Johnson -- Lindenstrauss lemma), so that for some small absolute constant β and every
(u, i), (v, j) ∈ Vt,r,
32 X(u,i)∈Vt,r
ε−1 degEε(u) +
(8)
γ
γ
16E +
γ
32Vt,r∆.
Pr((1 + β)−1ku − vk2 − β ≤ kϕ(u) − ϕ(v)k2 ≤ (1 + β)ku − vk2 + β) ≥ 1 −
β
k2 .
i′ ) − ϕ(u∗
i′′)k2 ≥ 2 − 3β for i′ 6= i′′. I.e., if u is good then all vectors ϕ(ui
We say that a vertex u ∈ V is good if for all i′, i′′ ∈ {i : (u, i) ∈ Vt,r}, the following inequality holds:
∗) are almost orthogonal.
kϕ(u∗
By Lemma 3.7, vertex u is good with probability at least 1 − β. We let u∗∗
i ), if u is good; and
u∗∗
i = 0, otherwise. To slightly simplify the proof, for every vertex u, we also zero out a random subset of
⌈βk/4⌉ vectors u∗∗
i }(u, i, v) = 1) ≥ 1 − 4β (unless u or
i}(u, i, v) = 1, then Pr(f1/4,{u∗∗
. Thus, if f1/8,{u∗
i = ϕ(u∗
i
18
v are not good vertices; ui or vπuv(i) has been zeroed; or the distance between u∗
distorted by φ). Hence, for some sample {u∗∗
i },
R ≡ X(u,i):ku∗∗
64(cid:16)E + X(u,i)∈Vt,r
i k=1 Xv:(u,v)∈Eε
i }(u, i, v) ≥
f1/4,{u∗∗
γ
i and v∗
πuv(i) is significantly
ε−1 degEε(u) + Vt,r∆(cid:17) ≡ S.
(9)
We denote the left hand side by R. We denote the right hand side by S.
We now fix the set of vectors {u∗∗
i } and a
random set of constraints {πuv} on edges (u, v) ∈ Eε inequality (9) holds. For each (u, v) ∈ Eε and
i }(u, i, v) = 1) ≤ 1/k (since for every i, there is at most one j ∈ [k] such that
i, Pr{πuv}(f1/4,{u∗∗
ku∗∗
i } and estimate the probability that for a given set {u∗∗
j k2 ≤ 1/4; here we use that all non-zero vectors vj are almost orthogonal). Thus,
ER = E X(u,i):ku∗∗
degEε(u) < 64S/(√kc).
i k2=1 Xv:(u,v)∈Eε
k X(u,i):ku∗∗
i }(u, i, v) ≤
f1/4,{u∗∗
i − v∗∗
i k2=1
1
In the last inequality, we used the bound γ ≥ cε/√k and (9). We assume that 64/c ≪ 1. We would like to
apply the Bernstein inequality, which would give us an upper bound of
(10)
ε−1 degEε(u) + γVt,r∆(cid:1) log(k)(cid:17)
exp(−Ω(S)) ≡ exp(cid:16) − Ω(cid:16)(cid:0)γE + γ X(u,i)∈Vt,r
on the probability that inequality (9) holds (note: E[R] ≪ S; log(S/E[R]) ≥ Θ(log k)). Formally, we
cannot use this inequality, since random variables f1/4,{u∗∗
i }(u, i, v) are not independent. So, instead, we
consider a random process, where we consequently pick edges (u, v) ∈ Eε and then for every i such that
u∗∗
6= 0 pick a random yet unchosen j ∈ [k] and set πuv(i) = j. For every u and i, there are at least ⌈βk/4⌉
i
different candidates j (since at least ⌈βk/4⌉ ui's equal 0) and for at most one of them ku∗∗
j k2 ≤ 1/4.
i }(u, i, v) = 1) ≤ 4/(βk). We now apply a martingale concentration
Hence, at every step Pr(f1/4,{u∗∗
inequality and get the same bound (10).
i } and then apply the union bound to show
that w.h.p.
the set Eε contains less than γE 1/16-short edges with respect to every SDP solution. Let
W = Vt,r be the number of pairs (u, i) such that u∗∗
6= 0. We choose W variables among {u∗∗
i }u,i and
assign them values from N as follows. Let au = {u∗∗
6= 0 : i}. First, we choose the values of au for all
vertices (note that Pu∈V au = W ). The number of ways to choose {au} is at most the number of ways
to write W as the sum of n numbers, which is (cid:0)n+W −1
n−1 (cid:1) < 2n+W . Then for each vertex u, we choose
iau 6= 0. The number of ways to choose labels is at most
Qu∈V kau = kW . Finally, we assign a vector from N to each chosen variable u∗∗
. The number of ways to
do so is NW . Thus there are at most
Finally, we bound the number of possible solutions {u∗∗
au labels i1, . . . , iau for which u∗∗
i1 6= 0, . . . , u∗∗
i − v∗∗
i
i
i
2n+W × NW × kW = exp(O(W log k + n))
ways to choose vectors {u∗∗
i } for given W . Since γVt,r∆ log k ≥ γW × (Cγ−1) log k ≥ CW log k
and γE log k ≥ Cn inequality (9) holds for some collection {u∗∗} with exponentially small probability
exp(−Ω(n)) for every fixed value of W . Since there are only n possible values on W , inequality (9) holds
for some value of W with exponentially small probability.
19
5 Random Corrupted Constraints in Linear Unique Games
In this section we study a special family of unique games, so-called Linear Unique Games (or MAX Γ-
Lin). In Linear Unique Games all labels i ∈ [k] are elements of the cyclic group Z/kZ; all constraints
have the form xu = xv − suv, where suv ∈ Z/kZ. It is known that if the Unique Games Conjecture holds
for general unique games, then it also holds for this family. We note that the results from the previous
subsection are not directly applicable for Linear Unique Games, since random permutations are restricted
to be shifts πuv : xu 7→ xv + suv and, thus, a random unique game from the "adversarial edges, random
constraints" model is most likely not a linear unique game. This is not the case in the "random edges,
adversarial constraints" model, since there the adversary may choose corrupted constraints in any fashion
she wants, particularly to be shifts. We note that the results of this section can be easily generalized to any
finite Abelian group using the decomposition of the group into a sum of cyclic subgroups. We omit the
details from the conference version of the paper.
For Linear Unique Games, we use a shift invariant variant of the SDP relaxation (see e.g., a paper by
Andersson, Engebretsen, and Hastad [2]). In this relaxation, all vectors ui are unit vectors, satisfying ℓ2
2
triangle inequalities, and an extra constraint kui − vjk = kui+s − vj+sk. The objective is to minimize
1
k ×
1
2E X(u,v)∈E Xi∈[k]
kui − vi+suvk2.
Given an arbitrary SDP solution {ui} one can obtain a uniform solution using symmetrization (this trick
works only for unique games on groups, not arbitrary unique games):
u′
i =
1
√k Mj∈Z/kZ
ui+j.
It is easy to verify (and it is well-known, see e.g., [8]) that vectors {u′
i} satisfy all SDP constraints and the
objective value of the SDP does not change. We adapt the definition of η -- short edges for this variant of SDP
as follows: we say that an edge (u, v) is η -- short with respect to the SDP solution {ui} if 1/2 kui−vi+sk2 ≤
η for every i ∈ [k] (or, equivalently, for some i ∈ [k], since kui − vi+sk = kuj − vj+sk).
We prove the following analog of Theorem 4.4, which immediately implies that our general algorithm
for the adversarial edges, random constraints model works for linear unique games.
Theorem 5.1. Let k ∈ N (k ≥ k0), ε ∈ (0, 1), η ∈ (ε/ log k, 1/ log k). Consider a graph G = (V, E)
with at least Cnγ−1 edges (k0, c, C are absolute constants) and a random linear unique game from the
"adversarial edges, random constraints" model. Let Eε be the set of corrupted edges; and let {ui} be the
optimal SDP solution. Then, with probability 1 − o(1), at most γE edges in Eε are 1/32 -- short.
Proof. As mentioned in the beginning of this section the result we would like to prove holds not only for
cyclic groups but for arbitrary Abelian groups. In fact, the proof is simpler if the group can be decomposed
into a direct sum of many cyclic groups of small size. So, in some sense, our plan is to represent the
cyclic group as a "pseudo-direct sum" of such groups. Before proceeding to the proof, we note that all
transformations described below are only required for the proof and are not performed by the approximation
algorithm.
Let S be a circle (or a one dimensional torus). We identify S with the segment [0, 1] with "glued" ends
0 and 1. Define an embedding θ : [k] → S as follows θ(i) = i/k. We denote the distance between two
adjacent labels by τ = 1/k. Below, all "+" and "−" operations on elements of S are "modulo 1" and all
operations on [k] (and later [T ]) are modulo k (and T respectively).
20
Recall, that in the "adversarial edges, random constraints" model of Linear Unique Games, the adversary
first chooses a set of edges Eε, then the "nature" randomly picks a shift suv ∈ [k] for every (u, v) ∈ Eε.
We assume that the nature picks suv in the following way: it chooses a random suv ∈ S, and then defines
suv = ⌊suvk⌋. Then, clearly, suv is distributed uniformly in [k].
If M is a subset of S, then M + s denotes the shift of M by s. We denote by L(M ) the set of labels
covered by M:
We let Ψu(M ) =Pi∈L(M ) ui, and Ψ′
Fix T = R2 (where R ∈ N) to be a sufficiently large absolute constant (not depending on k or any other
parameters). We represent each s ∈ S as
∞
L(M ) = {i ∈ [k] : θ(i) ∈ M}.
u(M ) = Ψu(M )/kΨu(M )k.
s =
sp/T p,
Xp=1
where every sp belongs to [T ], in other words, 0.s1s2s3 . . . is a representation of s in the numeral system
with base T . (This representation is not unique for a set of measure 0. If it is not unique, we pick the
representation that ends in zeros.) We define Dp(s) = sp. Note, that if s is uniformly distributed random
variable in S, then Dp(s) is uniformly distributed in [T ] and all Dp(s) are independent for different p. We
are interested in the first P = ⌈2 ln ln k⌉ digits of suv. We have chosen P so that T −P = log−Θ(1) k >
τ R ≡ R/k, but eP ≥ ln2 k.
For every integer p ∈ {1, . . . , P} and d ∈ [T ] define
Mp,d = {s ∈ S : Dp(s) ∈ [d, d + R − 1]},
that is, Mp,d is the set of numbers in [0, 1], whose p-th digit is in the range [d, d + R − 1]. Note, that the set
{s ∈ S : Dp(s) = d} is a union of segments of length T −P > τ R ≡ R/k (here and below, all segments are
left-closed and right-open).
Lemma 5.2. Let M ⊂ S be a union of segments of length at least Rτ and let s ∈ [0, 1]. Then,
(1 − 1/R) µ(M )/τ ≤ L(M ) ≤ (1 + 1/R) µ(M )/τ.
and
L(M ) △ L(M + s) ≤ min(L(M ),L(M + s))/R × (2s/τ + 2),
here △ denotes the symmetric difference of two sets; and µ is the uniform measure on [0, 1].
Proof. Let M be the disjoint union of segments I ∈ I of length at least Rτ each. Consider one of the
segments I ⊂ M. This segment covers at least ⌊µ(I)/τ⌋ ≥ (µ(I)/τ − 1) and at most ⌈µ(I)/τ⌉ ≤
(µ(I)/τ +1) points θ(i). In other words, µ(I)/τ−1 ≤ L(I) ≤ µ(I)/τ +1. Observe that 1/R·µ(I)/τ ≥ 1,
because µ(I) ≥ Rτ , thus
(1 − 1/R)µ(I)/τ ≤ L(I) ≤ (1 + 1/R)µ(I)/τ.
Using equalities L(M ) =PI∈I L(I) and µ(M ) =PI∈I µ(I), we get (1− 1/R)µ(M )/τ ≤ L(M ) ≤
(1 + 1/R)µ(M )/τ .
To prove the second inequality, observe that
L(M ) △ L(M + s) ⊂ [I∈I
(L(I) △ L(I + s)).
21
For each I ∈ I, L(I)△L(I+s) = L(I△(I+s)). The set I△(I+s) is the union of two segments of length at
most s, thus I△(I +s) covers at most 2(s/τ +1) points θ(i) (i ∈ [k]). Every interval covers at least R points.
Thus, the size of the family I is at most L(I)/R. Therefore, L(M )△L(M + s) ≤ L(M )/R×(2s/τ +2).
The same argument shows L(M ) △ L(M + s) ≤ L(M + s)/R × (2s/τ + 2).
Lemma 5.3. Let suv = ⌊ksuv⌋ (suv ∈ [k], suv ∈ S), p ∈ {1, . . . , P} and let d = Dp(suv) be the p-th digit
of suv. Then L(Mp,0),L(Mp,d) ∈ [(1 − 1/R)/(Rτ ), (1 + 1/R)/(Rτ )] and
L(Mp,0) △ (L(Mp,d) − suv) ≤ 8 min(L(Mp,0),L(Mp,d))/R.
Proof. We have µ(Mp,d) = R/T = 1/R, thus by Lemma 5.2, L(Mp,0) ∈ [(1 − 1/R)/(Rτ ), (1 +
1/R)/(Rτ )]. Write,
L(Mp,0) △ (L(Mp,d) − suv) = (L(Mp,0) + suv) △ L(Mp,d) = L(Mp,0 + θ(suv)) △ L(Mp,d)
≤ L(Mp,0 + θ(suv)) △ L(Mp,0 + suv) + L(Mp,0 + suv) △ L(Mp,d).
Since suv − θ(suv) ∈ [0, τ ], by Lemma 5.2,
L(Mp,0 + θ(suv))△ L(Mp,0 + suv) ≤ L(Mp,0 + θ(suv))/R· (2(suv − θ(suv))/τ + 2) ≤ 4L(Mp,0)/R.
The p-th digit of suv is Dp(suv) = d. Hence, the p-th digit of every number s in Mp,0 + suv is in the range
[d, (R − 1) + d + 1] = [d, R + d]. Moreover, all numbers with p-th digit in the interval [d + 1, (R − 1) + d]
are covered by Mp,0 + suv. Thus, (Mp,0 + suv) △ Mp,d ⊂ {s : Dp(s) ∈ {d, d + R}}. The measure of the
set {s : Dp(s) ∈ {d, d + R}} is 2/T . It is a union of segments of lengths T −p ≥ τ R. By Lemma 5.2,
L((Mp,0 + suv) △ Mp,d) ⊂ L({s : Dp(s) ∈ {d, d + R}}) ≤ (1 + 1/R) · 2/(T τ ) ≤ 4L(Mp,0)/R.
Lemma 5.4. Let suv = ⌊ksuv⌋ (suv ∈ [k], suv ∈ S), p ∈ {1, . . . , P}, and let d = Dp(suv) be the p-th
digit of suv. Suppose that an edge (u, v) ∈ Eε is 1/32 -- short, then
kΨu(Mp,0) − Ψv(Mp,d)k2 ≤ 1/8 min(kΨu(Mp,0)k2,kΨv(Mp,Dp(s))k2);
kΨ′
u(Mp,0) − Ψ′
v(Mp,d)k2 ≤ 1/8.
(11)
(12)
and
Proof. Write,
ui − Xi∈L(Mp,d)
kΨu(Mp,0) − Ψv(Mp,d)k2 = (cid:13)(cid:13) Xi∈L(Mp,0)
X
ui −
i∈L(Mp,0)\(L(Mp,d)−suv)
vi(cid:13)(cid:13)
2 =(cid:13)(cid:13)(cid:13)
X
X
i∈(L(Mp,d)−suv)\L(Mp,0)
2 triangle inequalities hui−vi+suv, uj−vj+suvi = −hui, vj+suvi−hvi+suv , uji ≤ 0, hui,−vj+suvi ≤
Using ℓ2
0 (for i 6= j) and then Lemma 5.3, we get (for sufficiently large R)
kΨu(Mp,0) − Ψv(Mp,d)k2 ≤
i∈L(Mp,0)∩(L(Mp,d)−suv)
kui − vi+suvk2 + L(Mp,0) △ (L(Mp,d) − suv)
≤ 1/16 L(Mp,0) + 8L(Mp,0)/R ≤ 1/8 L(Mp,0) = kΨu(Mp,0)k2/8.
X
i∈L(Mp,0)∩(L(Mp,d)−suv)
(ui − vi+suv) +
2
.
vi+suv(cid:13)(cid:13)(cid:13)
22
Similarly, kΨu(Mp,0) − Ψv(Mp,d)k2 ≤ kΨu(Mp,d)k2/8.
Inequality (12) immediately follows from inequality (11): let ψu = Ψu(Mp,0), ψv = Ψv(Mp,d) and
assume kψuk ≤ kψvk, then k ψu/kψuk − ψv/kψuk k2 ≤ 1/8. Vector ψu/kψuk has length 1, and vector
ψv/kψuk has length at least 1, hence k ψu/kψuk − ψv/kψvk k2 ≤ k ψu/kψuk − ψv/kψuk k2 ≤ 1/8.
Observe, that vectors Ψ′
v(Mp,d′), Ψ′
kΨ′
v(Mp,d′′) are orthogonal if d′′ − d′ > R, and thus
u(Mp,d′) − Ψ′
u(Mp,d′′)k2 = 2.
We now proceed the same way as in the proof of Theorem 4.4. We embed O(n log log k) vectors
Ψ′
u(Mp,d) (u ∈ V , p ∈ P , d ∈ T ) in a net N of size O(1) using a randomized mapping ϕ (see Lemma 3.7),
so that for some small absolute constant β and every u, v ∈ V ; d′, d′′ ∈ T ; and p ∈ {1, . . . , P},
Pr((1 + β)−1kΨ′
u(Mp,d′) − Ψ′
v(Mp,d′′)k2 − β ≤ kΦ(u, p, d′) − Φ(v, p, d′′)k2
v(Mp,d′′)k2 + β) ≥ 1 − β/T 2,
where Φ(u, p, d) = ϕ(Ψ′
u(Mp,d)). We say that a pair (u, p) ∈ V × {1, . . . , P} is good if the following
inequality holds: kΦ(u, p, d′) − Φ(u, p, d′′)k2 ≥ 2 − 3β for all d′, d′′ ∈ [T ] such that d′′ − d′ > R.
By Lemma 3.7, a pair (u, p) is good with probability at least 1 − β. Then, for a fixed 1/32 -- short edge
(u, v) ∈ Eε, the expected fraction of p's for which both pairs (u, p) and (v, p) are good and
u(Mp,d′) − Ψ′
≤ (1 + β)kΨ′
kΦ(u, p, 0) − Φ(v, p, Dp(suv))k2 ≤ 1/4
(13)
is at least 1 − 3β.
Assume that γE = γε−1Eε edges in Eε are 1/32 -- short. We say that an edge (u, v) ∈ Eε is good
with respect to the set {Φ(u′, p, d)u,p,d} if the following statement holds: "for at least (1 − 6β) fraction of
p's in {1, . . . , P}, the pairs (u, p) and (v, p) are good, and inequality (13) holds". By the Markov inequality,
there exists a realization of random variables Φ(u, p, d) (random with respect to a random embedding in the
net N) such that for at least γ/2 fraction of edges (u, v) ∈ Eε the previous statement holds.
Thus we have shown that for every linear unique game with γE 1/32 -- short edges there always exists
a witness -- a collection of vectors {Φ(u, p, d)} ⊂ N, such that at least γE/2 edges in Eε are good with
respect to this collection (Remark: so far we have not used that the instance is semi-random). Now, we
will prove that for a fixed witness, the probability that γE/2 edges in Eε is good in a semi-random unique
game is exponentially small.
Fix an edge (u, v) ∈ Eε and compute the probability that it is good with respect to a fixed witness
{Φ(u, p, d)} ⊂ N. The probability that
kΦ(u, p, 0) − Φ(v, p, Dp(suv))k2 ≤ 1/4
(14)
for a random (suv) is at most 1/R if pairs (u, p) and (v, p) are good, since among every R values d ∈
{d0, d0 + R, d0 + 2R,··· } there is at most one d satisfying
kΦ(u, p, 0) − Φ(v, p, d)k2 ≤ 1/4.
(Recall, that kΦ(v, p, d′)− Φ(v, p, d′′)k2 ≥ 2− 3β if d′ − d′′ > R). By the Chernoff bound the probability
that for (1 − 6β) fraction of p's the inequality (14) is satisfied and (u, p), (v, p) are good is at most
e−P ln((1−6β)R) ≤ e−P ≤ ln−2 k.
23
Hence, the expected number of good edges in Eε is at most Eε/ ln2 k = εE/ ln2 k, and the probability
that γE/2 edges in Eε are good is at most
exp(cid:0)−γE/2 · ln(cid:0)γε−1 ln2 k/2(cid:1)(cid:1) ≤ exp(−Cn/2 · ln(1/2 ln k)).
The total number of possible witnesses {Φ(u, p, d)} ⊂ N is exp(O(n log log k)). So by the union
bound (for sufficiently large absolute constant C) with probability 1 − exp(−n) = 1 − o(1), less than γE
edges in Eε are 1/32 -- short.
6 Random Initial Constraints
In this section, we consider the model, in which the initial set of constraints is chosen at random and other
steps are controlled by the adversary. Specifically, in this model the adversary chooses the constraint graph
G = (V, E) and a "planted solution" {xu}. Then for every edge (u, v) ∈ E, she randomly chooses a
permutation (constraint) πuv such that πuv(xu) = xv (each of (k − 1)! possible permutations is chosen with
the same probability 1/(k − 1)!; choices for different edges are independent). Then the adversary chooses
an arbitrary set Eε of edges of size at most εE and (adversarially) changes the corresponding constraints:
replaces constraint πuv with a constraint π′
uv for (u, v) ∈ Eε. Note that the obtained semi-random instance
is 1 − ε satisfiable since the "planted solution" xu satisfies constraints for edges in E \ Eε. The analysis of
this model is much simpler than the analysis of the other two models that we study.
Theorem 6.1. There exists a polynomial-time algorithm that given k ∈ N (k ≥ k0), ε ∈ (0, 1), η ∈
(c log k/√k, 1) and a semi-random instance of unique games from the "random initial instance" model
on graph G = (V, E) with at least Cη−1n log k edges finds a solution of value 1 − O(ε + η/ log k) with
probability 1 − o(1) (where c, C and k0 are some absolute constants).
Proof. We solve the standard SDP relaxation for the problem. Then we use a very simple rounding proce-
dure. For every vertex u, if kuik2 > 1/2 for some label i, we label u with i; otherwise, we label u with an
arbitrary label (since for every two labels i1 6= i2, kui1k2 +kui2k2 ≤ 1, we do not label any vertex with two
labels).
We now show that our labeling satisfies a 1 − O(ε + η/ log k) fraction of constraints w.h.p. Without
loss of generality, we assume that the planted solution is xu = 0 for every u ∈ V . For every t ∈ (1/2; 3/4),
let St = (cid:8)u : ku0k2 ≥ t(cid:9). Let t0 be the value of t that minimizes the size of the cut between St and
V \ St; let S = St0. Note that if u ∈ S then we label vertex u with 0. Therefore, our labeling satisfies all
constraints πuv for edges (u, v) within S (but not necessarily constraints π′
uv). We conservatively assume
that constraints for all edges from S to V \ S and edges within V \ S are not satisfied. We now estimate
their number. First, we bound the number of edges leaving S. Note that since the unique game instance is
1 − ε satisfiable the cost of the SDP solution is at most ε. In particular,
1
2 X(u,v)∈E(cid:12)(cid:12)(cid:12)ku0k2 − kv0k2(cid:12)(cid:12)(cid:12) ≤
1
2 X(u,v)∈E
and v /∈ St or u /∈ St and v ∈ St is at most 4(cid:12)(cid:12)(cid:12)ku0k2 − kv0k2(cid:12)(cid:12)(cid:12)
On the other hand, if we choose t uniformly at random from (1/2; 3/4), then the probability that u ∈ St
for every (u, v) ∈ E. Therefore, the expected
size of the cut between St and V \ St is at most 8εE. Hence the size of the cut between S and V \ S is at
most 8εE.
ku0 − v0k2 ≤ ε.
24
Now we estimate the number of edges within V \ S. We consider a new instance of Unique Games on
G with the label set {1, . . . , k − 1} and constraints π∗
uv = πuv{1,...,k−1} (the restriction of πuv to the set
{1, . . . , k − 1}). Note that each π∗
uv is a permutation of {1, . . . , k − 1} since πuv(0) = 0. Moreover, each
π∗
uv is a random permutation uniformly chosen among all permutations on {1, . . . , k − 1}. For each vertex
u and label i ∈ {1, . . . , k− 1}, we define a vector u∗
i = ui. Otherwise, we let
1 = e and u∗
u∗
i = 0 for i > 1, where e is a fixed unit vector orthogonal to all vectors vj in the SDP solution.
i } is a relaxed SDP solution if it satisfies all SDP conditions except possibly for the
i k2 = 1 for every vertex u. We require instead that 1/4 ≤ Pi ku∗
i k2 ≤ 1. Note that
i } is a relaxed SDP solution since for every u /∈ S, Pk−1
i k2 = 1 − ku0k2 ≥
i=1 ku∗
We say that {u∗
condition that Pi ku∗
the set of vectors {u∗
1 − t0 ≥ 1/4; for every u ∈ S,Pk−1
i k2 = kek2 = 1.
We now use a slightly modified version of Theorem 4.4.
i as follows. If u /∈ S, we let u∗
i=1 ku∗
Lemma 6.2. Consider a unique game on a graph G = (V, E) with at least Cη−1n log k edges with random
set of constraints π∗
i } be
a relaxed SDP solution. Then, there are at most O(η/ log k)E 1/64 -- short edges with probability 1− o(1).
The proof of the lemma almost exactly repeats the proof of Theorem 4.4 for ε = 1 (we only need to
uv, where η ∈ (c log k/√k, 1) (where c and C are some absolute constants). Let {u∗
change inequalities (6) and (7) slightly).
We apply this lemma to the solution {u∗
i }. We get that there are at most O(η/ log k)E 1/64 -- short
ku∗
i − v∗
uv(i)k2
π∗
edges. In particular, there are at most O(η/ log k)E 1/64 -- short edges in E(V \ S). Thus
1
Xi=1
2 X(u,v)∈E
2 X(u,v)∈E(V \S)
kui − vπuv(i)k2 =
kui − vπuv(i)k2 ≥
Xi=1
k−1
k−1
k−1
1
1
2 X(u,v)∈E(V \S)
≥ E(V \ S)
Xi=1
− O(η/ log k)E.
64
However, the left hand side is at most εE. Therefore, E(V \ S) = O(ε + η/ log k)E.
We conclude that the solution that our algorithm finds satisfies a 1 − O(ε + η/ log k) fraction of con-
straints πuv. Since there are at most εE corrupted constraints, the solution also satisfies a 1 − O(ε +
η/ log k) − ε = 1 − O(ε + η/ log k) fraction of corrupted constraints.
7 Distinguishing Between Semi-Random Unsatisfiable Games and Almost
Satisfiable Games
In this section, we study the following question, Is it possible to distinguish between (1 − ε) satisfiable
semi-random games and (1 − δ) satisfiable (non-random) games if δ ≪ ε? This question is interesting only
in the model where the corrupted constraints are chosen at random (i.e. step 4 is random), since in the other
two semi-random models (when the initial constraints are random, and when the set of corrupted edges Eε
is random), the semi-random instance can be 1 − δ satisfiable, therefore, the answer is trivially negative.
Specifically, we consider the following model. The adversary chooses a constraint graph G and a set
of constraints πuv. We do not require that this instance is completely satisfiable. Then she chooses a set of
edges Eε of size εE. She replaces constraints for edges in Eε with random constraints (each constraint is
chosen uniformly at random among all k! possible constraints).
We show that such semi-random instance can be distinguished w.h.p. from a (1 − δ) satisfiable instance
if δ < cε (where c is an absolute constant) if E ≥ Cn max(ε−1, log k). To this end, we consider the
25
standard SDP relaxation for Unique Games. We prove that the SDP value of a semi-random instance is at
least cε; whereas, of course, the SDP value of a (1 − δ) satisfiable instance is at most δ.
Theorem 7.1. Let k ∈ N (k ≥ k0) and ε ∈ (0, 1]. Consider a graph G with at least Cn max(ε−1, log k)
edges, and a semi-random unique games instance I on G with εE randomly corrupted constraints (k0 and
C are absolute constants). Then the SDP value of I is at least ε/32 with probability 1 − o(1).
Proof. We apply Theorem 4.4 to our instance of Unique Games, with γ = min(ε/2, 1/(2 log k)). We
get that at least half of all edges in Eε are 1/16 -- long w.h.p. The contribution of these edges to the sum
1
the SDP value is at least ε/32.
2P(u,v)Pi kui−vπuv(i)k2 in the SDP objective function is at least 1/16×(Eε/2) = εE/32. Therefore,
References
[1] M. Alekhnovich. More on Average Case vs Approximation Complexity. In Proceedings of the 44st
IEEE Symposium on Foundations of Computer Science: pp, 298 -- 307, 2003.
[2] G. Andersson, L. Engebretsen, and J. Hastad. A new way to use semidefinite programming with
applications to linear equations mod p, Journal of Algorithms, Vol. 39, 2001, pp. 162 -- 204.
[3] S. Arora, B. Barak, and D. Steurer. Subexponential Algorithms for Unique Games and Related prob-
lems. In Proceedings of the 51st IEEE Symposium on Foundations of Computer Science, 2010.
[4] S. Arora, R. Impagliazzo, W. Matthews, and D. Stuerer. Improved algorithms for unique games via
divide and conquer, In Electronic Colloquium on Computational Complexity, TR10-041, 2010.
[5] S. Arora, S. Khot, A. Kolla, D. Steurer, M. Tulsiani, and N. Vishnoi. Unique games on expanding
constraint graphs are easy. In Proceedings of the 40th ACM Symposium on Theory of Computing,
pp. 21 -- 28, 2008.
[6] B. Barak, M. Hardt, I. Haviv, A. Rao, O. Regev and D. Steurer. Rounding Parallel Repetitions of
Unique Games, In Proceedings of the 49th IEEE Symposium on Foundations of Computer Science,
pp. 374 -- 383, 2008.
[7] A. Blum and J. Spencer. Coloring Random and Semi-Random k-Colorable Graphs, J. Algorithms,
vol. 19, no. 2, pp. 204 -- 234, 1995.
[8] M. Charikar, K. Makarychev, and Y. Makarychev. Near-Optimal Algorithms for Unique Games. In
Proceedings of the 38th ACM Symposium on Theory of Computing, pp. 205 -- 214, 2006.
[9] E. Chlamtac, K. Makarychev, and Y. Makarychev. How to Play Unique Games Using Embeddings. In
Proceedings of the 47th IEEE Symposium on Foundations of Computer Science, pp. 687 -- 696, 2006.
[10] A. Gupta and K. Talwar. Approximating Unique Games. In Proceedings of the 17th ACM-SIAM
Symposium on Discrete Algorithms, pp. 99 -- 106, 2006.
[11] U. Feige. Relations Between Average Case Complexity and Approximation Complexity In Proceedings
of the 34th Annual ACM Symposium on Theory of Computing, pp. 534 -- 543, 2002.
26
[12] U. Feige and J. Kilian, Heuristics for Semirandom Graph Problems,
Journal of Computing and
System Sciences, vol. 63, pp. 639 -- 671, 2001.
[13] U. Feige and R. Krauthgamer. Finding and Certifying a Large Hidden Clique in a Semi-Random
Graph Random Structures and Algorithms, vol. 16(2), pp. 195 -- 208, 2000.
[14] V. Guruswami, R. Manokaran, and P. Raghavendra. Beating the Random Ordering is Hard: In-
In Proceedings of the 49th IEEE Symposium on
approximability of Maximum Acyclic Subgraph.
Foundations of Computer Science, pp. 573 -- 582, 2008.
[15] V. Guruswami and P. Raghavendra. Constraint satisfaction over a non-boolean domain: Approxima-
tion algorithms and unique-games hardness. In Proceedings of APPROX-RANDOM, 77 -- 90, 2008.
[16] M. Jerrum. Large Cliques Elude the Metropolis Process, Random Structures and Algorithm, vol. 3
(4), pp. 347 -- 359, 1992.
[17] S. Khot. On the power of unique 2-prover 1-round games. In Proceedings of the 34th ACM Sympo-
sium on Theory of Computing, pp. 767 -- 775, 2002.
[18] S. Khot, G. Kindler, E. Mossel, and R. O'Donnell. Optimal inapproximability results for MAX-CUT
and other 2-variable CSPs? ECCC Report TR05-101, 2005.
[19] S. Khot and O. Regev. Vertex cover might be hard to approximate to within 2 − ε. In Proceedings of
the 18th IEEE Annual Conference on Computational Complexity, 2003.
[20] S. Khot and N. Vishnoi. The Unique Games Conjecture, Integrality Gap for Cut Problems and
Embeddability of Negative Type Metrics into ℓ1, Foundations of Computer Science, pp. 53 -- 62, 2005.
[21] A. Kolla. Spectral Algorithms for Unique Games In Proceedings of the Conference on Computational
Complexity, pp. 122 -- 130, 2010.
[22] K. Makarychev and Y. Makarychev. How to Play Unique Games on Expanders, In Proceedings of
the eighth Workshop on Approximation and Online Algorithms, 2010.
[23] P. Raghavendra, Optimal algorithms and inapproximability results for every CSP?, Proceedings of
the 40th Annual ACM Symposium on Theory of Computing, pp. 245 -- 254, 2008.
[24] P. Raghavendra, D. Steurer, How to Round Any CSP, In Proceedings of the 50th IEEE Symposium
on Foundations of Computer Science, pp. 586 -- 594, 2009.
[25] A. Samorodnitsky and L. Trevisan. Gowers uniformity, influence of variables, and PCPs. In Proceed-
ings of the 38th annual ACM symposium on Theory of computing, pp. 11 -- 20, 2006.
[26] L. Trevisan. Approximation Algorithms for Unique Games. In Proceedings of the 46th IEEE Sympo-
sium on Foundations of Computer Science, pp. 197 -- 205, 2005.
27
|
1602.05622 | 1 | 1602 | 2016-02-17T22:56:59 | Compact Flow Diagrams for State Sequences | [
"cs.DS",
"cs.CC"
] | We introduce the concept of compactly representing a large number of state sequences, e.g., sequences of activities, as a flow diagram. We argue that the flow diagram representation gives an intuitive summary that allows the user to detect patterns among large sets of state sequences. Simplified, our aim is to generate a small flow diagram that models the flow of states of all the state sequences given as input. For a small number of state sequences we present efficient algorithms to compute a minimal flow diagram. For a large number of state sequences we show that it is unlikely that efficient algorithms exist. More specifically, the problem is W[1]-hard if the number of state sequences is taken as a parameter. We thus introduce several heuristics for this problem. We argue about the usefulness of the flow diagram by applying the algorithms to two problems in sports analysis. We evaluate the performance of our algorithms on a football data set and generated data. | cs.DS | cs |
Compact Flow Diagrams for State Sequences
Kevin Buchin, Maike Buchin, Joachim Gudmundsson,
Michael Horton, and Stef Sijben
Abstract. We introduce the concept of compactly representing a large
number of state sequences, e.g., sequences of activities, as a flow diagram.
We argue that the flow diagram representation gives an intuitive summary
that allows the user to detect patterns among large sets of state sequences.
Simplified, our aim is to generate a small flow diagram that models the
flow of states of all the state sequences given as input. For a small number
of state sequences we present efficient algorithms to compute a minimal
flow diagram. For a large number of state sequences we show that it is
unlikely that efficient algorithms exist. More specifically, the problem is
W [1]-hard if the number of state sequences is taken as a parameter. We
thus introduce several heuristics for this problem. We argue about the
usefulness of the flow diagram by applying the algorithms to two problems
in sports analysis. We evaluate the performance of our algorithms on a
football data set and generated data.
1
Introduction
Sensors are tracking the activity and movement of an increasing number of
objects, generating large data sets in many application domains, such as sports
analysis, traffic analysis and behavioural ecology. This leads to the question
of how large sets of sequences of activities can be represented compactly. We
introduce the concept of representing the "flow" of activities in a compact way
and argue that this is helpful to detect patterns in large sets of state sequences.
To describe the problem we start by giving a simple example. Consider three
objects (people) and their sequences of states, or activities, during a day. The set
of state sequences T = {τ1, τ2, τ3} are shown in Fig. 1(a). As input we are also
given a set of criteria C = {C1, . . . , Ck}, as listed in Fig. 1(b). Each criterion is
a Boolean function on a single subsequence of states, or a set of subsequences
of states. For example, in the given example the criterion C1 ="eating" is true
for Person 1 at time intervals 7 -- 8am and 7 -- 9pm, but false for all other time
intervals. Thus, a criterion partitions a sequence of states into subsequences, called
segments. In each segment the criterion is either true or false. A segmentation of
T is a partition of each sequence in T into true segments, which is represented
by the corresponding sequence of criteria. If a criterion C is true for a set of
subsequences, we say they fulfil C. Possible segments of T according to the set C
are shown in Fig. 1(c). The aim is to summarize segmentations of all sequences
efficiently; that is, build a flow diagram F, starting at a start state s and ending
at an end state t, with a small number of nodes such that for each sequence of
states τi, 1 ≤ i ≤ m, there exists a segmentation according to C which appears
Fig. 1: The input is (a) a set T = {τ1, . . . , τm} of sequences of states and (b)
a set of criteria C = {C1, . . . , , Ck}. (c) The criteria partition the states into a
segmentation. (d) A valid flow diagram for T according to C.
as an s -- t path in F. A possible flow diagram is shown in Fig. 1(d). This flow
diagram for T according to C can be validated by going through a segmentation
of each object while following a path in F from s to t. For example, for Person 1
the s -- t path s → C1 → C2 → C4 → C1 → t is a valid segmentation.
Now we give a formal description of the problem. A flow diagram is a node-
labelled DAG containing a source node s and sink node t, and where all other
nodes are labelled with a criterion. Given a set T of sequences of states and set
of criteria C, the goal is to construct a flow diagram with a minimum number of
nodes, such that a segmentation of each sequence of states in T is represented,
that is, included as an s -- t path, in the flow diagram. Furthermore (when criteria
depend on multiple state sequences, e.g. C7 in Fig. 1) we require that the
segmentations represented in the flow diagram are consistent, i.e. can be jointly
realized. The Flow Diagram problem thus requires the segmentations of each
sequence of states and the minimal flow diagram of the segmentations to be
computed. It can be stated as:
Problem 1. Flow Diagram (FD)
Instance: A set of sequences of states T = {τ1, . . . , τm}, each of length at most
n, a set of criteria C = {C1, . . . , Ck} and an integer λ > 2.
Question: Is there a flow diagram F with ≤ λ nodes, such that for each τi ∈ T ,
there exists a segmentation according to C which appears as an s -- t path in F?
Even the small example above shows that there can be considerable space
savings by representing a set of state sequences as a flow diagram. This is not
a lossless representation and comes at a cost. The flow diagram represents the
sequence of flow between states, however, the information about an individual
sequence of states is lost. As we will argue in Section 4, paths representing many
2
Person1Person2Person38-9amcycletoworkcycletoworkdrivetowork9am-5pmworkworkwork5-7pmstudydinnershop7-9pmdinnershopdinner7-8ambreakfastbreakfastgym(a)(b)sC3C1C4tC2C6C1C6(c)(d)8-9am9am-5pm5-7pm7-9pm7-8amPerson1[C2,C3][C4,C5][C4][C1,C7][C1,C7]Person2[C2][C4,C5][C1][C6][C3]Person3[C2,C3][C4,C5][C6][C1,C7][C1,C7]C1:Eating{breakfast,dinner}C3:Exercising{gym,cycletowork}C4:WorkingorstudyingC5:Workingforatleast4hoursC6:ShoppingC2:Commuting{cycle/drivetowork}C7:Atleast2peopleeatingsimultaneously segments in the obtained flow diagrams show interesting patterns. We will give
two examples. First we consider segmenting the morphology of formations of
a defensive line of football players during a match (Fig. 5). The obtained flow
diagram provides an intuitive summary of these formations. The second example
models attacking possessions as state sequences. The summary given by the flow
diagram gives intuitive information about differences in attacking tactics.
Properties of Criteria. The efficiency of the algorithms will depend on prop-
erties of the criteria on which the segmentations are based. Here we consider
four cases: (i) general criteria without restrictions; (ii) monotone decreasing and
independent criteria; (iii) monotone decreasing and dependent criteria; and (iv)
fixed criteria. To illustrate the properties we will again use the example in Fig. 1.
A criterion C is monotone decreasing [7] for a given sequence of states τ that
fulfils C, if all subsequences of τ also fulfil C. For example, if C4 is fulfilled by a
sequence τ then any subsequence τ(cid:48) of τ will also fulfil C4. This is in contrast to
criterion C5 which is not monotone decreasing.
A criterion C is independent if checking whether a subsequence τ(cid:48) of a
sequence τi ∈ T fulfils C can be achieved without reference to any other sequences
τj ∈ T , i (cid:54)= j. Conversely, C is dependent if checking that a subsequence τ(cid:48) of τi
requires reference to other state sequences in T . In the above example C4 is an
example of an independent criterion while C7 is a dependent criterion since it
requires that at least two objects fulfil the criterion at the same time.
Related work. To the best of our knowledge compactly representing sequences
of states as flow diagrams has not been considered before. The only related work
we are aware of comes from the area of trajectory analysis. Spatial trajectories
are a special case of state sequences. A spatial trajectory describes the movement
of an object through space over time, where the states are location points, which
may also include additional information such as heading, speed, and temperature.
For a single trajectory a common way to obtain a compact representation is
simplification [9]. Trajectory simplification asks to determine a subset of the
data that represents the trajectory well in terms of the location over time. If the
focus is on characteristics other than the location, then segmentation [1,2,7] is
used to partition a trajectory into a small number of subtrajectories, where each
subtrajectory is homogeneous with respect to some characteristic. This allows a
trajectory to be compactly represented as a sequence of characteristics.
For multiple trajectories other techniques apply. A large set of trajectories
might contain very unrelated trajectories, hence clustering may be used. Clus-
tering on complete trajectories will not represent information about interesting
parts of trajectories; for this clustering on subtrajectories is needed [5,12]. A
set of trajectories that forms different groups over time may be captured by a
grouping structure [6]. These approaches also focus on location over time.
For the special case of spatial trajectories, a flow diagram can be illustrated
by a simple example: trajectories of migrating geese, see [8]. The individual
trajectories can be segmented into phases of activities such as directed flight,
3
foraging and stop overs. This results in a flow diagram containing a path for the
segmentation of each trajectory. More complex criteria can be imagined that
depend on a group of geese, or frequent visits to the same area, resulting in
complex state sequences that are hard to analyze without computational tools.
Organization In Section 3 we present algorithms for the Flow Diagram problem
using criteria with the properties described above. These algorithms only run in
polynomial time if the number of state sequences m is constant. Below we observe
that this is essentially the best we can hope for by showing that the problem is
W [1]-hard. Both theorems are proved in Section 2. Unless W [1] = F P T , this
rules out the existence of algorithms with time complexity of O(f (m) · (nk)c) for
some constant c, where m, n and k are the number of state sequences, the length
of the state sequences and the number of criteria, respectively. To obtain flow
diagrams for larger groups of state sequences we propose two heuristics for the
problem in Section 3. We experimentally evaluate the algorithms and heuristics
in Section 4.
2 Hardness Results
In this section, the following hardness results are proven.
Theorem 2. The FD problem is NP-hard. This even holds when only two criteria
are used or when the length of every state sequence is 2. Furthermore, for any
0 < c < 1/4, the FD problem cannot be approximated within factor of c log m in
polynomial time unless N P ⊂ DT IM E(mpolylog m).
Also for bounded m the running times of our algorithms is rather high. Again,
we can show that there are good reasons for this.
Theorem 3. The FD problem parameterized in the number of state sequences is
W [1]-hard even when the number of criteria is constant.
To obtain the stated results we will perform two reductions; one from the
Shortest Common Supersequence problem and one from the Set Cover problem.
2.1 Reduction from SCS
Problem 4. Shortest Common Supersequence (SCS)
Instance: A set of strings R = {r1, r2, . . . , rk} over an alphabet Σ, a positive
integer λ.
Question: Does there exist a string s ⊂ Σ∗ of length at most λ, that is a
supersequence of each string in R?
The SCS problem has been extensively studied over the last 30 years (see [10]
and references therein). Several hardness results are known, we will use the
following two.
4
(a)
(b)
Fig. 2: Examples of flow diagrams produced by the reductions: (a) From Shortest
Common Supersequence. (b) From Set Cover.
Lemma 5 (Pietrzak [16]). The Shortest Common Supersequence problem pa-
rameterized in the number of Strings is W [1] hard even when the alphabet has
constant size.
Lemma 6 (Raiha and E. Ukkonen [18]). The Shortest Common Superse-
quence problem over a binary alphabet is NP-complete.
Given an instance I = (R = {r1, . . . , rm}, Σ) of SCS construct an instance of
FD as follows. Each character cl in the alphabet Σ corresponds to a criterion cl.
Each string ri corresponds to a state sequence Ti, where Ti[j] = cri[j]. Thus at
any step Ti fulfils exactly one criterion.
An algorithm for FD given an instance outputs a flow diagram F of size f .
Given F one can compute a linear sequence b of the vertices of F using topological
sort, as shown in Fig. 2a. The linear sequence b has f − 2 vertices (omitting the
start and end state of F ) and it is a supersequence of each string in R. It follows
that the size of F is λ if the number of characters in the SCS of I has length λ− 2.
Note that F contains a linear sequence of vertices (after topological sort), which
correspondence to a supersequence, and a set of directed edges. Consequently a
solution for the FD problem can easily be transformed to a solution for the SCS
problem but not vice versa.
From the above reduction, together with Lemmas 5-6, we obtain Theorem 3
and the following.
Lemma 7. The FD problem is NP-hard even for two criteria.
2.2 Reduction from Set Cover
Problem 8. Set Cover (SC)
Instance: A set of elements E = {e1, e2, . . . , em}, a set of n subsets of E,
S = {S1, S2, . . . , Sn} and a positive integer λ.
Question: Does there exist set of λ items in S whose union equals E?
5
stts Set Cover is well known to be NP-hard, and also hard to approximate:
Lemma 9 (Lund and Yannakakis [15]). For any 0 < c < 1/4, the Set
Covering problem cannot be approximated within factor of c log m in polynomial
time unless N P ⊂ DT IM E(mpolylogm).
Given an instance I = (E = {e1, e2, . . . , em}, S = {S1, S2, . . . , Sn}) of Set
Cover construct an instance of FD as follows. Each item ei in E corresponds to
a state sequence Ti of length two. Each subset Sj corresponds to a criterion Cj.
If a Sj contains ei then the whole state sequence Ti fulfils criterion Cj.
An algorithm for FD given the new instance outputs a flow diagram F of
size f . The output F is depicted in Fig. 2b. Given F the interior vertices of F
corresponds to a set of subsets in S whose union is E. The diagram F has f
vertices if and only there is f − 2 subsets in S that forms a Set Cover of E.
We obtain Theorem 2 from the above reduction, together with Lemma 9.
3 Algorithms
In this section, we present algorithms that compute a smallest flow diagram
representing a set of m state sequences of length n for a set of k criteria. First, we
present an algorithm for the general case, followed by more efficient algorithms for
the case of monotone increasing and independent criteria, the case of monotone
increasing and dependent criteria, and then two heuristic algorithms.
3.1 General criteria
Next, we present a dynamic programming algorithm for finding a smallest flow
diagram. Recall that a node v in the flow diagram represents a criterion Cj that
is fulfilled by a contiguous segment in some of the state sequences. Let τ [i, j],
i ≤ j, denote the subsequence of τ starting at the ith state of τ and ending at
the jth state, where τ [i, i] is the empty sequence. Construct an (n + 1)m grid
of vertices, where a vertex with coordinates (x1, . . . , xm), 0 ≤ x1, . . . , xm ≤ n,
represents (τ1[0, x1], . . . , τm[0, xm]). Construct a prefix graph G as follows:
1, . . . , x(cid:48)
m),
labeled by some criterion Cj, if and only if, for every i, 1 ≤ i ≤ m, one of the
following two conditions is fulfilled: (1) xi = x(cid:48)
i, or (2) all remaining τi[xi + 1, x(cid:48)
i]
jointly fulfil Cj. Consider the edge between (x1, x2) = (1, 0) and (x(cid:48)
2) = (1, 1)
in Fig. 3(b). Here x1 = x(cid:48)
There is an edge between two vertices v = (x1, . . . , xm) and v(cid:48) = (x(cid:48)
1, x(cid:48)
1 and τ2[x2 + 1, x(cid:48)
2] fulfils C2.
Finally, define vs to be the vertex in G with coordinates (0, . . . , 0) and add an
additional vertex vt outside the grid, which has an incoming edge from (n, . . . , n).
This completes the construction of the prefix graph G.
Now, a path in G from vs to a vertex v represents a valid segmentation of
some prefix of each state sequence, and defines a flow diagram that describes these
segmentations in the following way: the empty path represents the flow diagram
consisting only of the start node s. Every edge of the path adds one new node to
the flow diagram, labeled by the criterion that the segments fulfil. Additionally,
6
Fig. 3: (a) A segmentation of T = {τ1, τ2} according to C = {C1, C2, C3}. (b) The
prefix graph G of the segmentation, omitting all but four of the edges. (c) The
resulting flow diagram generated from the highlighted path in the prefix graph.
for each node the flow diagram contains an edge from every node representing a
previous segment, or from s if the node is the first in a segmentation. For a path
leading from vs to vt, the target node t is added to the flow diagram, together
with its incoming edges. This ensures that the flow diagram represents valid
segmentations and that each node represents at least one segment. An example
of this construction is shown in Fig. 3.
Hence the length of a path (where length is the number of edges on the path)
equals the number of nodes of the corresponding flow diagram, excluding s and t.
Thus, we find an optimal flow diagram by finding a shortest vs -- vt path in G.
Lemma 10. A smallest flow diagram for a given set of state sequences is repre-
sented by a shortest vs -- vt path in G.
Proof. We show that every vs -- vt path P in G represents a valid flow diagram
F , with the path length equal to the flow diagram's cost, and vice versa. Thus,
a shortest path represents a minimal valid flow diagram for the given state
sequences.
Let P := (vs =: v1, v2, . . . , v(cid:96) := vt) be a vs -- vt path of length (cid:96) − 1 in G. As
described in the text, every vs -- vt path in G represents a valid flow diagram, and
every vertex visited by the path contributes exactly one node to the flow diagram.
Thus, P represents a valid flow diagram with exactly (cid:96) nodes.
For the other direction, let F be a valid flow diagram of a set of state sequences
{T1, . . . , Tm}, each of length n. That is, there are segmentations S1, . . . ,Sm of the
state sequences such that every segmentation is represented in F in the following
way: assume the nodes of F are {s =: f1, f2, . . . , f(cid:96) := t} according to some
topological sorting. Let Sj consist of the segments sj,1, . . . , sj,σj , where σj is the
number of segments in Sj. Then there exists a path (s =: fj,0, fj,1, . . . , fj,σj , t)
in F such that each segment sj,i fulfils the criterion C(fj,i) associated with fj,i.
Let bj,i be the index in Tj at which sj,i ends, for 1 ≤ i ≤ σj, and let bj,0 := 1.
Since Sj is a segmentation of Tj, bj,σj = n. Let Fλ be the subdiagram of F induced
by (f1, . . . , fλ), for 1 ≤ λ ≤ (cid:96). We define xj,λ := max{bj,i fj,i ∈ {f1, . . . , fλ}}.
We show inductively that for each λ ∈ {1, 2, . . . , (cid:96)}, G contains a path from vs
7
(a)(b)(c)231τ1[C1][C3][C1]τ2[C1,C2][C3][C2]sτ2τ1001122vtvsC2C3C1C1C2C3t to the vertex vλ := (x1,λ, . . . , xm,λ) with length λ − 1, i.e. the number of nodes
in Fλ excluding s.
-- Base case λ = 1: Note that v1 = vs, and thus there is a path of length
λ − 1 = 0 from vs to v1.
-- Induction step: The node fλ+1 represents the segments
{Tj[xj,λ, xj,λ+1] 1 ≤ j ≤ m ∧ xj,λ (cid:54)= xj,λ+1}.
Since the flow diagram is valid, these segments fulfil the criterion C(fλ+1),
and thus G contains an edge from vλ to vλ+1. Since a path from vs to vλ of
length λ exists by the induction hypothesis, there is a path from vs to vλ+1
of length λ + 1.
For every state sequence Tj, there exists an index ϕj ∈ {1, . . . , (cid:96) − 1} such
that xj,λ = n for all λ ≥ ϕj. Thus, v(cid:96)−1 = (n, n, . . . , n) and G contains an edge
from v(cid:96)−1 to v(cid:96) = vt. So, there is a path from vs to vt of length (cid:96) − 1.
(cid:117)(cid:116)
Recall that G has (n + 1)m vertices. Each vertex has O(k(n + 1)m) outgoing
edges, thus, G has O(k(n + 1)2m) edges in total. To decide if an edge is present
in G, check if the nonempty segments the edge represents fulfil the criterion.
Thus, we need to perform O(k(n + 1)2m) of these checks. There are m segments
of length at most n, and we assume the cost for checking this is T (m, n). Thus,
the cost of constructing G is O(k(n + 1)2m · T (m, n)), and finding the shortest
path requires O(k(n + 1)2m) time.
Theorem 11. The algorithm described above computes a smallest flow diagram
for a set of m state sequences, each of length at most n, and k criteria in
O((n + 1)2mk · T (m, n)) time, where T (m, n) is the time required to check if a
set of m subsequences of length at most n fulfils a criterion.
3.2 Monotone decreasing and independent criteria
If all criteria are decreasing monotone and independent, we can use ideas similar
to those presented in [7] to avoid constructing the full graph. From a given vertex
with coordinates (x1, . . . , xm), we can greedily move as far as possible along the
sequences, since the monotonicity guarantees that this never leads to a solution
that is worse than one that represents shorter segments. For a given criterion Cj,
we can compute for each τi independently the maximum x(cid:48)
i such that τi[xi + 1, x(cid:48)
i]
fulfils Cj. This produces coordinates (x(cid:48)
m) for a new vertex, which is the
optimal next vertex using Cj. By considering all criteria we obtain k new vertices.
However, unlike the case with a single state sequence, there is not necessarily
one vertex that is better than all others (i.e. largest ending position), since there
is no total order on the vertices. Instead, we consider all vertices that are not
dominated by another vertex, where a vertex p dominates a vertex p(cid:48) if each
coordinate of p is at least as large as the corresponding coordinate of p(cid:48), and at
least one of p's coordinates is larger.
1, . . . , x(cid:48)
8
Let Vi be the set of vertices of G that are reachable from vs in exactly i
steps, and define M (V ) := {v ∈ V no vertex u ∈ V dominates v} to be the set
of maximal vertices of a vertex set V . Then a shortest vs -- vt path through G can
be computed by iteratively computing M (Vi) for increasing i, until a value of i
is found for which vt ∈ M (Vi). Observe that M (V ) = O((n + 1)m−1) for any
set V of vertices in the graph. Also note that V0 = M (V0) = vs.
Lemma 12. For each i ∈ {1, . . . , (cid:96) − 1}, every vertex in M (Vi) is reachable in
one step from a vertex in M (Vi−1). Here, (cid:96) is the distance from vs to vt.
Proof. Assume there exists a vertex v ∈ M (Vi) that has no edge from a vertex in
M (Vi−1). Since v ∈ M (Vi), v is also contained in Vi, and thus its distance from vs
is i. Thus, there must be a vertex v(cid:48) at distance i − 1 from vs, i.e. v(cid:48)
∈ Vi−1, that
has an edge to v representing a criterion Cj. By assumption, v(cid:48) is not contained
in M (Vi−1), and thus there is a vertex v(cid:48)(cid:48)
∈ M (Vi−1) that dominates v(cid:48). But
then, by the monotonicity of Cj, there must be a vertex reachable from v(cid:48)(cid:48) that
(cid:117)(cid:116)
is identical to v or dominates v. Both cases lead to a contradiction.
M (Vi) is computed by computing the farthest reachable vertex for each
v ∈ M (Vi−1) and criterion, thus yielding a set Di of O((n + 1)m−1k) vertices.
This set contains M (Vi) by Lemma 12, so we now need to remove all vertices
that are dominated by some other vertex in the set to obtain M (Vi).
We find M (Vi) using a copy of G. Each vertex may be marked as being in
Di or dominated by a vertex in Di. We process the vertices of Di in arbitrary
order. For a vertex v, if it is not yet marked, we mark it as being in Di. When a
vertex is newly marked, we mark its ≤ m immediate neighbours dominated by
it as being dominated. After processing all vertices, the grid is scanned for the
vertices still marked as being in Di. These vertices are exactly M (Vi).
When computing M (Vi), O((n + 1)m−1k) vertices need to be considered, and
the maximum distance from vs to vt is m(n + 1), so the algorithm considers
O(mk(n+1)m) vertices. We improve this bound by a factor m using the following:
Lemma 13. The total size of all Di, for 0 ≤ i ≤ (cid:96) − 1, is O(k(n + 1)m).
Proof. If a vertex v appears in M (Vi) for some i ∈ {0, . . . , (cid:96) − 1}, it generates
vertices for Di+1 that dominate v, and thus v (cid:54)∈ M (Vi+j) for any j > 0. So,
each of the nm vertices appears in at most one M (Vi) and generates k candidate
vertices for Di+1 (not all unique). Hence the total size of all Di is O(knm). (cid:117)(cid:116)
Using this result, we compute all M (Vi) in O((k + m)(n + 1)m) time, since
O(k(n + 1)m) vertices are marked directly, and each of the (n + 1)m vertices is
checked at most m times when a direct successor is marked. One copy of the grid
can be reused for each M (Vi), since each vertex of Di+1 dominates at least one
vertex of M (Vi) and is thus not yet marked while processing Dj for any j ≤ i.
Since the criteria are independent, the farthest reachable point for a given
starting point and criterion can be precomputed for each state sequence separately.
Using the monotonicity we can traverse each state sequence once per criterion and
thus need to test only O(nmk) times whether a subsequence fulfils a criterion.
9
Theorem 14. The algorithm described above computes a smallest flow diagram
for m state sequences of length n with k independent and monotone decreasing
criteria in O(mnk · T (1, n) + (k + m)(n + 1)m) time, where T (1, n) is the time
required to check if a subsequence of length at most n fulfils a criterion.
3.3 Monotone decreasing and dependent criteria
For monotone decreasing and dependent criteria, we can use a similar approach
to that described above, however, for a given start vertex v and criterion C,
there is not a single vertex v(cid:48) that dominates all vertices reachable from v
using this criterion. Instead there may be Θ((n + 1)m−1) maximal reachable
vertices from v for criterion C. The maximal vertices can be found by testing
O((n + 1)m−1) vertices on or near the upper envelope of the reachable vertices in
O((n + 1)m−1 · T (m, n)) time. Using a similar reasoning as in Lemma 13, we can
show that the total size of all Di (0 ≤ i ≤ (cid:96) − 1) is O(k(n + 1)2m−1), which gives:
Theorem 15. The algorithm from the previous section computes a smallest flow
diagram for m state sequences of length n with k monotone decreasing criteria in
O(k(n + 1)2m−1 · T (m, n) + m(n + 1)m) time, where T (m, n) is the time required
to check if a set of m subsequences of length at most n fulfils a criterion.
3.4 Heuristics
The hardness results presented in the introduction indicate that it is unlikely
that the performance of the algorithms will be acceptable in practical situations,
except for very small inputs. As such, we investigated heuristics that may produce
usable results that can be computed in reasonable time.
For monotone decreasing and independent criteria, the heuristics we consider
are based on the observation that by limiting Vi, the vertices that are reachable
from vs in i steps, to a fixed size, the complexity of the algorithm can be controlled.
Given that every path in a prefix graph represents a valid flow diagram, any path
chosen in the prefix graph will be valid, though not necessarily optimal. In the
worst case, a vertex that advances along a single state sequence a single time-step
(i.e. advancing only one state) will be selected, and for each vertex, all k criteria
must be evaluated, so O(kmn) vertices may be processed by the algorithm. We
consider two strategies for selecting the vertices in Vi to retain:
(1) For each vertex in Vi, determine the number of state sequences that are
advanced in step i and retain the top q vertices [sequence heuristic].
(2) For each vertex in Vi, determine the number of time-steps that are advanced
in all state sequences in step i and retain the top q vertices [time-step heuristic].
In our experiments we use q = 1 since any larger value would immediately
give an exponential worst-case running time.
4 Experiments
The objectives of the experiments were twofold: to determine whether compact
and useful flow diagrams could be produced in real application scenarios; and to
10
empirically investigate the performance of the algorithms on inputs of varying
sizes. We implemented the algorithms described in Section 3 using the Python
programming language. For the first objective, we considered the application
of flow diagrams to practical problems in football analysis in order to evaluate
their usefulness. For the second objective, the algorithms were run on generated
datasets of varying sizes to investigate the impact of different parameterisations
on the computation time required to produce the flow diagram and the complexity
of the flow diagram produced.
4.1 Tactical Analysis in Football
Sports teams will apply tactics to improve their performance, and computational
methods to detect, analyse and represent tactics have been the subject of several
recent research efforts [4,11,14,19,20,21]. Two manifestations of team tactics are
in persistent and repeated occurrence of spatial formations of players, and in plays
-- a coordinated sequence of actions by players. We posited that flow diagrams
would be a useful tool for compactly representing both these manifestations, and
we describe the approaches used in this section.
The input for the experiments is a database containing player trajectory
and event data from four home matches of the Arsenal Football Club from the
2007/08 season, provided by Prozone Sports Limited [17]. For each player and
match, there is a trajectory comprising a sequence of timestamped location points
in the plane, sampled at 10 Hz and accurate to 10 cm. In addition, for each match,
there is a log of all the match events, comprising the timestamp and location of
each event.
Defensive Formations. The spatial formations of players in football matches
are known to characterize a team's tactics [3], and a compact representation
of how formations change over time would be a useful tool for analysis. We
investigated whether a flow diagram could provide such a compact representation
of the defensive formation of a team, specifically to show how the formation
evolves during a phase of play. The trajectories of the four defensive players were
re-sampled at one-second intervals and used to compute a sequence of formation
states which were then segmented to model the formation.
The criteria were derived from those presented by Kim et al. [13]. The angles
between pairs of adjacent players (along the defence line) were used to compute
the formation criteria, see Fig 4. We extended this scheme to allow multiple
criteria to be applied where the angle between pairs of players is close to 10◦. The
reason for this was to facilitate compact results by allowing for smoothing of small
variations in contiguous time-steps. The criteria applied to each state is a triple
(x1, x2, x3), computed as follows. Given two player positions p and q, let ∠pq be
the angle between p and q relative to the goal-line. Let R(−1) = [−90◦,−5◦),
R(0) = (−15◦, +15◦), and R(+1) = (+5◦, +90◦] be three angular ranges. The
(x1, x2, x3) if ∠pipi+1 ∈ R(xi) for all i ∈ {1, 2, 3}.
positions of the four defenders satisfy the criteria (and thus have the formation)
11
Fig. 4: Segmentation of a single state sequence. The formation state sequence is
used to compute the segmentation representation, where segments corresponding
to criteria span the state sequence (bottom). The representation of this state
sequence in the movement flow diagram is shaded in Fig. 5.
Fig. 5: Flow diagram for formation morphologies of twelve defensive possessions.
The shaded nodes are the segmentation of the state sequence in Fig. 4.
The criteria in this experiment were monotone decreasing and independent,
and we ran the corresponding algorithm using randomly selected sets of the state
sequences as input. The size m of the input was increased until the running
time exceeded a threshold of 6 hours. The algorithm successfully processed up to
m = 12 state sequences, having a total of 112 assigned segments. The resulting
flow diagram, Fig. 5, has a total complexity of 12 nodes and 27 edges.
We believe that the flow diagram provides an intuitive summary of the
defensive formation, and several observations are apparent. There appears to
be a preference amongst the teams for the right-back to position himself in
advance of the right centre-half (i.e. the third component of the triple is +1).
Furthermore, the (0, 0, 0) triple, corresponding to a "flat back four" is not present
12
012345678910Trajectory Timestep(+1,0,+1)(+1,-1,+1)(0,+1,+1)(0,0,+1)(0,-1,+1)(-1,+1,+1)(-1,0,+1)(-1,-1,+1)Criteria3422222422st in the diagram. This is typically considered the optimal formation for teams
that utilise the offside trap, and thus may suggest that the defences here are
not employing this tactic. These observations were apparent to the authors as
laymen, and we would expect that a domain expert would be able to extract
further useful insights from the flow diagrams.
Attacking Plays. During a football match, the team in possession of the ball
is attempting to reach a position where they can take a shot at goal. Teams
will typically use a variety of tactics to achieve such a position, e.g. teams can
vary the intensity of an attack by pushing forward, moving laterally, making
long passes, or retreating. We modelled attacking possessions as state sequences,
segmented according to criteria representing the attacking intensity and tactics
employed, and computed flow diagrams for the possessions. In particular, we
were interested in determining whether differences in tactics employed by teams
when playing at home or away [4] are apparent in the flow diagrams.
We focus on ball events, where a player touches the ball, e.g. passes, touches,
dribbles, headers, and shots at goal. The event sequence for each match was
divided into sub-sequences where a single team was in possession, and then
filtered to include only those that end with a shot at goal.
We defined criteria that characterised the movement of the ball - relative to
the goal the team is attacking - between event states in the possession sequence.
The applied criteria are defined as follows. Let xi, yi, ti be the x-coordinate in
metres, y-coordinate in metres and time-stamp in seconds, respectively, for event
i. The velocity of the ball in the x-direction at event i, which is in the direction
of the goal, is thus xv = (xi+1 − xi)/(ti+i − ti) in m/s. The velocity yv of the
ball in the y direction is computed in a similar fashion, and together are used to
specify the following criteria.
-- Backward movement (BM): xv < 1, a sub-sequence of passes or touches that
move in a defensive direction.
lateral direction.
-- Lateral movement (LM): −5 < xv < 5, passes or touches that move in a
-- Forward movement (FM): −1 < xv < 12, passes or touches that move in an
attacking direction, at a velocity in the range achievable by humans, i.e. to
approximately 10m/s.
-- Fast forward movement (FFM): 8 < xv, passes or touches moving in an
attacking direction at a velocity generally in exceeds of maximum human
velocity.
-- Long ball (LB): a pass travelling 30m in the attacking direction.
-- Cross-field ball (CFB): a pass travelling 20m in the cross-field direction, and
that has angle in range [80, 100] or [−80,−100].
-- Shot resulting in goal (SG): a successful shot resulting in a goal.
-- Shot not resulting in goal (SNG): an unsuccessful shot that does not produce
a goal.
13
Fig. 6: Flow diagram produced for the home team. The edge weights are the
number of possessions that span the edge, and the nodes with grey background
are event types that are significant, as defined in Section 4.1.
For a football analyst, the first four criteria are simple movements, and are
not particularly interesting. The last four events are significant: the long ball
and cross-field ball change the locus of attack; and the shot criteria represent the
objective of an attack.
The possession state sequences for the home and visiting teams were segmented
according to the criteria and the time-step heuristic algorithm was used to
compute the flow diagrams. The home-team input consisted of 66 sequences
covered by a total of 866 segments, and resulted in a flow diagram with 25 nodes
and 65 edges, see Fig. 6. Similarly, the visiting-team input consisted of 39 state
sequences covered by 358 segments and the output flow diagram complexity was
22 nodes and 47 edges, as shown in Fig. 7.
At first glance, the differences between these flow diagrams may be difficult
to appreciate, however closer inspection reveals several interesting observations.
The s -- t paths in the home-team flow diagram tend to be longer than those in
the visiting team's, suggesting that the home team tends to retain possession of
the ball for longer, and varies the intensity of attack more often. Moreover, the
14
sBM41FM21FFM4303CFB2SG2SNG47126BM351LM15BM4FM2FM7t759318FM23218BM10LB231114BM2FM3113274FFM129110BM4FM221131BM41FM31BM2222BM112 Fig. 7: Flow diagram produced for the visiting team. The edge weights are the
number of possessions that span the edge, and the nodes with grey background
are event types that are significant, as defined in Section 4.1.
nodes for cross-field passes and long-ball passes tend to occur earlier in the s -- t
paths in the visiting team's flow diagram. These are both useful tactics as they
alter the locus of attack, however they also carry a higher risk. This suggests that
the home team is more confident in its ability to maintain possession for long
attack possessions, and will only resort to such risky tactics later in a possession.
Furthermore, the tactics used by the team in possession are also impacted by
the defensive tactics. As Bialkowski et al [4] found, visiting teams tend to set
their defence deeper, i.e. closer to the goal they are defending. When the visiting
team is in possession, there is thus likely to be more space behind the home
team's defensive line, and the long ball may appear to be a more appealing tactic.
The observations made from these are consistent with our basic understanding
of football tactics, and suggest that the flow diagrams are interpretable in this
application domain.
4.2 Performance Testing
In the second experiment, we used a generator that outputs synthetic state
sequences and segmentations, and tested the performance of the algorithms on
inputs of varying sizes.
The segmentations were generated using Markov Chain Monte Carlo sampling.
Nodes representing the criteria set of size k were arranged in a ring and a Markov
chain constructed, such that each node had a transition probability of 0.7 to
15
sBM28FM1121SNG3FFM2LB1SG1733BM17CFB2t373BM5FM4252FM10BM1LM12BM811LM5FM1411BM3FFM124111FM22LM2111212 remain at the node, 0.1 to move to the adjacent node, and 0.05 to move to the
node two places away. Segmentations were computed by sampling the Markov
chain starting at a random node. Thus, simulated datasets of arbitrary size m,
state sequence length n, criteria set size k were generated.
We performed two tests on the generated segmentations. In the first, ex-
periments were run on the four algorithms described in Section 3 with varying
configurations of m, n and k to investigate the impact of input size on the
algorithm's performance. The evaluation metric used was the CPU time required
to generate the flow diagram for the input. In the second test, we compared
the total complexity of the output flow diagram produced by the two heuristic
algorithms with the baseline complexity of the flow diagram produced by the
exact algorithm for monotone increasing and independent criteria.
We repeated each experiment five times with different input sequences for
each trial, and the results presented are the mean values of the metrics over the
trials. Limits were set such that the process was terminated if the CPU time
exceeded 1 hour, or the memory required exceeded 8GB.
The results of the first test showed empirically that the exact algorithms have
time and storage complexity consistent with the theoretical worst-case bounds,
Fig. 8 (top). The heuristic algorithms were subsequently run against larger test
data sets to examine the practical limits of the input sizes, and were able to
process larger input -- for example, an input of k = 128, m = 32 and n = 1024 was
tractable -- although the cost is that the resulting flow diagrams were suboptimal,
but correct, in terms of their total complexity.
For the second test, we investigated the complexity of the flow diagram induced
by inputs of varying parameterisations when using the heuristic algorithms. The
objective was to examine how close the complexity was to the optimal complexity
produced using an exact algorithm. The inputs exhibited monotone decreasing
and independent criteria, and thus the corresponding algorithm was used to
produce the baseline. Fig. 8 (bottom) summarises the results for varying input
parameterisations. The complexity of the flow diagrams produced by the two
heuristic algorithms are broadly similar, and increase at worst linearly as the input
size increases. Moreover, while the complexity is not optimal it appears to remain
within a constant factor of the optimal, suggesting that the heuristic algorithms
could produce usable flow diagrams for inputs where the exact algorithms are
not tractable.
5 Concluding Remarks
We introduced flow diagrams as a compact representation of a large number of
state sequences. We argued that this representation gives an intuitive summary
allowing the user to detect patterns among large sets of state sequences, and
gave several algorithms depending on the properties of the segmentation criteria.
These algorithms only run in polynomial time if the number of state sequences
m is constant, which is the best we can hope for given the problem is W [1]-hard.
As a result we considered two heuristics capable of processing large data sets in
16
Fig. 8: Runtime statistics for generating flow diagram (top), and total complexity
of flow diagrams produced (bottom). Default values of m = 4, n = 4 and k = 10
were used. The data points are the mean value and the error bars delimit the
range of values over the five trials run for each input size.
reasonable time, however we were unable to give an approximation bound. We
tested the algorithms experimentally to assess the utility of the flow diagram
representation in a sports analysis context, and also analysed the performance of
the algorithms of inputs of varying parameterisations.
References
1. S. P. A. Alewijnse, K. Buchin, M. Buchin, A. Kolzsch, H. Kruckenberg, and
M. Westenberg. A framework for trajectory segmentation by stable criteria. In
Proc. 22nd ACM SIGSPATIAL/GIS, pages 351 -- 360. ACM, 2014.
2. B. Aronov, A. Driemel, M. J. van Kreveld, M. Loffler, and F. Staals. Segmentation
of trajectories for non-monotone criteria. In Proc. 24th ACM-SIAM SODA, pages
1897 -- 1911, 2013.
3. A. Bialkowski, P. Lucey, G. P. K. Carr, Y. Yue, S. Sridharan, and I. Matthews.
Identifying team style in soccer using formations learned from spatiotemporal
tracking data. In ICDM Workshops, pages 9 -- 14. IEEE, 2014.
4. A. Bialkowski, P. Lucey, P. Carr, Y. Yue, and I. Matthews. Win at home and draw
away: automatic formation analysis highlighting the differences in home and away
team behaviors. In Proc. 8th Annual MIT Sloan Sports Analytics Conference, 2014.
17
2468101214161820Criteria Set Size k10-410-310-210-1100101102103104Execution time in sec (log)2468101214161820Number of State Sequences m10-410-310-210-1100101102103104Execution time in sec (log)12345678910State Sequence Length n10-410-310-210-1100101102103104Execution time in sec (log)General CriteriaMonotone Decreasing and Independent CriteriaSegment HeuristicTimestep Heuristic212223242526272829210Criteria Set Size k (log2)10121416182022Flow Diagram Complexity212223242526272829210Number of State Sequences m (log2)020406080100120140160Flow Diagram Complexity212223242526272829210State Sequence Length n (log2)05001000150020002500Flow Diagram Complexity 5. K. Buchin, M. Buchin, J. Gudmundsson, M. Loffler, and J. Luo. Detecting com-
muting patterns by clustering subtrajectories. Int. J. Comput. Geometry Appl.,
21(3):253 -- 282, 2011.
6. K. Buchin, M. Buchin, M. J. van Kreveld, B. Speckmann, and F. Staals. Trajectory
grouping structure. In Proc. 13th WADS, pages 219 -- 230, 2013.
7. M. Buchin, A. Driemel, M. van Kreveld, and V. Sacristan. Segmenting trajectories:
A framework and algorithms using spatiotemporal criteria. Journal of Spatial
Information Science, 3:33 -- 63, 2011.
8. M. Buchin, H. Kruckenberg, and A. Kolzsch. Segmenting trajectories based on
movement states. In Proc. 15th SDH, pages 15 -- 25. Springer-Verlag, 2012.
9. H. Cao, O. Wolfson, and G. Trajcevski. Spatio-temporal data reduction with
deterministic error bounds. The VLDB Journal, 15(3):211 -- 228, 2006.
10. C. B. Fraser and R. W. Irving. Approximation algorithms for the shortest common
supersequence. Nordic Journal of Computing, 2(3):303 -- 325, 1995.
11. J. Gudmundsson and T. Wolle. Football analysis using spatio-temporal tools.
Computers, Environment and Urban Systems, 47:16 -- 27, 2014.
12. C.-S. Han, S.-X. Jia, L. Zhang, and C.-C. Shu. Sub-trajectory clustering algorithm
based on speed restriction. Computer Engineering, 37(7), 2011.
13. H.-C. Kim, O. Kwon, and K.-J. Li. Spatial and spatiotemporal analysis of soccer.
In Proc. 19th ACM SIGSPATIAL/GIS, pages 385 -- 388. ACM, 2011.
14. P. Lucey, A. Bialkowski, G. P. K. Carr, S. Morgan, I. Matthews, and Y. Sheikh.
Representing and Discovering Adversarial Team Behaviors Using Player Roles. In
Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR'13),
pages 2706 -- 2713, Portland, OR, jun 2013. IEEE.
15. C. Lund and M. Yannakakis. On the hardness of approximating minimization
problems. In Proc. 25th ACM STOC, pages 286 -- 293. ACM, 1993.
16. K. Pietrzak. On the parameterized complexity of the fixed alphabet shortest
common supersequence and longest common subsequence problems. Journal of
Computer and System Sciences, 67(4):757 -- 771, 2003.
17. Prozone Sports Ltd. Prozone Sports - Our technology. http://prozonesports.
stats.com/about/technology/, 2015.
18. K.-J. Raiha and E. Ukkonen. The shortest common supersequence problem over
binary alphabet is NP-complete. Theoretical Computer Sci., 16(2):187 -- 198, 1981.
19. J. Van Haaren, V. Dzyuba, S. Hannosset, and J. Davis. Automatically Discovering
Offensive Patterns in Soccer Match Data. In Advances in Intelligent Data Analysis
XIV - 14th International Symposium, IDA 2015, volume 9385 of Lecture Notes in
Computer Science, pages 286 -- 297, Saint Etienne, oct 2015. Springer.
20. Q. Wang, H. Zhu, W. Hu, Z. Shen, and Y. Yao. Discerning Tactical Patterns for
Professional Soccer Teams. In Proceedings of the 21th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining - KDD '15, pages 2197 -- 2206,
Sydney, aug 2015. ACM Press.
21. X. Wei, L. Sha, P. Lucey, S. Morgan, and S. Sridharan. Large-Scale Analysis
In 2013 International Conference on Digital Image
of Formations in Soccer.
Computing: Techniques and Applications (DICTA), pages 1 -- 8, Hobart, nov 2013.
IEEE.
18
|
1408.0596 | 3 | 1408 | 2015-10-31T14:58:04 | Approximation Bounds For Minimum Degree Matching | [
"cs.DS"
] | We consider the MINGREEDY strategy for Maximum Cardinality Matching. MINGREEDY repeatedly selects an edge incident with a node of minimum degree. For graphs of degree at most $\Delta$ we show that MINGREEDY achieves approximation ratio at least $ \frac{\Delta-1}{2\Delta-3} $ in the worst case and that this performance is optimal among adaptive priority algorithms in the vertex model, which include many prominent greedy matching heuristics. Even when considering expected approximation ratios of randomized greedy strategies, no better worst case bounds are known for graphs of small degrees. | cs.DS | cs |
Approximation Bounds For Minimum Degree Matching
Bert Besser⋆
Institut für Informatik, Goethe-Universität Frankfurt am Main, Germany
Abstract. We consider the MinGreedy strategy for Maximum Cardi-
nality Matching. MinGreedy repeatedly selects an edge incident with a
node of minimum degree. For graphs of degree at most ∆ we show that
MinGreedy achieves approximation ratio at least ∆−1
2∆−3 in the worst
case and that this performance is optimal among adaptive priority al-
gorithms in the vertex model, which include many prominent greedy
matching heuristics.
Even when considering expected approximation ratios of randomized
greedy strategies, no better worst case bounds are known for graphs
of small degrees.
Keywords: matching, greedy, approximation, priority algorithm.
1 Introduction
In the Maximum Cardinality Matching Problem a node disjoint subset
of edges of maximum size is to be determined. Matching problems have many
applications, e.g. image feature matching in computer vision or protein structure
comparison in computational biology.
A maximum matching can be found in polynomial time, e.g. by the algorithm
of Micali and Vazirani [MV80] running in time O(pV · E · α(E,V )) where
α(E,V ) is the inverse Ackermann function [Vaz12]. The algorithm of Mucha
and Sankowski [MS04] runs asymptotically faster on dense graphs, its runtime is
O(V ω) where ω < 2.38 is the exponent needed to perform matrix multiplication.
However, there are much faster greedy algorithms that in practice compute
very large matchings, even near optimal ones. Very good approximate solutions
may already be satisfactory in some applications. If maximum matchings are
needed, one can save lots of runtime when feeding large greedy matchings into
optimal algorithms which iteratively improve an initial solution.
MinGreedy. The (randomized) MinGreedy algorithm computes a match-
ing M by repeatedly picking edges incident to nodes of currently minimum de-
gree, see Figure 1. MinGreedy can be implemented in linear time O(V +E).
The experimentally observed approximation performance is quite impres-
sive. Tinhofer [Tin84] observed that on random graphs of varying density Min-
Greedy performed superior to Greedy (which randomly selects an edge) and
to MRG (which randomly selects a node and subsequently an incident edge). In
experiments of Frieze et al. [FRS93] on random cubic graphs MinGreedy left
only about 10 out of 106 nodes unmatched. On random graphs of small constant
⋆ Partially supported by DFG SCHN 503/6-1.
M = ∅
repeat until all edges removed from input graph:
select (random) node u of minimum non-zero degree
select (random) neighbor v of u
pick edge {u, v}, i.e. set M = M ∪ {{u, v}}
remove all edges incident with u and v from input graph
return M
Fig. 1. The (randomized) MinGreedy algorithm
average degree Magun [Mag97] observed that MinGreedy produces extremely
few "lost edges" in comparison with an optimal solution.
1
In an involved argument, Frieze et al. [FRS93] showed that c1 · n
5 ≤ λn ≤
c2 · n
5 · ln(n) holds (c1, c2 being constants) for the expected number λn of nodes
not being matched by MinGreedy on random n-node cubic graphs.
Whereas the performance of MinGreedy on random instances is very good,
its worst case performance is poor. Poloczek [Pol12] constructed hard input
instances on which MinGreedy (and common variations of the algorithm)
achieves approximation ratio at most 1
2 + o(1) w.h.p.
1
MRG, Greedy & Shuffle. The Modified Random Greedy algorithm,
abbreviated MRG in literature, ignores node degrees and repeatedly selects a
node and then a neighbor uniformly at random. The expected approximation
ratio was shown to be at least 1
400.000 by Aronson et al. [ADFS95].
2 + 1
The random edge algorithm Greedy repeatedly selects an edge uniformly
at random. For graphs with degrees bounded by ∆ an expected lower bound on
the approximation ratio of ∆
2∆−1 was shown by Dyer and Frieze [DF91] and later
improved by Miller and Pritikin [MP97] to 1
prefers edges of degree-1 nodes, the KarpSipser algorithm is obtained, which
is asymptotically optimal w.h.p. on large sparse random graphs [KS81].
2 (p(∆ − 1)2 + 1−∆+2). If Greedy
The Shuffle algorithm, proposed by Goel and Tripathi [GT12], is an adap-
tation of the Ranking algorithm of Karp et al. [KVV90] to non-bipartite graphs.
Shuffle selects a random permutation π of the nodes and repeatedly matches
the, according to π, first non-isolated node to its first unmatched neighbor. Chan
et al. [CCWZ14] showed that Shuffle achieves an approximation ratio of at
least 2 · (5 − √7)/9 ≈ 0.523.
Inapproximability. To show performance bounds for greedy algorithms,
Borodin et al. [BNR02] proposed the model of adaptive priority algorithms. The
model formalizes the greedy nature of an algorithm: while gathering knowledge
about the input, irrevocable decisions have to be made to construct a solution.
Davis and Impagliazzo [DI04] introduced the vertex model to study adaptive
priority algorithms for graph problems. Adaptive priority algorithms in the ver-
tex model for the matching problem, which we call APV-algorithms, implement
powerful node and edge selection routines. In particular, in each step a node v
and an incident edge is not picked arbitrarily but based on all knowledge already
gathered about v and its neighbors, e.g. is a neighbor matched or unmatched,
2
what is the degree of a neighbor, what are the neighbors of a neighbor, etc.
APV-algorithms include Greedy, KarpSipser, MRG, MinGreedy, Shuf-
fle and all vertex iterative algorithms, a class of algorithms defined in [GT12]
as a generalization of Shuffle.
Despite the strength of APV-algorithms, Poloczek [Pol12] constructed rather
simple graphs with worst case approximation ratio at most 2
3 . (He also showed
an inapproximability bound of 5
6 for randomized priority algorithms.) Presented
in the same thesis, for graphs with arbitrarily large degree Besser and Poloczek
showed that no APV-algorithm achieves worst case approximation ratio better
than 1
2 + o(1).
Contributions. From now on we reserve the name MinGreedy for the
deterministic version of MinGreedy in which a node of minimum degree and
an incident edge is picked by a worst case adversary.
We show that MinGreedy approximates an optimal matching within a fac-
tor of ∆−1
2∆−3 for graphs in which degrees are bounded by at most ∆. In the proof
we analyze a variant of MinGreedy which is also related with the KarpSipser
algorithm.
We also show that the worst case approximation performance of MinGreedy
is optimal for (deterministic) APV-algorithms. In particular, we improve the
construction of Besser and Poloczek given in [Pol12] and present hard input in-
stances of degree at most ∆ for which any APV-algorithm computes a matching
of size at most ∆−1
2∆−3 + o(1) times optimal.
Our worst case performance guarantees are stronger than the best known
400.000 resp.
bounds on the expected performance of MRG and Shuffle ( 1
≈0.523), for small ∆, and of Greedy (≈0.618 for ∆=3), for all ∆.
2 + 1
Techniques. For our performance guarantees for MinGreedy we study
the matching graph composed of the edges of a matching M , computed by
MinGreedy, and of a maximum matching M ∗. The connected components
are alternating paths and cycles. Only paths of length three have poor "local"
approximation ratio (of M -edges to M ∗-edges). To obtain a global performance
guarantee, we balance local approximation ratios by transferring "M -funds" from
rich to poor components using edges of the input graph.
Incorporating the properties of MinGreedy within an amortized analysis is
our technical contribution.
Overview. In Section 2 we present the charging scheme used to prove the
performance guarantees for MinGreedy. In Section 3 we show our 2
3 bound for
graphs of degree at most ∆ = 3. For graphs of degree at most ∆ ≥ 4 we present
in Section 4 our performance guarantee of ∆−1
2∆−3 . Our inapproximability results
for APV-algorithms are given in Section 5.
3
2 The Charging Scheme
The Matching Graph. Let G = (V, E) be a connected graph, M ∗ a maximum
matching in G and M a matching computed by MinGreedy when applied to
input G. To analyze the worst case approximation ratio of MinGreedy we
investigate the graph
H = (V, M ∪ M ∗) .
The connected components of H are paths and cycles composed of edges of M
and M ∗. For example, H contains so-called (M -)augmenting paths, or paths for
short: an augmenting path X has mX edges of M and m∗
X = mX + 1 edges of
M ∗ and starts and ends with an M ∗-edge:
We call the two nodes of a path X which do not have an incident M -edge
the endpoints of X (the leftmost and the rightmost node in the figure). Other
connected components of H are edges of M ∩ M ∗
,
which we call singletons. For a singleton X we have mX = m∗
focus on these two component types:
X = 1. We may
Lemma 1. There is a maximum matching M ∗ in G such that each connected
component of H is a singleton or an augmenting path.
Proof. Let M ′ be a maximum matching in G = (V, E) and let MinGreedy com-
pute the matching M . We show how to transform M ′ into M ∗. The connected
components of the graph (V, M ∪ M ′) are singletons and alternating paths and
cycles, where a path starts and ends with an M -edge or M ′-edge and a cycle
does not have path endpoints. To prove the statement we eliminate paths start-
ing and ending both with an M -edge, paths starting and ending with different
types of edges, and cycles. There is no path X of the first type, since if there
was, then we could replace the m′
X + 1 many
M -edges of X to obtain a matching larger than M ′. In a component X of the
latter types we replace the M ′-edges of X with the M -edges of X: component
⊓⊔
X is replaced by mX many singletons.
X many M ′-edges of X with the m′
Local Approximation Ratios. To bound the approximation ratio
of MinGreedy, our approach is to bound local approximation ratios
α = M/M ∗
of components X of H. Observe that a 1
2 -path X
αX = mX /m∗
X ,
,
4
X = 2 edges of M ∗,
i.e. an augmenting path with mX = 1 edge of M and m∗
has local approximation ratio αX = 1
2 , while all other components have local
approximation ratios at least 2
3 . In particular, a singleton X has optimal local
approximation ratio αX = 1. Thus we have to balance local approximation
ratios. For any component X, we say that X has M -funds mX and we introduce
a change tX to the M -funds of X such that the changed local approximation
ratio of X is lower bounded by
αX =
mX + tX
m∗
X
!
≥ β
for appropriately chosen β. IfPX tX = 0 holds, then the total M -fundsPX mX +
tX = PX mX = M are unchanged and hence MinGreedy achieves approxi-
mation ratio at least
α = M/M ∗ = XX
mX + tX! /M ∗ ≥ XX
βm∗
X! /M ∗ = β .
w
Fig. 2. An augmenting path with transfers to its path endpoints and from its M -
covered nodes. Not all F -edges are used to move M -funds.
Transferring M -Funds. The idea is to transfer M -funds from rich com-
ponents to poor components: as an example, we could transfer M -funds from a
singleton to a 1
2 -path. But which components should be involved in a transfer?
Observe that an augmenting path endpoint w is detrimental to its augment-
ing path X, since w decreases the local approximation ratio of X. So a transfer
should push M -funds towards poor w from a rich M -covered node.
Since the G-neighbors of w are M -covered (otherwise M would not be maxi-
mal), our approach is to move M -funds to w from G-neighbors of w which belong
to other components: M -funds are moved over certain edges in
F := E \ (M ∪ M ∗)
which connect the components of H, see Figure 2 for an illustration. We verify
that a poor component, in particular a 1
2 -path, is able to receive M -funds:
5
Lemma 2. An augmenting path endpoint w is incident with F -edge, since the
degree of w in G is at least
dG(w) ≥ 2 .
Proof. When MinGreedy picks the first M -edge of the augmenting path of w,
a node of degree at least two is selected (with an incident M -edge and M ∗-edge).
So w also has degree at least two and is incident not only with its M ∗-edge. ⊓⊔
When transferring M -funds over all F -edges we face the danger of augment-
ing path endpoints having large degree and drawing very large amounts of M -
funds: rich components might become poor and now have too small local ap-
proximation ratio. The following definition prevents high degree path endpoints
from wasting M -funds.
Definition 1. Let w be an augmenting path endpoint. Edge {v, w} ∈ F is a
transfer if in the step of MinGreedy matching v the degree of w drops to at
most d(w) ≤ 1. A transfer moves one coin with value θ from the component of v
to the component of w.
We frequently denote a transfer {v, w} as (v, w) to stress its direction. In
order to refer to transfers from/to a given component X, we also call (v, w)
a debit from v and a credit to w. I.e., debits from a component are transfers
directed from its M -covered nodes, credits to an augmenting path are transfers
directed to its path endpoints.
Bounding Local Approximation Ratios. Let X be a component of H
and denote the numbers of debits and credits of nodes of X by dX respectively
cX . We call cX −dX the balance of X.1 Given an upper bound on the debits from
X and a lower bound on the credits to X, we obtain a bound on the balance of
X of at least
cX − dX ≥ TX .
Whenever we move M -funds over an edge {v, w} ∈ F , we transfer an amount θ
of M -funds to the augmenting path endpoint w using one coin. Hence the local
approximation ratio of X is at least
αX =
mX + θcX − θdX
m∗
X
≥
mX + θTX
m∗
X
.
In the analysis we find TX and θ such that αX ≥ β for all components. Hence
MinGreedy computes matchings of size at least β times optimal.
1 Previous versions of this work use the notion of total debits of a component X, which
are defined as dX −cX . This notion is equivalent to the notion of the balance of X.
6
2.1 Balance Bounds
For each path X we demand the following: both nodes of each M -edge {x, x′}
of X pay at most
coins and X receives at least
d{x,x′} ≤ 2(∆ − 2)
(−Edge)
cX ≥ 2
(+Endpoints)
coins. Therefore, the minimum balance of path X is 2− DX for DX := 2mX (∆−
2). In particular, bounds (−Edge) and (+Endpoints) are bounds on numbers of
coins and not on numbers of transfers: we demand that both hold w.r.t. transfers
as well as donations, the latter of which are another way to move coins that we
introduce later.
To show that all components have local approximation ratio at least ∆−1
2∆−3
we will verify balance bounds
−dX ≥ −2(∆ − 1) + 2
for any singleton X and (±Singleton)
(±Path)
cX − dX ≥ 2 − DX
Lemma 3. If balance bounds (±Singleton) and (±Path) hold then all compo-
nents have local approximation ratio at least ∆−1
2∆−3 .
+ 2(∆ − 2)
for any path X.
Proof. Choose θ =
of a singleton X is at least 1+θ(cX −dX )
1
2(2∆−3) . Using (±Singleton), the local approximation ratio
≥ 1 − 2θ(∆ − 2), which in turn equals
2(∆ − 2)
2(2∆ − 3)
∆ − 1
2∆ − 3
1−2θ(∆ − 2) = 1−
(1)
1
=
.
Therefore each singleton has local approximation ratio at least ∆−1
2∆−3 .
Using (±Path), we get that the local approximation ratio of a path X is at
least
mX + θ(cX − dX )
m∗
X
≥
mX − 2θmX(∆ − 2) + 2κ(∆ − 1)
mX + 1
= 1 − 2θ(∆ − 2) +
= 1 − 2θ(∆ − 2) ,
2θ(2∆ − 3) − 1
mX + 1
which by (1) is bounded by at least ∆−1
proximation ratio at least ∆−1
2∆−3 .
2∆−3 as well, i.e. all paths have local ap-
⊓⊔
7
2.2 Organisation of the Proof of (±Singleton) and (±Path)
Consider the 1-2-MinGreedy algorithm given in Figure 3. Observe that a se-
quence of edges picked by the (deterministic) MinGreedy algorithm is a valid
sequence of edges to be picked by 1-2-MinGreedy. Therefore 1-2-MinGreedy
achieves no better worst case approximation ratio than MinGreedy and obtain-
ing performance guarantees for 1-2-MinGreedy implies the same approxima-
tion performance for MinGreedy.
Note that 1-2-MinGreedy can also be understood as a refined variant of the
(deterministic) KarpSipser algorithm, which selects an arbitrary edge when-
ever all degrees are at least two and otherwise selects an arbitrary node of min-
imum degree, i.e. of degree one, and matches it with an arbitrary neighbor.
M = ∅
repeat until all edges removed from input graph:
if each node has degree at least 3:
select (arbitrary) edge {u, v}
else:
select (arbitrary) node u of minimum non-zero degree
select (arbitrary) neighbor v of u
pick edge {u, v}, i.e. set M = M ∪ {{u, v}}
remove all edges incident with u and v from input graph
return M
Fig. 3. The (deterministic) 1-2-MinGreedy algorithm
We prepare the proof with some basic observations in Section 2.3. Then,
for 1-2-MinGreedy we verify bounds (−Edge), (+Endpoints), (±Singleton),
and (±Path) for graphs of maximum degree ∆ = 3 in Section 3, where our
argument will rely on transfers only. In Section 4 we analyze graphs of maximum
degree ∆ ≥ 4, where we will introduce donations as an additional way to move
coins.
2.3 Preparations
Since transfers move coins from M -covered nodes over F -edges to adjacent path
endpoints, the following bounds on the number of outgoing and incoming trans-
fers are an immediate consequence.
In particular, the following bounds imply (−Edge) as well as (+Endpoints)
in case coins are moved only in transfers but not in donations.
Lemma 4. Let an M -edge {u, v} be given.
If {u, v} is the edge of a singleton, then both nodes pay at most 2(∆ − 1) debits.
If {u, v} is a path M -edge, then both nodes pay at most 2(∆ − 2) debits.
8
Proof. Since only F -edges can be transfers, it suffices to bound the number
of F -edges incident with each of u and v. Each M -covered node of a singleton
is incident with at most ∆ − 1 many F -edges. Each M -covered node of a path
is incident with at most ∆ − 2 many F -edges.
⊓⊔
Lemma 5. Let w be an endpoint of path X. Endpoint w receives at least one
credit. Path X receives at least cX ≥ 2 credits.
Proof. Let dG(w) denote the degree of w in G and denote the current degree
of w by d(w). By Lemma 2 we have dG(w) ≥ 2.
when the degree of w drops from d(w) to 0, where d(w) ∈ {1, 2}.
obtain cw ≥ 1.
If d(w) = 1 holds, then either an F -edge incident with w gets removed and
we have cw ≥ 1, or the M ∗-edge of w gets removed. In the latter case the degree
of w dropped from dG(w) ≥ 2 to d(w) = 1 when an F -edge incident with w was
removed, and again we have cw ≥ 1.
Finally, observe that since path X has two endpoints, path X receives at
⊓⊔
least cX ≥ 2 credits.
Since w is a path endpoint, node w never gets matched. Consider the step
If d(w) = 2 holds, then an F -edge incident with w gets removed and we
3 Maximum Degree ∆ = 3
In this section we show
Theorem 1. The 1-2-MinGreedy algorithm achieves approximation ratio at
least ∆−1
3 for graphs of maximum degree ∆ = 3.
2∆−3 = 2
−dX ≥ −2
In the proof we verify (±Singleton) and (±Path), which for ∆ = 3 reduce to
(±Singleton3)
(±Path3)
Coins will be moved using transfers only. Therefore (−Edge) and (+Endpoints)
hold: in case only transfers are used, we have already verified both bounds in
Lemmas 4 and 5.
for each singleton X and
for each path X.
cX − dX ≥ 2 − 2mX + 2
3.1 Two Debits are Missing
Recall that the minimum balance of a singleton X is −2(∆ − 1)= − 4, i.e. sin-
gleton X pays at most four debits. Our plan to verify (±Singleton3) is to show
that X pays two debits less than maximum: we say that two debits are missing.
Similarly, the minimum balance of a path X is 2− DX = 2− 2mX, where DX
is the maximum number of debits payed by X. Again, if X has two missing debits
then X is balanced as required in (±Path3). However, it might be the case that X
has less than two missing debits. In this case the balance of X will be increased
by additional credits.
9
Organization of the Proof. We start with the proof of two missing debits for
singletons in Lemma 6 (in Lemma 6 c), in particular); this result concludes the
proof of (±Singleton3).
More generally, Lemma 6 also applies to paths and identifies those cases in
which paths have two missing debits as well. Lemma 6 also prepares the analysis
of the special case in which a path X has less than two missing debits: we will
argue that X has exactly one missing debit and in Lemma 7 in Section 3.2 we
prove that the extra debit is compensated by an additional credit to X. Hence
Lemmas 6 and 7 conclude the proof of (±Path3).
Lemma 6. Assume that degrees are bounded by at most ∆ = 3. Consider the
creation step of component X, and assume that the 1-2-MinGreedy algorithm
selects node u with current degree d(u). Component X has two missing debits if
a) we have d(u) = 1
b) we have d(u) = ∆ = 3
c) component X is a singleton.
Proof. Let X be created in step s when u is matched with, say, node v. If
neither u nor v pays a debit, then two debits are missing and we are done.
Hence in the rest of the proof we may assume that at least one of u and v pays
a debit, which we denote by (x, w) for x∈{u, v}.
We prove a). If d(u) = 1 holds at step s, then no F -edges are incident with u.
Hence u has two missing debits by Definition 1.
We prove b). Since (x, w) is a transfer, step s removes edge {x, w} from G and
after step s the degree of w is d′(w) ≤ 1. Observe that in step s at most two edges
incident with w are removed. Since the degree of w is at least d(w) ≥ d(u) = 3
before step s, we have d′(w) = 1 at step s + 1. But since endpoint w is never
selected, another degree-1 node y 6= w is selected next in step s + 1. Observe
that the degree of y also drops from d(y) = 3 to d′(y) = 1 in step s, namely
when incident edges {u, y} and {v, y} are removed.
-- If y belongs to a component Y 6= X other than X, then both {u, y}
and {v, y} are F -edges. But {u, y} and {v, y} are not debits, since u, v and y
are M -covered. Hence we have found a missing debit for each of u and v.
-- If y is a node of X (note that in this case X must be a path), then
nodes u, v, and y form a triangle and one of {u, y} and {v, y}, say {u, y}, is
an F -edge. Since both u and y are M -covered, edge {u, y} is not a transfer
and hence both u and y have a missing debit.
We prove c). We may assume that d(u) = 2 holds at creation of singleton X,
since a) and b) apply in particular to singletons. Observe that a debit is missing
from node u. In order to obtain a contradiction, we assume that this is the only
missing debit from X. Consequently, node u pays exactly one debit, say to wu,
and v pays exactly two debits, say to wv and w′
v.
10
Since at creation of X node u has degree d(u) = 2 and is adjacent to v,
v. Since edges {u, wu}, {v, wv},
v} are transfers, Definition 1 implies that the degrees of wu, wv and, w′
node u is adjacent to at most one of wv and w′
and {v, w′
are at most 1 after step s. We distinguish two cases.
v
-- If u is adjacent to neither wv nor w′
v, then wu, wv, and w′
v have de-
gree exactly 1 after step s: their degrees were at least the minimum degree
of d(u) = 2 before {u, v} was picked, and their degrees dropped to at most
one afterwards.
v, say wu = wv holds.
v drops by at most one when edge {u, v} is picked, and
-- Now consider the case that u is adjacent to wv or w′
Then the degree of w′
hence w′
v has degree exactly 1 afterwards.
In both cases, no other degrees than those of wu, wv, and w′
v dropped in step
s. Since in step s+1 one of these endpoints has degree exactly one, one of wu, wv,
or w′
v is selected and matched next. A contradiction, since path endpoints are
⊓⊔
never matched by 1-2-MinGreedy.
3.2 One Missing Debit and an Additional Credit
To finish the proof of Theorem 1 it remains to verify (±Path3) for any given
path X. Recall that the balance of X is at least cX − dX ≥ 2 − DX = 2 − 2mX .
≥ 2 − 2mX + 2 it suffices to show that
To prove a balance of at least cX − dX
two debits are missing for X.
!
Since by Lemma 2 path X is created in a step selecting a node u of cur-
rent degree d(u) ≥ 2, we analyze the cases that in the creation step of X we
have d(u) = 2 or d(u) = 3.
In the latter case, two debits are missing by Lemma 6 b). Hence, for the rest
of the proof we may assume that d(u) = 2 holds. Observe that no F -edge is
incident with u when being selected, hence a debit is missing for u.
If an additional debit is missing for X, then we are done once again. So assume
from here on that the only missing debit for X is missing for u. Then dX =
DX − 1 = 2mX − 1 holds. Rather than proving a contradiction, in Lemma 7 we
show that X receives at least three credits: we obtain cX − dX ≥ 3 − 2mX + 1,
as required in (±Path3).
Lemma 7. Assume that degrees are bounded by at most ∆ = 3. Let X be a path
which pays dX = 2mX − 1 debits. Then X receives at least cX ≥ 3 credits.
Proof. We assume that cX < 3 holds and show a contradiction.
By assumption and as a consequence of Lemma 5, each endpoint of X receives
exactly one credit. Moreover, we have already argued before Lemma 7 that the
node u selected to create X has degree d(u) = 2 at creation, and that u has
the only missing debit of X. Consequently, all other M -covered nodes of X have
exactly one debit, since dX = 2mX − 1 holds. Let u be matched with node v.
11
First we consider the case that X is a 1
2 -path with mX = 1. Let wu and wv
be the endpoints of X such that we have {u, wu},{v, wv} ∈ M ∗. Note that
this implies wu 6= wv. Node v pays the only debit, say to the path endpoint y.
Observe that in the creation step of X the degrees of wu, wv, and y all drop, and
no other degrees drop.
If the degree of wv is still at least two after X was created, then wv receives
two credits over two still incident F -edges. A contradiction to our assumption
that cX < 3 holds.
Hence we may assume from here on that after creation of X the degree
of wv is at most d′(w) ≤ 1. But when {u, v} is picked we have d(u) = 2, i.e.
endpoint wv is not incident with u and the degree of wv drops by at most one.
Consequently, endpoint wv has degree exactly d′(w) = 1 after creation, which
is the new minimum degree in the graph. But since only the degrees of wu, wv,
and y dropped in the creation step, one of wu, wv and y is matched in the next
step. A contradiction is obtained since path endpoints are never matched. The
possible configurations are depicted in Figure 4.
y
v
wu
u
wv
wu
u
v
wv
Fig. 4. Creating a 1
2 -path which pays one debit and receives exactly two credits:
dashed endpoints have degree one after u and v are matched (not all edges are drawn)
Now consider the case that path X has mX ≥ 2 edges of M . After edge {u, v}
is picked in the creation step of X, at least one path endpoint w of X is still
connected with its unique M ∗-neighbor, call it x. We consider the step when x
becomes matched, say with its M -neighbor x′. In this step, we denote the current
degree of a node z as d(z).
Recall that x pays a debit, say to endpoint y. Since w and y are endpoints
and are thus never matched, both w and y are not yet isolated. Thus node x is
still connected with x′, w, and y and has degree at least d(x) ≥ 3.
Also, we may assume that w has degree at most d(w) ≤ 2. To see this, assume
that the degree of w is still ∆ = 3 and observe that the two F -edges incident
with w become credits, which contradicts our assumption that cX < 3 holds.
Moreover, we argue that the degree of w is exactly d(w) = 2. Why? Recall
that u has the only missing debit from X. Hence x′ pays a debit, i.e. node x′
is adjacent to an endpoint of an path. Since x′ is also adjacent to x, node x′
has degree at least d(x′) ≥ 2. But the degree of x′ is not larger than the de-
gree d(w) of w, i.e. it is at most d(x′) ≤ d(w) ≤ 2. Therefore the degree of x′
is exactly d(x′) = 2. Consequently, endpoint w has degree exactly d(w) = 2 as
well.
12
All still present neighbors of x and x′ are endpoints of paths, and only their
degrees drop when edge {x′, x} is picked. The possible configurations are depicted
in Figure 5. We distinguish two cases.
-- If x′ is not adjacent to w, then the degree of w drops by exactly one to
exactly (the three leftmost configurations in Figure 5). Hence either w or
another path endpoint is selected in the next step. A contradiction, since a
path endpoint is never matched.
-- Lastly, assume that x′ and w are adjacent (the rightmost configuration
in Figure 5). Then w becomes isolated in that step. We consider the recip-
ient y of the debit from x. Since x′ is adjacent to w and x and has degree
exactly d(x′) = 2, endpoint y is not adjacent to x′. Therefore the degree of y
drops by exactly one. Since the degree of y drops from at least d(y) ≥ 2 to
at most 1, by Definition 1 we get that the degree of y is exactly 1 in the
next step. Furthermore, endpoint y is now the only degree-1 node. Hence y
is selected and matched next. A contradiction, since y is a path endpoint.
⊓⊔
w′
w′
w′
w′
x′
x
w
x′
x
w
x′
x
w
x′
x
w
Fig. 5. Removing the last M ∗-edge at an end of a path: dashed endpoints have degree
one after x and x′ are matched (not all edges are drawn)
4 Maximum Degree ∆ ≥ 4
In this section we show
Theorem 2. The 1-2-MinGreedy algorithm achieves approximation ratio at
least ∆−1
2∆−3 for graphs of maximum degree ∆ ≥ 4.
In the proof we have to verify bounds (−Edge), (+Endpoints), (±Singleton),
and (±Path). First, we recall these bounds in an overview of how we refine the
analysis applied for ∆ = 3. In particular, the overview is structured so as to
highlight the two main concepts used in our analysis.
-- Canceling Transfers
Since ∆ is now larger than 3, a path endpoint might receive more credits
than for ∆ = 3. However, extra transfers degrade the balance of their source
components.
To annihilate this drawback we cancel extra transfers: certain transfers as per
Definition 1 will no longer move coins. The discussion is found in Section 4.1.
13
In particular, we carefully cancel transfers such that Lemma 5 still applies:
each endpoint of a path X still receives at least one credit. Consequently,
bound (+Endpoints) of at least
cX ≥ 2
(+Endpoints)
credits to X will be satisfied.
-- Donations
We discuss how to make a path X balanced. Assuming that (+Endpoints)
and (−Edge) hold, i.e. that the minimum balance of X is cX − dX ≥ 2− DX,
observe that
cX − dX ≥ 2 − DX + 2(∆ − 2)
(±Path)
holds if X has at least 2(∆− 2) missing debits. However, we will see that the
number of missing debits from X might smaller, even as small as a constant.
To increase the balance of X we introduce a donation to X, which will
compensate excessively payed debits from X. Summing up, path X will
have to pay 2(∆ − 2) debits less than maximum, as desired. Like transfers,
donations are edges in F . The discussion is found in Section 4.2.
The balance of the source component Y of a donation is reduced, since coins
have to be payed. Is Y balanced? If Y is a singleton, then we show that both
nodes of Y pay at most
(±Singleton)
coins even if a donation must be payed. For a path Y we have to verify (−Edge)
for each M -edge {y, y′} of Y , i.e. we have to show that nodes y and y′ pay at
most
dY ≤ 2(∆ − 2)
coins even if a donation must be payed.
2(∆ − 2)
(−Edge)
Organization of the Proof. In Section 4.1 we introduce transfer cancellations
and verify bound (+Endpoints). Thereafter we discuss donations in Section 4.2.
In Section 4.3 we develop bounds on the number of coins payed by edges with
an outgoing donation. Finally, in Section 4.4 we show that all components are
balanced by proving (−Edge), (±Singleton), and (±Path).
4.1 Canceling Extra Transfers
Unlike for ∆ = 3 where each path endpoint has at most two incident F -edges and
therefore receives at most two credits, for ∆ ≥ 4 we will see that a path endpoint
might receive up to three credits. But extra credits decrease the balances of their
source components. In order to push up their balances, we show how to cancel
extra credits in Definition 2. Thereafter we verify (+Endpoints) in Lemma 9.
14
Example. We prepare an example of an endpoint w which receives an extra
credit. Assume that w receives k credits from nodes matched in steps 1, . . . , s:
in any step s + l with l ≥ 1 we say that w already receives k credits.
Consider Figure 6, where 1-2-MinGreedy begins by creating the ∆ − 5
many 1
2 -paths drawn above path X, then proceeds to pick the two edges of X in
steps ∆−4 and ∆−3, and eventually picks the two singletons drawn left and right
of X. Endpoint w receives three credits in total: in step ∆ − 3 node w already
receives two credits from the nodes matched to create path X, and afterwards w
receives an additional credit.
1
∆
−
5
. . .
w
2
−
∆
...
. . .
...
∆−4
∆−3
1
−
∆
X
Fig. 6. Endpoint w receives three credits from path X (M -edges are numbered to
indicate the order in which they are picked, where 1-2-MinGreedy selects fat nodes)
However, path X is not balanced. Why? Path X pays three debits to each of
the ∆−5 topmost 1
2 -paths, and receives only four credits from the two singletons.
Hence the balance of X is as small as cX − dX = 4 − 3(∆ − 5), whereas balance
at least cX − dX ≥ 2 − 2mX (∆ − 2) + 2(∆ − 2) = 2 − 2(∆ − 2) is required
by (±Path).
The Definition. Therefore we cancel an extra credit to w, thereby limiting the
number of credits to w to at most two (like for ∆ = 3). To prepare the definition
we first analyze the exact "configuration" of a path endpoint with more than two
credits.
15
Lemma 8. As a consequence of Definition 1 of transfers,
a) each path endpoint receives at most three credits, and
b) path endpoint w receives three credits if and only if the the following holds:
the degree of w drops
i. from d(w) = 3 to d(w) = 1 when two incident F -edges are removed and
ii. from d(w) = 1 to d(w) = 0 when the last incident F -edge is removed.
Proof. First we argue that an endpoint w receives less than three credits if there
is a step s when the current degree of w is d(w) = 2. By Definition 1, no F -
edge of w removed before step s is a credit to w. Moreover, in step s at most
two F -edges are still incident with w.
Consequently, if endpoint w receives at least three credits, then w never
has current degree d(w) = 2. This implies the following two facts. First, the
degree of w in G is dG(w) ≥ 3, since the degree of w in G is dG(w) ≥ 2 by
Lemma 2. Secondly, there is a step when the current degree of w is d(w) = 3:
the degree d(w) of w never equals two, hence we get that d(w) drops from larger
than d(w) > 2 to below d(w) < 2, which is only possible if d(w) drops from
exactly d(w) = 3, since in each step at most two edges incident with w are
removed.
a) In the step when d(w) = 3 holds, by Definition 1 of transfers endpoint w
still has zero credits. Hence only the remaining three incident edges can be
transfer F -edges. Thus w receives at most 3 credits.
b) In the step when d(w) = 3 holds, all remaining edges incident with w
must be F -edges, since otherwise w would receive less than three credits. The
statement follows, since we have already argued that w never has degree d(w)=2.
⊓⊔
If endpoint w receives the maximum of three credits, then we cancel the
"third" credit, i.e. the one coming from a node getting matched in the step when
the degree of w drops from one to zero, cf. Lemma 8 b) ii. (Observe that path X
in Figure 6 is now balanced, since the debits from its rightmost M -covered node
are canceled and therefore the balance cX − dX ≥ 4 − 2(∆ − 5) > 2 − 2(∆ − 2)
is now strictly larger than (±Path) requires.)
Definition 2. Let w be a path endpoint with current degree d(w) = 1 and one
incident F -edge {v, w}. If w already receives two credits, then we cancel trans-
fer (v, w), i.e. edge {v, w} does not move coins.
The final set of transfers is given by Definitions 1 and 2. We are now ready
to prove (+Endpoints).
Lemma 9. a) Each path endpoint w receives at most two credits. An endpoint
for which a credit was canceled receives exactly two credits.
b) Each path endpoint w receives at least one credit. Therefore (+Endpoints)
holds.
16
Proof. a) is a direct consequence of Lemma 8 and Definition 2, since w receives
at most three credits, and if so then the third credit is canceled. To prove b) we
only have to consider path endpoints with less than two credits. In particular,
we only have to study an endpoint w for which no credit was canceled, since
by a) only such an endpoint can have less than two credits. Now observe that
Lemma 5 applies to w, no matter if a credit was canceled for w or not: by
⊓⊔
Lemma 5 endpoint w receives at least one credit.
4.2 Moving Coins to Paths in Donations
We prepare the discussion and introduce some notation, which will also be used
in the remainder of Section 4. Let X be a path and consider the creation step
of X. We denote the current degree of a node x as d(x). The nodes matched
in the creation step are called u and v, where we assume that 1-2-MinGreedy
selects u and matches u with neighbor v. In the next step, we denote the current
degree of x as d′(x), and we call the matched nodes u′ and v′, where we assume
that u′ is selected and matched with neighbor v′.
Organization of this Section. We begin this section with an example of a path
which pays many debits from nodes u and v. Thereafter, we sketch that nodes u′
and v′ are good candidates to compensate for the excessively payed debits from u
and v. We conclude with the definition of donations -- which move coins over F -
edges -- and a discussion of donation steps, i.e. the steps matching source nodes
of donations.
Example. Recall that to verify (±Path) we would like to show that for each path
at least 2(∆ − 2) debits are missing. However, even after canceling transfers a
path might pay too many debits. In particular, the example in Figure 6 can be
changed slightly as follows, see Figure 7: like in Figure 6 the 1-2-MinGreedy
algorithm first creates the topmost ∆ − 5 many 1
2 -path end creates the left and
right singletons in the last two steps; however, in steps ∆ − 4 and ∆ − 3 the
algorithm creates a 1
2 -path X and another singleton. Path X receives cX = 4
credits and pays dX = 2(∆− 5) debits, i.e. its balance is cX − dX = 4− 2(∆− 5),
whereas balance at least cX − dX ≥ 2 − 2mX (∆ − 2) + 2(∆ − 2) = 2 is required
by (±Path).
Sketch. Hence we have to provide more coins to an unbalanced path X. Therefore
we focus on the creation step of X as well as on the following step. Recall
that (±Path) holds if nodes u and v pay k = 0 debits, since then at least 2(∆−2)
debits are missing in total. If nodes u and v pay
k > 0
coins over debits, then it will turn out that u′ and v′ have missing debits, which
can be used to compensate for excessively payed debits from u and v. Depending
on whether u′ and v′ belong to X or not we proceed as follows:
17
1
. . .
∆
−
5
2
−
∆
...
. . .
∆−4
.
.
.
∆−3
X
1
−
∆
Fig. 7. A 1
the order in which they are picked, where 1-2-MinGreedy selects fat nodes)
2 -path X which pays too many debits (M -edges are numbered to indicate
-- If u′ and v′ belong to X, then the numbers of missing debits for u, v, u′
and v′ will add up to at least 2(∆ − 2): the balance of X is cX − dX ≥
2 − DX + 2(∆ − 2), as required in (±Path).
-- If otherwise u′ and v′ belong to another component Y 6= X, then without Y
becoming unbalanced component Y will be able to donate k coins to u and v:
the k debits payed by u and v are compensated and again the balance of X
is increased by at least 2(∆ − 2) over 2 − DX.
We call the step in which u′ and v′ are matched the donation step of X.
Degree-1 Endpoints After Creation. We argue that debits are missing for u′
and v′. Recall that we need only consider the case that a debit leaves u or v, say
to endpoint w, since otherwise we have found sufficiently many missing debits
for (±Path) to hold. Definition 1 of transfers requires that w has degree at
most d′(w) ≤ 1 after the creation step of X. The key to our proof is:
endpoint w might have degree exactly d′(w) = 1 after creation of X.
Definition 3. Let X be a path being created when 1-2-MinGreedy selects
node u and matches it with v. Assume that u or v pays a debit, say to end-
point w. If after creation of X endpoint w has degree exactly d′(w) = 1, then we
say that a degree-1 endpoint exists after creation of X.
18
Assume that a degree-1 endpoint exists after creation of X, call it w. Since end-
point w will never be matched, the 1-2-MinGreedy algorithm selects another
degree-1 node, namely node u′, next. As desired, we have found missing debits
for u′: since u′ has degree d′(u′) = 1 when being selected, node u′ pays zero
debits by Definition 1.
In Lemma 10 we identify the "configurations" in which a degree-1 endpoint
exists after creation of a path. In particular, in Lemma 10 b) we identify the
exact configurations in which no degree-1 endpoint exists after creation.
Lemma 10. Consider the creation step of path X, where nodes u and v are
matched and 1-2-MinGreedy selects u with current degree d(u). Assume that
one of u or v pays a debit t, say to node w.
a) If d(u)≥3 holds, then a degree-1 endpoint exists after creation of X.
b) If d(u)=2 holds, then a degree-1 endpoint exists after creation of X, unless
the following holds (see Figure 8 for an illustration of this exception):
i. after creation of X endpoint w has degree d′(w) = 0,
ii. at creation of X endpoint w has degree d(w) = 2,
iii. no debit leaves u,
iv. debit t leaves v, i.e. we have t = (v, w),
v. we have {u, w} ∈ M ∗,
vi. at creation of X all other endpoints w1, . . . , wk neighboring v (i.e. we
have w /∈ {w1, . . . , wk}) have degree at least three, and
vii. node v pays no other debits besides t.
. . .
. . .
w1
. . .
wk
v
. . .
w
u
X
Fig. 8. No degree-1 endpoint exists after creation of path X (the M ∗-neighbor of v
could be one of the wi)
Proof. By Definition 1, endpoint w has degree at most d′(w) ≤ 1 in the step
after creation of X. Observe that before creation of X endpoint w is not isolated,
since t is an incoming transfer and hence edge t is still incident with w.
a) If d(u) ≥ 3 holds, then the degree of w drops from at least d(w) ≥ 3 to at
least d′(w) ≥ 1, since at most two edges incident with w are removed. Hence w
has degree exactly d′(w) = 1 after creation of X.
19
b) Now assume that d(u) = 2 holds. Since we have d′(w) ≤ 1 after creation
of X, to prove the statement we only need to study the case that no degree-1
endpoint exists after creation, i.e. that endpoint w has degree d′(w) = 0 after
creation. Observe that i. holds and it remains to verify ii. to vii.
ii. Endpoint w has degree d(w) = 2 at creation, since the degree is d(w) ≥ 2
by Lemma 2 and if the degree was at least d(w) ≥ 3 then it could not drop
to d′(w) = 0 after creation, as is fact by i.
iii. Since u has degree d(u) = 2 at creation of X, node u is not incident with
an F -edge, as would be required in order to pay a debit.
iv. Since debit t does not leave u by iii., debit t leaves v.
v. Since in the creation step of X the degree of w drops by two, an edge con-
necting w with each of u and v is removed, and one of both edges must be
the M ∗-edge incident with w. But we have t = (v, w), i.e. edge {v, w} is
an F -edge. We obtain {u, w} ∈ M ∗.
vi. Observe that for each wi we have d′(wi) ≥ 1 after creation of X, since before
creation of X we have d(wi) ≥ d(u) = 2 and wi is not connected with u.
Since we assumed that no degree-1 endpoint exists after creation of X and wi
is not isolated then, we obtain d′(wi) ≥ 2. But an edge incident with wi is
removed in the creation step of X. Hence we get d(wi) ≥ 3.
vii. This follows from vi. Why? If all wi have degree at least three at creation
of X, then no wi receives a transfer from v, since the degree of each wi drops
by exactly one (recall that u is adjacent only with v and w). Consequently,
⊓⊔
endpoint w receives the only transfer from v.
The Definition of Donations. Donations will only be needed to move coins to
paths for which a degree-1 endpoint exists after creation. Hence in the rest of
this section we focus on this case. (In Lemma 18 we analyze the case that no
degree-1 endpoint exists after creation.)
Assume that a degree-1 endpoint exists after creation of path X. Since u′ is
selected with current degree d′(u′) = 1 and hence pays zero debits, node u′ is a
preferred candidate to compensate the debits payed by u and v.
To prepare the definition of donations, in Lemma 11 we show that an edge e
connecting u′ with u or v is removed in the creation step of X. In particular, if
a donation needs to be given to X from another component Y 6= X, then e is
an edge in F .
Lemma 11. Assume that a degree-1 endpoint exists after creation of path X.
The node u′ selected next by 1-2-MinGreedy is
a) an M ∗-neighbor of u or v in X or
b) an F -neighbor of u or v in a component Y 6= X.
Proof. After creation of X let w be a degree-1 endpoint. Since at creation of X
all degrees are at least two, thereafter both degree-1 nodes w and u′ lost an edge
connecting to u or v. Hence node u′ was connected to u or v by an M ∗-edge or
by an F -edge.
If u′ was connected by an M ∗-edge, then node u′ belongs to X. This proves a).
20
To prove b), assume that u′ is not an M ∗-neighbor of u or v. Then u′ was
connected only by an F -edge with u or v. It remains to show that u′ is not a
node of X. To see this, observe that after creation of X all M -covered nodes
of X (other than the M ∗-neighbors of u and v) have degree at least two and
recall that u′ is selected when it has degree d′(u′) = 1.
⊓⊔
Definition 4. Let nodes u and v be matched in the creation step of path X,
where 1-2-MinGreedy selects node u with current degree d(u). Assume that u
and v pay k > 0 debits. Let nodes u′ and v′ of component Y 6= X be matched
in the donation step of X, where 1-2-MinGreedy selects u′ with current de-
gree d′(u′)=1. Denote the F -edge connecting u′ with u or v as {u′, x} for x ∈
{u, v}.
-- If d(u) = 2 holds, then edge {u′, x} = {u′, v} is called a static donation and
-- If d(u) ≥ 3 holds, then edge {u′, x} is called a dynamic donation and moves k
moves ∆−3 coins to v.
coins to x.
Donation {u′, x} is denoted as a bold directed edge (u′, x) from the paying
node u′ of Y to the receiving node x of X.
Soundness of the Definition. Recall that our proof of (±Path) crucially relies on
bound (−Edge), i.e. on the (yet unproven) fact that each M -edge of path X pays
at most 2(∆−2) coins over outgoing debits and donations: in total, all M -edges
of X pay at most DX = mX · 2(∆ − 2) coins. Since the endpoints of X receive
at least cX ≥ 2 credits by (+Endpoints), the balance of X is at least cX − dX ≥
2 − DX . Our approach to use a donation to compensate all coins payed by
nodes u and v implies that the balance of X is increased by at least 2(∆ − 2):
the balance is at least cX − dX ≥ 2 − DX + 2(∆ − 2), as required in (±Path).
But to develop our Definition 4 of donations we only considered coins payed
by u and v over debits: if u or v pays a donation whose coins are not compensated,
then our approach to verify (±Path) fails. However, Lemma 12 a) implies that
neither u nor v pays a donation, i.e. our Definition 4 of donations is consistent
with our approach to analyze the creation step of X. Moreover, Lemma 12 b)
shows that coins donated to X need not be shared with another path.
Lemma 12. Let set C = {s1, s2, . . .} with s1 < s2 < . . . be the set of path
creation steps, and set D = {s1+1, s2+1, . . .} the set of donation steps.
a) We have C ∩ D = ∅.
b) The node selected in donation step si+1 donates coins to exactly one node,
which was matched in the creation step si.
Proof. a) Consider an arbitrary donation step si+1. Assume that 1-2-MinGreedy
selects node u′
i) = 1. How-
ever, Lemma 2 shows that in each path creation step sj a node uj of current
degree at least d(uj) ≥ 2 is selected. We get si+1 /∈ C and the statement follows.
i is selected with current degree d′(u′
i, and recall that u′
21
b) It suffices to show that si−1+1 < si holds, since then creation step si is
the only creation step between donation step si+1 and the previous donation
step si−1+1. The previous donation step si−1+1 selects a node of degree one,
whereas step si selects a node of degree at least two in order to create a path.
Therefore si−1+1 6= si holds. Using si−1 < si we obtain si−1+1 < si.
⊓⊔
4.3 Combined Coins of Transfers and Donations
In this section we prove the following result, which will be applied in Section 4.4
to verify bounds (±Path), (±Singleton), and (−Edge).
Lemma 13. Let X be a path for which a degree-1 endpoint exists after creation,
and assume that nodes u and v are matched in the creation step of X and nodes u′
and v′ are matched in the next step.
Denote by du,v the number of debits payed by u and v, and by dv′ the numbers
of debits payed by v′ (recall that u′ pays no debits since u′ is selected with current
degree one). There exist (non-negative) integers k and l with
such that
du,v ≤ k
and
dv′ ≤ l
k + l ≤ 2(∆ − 2)
holds. In particular, if u′ belongs to another component Y 6= X and pays a
donation to X, then we may choose k as the number of donated coins.
Organization of the Proof. To verify Lemma 13 we distinguish the following
cases. Assume that path X is created when u is selected with current de-
gree d(u)=2. Then we bound k + l whether u′ is a node of X or of another
component Y 6= X, where by Definition 4 the analysis of the latter case involves
the use of a static donation. Assuming that path X is created when u is selected
with current degree d(u) ≥ 3, then again we bound k + l whether u′ belongs
to X or not, where by Definition 4 in the latter case we have to take into account
a dynamic donation. Cases d(u) = 2 and d(u) ≥ 3 are analyzed in Lemma 14
resp. Lemma 15.
In the proof of Lemma 14 it suffices to analyze (the number of debits leaving)
each of nodes u, v, u′, and v′ individually. However, in order to show Lemma 15
we have to take into account that the bound l depends"dynamically" on the num-
ber k of debits payed by u and v. Consequently, the proof is more involved than
that of Lemma 14. In particular, we study the endpoints neighboring nodes u
and v as well as nodes u′ and v′ (in G). Therefore we introduce some helpful
notation in Definition 5.
22
Lemma 14. Lemma 13 holds in case 1-2-MinGreedy selects u with current
degree d(u) = 2. In particular, we have the following:
a) If u′ is an M ∗-neighbor of u or v, i.e. node u′ belongs to X, then we may
choose k and l such that we have du,v = k ≤ ∆−2 as well as dv′ = l ≤ ∆−2.
b) If u′ belongs to component Y 6= X, then we may choose k and l such that
we have du,v ≤ k = ∆ − 3 as well as dv′ = l ≤ ∆ − 1.
In both cases we have k + l ≤ 2(∆ − 2).
Proof. To prepare the proof, consider the creation step of X. Observe that u is
not incident with an F -edge, since we have d(u) = 2. Hence u pays zero debits.
Recall that also u′ pays no debits since 1-2-MinGreedy selects u′ with current
degree d′(u′) = 1.
We prove a). At creation of X node v is incident with at most ∆ − 2 edges
of F , since v is also incident with its M -edge and M ∗-edge. Hence v pays at
most ∆− 2 debits. We obtain du,v ≤ ∆− 2, since u pays no debits. Analogously,
after creation of X node v′ is incident with at most ∆ − 2 edges of F , since v′ is
also a node of path X. Here we get dv′ ≤ ∆−2. Now choose k = du,v and l = dv′ .
We prove b). Since u′ does not belong to X and is selected with degree d′(u′) =
1, an F -edge e incident with u′ is removed in the creation step of X. Edge e
connects u′ with v, since u is not incident with any F -edges when being selected
by 1-2-MinGreedy. In particular, edge e is a static donation from u′ to v by
Definition 4. But both u′ and v are M -covered, thus e is not a debit from v and
at most ∆− 3 debits leave v. We get du,v ≤ ∆− 3. Now observe that the M -edge
incident with v′ is not a debit from v′. Hence we have dv′ ≤ ∆− 1. Now choose k
⊓⊔
as the number of donated coins k = ∆ − 3 as well as l = dv′ .
Definition 5. Consider the creation step of path X, when 1-2-MinGreedy
selects node u with current degree d(u) ≥ 3 and matches u with v. We denote
-- by W the set of path endpoints w with current degree at least d(w) ≥ 3,
-- by W ⊆ W the set of W-endpoints neighboring u or v,
-- by Wδ = {w ∈ W : d′(w) = δ} the set of W -endpoints which have degree δ
-- by W≥3 = W \ (W1∪ W2) the set of W -endpoints which have degree at least 3
-- by W f
1 = {w ∈ W1 : {{w, u},{w, v}} ∩ F = f } the set of W1-endpoints
-- by E(W ) = { {x, y} ∈ E : x ∈ {u, v}, y ∈ W } the set of edges connecting
Note that W 1
1 form a partition of W1 and that W1, W2, and W≥3 form
a partition of W , i.e. sets W \ W, W≥3, W2, W 2
1 are pairwise disjoint.
Lemma 15. Lemma 13 holds in case 1-2-MinGreedy selects u with current
degree d(u) ≥ 3. In particular, we may choose k and l such that we have
connected to u and v by f edges of F for f ∈ {1, 2}, and
nodes u and v with W -endpoints.
after creation of X for δ ∈ {1, 2},
after creation of X,
1 and W 2
1 , and W 1
du,v = k = 2W 2
dv′ = l ≤ W 1
1 + W 1
1
1 + W2
and
(a)
(b)
23
as well as
1 + 2W 1
1 + W2
k + l ≤ 2W 2
≤ E(W )
≤ 2(∆ − 2) .
(c)
(d)
Proof. We prove (a). Therefore we show that nodes u and v pay du,v = 2W 2
1 +
W 1
1 debits. To prove that we may choose k = du,v we have to verify that if X
receives a donation then this donation moves du,v coins. By Definition 4, path X
receives a dynamic donation since u was selected with current degree d(u) ≥ 3.
Now observe that a dynamic donation moves exactly du,v coins.
To count the number of debits payed by u and v, we first study which
endpoints do not receive a transfer from u or v. First, recall that each end-
point w ∈ W has degree at least three before creation of X and observe that w
has degree at least one thereafter, since degrees drop by at most two when X is
created. Endpoints in W \ W , i.e. endpoints which are not adjacent to u or v,
as well as endpoints in W≥3 still have degree at least three after creation of X
and hence do not receive credits from u or v. Also, endpoints in W2 receive no
credits from u or v, since their degree is at least two after creation.
1 debits.
1 + W 1
To show (b), we prove that node v′ pays at most dv′ ≤ W 1
Hence only endpoints in W1 might receive credits from u or v, since their
degree is exactly one after creation. In the creation step of X, for each w ∈ W1,
either two incident F -edges are removed, i.e. we have w ∈ W 2
1 , or one incident F -
edge and an incident M ∗-edge is removed, i.e. we have w ∈ W 1
1 : since all these
removed F -edges are transfers, we get that nodes u and v pay exactly du,v =
2W 2
1 +W2 debits and
set l = dv′ . Which endpoints do not receive transfers from v′? (Recall that u′
pays no debits since u′ is selected when it has degree d′(u′) = 1.) Here, our
Definition 2 of transfer cancellations is crucial: observe that v′ does not pay
transfers to endpoints in W 2
1 , since each of these endpoints already receives two
credits from u and v and any additional credit from v′ is canceled by Definition 2.
Furthermore, endpoints in W \ W have degree at least three after creation of X
and do not receive transfers from v′: when u′ and v′ are matched, node u′
is selected with current degree d′(u′) = 1, thus the degree of each endpoint
in W \ W drops by at most one and to at least two. Analogously, no endpoint
in W≥3 receives a transfer from v′.
1 and W2, and (b) follows.
We note that, in particular, an endpoint w ∈ W 1
1 ∪ W2 receives less than two
credits from u or v, i.e. further credits are not canceled, and we have d′(w) ≤ 2,
i.e. the degree of w might drop to at most one when edge {v′, w} is removed.
We prove (c), i.e. that k + l is bounded from above by the number E(W ) of
edges connecting u or v with an endpoint before creation of X. First, we argue
that
Hence v′ might pay debits only to endpoints in W 1
W2 ≤ E(W ) − 2(W 1
1 + W 2
1 )
24
1 as well as for each w ∈ W 2
holds: for each w ∈ W 1
1 two edges connecting w
with u and v are removed when in the creation step of X the degree of w drops
from at least d(w) ≥ 3 to at most d′(w) ≤ 1; moreover, each endpoint in W2 is
connected with u or v by at least one of the remaining edges. Now we apply (a)
and (b) to obtain k + l ≤ 2W 2
We prove (d), i.e. that E(W ) is bound from above by 2(∆ − 2). To see this,
observe first that before creation of X two edges incident with u do not connect u
with an endpoint: edge {u, v} is an M -edge and edge {u, u′} is removed when
in the creation step of X the degree of u′ drops from at least d(u′) ≥ d(u) ≥ 3
to d′(u′) = 1. Analogously, edges {v, u} and {v, u′} do not connect v with an
endpoint. Consequently, at most ∆ − 2 edges connect each of u and v with an
⊓⊔
endpoint. The statement follows.
1 + 2W 1
1 + W2 ≤ E(W ), as claimed.
4.4 All Components Are Balanced
Recall that in order to prove Theorem 2 we have to verify (+Endpoints), (−Edge),
(±Path), and (±Singleton). Recall also that in Lemma 9 we have already shown
(+Endpoints), i.e. that each path receives at least two credits at its endpoints.
In Lemma 16 we show (±Singleton) for each singleton.
Also in Lemma 16, we show (−Edge) for path M -edges. Thereafter it remains
to prove (±Path) for each path. In particular, we prove (±Path) for each path
for which a degree-1 endpoint exists after creation as well as for each path for
which there does not exist a degree-1 endpoint after creation: these proofs are
provided in Lemma 17 resp. Lemma 18.
In the proof of (±Path) we may assume that the nodes u and v matched
to create a path X pay at least one debit, since otherwise at least 2(∆ − 2)
debits are missing for X and as required in (±Path) the balance of X is at
least cX − dX ≥ 2 − DX + 2(∆ − 2), by (+Endpoints) and (−Edge).
We begin with the proof of (−Edge) and (±Singleton).
Lemma 16. Bounds (−Edge) and (±Singleton) hold, since for each M -edge {y, y′}
of a path or a singleton, nodes y and y′ pay at most 2(∆ − 2) coins.
Proof. We have to verify that y and y′ pay at most 2(∆ − 2) coins whether a
donation must be payed or not. Assume that 1-2-MinGreedy selects node y
and matches y with neighbor y′. Hence a donation might be payed from y, but
not from y′.
Assume that y pays a donation. The following argument applies whether {y, y′}
is a path M -edge or a singleton. By Definition 4 of donations, edge {y, y′} is
picked in the donation step of a path X for which a degree-1 endpoint exists
after creation. By Lemma 13, we have k + l ≤ 2(∆ − 2) for the number k of
coins donated by y and the number l of debits payed by y′, i.e. both nodes pay
at most 2(∆ − 2) coins.
25
From here on we assume that nodes y and y′ pay only debits. If {y, y′} is a
path M -edge then observe that each of y and y′ is incident with at most ∆ − 2
edges of F and recall that only F -edges can be transfers. Thus y and y′ pay at
most 2(∆ − 2) coins.
If {y, y′} is the edge of a singleton Y , then each of y and y′ has at most ∆− 1
incident F -edges, i.e. we have dY ≤ 2(∆ − 1). In order to prove (±Singleton)
with balance at least −dY ≥ −2(∆ − 2) it suffices to show that at least two
debits are missing. Assume otherwise, i.e. that y and y′ pay at least 2(∆− 2) + 1
debits. Then each of y and y′ has degree at least ∆ − 1 when {y, y′} is picked,
since otherwise one of y and y′ would have at most ∆ − 3 incident F -edges
and hence at least two missing debits. But since both y and y′ have degree at
least ∆ − 1 ≥ 3 before creation and a debit leaves y or y′, say to endpoint w,
before creation endpoint w has degree at least d(w) ≥ 3 as well and we obtain
that d′(w) = 1 holds after creation. Hence a node other than w is selected next
after creation of Y , call it z. But z had degree at least 3 before creation of Y
as well, hence z must be a degree-1 node after creation: two edges connecting y
and y′ with z are removed from the graph in the creation step of Y . Both edges
are not transfers, since y, y′, and z are M -covered. Consequently, each of y and y′
⊓⊔
has a missing debit, a contradiction.
To complete the proof of Theorem 2, it remains to verify (±Path).
Lemma 17. Bound (±Path) holds for each path for which a degree-1 path end-
point exists after creation.
Proof. Let X be a path for which a degree-1 endpoint exists after creation.
Since X receives at least cX ≥ 2 credits by (+Endpoints) and each M -edge of X
pays at most 2(∆− 2) coins by (−Edge), to show (±Path) we have to prove that
the balance of X is increased above 2 − DX by at least 2(∆ − 2).
Let u′ be the node selected next after creation of X, and recall that u′ is
selected with degree d′(u) = 1 since there is a degree-1 endpoint after creation
of X. By Lemma 11, node u′ is the M ∗-neighbor of u or v, or u′ belongs to
another component Y 6= X. We distinguish these two cases.
Assume that u′ is the M ∗-neighbor of u or v, i.e. edge {u′, v′} is an M -edge
of X. Observe that by (−Edge) the number of coins payed by nodes of the X-
edges {u, v} and {u′, v′} is bounded from above by 4(∆ − 2). However, letting k
be the number of debits payed by nodes u and v, and letting l be the number of
debits payed by nodes u′ and v′, Lemma 13 shows that k + l ≤ 2(∆ − 2) holds,
i.e. edges {u, v} and {u′, v′} pay at least 2(∆ − 2) coins less than maximum.
Therefore we obtain dX ≤ DX − 2(∆ − 2) and hence the balance of X is at
least cX − dX ≥ 2 − DX + 2(∆ − 2).
Now assume that u′ belongs to another component Y 6= X. Hence X receives
a donation from u′. Here, Lemma 13 shows that the number k of coins received
by X over the donation satisfies k ≥ du,v for the number du,v of debits payed
by u and v. Since we have du,v − k ≤ 0, we obtain dX ≤ (mX − 1) · 2(∆ − 2) +
du,v−k ≤ (mX − 1) · 2(∆ − 2) = DX − 2(∆ − 2) and hence the balance of X is
at least cX − dX ≥ 2 − DX + 2(∆ − 2) again.
⊓⊔
26
No Degree-1 Endpoint Exists After Creation. By ¯X we denote path such that
after creation of ¯X there does not exist a degree-1 endpoint. Assume that
1-2-MinGreedy selects node ¯u to create ¯X and matches ¯u with ¯v. Lemma 10 iii.
shows that node ¯u pays no debits, Lemma 10 iv. and vii. show that node ¯v pays
exactly one debit. Therefore we can only bound the number of missing debits
from ¯u and ¯v by at least 2(∆ − 2) − 1. Thus the number of coins payed by ¯X is
at most
d ¯X ≤ D ¯X − ( 2(∆ − 2) − 1 ) ,
where D ¯X = m ¯X · 2(∆ − 2) is the maximum possible number of coins payed by
¯X, since each M -edge of ¯X pays at most 2(∆ − 2) coins by (−Edge).
Can we find an additional missing debit However, unlike in the case that a
degree-1 endpoint exists after creation, in order to find an additional missing
debit for ¯X we cannot rely on the analysis of a degree-1 node matched after
creation of ¯X. In particular, no coins are donated to ¯X.
Therefore, using (+Endpoints) we obtain that the balance of ¯X is at least
c ¯X − d ¯X ≥ 2 − (D ¯X − ( 2(∆ − 2) − 1 )) = 1 − D ¯X + 2(∆ − 2) ,
which fails to satisfy bound (±Path) by one coin. Thus, if we can find an addi-
tional credit or an additional missing debit, then ¯X is balanced and we are done.
Hence the following result completes the proof of Theorem 2.
Lemma 18. Bound (±Path) holds for each path ¯X for which there does not
exist a degree-1 path endpoint after creation, since ¯X receives c ¯X ≥ 3 credits or
an M -edge {¯x, ¯x′} of ¯X pays at most 2(∆ − 2) − 1 coins, where {¯x, ¯x′} 6= {¯u, ¯v}
is not the M -edge of ¯X picked in the creation step.
Proof. To prepare the proof, observe that to show c ¯X ≥ 3 it suffices to identify
an endpoint of ¯X which receives at least two credits, since by Lemma 5 the other
endpoint of ¯X receives an additional credit.
We distinguish cases by the number m ¯X of M -edges of ¯X.
Assume that ¯X is a 1
2 -path with m ¯X = 1. We show that an endpoint of ¯X
receives two credits, which proves c ¯X ≥ 3. First, we consider the creation step
of ¯X and argue that in the subsequent step there is an endpoint of ¯X which is
not isolated. Since after creation of ¯X there does not exist a degree-1 endpoint,
by Lemma 10 b) the 1-2-MinGreedy algorithm selects a node ¯u of current
degree exactly ¯d(¯u) = 2. Let ¯v denote the node matched with ¯u and observe
that the M ∗-neighbor of ¯v, call it ¯w, has degree at least ¯d( ¯w) ≥ ¯d(¯u) = 2 as
well. Since m ¯X = 1 holds, node ¯w must be an endpoint of ¯X, and since we
have ¯d(¯u) = 2 endpoint ¯w is not adjacent to ¯u. Thus the degree of ¯w drops by
at most one and to at least ¯d′( ¯w) ≥ 1 in the step after creation of ¯X.
Moreover, after creation of ¯X all non-isolated path endpoints have degree at
least two, since there does not exist a degree-1 endpoint. This holds in particular
for ¯w. But since after creation of ¯X endpoint ¯w is incident with at least two F -
edges, endpoint ¯w eventually receives at least two credits, namely over those
two F -edges of ¯w which are removed last from the graph.
27
From here on, assume that ¯X has m ¯X ≥ 2 edges of M . We again denote the
nodes matched to create ¯X by ¯u and ¯v, where we assume that 1-2-MinGreedy
selects ¯u. Recall that since no degree-1 endpoint exists after creation of ¯X, by
Lemma 10 b) an M ∗-edge of ¯X incident with an endpoint of ¯X is removed from
the graph.
We consider the step s when for the second time an M ∗-edge incident with
an endpoint of ¯X is removed from the graph. Denote by ¯x the selected node and
by ¯x′ the neighbor matched with ¯x. Since m¯x ≥ 2 holds we have {¯u, ¯v} 6= {¯x, ¯x′}.
We distinguish the cases that ¯x and ¯x′ pay only debits, or ¯x pays a static or a
dynamic donation.
First, assume that ¯x and ¯x′ pay no donation but only debits. To show that
edge {¯x, ¯x′} pays at most 2(∆ − 2) − 1 debits, assume the opposite. Then each
of ¯x and ¯x′ pays ∆−2 debits, since each is incident with at most ∆−2 edges of F .
Hence each of ¯x and ¯x′ has degree at least ∆− 1 at step s, since each is incident
with ∆ − 2 edges of F as well as with edge {¯x, ¯x′}. But a debit leaves ¯x or ¯x′,
say to endpoint w, hence by Definition 1 of transfers after step s endpoint w has
degree at most one. In particular, endpoint w has degree exactly one after step s,
since before step s node ¯x has degree at least ∆ − 1 ≥ 3 and hence endpoint w
has degree at least 3 as well.
Thus a node other than w is selected next, call it y. But node y had de-
gree at least 3 before step s as well and since thereafter endpoint w has degree
exactly one, node y is selected with current degree exactly one. Consequently,
edges {¯x, y} and {¯x′, y} are removed from the graph in step s. Both {¯x, y}
and {¯x′, y} are not transfers, since ¯x, ¯x′, and y are M -covered. Now observe that
at most one of {¯x, y} and {¯x′, y} can be an M ∗-edge, since otherwise two M ∗-
edges would be incident with y: at least one of {¯x, y} and {¯x′, y} is an F -edge,
say {¯x, y}. Since edge {¯x, y} is not a transfer, there is a debit missing for ¯x and
edge {¯x, ¯x′} pays at most 2(∆ − 2) − 1 debits, as claimed.
Now assume that ¯x pays a static donation. By Definition 4, path ¯X pays ∆−
3 coins over the donation from ¯x. Furthermore, node ¯x′ pays at most ∆−2 debits,
since at most ∆ − 2 many F -edges are incident with ¯x′ when 1-2-MinGreedy
picks edge {¯x, ¯x′}. Again, the nodes in edge {¯x, ¯x′} pay at most 2(∆ − 2) − 1
coins.
Lastly, assume that ¯x pays a dynamic donation (¯x, u) of k coins. By Def-
inition 4 of donations, node u belongs to a path X 6= ¯X for which a degree-1
endpoint exists after creation, and 1-2-MinGreedy selects u with current de-
gree d(u) ≥ 3 when creating X. Edge {¯x, ¯x′} is picked in the donation step of X
when ¯x has current degree d′(¯x) = 1.
Assume that u is matched with v, and recall the definition of the sets W, W1,
1 , W 2
W 1
1 , W2, and W≥3 of endpoints neighboring u and v, see Definition 5. By
Lemmas 13 and 15 we have the following: nodes u and v pay du,v = k = 2W 2
1 +
1 debits, node ¯x donates k coins to u, node ¯x′ pays at most d¯x′ = l ≤
W 1
W 1
1 + W2 debits. Hence ¯x and ¯x′ pay at most k + l ≤ 2(∆ − 2) coins in total.
If k + l < 2(∆ − 2) holds then an additional debit is missing from ¯x or ¯x′ and
we are done.
28
So assume from here on that ¯x and ¯x′ pay k + l = 2(∆ − 2) coins. As a
consequence of Lemma 15 (c) and (d) we get
1 + 2W 1
1 + W2 = E(W ) = 2(∆ − 2) .
k + l = 2W 2
1 + W 1
Since k = 2W 2
1 holds, we get that ¯x′ pays exactly
d¯x′ = l = W 1
1 + W2
debits. To which endpoints? Node ¯x′ pays zero debits to endpoints in W \ W
or W≥3: in the step when 1-2-MinGreedy picks edge {¯x, ¯x′} these endpoints
have degree at least three, and since these endpoints are not adjacent to ¯x
when ¯x is selected with current degree d′(¯x) = 1 in the donation step of X, the
degrees of these endpoints drop by at most one and to at least two, i.e. these
endpoints do not receive transfers from ¯x′. Therefore ¯x′ might only pay debits to
endpoints in W2 ∪ W1. But node ¯x′ also pays no debits to nodes in W 2
1 , since in
the step when 1-2-MinGreedy picks edge {¯x, ¯x′} each endpoint in W 2
1 already
receives two credits from nodes u and v and any further credits from ¯x′ are
canceled. Consequently, node ¯x′ might pay debits only to nodes in W2 or W 1
1 .
Since sets W \ W, W≥3, W2, W 2
1 are pairwise disjoint and since we
1 +W2, we obtain that ¯x′ pays exactly one debit to each endpoint
have d¯x′ = W 1
in W 1
1 as well as to each endpoint in W2.
1 , and W 1
In the rest of the proof we proceed as follows. We show that either an end-
point of ¯X receives at least two credits, which proves cX ≥ 3, or we show a
contradiction to our assumption that nodes ¯x and ¯x′ pay k + l = 2(∆− 2) coins,
which proves that edge {¯x, ¯x′} pays at most 2(∆ − 2) coins.
Consider the step when 1-2-MinGreedy picks edge {¯x, ¯x′}. Recall that ¯x
is selected with current degree d′(¯x) = 1, i.e. when the M ∗-edge of ¯x is already
removed from the graph. Recall also that an M ∗-edge incident with a path
endpoint of ¯X is removed in this step. This M ∗-edge must be incident with ¯x′
and we call it edge {¯x′, ¯w}.
1 , or W2 endpoint ¯w
belongs to. (Recall that sets W 1
1 form a partition of W1 and sets W1, W2,
and W≥3 form a partition of W .) Path X is created when the nodes u and v
being matched have degree at least d(u) ≥ 3 resp. d(v) ≥ 3, and after creation
of X the degree of ¯x is exactly d′(¯x) = 1. Thus two edges {u, ¯x} and {v, ¯x} are
removed from the graph in the creation step of X.
We conduct a case analysis based on which of W 1
1 and W 2
1 , W 2
¯w ∈ W \ W or ¯w ∈ W≥3 : Here we show that ¯w receives at least two cred-
its due to its large degree after creation of X.
Since X receives a dynamic donation, by Definition 4 of donations
the degrees of u and v are at least three in the creation step of X.
Consequently, the degree of ¯w is at least d( ¯w) ≥ 3 as well. If we
have ¯w ∈ W \ W , then ¯w is not adjacent to u or v and hence no
edges incident with ¯w are removed in the creation step of X. Thus
the degree of ¯w is at least d′(w) ≥ 3 after creation of X. If ¯w ∈ W≥3
29
holds, then the degree of ¯w is at least d′(w) ≥ 3 after creation of X
by Definition 5.
In the step after creation of X, i.e. in the donation step of X, when
1-2-MinGreedy picks edge {¯x, ¯x′}, the degree of endpoint ¯w drops
by at most one, since ¯w is not adjacent with node ¯x which is selected
with current degree d′(¯x) = 1.
Consider the step after edge {¯x, ¯x′} is picked. The degree of ¯w is at
least two and the M ∗-edge of ¯w is removed from the graph. Hence ¯w
is incident with at least two F -edges. Now observe that ¯w eventually
receives at least two credits, namely over the two of the F -edges
of ¯w which are removed last from the graph. Consequently, path ¯X
receives at least c ¯X ≥ 3 credits.
1 : We show a contradiction to our assumption that k + l = 2(∆ − 2)
holds.
Consider the step when path X is created. Recall by Definition 5
of W 1
1 , that at least two edges incident with ¯w are removed, since ¯w
has degree at least d( ¯w) ≥ 3 before and degree at most d′( ¯w) ≤
1 afterwards. Also by definition of W 1
1 , only one F -edge incident
with ¯w is removed, hence the M ∗-edge incident with ¯w is removed
as well. Consequently, endpoint ¯w belongs to X.
We obtain a contradiction, since endpoint ¯w belongs to path ¯X
and ¯X6=X holds by definition of the donation (¯x, u).
1 , two F -edges e1 and e2 incident with ¯w are
removed when path X is created. In particular, since after creation
of X the degree of ¯w is exactly one, edges e1 and e2 are credits
to ¯w and both credits are never canceled. Hence path ¯X receives
at least c ¯X ≥ 3 credits.
1 + W2 holds and that node ¯x′ pays a
¯w ∈ W2 : Recall that d¯x′ = l = W 1
debit to each endpoint in W 1
1 and to each endpoint in W2. Since we
have ¯w ∈ W2, node ¯x′ pays a debit to ¯w. But only F -edges can be
transfers, thus we have {¯x′, ¯w} ∈ F . A contradiction to endpoint ¯w
being the M ∗-neighbor of ¯x′, i.e. to {¯x′, ¯w} ∈ M ∗.
⊓⊔
1 : By Definition 5 of set W 2
¯w ∈ W 1
¯w ∈ W 2
5 Inapproximability Results
MinGreedy does not exploit knowledge gathered about the input in previous
steps: e.g. the neighbors of the selected node u are not remembered in order to
"explore" the neighborhood of u later. In a step of MinGreedy, an arbitrary
node of minimum degree, who is located in an unknown place in the graph, is
matched to an arbitrary neighbor.
A question arises naturally: are the worst case performance guarantees given
above for MinGreedy optimal, i.e. is there a greedy matching algorithm which
always computes larger matchings than proven for MinGreedy? In particular,
a greedy matching algorithm in question may in each step utilize all previously
gathered knowledge in very sophisticated node and edge selection routines.
30
Adaptive priority algorithms [BNR02] in the vertex model
[DI04] define a
large class of deterministic greedy matching algorithms, which we denote as
APV. APV-algorithms do not have resource constraints and formalize the es-
sential properties of greedy algorithms: to what extent can the input be unveiled
in a single step, what are the possible irrevocable decisions for the constructed
solution to be done after part of the input is revealed? An APV-algorithm may
gather and process much data about its input instances and deduce knowledge
to be used in clever future steps. Therefore APV-algorithms seem much stronger
than MinGreedy. In particular, APV contains (the deterministic variants of)
many prominent greedy matching algorithms, see Lemma 19.
Nevertheless, we construct graphs with degrees bounded by ∆ for which a
matching of size at most ∆−1
2∆−3 + o(1) times optimal is computed. So our 2
3 lower
bound for ∆ = 3 and our ∆−1
2∆−3 lower bound for ∆-regular graphs are tight: the
very simple MinGreedy algorithm has optimal worst case performance among
APV-algorithms. For graphs of degree at most ∆ our ∆−1/2
2∆−2 lower bound shows
that MinGreedy has good worst case performance.
For an APV algorithm A the input graph is represented as a set of adjacency
lists, e.g. {hu; v, wi,hv; u, wi,hw; u, vi} is the triangle on u, v, w (where an arbi-
trarily ordered list of neighbors appears after a semicolon). An APV-algorithm A
has a priori access to the number of nodes and starts with an empty matching.
In each step, algorithm A selects a node by specifying a total priority order
on all possible adjacency lists. From the given order, algorithm A receives the
adjacency list which has highest priority and corresponds to a still non-isolated
node u, say hu; v, w, . . .i. (A node is called isolated, if it, or each of its neighbors,
is matched.) Lastly, algorithm A selects a non-isolated matching partner for u
from the neighbor set {v, w, . . .} and then changes to the next step.
Matched nodes are not removed from the adjacency lists in G: Observation 1
shows that if A remembers already matched nodes, then A can submit priority
orders on adjacency lists w.r.t. the "reduced" graph, i.e. w.r.t. the set of nodes
which are not yet isolated.
Observation 1. If an APV-algorithm receives a data item hu; v, w, x, . . .i, then
each neighbor v, w, x, . . . of u is either matched or not isolated.
Proof. Let y ∈ {v, w, x, . . .} and assume that y is not matched but isolated.
Each neighbor of y is matched and thus isolated. Since u is a neighbor of y and
⊓⊔
the data item of u is received, node u is not isolated. A contradiction.
Leaving adjacency lists unchanged increases the power of A: a neighbor of
an already matched node v may be requested and hence A is able to explore the
neighborhood of v and is not oblivious to the parts of G being processed.
31
Lemma 19. The class APV contains (all deterministic variants of ) Greedy,
KarpSipser, MRG, MinGreedy, Shuffle, and all vertex iterative algorithms.
Proof. We have to implement any deterministic variant of one of the given al-
gorithms as an APV-algorithm.
In a given round of Greedy, a priority order e1, e2, e3, . . . on edges has
to be built adaptively depending on previous computations. This list can be
built edge by edge like this: The sub-list e2, e3, . . . is built under the assumption
that e1 is not in the graph, the sub-list e3, . . . is built under the assumption that
both e1, e2 are not in the graph, etc. After Greedy adds the highest priority edge
in the graph to the solution, a new round starts and a new priority list is built
adaptively like above. We have to translate e1, e2, e3, . . .
into a priority order
on adjacency lists. Therefore edge ei = {u, v} is replaced by a list containing
an adjacency list hu; v, w, x, . . . i for each possible choice of w, x, . . . , making
sure that if ei is in the graph it is picked before ei+1. If Greedy receives the
adjacency list hu; v, w, x, . . .i, node u is matched with v.
The argument for the other algorithms in analogous. In each round, a priority
list is built recursively under the assumptions that high priority entries are not
in the graph, and then translated to a list of adjacency lists.
KarpSipser works like Greedy but prefers an edge incident with a degree-
1 node, if such an edge exists. So edges incident with degree-1 nodes are moved
to the front of the priority order, i.e. the priority order on adjacency lists starts
with adjacency lists hu; v, w, x, . . .i for all possible choices of u and v, w, x, . . .
where only v is not already matched.
To implement MRG, a priority list on nodes u1, u2, . . . has to be translated.
Node ui is replaced by a list of adjacency lists hui; v, w, x, . . .i for each possi-
ble choice of v, w, x, . . . If MRG receives an adjacency list, say for node u, an
arbitrary non-isolated neighbor of u is to be matched with u.
To implement MinGreedy, a priority list on node degrees 1, 2, . . . has to be
translated. Degree i is replaced by a list of adjacency lists hu; v1, . . . , vi, w, x, . . . i
for each possible combination of u and v1, . . . , vi, w, x, . . . where only v1, . . . , vi
are not already matched. As for MRG, for a received adjacency list, say for
node u, an arbitrary non-isolated neighbor of u is matched with u.
Shuffle does not compute priority lists in each round, but a node permuta-
tion u1, u2, . . . , un is computed once at the start, using the number n of nodes in
the graph. A node ui is replaced by a list of adjacency lists hui; v, w, x, . . . i for
each possible choice of v, w, x, . . . This priority order is used in each round. If
Shuffle receives an adjacency list for a node ui, then from the still non-isolated
neighbors of ui the first one in u1, u2, . . . , un is matched with ui.
A vertex iterative algorithm [GT12] considers nodes one at a time, and probes
each node u for neighbors. The probing for u ends after the first successful probe,
say for neighbor v. Then nodes u and v are matched. The probing for u also ends
after all possible neighbors have been tested without success. Furthermore, the
probing for u ends if the algorithm decides to stop probing for further neighbors
of u. Once the probing for u ends, node u is never considered again and u is
never probed as the neighbor of any other node considered later. In a round of
32
of the algorithm, we denote by (a, b), (a, c), . . . , (a, d), (e, f ), (e, g), . . . , (e, h), . . .
that node a is to be probed for neighbors b, c, . . . , d, then e is to be probed
for neighbors f, g, . . . , h, etc. A pair (u, v) is translated to a list of adjacency
lists hu; v, w, x, . . .i for each possible choice of w, x, . . . . If an adjacency list
hu; v, w, x, . . .i is received, then node u is matched with v.
⊓⊔
To proof that no APV-algorithm can guarantee approximation ratio bet-
ter than ∆−1
2∆−3 + o(1), we first present in Lemma 20 a construction for APV-
algorithms without access to the number of nodes in the graph, and then adapt
the construction in Theorem 3 to the full class of APV-algorithms.
Lemma 20. Let A be an APV-algorithm without access to the number of nodes.
There is an input graph of degree at most ∆ for which A computes a matching
of size at most ∆−1
2∆−3 times optimal.
Proof. We describe the construction of a hard input instance G for algorithm A
as a game played between A and an adversary B. As A unveils G only bit by
bit, adversary B may actually construct G on the fly, thereby reacting to the
various moves of A such that G has a much larger matching than the solution
of A. Of course, all adjacency lists presented by B during the whole game have
to be consistent with the final graph G constructed by B.
followed by the endgame, which has two steps.
The game consists of the regular game, which lasts for s = ∆ − 3 steps,
During the regular game, adversary B maintains the following invariant: each
node v that is not yet isolated has an adjacency list of one of the following types.
occur in a previously received adjacency list, and 3 ≤ d ≤ ∆.
Type 1: hv; v1, . . . , vdi where v and v1, . . . , vd are unknown, i.e. they did not
Type 2: hv; v1, v2i where v and v1, v2 are unknown.
Type 3: hv; v1, v2, v3i where v and v1, v2 are unknown and v3 is known, i.e. node
v3 was received by A in a previous adjacency list.
Observe, that all nodes in G have degree at least two.
Consider the very first step of A. Since all nodes are still unknown, all nodes
have adjacency lists of type 1 or 2. Hence the invariant holds. Consider step i
and assume that the invariant holds. Adversary B presents the highest ranked
adjacency list that is of type 1, 2 or 3. Call that adjacency list ai.
Case 1: ai = hv; v1, . . . , vdi is a type-1 adjacency list. Since all nodes in ai
are unknown, we may w.l.o.g. assume that A matches v with v1. Adversary B
constructs the connected component C depicted in Figure 9 which consists only
of nodes of types 1 and 2. All nodes of C are isolated in the next step, hence
the invariant is maintained. Observe that within C the maximum matching M ∗
scores two edges (the double edges {v, v2},{v1, vd} in Figure 9) whereas the
matching M computed by A scores just one edge (the crossed edge {v, v1}).
Since in Case 1 algorithm A requests only unknown nodes, adversary B is
able to trick A into unveiling part of G from which A cannot gather knowledge
about the rest of G. Can A act smarter? Assume that A has already received
33
v
v1
v2
v3
. . .
vd
l1
...
lk
m1
r1
u1
...
mk
rk
uk
a
c
r
e
t
n
e
c
b
d
Fig. 9. A connected com-
ponent of a hard instance
Fig. 10. The core of a hard instance (Gray nodes are
unknown, fat frontier nodes. The dashed edge is an ex-
ample for {mi = v, rj = v3} in case 3.)
the adjacency lists of the middle nodes m1, . . . , mk of the triangles {lj, mj, rj}
of known nodes connected by frontier nodes rj and unknown nodes uj to the
still unknown center of G, see Figure 10. If A requests an unknown node with
two unknown neighbors, then B easily tricks A by constructing a new triangle
{li, mi, ri}.
Case 2: ai = hv; v1, v2i is a type-2 adjacency list. Again, all nodes of ai are
unknown and we may assume that A matches v with v1. Adversary B constructs
a triangle {li, mi, ri} with li = v2, mi = v, ri = v1 and inserts the edge {ri, ui},
with a new unknown node ui, to connect the triangle to the unknown center.
Observe that before nodes mi, ri are matched, nodes mi, li are of type 2 and
ri, ui are of type 1. After matching mi, ri, nodes li, mi, ri are isolated and ui
turns into a type-3 node. Hence the invariant still holds. Again, M ∗ scores two
edges, namely {li, mi},{ri, ui}, and M scores the edge {mi, ri}.
Now assume that A tries to explore the neighborhood of known nodes. Ob-
serve that the only adjacency lists with a known node are of type 3 and have
exactly one unknown node: since the known nodes lj, mj, rj are already isolated,
an unknown node can only be explored in the neighborhood of frontier nodes.
Again, adversary B tricks A with a new triangle {li, mi, ri}.
Case 3: ai = hv; v1, v2, v3i is a type-3 adjacency list. Since v3 is known,
v3 occurred in a previously presented adjacency list. Observe that in our con-
struction so far, the only type-3 nodes are unknown neighbors of known frontier
nodes. So v is the neighbor of a frontier node rj = v3 with j < i.
Is algorithm successful in exploring the unknown neighbor uj of rj, i.e. does
v = uj hold? Not necessarily, since B may on the fly construct further neighbors
of rj . Why? Since rj gets matched as soon as it becomes known, algorithm A
never gets to see the adjacency list of rj and consequently A can never tell if it
already knows all neighbors of rj. (Adversary B uses this trick here as well as
in the end game.)
Since v3 = rj is matched and v1, v2 are unknown, we may assume that A
matches v with v1. Adversary B behaves exactly as in case 2 and constructs the
triangle {li = v2, mi = v, ri = v1} and inserts the edge {ri, ui} where ui is a
new unknown node. To complete the devious trick, adversary B also inserts the
edge {mi = v, rj = v3} (see e.g. the dashed edge in Figure 10). Before mi, ri are
matched, node li is of type 2, nodes ri, ui are of type 1 and mi = v is of type
34
3. After matching mi, ri, nodes li, mi, ri are isolated and ui turns into a type-3
node. uj is still of type 3. Hence the invariant still holds. As in case 2, M ∗ scores
{li, mi},{ri, ui} and M scores {mi, ri}.
This concludes the regular game. In the first step of the endgame adversary
B makes algorithm A match a with b. Hence in the next and last step algorithm
A matches c. So algorithm A scores two edges in the center, whereas three edges
are optimal. As desired, we get
M = s + 2 = ∆ − 1
and
M ∗ = 2s + 3 = 2∆ − 3 .
Observe that our invariant still holds in the first step of the endgame. Again,
adversary B presents the highest ranked adjacency list of type 1, 2 or 3. Observe
that a and c are the only type-1 nodes left, since the uj have known neighbors
and are of type 3 and all other nodes have degree two and are of type 2. The
degree of a and c is δ ≤ 3 + s, since both a, c have three center neighbors and
each step of the regular game adds at most one neighbor to a respectively c. Let
a∆−2 be the adjacency list received in step s + 1 = ∆ − 2.
Case 4a: a∆−2 = hv; v1, . . . , vδi is a type-1 adjacency list. Since all nodes
of a∆−2 are unknown we may assume that A matches v with v1. Adversary B
chooses v = a, v1 = b and v2, . . . , vδ as the remaining neighbors of a.
Case 4b: a∆−2 = hv; v1, v2i is a type-2 adjacency list. Since all nodes of
a∆−2 are unknown we may assume that A matches v with v1. Adversary B sets
v = b, v1 = a and v2 = d.
Case 4c: a∆−2 = hv; v1, v2, v3i is a type-3 adjacency list. As in case 3, the
known node v3 is some matched frontier node rj, j < ∆− 2 and we may assume
that A matches v with v1, since v1, v2 are unknown. As in case 3, adversary B
does not present the adjacency list of the unknown node uj. Instead, B makes
b a neighbor of rj by inserting {v3=rj , b} -- now b has three neighbors -- and sets
v=b, v1=a and v2=d.
Adversary B does not violate degree constraints. Nodes introduced in case 1
have degree at most d ≤ ∆. All other degrees are at most three, but for a, c and
frontier nodes rj. As discussed, nodes a, c have degree at most δ = 3 + s ≤ ∆.
Frontier node have degree at most 3 + (s− 1) + 1 = ∆, since in each but the first
step of the regular game and in step s+ 1 at most one incident edge is added. ⊓⊔
Theorem 3. Let A be an APV-algorithm. There is a graph G of degree at most
∆ for which A computes a matching of size at most ∆−1
2∆−3 + ε times optimal for
any ε > 0.
Proof. We modify the adversary B who constructs hard inputs in the proof
of Lemma 20. First, the modified adversary B′ announces the number t∆ of
nodes, where t is a large integer. Using so many nodes B′ constructs Ω(t) con-
nected components, each with O(∆) nodes and approximation ratio no better
than ∆−1
2∆−3 . A negligible portion of the graph might be solved optimally by A.
In the proof we frequently refer to the adjacency list types and cases found in
the proof of Lemma 20 (types 1, 2, and 3, and cases 1, 2, 3, and 4a-c).
35
Again, the game between A and B′ is split up into the regular game and the
endgame. Like the adversary B from the proof of Lemma 20, in each round of the
regular game the construction of B′ keeps up the invariant that all non-isolated
nodes in the graph have adjacency lists of type 1, 2 or, 3 and B′ returns the high-
est priority adjacency list having one of these types. However, depending on the
requests of A, not only one but several centers C1, C2, . . . might be constructed,
each in its own connected component of G. Each center Ci is defined as in the
proof of Lemma 20, with nodes ai, bi, ci, di and two more nodes unique to Ci,
see Figure 10. Each Ci will get attached to it a maximum number of triangles,
which are not connected to any other center. Thereafter, the nodes of Ci are
supposed to be matched in the same order as in the proof of Lemma 20, i.e.
when no more triangles are attached, nodes ai and bi are matched to each other
before ci is matched. Once ci is matched, all edges in the connected component
of Ci are removed.
Assume that B′ has already created the centers C1, . . . , Cl. Call Ci active
if ai is not yet matched with bi, and inactive otherwise. The construction will
ensure that C1, . . . , Cl−1 are inactive; Cl might still be active. (We note here that
after center Cl becomes inactive, there are nodes in the rest of the connected
component K of Cl which do not have adjacency lists of types 1, 2 or 3. However,
all these nodes are neighbors of cl and adversary B′ adds no more nodes to K.
Thus A scores at most one more edge in K. Therefore, the additional adjacency
list types do not have effect on the rest of the construction and we do not discuss
these types explicitly.)
Assume that in the next round A receives a type-2 adjacency list of a node
with two unknown neighbors. Assume that Cl is already inactive, then B′ creates
the next center Cl+1 and connects a new type-2 triangle to al+1, cl+1 as described
in case 2. If otherwise Cl is still active, let δ be the number of neighbors of al, cl
constructed so far and recall that we demand δ ≤ ∆. If δ < ∆, then B′ connects
a new type-2 triangle to al, cl as described in case 2. If δ = ∆, then B′ makes A
match al with bl as described in case 4b, thereby inactivating Cl.
Assume that in the next round A receives a type-3 adjacency list. By con-
struction, the received node is unknown and among its three neighbors there is
exactly one known node v3, where v3 = r is a frontier node r in the connected
component of the still active center Cl. Let δ be the number of neighbors of al, cl
constructed so far. If δ < ∆, then B′ connects a new type-3 triangle to al, cl as
described in case 3. If δ = ∆, then B′ makes A match al with bl as described in
case 4c, and inactivates Cl.
Assume that in the next round A receives a type-1 adjacency list. If the
degree of the received node is smaller than ∆, then B′ proceeds as in case 1 and
creates a new type-1 connected component. Now assume that the received node
has degree ∆. If nodes al, cl have degree less than ∆, then again B′ creates a
new type-1 connected component. If nodes al, cl have degree δ = ∆, then B′
makes A match al with bl as described in case 4a, and thereby inactivates Cl.
Why is Cl inactivated before B′ constructs the next active center? In type-2
and type-3 rounds an increasing number of triangles is connected to Cl until
36
the degrees of al, cl are ∆. (In intermediate type-1 rounds only type-1 connected
components are constructed.) Thereafter, Cl is inactivated in the first type-2 or
type-3 round or the first type-1 round in which an adjacency list of a degree-∆
node is received. (In intermediate type-1 rounds with nodes of degree less than ∆
only type-1 connected components are constructed.)
The endgame begins as soon as B′ has constructed k ≥ t∆ − 6∆ nodes.
Let ν = t∆ − k be the number of nodes still to be constructed. Observe that
6∆ ≥ ν ≥ 2∆ holds, since in each round no more than 4∆ additional nodes are
introduced (e.g. if ∆ = 3 holds and a new triangle is connected to a new center).
Since B′ has committed to a number of exactly t∆ nodes, in the first round of
the endgame B′ utilizes all remaining ν nodes to create additional connected
components Γ1, . . . , Γc, each being a complete bipartite graph with 2 nodes on
the right side and between 2 and ∆ (both 2 and ∆ included) nodes on the left.
The left sides are as large as possible such that c = O(1) is constant and all nodes
in the Γi have degree at least two. All nodes in Γ1, . . . , Γc are still unknown, in
particular all nodes have adjacency lists only of types 1 or 2. Since right sides
have two nodes, each Γi has at most two edges in a maximum matching, making
an additional constant number 2c of optimal edges in total. We assume that A
performs optimally in all Γi, thereby scoring 2c edges.
Also in the endgame, algorithm A matches still unmatched nodes in the
already inactive centers C1, . . . , Cl−1. We assume that A performs optimally
also in the connected component of the last center Cl.
What is the approximation ratio of A? Observe that each type-1 component
has O(∆) nodes. The same is true for each connected component with a (con-
stant size) center Ci, since O(∆) triangles (of constant size) are attached to Ci.
So as claimed, Ω(t) connected components are constructed. Recall that A scores
only one out of two edges in each type-1 connected component. On the other
hand, in the connected component of an inactive center A achieves approxima-
tion ratio exactly ∆−1
2∆−3 . In all other components, algorithm A even performs
optimally. Therefore, to bound the performance of A we may assume that no
type-1 components are constructed. Since only Cl might be active at the end
of the regular game, there are at least l − 1 = Ω(t) inactive centers, hence the
approximation ratio of A is at most
(l − 1) ·
(∆ − 1) + (2∆ − 3) + 2c
(l − 1) · (2∆ − 3) + (2∆ − 3) + 2c
,
since A performs optimally in Cl and in Γ1, . . . , Γc. Letting t → ∞ we get l → ∞
and this ratio is dominated by (l−1)·(∆−1)
⊓⊔
(l−1)·(2∆−3) . The statement follows.
6 Conclusion
We have analyzed the worst case approximation ratio of the well-known Min-
Greedy algorithm on graphs of bounded degree. Our performance guarantees
of 2
2∆−3 for ∆-regular graphs are
tight. In particular, MinGreedy is optimal in the large class of APV-algorithms,
3 for graphs of degree at most three and of ∆−1
37
which contains many prominent greedy matching algorithms. We also proved a
performance guarantee of ∆−1/2
2∆−2 for graphs of degree at most ∆, and we conjec-
ture that also in this case MinGreedy is optimal among APV-algorithms and
achieves a worst case approximation ratio of at least ∆−1
2∆−3 .
Our worst case performance guarantees are stronger than the best known
worst case bounds on the expected approximation ratio for the well-known
greedy matching algorithms Greedy, MRG and Shuffle, if degrees are small.
Open Questions. Is MinGreedy optimal among APV-algorithms on graphs
What bounds for MinGreedy can be shown for more restricted graph classes,
of degree at most ∆?
e.g. bipartite graphs?
Recall that the expected approximation ratio of the randomized MinGreedy
algorithm is 1
2 + o(1) w.h.p. on graphs of arbitrarily large degree. Does random-
ized MinGreedy have an expected approximation ratio strictly better than
∆−1
2∆−3 if degrees are bounded by ∆?
Acknowledgements. I thank Georg Schnitger for many helpful discussions.
References
ADFS95. Jonathan Aronson, Martin Dyer, Alan Frieze, and Stephen Suen. Ran-
domized greedy matching. ii. Random Structures & Algorithms, 6(1):55 -- 73,
1995.
BNR02. Allan Borodin, Morten N. Nielsen, and Charles Rackoff. (incremental) pri-
ority algorithms. In Proc. 13th SODA, pages 752 -- 761, 2002.
DF91.
CCWZ14. T.-H. Hubert Chan, Fei Chen, Xiaowei Wu, and Zhichao Zhao. Ranking on
arbitrary graphs: Rematch via continuous lp with monotone and boundary
condition constraints. In Chandra Chekuri, editor, SODA, pages 1112 -- 1122.
SIAM, 2014.
Martin E. Dyer and Alan M. Frieze. Randomized greedy matching. Random
Structures & Algorithms, 2(1):29 -- 46, 1991.
Sashka Davis and Russell Impagliazzo. Models of greedy algorithms for
graph problems. In Proc. 15th SODA, pages 381 -- 390, 2004.
Alan Frieze, A. J. Radcliffe, and Stephen Suen. Analysis of a simple greedy
matching algorithm on random cubic graphs. In Proc. 4th SODA, pages
341 -- 351, 1993.
Gagan Goel and Pushkar Tripathi. Matching with our eyes closed. In Proc.
53rd FOCS, pages 718 -- 727, 2012.
Richard M. Karp and Michael Sipser. Maximum matchings in sparse ran-
dom graphs. In Proc. 22nd FOCS, pages 364 -- 375, 1981.
FRS93.
DI04.
GT12.
KS81.
KVV90. Richard M. Karp, Umesh V. Vazirani, and Vijay V. Vazirani. An optimal
algorithm for on-line bipartite matching. In Proc. 22nd STOC, pages 352 --
358, 1990.
Jakob Magun. Greedy matching algorithms, an experimental study.
In
Proceedings of the 1st Workshop on Algorithm Engineering, volume 6, pages
22 -- 31, 1997.
Mag97.
38
MP97.
MS04.
MV80.
Pol12.
Tin84.
Vaz12.
Zevi Miller and Dan Pritikin. On randomized greedy matchings. Random
Struct. Algorithms, 10(3):353 -- 383, 1997.
Marcin Mucha and Piotr Sankowski. Maximum matchings via gaussian
elimination. In FOCS, pages 248 -- 255. IEEE Computer Society, 2004.
Silvio Micali and Vijay V. Vazirani. An o(sqrt(v) e) algorithm for find-
ing maximum matching in general graphs. In FOCS, pages 17 -- 27. IEEE
Computer Society, 1980.
Matthias Poloczek. Greedy algorithms for max sat and maximum matching
: their power and limitations. PhD thesis, Institut für Informatik, Goethe-
Universität Frankfurt am Main, 2012. http://d-nb.info/1036608425.
G. Tinhofer. A probabilistic analysis of some greedy cardinality matching
algorithms. Annals of Operations Research, 1(3):239 -- 254, 1984.
Vijay V. Vazirani. An improved definition of blossoms and a simpler proof
of the mv matching algorithm. CoRR, abs/1210.4594, 2012.
39
|
1209.5608 | 1 | 1209 | 2012-09-25T13:42:56 | Faster Deterministic Fully-Dynamic Graph Connectivity | [
"cs.DS",
"cs.DM"
] | We give new deterministic bounds for fully-dynamic graph connectivity. Our data structure supports updates (edge insertions/deletions) in $O(\log^2n/\log\log n)$ amortized time and connectivity queries in $O(\log n/\log\log n)$ worst-case time, where $n$ is the number of vertices of the graph. This improves the deterministic data structures of Holm, de Lichtenberg, and Thorup (STOC 1998, J.ACM 2001) and Thorup (STOC 2000) which both have $O(\log^2n)$ amortized update time and $O(\log n/\log\log n)$ worst-case query time. Our model of computation is the same as that of Thorup, i.e., a pointer machine with standard $AC^0$ instructions. | cs.DS | cs | Faster Deterministic Fully-Dynamic Graph Connectivity
Christian Wulff-Nilsen ∗
Abstract
We give new deterministic bounds for fully-dynamic graph connectivity. Our data
structure supports updates (edge insertions/deletions) in O(log2 n/ log log n) amortized
time and connectivity queries in O(log n/ log log n) worst-case time, where n is the number
of vertices of the graph. This improves the deterministic data structures of Holm, de
Lichtenberg, and Thorup (STOC 1998, J.ACM 2001) and Thorup (STOC 2000) which
both have O(log2 n) amortized update time and O(log n/ log log n) worst-case query time.
Our model of computation is the same as that of Thorup, i.e., a pointer machine with
standard AC 0 instructions.
2
1
0
2
p
e
S
5
2
]
S
D
.
s
c
[
1
v
8
0
6
5
.
9
0
2
1
:
v
i
X
r
a
∗Department
of
Computer
Science,
University
of
Copenhagen,
[email protected],
http://www.diku.dk/ e koolooz/.
1
1
Introduction
The dynamic graph connectivity problem is perhaps the most fundamental dynamic graph
problem and has received considerable attention from the algorithms community for decades.
The goal is to build an efficient data structure that supports one or more of the following
operations in a dynamic graph G:
• connected(u, v): determines whether vertices u and v are connected in G,
• insert(u, v): inserts edge (u, v) in G,
• delete(u, v): deletes edge (u, v) from G.
The fully-dynamic graph connectivity problem supports all three operations, whereas the
simpler decremental and incremental variants do not support insert and delete, respectively.
In the following, we refer to both insert and delete as update operations.
The first non-trivial data structure for fully-dynamic graph connectivity is due to Fred-
erickson [2] who showed how to support updates in O(√m) time and connectivity queries
in O(1) time, where m is the number of edges of the graph. Using a general sparsification
technique, Eppstein, Galil, Italiano, and Nissenzweig [1] improved update time to O(√n),
where n is the number of vertices. Both of these data structures are deterministic and the
time bounds are worst-case.
Henzinger and King [3] significantly improved update time to O(log3 n) with only a small
increase in query time to O(log n/ log log n). However, their bounds are randomized expected
and update time is amortized. Using a clever sampling technique, Henzinger and Thorup [4]
shaved off a factor of log n in the update time. A simple and elegant deterministic data
structure with the same bounds as in [4] was given by Holm, de Lichtenberg, and Thorup [5].
Its space requirement was later improved to linear by Thorup [10] who also gave a randomized
data structure with a faster update time of O(log n(log log n)3) and marginally slower query
time of O(log n/ log log log n).
A general cell-prove lower bound of Ω(log n) for fully-dynamic graph connectivity was
provided by Patra¸scu and Demaine [6]. Hence, the data structures above are near-optimal.
As shown by Tarjan [8], incremental connectivity is the union-find problem which can be
solved in O(α(m, n)) time over m updates. Thorup [9] gave an O(log n) bound for decremental
connectivity if the initial graph has Ω(n log5 n) edges. He also gave an O(1) bound if the initial
graph is dense.
Our contribution is a deterministic data structure that improves the update time of the
deterministic data structures in [5, 10] by a factor of log log n. We use several ingredients of
Thorup [10], including his structural forest (which we refer to as a cluster forest) as well as
lazy local trees and shortcuts both of which we modify to suit our needs. We also introduce
an additional system of shortcuts that allows us to more quickly walk up trees of the cluster
forest. Together, these changes and additions give an order log log n speed-up in the update
time.
Table 1 summarizes the results for fully-dynamic graph connectivity.
Our paper is organized as follows. In Section 2, we introduce basic definitions and notation.
Section 3 gives a simple data structure with O(log2 n) update time and O(log n) query time. It
is essentially the deterministic data structure of Thorup [10] but a slightly more minimalistic
variant that does not need to maintain spanning trees of connected components. In Section 4,
2
Update time
Query time
Type
Reference
O(√m)
O(√n)
O(log3 n)
O(log2 n)
O(log2 n)
O(1)
O(1)
deterministic; worst-case time
deterministic; worst-case time
O(log n/ log log n)
randomized; amortized update time
O(log n/ log log n)
randomized; amortized update time
[2]
[1]
[3]
[4]
O(log n/ log log n)
deterministic; amortized update time
[5, 10]
O(log n(log log n)3) O(log n/ log log log n)
randomized; amortized update time
[10]
O(log2 n/ log log n)
O(log n/ log log n)
deterministic; amortized update time This paper
Table 1: Performance of data structures for fully-dynamic graph connectivity.
we add two systems of shortcuts to our data structure that together improve both update
and query time by a factor of log log n. The simplification given in Section 3 is not important
in order to get our improvement in Section 4; indeed, our result can easily be extended to
maintain a spanning forest. However, we believe that our approach gives a slightly cleaner
analysis and it should give a small improvement in practice. For instance, Thorup's data
structure needs to maintain two types of bitmaps for edges, one for tree edges and one for
non-tree edges whereas our data structure only needs to maintain one type; see Section 3.5
for details. Finally, we give some concluding remarks in Section 5.
2 Preliminaries
Let log denote the base 2 logarithm. We assume the same model of computation as in [10],
i.e., a pointer machine with words (bitmaps) containing at least ⌊log n⌋ + 1 bits and with
the following standard AC 0 instructions: addition, subtraction, bitwise 'and', 'or', and 'not',
and bit shifts. Our data structure also needs to perform division x/y and multiplication xy
which are not AC 0 instructions. To handle this, we assume that y is a power of 2 so that a
bit shift operation can be used instead; we can always round y to the nearest such value and
the constant multiplicative error introduced will not affect correctness or running time. For
a bitmap m, we denote by m[i] the ith bit of m, i ≥ 0.
We let G denote the input graph and it is assumed to contain no edges initially. To distin-
guish between vertices of G and other vertices (such as those in trees of our data structure),
we refer to the latter as nodes. For a path P and nodes a, b ∈ P , P [a, b] is the subpath of P
between a and b. We abbreviate balanced binary search tree as BBST and depth-first search
as DFS.
3 A Simple Data Structure
We first give a simplified version of our data structure which is no better than the deterministic
data structures in [5] and [10]. In fact, it has a slower query time of O(log n). In Section 4,
we shall speed up both query and update time by a factor of log log n.
3
3.1 The cluster forest
As in [10], we assign to each edge e of G a level ℓ(e) between 0 and ℓmax = ⌊log n⌋ and for
0 ≤ i ≤ lmax, we denote by Gi the subgraph of G induced by edges e with ℓ(e) ≥ i. We refer
to the connected components of Gi as level i clusters or just clusters. The following invariant
will be maintained by our data structure:
Invariant: For each i, any level i cluster spans at most ⌊n/2i⌋ vertices.
The cluster forest of G is a forest C of rooted trees where each node u corresponds to a
cluster C(u). The level ℓ(u) of u is its depth in C (between 0 and ℓmax) and a level i node
corresponds to a level i cluster. In particular, roots of C correspond to level 0 clusters and
hence to the connected components of G0 = G. By the invariant, each leaf of C corresponds
to a vertex of G; we often identify the two and our data structure keeps bidirected pointers
between them. A node u at a level i < ℓmax has as children the level (i + 1) nodes v such that
C(v) ⊆ C(u).
number of leaves in the subtree of C rooted at u.
vertices of G spanned by C(u). This completes the description of the cluster forest.
Our data structure will maintain, for each node u of C, an integer n(u) denoting the
In other words, n(u) is the number of
Given C, we can determine whether two vertices u and v are connected in G in O(log n)
time as follows. Traverse paths from u and v to roots ru and rv of trees of C containing u and
v, respectively. Then u and v are connected iff ru = rv.
3.2 Handling insertions
We need to maintain C as G changes. First we describe how to update C after an operation
insert(u, v). We initialize ℓ(u, v) ← 0. Letting ru and rv be defined as above, if ru = rv,
no update of C is required since u and v were already connected in G = G0. If ru 6= rv, we
update C by merging ru and rv into ru, meaning that ru inherits the children of rv, n(ru) is
increased by n(rv), and rv is deleted. This update corresponds to merging C(ru) and C(rv).
Thus C is correctly updated and the invariant still holds.
3.3 Handling deletions
Now consider the update delete(u, v). Let i = ℓ(u, v) and let Cu and Cv be the level
(i + 1) clusters containing u and v, respectively. Assume that Cu 6= Cv since otherwise, the
connectivity in Gi is not affected (there is a uv-path in Gi+1 ⊂ Gi connecting u and v). Let
Mi be the multigraph with level (i + 1) clusters as vertices and with level i-edges of G as
edges (so an edge of Mi connects two vertices if that edge connects the corresponding level
(i + 1) clusters in Gi). Our algorithm will not actually keep Mi but it will help to simplify
the description in this subsection.
We now execute two standard search procedures (say, DFS) in Mi, one, Pu, starting from
vertex Cu and another, Pv, starting from Cv. The two procedures are executed in "parallel"
by alternating between the two, i.e., one unit of time is spent on Pu, then one unit on Pv,
then one unit on Pu, and so on. We terminate both procedures as soon as we are in one of
the following two cases (which must happen at some point):
1. a vertex of Mi is explored by both procedures,
2. one of the procedures has no more edges to explore and we are not in case 1.
4
In the following, we show how to deal with these two cases.
n(C) and nv = PC∈Cv
Case 1: Let Cuv be the vertex (level (i + 1) cluster) of Mi explored by both procedures.
Assume w.l.o.g. that Pv was the last to explore Cuv. Let Cu be the set of level (i + 1) clusters
explored by Pu and let Cv be the set of level (i+1) clusters explored by Pv, excluding Cuv. If we
n(C) then since Cu ∩Cv = ∅, we have nu + nv ≤ ⌊n/2i⌋
let nu = PC∈Cu
by our invariant and thus min{nu, nv} ≤ ⌊n/2i+1⌋.
Assume first that nu ≤ nv. Then we can increase the level of every edge explored by
Pu from i to i + 1 without violating the invariant. To see this, note that the level updates
correspond to merging clusters of Cu into one level (i + 1) cluster spanning nu ≤ ⌊n/2i+1⌋
vertices. The idea is that the search performed by Pu is paid for by these level increases. As
Pv spent the same amount of time as Pu (up to an additive O(1) term), the level increases
also pay for the search by Pv.
We need to update cluster forest C accordingly. When increasing the level of an edge e
from i to i + 1, we identify the level (i + 1)-ancestors a and b of the endpoints of e. Clusters
C(a) and C(b) need to be merged and we do this by merging a and b into b and updating
n(a) ← n(a) + n(b). As we will see later, this update can also be paid for by the level increase
of e. Note that the procedures have found a replacement path in Gi for deleted edge (u, v)
so no further updates are required in C, and we terminate.
We assumed above that nu ≤ nv. We do exactly the same for clusters in Cv when nv ≤ nu
except that we do not increase the level of the last edge explored by Pv as it connects to a
cluster in Cu. If this was the only edge explored, there are no edges to pay for it but in this
case we have found a replacement path for edge (u, v) and the delete(u, v)-operation can
pay.
Case 2: Now assume that one of the procedures, say Pu, explores all edges in the connected
component of Mi containing Cu and that we are not in case 1. Let us assume that nu ≤
⌊n/2i+1⌋; if not, we fully explore the connected component of Mi containing Cv and update
nv which will be at most ⌊n/2i+1⌋ by our invariant; the description below then applies if we
swap the roles of u and v.
We can conclude that no replacement path for (u, v) exists in Gi. All edges explored by
Pu have their level increased to i + 1 and we update C accordingly by merging all level (i + 1)
nodes explored by Pu into one, w, and setting n(w) to the sum of n(w′) for nodes w′ explored
by Pu. These level increases pay for the two searches. Since Cu and the component of Mi
containing Cv are no longer connected in Gi, we further update C as follows:
let p be the
parent of w in C. We remove w as a child of p, decrease n(p) by n(w), add w as a child of a
new level i node p′, set n(p′) = n(w), and add p′ as a child of the parent of p. This correctly
updates C and the invariant is maintained.
If i > 0, it may still be possible to reconnect u and v in Gj for some j < i. We thus
execute the above algorithm recursively with i ← i − 1, Cu ← C(p′), and Cv ← C(p). Should
we end up in case 2 with i = 0, no replacement path in G between u and v could be found.
Then p′ becomes a new root of C and we terminate.
3.4 Local trees
We now extend our data structure to allow the search procedures to explore edges and vertices
of multigraph Mi in a more efficient way.
5
First we shall convert C into a forest of binary trees by adding local trees as in [10]. Let u
be a non-leaf node of C. We form a binary tree L(u) with u as root and with the children of
u as leaves as follows. First assign a rank rank(v) ← ⌊log n(v)⌋ to each child v of u. Initially,
each such v is regarded as a tree consisting just of v. While there are trees T and T ′ whose
roots r and r′ have the same rank, we pair them by attaching r and r′ to a new root r′′ with
rank(r′′) ← rank(r) + 1. We end up with at most log n trees T1, . . . , Tk, called rank trees,
whose roots r1, . . . , rk have pairwise distinct ranks: rank(r1) > rank(r2) > ··· > rank(rk).
We connect the rank trees into a single local tree L(u) rooted at u by adding a rank path
v1v2 . . . vk−1 down from v1 = u and connecting ri as a child to vi for i = 1, . . . , k − 1 and rk
as a child to vk−1. We define rank(u) ← ⌊log n(u)⌋.
The edges in C from u to its children are replaced by local tree L(u); let CL be C after all
local trees have been added. As shown by Thorup, for a child v of a node u in C, the depth of
v in L(u) is at most log(n(u)/n(v)) + 1. Since any leaf of C has depth at most ℓmax = ⌊log n⌋,
a telescoping sums argument implies that any leaf of CL has depth O(log n).
Refer to nodes of CL that are also nodes of C as C-nodes. Our data structure will maintain
CL as well as n(u) for each C-node u and rank(v) for each node v ∈ CL.
3.5 Searching for edges
We shall use CL to search for edges in Mi. To facilitate this, we associate a bitmap edge(u)
with each node u of CL where edge(u)[i] = 1 iff a level i-edge is incident to a leaf of the subtree
of CL rooted at u.1
We can use these bitmaps to search for the edges of Mi. Consider one of the search
procedures, say Pu, described above. At any point in the search, a set of vertices of Mi have
been explored and these correspond to level (i + 1) nodes in CL that we mark as explored.
With the bitmaps, we identify unexplored descendant leaves of marked nodes in CL that are
incident to level i-edges and hence to edges of Mi that should be explored by Pu. At each
leaf, we have all incident edges grouped according to their level. A BBST is kept which allows
us to get down to a particular group in O(log ℓmax) = O(log log n) time. When a level i-edge
(a, b) is explored in the direction from a to b, we determine the endpoint in Mi corresponding
to b by moving up from leaf b to the ancestor level (i + 1) node in CL. Finally, we mark this
level (i + 1) node as explored. Since CL has O(log n) height, we can execute Pu in O(log n)
time per edge explored.
3.6 Maintaining CL
We now describe how to maintain CL as C is updated. Let us consider the update in C of
merging nodes u and v into u. In CL, this is done by first removing the rank paths in L(u)
and L(v), leaving at most log n rank trees of distinct ranks for each of the nodes u and v. We
may assume that rank trees are kept in two lists sorted by the ranks of their roots and we
merge the two lists into one and start pairing up trees whose roots have the same rank, in
the same way as above. We connect their roots with a new rank path, thereby creating the
new L(u) and we identify its root with u. Total time for a merge is O(log n).
We also need to update CL when a child b in C is added to or removed from a node a (we
need this when failing to find a replacement path at some level). If b is to be added, we can
1Thorup's data structure needs two bitmaps in order to distinguish between tree edges and non-tree edges
whereas we only need one; edge(u) can be regarded as the bitwise 'or' of his two bitmaps.
6
regard it as a trivial rank tree that should be added to L(a). This can be done in O(log n)
time using the same approach as for merging. If b is to be removed, we first remove the rank
path of L(a) and identify the rank tree Tb containing b. We delete the path from b to the root
of Tb, thereby partitioning this rank tree into O(log n) smaller rank trees, sorted by ranks.
We pair up rank trees as described above and add a new rank path to form the updated L(a).
All of this can be done in O(log n) time.
3.7 Maintaining bitmaps
Finally, we need to update integers n(u) for C-nodes u as well as the edge-bitmaps. The former
is done exactly as in Sections 3.2 and 3.3 so let us focus on the bitmaps. If a level i-edge e
is removed, we do the following for each of its endpoints a. In the leaf a of CL, we check in
O(log log n) time if e was the only level i-edge incident to a. If so, we set edge(a)[i] ← 0 and
we move up CL, updating the bitmap of the current node as the bitwise 'or' of its children.
Since CL has O(log n) height, only O(log n) bitmaps need to be updated. Similarly, if e is
added, we set edge(a)[i] ← 1 and update bitmaps for ancestors in the same way. For nodes of
CL whose children change, we also update their bitmaps bottom up by taking the bitwise 'or'
of their children. Only O(log n) nodes are affected in CL after an update in C so total time is
O(log n).
This completes the description of the first version of our data structure. Correctness
follows since the data structure is a simple variation of that of Thorup where spanning trees
of clusters are not kept; rather, our search procedures certify that a spanning tree exists for an
explored component and this suffices to maintain the cluster forest. From the analysis above,
our data structure handles updates in O(log2 n) amortized time and queries in O(log n) time.
In the next section, we speed up both bounds by a factor of log log n.
4 An Improved Data Structure
In this section, we give our improved data structure. Before going into details, let us highlight
the main differences between this structure and that of the previous section. One ingredient is
to add shortcuts to CL. Each shortcut skips O(log log n) nodes and this will allow our search
procedures to walk up trees of CL in O(log n/ log log n) time per traversal when identifying
visited nodes of a multigraph Mi. Adding these shortcuts essentially corresponds to turning
the forest CL of binary trees into one consisting of trees with a branching factor of order
log n, and reducing the height of the trees to order log n/ log log n. Furthermore, we will
modify CL by using lazy local trees similar to those of Thorup [10] instead of the local trees
presented in the previous section. This is done to maintain CL more efficiently during changes.
Unfortunately, Thorup's lazy local trees increase the height of CL to O(log n log log n) so our
shortcuts will not give any speed-up over the data structure in the previous section. Instead,
we shall use a slightly more complicated type of lazy local tree which has the properties we
need while keeping the height of CL bounded by O(log n). The idea is to partition the children
of each node of C into so called heavy children and light children and construct the lazy local
tree only for the light children and the local tree of the previous section for the heavy children.
Balancing this in the right way will ensure a logarithmic depth of trees while still getting the
speed-up from lazy local trees. Finally, we will need another system of shortcuts for quickly
identifying edges to be explored by the search procedures; Thorup uses a similar system but
it does not fit into our framework as our lazy local trees are different from his. As shown in
7
Th(u)
u
Tl(u)
top tree
Figure 1: Lazy local tree L(u). Rank trees are black, top and bottom trees are white, the
buffer tree is grey, nodes of L(u) ∩ C are white, and nodes of L(u) \ C are grey.
Lemma 6 in Section 4.7, with these shortcuts, the search procedures can visit edges in only
O(log n/ log log n) time per edge plus an additive O(log n) if we are in case 1 in Section 3.3;
note that the latter can be paid for by the delete operation since a replacement path for the
deleted edge has been found. In Section 4.6, we define these shortcuts and in Section 4.7, we
give an algorithm that uses these shortcuts to explore level i-edges; we refer to it as a level
i-iterator or just iterator.
4.1 Lazy local trees
Thorup [10] introduced lazy local trees and showed how they can be maintained more effi-
ciently than the local trees in Section 3.4. Let u be a non-leaf node of C and let L be the
set of children of u. To form the lazy local tree of u, L is divided into groups each of size at
most 2(log n)α, where α is a constant that we may pick as large as we like. The nodes in each
group are kept in a BBST ordered by n(v)-values. One of these trees is the buffer tree while
the others are the bottom trees. The root of a bottom tree has rank equal to the maximum
rank of its leaves. These bottom trees are paired up to form at most log n rank trees, as
described in Section 3.4. The roots of the rank trees together with the root of the buffer tree
are leaves of a BBST called the top tree where leaves are ordered according to rank (and the
root of the buffer tree is regarded as, say, the smallest element). Together, these trees form
the lazy local tree of u which is rooted at the root of the top tree. Note that bottom, buffer,
and top trees have polylogarithmic size only. It is ensured by the data structure of Thorup
that for each bottom tree B, new leaves are never added to B and ranks of leaves in B are
not changed. This will also hold for our data structure.
We shall use these lazy local trees to improve update time to O(log2 n/ log log n). However,
it is easy to see that due to the BBSTs in lazy local trees, if we form CL with these trees,
the depth of CL becomes O(log n log log n). If we use the same approach as in the previous
section, we thus increase query and update time by a factor of log log n. Adding shortcuts to
CL will avoid this slowdown but this gives a data structure with the same bounds as in the
previous section.
To handle this, we introduce a new type of lazy local trees. Let u be a non-leaf node of
C. A child v of u in C is said to be heavy if n(v) ≥ n(u)/ logǫ n and otherwise it is light; here
ǫ > 0 is a constant that we may pick as small as we like.
It is rooted at u and has two
children. One child is the root of a tree Th(u) having the heavy children of u as leaves and
the other child is the root of a tree Tl(u) having the light children of u as leaves (to simplify
Our lazy local tree L(u) of u is illustrated in Figure 1.
8
the description, we assume that u has both light and heavy children; if there were no light
(resp. heavy) children, we would simply identify L(u) with Th(u) (resp. Tl(u))). We call Th(u)
the heavy tree (of u) and it is defined as the local tree from the previous section over the heavy
children of u; note that its size is asymptotically bounded by the number of heavy children
of u which is at most logǫ n. The tree Tl(u) is called the light tree (of u) and it is Thorup's
lazy local tree for the set of light children of u.
The following lemma shows that when CL is formed from C by inserting these lazy local
trees, the height of trees in CL is only a constant factor larger than that in Section 3.
Lemma 1. The height of CL (with lazy local trees) is O(log n).
Proof. Let u be a non-leaf node in C and let v be one of its children, also in C.
If v is a
heavy child of u then v ∈ Th(u) so its depth in L(u) is at most log(n(u)/n(v)) + 1. Now
assume that v is a light child of u. Then v belongs to either a bottom tree or the buffer tree
of L(u). In the latter case, the depth of v in L(u) is O(log log n). In the former case, let B
be the bottom tree containing v and let w be a leaf of B maximizing n(w). By definition,
the root of B has rank rank(w). Hence the depth of v in L(u) is at most log(n(u)/n(w)) +
O(log log n) ≤ log(n(u)/n(v))+O(log log n). This is O(log(n(u)/n(v))) as n(v) < n(u)/ logǫ n
implies log log n = O(log(n(u)/n(v))).
It follows that in both cases, v has depth O(1 + log(n(u)/n(v))) in L(u). The height of C
is at most log n so by a telescoping sums argument, CL has height O(log n), as desired.
4.2 Maintaining lazy local trees
Now, let us describe how to maintain lazy local trees of CL corresponding to changes in C. For
technical reasons, we assign ranks to rank path nodes in heavy trees by rank(vi) = rank(ri)
for i = 1, . . . , k − 1, where vi and ri are defined as in Section 3.4. In the following, rank nodes
are nodes that are assigned a rank. Note that every node of CL is a rank node except non-leaf
nodes of a buffer or top tree and nodes of a bottom tree B that are neither leaves nor the
root of B.
4.2.1 Merging
We need to maintain lazy local trees when merging C-nodes and when adding and removing
children from C-nodes. We start with merging. Consider two C-nodes u and v that are to be
merged into u. Denote by u′ the updated u after the merge. Note that every heavy child of
u′ must be a heavy child of either u or v. Hence, we can form Th(u′) by traversing every leaf
w of Th(u) and Th(v) and adding it as a leaf of Th(u′) iff n(w) ≥ n(u′)/ logǫ n. Total time for
this is O(Th(u) + Th(v)) = O(logǫ n).
Forming Tl(u′) is done as in [10]. The two buffer trees are merged in time bounded by the
smaller of the two trees. If the number of leaves of the merged buffer tree exceeds logα n, it
is turned into a bottom tree for u′, leaving an empty buffer tree. The root of the new bottom
tree is paired up with other rank nodes, if needed. We can pay for all buffer tree merges by
giving a node (log log n)2 credits when it switches from not belonging to belonging to a buffer
tree: every time it is moved to another buffer tree, we spend O(log log n) time for the node
and the size of the buffer tree containing the node grows by a factor of at least 2; hence the
node is moved at most O(log log n) times before either being deleted or being moved to a
bottom tree.
9
We propose a different approach for merging top trees than that of Thorup; we feel ours
is simpler as it avoids keeping a special bitmap associated with each top tree node. Assume
w.l.o.g. that the top tree T (u) in Tl(u) is no bigger than the top tree T (v) in Tl(v). For each
leaf of T (u) we binary search for a leaf with the same rank in T (v) in O(log log n) time. While
there are roots with equal ranks, we pair them up as before, finally obtaining the top tree for
Tl(w). The time for this is O(T (u) log log n) plus time bounded by the number of new rank
nodes created. Below we will show how the creation of rank nodes are paid for when they
are deleted. With an amortized analysis similar to that above for buffer trees, we can pay
for all top tree updates if we assign (log log n)2 credits to a node when it switches from not
appearing to appearing as a leaf in a top tree, and if we borrow (log log n)2 credits whenever
we delete a leaf of a top tree (thereby borrowing from a new rank node) and distribute these
borrowed credits evenly among the remaining leaves.
Since n(u′) = n(u) + n(v), we may have some leaves w ∈ Th(u) ∪ Th(v) with n(w) <
n(u′)/ logǫ n and hence w should belong to Tl(u′). All such nodes are added to the buffer
tree; as before, this tree is turned into a bottom tree if it gets more than logα n leaves.
Total time for this is O(logǫ n log log n) (plus time bounded by the number of new rank nodes
created) since at most 2 logǫ n nodes need to be moved from Th(u) ∪ Th(v).
Finally, let p be the parent of u and v in C; below we will add shortcuts that allow us
to identify p from u (equivalently from v) in O(log n/ log log n) time (Lemma 3). We remove
u and v as leaves of L(p) and then we add u′ as a leaf of Th(p) if n(u′) ≥ n(p)/ logǫ n and
otherwise we add u′ as leaf of the buffer tree in Tl(p). This takes O(log log n) time since both
the buffer tree and Th(p) have poly-logarithmic size and their roots have depth O(log log n)
in L(p).
4.3 Removing a child
Consider removing a C-node child v of a C-node u and adding it as a child of a new C-node w
which is added as a child of the parent C-node p of u. We first focus on removing v and we
let u′ denote u after this update.
Assume first that v ∈ Th(u). After removing v from Th(u), we have n(u′) = n(u) − n(v)
and hence some nodes may need to be moved from Tl(u) to Th(u) in order to form L(u′).
Identifying such nodes in the buffer tree of Tl(u) can be done in O(log log n) time per node.
Now suppose w is a leaf of a bottom tree B that needs to be moved to Th(u). Let b be
the root of B and let a be the leaf of the top tree of Tl(u) having b as descendant. Since
rank(b) ≥ rank(w) ≥ ⌊log(n(u′)/ logǫ n)⌋ > ⌊rank(a) − ǫ log log n⌋ and since ranks are strictly
decreasing on the path from a to b, b has depth at most ⌈ǫ log log n⌉ in the subtree of Tl(u)
rooted at a. A DFS from a identifies all bottom tree roots with at most this depth in O(logǫ n)
time; let La be the set of leaves in these bottom trees that need to be moved to Th(u). Using
binary search in the O(logǫ n) bottom trees, we identify La and move it to Th(u) in time
O((logǫ n + La) log log n). Since all rank leaves of the top tree of Tl(u) have distinct ranks,
there are only O(log log n) choices for a. Also, the maximum number of leaves to be moved
is bounded by the number of leaves of Th(u′) which is at most logǫ n. Total time is thus
O(logǫ n(log log n)2).
We may also need to move u in L(p). If u belongs to a bottom tree in Tl(p), we move it
to the buffer tree as we do not allow ranks of leaves in bottom trees to change. We also move
u to the buffer tree if u ∈ Th(p) and n(u′) < n(p)/ logǫ n. As we saw for merge, the time for
this is O(log n/ log log n).
10
As in [10], we need to do more global updates whenever removing a leaf from a bottom
tree B of Tl(u) reduces the maximum rank of leaves in B and hence the rank of the root b of
B. We use a similar approach and amortized analysis as Thorup here: first, delete all rank
nodes from b to the ancestor leaf a of the top tree of Tl(u). Then pair nodes of equal rank as
before. For the amortized analysis, we can assume that the graph G ends with no edges so
all rank nodes end up being deleted and we can amortize creation of rank nodes in Tl(u) over
deletion of rank nodes in Tl(u). A rank node is only deleted when a bottom tree root has its
rank reduced. Since a rank is at most log n, a particular bottom tree can have its root rank
reduced at most log n times (nodes are never added to a bottom tree and ranks of bottom tree
leaves do not change) so in total it gives rise to at most log2 n rank node deletions for that
bottom tree. But since a bottom tree starts out with (log n)α leaves that will all be deleted
eventually, we can amortize each rank node deletion over (log n)α−2 deletions of bottom tree
leaves. When removing child v, we delete at most logǫ n leaves from bottom trees of Tl(u) in
order to form L(u′) so we can amortize each rank node deletion over (log n)α−2−ǫ deletions
of children in C. Hence if we pick constant α ≥ 2 + ǫ, we can afford to pay for rank node
deletions and also to pay for the (log log n)2 credits that may have been borrowed from a rank
node during a merge.
It remains to consider the case v ∈ Tl(u). Above we showed how to efficiently remove up
to order logǫ n leaves from Tl(u) so clearly the single leaf v can also be removed efficiently.
We then move additional leaves from Tl(u) to Th(u) and move u to Th(p) or to the buffer tree
of Tl(p), as above.
4.4 Adding a child
Now consider adding v as a child of w and w as a child of p. The former is trivial as w has
no children before adding v. If n(v) ≥ n(p)/ logǫ n, we add w to Th(p) and otherwise we add
it to the buffer tree of Tl(p). Given p, total time for this is O(log log n).
4.5 Shortcutting
In order to get our log log n speed-up for updates and queries, we need to be able to traverse
CL faster. Thorup [10] introduced a system of shortcuts for quickly identifying certain edges
incident to clusters. This will not suffice in our approach since for our search procedures, we
also need to move quickly from a leaf of CL to its ancestor level i node in order to identify the
associated level i cluster, for some i. We therefore introduce a different system of shortcuts in
the following. To avoid skipping past a level i node with these shortcuts, our data structure
associates, for each node of a heavy tree Th(u), the level ℓ(u) of u. We can easily extend the
data structure for lazy local trees to maintain these values within the same time bound since
Th(u) has only size O(logǫ n).
Let us color each node of CL either white or black. For the coloring below, we define a
black-induced child of a node u ∈ CL to be a black descendant v of u such that all interior
nodes on the path from u to v in CL are white. If u is black, we add a shortcut between u and
each of its black-induced children. The shortcut is directed to u, allowing us to move quickly
up in CL. The black-induced parent of a node is defined similarly. Note that the shortcuts
(with directions reversed) form a forest of rooted trees over the black nodes.
Now, let us define the coloring of nodes of CL. The following nodes are colored black:
1. every C-node u with ℓ(u) = i⌊ǫ log log n⌋ for some integer i,
11
2. every rank node u having a parent rank node v in CL such that rank(u) ≤ i⌊ǫ log log n⌋ <
rank(v) for some integer i,
3. every leaf of CL and of every buffer, bottom, and top tree, and
4. every node of a buffer, bottom, and top tree whose depth in that tree is divisible by
⌊ǫ log log n⌋ (in particular, every root of such a tree is black).
A black node is of type 1, 2, 3, and/or 4, depending on these four cases. All other nodes
are colored white. For performance reasons, we shall only maintain edge-bitmaps for black
nodes. Lemma 3 below shows that these shortcuts give a log log n speed-up when moving up
a tree of CL. We first need the following result.
Lemma 2. Ranks are non-decreasing along any simple leaf-to-root path in CL. Between any
two consecutive C-nodes on such a path, there are at most two pairs of nodes of equal rank.
Proof. The first part of the lemma will follow if we can show that ranks are non-decreasing
along any simple leaf-to-u path P in a lazy local tree L(u). This is clearly the case for leaves
in Th(u). A leaf v in Tl(u) either belongs to a bottom or buffer tree T . Assume the former
since otherwise, u and v are the only rank nodes on the v-to-u path P and since both are
C-nodes, rank(u) ≥ rank(v).
Only the first node v and last node r of subpath P [v, r] = T ∩ P are rank nodes. Since r
is the root of T , it has maximum rank among leaves in T so rank(v) ≤ rank(r). Let l be the
leaf of the top tree of L(u) belonging to P . All nodes on P [r, l] belong to a rank tree so ranks
are increasing along this subpath. For the subpath P [l, u], only l and u are rank nodes. Let L
be the set of leaves of Tl(u) formed by picking one of maximum rank from each bottom tree
descending from l. Then rank(l) ≤ ⌊log(Pu′∈L n(u′))⌋ ≤ ⌊log n(u)⌋ = rank(u). This shows
the first part of the lemma.
For the second part, let u and v be C-nodes where v is a child of u. Assume first that v is
a leaf of Th(u). Ranks are strictly increasing on the path from v to the root r of the rank tree
containing v. Ranks are also strictly increasing along the rank path in Th(u). Hence, there
are at most two pairs of equal rank nodes in CL between u and v, namely r and its parent and
the root of Th(u) and u. Now consider the case where v is a leaf of Tl(u) and again assume it
belongs to a bottom tree B. Let r be the root of B and let l be the leaf of the top tree which
is an ancestor of v. Then again, since ranks are strictly increasing along any leaf-to-root path
in a rank tree, there can be at most two equal-rank pairs between u and v, namely (v, r) and
(l, u). This completes the proof.
Lemma 3. Given CL with shortcuts, given a level i, and given a C-node of CL with an ancestor
level i node, we can identify this ancestor in O(log n/ log log n) time.
Proof. Let v be the given node. To identify the ancestor level i node u of v, we start by
traversing the v-to-root path of the tree in CL containing v and we stop if we reach u or a
black node. Since v is a C-node and since light trees have black leaves, all nodes visited are
rank nodes. Lemma 2 then implies that we visit at most O(log log n) nodes before stopping.
Hence, the traversal takes O(log log n) time. Assume that we reach a black node b1 as we are
done if we reach u.
From b1 we traverse shortcuts until we get to the lowest-depth black node b2 having u as
ancestor. Finally we traverse the b2-to-root path in CL until we reach u. The latter takes
O(log log n) time by an argument similar to the above.
12
Next we show that there are O(log n/ log log n) shortcuts between b1 and b2. Since C has
logarithmic height, there are only O(log n/ log log n) shortcut endpoints of type 1. Lemma 2
implies the same bound for shortcut endpoints of type 2.
If a shortcut ends at a type 3
node b which is a leaf of a buffer or bottom tree, it means that we enter a light tree Tl(w).
We encounter only one additional type 3 node in Tl(w), namely a leaf of a top tree. Since
n(w) > n(b) logǫ n we have rank(w) > rank(b) +⌊ǫ log log n⌋ and since a rank is at most log n,
Lemma 2 implies that we encounter no more than O(log n/ log log n) type 3 nodes between
b1 and b2. Finally, this bound on the number of type 3 nodes and Lemma 1 give the same
asymptotic bound on the number of type 4 nodes.
What remains is to describe how to avoid jumping past u when traversing the shortcuts.
Let (b2, b3) be the shortcut that jumps past u, if any. Since leaves of light trees are black,
b3 must belong to some heavy tree Th(wb3 ). If b2 belongs to a light tree, it must belong to
the root of the top tree in Tl(u) since that root is black. We can avoid this case as follows:
whenever we reach the root of a top tree, its parent is a C-node and we compare its level with
i to determine whether we should continue with the shortcuts.
Now, consider the case where b2 belongs to a heavy tree Th(wb2 ). Recalling that for every
node of a heavy tree Th(w) we keep the level ℓ(w) of w, we can check that ℓ(wb2) ≤ i < ℓ(wb3 )
to detect that a is the last node that we should visit with shortcuts. This completes the
proof.
4.6
Induced shortcuts
Lemma 3 allows us to speed up part of our search procedure, namely identifying the endpoints
(level (i + 1) clusters) of an edge in a multigraph Mi from the endpoints of the corresponding
edge in G; we can do this in O(log n/ log log n) time per endpoint. We also need a faster
iterator for level i-edges incident to explored level (i + 1) clusters. We focus on this in the
following.
Define an i-induced forest Fi as in [10]: its i-induced leaves are the leaves of CL with an
incident level i-edge. Its i-induced roots are the level (i + 1) nodes of CL having descendant
Its i-induced branch nodes are the nodes of CL with both children have
i-induced leaves.
descending i-induced leaves. The i-induced parent of an i-induced node is its nearest i-
induced ancestor. This defines Fi.
A straightforward level i-iterator performs a DFS in a tree of Fi. However, maintaining
the edges of Fi will be too expensive.
Instead, we introduce a new system of shortcuts
in CL that will allow the DFS to move between any two incident i-induced nodes of Fi
in O(log n/ log log n) time. Since a tree of Fi is binary, the number of branch nodes of T is
bounded by the number of leaves of T , so this will give a level i-iterator with O(log n/ log log n)
amortized time per level i-edge. In the following, we define the new shortcuts. Refer to the
following types of nodes of CL as special:
1. every C-node u with ℓ(u) = i⌊log log n⌋⌊ǫ log log n⌋ for some integer i,
2. every leaf of CL, and
3. every rank node u of a light tree with rank(u) = i⌊log log n⌋⌊ǫ log log n⌋ for some integer
i.
Note that every special node is black. Also note that we defined type 3 special nodes using
equality rather than inequality as for type 2 black nodes. This suffices since ranks increase
13
by 1 as we move up rank nodes of a light tree; this is not the case in heavy trees where ranks
can increase by larger values along a rank path. For a special node u, define a special child
of u to be a descendant special node v such that all nodes between u and v are not special.
Special parents are defined similarly. For any level i, if there is a unique special child v of u for
which edge(v)[i] = 1, we add a shortcut (bidirected pointer) between u and v. To distinguish
these shortcuts from those of Section 4.5, we refer to the former as i-induced shortcuts or just
induced shortcuts and the latter as standard i-shortcuts or just standard shortcuts. Observe
that for all i-induced shortcuts (a, b), where b is a special child of a, there is an edge in Fi
from a or an ancestor of a to b or a descendant of b. For each special node u, we keep a BBST
with a leaf for each i containing the i-induced shortcuts to a special child and/or parent (if
they exist).
4.7 Faster iterator
Now let us present the level i-iterator. It starts at the root v of a tree in Fi, i.e., v is a level
(i + 1) node of CL. It performs a DFS of the subtree of CL rooted at v with the following
if it visits a black node w for which edge(w)[i] = 0, it backtracks; if it visits
modification:
a special node w′ with an i-induced shortcut to a special child, it visits this special child
instead of the children of w′ in CL. When it reaches a leaf l of CL, it identifies the group of
incident level i-edges with a binary search in the BBST associated with l and then iterates
over these edges. This completes the description of the level i-iterator. Lemma 6 below shows
the performance of the level i-iterator. To prove it, we need two additional lemmas.
Lemma 4. Any node of CL has only O((log n)3ǫ) black-induced children.
Proof. Let u be a node of CL. If u is a non-leaf node of a bottom, buffer, or top tree, the
lemma follows from the definition of type 3 and 4 black nodes. Otherwise, u is a rank node.
For any black-induced child v of u, Lemma 2 and the definition of type 1 and 2 black nodes
imply that v has depth at most 3ǫ log log n in the subtree of CL rooted at u. As CL is binary,
the lemma follows.
Lemma 5. For any i-induced shortcut (a, b), the simple a-to-b path in CL has length O((log log n)4).
Proof. Let P be the simple a-to-b path in CL. Clearly, P contains only O((log log n)2) C-
nodes. Let P ′ be a subpath of P containing no C-nodes. Then P ′ is either contained in a
heavy or a light tree. In the former case, P ′ = O(log log n). In the latter case, we encounter
at most O(log log n) nodes of buffer, bottom, and top trees on P ′. Since consecutive rank
nodes of P ′ differ in rank by exactly 1 (as they all belong to a light tree and hence to a rank
tree), we encounter at most O((log log n)2) rank nodes on P ′ so P ′ = O((log log n)2).
Lemma 6. The level i-iterator above traverses a tree in Fi with k leaves in O(k log n/ log log n)
time. The time to visit the first k′ leaves is O(k′ log n/ log log n + log n).
Proof. Correctness follows easily from the definition of edge-bitmaps, i-induced shortcuts,
and BBSTs associated with leaves of CL so let us focus on the time bound to traverse a
k-leaf tree T in Fi. Let TL be the tree in CL obtained by replacing each edge (a, b) ∈ T with
the corresponding simple path P in CL between a and b. By Lemmas 1 and 2, there can
only be O(log n/(log log n)2) special nodes on such a path P . Hence, since T has no degree
2-vertices, the total number of special nodes and hence i-induced shortcuts traversed by the
14
level i-iterator in TL is O(k log n/(log log n)2). For each special node visited, O(log log n) time
is spent on binary search to find the next i-induced shortcut, if it exists. Hence, the total time
spent on visiting special nodes and traversing i-induced shortcuts is O(k log n/ log log n).
We will now show that the number of additional nodes visited by the DFS is O(k(log n)3ǫ(log log n)4).
Since only constant time is spent for each such node, this will show the first part of the lemma.
First we bound the number of visited nodes of TL which are not special. Let (a, b) and P be as
above. If we traverse P from a then it follows from Lemma 5 that after at most O((log log n)4)
nodes, we will reach either b or a special node a′. Similarly, if we traverse P from b then after
at most O((log log n)4) nodes, we will reach either a or a special node b′. If a′ and b′ exist then
all nodes visited by the DFS on P [a′, b′] are special nodes connected by i-induced shortcuts.
Summing over all such paths P , it follows that the total number of nodes visited on TL which
are not special is O(k(log log n)4).
Finally, let us bound the number of nodes of CL visited by the DFS which are not on TL.
Consider a visited node u ∈ TL and let v /∈ TL be a visited node such that u is the nearest
ancestor of v belonging to TL. Note that there is no i-induced shortcut from u to a special
child since then the DFS would have traversed this shortcut instead of the children of u in CL.
In particular, there are only O(k(log log n)4) choices for u. Furthermore, all interior nodes on
the simple path from u to v in CL are white since any black node w would have edge(w)[i] = 0
(as w /∈ TL), meaning that the DFS would have backtracked before reaching v. By Lemma 4,
there are only O((log n)3ǫ) choices for v for each u. Hence, the total number of nodes visited
which are not on TL is O(k(log n)3ǫ(log log n)4). This shows the first part of the lemma.
For the second part, consider a partially grown DFS tree T ′ which has visited k′ leaves.
For every node of T ′ having two children, at least one of the two subtrees rooted at the
children is fully explored. Hence, T ′ consists of a path P from the root of T to a leaf of T
with fully explored subtrees attached to P . The same argument as above shows that the total
time to explore these subtrees is O(k′ log n/ log log n). By Lemma 1, it takes O(log n) time to
explore P (the number of special nodes on P is O(log n/(log log n)2) so we only spend a total
of O(log n/ log log n) time on binary searches for these nodes).
It follows from Lemma 6 that the level i-iterator spends O(log n/ log log n) amortized time
per edge visited plus additional O(log n) time if a replacement path was found (if such a path
is not found, an entire tree in Fi is traversed). The O(log n/ log log n) amortized time per
edge is paid for by the increase in the level of the edge and the O(log n) time is paid for by
the deletion of an edge in G since at most one replacement path is found for such an edge.
It remains to describe how colors, shortcuts (standard and induced) and edge-bitmaps are
maintained when C (and hence CL) is updated and when edges of G are added/removed or
change level. First we deal with changes to C. The following lemma will prove useful.
Lemma 7. Given edge-bitmaps of black nodes and given a special node u, we can find the
induced shortcuts between u and its special parent (if any) in O(log n) time. For any i, we
can find the i-induced shortcut from u to a special child or determine that no such shortcut
exists in O((log n)3ǫ(log log n)4) time.
Proof. We first walk up CL from u to identify its special parent p. By Lemma 5, this takes
O((log log n)4) time. Then we perform a DFS in the subtree of CL rooted at p and backtrack
if we encounter u or a black node which is not an ancestor of u. If any such black node is
encountered for which the ith bit of its edge-bitmap is 1 then we know that there should
not be an i-induced shortcut between u and p. Otherwise there should be iff edge(u)[i] = 1.
15
Let m be the bitmap obtained by taking the bitwise 'or' of the edge-bitmaps of visited black
nodes not on the u-to-p path. By Lemmas 4 and 5, finding m takes O((log n)3ǫ(log log n)4)
time. Now, there is an i-induced shortcut between u and p iff m[i] = 0 and edge(u)[i] = 1.
Hence, all induced shortcuts between u and p can be found in O(log n) time.
To find the i-induced shortcut (if any) to a special child of u, we make a DFS from u
which backtracks when reaching a black node. Suppose exactly one visited black node w has
edge(w)[i] = 1 (otherwise, there is no i-induced shortcut). If w is special, we have identified
the i-induced shortcut (w, u). Otherwise, we recurse on w. As above, total time for this is
O((log n)3ǫ(log log n)4).
4.8 Structural changes
Let us now describe how shortcuts and edge-bitmaps are maintained after structural changes
to CL. It follows from Lemma 4 that for each update to CL, we can update colors, standard
shortcuts and edge-bitmaps in O((log n)3ǫ) time. From the results in Section 4.2, this will
not affect the overall time bound (if we pick constant α sufficiently large). In the following,
we thus only consider updating induced shortcuts.
We shall restrict our attention to structural changes caused by a delete-operation as
insert corresponds to merging two clusters (or none), a type of update that needs to be
supported during a delete.
Recall that after a delete-operation, C is updated as follows: some children of a node
u are removed and merged into a single node; this node is either added as a child of u (if a
replacement path was found) or it is added as a child of a new node u′ which is added as a
child of the parent p of u and the process is repeated recursively on p (if a replacement path
was not found). We observe that all C-nodes whose children are updated are contained in two
leaf-to-root paths in C after the delete-operation has been executed.
There are two types of induced shortcuts that need to be updated, those incident to a
type 1 special node and those descending from a type 3 special node and not ascending from
a type 1 special node (see definitions of types in Section 4.6). Below we show how to update
the latter.
For the former, it follows from the above that we only need to focus on type 1 special
nodes on a leaf-to-root path P in CL (there are two paths but they are handled in the
same manner). Let u1, . . . , uk be the sequence of special nodes as we move up P during
the delete-operation (some of them may be new or merged nodes and hence do not exist
before the delete-operation). By Lemma 7, we can find all induced shortcuts descending
from u1 in O((log n)1+3ǫ(log log n)4) time. When we reach uj, j > 1, we compute induced
shortcuts between uj−1 and uj. By Lemma 7, this takes O(log n log log n) time (including
binary searches in the BBSTs of uj−1 and uj) for a total of O(log2 n/ log log n) over all j
which the delete-operation can pay for. We also compute induced shortcuts descending
from uj for those i for which edge(uj)[i] = 1 and edge(uj−1)[i] = 0. Total time over all j is
O((log n)1+3ǫ(log log n)4) since if edge(uj)[i] = 1 then edge(uj ′)[i] = 1 for all j′ > j, implying
that the second part of Lemma 7 is applied at most once for each i. Note that all i-induced
shortcuts from uj to a special child which have not been identified by the second part of
Lemma 7 must have edge(uj)[i] = edge(uj−1)[i] = 1 and hence must connect uj to uj−1
which we have computed above. Hence, we correctly compute all induced shortcuts incident
to special nodes on P and we do so within the desired time bound.
Now we show how to maintain induced shortcuts descending from a type 3 special node
16
p and not ascending from a type 1 special node. We may assume that p is not on one of the
leaf-to-root paths considered above. Then the only structural changes to CL that may require
such shortcuts to be updated are
1. a type 3 special node is created or deleted, or
2. a leaf is removed from a bottom tree in the light tree containing p.
A type 3 special node can pay any polylogarithmic amount when it is created/deleted so
consider updating induced shortcuts descending from p when a leaf u is removed from a
bottom tree B in the light tree Tl(w) containing p. To handle this case, we will assume that
each rank node of Tl(w) has log n/ log log n credits for each i for which it is an i-induced
branch node. To see that this assumption can be made, first observe that when a buffer tree
is turned into a bottom tree, it can pay any polylogarithmic amount if we pick α sufficiently
big. This is also the case when a new rank node of Tl(w) is created/deleted. Since we never
add but only remove leaves from bottom trees, the only other way a rank node of Tl(w) can
become an i-induced branch node is if some edge of G has its level increased to i. Such an
edge can only contribute with two i-induced branch nodes to Fi (one for each of its endpoints)
so we may add log n/ log log n credits to the two new i-induced branch nodes which the level
increase can pay for. This shows the desired.
Let branch(u) be the bitmap where branch(u)[i] = 1 iff there is an i-induced branch
node v 6= u on the path from u to p. By Lemmas 4 and 5, we can form branch(u) in
O((log n)3ǫ(log log n)4) time which the removal of u from B can pay for. For each i for
which branch(u)[i] = 1, the removal of u may require us to compute an i-induced shortcut
descending from p. Since a branch node is removed in the process, we can spend its credits
to pay for applying Lemma 7 to find this shortcut. We can binary search for each of the k
1-entries of branch(u) in O(k log log n) time; for instance, to determine whether the lower half
of branch(u) has any 1-bits, we can take the bitwise 'and' of branch(u) and a precomputed
bitmap having 1-bits in the lower half and 0-bits in the upper half. The 'and' is 1 iff there
are 1-bits in the lower half of branch(u).
Now consider an i for which branch(u)[i] = 0. We may assume that a bitmap induced(p)
associated with p is maintained where induced(p)[j] = 1 iff p has a j-induced shortcut to a
special child. If edge(u)[i] = 0 or induced(p)[i] = 0, nothing needs to be done for i so assume
edge(u)[i] = 1 and induced(p)[i] = 1. Then we delete the i-induced shortcut descending from
p. If there is an i-induced shortcut from p to its special parent p′, we remove it too and recurse
on p′; the recursion stops when we reach a special node q without an i-induced shortcut to
its special parent q′. Each induced shortcut can be removed in O(log log n) time using binary
search in the associated BBSTs. By adding log log n credits to an induced shortcut when it
is created, we can pay for all deletions of induced shortcuts. If q′ is a rank node of Tl(w), we
may need to add an i-induced shortcut from q′ to a special child. This can only happen if an
i-induced branch node between q and q′ disappears and as above, we can spend its credits to
pay for finding this shortcut.
4.9 Non-structural changes
Above we dealt with updates of shortcuts due to structural changes in CL. We now handle
updates when leaves of Fi appear or disappear due to edge level changes.
17
Edge deletions: When a level i-edge e is deleted (possibly due to its level being increased
to i + 1), edge(u)[i] might change from 1 to 0 for one of its endpoints u which will then no
longer be an i-induced leaf of Fi. We describe how to update other edge-bitmaps accordingly
and remove some of the i-induced shortcuts. The following is done for u. If there are still
level i-edges incident to u then no updates are needed. Otherwise, all i-induced shortcuts on
the simple path in CL from u to its i-induced parent p should be removed. Since u is a leaf of
CL, it is a special node. We traverse i-induced shortcuts from u to ancestors until we reach
a special node v without an i-induced shortcut to an ancestor. Since u is the only i-induced
leaf below v, we delete all shortcuts traversed. We also set edge(v)[i] ← 0 for all black nodes
v between u and v (including v) in O(log n/ log log n) time by traversing standard i-shortcuts
between u and v.
We then traverse black nodes up from v in CL and stop if we find a special node w.
Whenever we visit a black node b, we perform a DFS in the subtree of CL rooted at b,
backtracking at descending black nodes. If each black node b′ visited below b has edge(b′)[i] =
0, u was the only leaf of CL below b with an incident level i-edge so we set edge(b)[i] ← 0
and proceed up to the next black node. Conversely, if some black node b′ visited below b has
edge(b′)[i] = 1, p must be below b and no more bitmaps need to be updated.
Having updated the bitmaps and removed all i-induced shortcuts below v, we need to check
if an i-induced shortcut should be added from w to one of its descendants. By Lemma 7, this
takes O((log n)3ǫ(log log n)4) time which can be paid for by the deletion of e; here we can also
afford to add log log n credits to the shortcut if it was added.
Edge insertions Now suppose a level i-edge e is inserted. To update edge-bitmaps and
add new i-induced shortcuts, we do the following for each endpoint u of e.
If other level
i-edges are incident to u then nothing needs to be done as u is already a leaf of Fi so assume
otherwise. Inserting e corresponds to updating Fi by adding a new edge (u, p), where p is the
i-induced parent of u. Hence, we need to add i-induced shortcuts between u and p. Suppose
u has an ancestor black node v such that edge(v)[i] = 1. We traverse standard shortcuts up
from u and stop when we identify the first such v. Let u1, . . . , uk be the ordered (possibly
empty) sequence of special nodes visited from u to v. Since v already has a descending leaf
incident to a level i-edge and uk does not, p must be on the uk-to-v path in CL. Hence, the
new i-induced shortcuts to be added are (u1, u2), (u2, u3), . . . , (uk−1, uk). We also set to 1 the
ith bit of the edge-bitmaps of all black nodes visited. By Lemmas 1 and 2, all of this can be
done in O(log n/ log log n) time and we can also afford to add log log n credits to each of the
O(log n/(log log n)2) new i-induced shortcuts.
We assumed that a node v with edge(v)[i] = 1 was encountered. If this is not the case,
it means that u should not be added to an existing tree in Fi. Rather, a new tree should be
formed consisting of a single edge (u, p), where p is the level (i+1)-ancestor of u in CL. Clearly,
the corresponding i-induced shortcuts can be added and edge-bitmaps updated within the
same O(log n/ log log n) time bound.
We can now conclude with the following theorem.
Theorem 1. There is a deterministic data structure for fully dynamic graph connectivity
which supports edge insertions/deletions in O(log2 n/ log log n) amortized time per update and
connectivity queries in O(log n/ log log n) worst case time, where n is the number of vertices
of the graph.
18
5 Concluding Remarks
We gave a deterministic data structure for fully-dynamic graph connectivity that achieves an
amortized update time of O(log2 n/ log log n) and a worst-case query time of O(log n/ log log n),
where n is the number of vertices of the graph. This improves the update time of Holm, de
Lichtenberg, and Thorup [5] and Thorup [10] by a factor of log log n. We believe our improve-
ment may extend to fully-dynamic minimum spanning tree, 2-edge, and/or biconnectivity.
There is still a small gap between upper and lower bounds. For instance, for O(log n/ log log n)
query time, [6] gives an Ω((log n)1+ǫ) cell-probe lower bound for updates for constant ǫ > 0.
Simultaneously getting O(log n) update and query time and improving the O(√n) worst-case
update bound in [1] remain the main open problems for fully-dynamic graph connectivity.
References
[1] D. Eppstein, Z. Galil, G. F. Italiano, and A. Nissenzweig. Sparsification - a technique for
speeding up dynamic graph algorithms. J. ACM, 44(5):669–696, 1997. See also FOCS'92.
[2] G. N. Frederickson. Data structures for on-line updating of minimum spanning trees,
with applications. SIAM J. Comput., 14(4):781–798, 1985. See also STOC'83.
[3] M. R. Henzinger and V. King. Randomized dynamic graph algorithms with polyloga-
rithmic time per operation. Proc. twenty-seventh annual ACM symposium on Theory of
computing (STOC), 1995, pp. 519–527.
[4] M. R. Henzinger and M. Thorup. Sampling to provide or to bound: With applications
to fully dynamic graph algorithms. Random Structures and Algorithms, 11(4):369–379,
1997. See also ICALP'96.
[5] J. Holm, K. de Lichtenberg, and M. Thorup. Poly-logarithmic deterministic fully-
dynamic algorithms for connectivity, minimum spanning tree, 2-edge, and biconnectivity.
J. ACM, 48(4): 723–760, 2001. See also STOC'98.
[6] M. Patra¸scu and E. Demaine. Logarithmic Lower Bounds in the Cell-Probe Model. SIAM
J. Comput., 35(4): 2006. Special issue 36th ACM Symposium on Theory of Computing
(STOC 2004).
[7] D. Sleator and R. E. Tarjan. A data structure for dynamic trees. J. Comput. Syst. Sc.,
26(3):362–391, 1983. See also STOC'81.
[8] R. E. Tarjan. Efficiency of a good but not linear set union algorithms. J. ACM, 22:215–
225, 1975.
[9] M. Thorup. Decremental dynamic connectivity. Proc. 8th SODA, pp. 305–313, 1997.
[10] M. Thorup. Near-optimal fully-dynamic graph connectivity. Proc. thirty-second annual
ACM symposium on Theory of computing (STOC), 2000, pp. 343–350.
19
|
0812.2291 | 7 | 0812 | 2013-06-03T21:03:36 | Characterizing Truthful Multi-Armed Bandit Mechanisms | [
"cs.DS",
"cs.GT",
"cs.LG"
] | We consider a multi-round auction setting motivated by pay-per-click auctions for Internet advertising. In each round the auctioneer selects an advertiser and shows her ad, which is then either clicked or not. An advertiser derives value from clicks; the value of a click is her private information. Initially, neither the auctioneer nor the advertisers have any information about the likelihood of clicks on the advertisements. The auctioneer's goal is to design a (dominant strategies) truthful mechanism that (approximately) maximizes the social welfare.
If the advertisers bid their true private values, our problem is equivalent to the "multi-armed bandit problem", and thus can be viewed as a strategic version of the latter. In particular, for both problems the quality of an algorithm can be characterized by "regret", the difference in social welfare between the algorithm and the benchmark which always selects the same "best" advertisement. We investigate how the design of multi-armed bandit algorithms is affected by the restriction that the resulting mechanism must be truthful. We find that truthful mechanisms have certain strong structural properties -- essentially, they must separate exploration from exploitation -- and they incur much higher regret than the optimal multi-armed bandit algorithms. Moreover, we provide a truthful mechanism which (essentially) matches our lower bound on regret. | cs.DS | cs |
Characterizing Truthful Multi-Armed Bandit Mechanisms∗
Moshe Babaioff
Microsoft Research Silicon Valley
Mountain View, CA 94043, USA
[email protected]
Yogeshwer Sharma†
Facebook
Palo Alto, CA 94301, USA
[email protected]
Aleksandrs Slivkins
Microsoft Research Silicon Valley
Mountain View, CA 94043, USA
[email protected]
December 2008
This version: May 2013
Abstract
We consider a multi-round auction setting motivated by pay-per-click auctions for Internet advertis-
ing. In each round the auctioneer selects an advertiser and shows her ad, which is then either clicked
or not. An advertiser derives value from clicks; the value of a click is her private information. Ini-
tially, neither the auctioneer nor the advertisers have any information about the likelihood of clicks on
the advertisements. The auctioneer’s goal is to design a (dominant strategies) truthful mechanism that
(approximately) maximizes the social welfare.
If the advertisers bid their true private values, our problem is equivalent to the multi-armed bandit
problem, and thus can be viewed as a strategic version of the latter. In particular, for both problems
the quality of an algorithm can be characterized by regret, the difference in social welfare between the
algorithm and the benchmark which always selects the same “best” advertisement. We investigate how
the design of multi-armed bandit algorithms is affected by the restriction that the resulting mechanism
must be truthful. We find that deterministic truthful mechanisms have certain strong structural properties
– essentially, they must separate exploration from exploitation – and they incur much higher regret
than the optimal multi-armed bandit algorithms. Moreover, we provide a truthful mechanism which
(essentially) matches our lower bound on regret.
ACM Categories and subject descriptors: F.2.2 [Analysis of Algorithms and Problem Complexity]:
Nonnumerical Algorithms and Problems; K.4.4 [Computers and Society]: Electronic Commerce; F.1.2
[Computation by Abstract Devices]: Modes of Computation—Online computation; J.4 [Social and Be-
havioral Sciences]: Economics
General Terms: theory, algorithms, economics.
Keywords: mechanism design, truthful mechanisms, single-parameter auctions, pay-per-click auctions,
multi-armed bandits, regret.
∗This is a full version of a conference paper published in 10th ACM Conf. on Electronic Commerce (EC), 2009. Apart from the
revised presentation, this version is updated to reflect the follow-up work [9, 56, 22, 48] and the current snapshot of open questions.
†This research was done while Y. Sharma was a student at Cornell University and an intern at Microsoft Research Silicon Valley.
1
1 Introduction
In recent years there has been much interest in understanding the implication of strategic behavior on the
performance of algorithms whose input is distributed among selfish agents. This study was mainly moti-
vated by the Internet, the main arena of large scale interaction of agents with conflicting goals. The field
of Algorithmic Mechanism Design [40] studies the design of mechanisms in computational settings (for
background see the recent book [41] and survey [47]).
Much attention has been drawn to the market for sponsored search (e.g. [31, 19, 55, 36, 2]), a multi-
billion dollar market with numerous auctions running every second. Research on sponsored search mostly
focus on equilibria of the Generalized Second Price (GSP) auction [19, 55], the auction that is most com-
monly used in practice (e.g. by Google and Bing), or on the design of truthful auctions [1]. All these auctions
rely on knowing the rates at which users click on the different advertisements (a.k.a. click-through rates, or
CTRs), and do not consider the process in which these CTRs are learned or refined over time by observing
users’ behavior. We argue that strategic agents would take this process into account, as it influences their
utility. While prior work [24] focused on the influence of click fraud on methods for learning CTRs, we
are interested in the implications of the strategic bidding by the agents. Thus, we consider the problem of
designing truthful sponsored search auctions when the process of learning the CTRs is a part of the game.
We are mainly interested in the interplay between the online learning and the strategic bidding. To
isolate this issue, we consider the following setting, which is a natural strategic version of the multi-armed
bandit (MAB) problem. In this setting, there are k ≥ 2 agents. Each agent i has a single advertisement, and
a private value vi > 0 for every click she gets. The mechanism is an online algorithm that first solicits bids
from the agents, and then runs for T rounds. In each round the mechanism picks an agent (using the bids
and the clicks observed in the past rounds), displays her advertisement, and receives a feedback – if there
was a click or not. Payments are charged after round T . Each agent tries to maximize her own utility: the
value that she derives from clicks minus the payment she pays. We assume that initially no information is
known about the likelihood of each agent to be clicked, and in particular there are no Bayesian priors.
We are interested in designing mechanisms which are truthful (in dominant strategies): every agent
maximizes her utility by bidding truthfully, for any bids of the others and for any clicks that would have
been received (that is, for any realization of the clicks an agent never regrets being truthful in retrospect).
The goal is to maximize the social welfare.1 Since the payments cancel out, this is equivalent to maximizing
the total value derived from clicks, where an agent’s contribution to that total is her private value times the
number of clicks she receives. We call this setting the MAB mechanism design problem.
In the absence of strategic behavior this problem reduces to a standard MAB formulation in which an
the
algorithm repeatedly chooses one of the k alternatives (“arms”) and observes the associated payoff:
value-per-click of the corresponding ad if the ad is clicked, and 0 otherwise. The crucial aspect in MAB
problems is the tradeoff between acquiring more information (exploration) and using the current information
to choose a good agent (exploitation). MAB problems have been studied intensively for the past three
decades. In particular, the above formulation is well-understood [6, 7, 16] in terms of regret relative to the
benchmark which always chooses the same “best” alternative (time-invariant benchmark). This notion of
regret naturally extends to the strategic setting outlined above, the total payoff being exactly equal to the
social welfare, and the regret being exactly the loss in social welfare relative to the time-invariant benchmark.
Thus one can directly compare MAB algorithms and MAB mechanisms in terms of welfare loss (regret).
Broadly, we ask how the design of MAB algorithms is affected by the restriction of truthfulness: what is
the difference between the best algorithms and the best truthful mechanisms? We are interested both in terms
of the structural properties and the gap in performance (in terms of regret). In short, we establish that the
1Social welfare includes both the auctioneer’s revenue and the agents’ utility. Since in practice different sponsored search plat-
forms compete against one another, taking into account the agents’ utility increases the platform’s attractiveness to the advertisers.
2
additional constraints imposed by truthfulness severely limit the structure and performance of online learn-
ing algorithms. We are not aware of any prior work that characterizes truthful online learning algorithms or
proves negative results on their performance.
Discussion. We believe that the fundamental limitations of truthfulness are best studied in simple models
such as the one defined above. We did not attempt to incorporate many additional aspects of pay-per-click
ad auctions such as information that is revealed to and by agents over time, multiple ad slots, user contexts,
ad features, etc. However, intuition from our impossibility results applies to richer models, and for some of
these models it is not difficult to produce precise corollaries. The key idea in the simple truthful mechanism
that we present (separating exploration and exploitation) can be easily extended as well.
We consider a strong notion of truthfulness: bidding truthfully is optimal for every possible click real-
ization (and bids of others). This notion is attractive as it does not require the agents to be risk neutral with
respect to the randomness inherent in clicks, or consider their beliefs about the CTRs. It allows for the CTRs
to change over time, and still incentivizes agents to be truthful. Moreover, an agent never regrets truthful
bidding in retrospect. It is desirable to understand what can be achieved with this notion before moving to
weaker notions, and thus we focus on this notion in this paper.
1.1 Our contributions
We present two main contributions: structural characterizations of (dominant-strategy) deterministic truthful
mechanisms, and lower bounds on the regret that such mechanisms must suffer. The regret suffered by
truthful mechanisms is significantly larger than the regret of the best MAB algorithms. We emphasize
that our characterization results hold regardless of whether the mechanism’s goal is to maximize welfare,
revenue, or any other objective.
Formally, a mechanism for the MAB mechanism design problem is a pair (A,P), where A is the al-
location rule (essentially, an MAB algorithm which also gets the bids as input), and P is the payment rule
that determines how much to charge each agent. Both rules can depend only on the observable quantities:
submitted bids and click events (clicks or non-clicks) for ads that have been displayed by the algorithm.
Since the allocation rule is an online algorithm, its decision in a given round can only depend on the click
events observed in the past.
The distinction between an allocation rule and a payment rule is essential in prior work on Mechanism
Design, and it is also essential for this paper. In particular, social welfare (and therefore regret) is completely
determined by the allocation rule. This is because welfare includes each payment twice, with opposite signs:
amount paid by an advertiser and amount received by the mechanism, and the two cancel out.
Characterization. The MAB mechanisms setting is a single-parameter auction, the most studied and
well-understood type of auctions. For such settings truthful mechanisms are fully characterized [38, 3]: a
mechanism is truthful if and only if the allocation rule is monotone (by increasing her bid an agent cannot
cause a decrease in the number of clicks she gets), and the payment rule is defined in a specific and, essen-
tially, unique way. Yet, we observe that this characterization is not the right characterization for the MAB
setting! The main problem is that if an agent is not chosen in a given round then the corresponding click
event is not observed by the mechanism, in the sense that the mechanism does not know whether this agent
would have received a click had it been selected in this round. Therefore the payment cannot depend on
any such unobserved click events. This is a non-trivial restriction because the naive payment computation
according to the formula mandated by [38, 3] requires simulating the run of the allocation rule for bids
different than the ones actually submitted, which in turn may depend on unobserved click events. We show
that this restriction has severe implications on the structure of truthful mechanisms.
3
The first notable necessary property of a truthful MAB mechanism is a much stronger version of mono-
tonicity which we call “pointwise monotonicity”:
Definition 1.1. A click realization consists of click information for all agents and all rounds: it specifies
whether a given agent receives a click if it is selected in a given round.2 An allocation rule is pointwise
monotone if for each click realization, each bid profile and each round, if an agent is selected at this round,
then she is also selected after increasing her bid (fixing everything else).
We first consider the case of two agents and show that truthful MAB mechanisms must have a strict
separation between exploration and exploitation, in the following sense. A crucial feature of exploration is
the ability to influence the allocation in forthcoming rounds. To make this point more concrete, we call a
round t influential for a given click realization, with influenced agent j, if for some bid profile changing the
click realization for this round can affect the allocation of agent j in some future round. We show that in
any influential round, the allocation can not depend on the bids. Thus, we show that influential rounds are
essentially useless for exploitation.
Definition 1.2. An MAB allocation rule A is called exploration-separated if for any click realization, the
allocation in any influential round does not depend on the bids.
In our model, agents derive value from clicks. In particular, an agent with zero value per click receives no
value. We focus on mechanisms in which a truthfully bidding agent with zero value-per-click pays exactly
zero; we call such mechanisms normalized. Among truthful single-parameter mechanisms, normalized
mechanisms are precisely the ones that satisfy two desirable properties: voluntary participation (truthfully
bidding agents never lose from participating), and no positive transfers (advertisers are charged, not paid).
We also make a mild assumption that an allocation rule is scale-free: invariant under multiplying all
bids by the same positive number, i.e. does not depend on the choice of the currency unit. Many MAB
algorithms from prior work can be easily converted into scale-free MAB allocation rules via some generic
ways to incorporate bids into algorithms’ specification.3
We are now ready to present our main structural result for two agents.
Theorem 1.3. Consider the MAB mechanism design problemwith two agents. Let A be a non-degenerate,4
deterministic, scale-free allocation rule. Then a mechanism (A,P) is normalized and truthful for some
payment rule P if and only if A is pointwise monotone and exploration-separated.
The case of more than two agents requires slightly more refined notions.
Definition 1.4. For a given realization and bid profile, a round is secured from an agent if that agent cannot
change the allocation at that round by increasing his bid. A deterministic MAB allocation rule is called
weakly separated if for every click realization and bid profile, if a round is influential for this realization and
bid profile, then it is secured from every agent that this round influences.
2Note that an MAB mechanism does not observe the entire click realization: it only observes click information for one agent
per round, the agent that was selected in this round.
3Many algorithms from prior work on stochastic MAB maintain an estimate νi of the expected reward for each arm i, such as
an upper confidence bound in UCB1 [6] or an independent sample from Bayesian posterior in Thompson’s Heuristic [54], so that
the algorithms’ decisions depend only on these estimates. An allocation rule can interpret νi as an estimate of the CTR, and use
i = bi νi instead of νi for all decisions. Moreover, any MAB algorithm can be converted to a scale-free MAB allocation rule by
ν ′
assigning a reward of bi/(maxj bj) to each agent i for each click on her ad. We use both approaches in this paper, in Section 5 and
Section 6.1, respectively.
4Non-degeneracy is a mild technical assumption, formally defined in “preliminaries”, which ensures that (essentially) if a given
allocation happens for some bid profile (bi, b−i) then the same allocation happens for all bid profiles (x, b−i), where x ranges over
some non-degenerate interval. Without this assumption, all structural results hold (essentially) almost surely w.r.t the k-dimensional
Lebesgue measure on the bid vectors. Exposition becomes significantly more cumbersome, yet leads to the same lower bounds on
regret. For clarity, we assume non-degeneracy throughout this paper.
4
The “weakly separated” condition is weaker than “exploration-separated”: while the latter ensures that
all agents cannot change the allocation at any given influential round t, the former only requires this for each
agent that is influenced by round t, fixing the bids of all other agents. For two agents and a scale-free MAB
allocation rule, the two conditions are equivalent.
Our complete characterization for any number of agents follows.
Theorem 1.5. Consider the MAB mechanism design problem. Let A be a non-degenerate deterministic
allocation rule. Then a mechanism (A,P) is normalized and truthful for some payment rule P if and only
if A is pointwise monotone and weakly separated.
Note that the general characterization does not require the allocation rule to be scale-free. In the special
case of two agents and scale-free allocation rules it implies Theorem 1.3.
We also investigate under which assumptions a weakly separated MAB allocation rule is exploration-
separated, as the latter condition is sufficient for proving performance limitations (bounds on regret). To this
end, we adapt a well-known notion from the literature on Social Choice, called Independence of Irrelevant
Alternatives (IIA, for short): an MAB allocation rule is IIA if for any given click realization, bid profile and
round, a change of bid of agent i cannot transfer the allocation in this round from agent j to agent l, where
these are three distinct agents. Note that the IIA condition trivially holds if there are only two agents. We
prove that for a non-degenerate deterministic allocation rule which is scalefree, pointwise monotone, and
satisfies IIA it holds that the rule is exploration-separated if and only if it is weakly separated. Technically,
assuming IIA allows us to extend our performance limitations results to more than two agents.5
Lower bounds on regret.
In view of the characterizations of truthful mechanisms, we present a lower
bound on the performance of exploration-separated algorithms. We consider a setting, termed the stochastic
MAB mechanism design problem, in which each click on a given advertisement is an independent random
event which happens with a fixed probability, a.k.a. the CTR. The expected “payoff” from choosing a given
agent is her private value times her CTR. For the ease of exposition, assume that the bids lie in the interval
[0, 1]. Then the non-strategic version is the stochastic MAB problem in which the payoff from choosing a
given arm i is an independent sample in [0, 1] with a fixed mean µi. In both versions, we compete with
the best-fixed-arm benchmark:
the hypothetical allocation rule (resp. algorithm) that always chooses an
arm with the maximal expected payoff. This benchmark is standard in the literature on stochastic MAB; it is
optimal among all MAB algorithms that are given the expected rewards for each arms (resp., among all MAB
allocation rules that are given the bids and the CTRs). We define regret as the expected difference between
the social welfare (resp. total payoff) of the benchmark and that of the allocation rule (resp. algorithm). The
algorithm’s goal is to minimize R(T ), worst-case regret over all problem instances on T rounds.
We show that the worst-case regret of any exploration-separated algorithm is larger than that of the
optimal MAB algorithm [7]: Ω(T 2/3) vs. O(√T ) for a fixed number of agents. We obtain an even more
pronounced difference if we restrict our attention to the δ-gap problem instances: instances for which the
best agent is better than the second-best by a (comparatively large) amount δ, that is µ1v1 − µ2v2 = δ ·
(maxi vi), where arms are arranged such that µ1v1 ≥ µ2v2 ≥ ··· ≥ µkvk. Such problem instances are
known to be easy for the MAB algorithms. Namely, an MAB algorithm can concurrently achieve the optimal
worst-case regret O(√kT log T ) and regret O( k
δ log T ) on δ-gap instances [32, 6]. However, we show
5Since prior work on MAB algorithms did not address strategic issues, these algorithms were not designed to satisfy properties
like (pointwise) monotonicity and IIA (and besides, these properties are not even well-defined for MAB algorithms, only for MAB
allocation rules). So it is not yet clear how limiting are these properties. The simple pointwise monotone MAB allocation rule
described later in the Introduction does satisfy IIA, but suffers from high regret. Designing better-performing MAB allocation
rules that are (pointwise) monotone appears quite challenging. For instance, such allocation rule is one of the main results in the
follow-up paper [9]. We leave open the question of existence of low-regret MAB allocation rules that are both pointwise-monotone
and IIA.
5
that for exploration-separated allocation algorithms the worst-case regret Rδ(T ) over the δ-gap instances
is polynomial in T (rather than poly-logarithmic in T ) as long as worst-case regret is even remotely non-
trivial (i.e., sublinear). Thus, for the δ-gap instances the gap in the worst-case regret between unrestricted
algorithms and exploration-separated algorithms is exponential in T .
Theorem 1.6. Consider the stochastic MAB mechanism design problem with k ≥ 2 agents. Let A be a deter-
ministic allocation rule that is exploration-separated. Then A has worst-case regret R(T ) = Ω(k1/3 T 2/3).
Moreover, if R(T ) = O(T γ) for some γ < 1 then for every fixed δ ≤ 1
4 and any ǫ > 0 the worst-case regret
over the δ-gap instances is Rδ(T ) = Ω(δ T 2(1−γ)−ǫ).
For two agents, Theorem 1.6 implies a significant gap in performance between truthful MAB mecha-
nisms and the best MAB algorithms, since truthful MAB mechanisms are necessarily exploration-separated.6
For example, while truthful MAB mechanisms suffer regret of Ω(T 2/3), the best algorithms have regret of
only O(√T ); as we described above, for δ-gap distances the difference in regret is even more pronounced.
For more than two agents, Theorem 1.6 does not immediately imply any regret bounds for truthful
MAB mechanisms. This is because the theorem requires the “exploration-separated” condition, whereas
the corresponding characterization result in Theorem 1.5 only guarantees the “weakly separated” condition.
Recall that one way to guarantee the “exploration-separated” condition (and therefore the regret bound) is
to furthermore assume IIA. It is an open question whether one can prove similar regret bounds for weakly
separated MAB allocation rules without assuming IIA.
We note that our lower bounds hold for a more general setting in which the values-per-click can change
over time, and the advertisers are allowed to change their bids at every time step.
Somewhat counter-intuitively, the lower bound on regret for k = 2 agents does not immediately imply
the same lower bound for any constant k > 2. This is, essentially, because our setting requires a mechanism
to show an ad in each round. A seemingly obvious approach to extend the lower bound from k = 2 to
(say) k = 3 is to assume, for the sake of contradiction, that there exists a truthful MAB mechanism M for
3 agents whose regret is less than the lower bound for two agents, and use M construct a truthful MAB
mechanism M′ for two agents with the same regret. (This would yield a contradiction, and hence prove
the lower bound for three agents.) The derived two-agent mechanism M′ adds a fictitious third agent (a
dummy) that never receives any clicks, and runs the original three-agent mechanism M. However, when M
picks the dummy agent, the two-agent mechanism must pick one of the two real agents. These additional
allocations may distort the agents’ incentives, so M′ is not guaranteed to be truthful. Hence, this reduction
is not guaranteed to work. Likewise, the allocation rule of M′ is not guaranteed to be weakly separated even
if the allocation rule of M is exploration-separated. Thus, we cannot immediately obtain a lower bound on
regret for more than two agents simply by combining the two-agent characterization in Theorem 1.3 and the
two-agent regret bound of Theorem 1.6.
Tightness: a positive result. To complete the picture for exploration-separated MAB allocation rules, we
present a very simple deterministic mechanism that is truthful and normalized, and matches the lower bound
R(T ) = Ω(k1/3 T 2/3) up to logarithmic factors. The allocation rule in this mechanism is exploration-
separated; it consists of two phases: an exploration phase in which agents are chosen in a round-robin
fashion, followed by an exploitation phase which allocates all rounds to the agent with the best empirical
performance in the exploration phase. Crucially, the duration of the exploration phase is fixed in advance
(and optimized given k and T ).
6Formally, this holds for truthful MAB allocation rules with allocation rules that satisfy the mild assumptions of non-degeneracy
and scale-freeness. We remove the latter assumption in one of the extensions.
6
Extensions. We extend our main results in several directions.
1. We derive a lower bound on regret for deterministic truthful mechanisms without assuming that the
allocations are scale-free. In particular, for two agents there are no assumptions. This lower bound
holds for any k (the number of agents) assuming IIA. However, the value of the lower bound does not
increase with k; in this sense this lower bound is weaker than the one in Theorem 1.6.
2. We consider randomized MAB mechanisms that are universally truthful, i.e. truthful for each realiza-
tion of the internal random seed. We extend the Ω(k1/3 T 2/3) lower bounds on regret to mechanisms
that randomize over exploration-separated deterministic MAB allocation rules.
3. We consider randomized MAB mechanisms under a weaker (less restrictive) version of truthfulness: a
mechanism is weakly truthful if for each click realization, it is truthful in expectation over its random
seed. We show that any randomized allocation that is pointwise monotone and satisfies a certain stong
notion of “separation between exploration and exploitation” can be turned into a mechanism that is
weakly truthful and normalized.
We apply this result to the version of the MAB mechanism design problem in which the clicks are
chosen by an oblivious adversary.7 (The corresponding algorithmic version is the adversarial MAB
problem [7, 14].) Using an MAB algorithm from the literature [8, 28], we obtain a weakly truthful
MAB mechanism for this problem with regret O((k log k)1/3 · T 2/3). This matches our lower bound
for deterministic MAB mechanisms up to (log k)1/3 factor.
4. The stochastic MAB mechanism design problem admits a very reasonable notion of truthfulness that is
even weaker: truthfulness in expectation, where for each vector of CTRs the expectation is taken over
clicks (and the internal randomness in the mechanism, if the latter is not deterministic).8 Following
our line of investigation, we ask whether restricting a mechanism to be truthful in expectation has
any implications on the structure and regret thereof. Given our negative results on mechanisms that
are truthful and normalized, it is tempting to seek similar results for mechanisms that are truthful in
expectation and normalized in expectation. We show that such approach is not likely to be fruitful.
Surprisingly, we prove that any monotone-in-expectation MAB allocation rule gives rise to an MAB
mechanism that is truthful in expectation and normalized in expectation, with a very minor increase in
regret. The key idea is to view the expected payments as multivariate polynomials over the CTRs, and
argue that any such polynomial can be “implemented” by a suitable payment rule. While this result is
purely theoretical, e.g. because the payments have very high variance, it implies that any impossibility
result for truthful-in-expectation MAB mechanisms must either follow directly from monotonicity-in-
expectation of the allocation rule, or requires bounds on the variability of the payments.
Informational obstacle. Our paper exposes a new kind of obstacle which might stands in the way of
designing truthful mechanisms: insufficient observable information to compute payments; we will term it
“informational obstacle” from here on.
Interestingly, this obstacle appears more general than the current setting. First, it would still feature
prominently in any mechanism design setting which can be modeled as one of the numerous MAB settings
studied in the literature. Second, and perhaps more importantly, we conjecture that it can be extended to a
very general class of mechanisms that interact with the environment. The follow-up work [56, 48] provides
some evidence to this conjecture, see Section 1.3 for more details.
7An oblivious adversary chooses the entire click realization in advance, without observing algorithm’s behavior.
8Normalized-in-expectation and monotone-in-expectation properties are defined similarly.
7
1.2 Additional related work
Mechanism Design. The question of how the performance of a truthful mechanism compares to that of the
optimal algorithm for the corresponding non-strategic problem is one of the central themes in Algorithmic
Mechanism Design. Performance gaps have been shown for various scheduling problems [3, 40, 18] and for
online auction for expiring goods [35]. Other papers presented approximation gaps due to computational
constraints, e.g. for combinatorial auctions [34, 18] and combinatorial public projects [43], showing a gap
via a structural result for truthful mechanisms.
The intersection of Machine Learning and Mechanism Design is an active research area which includes
work in various topics such as online mechanisms [35], dynamic auctions [13, 4], dynamic pricing [46],
secretary problems [21], offline learning from self-interested data sources [10, 37] and a number of others.
A more detailed review of this area, or any of the topics listed above, is beyond the scope of this paper.
MAB mechanisms. MAB algorithms were used in the design of Cost-Per-Action sponsored search auc-
tions in Nazerzadeh et al. [39], where the authors construct a mechanism with approximate (asymptotic)
properties of truthfulness and individual rationality. However, even if the gains from lying are small, it may
still be rational for the agents to deviate from being truthful, perhaps significantly. Moreover, as truthful
bidding is not a Nash equilibrium, an agent may speculate that other agents will deviate, which in turn may
increase her own incentives to deviate. All of that may result in unpredictable, and possibly highly subopti-
mal outcomes. On the other hand, approximate truthfulness guarantees suffice whenever it is reasonable to
assume that the agents would not lie unless it leads to significant gains.
In a concurrent and independent work with respect to this paper, Devanur and Kakade [17] considered
the same setting: deterministic truthful MAB mechanisms. They focus on maximizing the revenue of the
mechanism (as opposed to the social welfare). They present an impossibility result for the two-agent case:
a lower bound of Ω(T 2/3) on the loss in revenue with respect to the VCG payments; this bound is extended
to deterministic MAB mechanisms that are truthful with high probability. They also provide a deterministic
truthful mechanism which matches the above lower bound, and is almost identical to our simple two-phase
mechanism described in Section 1.1.9
A closely related line of work on dynamic auctions [13, 4, 44, 25] considers a more general setting in
which private information is revealed to agents over time. The mechanism needs to create the right incentives
for the agents to reveal all the information they receive over time, and to stay in the auction after every round;
these challenges do not exist in our setting, in which all private information is known to the agents upfront.
On the other hand, these papers study fully Bayesian settings in which Bayesian priors on CTRs are known
and VCG-like social welfare-maximizing mechanisms are therefore feasible. In our setting – with no priors
on CTRs – VCG-style mechanisms cannot be applied as such mechanisms require the allocation to exactly
maximize the expected social welfare, which is impossible (and even not well-defined) without a prior.
Moreover, even if applied to MAB mechanisms with Baeysian priors over CTRs, the techniques from this
line of work can only guarantee truthfulness in expectation over the Bayesian prior, which is a much weaker
notion compared to the “prior-independent” notions of truthfulness that are studied in this paper.
Multi-armed bandits (MAB). Absent the strategic constraint, our problem fits into the framework of
MAB algorithms. MAB has a rich literature in Statistics, Operations Research, Computer Science and
Economics; a reader can refer to [14, 12] for background. Most relevant to the present paper is the work on
stochastic MAB [32, 6] and adversarial MAB [7]. Both directions have spawned vast amounts of follow-up
research. Results used in this paper come from [6, 32, 7, 5, 8, 28].
9This mechanism is for a more general setting in which values-per-click change over time and the agents are allowed to submit
a different bid at every round. Instead of assigning all impressions to the same agent in the exploitation phase, their mechanism
runs the same allocation and payment procedure for each exploitation round separately, with the bids submitted in this round.
8
Our lower bounds on regret use (a novel application of) the relative entropy technique from [32, 7],
see [29] for an account. This is the technique typically used to prove lower bound on regret for MAB and
related problems. For other application of this technique, see e.g. [16, 26, 30, 11].
The prior work on MAB algorithms considered numerous MAB settings with various assumptions on
payoff evolution over time (e.g., [7, 51, 23]), dependencies between arms (e.g., [20, 42, 30, 52]), side
information available to an algorithm (e.g., [30, 33, 49]), etc. Many of these settings are motivated by
pay-per-click ad auctions. For every such MAB setting one could define the corresponding version of the
MAB mechanism design problem.
1.3 Follow-up work
The conference publication of this paper gave rise to a several follow-up papers [9, 56, 22, 48] which have
addressed some of the questions left open by this paper and posed some new ones. Below we present the
current snapshot of this line of work.
One direction concerns weakly truthful, randomized MAB mechanisms. Informally, the main question
here is whether they are significantly more powerful than their deterministic counterparts. Babaioff, Klein-
berg and Slivkins [9] resolve this question in the affirmative: they prove that there exist weakly truthful
randomized MAB mechanisms whose regret bounds for the stochastic MAB setting are optimal for MAB
algorithms, both in the worst case and for δ-gap instances. A major component of this result, henceforth
called the BKS reduction, reduces designing weakly truthful MAB mechanisms to designing MAB alloca-
tion rules that satisfy the appropriate notion of monotonicity called weak monotonicity: an MAB allocation
is weakly monotone if for each click realization, it is monotone in expectation over its random seed.10 The
BKS reduction subsumes and generalizes our result on truthfulness in expectation (using a very different
technique). Moreover, it is not specific to the stochastic MAB setting: it extends beyond MAB mechanisms
to arbitrary single-parameter domains (see [41] for more background). In particular, the BKS reduction
applies to MAB mechanisms with clicks chosen by an oblivious adversary, and to MAB mechanism design
problems based on most other settings studied in the vast literature on MAB algorithms.
Our truthful-in-expectation construction and the BKS reduction suffer from a very high variance in
payments. Both results include an explicit tradeoff between the variance in payments and the loss in perfor-
mance. Very recently, Wilkens and Sivan [56] have proved that the tradeoff in the BKS reduction is optimal
in a certain worst-case sense: the BKS reduction achieves the optimal worst-case variance in payments for
any given worst-case loss in performance, where the worst case is over all monotone MAB allocation rules.
(More generally, the optimality result in [56] applies to any given single-parameter problem.)
Additional developments in [9] concern MAB allocation rules. First, they prove that an MAB allocation
rule based on UCB1 satisfies monotonicity-in-expectation, and therefore can be transformed (using our result
from Section 7 or the BKS reduction) to a truthful-in-expectation MAB mechanism with essentially the same
regret. Second, they provide a new deterministic MAB allocation rule called NewCB which has optimal
regret and is monotone. In conjunction with the BKS reduction, NewCB yields the weakly truthful MAB
mechanism discussed above.
The analysis in this paper provides a strong intuition that the crucial obstacle for deterministic MAB
mechanisms is not the monotonicity of an allocation rule but instead the “informational obstacle”: insuffi-
cient observable information to compute payments. The analysis of NewCB in [9] makes this point rigorous.
Moreover, [56, 48] describe some additional settings, different from MAB mechanisms, where this “infor-
mational obstacle” arises. Wilkens and Sivan [56] provide two variants of offline pay-per-click ad auctions
with multiple ad slots. Shneider et al. [48] describe a packet scheduling problem in a network router, where
the potentially non-observable information is the packet arrival times (rather than the click events). They
10[9] uses a somewhat different (and perhaps more systematic) terminology regarding the different notions of truthfulness, mono-
tonicity and normalization. We discuss the results from [9] using the terminology of the present paper.
9
observe that in the network router setting information about packet arrival times may be missing not only
because it is not observed by the router but also because the router does not have much space to store it.
Finally, a very recent paper by Gatti, Lazaric and Trovo [22] considers multi-slot MAB mechanisms, i.e.
pay-per-click ad auctions with multiple ad slots and unknown CTRs. This setting combines multi-slot pay-
per-click ad auctions [55, 19] on the mechanism design side, and multi-slot MAB [45, 53] on the learning
side. The authors provide truthful multi-slot MAB mechanisms based on the simple MAB mechanism
presented in this paper and (independently) in Devanur and Kakade [17].
Despite all these exciting development, MAB mechanisms are not well-understood; see Section 8 for
the current snapshot of open questions.
1.4 Map of the paper
Section 2 is preliminaries. Truthfulness characterization is developed and proved in Section 3 and Section A.
The lower bounds on regret are presented in Section 4. The simple mechanism that matches these lower
bounds is in Section 5. Weakly truthful randomized allocations for adversarial clicks are derived in Section 6.
Truthfulness in expectation is discussed in Section 7. Open questions are in Section 8.
2 Definitions and preliminaries
In the MAB mechanism design problem, there is a set K of k agents numbered from 1 to k. Each agent
i has a value vi > 0 for every click she gets; this value is known only to agent i. Initially, each agent i
submits a bid bi > 0, possibly different from vi. 11 12 The “game” lasts for T rounds, where T is the given
time horizon. A click realization represents the click information for all agents and all rounds. Formally,
it is a tuple ρ = (ρ1 , . . . , ρk) such that for every agent i and round t, the bit ρi(t) ∈ {0, 1} indicates
whether i gets a click if selected at round t. An instance of the MAB mechanism design problem consists
of the number of agents k, time horizon T , a vector of private values v = (v1, . . . , vk), a vector of bids (bid
profile) b = (b1, . . . , bk), and click realization ρ.
A mechanism is a pair (A,P), where A is allocation rule and P is the payment rule. An allocation rule
is represented by a function A that maps bid profile b, click realization ρ and a round t to the agent i that
is chosen (receives an impression) in this round: A(b; ρ; t) = i. We also denote Ai(b; ρ; t) = 1{A(b;ρ;t)=i}.
The allocation is online in the sense that at each round it can only depend on clicks observed prior to that
round. Moreover, it does not know the click realization in advance; in every round it only observes the click
realization for the agent that is shown in that round. A payment rule is a tuple P = (P1 , . . . ,Pk), where
Pi(b; ρ) ∈ R denotes the payment charged to agent i when the bids are b and the click realization is ρ. 13
Again, the payment can only depend on observed clicks.
A mechanism is called normalized if for any agent i, bids b−i of the other agents, and click realization
ρ it holds that Pi(bi, b−i; ρ) → 0 as bi → 0. For any single-parameter, truthful mechanism, this limit exists
and is independent of bi [38, 3]; further, this limit is always 0, for a given agent i, if and only if the payment
per click is between 0 and bi.
11One can also consider a more realistic and general model in which the value-per-click of an agent changes over time and the
agents are allowed to change their bid at every round. The case that the value-per-click of each agent does not change over time
is a special case. In that case truthfulness implies that each agent basically submits one bid as in our model (the same bid at every
round), thus our main results (necessary conditions for truthfulness and regret lower bounds) also hold for the more general model.
12Since private values vi are strictly positive, there is no need to allow zero bids. Also, this avoids some technical complications
in the proofs. Accordingly, we define “normalized mechanisms” in terms of the payment as bi → 0.
13We allow the mechanism to determine the payments at the end of the T rounds, and not after every round. This makes that task
of designing a truthful mechanism easier and thus strengthen our necessary condition for truthfulness (the condition used to derive
the lower bounds on regret.)
10
(2.1)
For given click realization ρ and bid profile b, the number of clicks received by agent i is denoted
Ci(b; ρ). Call C = (C1 , . . . ,Ck) the click-allocation for A. The utility that agent i with value vi gets from
the mechanism (A,P) when the bids are b and the click realization is ρ is Ui(vi; b; ρ) = vi·Ci(b; ρ)−Pi(b; ρ)
(quasi-linear utility). The mechanism is truthful if for any agent i, value vi, bid profile b and click realization
ρ it is the case that Ui(vi; vi, b−i; ρ) ≥ Ui(vi; bi, b−i; ρ).
In the stochastic MAB mechanism design problem, an adversary specifies a vector µ = (µ1 , . . . , µk) of
CTRs (concealed from A), then for each agent i and round t, click realization ρi(t) is chosen independently
with mean µi. Thus, an instance of the problem includes µ rather than a fixed click realization. For a given
problem instance I, let i∗ ∈ argmaxi µi vi, then regret on this instance is defined as
i=1 µi vi Ai(b; ρ; t)i .
RI(T ) = T vi∗µi∗ − EhPT
t=1Pk
For a given parameter vmax, the worst-case regret14 R(T ; vmax) denotes the supremum of RI(T ) over all
problem instances I in which all private values are at most vmax. Similarly, we define Rδ(T ; vmax), the
worst-case δ-regret, by taking the supremum only on instances with δ-gap.
Most of our results are stated for non-degenerate allocation rules, defined as follows. An interval is
called non-degenerate if it has positive length. Fix bid profile b, click realization ρ, and rounds t and t′ with
t ≤ t′. Let i = A(b; ρ; t) and ρ′ be the allocation obtained from ρ by flipping the bit ρi(t). An allocation
rule A is non-degenerate w.r.t. (b, ρ, t, t′) if there exists a non-degenerate interval I containing bi such that
Ai(x, b−i; ϕ; s) = Ai(b; ϕ; s)
for each ϕ ∈ {ρ, ρ′}, each s ∈ {t, t′}, and all x ∈ I.
An allocation rule is non-degenerate if it is non-degenerate w.r.t. each tuple (b, ρ, t, t′).
3 Truthfulness characterization
Before presenting our characterization we begin by describing some related background. The click alloca-
tion C is non-decreasing if for each agent i, increasing her bid (and keeping everything else fixed) does not
decrease Ci. Prior work has established a characterization of truthful mechanisms for single-parameter do-
mains (domains in which the private information of each agent is one-dimensional), relating click allocation
monotonicity and truthfulness (see below). For our problem, this result is a characterization of MAB algo-
rithms that are truthful for a given click realization ρ, assuming that the entire click realization ρ can be used
to compute payments (when computing payments one can use click information for every round and every
agent, even if the agent was not shown at that round.) One of our main contributions is a characterization
of MAB allocation rules that can be truthfully implemented when payment computation is restricted to only
use clicks information of the actual impressions assigned by the allocation rule.
3.1 Monotonicity
An MAB allocation rule A is truthful with unrestricted payment computation if it is truthful with a payment
rule that can use the entire click realization ρ in it computation. We next present the prior result character-
izing truthful mechanisms with unrestricted payment computation.
Theorem 3.1 (Myerson [38], Archer and Tardos [3]). Let (A,P) be a normalized mechanism for the MAB
mechanism design problem. It is truthful with unrestricted payment computation if and only if for any given
click realization ρ the corresponding click-allocation C is non-decreasing and the payment rule is given by
(3.1)
Pi(bi, b−i; ρ) = bi · Ci(bi, b−i; ρ) −R bi
0 Ci(x, b−i; ρ) dx.
14By abuse of notation, when clear from the context, the “worst-case regret” is sometimes simply called “regret”.
11
We can now move to characterize truthful MAB mechanisms when the payment computation is re-
stricted. The following notation will be useful: for a given click realization ρ, let ρ ⊕ 1(i, t), be the click
realization that coincides with ρ everywhere, except that the bit ρi(t) is flipped.
The first notable property of truthful mechanisms is a stronger version of monotonicity. Recall (see
Definition 1.1) that an allocation rule A is pointwise monotone if for each click realization ρ, bid profile b,
round t and agent i, if Ai(bi, b−i; ρ; t) = 1 then Ai(b+
i > bi. In words, increasing
a bid cannot cause a loss of an impression.
Lemma 3.2. Consider the MAB mechanism design problem. Let (A,P) be a normalized truthful mechanism
such that A is a non-degenerate deterministic allocation rule. Then A is pointwise-monotone.
Proof. For a contradiction, assume not. Then there is a click realization ρ, a bid profile b, a round t and
agent i such that agent i loses an impression in round t by increasing her bid from bi to some larger value
b+
i . In other words, we have Ai(b+
i , b−i; ρ; t) < Ai(bi, b−i; ρ; t). Without loss of generality, let us assume
that there are no clicks after round t, that is ρj(t′) = 0 for any agent j and any round t′ > t (since changes
in ρ after round t does not affect anything before round t).
i , b−i; ρ; t) = 1 for any b+
Let ρ′ = ρ ⊕ 1(i, t). The allocation in round t cannot depend on this bit, so it must be the same
for both click realizations. Now, for each click realization ϕ ∈ {ρ, ρ′} the mechanism must be able to
compute the price for agent i when bids are (b+
i , b−i). That involves computing the integral Ii(ϕ) =
Rx≤b+
i Ci(x, b−i; ϕ) dx from (3.1). We claim that Ii(ρ) 6= Ii(ρ′). However, the mechanism cannot dis-
tinguish between ρ and ρ′ since they only differ in bit (i, t) and agent i does not get an impression in round
t. This is a contradiction.
It remains to prove the claim. Without loss of generality, assume that ρi(t) = 0 (otherwise interchange
the role of ρ and ρ′). We first note that Ci(x, b−i; ρ) ≤ Ci(x, b−i; ρ′) for every x. This is because everything
is same in ρ and ρ′ until round t (so the impressions are same too), there are no clicks after round t, and in
round t the behavior of A on the two click realizations can be different only if that agent i gets an impression,
in which case she is clicked under ρ′ and not clicked under ρ.
Since A is non-degenerate, there exists a non-degenerate interval I containing bi such that changing bid
of agent i to any value in this interval does not change the allocation at round t (both for ρ and for ρ′). For
any x ∈ I we have Ci(x, b−i; ρ) < Ci(x, b−i; ρ′), where the difference is due to the click in round t. It
follows that Ii(ρ) < Ii(ρ′). Claim proved. Hence, the mechanism cannot be implemented truthfully.
3.2 Structural definitions
Let us restate the structural definitions from the Introduction in a more detailed fashion.
Definition 3.3. Fix click realization ρ, bid vector b, and round t.
(a) Round t is called (b; ρ)-secured from agent i if A(b+
(b) Round t is called bid-independent w.r.t. ρ if the allocation A(b; ρ; t) is a constant function of b.
(c) Round t is called (b; ρ)-influential if for some round t′ > t it holds that A(b; ρ; t′) 6= A(b; ρ′; t′) for
click realization ρ′ = ρ ⊕ 1(j, t) such that j = A(b; ρ; t). 15 In words: changing the relevant part of
the click realization at round t affects the allocation in some future round t′.
i , b−i; ρ; t) = A(bi, b−i; ρ; t) for any b+
i > bi.
(d) In part (c), round t′ is called the influenced round and j is called the influencing agent of round t. The
agent i is called an influenced agent of round t if i ∈ {A(b; ρ; t′), A(b; ρ′; t′)}.
15Note that click realizations ρ and ρ′ are interchangeable.
12
(e) Round t is called influential w.r.t. click realization ρ if and only if it is (b, ρ)-influential for some b.
Definition 3.4. Let A be a deterministic MAB allocation rule.
• A is called exploration-separated if for every click realization ρ and round t that is influential for ρ,
it holds that A(b; ρ; t) = A(b′; ρ; t) for any two bid vectors b, b′ (in words: allocation at round t does
not depend on the bids).
• A is called weakly separated if for every click realization ρ and bid vector b, it holds that if round t is
(b; ρ)-influential with influenced agent i then it is (b; ρ)-secured from i.
Observation 3.5. Any deterministic, exploration-separated MAB allocation rule is weakly separated.
Proof. It follows from the definitions. Fix click realization ρ and bid vector b, let t be a (b; ρ)-influential
round with influenced agent i. We need to show that t is (b; ρ)-secured from i. Round t is (b; ρ)-influential,
thus influential w.r.t. ρ, thus (since the allocation is exploration-separated) it is bid-independent w.r.t. ρ, thus
agent i cannot change allocation in round t by increasing her bid.
Observation 3.6. Let A be a scale-free, weakly separated MAB allocation rule for two agents. Then A is
exploration-separated.
The proof of this observation is fairly straightforward, but it requires to carefully unwind the definitions.
To provide some intuition with these definitions, we write it out in detail.
Proof of Observation 3.6. Fix a click realization ρ and round t that is influential for ρ. Let b, b′ be two bid
vectors. We need to conclude that A(b; ρ; t) = A(b′; ρ; t).
By definition of “influential round”, there exists some bid vector b∗ such that t is (b∗, ρ)-influential with
influenced agent i. Since there are only two agents, the other agent is influenced, too. By definition of
“weakly separated”, round t is (b∗, ρ)-secured from both agents. By definition of “secured”, we have:
A(b∗; ρ; t) = A(b+
= A(b∗
2; ρ; t) for any b+
2 ; ρ; t) for any b+
Let us prove that A(b; ρ; t) = A(b∗; ρ; t). We consider two cases.
• Suppose b1/b2 ≥ b∗
1 , b∗
1, b+
2; ρ; t). Since λb1 > b∗
1, then we are done by taking b+
A(λb1, b∗
1/b∗
2. Then by definition of “scale-free”, letting λ = b∗
1 > b∗
1
2 > b∗
2.
(3.2)
(3.3)
2/b2 we have A(b; ρ; t) =
1 = λb1 and using (3.2).
• Suppose b1/b2 < b∗
A(b∗
1/b∗
2. Then by definition of “scale-free”, letting λ = b∗
1, λb2; ρ; t). Since λb2 > b∗
2, then we are done by taking b+
2 = λb2 and using (3.3).
1/b1 we have A(b; ρ; t) =
Claim proved. Similarly, A(b′; ρ; t) = A(b∗; ρ; t).
3.3 The two agents case (Theorem 1.3)
The two-agent structural characterization in Theorem 1.3 follows from the general characterization in The-
orem 1.5. More precisely, the “if” direction of Theorem 1.3 follows from the “if” direction of Theorem 1.5
and Observation 3.5; the “only if” direction of Theorem 1.3 follows from the “only if” direction of Theo-
rem 1.5 and Observation 3.6.
The main structural implication in both theorems is that truthfulness implies the corresponding structural
condition (either that the allocation rule is exploration separated or that it is weakly separated.) To illustrate
the ideas behind this implication, we prove the two-agent case directly.
13
Proposition 3.7. Consider the MAB mechanism design problem with two agents. Let A be a non-degenerate
scale-free deterministic allocation rule. If (A,P) is a normalized truthful mechanism for some P, then it is
exploration separated.
Proof. Assume A is not exploration-separated. Then there is a counterexample (ρ, t): a click realization ρ
and a round t such that round t is influential and allocation in round t depends on bids. We want to prove
that this leads to a contradiction.
Let us pick a counterexample (ρ, t) with some useful properties. Since round t is influential, there exists
a click realization ρ and bid profile b such that the allocation at some round t′ > t (the influenced round)
is different under click realization ρ and another click realization ρ′ = ρ ⊕ 1(j, t), where j = A(b; ρ; t) is
the agent chosen at round t under ρ. Without loss of generality, let us pick a counterexample with minimum
value of t′ over all choices of (b, ρ, t). For ease of exposition, from this point on let us assume that j = 2.
For the counterexample we can also assume that ρ1(t′) = 1, and that there are no clicks after round t′, that
is ρl(t′′) = ρ′
l(t′′) = 0 for all t′′ > t′ and for all l ∈ {1, 2}.
We know that the allocation in round t depends on bids. This means that agent 1 gets an impression in
round t for some bid profile b = (b1, b2) under click realization ρ, that is A(b; ρ; t) = 1. As the mechanism
is scale-free this means that, denoting b+
1 , b2; ρ; t) = 1. Since A(b1, b2; ρ; t) = 2
and A(b+
1 > b1. We conclude that there
exists a bid b+
1 , b2; ρ; t) = 1, pointwise monotonicity (Lemma 3.2) implies that b+
1 = b1 b2/b2 we have A(b+
1 , b2; ρ; t) = 1.
1 , b2; ρ) and P1(b+
Now, the mechanism needs to compute prices for agent 1 for bids (b+
1 , b2) under click realizations ρ
i , b2; ρ′). Therefore, the mechanism needs to compute the integral
1 C1(x, b2; ϕ) dx for both click realizations ϕ ∈ {ρ, ρ′}.
and ρ′, that is P1(b+
I1(ϕ) =Rx≤b+
First of all, for all x ≤ b+
1 and for all t′′ < t′, A(x, b2; ρ; t′′) = A(x, b2; ρ′; t′′), since otherwise the
Let us assume A1(b1, b2; ρ; t′) < A1(b1, b2; ρ′, t′) (otherwise, we can swap ρ and ρ′). We make the
minimality of t′ will be violated. The only difference in the allocation can occur in round t′.
1 > b1 for agent 1 such that A(b+
claim that for all bids x ≤ b+
1 of agent 1, the influence of round t on round t′ is in the same “direction”:
A1(x, b2; ρ; t′) ≤ A1(x, b2; ρ′; t′) for all x ≤ b+
1 .
(3.4)
Suppose (3.4) does not hold. Then there is an x < b+
1 such that 1 = A1(x, b2; ρ; t′) > A1(x, b2; ρ′; t′) = 0.
(Note that we have used the fact that the mechanism is deterministic.) If x < b1 then pointwise monotonicity
is violated under click realization ρ, since A1(x, b2; ρ; t′) > A1(b1, b2; ρ; t′); otherwise it is violated under
click realization ρ′, giving a contradiction in both cases. The claim (3.4) follows.
Since A is non-degenerate, there exists a non-degenerate interval I containing bi such that if agent 1
bids any value x ∈ I then A1(x, b2; ρ; t′) < A1(x, b2; ρ′; t′). Now by (3.4) it follows that I1(ρ) < I2(ρ′).
However, the mechanism cannot distinguish between ρ and ρ′ when the bid of agent 1 is b+
1 , since the
differing bit ρ2(t) is not observed. Therefore the mechanism cannot compute prices, contradiction.
3.4 The general case (Theorem 1.5)
Let us prove the general characterization (Theorem 1.5). We restate it here for convenience.
Theorem (Theorem 1.5, restated). Consider the MAB mechanism design problem. Let A be a non-degenerate
deterministic allocation rule. Then a mechanism (A,P) is normalized and truthful for some payment rule
P if and only if A is pointwise monotone and weakly separated.
Proof of Theorem 1.5: the “only if” direction. Suppose (A,P) be a normalized truthful mechanism, for
some payment rule P. Then A is pointwise-monotone by Lemma 3.2. The fact that A is weakly sepa-
rated is proved similarly to Proposition 3.7, albeit with a few extra details.
14
Assume A is not weakly separated. Then there is a counterexample (ρ, b, t, t′, i): a click realization
ρ, bid vector b, rounds t, t′ and agent i such that round t is (b; ρ)-influential with influenced agent i and
influenced round t′ and it does not holds that round t is (b; ρ)-secured from i. We prove that this leads to a
contradiction..
there exists a bid b+
i , b−i; ρ; t) 6= j.
Let us pick a counterexample (ρ, b, t, t′, i) with a minimum value of t′ over all choices of (ρ, b, t, i).
Without loss of generality, let us assume that ρi(t′) = 1 and ρj(t′′) = 0 for all t′′ > t′ and for all agents j.
Let j = A(b; ρ; t). As it does not holds that round t is (b; ρ)-secured from i, this means that j 6= i, and
i > bi such that A(b+
Let ρ′ = ρ ⊕ 1(j, t). The mechanism needs to compute prices for agent i when her bid is b+
i under click
realizations ρ and ρ′, that is to compute Pi(b+
i , b−i; ρ′). Therefore, the mechanism needs
to compute the integral Ii(ϕ) =Rx≤b+
1 Ci(x, b−i; ϕ) dx for both click realizations ϕ ∈ {ρ, ρ′}.
First of all, for all x ≤ b+
i and for all t′′ < t′, Ai(x, b−i; ρ; t′′) = Ai(x, b−i; ρ′; t′′). If not,then the
minimality of t′ will be violated. This is because, if there were such an x and t′′ < t′ with Ai(x, b−i; ρ; t′′) 6=
Ai(x, b−i; ρ′; t′′), then round t will still be (b, ρ)-influential with influenced agent i, and influenced round
t′′ < t′, violating the minimality of t′′. Therefore, when we decrease the bid of agent i, the only difference
in the allocation can occur at time round t′.
i , b−i; ρ) and Pi(b+
As i is the influenced agent at round t′ it must hold that Ai(bi, b−i; ρ; t′) 6= Ai(bi, b−i; ρ′, t′). Let us
assume 0 = Ai(bi, b−i; ρ; t′) < Ai(bi, b−i; ρ′, t′) = 1 (otherwise, we can swap ρ and ρ′). Note that we have
made use of the fact that the mechanism is deterministic. Let us make the the claim that for all bids x ≤ b+
the influence of round t on round t′ is in the same “direction.”
i
Ai(x, b−i; ρ; t′) ≤ Ai(x, b−i; ρ′; t′) for all x ≤ b+
i .
(3.5)
Suppose (3.5) does not hold. Then there is an x ≤ b+
i such that 1 = Ai(x, b−i; ρ; t′) > Ai(x, b−i; ρ′; t′) = 0.
(Note that we have used the fact that the mechanism is deterministic.) If x > bi, then pointwise monotonicity
is violated in ρ′, since 0 = Ai(x, b−i; ρ′; t′) < Ai(bi, b−i; ρ′; t′) = 1. If x < bi on the other hand, then
the pointwise-monotonicity is violated in ρ, since 1 = Ai(x, b−i; ρ; t′) > Ai(bi, b−i; ρ; t′) = 0, giving a
contradiction in both cases. The claim (3.5) follows.
By the non-degeneracy of A, there exists a non-degenerate interval I containing bi such that
Ai(x, b−i; ρ; t′) < Ai(x, b−i; ρ′; t′) for all x ∈ I.
(3.6)
i , since the differing bit ρj(t) is not seen. Contradiction.
By (3.5) and (3.6) it follows that Ii(ρ) < Ii(ρ′). However, the mechanism cannot distinguish between ρ
and ρ′ when agent i’s bid is b+
Proof of Theorem 1.5: the “if” direction. Let A be a deterministic allocation rule which is pointwise mono-
tone and weakly separated. We need to provide a payment rule P such that the resulting mechanism (A,P)
is truthful and normalized. Since A is pointwise monotone, it immediately follows that it is monotone (i.e.,
as an agent increases her bid, the number of clicks that she gets cannot decrease). Therefore it follows from
Theorem 3.1 that mechanism (A,P) is truthful and normalized if and only if P is given by (3.1). We need
to show that P can be computed using only the knowledge of the clicks (bits from the click realization) that
were revealed during the execution of A.
Assume we want to compute the payment for agent i in bid profile (bi, b−i) and click realization ρ. We
will prove that we can compute Ci(x) := Ci(x, b−i; ρ) for all x ≤ bi. To compute Ci(x), we show that it
is possible to simulate the execution of the mechanism with bidi = x. In some rounds, the agent i loses
an impression, and in others it retains the impression (pointwise monotonicity ensures that agent i cannot
gain an impression when decreasing her bid). In rounds that it loses an impression, the mechanism does
not observe the bits of ρ in those rounds, so we prove that those bits are irrelevant while computing Ci(x).
In other words, while running with bidi = x, if mechanism needs to observe the bit that was not revealed
15
when running with bidi = bi, we arbitrarily put that bit equal to 1 and simulate the execution of A. We
want to prove that this computes Ci(x) correctly.
Let t1 < t2 < ··· < tn be the rounds in which agent i did not get an impression while bidding x, but did
get an impression while bidding bi. Let ρ0 := ρ, and let us define click realization ρl inductively for every
l ∈ [n] by setting ρl := ρl−1 ⊕ 1(jl, tl), where jl = A(x, b−i; ρl−1; tl) is the agent that got the impression
at round tl with click realization ρl−1 and bids (x, b−i).
First, we claim that jl 6= i for any l. Indeed, suppose not, and pick the smallest l such that jl+1 = i.
Then tl is a (x, b−i; ρl)-influential round, with influenced agent jl+1 = i. Thus tl is (x, b−i; ρl)-secured
from i. Since A(x, b−i; ρl; tl) = A(x, b−i; ρl−1; tl) = jl 6= i by minimality of l, agent i does not get an
impression in round tl if she raises her bid to bi. That is, A(b; ρl; tl) 6= i. However, the changes in click
realizations ρ0 , . . . , ρl−1 only concern the rounds in which agent i is chosen, so they are not seen by the
allocation if the bid profile is b (to prove this formally, use induction). Thus, A(b; ρl; tl) = A(b; ρ; tl) = i,
contradiction. Claim proved. It follows that A(b; ρ; tl) = i for each l. (This is because by induction, the
change from ρl−1 to ρl is not seen by the allocation if the bid profile is b.)
We claim that Ai(x, b−i; ρ; t′) = Ai(x, b−i; ρn; t′) for every round t′, which will prove the theorem. If
not, then there exists l such that Ai(x, b−i; ρl; t′) 6= Ai(x, b−i; ρl−1; t′) for some t′ (and of course t′ > tl).
Round tl is thus (x, b−i; ρl)-influential with influenced round t′ and influenced agent i. Moreover, the
influencing agent of that round is jl, and we already proved that jl 6= i. Since round tl is (x, b−i; ρl)-secured
from agent i due to the “weakly separated” condition, it follows that agent i does not get an impression in
round tl if she raises her bid to bi. That is, A(b; ρl; tl) 6= i, contradiction.
Let us argue that the non-degeneracy assumption in Theorem 1.5 is indeed necessary.
Claim 3.8. There exists a deterministic mechanism (A,P) for two agents that is truthful and normalized,
such that the allocation rule A is pointwise monotone, scale-free and yet not weakly separated.
Proof. There are only two rounds. Agent 1 allocated at round 1 if and only if b1 ≥ b2. Agent 1 allocated at
round 2 if b1 > b2 or if b1 = b2 and ρ1(1) = 1; otherwise agent 2 is shown. This completes the description
of the allocation rule. To obtain a payment rule P which makes the mechanism normalized and truthful,
consider an alternate allocation rule A′ which in each round selects agent 1 if and only if b1 ≥ b2. (Note that
A′ = A except when b1 = b2.) Use Theorem 1.5 for A′ to obtain a normalized truthful mechanism (A′,P ′),
and set P = P ′. The payment rule P is well-defined since the observed clicks for P and P ′ coincide unless
b1 = b2, in which case both payment rules charge 0 to both agents. The resulting mechanism (A,P) is
normalized and truthful because the integral in (3.1) remains the same even if we change the value at a
single point. It is easy to see that the allocation rule A has all the claimed properties; it fails to be non-
degenerate because round t is influential only when b1 = b2.
3.5 Scalefree and IIA allocation rules
We show that under the right assumptions, an MAB allocation rule is exploration-separated if and only if it
is weakly separated.
Lemma 3.9. Consider the MAB mechanism design problem. Let A be a non-degenerate deterministic
allocation rule which is scalefree, pointwise monotone, and satisfies IIA. Then it is exploration-separated if
and only if it is weakly separated.
The proof of Lemma 3.9 is very technical. We precede it with a proof sketch. To preserve the flow, we
place the full proof in Appendix A.
16
t
∗
(b+
i′ , b−i′ )
∗
(b+
i , b−i)
(b−
l , b−l)
. . .
2
bi′ ↑
3
bi ↑
or
bl ↓
ρ
ρ′
4
∗
(b; ρ)
(b; ρ′)
11
Time rounds
t′
. . .
. . .
i′
i
j
8
8
i = b+ρ′
b+ρ
i
∗ (b+ρ
i
; ρ)
∗ (b+ρ′
i
; ρ′)
7
in ρ, bi ↑
6
∗
(b; ρ)
in ρ′, bi ↑
∗
(b; ρ′)
l = j ′
5
Figure 1: This figure explains all the steps in the proof of Lemma 3.9. The rows correspond to agents (whose identity
is shown on the right side), and columns correspond to time rounds. The asterisks show the impressions. The arrows
show how the impressions get transferred, and labels on the arrows show what causes the transfer. In labels, “in ρ,
bi ↑” denotes that a particular transfer of impression is caused in click realization ρ when bid bi in increased.
Proof Sketch. We sketch the proof of Lemma 3.9 at a very high level. The “only if” direction was observed
in Observation 3.5; we focus on the “if” direction. Let A be a weakly-separated mechanism. We prove by
a contradiction that it is exploration-separated. If not, then there is a click realization ρ and a round t such
that t is influencial w.r.t. ρ as well as not bid-dependent w.r.t. ρ. Let round t be influencial with bid vector b,
influencing agent l, and influenced agents j and j′ 6= j in influenced round t′ (see 1 in Figure 1; all boxed
numbers in this sketch will refer to this figure).
From the assumption, t is not bid-dependent w.r.t. ρ, which means that there exists a bid profile b′ such
6= l is selected in round t with bids b′. Using scalefreeness, IIA, and pointwise-monotonicity, we
that i′
can prove that there exists a sufficiently large bid b+
i′ of agent i′ such that she gets an impression in round t
with bids (b+
i′ , b−i′) (see 2 ). Using the properties of the mechanism, it can further be proved that there is an
agent i such that she gets the impression in round t when either i increases her bid, or l decreases her bid
(see 3 ). When i increases her bid to b+
i , she also gets an impression in round t′, since impressions cannot
differ in round t′ in the case when l is not selected in round t and they must get transferred from j and j′ to
somebody in round t′, and IIA implies that this somebody should be i.
Recall that two different agents j and j′ get the impression in round t′ under ρ and ρ′ respectively (see
4 ). We prove that either agent j′ or agent j must be equal to l (this is done by looking at how the allocation
in round t′ changes when l decreases her bid). Let us break the symmetry and assume j′ = l (see box 5 ).
It is also easy to see that when i increases her bid, impression in round t′ get transferred to her in ρ (at
some minimum value b+ρ
, see 6 ), and impression in round t′ gets transferred to her also in ρ′ (as some
possibly different minimum value b+ρ′
, see 7 ). Using the assumptions of weakly-separatedness, we prove
that b+ρ
}, and then using
weakly-separatedness of A. Since these two bids were at a “threshold value” (these were the minimum
(see 8 ). This can be proved by observing that b+
i ≥ max{b+ρ
i = b+ρ′
i
i
i
, b+ρ′
i
i
17
values of bids to have transferred the impression in ρ and ρ′ from j and l respectively), we are able to prove
that the ratio of bj/bl must be some fixed number dependent on ρ, ρ′, and t′. In particular, it follows that bl
belongs to a finite set S(b−l) which depends only on b−l. However, by non-degeneracy of A there must be
infinitely many such bl’s, which leads to a contradiction.
4 Lower bounds on regret
In this section we use structural results from the previous section to derive lower bounds on regret.
Theorem 4.1. Consider the stochastic MAB mechanism design problem with k agents. Let A be an
exploration-separated deterministic allocation rule. Then its regret is R(T ; vmax) = Ω(vmax k1/3 T 2/3).
2 , . . . , 1
Let ~µ0 = ( 1
2 ) ∈ [0, 1]k be the vector of CTRs in which for each agent the CTR is 1
2. For each
agent i, let ~µi = (µi1, . . . , µik) ∈ [0, 1]k be the vector of CTRs in which agent i has CTR µii = 1
2 + ǫ,
ǫ = k1/3 T −1/3, and every other agent j 6= i has CTR µij = 1
2 . As a notational convention, denote by Pi[·]
and Ei[·] respectively the probability and expectation induced by the algorithm when clicks are given by ~µi.
Let Ii be the problem instance in which CTRs are given by ~µi and all bids are vmax. For each agent i, let Ji
be the problem instance in which CTRs are given by ~µ0, the bid of agent i is vmax, and the bids of all other
agents are vmax/2. We will show that for any exploration-separated deterministic allocation rule A, one of
these 2k instances causes high regret.
Let Ni be the number of bid-independent rounds in which agent i is selected. Note that Ni does not
depend on the bids. It is a random variable in the probability space induced by the clicks; its distribution
is completely specified by the CTRs. We show that (in a certain sense) the allocation cannot distinguish
between ~µ0 and ~µi if Ni is too small. Specifically, let At be the allocation in round t. Once the bids
are fixed, this is a random variable in the probability space induced by the clicks. For a given set S of
agents, we consider the event {At ∈ S} for some fixed round t, and upper-bound the difference between the
probability of this event under ~µ0 and ~µi in terms of Ei[Ni], in the following crucial claim, which is proved
in Section 4.1 via relative entropy techniques.
Claim 4.2. For any fixed vector of bids, each round t, each agent i and each set of agents S, we have
P0[At ∈ S] − Pi[At ∈ S] ≤ O(ǫ2 E0[Ni]).
(4.1)
2. Then Pi[At 6= i] ≥ 1
Proof of Theorem 4.1: Fix a positive constant β to be specified later. Consider the case k = 2 first. If
E0[Ni] > β T 2/3 for some agent i, then on the problem instance Ji, regret is Ω(T 2/3). So without loss of
generality let us assume E0[Ni] ≤ β T 2/3 for each agent i. Then, plugging in the values for ǫ and E0[Ni],
the right-hand side of (4.1) is at most O(β). Take β so that the right-hand side of (4.1) is at most 1
4. For
each round t there is an agent i such that P0[At 6= i] ≥ 1
4 by Claim 4.2, and therefore
in this round algorithm A incurs regret Ω(ǫ vmax) under problem instance Ii. By Pigeonhole Principle there
exists an i such that this happens for at least half of the rounds t, which gives the desired lower-bound.
Case k ≥ 3 requires a different (and somewhat more complicated) argument. Let R = β k1/3 T 2/3 and
N be the number of bid-independent rounds. Assume E0[N ] > R. Then E0[Ni] ≤ 1
E0[N ] for some agent
i. For the problem instance Ji there are, in expectation, E[N − Ni] = Ω(R) bid-independent rounds in
which agent i is not selected; each of which contributes Ω(vmax) to regret, so the total regret is Ω(vmax R).
2 agents
5 for more than
2 good agents. Suppose not. If agent i is not good
i such that E0[Ni] ≤ 2R/k. Furthermore, let us say that an agent i is good if P0[At = i] ≤ 4
T /6 different rounds t. We claim that there are more than k
then P0[At = i] > 4
From now on assume that E0[N ] ≤ R. Note that by Pigeonhole Principle, there are more than k
6 T different rounds t, so if there are at least k/2 such agents then
5 for at least 5
k
T =PT
t=1Pk
i=1
2 × ( 5
6 T ) × 4
5 ≥ kT /3 ≥ T,
P0[At = i] > k
18
contradiction. Claim proved. It follows that there exists a good agent i such that E0[Ni] ≤ 2R/k. Therefore
the right-hand side of (4.1) is at most O(β). Pick β so that the right-hand side of (4.1) is at most 1
10. Then
by Claim 4.2 for at least T /6 different rounds t we have Pi[At = i] ≤ 9
10 . In each such round, if agent i is
not selected then algorithm A incurs regret Ω(ǫ vmax) on problem instance Ii. Therefore, the (total) regret
of A on problem instance Ii is Ω(ǫ vmax T ) = Ω(vmax k1/3 T 2/3).
Theorem 4.3. In the setting of Theorem 4.1, fix k and vmax and assume that R(T ; vmax) = O(vmax T γ)
for some γ < 1. Then for every fixed δ ≤ 1
Proof. Fix λ ∈ (0, 2(1 − γ)). Redefine ~µi’s with respect to a different ǫ, namely ǫ = T −λ/2. Define the
problem instances Ii in the same way as before: all bids are vmax, the CTRs are given by ~µi.
Let us focus on agents 1 and 2. We claim that E1[N1] + E2[N2] ≥ β T λ, where β > 0 is a constant to
be defined later. Suppose not. Fix all bids to be vmax. For each round t, consider event St = {At = 1}.
Then by Claim 4.2 we have
4 and λ < 2(1 − γ) we have Rδ(T ; vmax) = Ω(δ vmax T λ).
(cid:12)(cid:12)P1[St] − P2[St](cid:12)(cid:12) ≤(cid:12)(cid:12)P0[St] − P1[St](cid:12)(cid:12) +(cid:12)(cid:12)P0[St] − P2[St](cid:12)(cid:12) ≤ O(cid:0)ǫ2(cid:1) (E1[N1] + E2[N2]) ≤ 1
for a sufficiently small β. Now, P1[St] ≥ 1
2 for at least T /2 rounds t. This is because otherwise on problem
instance Ii regret would be R(T ) ≥ Ω(ǫ T vmax) = Ω(vmax T 1−λ/2), which contradicts the assumption
R(T ) = O(vmax T γ). Therefore P2[St] ≥ 1
4 for at least T /2 rounds t, hence on problem instance I2 regret
is at least Ω(ǫ T vmax), contradiction. Claim proved.
2 T λ. Consider the problem instance in
which CTRs given by ~µ1, bid of agent 2 is vmax, and all other bids are vmax(1 − 2δ)/(1 + 2ǫ). It is easy to
see that this problem instance has δ-gap. Each time agent 1 is selected, algorithm incurs regret Ω(δvmax).
Thus the total regret is at least Ω(δN1 vmax) = Ω(δ vmax T λ).
Now without loss of generality let us assume that E1[N1] ≥ β
4
4.1 Relative entropy technique: proof of Claim 4.2
We extend the relative entropy technique from [7]. All relevant facts about relative entropy are summarized
in the theorem below. We will need the following definition: given a random variable X on a probability
space (Ω,F, P), let PX be the distribution of X, i.e. a measure on R defined by PX(x) = P[X = x].
Theorem 4.4 (Some standard facts about relative entropy, e.g. [15, 27, 29]).
Let p and q be two probability measures on a finite set U, and let Y and Z be functions on U. There exists
a function F (p; qY ) : U → R with the following properties:
(i) Ep F (p; qY ) = Ep F (p; q(Y, Z)) + Ep F (pZ; qZY )
(chain rule),
2D(pkq) for any event U ′ ⊂ U, where D(pkq) = Ep F (p; q1)
(ii) (cid:12)(cid:12)p(U ′) − q(U ′)(cid:12)(cid:12) ≤q 1
(iii) for each x ∈ U, if conditional on the event {Z = Z(x)} p coincides with q, then F (p; qZ)(x) = 0.
(iv) for each x ∈ U, if conditional on the event {Z = Z(x)} p and q are fair and ( 1
2 + ǫ)-biased coins,
respectively, then it is the case that F (p; qZ)(x) ≤ 4ǫ2.
Remark. This theorem summarizes several well-known facts about relative entropy, albeit in a somewhat
non-standard notation. For the proofs, see [15, 27, 29]. In the proofs, one defines F = F (p; qY ) as a
function F : U → R which is specified by F (x) = Px′∈U p(x′Ux) lg p(x′Ux)
q(x′Ux) , where Ux is the event
{Y = Y (x)}.16 Note that the quantity Ep F (p; q1) is precisely the relative entropy (a.k.a. KL-divergence),
commonly denoted D(pkq), and Ep F (p; qY ) is the corresponding conditional relative entropy.
16We use the convention that p(x) log(p(x)/q(x)) is 0 when p(x) = 0, and +∞ when p(x) > 0 and q(x) = 0.
19
be considered a random variable under p or q.
pletely determines whether round t is bid-independent, and if so, which arm is chosen in this round.
In what follows we use Theorem 4.4 to prove Claim 4.2. For simplicity we will prove (4.1) for i = 1.
The history up to round t is Ht = (h1, h2 , . . . , ht) where hs ∈ {0, 1} is the click or no click event re-
ceived by the algorithm at round s. Let Ct be the indicator function of the event “round t is bid-independent”.
Define the bid-independent history as bHt = (bh1,bh2 , . . . ,bht), where bht = htCt. For any exploration-
separated deterministic allocation rule and each round t, the bid-independent history bHt−1 and the bids
completely determine which arm is chosen in this round. Moreover, bHt−1 alone (without the bids) com-
Recall the CTR vectors ~µi as defined in Section 4. Let p and q be the distributions induced on bHT by
~µ0 and ~µ1, respectively. Let pt and qt be the distributions induced onbht by ~µ0 and ~µ1, respectively. Let
Ht the support of bHt, i.e. the set of all t-bit vectors. In the forthcoming applications of Theorem 4.4, the
universe will be U = HT . By abuse of notation, we will treat bHt as a projection HT → Ht, so that it can
Claim 4.5. D(pkq) = Ep F (p; q bHt) + Pt
Proof. Use induction on t ≥ 0 (set bH0 = 1). In order to obtain the claim for a given t assuming that it holds
for t − 1, apply Theorem 4.4(i) with Y = bHt−1 and Z =bht.
Claim 4.6. F (pt; qt bHt−1) ≤ 4ǫ2 Ct 1{At=1} for each round t.
Proof. We are interested in the function F = F (pt; qt bHt−1) : HT → R. Given bHt−1, one of the following
• round t is not bid-independent. Thenbht = 0, hence F (·) = 0 by Theorem 4.4(iii),
• round t is bid-independent and arm 1 is not selected. Thenbht is distributed as a fair coin under both
p and q, so again F (·) = 0.
• round t is bid-independent and arm 1 is selected. Then F (·) ≤ 4ǫ2 by Theorem 4.4(iv).
Given the full bid-independent history bHT , p and q become (the same) point measure, so by Theo-
rem 4.4(iii) Ep F (p; q bHT ) = 0. Therefore taking Claim 4.5 with t = T we obtain
s=1 Ep F (ps; qs bHs−1) for any t > 1.
three cases occurs:
D(pkq) =
TXt=1
Ep F (pt; qt bHt−1) = 4ǫ2
TXt=1
Ep [Ct 1{At=1}] = 4ǫ2 Ep[N1].
(4.2)
For a given round t and fixed bids, the allocation at round t is completely determined by the bid-independent
history bHt−1. Thus, we can treat {At ∈ S} as an event in HT . Now (4.1) follows from (4.2) via an
application of Theorem 4.4(ii) with U ′ = {At ∈ S}.
4.2 Lower bound for non-scalefree allocations
In this subsection we derive a regret lower bound for deterministic truthful mechanisms without assuming
that the allocations are scale-free. In particular, for two agents there are no assumptions. This lower bound
holds for any k (the number of agents) assuming that the allocation satisfies IIA, but unlike the one in
Theorem 4.1 it does not depend on k.
Theorem 4.7. Consider the stochastic MAB mechanism design problem with k agents. Let (A,P) be a
normalized truthful mechanism such that A is a non-degenerate deterministic allocation rule. Suppose A
satisfies IIA. Then its regret is R(T ; vmax) = Ω(vmax T 2/3) for any sufficiently large vmax.
Let us sketch the proof. Fix an allocation A. In Definition 3.3, if round t is (b, ρ) influential, for some
click realization ρ and bid vector b, an agent i is called strongly influenced by round t if it is one of the
20
two agents that are “influenced” by round t but is not the “influencing agent” of round t. In particular, it
holds that A(b, ρ, t) 6= i. For each click realization ρ, round t and agent i, if there exists a bid vector b
such that round t is (b, ρ)-influential with strongly influenced agent i, then fix any one such b, and define
b∗
i = b∗
i (ρ, t), where the maximum is taken over all
click realizations ρ, all rounds t, and all agents i. Let us say that round t is B∗-free from agent i w.r.t click
realization ρ, if for this click realization the following property holds: agent i is not selected in round t as
long as each bid is at least B∗.
i (ρ, t) := maxj6=i bj. Let us define B∗
A = maxρ,t,i b∗
Lemma 4.8. In the setting of Theorem 4.7, for any click realization ρ, any influential round t is B∗
from some agent w.r.t. ρ.
A-free
i (ρ, t) ≤ B∗
Suppose round t is not B∗
Proof. Fix click realization ρ. Since round t is influential, for some bid profile b and agent i it is (b, ρ)-
influential with a strongly influenced agent i. By definition of b∗
i (ρ, t), without loss of generality each bid in
b (other than i’s bid) is at most b∗
A. Then A(b, ρ, t) 6= i, and round t is (b, ρ)-secured from agent
i.
A-free from agent i w.r.t ρ. Then there exists a bid profile b′ in which each bid
(other than i’s bid) is at least B∗
A such that A(b′, ρ, t) = i. To derive a contradiction, let us transform b to
b′ by adjusting first the bid of agent i and then bids of agents j 6= i one agent at a time. Initially agent i is
not chosen in round t, and after the last step of this transformation agent i is chosen. Thus it is chosen at
some step, say when we adjust the bid of agent i or some agent j 6= i. This transfer of impression to agent
i cannot happen when bid of agent i is adjusted from bi to b′
i (since round t is (b; ρ)-secured from i), and
it cannot happen when bid of agent j 6= i is adjusted from bj to b′
j ≥ bj (this is because, the transfer to i
cannot happen from j because of pointwise-monotonicity and the transfer to i cannot happen from l 6= j
because of IIA). This is a contradiction.
let ~µ0 be the vector of CTRs in which all CTRs are 1
A. Let N (ρ) be the number of influential rounds w.r.t
click realization ρ. Let Ni(ρ) be the number of influential rounds w.r.t. click realization ρ that are B∗
A-free
from agent i w.r.t. ρ. Then N and the Ni’s are random variables in the probability space induced by the
Let T be the time horizon. Assume vmax ≥ 2B∗
clicks. By Lemma 4.8 we have thatPi Ni(ρ) is at least the number of influential rounds. As in Section 4,
Fix a constant β > 0 to be specified later. If E0[N ] ≥ βk T 2/3 then E0[Ni] ≥ β T 2/3 for some agent
i, so the allocation incurs expected regret R(T ; vmax) ≥ Ω(vmax T 2/3) on any problem instance Jj, j 6= i.
(In this problem instance, CTRs given by ~µ0, the bid of agent j is vmax, and all other bids are vmax/2.) Now
suppose E0[N ] ≤ βk T 2/3. Then the desired regret bound follows by an argument very similar to the one in
the last paragraph of the proof of Theorem 4.1.
2 , and let E0[·] denote expectation w.r.t. ~µ0.
4.3 Universally truthful randomized MAB mechanisms
Consider randomized mechanisms that are universally truthful, i.e. truthful for each realization of the inter-
nal random seed. Our goal here is to extend the Ω(vmax T 2/3) regret bounds for deterministic mechanisms
to universally truthful randomized mechanisms, under relatively mild assumptions.
Note that lower bounds on regret for universally truthful MAB mechanisms do not immediately follow
from those for deterministic truthful MAB mechanisms. To see this, consider a randomized MAB mecha-
nism A that randomizes over some deterministic truthful mechanisms, each with regret at least R. Then for
each deterministic mechanism A′ in the support of A there is a problem instance on which A′ has regret at
least R; it could be a different problem instance for different A′. Whereas to lower-bound the regret of A
we need to provide one problem instance with high regret in expectation over all A′.
We consider mechanisms that randomize over exploration-separated deterministic allocation rules. As
per the discussion above, it does not suffice to quote Theorem 4.1; instead, we need to extend its proof.
21
Lemma 4.9. Consider the MAB mechanism design problem. Let D be a distribution over exploration-
separated deterministic allocation rules. Then
EA∈D [RA(T ; vmax)] = Ω(vmax k1/3 T 2/3).
Proof. Recall that in the proof of Theorem 4.1 we define a family F of 2k problem instances, and show
that if A is an exploration-separated deterministic allocation rule, then on one of these instances its regret
In fact, we can extend this analysis to show that the regret is “high”, that is at least R∗ =
is “high”.
Ω(vmax k1/3 T 2/3), on an instance I ∈ F chosen uniformly at random from F; here regret is in expectation
over the choice of I. 17 Once this is proved, it follows that regret is R∗/2 for any distribution over such
A, in expectation over both the choice of A and the choice of I. Thus there exists a single (deterministic)
instance I such that EA∈D [RA,I (T )] ≥ R∗/2.
Theorem 4.3 can be extended similarly.
5 A matching upper bound
Let us describe a very simple mechanism, called the naive MAB mechanism, which matches the lower bound
from Theorem 4.1 up to polylogarithmic factors (and also the lower bound from Theorem 4.3, for γ = λ = 2
3
and constant δ).
Fix the number of agents k, the time horizon T , and the bid vector b. The mechanism has two phases.
In the exploration phase, each agent is selected for T0 := k−2/3 T 2/3(log T )1/3 rounds, in a round robin
fashion. Let ci be the number of clicks on agent i in the exploration phase. In the exploitation phase, an agent
i∗ ∈ argmaxi cibi is chosen and selected in all remaining rounds. Payments are defined as follows: agent i∗
pays maxi∈[k]\{i∗} cibi/ci∗ for every click she gets in exploitation phase, and all others pay 0. (Exploration
rounds are free for every agent.) This completes the description of the mechanism.
Lemma 5.1. Consider the stochastic MAB mechanism design problem with k agents. The naive mechanism
is normalized, truthful and has worst-case regret R(T ; vmax) = O(vmax k1/3 T 2/3 log2/3 T ).
Proof. The mechanism is truthful by a simple second-price argument.18 Recall that ci is the number of
clicks i got in the exploration phase. Let pi = maxj6=i cjbj/ci be the price paid (per click) by agent i if she
wins (all) rounds in exploitation phase. If vi ≥ pi, then by bidding anything greater than pi agent i gains
vi − pi utility each click irrespective of her bid, and bidding less than vi, she gains 0, so bidding vi is weakly
dominant. Similarly, if vi < pi, then by bidding anything less than pi she gains 0, while bidding bi > pi,
she loses bi − pi each click. So bidding vi is weakly dominant in this case too.
By Chernoff bounds, for each agent i we have Pr [¯µi − µi > r] ≤ T −4, for r = p8 log(T )/T0. If in
For the regret bound, let (µ1 , . . . , µk) be the vector of CTRs, and let ¯µi = ci/T0 be the sample CTRs.
a given run of the mechanism all estimates ¯µi lie in the intervals specified above, call the run clean. The
expected regret from the runs that are not clean is at most O(vmax), and can thus be ignored. From now on
let us assume that the run is clean.
The regret in the exploration phase is at most k T0 vmax = O(vmax k1/3 T 2/3 log1/3 T ). For the ex-
ploitation phase, let j = argmaxi µibi. Then (since we assume that the run is clean) we have
(µi∗ + r) bi∗ ≥ ¯µi∗ bi∗ ≥ ¯µj bj ≥ (µj − r) bj,
17This extension requires but minor modifications to the proof of Theorem 4.1. For instance, for the case k ≥ 3 we argue that
2 agents i (and so on), and if E0[N ] ≤ R then (omitting some details)
18Alternatively, one can use Theorem 1.5 since all exploration rounds are bid-independent, and only exploration rounds are
first, if E0[N ] > R then E0[Ni] ≤ 2
there are Ω(k) good agents i such that E0[Ni] ≤ 2R/k (and so on).
k E0[N ] for at least k
influential, and the payments are exactly as defined in Theorem 3.1.
22
which implies µjvj − µi∗vi∗ ≤ r(vj + vi∗) ≤ 2r vmax. Therefore, the regret in exploitation phase is at most
2r vmax T = O(vmax k1/3 T 2/3 log2/3 T ). Therefore the total regret is as claimed.
6 Randomized allocations and adversarially chosen clicks
In this section we discuss randomized allocations. We apply them to a version of the MAB mechanism
design problem in which clicks are generated adversarially.19 The objective is to optimize the worst-case
regret over all values v = (v1 , . . . , vk) such that vi ∈ [0, vmax] for each i, and all click realizations ρ:
R(T ; v; ρ) =hmaxi viPT
R(T ; vmax) = max{R(T ; v; ρ) : all click realizations ρ, all v such that vi ∈ [0, vmax] for each i}.
t=1ρi(t)i −PT
t=1Pk
i=1 vi ρi(t) E [Ai(v; ρ; t)]
(6.1)
The first term in (6.1) is the social welfare from the best time-invariant allocation, the second term is the
social welfare generated by A.
Let us make a few definitions related to truthfulness. Recall that a mechanism is called weakly truthful
if for each click realization, it is truthful in expectation over its random seed. A randomized allocation is
pointwise monotone if for each click realization and each bid profile, increasing the bid of any one agent
does not decrease the probability of this agent being allocated in any given round. For a set S of rounds
and a function σ : S → {agents}, an allocation is (S, σ)-separated if (i) it coincides with σ on S, (ii)
the clicks from the rounds not in S are discarded (not reported to the algorithm). An allocation is strongly
separated if before round 1, without looking at the bids, it randomly chooses a set S of rounds and a function
σ : S → {agents}, and then runs a pointwise monotone (S, σ)-separated allocation. Note that the choice of
S and σ is independent of the clicks, by definition.
We obtain a structural result: for any (randomized) strongly separated allocation rule A there exists a
mechanism that is normalized and weakly truthful.
Lemma 6.1. Consider the MAB mechanism design problem. Let A be a (randomized) strongly separated
allocation rule. Then there exists a payment rule P such that the resulting mechanism (A,P) is normalized
and weakly truthful.
We consider PSIM [8, 28], a randomized MAB algorithm from the literature which we here interpret
as an MAB allocation rule. It follows from [8, 28], that PSIM has strong regret guarantees for the adver-
sarial MAB mechanism design problem: it obtains regret R(T, vmax) = O(vmax k1/3 (log k)1/3 T 2/3). In
Section 6.1 we state PSIM and show that it is strongly separated. Thus, we obtain the following result.
Theorem 6.2. There exists a weakly truthful normalized mechanism for the adversarial MAB problem
(against oblivious adversary) whose regret grows as O((k log k)1/3 · T 2/3 · vmax).
Remark. For the adversarial MAB problem (i.e., without the restriction of truthfulness), the regret bound can
be improved to O(√kT · vmax) [7, 5]. However, the algorithms that achieve this bound do not immediately
yield MAB allocation rules that are strongly separated. It is an open question whether the regret bound in
Corollary 6.2 can be improved.
Proof of Lemma 6.1: Throughout the proof, let us fix a click realization ρ, time horizon T , bid vector b,
and agent i. We will consider the payment of agent i. We will vary the bid of agent i on the interval [0, bi];
the bids b−i of all other agents always stay the same.
19We focus on the oblivious adversary which (unlike the more difficult “adaptive adversary”) specifies all clicks in advance.
23
Let ci(x) be the number of clicks received by agent i given that her bid is x. Then by (the appropriate
version of) Theorem 3.1 the payment of agent i must be Pi(b) such that
EA[Pi(b)] = EAhbi ci(bi) −R bi
x=0 ci(x) dxi ,
(6.2)
where the expectation is taken over the internal randomness in the algorithm.
Recall that initially A randomly selects, without looking at the bids, a set S of rounds and a function
σ : S → {agents}, and then runs some pointwise monotone (S, σ)-separated allocation A(S,σ). In what
follows, let us fix S and σ, and denote A∗ = A(S,σ). We will refer to the rounds in S as exploration
rounds, and to the rounds not in S as exploitation rounds. Let γ∗
i (x, t) be the probability that algorithm A∗
allocates agent i in round t given that agent i bids x. Note that for fixed value of internal random seed of
A∗ this probability can only depend on the clicks observed in exploration rounds, which are known to the
mechanism. Therefore, abstracting away the computational issues, we can assume that it is known to the
mechanism. Define the payment rule as follows: in each exploitation round t in which agent i is chosen and
clicked, charge
P ∗
i (b, t) = bi −
Then the total payment assigned to agent i is
1
i (bi, t)Z bi
0
γ∗
γ∗
i (x, t) dx.
(6.3)
i (b) =Pt6∈S ρi(t) A∗
P ∗
i (b; ρ; t) P ∗
Since allocation A∗ is pointwise monotone, the probability γ∗
i (x, t) is non-decreasing in x. Therefore
P ∗
i (b, t) ∈ [0, bi] for each round t. It follows that the mechanism is normalized (for any realization of the
random seed of allocation A).
It remains to check that the payment rule (6.3) results in (6.2). Let c∗
i (x) be the number of clicks
allocated to agent i by allocation A∗ given that her bid is x. Let c
(x) be the corresponding number of
clicks in exploitation rounds only. Since A∗ is (S, σ)-separated, we have
i (b, t).
(6.4)
expl
i
E[c∗
i (x) − c
expl
i
(x)] =Pt∈S ρσ(t)(t) = const(x).
Taking expectations in (6.4) over the random seed of AS and using (6.5), we obtain
(6.5)
E[P ∗
i (b, t)
i (b)] =Pt6∈S ρi(t) γ∗
=Pt6∈S ρi(t) hbi γ∗
= bihPt6∈S ρi(t) γ∗
(bi)] −R bi
= Ehbi c∗
i (bi) −R bi
i (bi, t) P ∗
i (bi, t) −R bi
i (bi, t)i −R bi
i (x) dxi .
= bi E [cexpl
0
0 c∗
E[cexpl
i
i
0 γ∗
i (x, t) dxi
0 hPt6∈S ρi(t) γ∗
(x)] dx
i (x, t)i dx
Finally, taking expectations over the choice of S and σ, we obtain (6.2).
6.1 Algorithm PSIM is strongly separated
In this subsection we interpret PSIM [8, 28] as an MAB allocation rule and show that it is strongly separated
(which implies Theorem 6.2). For the sake of completeness, we present PSIM below. As usual, k denotes
the number of agents; let [k] denote the set of agents.
24
Input: Time horizon T , bid vector b. Let vmax = maxi bi.
Output: For each round t ≤ T , a distribution on [k].
1. Divide the time horizon into P phases of T /P consecutive rounds each.
2. From rounds of each phase p, pick without replacement k rounds at random (called the exploration
rounds) and assign them randomly to k arms. Let S denote the set of all exploration rounds (of all
phases). Let f : S → [k] be the function which tells which arm is assigned to an exploration round in
S. The rounds in [T ] \ S are called the exploitation rounds.
3. Let wi(0) = 1 for all i ∈ [k].
4. For each phase p = 1, 2, . . . , P
(a) For each round t in phase p
i. If t ∈ S and f (t) = i, then define the distribution γ(b; t; S, f ) such that γi(b; t; S, f ) = 1.
Pick an agent according to this distribution (equivalently, pick agent i), observe the click
ρi(t), and update wi(p) multiplicatively,
wi(p) = wi(p − 1) · (1 + ǫ)ρi(t)bi/vmax.
ii. If t 6∈ S, then define the distribution γ(b; t; S, f ) such that γi(b; t; S, f ) = wi(p−1)
an agent according to γ(b; t; S, f ), observe the feedback, and discard the feedback.
Pj wj(p−1) . Pick
If we pick the values ǫ = (k log k/T )1/3 and P = (log k)1/3(T /k)2/3, then the regret of PSIM is
Regret.
bounded by O((k log k)1/3T 2/3vmax) against any oblivious adversary (see [8, 28]).
Claim 6.3. PSIM is strongly-separated.
Proof. It is clear from the structure of PSIM above that it chooses a set S of exploration rounds and a
function f : S → [k] in the beginning without looking at the bids and then runs an (S, f )-separated
allocation. We need to prove that the (S, f )-separated allocation is pointwise monotone. For this we need
prove that the probability γi(b; t; S, f ) is monotone in the bid of agent i, where γi(b; t; S, f ) denotes the
probability of picking agent i in round t when bids are b given the choice of S and f .
If t ∈ S, the
γi(b; t; S, f ) is independent of bids, and hence is monotone in bi. Let t 6∈ S and t is a round in phase p. Let
us denote by f −1(i, p) the (unique) exploration round in phase p assigned to agent i. We then have
γi(b; t; S, f ) = (1 + ǫ)
bi
vmax
Pp−1
q=1 ρi(f −1(i,q)) (cid:30)Xj
(1 + ǫ)
bj
vmax
Pp−1
q=1 ρj (f −1(j,q)).
We split the denominator into the term for agent i and all other terms. It is then not hard to see that this is a
non-decreasing function of bi.
7 Truthfulness in expectation over CTRs
We consider the stochastic MAB mechanism design problem under a more relaxed notion of truthfulness:
truthfulness in expectation, where for each vector of CTRs the expectation is taken over clicks (and the
internal randomness in the mechanism, if the latter is not deterministic).20 We show that any MAB allocation
A∗ that is monotone in expectation, can be converted to an MAB mechanism that is truthful in expectation
and normalized in expectation, with minor changes and a very minor increase in regret. As discussed in the
Introduction, this result rules out a natural lower-bounding approach.
20 Normalized-in-expectation and monotone-in-expectation properties are defined similarly. An allocation rule is monotone in
expectation if for each agent i and fixed bid profile b−i, the corresponding expected click-allocation is a non-decreasing function
of bi. A mechanism is normalized in expectation if in expectation each agent is charged an amount between 0 and her bid for each
click she receives. In both cases, the expectation is taken over the clicks and possibly the allocation’s random seed.
25
Remark. The follow-up work [9] has established that there exist MAB allocations that are monotone in
expectation whose regret matches the optimal upper bounds for MAB algorithms. In fact, [9] defined a
rather natural class of “well-formed MAB algorithms” that, e.g., includes (a version of) algorithm UCB1 [6],
and proved that any algorithm in this class gives rise to a monotone-in-expectation MAB allocation.
We will show that for any allocation A∗ that is monotone in expectation, any time horizon T , and any
parameter γ ∈ (0, 1) there exists a mechanism (A,P) such that the mechanism is truthful in expectation and
normalized in expectation, and allocation A initially makes a random choice between A∗ and some other
allocation, choosing A∗ with probability at least γ. We call such allocation A a γ-approximation of A∗.
Clearly, on any problem instance we have RA(T ) ≤ γ RA∗(T ) + (1 − γ)T . The extra additive factor of
(1 − γ)T is not significant if e.g. γ = 1 − 1
T . The problem with this mechanism is that it is not ex-post
normalized; moreover, in some click realizations payments may be very large in absolute value.
Theorem 7.1. Consider the stochastic MAB mechanism design problem with k agents and a fixed time
horizon T . For each γ ∈ (0, 1) and each allocation rule A∗ that is monotone in expectation, there exists
a mechanism (A,P) such that A is a γ-approximation of A∗, and the mechanism is truthful in expectation
and normalized in expectation.
Remark. The key idea is to view the Myerson payments (see Theorem 3.1) as multivariate polynomials
over the CTRs, and argue that any such polynomial can be “implemented” by a suitable payment rule. The
payment rule P will be well-defined as a mapping from histories to numbers; we do not make any claims
on the efficient computability thereof.
Proof. Let Aexpl be the allocation rule where in each round an agent is chosen independently and uniformly
at random. Allocation A is defined as follows: use A∗ with probability γ; otherwise use Aexpl. Fix an
instance (b, µ) of the stochastic MAB mechanism design problem, where b = (b1 , . . .
, bk) and µ =
(µ1 , . . . , µk) are vectors of bids and CTRs, respectively. Let Ci = Ci(bi; b−i) be the expected number of
clicks for agent i under the original allocation A∗. Then by Myerson [38] the expected payment of agent i
must be
i = γhbi Ci(bi; b−i) −R bi
PM
0 Ci(x; b−i) dxi .
i
We treat the expected payment as a multivariate polynomial over µ1 , . . . , µk.
Claim 7.2. PM
Proof. Fix the bid profile. Let Xt be allocation of algorithm A∗. Let poly(T ) be the set of all polynomials
over µ1 , . . . , µk of degree at most T . Consider a fixed history h = (x1, y1; . . . ; xT , yT ), and let ht be the
corresponding history up to (and including) round t. Then
is a polynomial of degree ≤ T in variables µ1 , . . . , µk.
(7.1)
(7.2)
(7.3)
P[h] =QT
Ci(bi; b−i) =Ph∈H
t=1 Pr[Xt = xt ht−1] µyt
xt (1 − µxt)1−yt ∈ poly(T )
P[h] #clicksi(h) ∈ poly(T ).
Therefore PM
monomial of Ci(x; b−i).
i ∈ poly(T ), since one can take an integral in (7.1) separately over the coefficient of each
Fix time horizon T . For a given run of an allocation rule, the history is defined as h = (x1, y1; . . . ; xT , yT ),
where xt is the allocation in round t, and yt ∈ {0, 1} is the corresponding click. Let H be the set of all
possible histories.
26
i
Our payment rule P is a deterministic function of history. For each agent i, we define the payment Pi =
Pi(h) for each history h such that Eh[Pi(h)] = PM
for any choice of CTRs, and hence Eh[Pi(h)] ≡ PM
i ,
where ≡ denotes an equality between polynomials over µ1 , . . . , µk.
Fix the bid vector and fix agent i. We define the payment Pi as follows. Charge nothing if allocation
A∗ is used. If allocation Aexpl is used, charge per monomial. Specifically, let mono(T ) be the set of all
monomials over µ1 , . . . , µk of degree at most T . For each monomial Q ∈ mono(T ) we define a subset of
relevant histories Hi(Q) ⊂ H. (We defer the definition till later in the proof.) For a given history h ∈ H
we charge a (possibly negative) amount
Pi(h) = 1
i (Q),
(7.4)
where deg(Q) is the degree of Q, and PM
histories induced by Aexpl. Then the expected payment is
1−γ PQ∈mono(T ): h∈Hi(Q) kdeg(Q) PM
i (Q) is the coefficient of Q in PM
Eh[Pi(h)] =PQ∈mono(T ) kdeg(Q) Pexpl[Hi(Q)] PM
Therefore in order to guarantee that Eh[Pi(h)] ≡ PM
i
it suffices to choose Hi(Q) for each Q so that
(7.5)
Consider a monomial Q = µα1
k . Let Hi(Q) consist of all histories such that first agent 1 is selected
1
α1 times in a row, and clicked every time, then agent 2 is selected α2 times in a row, and clicked every time,
and so on till agent k. In the remaining T − deg(Q) rounds, any agent can be chosen, and any outcome
(click or no click) can be received. It is clear that (7.5) holds.
kdeg(Q) Pexpl[Hi(Q)] ≡ Q.
. . . µαk
i . Let Pexpl be the distribution on
i (Q).
8 Open questions
Despite the exciting developments in the follow-up work [9, 56, 22, 48] (discussed in Section 1.3), MAB
mechanisms are not well-understood. Below is a snapshot of the open questions, current as of this writing.
Impossibility results for deterministic MAB mechanisms.
1. For deterministic MAB mechanisms with k > 2 agents, is it possible to obtain lower bounds on regret
for weakly separated MAB allocation rules, without assuming IIA?
2. We conjecture that the “informational obstacle” – insufficient observable information to compute
payments – can be meaningfully extended to a very general class of mechanisms in which an alloca-
tion rule interacts with the environment. As mentioned in Section 1.3, the follow-up work [56, 48]
suggested settings other than MAB mechanisms in which this obstacle arises. To conclude that the
“informational obstacle” is prominent in a given setting, one needs to prove that unrestricted payment
computation makes truthful mechanisms strictly more powerful.
3. Surprisingly, we still do not understand the limitations of deterministic truthful-in-expectation mecha-
nisms. While, according to [9], there exist regret-optimal MAB allocation rules that are deterministic
and monotone-in-expectation (e.g., the allocation rule based on UCB1), it is not clear whether any such
allocation rule can be extended to a deterministic truthful-in-expectation MAB mechanism.
4. It would be interesting to analyze a slightly more permissive model in which an MAB mechanism
can decide to “skip” a round without displaying an ad. In particular, in such model we could trivially
extend the lower bounds on regret from the special case of k = 2 agents to k > 2 agents. However,
our negative results for two agents do not immediately extend to this new model, and moreover the
structural results for k > 2 agents do not immediately follow either.
27
Randomized MAB mechanisms.
1. Recall that the “BKS reduction” from Babaioff, Kleinberg and Slivkins [9] exhibits a tradeoff between
variance in payments and loss in performance. Since the variance in payments can be very high,
optimizing this tradeoff is crucial.
This question is not resolved by the worst-case optimality result in Wilkens and Sivan [56]. While no
other reduction can achieve a better tradeoff for all monotone MAB allocation rules simultaneously,
the result in [56] does not rule out a reduction with better tradeoff for some monotone MAB allocation
rules, and therefore it does not rule out an MAB mechanism with better tradeoff. Furthermore, it is
possible that an MAB mechanism with optimal tradeoff cannot be represented as a reduction from a
regret-optimal allocation rule, in which case results about reductions simply do not apply.
2. Consider weakly truthful MAB mechanisms in the setting with adversarially chosen clicks.21 The
weakly truthful MAB mechanism in the present paper achieves regret O(k1/3 T 2/3), whereas the best
known MAB algorithms achieve regret O(√kT ) [7, 5]. It is not clear what should be the tight regret
bound. In particular, neither our reduction in Section 6 nor the BKS reduction from [9] immediately
apply to the algorithms in [7, 5].
3. More generally, as discussed in Section 1.2, pay-per-click ad auctions motivate many other versions
of the MAB mechanism design problem, corresponding to the various MAB settings studied in the
literature. For every such version one could compare the performance of weakly truthful MAB mech-
anisms with that of the best MAB algorithms. The positive direction here reduces (using the BKS
reduction) to designing weakly monotone MAB allocations. This type of question is a new angle in
the MAB literature, see [50] for a self-contained account.
Multi-slot MAB mechanisms:
pay-per-click auctions with multiple ad slots and unknown CTRs.
1. Intuitively it seems that the negative results from this paper should extend to the setting with two or
more ad slots. However, the precise characterization results and regret bounds remain elusive. Also,
such results would probably depend on the specific multi-slot model, i.e. on on how clicks in different
slots are correlated, and how CTRs of the same ad in different slots are related to one another.
2. Recall that Gatti, Lazaric and Trovo [22] provide truthful multi-slot MAB mechanisms based on the
simple MAB mechanism presented in this paper and (independently) in Devanur and Kakade [17].
It remains to be seen if one can obtain weakly truthful mechanisms with better regret, e.g. using a
more efficient multi-slot MAB algorithm with an extension of the BKS reduction. Note that even the
algorithmic (i.e., non-strategic) version of multi-slot MAB is not fully understood.
Acknowledgements
We thank Jason Hartline, Robert Kleinberg and Ilya Segal for helpful discussions.
References
[1] Gagan Aggarwal, Ashish Goel, and Rajeev Motwani. Truthful auctions for pricing search keywords. In ACM
Conf. on Electronic Commerce (EC), pages 1–7, 2006.
21Recall that an MAB mechanism is weakly truthful if for each click realization, it is truthful in expectation over its random seed.
Weakly monotone MAB allocation rules are defined similarly.
28
[2] Gagan Aggarwal and S. Muthukrishnan. Tutorial on theory of sponsored search auctions. In IEEE Symp. on
Foundations of Computer Science (FOCS), 2008.
[3] Aaron Archer and ´Eva Tardos. Truthful mechanisms for one-parameter agents. In IEEE Symp. on Foundations
of Computer Science (FOCS), pages 482–491, 2001.
[4] Susan Athey
and
Ilya
Segal.
An
efficient
dynamic mechanism.
Available
from
http://www.stanford.edu/isegal/agv.pdf, March 2007.
[5] J.Y. Audibert and S. Bubeck. Regret Bounds and Minimax Policies under Partial Monitoring. J. of Machine
Learning Research (JMLR), 11:2785–2836, 2010. A preliminary version has been published in COLT 2009.
[6] Peter Auer, Nicol`o Cesa-Bianchi, and Paul Fischer. Finite-time analysis of the multiarmed bandit problem.
Machine Learning, 47(2-3):235–256, 2002. Preliminary version in 15th ICML, 1998.
[7] Peter Auer, Nicol`o Cesa-Bianchi, Yoav Freund, and Robert E. Schapire. The nonstochastic multiarmed bandit
problem. SIAM J. Comput., 32(1):48–77, 2002. Preliminary version in 36th IEEE FOCS, 1995.
[8] Baruch Awerbuch and Robert Kleinberg. Online linear optimization and adaptive routing. J. of Computer and
System Sciences, 74(1):97–114, February 2008. Preliminary version in 36th ACM STOC, 2004.
[9] Moshe Babaioff, Robert Kleinberg, and Aleksandrs Slivkins. Truthful mechanisms with implicit payment com-
putation. In 11th ACM Conf. on Electronic Commerce (EC), pages 43–52, 2010. Best Paper Award.
[10] Maria-Florina Balcan, Avrim Blum, Jason D. Hartline, and Yishay Mansour. Reducing mechanism design to al-
gorithm design via machine learning. J. of Computer and System Sciences, 74(8):1245–1270, 2008. Preliminary
version in IEEE FOCS 2005.
[11] Michael Ben-Or and Avinatan Hassidim. The Bayesian Learner is Optimal for Noisy Binary Search (and Pretty
Good for Quantum as Well). In IEEE Symp. on Foundations of Computer Science (FOCS), 2008.
[12] Dirk Bergemann and Juuso Valimaki. Bandit Problems. In Steven Durlauf and Larry Blume, editors, The New
Palgrave Dictionary of Economics, 2nd ed. Macmillan Press, 2006.
[13] Dirk Bergemann
and
dynamic
cowles.econ.yale.edu/P/cd/d15b/d1584.pdf, October 2006.
Juuso Valimaki.
Efficient
auctions.
Available
from
[14] Nicol`o Cesa-Bianchi and G´abor Lugosi. Prediction, learning, and games. Cambridge Univ. Press, 2006.
[15] Thomas M. Cover and Joy A. Thomas. Elements of Information Theory. John Wiley & Sons, New York, 1991.
[16] Varsha Dani and Thomas P. Hayes. Robbing the bandit: less regret in online geometric optimization against an
adaptive adversary. In 17th ACM-SIAM Symp. on Discrete Algorithms (SODA), pages 937–943, 2006.
[17] Nikhil Devanur and Sham M. Kakade. The price of truthfulness for pay-per-click auctions. In 10th ACM Conf.
on Electronic Commerce (EC), pages 99–106, 2009.
[18] Shahar Dobzinski and Mukund Sundararajan. On characterizations of truthful mechanisms for combinatorial
auctions and scheduling. In ACM Conf. on Electronic Commerce (EC), pages 38–47, 2008.
[19] Benjamin Edelman, Michael Ostrovsky, and Michael Schwarz. Internet advertising and the generalized second-
price auction: Selling billions of dollars worth of keywords. American Economic Review, 97(1):242–259, March
2007.
[20] Abraham Flaxman, Adam Kalai, and H. Brendan McMahan. Online Convex Optimization in the Bandit Setting:
Gradient Descent without a Gradient. In 16th ACM-SIAM Symp. on Discrete Algorithms (SODA), pages 385–394,
2005.
[21] P.R. Freeman. The secretary problem and its extensions: a review. Intl. Statistics Review, 51(2):189–206, 1983.
[22] Nicola Gatti, Alessandro Lazaric, and Francesco Trovo. A Truthful Learning Mechanism for Contextual Multi-
Slot Sponsored Search Auctions with Externalities. In 13th ACM Conf. on Electronic Commerce (EC), 2012.
[23] Elad Hazan and Satyen Kale. Better algorithms for benign bandits.
In 20th ACM-SIAM Symp. on Discrete
Algorithms (SODA), pages 38–47, 2009.
29
[24] Nicole Immorlica, Kamal Jain, Mohammad Mahdian, and Kunal Talwar. Click fraud resistant methods for
learning click-through rates. In Workshop on Internet & Network Economics (WINE), pages 34–45, 2005.
[25] Sham M. Kakade, Ilan Lobel, and Hamid Nazerzadeh. Optimal Dynamic Mechanism Design and the Virtual
Pivot Mechanism. SSRN Report, SSRN ID 1782211, 2011.
[26] Richard Karp and Robert Kleinberg. Noisy binary search and its applications. In 18th ACM-SIAM Symp. on
Discrete Algorithms (SODA), pages 881–890, 2007.
[27] Robert Kleinberg. Online Decision Problems with Large Strategy Sets. PhD thesis, MIT, 2005.
[28] Robert Kleinberg. Lecture notes: CS683: Learning, Games, and Electronic Markets (week 8), Spring 2007.
Available at http://www.cs.cornell.edu/courses/cs683/2007sp/lecnotes/week8.pdf.
[29] Robert Kleinberg. Lecture notes: CS683: Learning, Games, and Electronic Markets (week 9), Spring 2007.
Available at http://www.cs.cornell.edu/courses/cs683/2007sp/lecnotes/week9.pdf.
[30] Robert Kleinberg, Aleksandrs Slivkins, and Eli Upfal. Multi-Armed Bandits in Metric Spaces. In 40th ACM
Symp. on Theory of Computing (STOC), pages 681–690, 2008.
[31] Sebastien Lahaie, David M. Pennock, Amin Saberi, and Rakesh V. Vohra.
In N. Nisan, T. Roughgarden, E.
Tardos, and V. Vazirani (eds.) Chapter 28, Sponsored search auctions. Cambridge University Press., 2007.
[32] Tze Leung Lai and Herbert Robbins. Asymptotically efficient Adaptive Allocation Rules. Advances in Applied
Mathematics, 6:4–22, 1985.
[33] John Langford and Tong Zhang. The Epoch-Greedy Algorithm for Contextual Multi-armed Bandits. In 21st
Advances in Neural Information Processing Systems (NIPS), 2007.
[34] Ron Lavi, Ahuva Mu’alem, and Noam Nisan. Towards a characterization of truthful combinatorial auctions. In
IEEE Symp. on Foundations of Computer Science (FOCS), page 574, 2003.
[35] Ron Lavi and Noam Nisan. Online ascending auctions for gradually expiring items. In 16th ACM-SIAM Symp.
on Discrete Algorithms (SODA), pages 1146–1155, 2005.
[36] Aranyak Mehta, Amin Saberi, Umesh Vazirani, and Vijay Vazirani. Adwords and generalized online matching.
J. ACM, 54(5):22, 2007.
[37] Reshef Meir, Ariel D. Procaccia, and Jeffrey S. Rosenschein. Algorithms for Strategyproof Classification. Arti-
ficial Intelligence, 186:123–156, 2012.
[38] Roger B. Myerson. Optimal Auction Design. Mathematics of Operations Research, 6:58–73, 1981.
[39] Hamid Nazerzadeh, Amin Saberi, and Rakesh Vohra. Dynamic cost-per-action mechanisms and applications to
online advertising. In 17th Intl. World Wide Web Conf. (WWW), 2008.
[40] N. Nisan and A. Ronen. Algorithmic Mechanism Design. Games and Economic Behavior, 35(1-2):166–196,
2001.
[41] N. Nisan, T. Roughgarden, E. Tardos, and V. Vazirani (eds.). Algorithmic Game Theory. Cambridge University
Press., 2007.
[42] Sandeep Pandey, Deepayan Chakrabarti, and Deepak Agarwal. Multi-armed Bandit Problems with Dependent
Arms. In 24th Intl. Conf. on Machine Learning (ICML), 2007.
[43] Christos Papadimitriou, Michael Schapira, and Yaron Singer. On the hardness of being truthful. In IEEE Symp.
on Foundations of Computer Science (FOCS), 2008.
[44] Alessandro Pavan, Ilya Segal, and Juuso Toikka. Dynamic Mechanism Design: Revenue Equivalence, Profit
Maximization, and Information Disclosure. Working paper, 2011.
[45] Filip Radlinski, Robert Kleinberg, and Thorsten Joachims. Learning diverse rankings with multi-armed bandits.
In 25th Intl. Conf. on Machine Learning (ICML), pages 784–791, 2008.
[46] Michael Rothschild. A two-armed bandit theory of market pricing. Journal of Economic Theory, 9:185–202,
1974.
30
[47] Tim Roughgarden. An algorithmic game theory primer. IFIP International Conference on Theoretical Computer
Science (TCS). An invited survey., 2008.
[48] Victor Shnayder, Jeremy Hoon, David Parkes, and Vikas Kawadia. Truthful Prioritization Schemes for Spectrum
Sharing. In 7th Workshop on the Economics of Networks, Systems and Computation (NetEcon), 2012.
[49] Aleksandrs Slivkins. Contextual Bandits with Similarity Information. In 24th Conf. on Learning Theory (COLT),
2011.
[50] Aleksandrs Slivkins. Monotone multi-armed bandit allocations. Open Problem Session at COLT 2011 (Conf. on
Learning Theory), 2011.
[51] Aleksandrs Slivkins and Eli Upfal. Adapting to a Changing Environment: the Brownian Restless Bandits. In
21st Conf. on Learning Theory (COLT), pages 343–354, 2008.
[52] Niranjan Srinivas, Andreas Krause, Sham Kakade, and Matthias Seeger. Gaussian Process Optimization in the
Bandit Setting: No Regret and Experimental Design. In 27th Intl. Conf. on Machine Learning (ICML), pages
1015–1022, 2010.
[53] Matthew Streeter and Daniel Golovin. An online algorithm for maximizing submodular functions. In Advances
in Neural Information Processing Systems (NIPS), pages 1577–1584, 2008.
[54] William R. Thompson. On the likelihood that one unknown probability exceeds another in view of the evidence
of two samples. Biometrika, 25(3-4):285294, 1933.
[55] Hal R. Varian. Position auctions. International Journal of Industrial Organization, 25(6):1163–1178, December
2007.
[56] Chris Wilkens and Balasubramanian Sivan. Single-call mechanisms. In 13th ACM Conf. on Electronic Commerce
(EC), 2012.
Appendix A: Proof of Lemma 3.9
In this section we present the full proof of Lemma 3.9. Recall that the “only if” direction is a conse-
quence of Observation 3.5. We focus on the “if” direction.
For bid profile b, click realization ρ, agent l and round t, the tuple (b; ρ; l; t) is called an influence-tuple
if round t is (b, ρ)-influential with influencing agent l. Suppose allocation A is weakly separated but not
exploration-separated. Then there is a counterexample: an influence-tuple (b; ρ; l; t) such that round t is not
bid-independent w.r.t. click realization ρ. We prove that such counterexample can occur only if bl ∈ Sl(b−l),
for some finite set Sl(b−l) ⊂ R that depends only on b−l.
Proposition A.1. Let A be as in Lemma 3.9. Assume A is weakly separated. Then for each agent l and each
bid profile b−l there exists a finite set Sl(b−l) ⊂ R with the following property: for each counterexample
(bl, b−l; ρ; l; t) it is the case that bl ∈ Sl(b−l).
Once this proposition is proved, we obtain a contradiction with the non-degeneracy of A. Indeed, sup-
pose (b; ρ; l; t) is a counterexample. Then (b; ρ; l; t) is an influence-tuple. Since A is non-degenerate, there
exists a non-degenerate interval I such that for each x ∈ I it holds that (x, b−l; ρ; l; t) is an influence-tuple,
and therefore a counterexample. Thus the set Sl(b−l) in Proposition A.1 cannot be finite, contradiction.
In the rest of this section we prove Proposition A.1. Fix a counterexample (b; ρ; l; t); let t′ > t be
the influenced round. In particular, A(b; ρ; t) = l (see 1 in Figure 1 on page 17; all boxed numbers will
refer to this figure). Then by the assumption there exist bids b′ such that A(b′; ρ; t) = i′
6= l. We claim
that this implies that there exists a bid b+
i′ , b−i′; ρ; t) = i′ (see 2 ). This is proven in
Lemma A.3 below, and in order to prove it we first present the following lemma, which essentially states
that if the mechanism makes a choice between i and j of who to be show, then it can only depend on the
ratio of their bids bidi/bidj, and not on the bids of other agents.
i′ > bi′ such that A(b+
31
Lemma A.2. Let A be an MAB (deterministic) allocation rule that is pointwise-monotone, scalefree, and
satisfies IIA. Let there be two bid profiles α and β such that A(α; ρ; t) ∈ {i, j}, A(β; ρ; t) ∈ {i, j}, and
αi/αj = βi/βj. Then it must be the case that A(α; ρ; t) = A(β; ρ; t).
Proof. As A is scalefree we assume that αi = βi and αj = βj by scaling bids in β by a factor of αi/βi (or
a factor of αj/βj), without changing the allocation.
Assume for the sake of a contradiction that A(β; ρ; t) 6= A(α; ρ; t). Let us number the agents as follows.
Agents i and j are numbered 1 and 2, respectively. The rest of the agents are arbitrarily numbered 3 to k.
Consider the following sequence of bid vectors. α(1) = α(2) = α and α(m) = (βm, α(m − 1)−m) for
m ∈ {3, . . . , k}. As α(1) = α and α(k) = β, A(α(1); ρ; t) = A(α; ρ; t) and A(α(k); ρ; t) = A(β; ρ; t).
Since A(α(k); ρ; t) = A(β; ρ; t) 6= A(α; ρ; t) = A(α(1); ρ; t) there exists m ∈ {3, . . . , k} such that
A(α(m − 1); ρ; t) = A(α; ρ; t) ∈ {i, j} while A(α(m); ρ; t) 6= A(α(m − 1); ρ; t). As m 6= i and m 6= j,
IIA implies that A(α(m); ρ; t) = m and given that, IIA also implies that A(α(k); ρ; t) ∈ {m, m + 1, . . . k}
(note that i, j are not in this set). But as A(α(k); ρ; t) = A(β; ρ; t) ∈ {i, j} this yields a contradiction.
Lemma A.3. Let A be an MAB (deterministic) allocation rule that is pointwise-monotone, scalefree, and
satisfies IIA. Let there be two bid profiles α and β such that A(α; ρ; t) = i and A(β; ρ; t) = j 6= i. Then
there exists β+
In other words, if it is possible for i to get the impression in round t at all, then it is possible for her to
i > βi such that A(β+
i , β−i; ρ; t) = i.
get the impression starting from any bid profile and raising her bid high enough.
βj
βj
αj ≥ βi
. In the bid profile (α+
Proof. We first note that αi
αj · βi
contradiction to Lemma A.2, since A(α+
but A(α+
i , α−i; ρ; t) 6= A(β; ρ; t).
< βi
βj
. If not, then αi
αj
. Consider a raised bid of i from αi to α+
i =
i , α−i), i must get the impression (by pointwise monotonicity). This gives a
= βi
,
βj
i , α−i; ρ; t) = i ∈ {i, j}, A(β; ρ; t) = j ∈ {i, j}, and α+
i
αj
αj
αj
= β+
i
βj
A(β+
A(β+
i = βj · αi
. Now, A(α; ρ; t) = i ∈ {i, j},
. We can apply Lemma A.2 to deduce that A(α; ρ; t) =
Now, consider i increasing her bid in profile β to β+
i , β−i; ρ; t) ∈ {i, j} (from IIA), and αi
i , β−i; ρ; t) and both are equal to i since the first allocation is equal to i.
From the lemma above, it follows that agent i′ can increase her bid (in bid profile b) and get the im-
pression in click realization ρ, round t. To quantify by how much agent i′ needs to raise her bid to get the
impression, we introduce the notion of threshold Θi,j(ρ; t) in the next lemma.
Lemma A.4. Let A be an MAB (deterministic) allocation rule that is pointwise monotone, scalefree and
satisfies IIA. For click realization ρ, round t, two agents i and j 6= i, let bids b−i−j be such that there
exist x0 and y satisfying A(x0, y, b−i−j; ρ; t) = j, and there exists x (possibly dependent on y) satisfying
A(x, y, b−i−j; ρ; t) = i. Let us fix such a y and define22
Θb−i−j
i,j
(ρ, t) = 1
y inf
x (cid:8)x(cid:12)(cid:12) A(x, y, b−i; ρ; t) = i(cid:9).
(ρ, t) is well defined and satisfies Θ
b′
−i−j
i,j
(ρ, t) = Θb−i−j
i,j
(ρ, t). We denote it
Then for any bids b′
by Θi,j(ρ, t), as Θ
−i−j, Θ
b−i−j
i,j
b′
−i−j
i,j
(ρ, t) is independent of b−i−j.
22Note that if there are no values of bids of i (x0 and x) and j (equal to y) such that j can get an impression with small enough
(ρ; t) at all. We will be careful
bid (x0) of agent i and i can get an impression by raising her bid (to x), then we don’t define Θ
not to use such undefined Θ’s. It is not hard to see that if bids are nonzero, then Θi,j(ρ; t) is defined if and only if Θj,i(ρ; t) is.
Moreover 0 < Θi,j(ρ; t) < ∞, and Θj,i(ρ; t) = (Θi,j(ρ; t))−1.
b−i−j
i,j
32
−i−j; ρ; t) = j then existence of y′ is proved for b′
Proof. We first prove that if the conditions of the definition of Θb−i−j
(ρ; t) are satisfied for b−i−j, then are
also satisfied for any other b′
−i−j. Let us say they are satisfied for b−i−j, that is there exists x0, x and y,
such that A(x0, y, b−i−j; ρ; t) = j and A(x, y, b−i; ρ; t) = i. We want to prove existence of x′ and y′ for
b′
−i−j. If A(x0, y, b′
−i−j too, since y′ = y works. If not,
−i−j; ρ; t) = j′ 6= j and A(x0, y, b−i−j; ρ; t) = j, and by Lemma A.3, there exists a y′ > y
then A(x0, y, b′
such that A(x0, y′, b′
−i−j; ρ; t) = j. Once the existence of y′ is proved, we now prove the existence of x′.
Let x′ = x · y′
y ≥ x. We have A(x, y, b−i−j; ρ; t) = i ∈ {i, j} and A(x′, y′, b′
−i−j; ρ; t) ∈ {i, j} by IIA
(i can only transfer impression to her by changing her bid) and x′/y′ = x/y. From Lemma A.2, we get
i = A(x, y, b−i−j; ρ; t) = A(x′, y′, b′
−i−j; ρ; t). Hence the existence of x′ is proved too.
For the sake of contradiction, let us assume that θ := Θb−i−j
(ρ; t) =: θ′. Let us scale
−i−j) by a factor such that the factor times y′ is equal to y. We can hence assume that
the bids in (x′, y′, b′
y′ = y. Let us pick a bid x′′ ∈ (θy, θ′y). We have A(x′′, y, b−i−j; ρ; t) = i (since x′′/y is past the threshold
θ), A(x′′, y′ = y, b′
−i−j; ρ; t) = j (x′′/y′ is yet not past the threshold θ′), and x′′/y = x′′/y′. This is a
contradiction to the Lemma A.2. Therefore, θ = θ′.
(ρ; t) < Θ
b′
−i−j
i,j
i,j
i,j
We conclude that if b+
i′ > bl · Θi′,l(ρ, t) then A(b+
i′ , b−i′; ρ; t) = i′ 6= l (see 2 again). Note that we are
using Θi′,l(ρ; t) since this is well-defined. Define ρ′ = ρ ⊕ 1(l, t).
Let us think about decreasing the bid of agent l from bl (it is positive, since all bids are assumed to be
positive). When the bid of agent l is bl, she gets the impression in round t, but when her bid is small enough
(in particular as low as bi′/Θi′,l(ρ; t)), then she must not get the impression in round t (see Lemma A.2).
When the bid of l decreases, some other agent gets the impression in round t, let us call that agent i (note
that this agent may not be the same as agent i′ above). See 3 .
i > blΘi,l(ρ; t), agent i gets the impression in round t (see 3 again). Note that A(b+
Now, starting from bid profile b, let us increase the bid of agent i. When the bid of agent i is large
enough (in particular as large as biΘi′,l(ρ; t)bl/bi′), then l can no longer get the impression in round t (see
Lemma A.2). From IIA, the impression must get transferred to i. Therefore we can define Θi,l(ρ; t), and
when b+
i , b−i; ρ; t) =
A(b+
i , b−i; ρ′; t) = i (click information for l at round t cannot influence the impression decision at round t).
Recall that t′ is the influenced round. Let A(b; ρ; t′) = j and let A(b; ρ′; t′) = j′ 6= j (see 4 ). As A is
pointwise monotone and IIA, A(b+
i , b−i; ρ′; t′) ∈ {i, j′}. It must be the case
that A(b+
i , b−i; ρ′; t′), as l does not get an impression at round t (and the algorithm does
not see the difference between ρ and ρ′). As j′ 6= j we conclude that
i , b−i; ρ; t′) ∈ {i, j} and A(b+
i , b−i; ρ; t′) = A(b+
A(b+
i , b−i; ρ; t′) = A(b+
i , b−i; ρ′; t′) = i.
Next we note that i 6= j and i 6= j′. This is because if i = j (respectively i = j′), then round t
would be (b; ρ)-influential (respectively (b; ρ′)-influential) with influenced agent i but it is not (b; ρ)-secured
(respectively (b; ρ′)-secured) from i, in contradiction to the assumption.
l , b−l; ρ; t) = A(b−
We also note that l ∈ {j, j′} (see 5 ). Assume for the sake of contradiction that l 6= j and l 6= j′. For
l < bi · Θl,i(ρ, t) it holds that A(b−
b−
l , b−l; ρ′; t) = i (since i was defined such that i
l , b−l; ρ; t′) = A(b−
gets the impression in round t when l decreases her bid) thus A(b−
l , b−l; ρ′; t′) (as click
information for l at round t is not observed). (Also, as a side note, observe that b−
l < bl by pointwise-
monotonicity since agent l was getting an impression in round t with bid bl and lost it when her bid is b−
l .)
Let A(b−
6= l, since otherwise, Al(x, b−l; ρ; t′) is not a
monotone function of x: it is 0 when x = bl (since j gets an impression), and 1 when x = b−
l < bl, a
contradiction to pointwise-monotonicity. Now, note that the impression in ρ′ at time t′ transfers from j′ to
l′, and impression in ρ at time t′ transfers from j to l′, none of which ({j, j′, l′}) are equal to l and j 6= j′.
Let us write this in equations:
l , b−l; ρ′; t′) = l′. Note that l′
l , b−l; ρ; t′) = A(b−
A(bl, b−l; ρ; t′) = j
A(b−
l , b−l; ρ; t′) = l′
33
A(bl, b−l; ρ′; t′) = j′
A(b−
l , b−l; ρ′; t′) = l′.
It must be the case that either j 6= l′ or j′ 6= l′ (since j 6= j′). If j 6= l′, then in ρ at time t′, reducing the
bid of l transfers impression from j to l′ (both of them are different from l), thus violating IIA. Similarly, if
j′ 6= l′, then in ρ′ at time t′, reducing the bid of l transfers impression from j′ to l′ (both of them are different
from l), thus violating IIA. We thus have l ∈ {j, j′}. Let l = j′ (since otherwise, we can swap the roles of ρ
and ρ′).
To summarize what we have proved so far: there are 3 distinct agents i, j, l such that
A(b; ρ; t) = A(b; ρ′; t) = A(b; ρ′; t′) = l
A(b; ρ; t′) = j
A(b+
i , b−i; ρ; t) = A(b+
and
i , b−i; ρ; t′) = A(b+
i , b−i; ρ′; t) = A(b+
i , b−i; ρ′; t′) = i.
(since A(b; ρ′; t′) = j′ = l),
Observe also that Θi,l(ρ, t) = Θi,l(ρ′, t) as ρ and ρ′ only differ at a click at round t, and such a click cannot
determine the allocation decision at round t. Also, max{Θi,j(ρ, t′) · bj, Θi,l(ρ′, t′) · bl} ≤ Θi,l(ρ, t) · bl as
the allocation at round t′, which is different for ρ and ρ′ (at b), depends on l getting the impression at round
t.23 Finally we prove that Θi,j(ρ, t′) · bj= Θi,l(ρ′, t′) · bl (see 8 ).
Claim A.5. Θi,j(ρ, t′) · bj = Θi,l(ρ′, t′) · bl
Proof. First of all, note that Θi,j(ρ; t′) and Θi,l(ρ′, t′) are well-defined. Let ¯bi = (Θi,j(ρ, t′)·bj +Θi,l(ρ′, t′)·
bl)/2. Consider the following two cases.
If Θi,j(ρ, t′) · bj < Θi,l(ρ′, t′) · bl then round t is (¯bi, b−i; ρ)-influential (as A(¯bi, b−i; ρ; t′) = i
and A(¯bi, b−i; ρ′; t′) = l) with influencing agent l (A(¯bi, b−i; ρ; t) = A(¯bi, b−i; ρ′; t) = l since ¯bi <
Θi,l(ρ, t)·bl) and influenced agent i. Additionally, t it is not (¯bi, b−i; ρ)-secured from i (as A(b+
i , b−i; ρ; t) =
A(b+
i , b−i; ρ′; t) = i). A contradiction to first condition in the theorem.
Similarly, if Θi,j(ρ, t′)·bj > Θi,l(ρ′, t′)·bl then round t is (¯bi, b−i; ρ)-influential (as now A(¯bi, b−i; ρ; t′) =
j and A(¯bi, b−i; ρ′; t′) = i) with influencing agent l and influenced agent i. Additionally, t it is not
(¯bi, b−i; ρ)-secured from i. Again, a contradiction to the first condition in the theorem.
The lemma implies that bl ∈ Sl(b−l), where a finite set Sl(b−l) is defined by
Sl(b−l) =(cid:26)bj
: all agents i, j 6= l, all click realizations ρ, ρ′ and all t′ s.t.
Θi,j(ρ, t′)
Θi,l(ρ′, t′)
Θi,j(ρ, t′)
Θi,l(ρ′, t′)
is well-defined(cid:27) .
This completes the proof of Proposition A.1.
23In Figure 1 we defined b+ρ
:= Θi,j (ρ; t′)bj and b+ρ′
i
:= Θi,l(ρ′; t′)bl. These are the bids of agent i at which impression
transfers to her in round t′ in ρ and ρ′ respectively. See 6 and 7 in the figure.
i
34
|
1010.5717 | 1 | 1010 | 2010-10-27T15:38:24 | PPZ For More Than Two Truth Values - An Algorithm for Constraint Satisfaction Problems | [
"cs.DS"
] | We analyze the so-called ppz algorithm for (d,k)-CSP problems for general values of d (number of values a variable can take) and k (number of literals per constraint). To analyze its success probability, we prove a correlation inequality for submodular functions. | cs.DS | cs |
PPZ For More Than Two Truth Values -- An
Algorithm for Constraint Satisfaction Problems
Dominik Scheder
Theoretical Computer Science, ETH Zurich
CH-8092 Zurich, Switzerland
[email protected]
November 5, 2018
Abstract. We analyze the so-called ppz algorithm for (d, k)-CSP prob-
lems for general values of d (number of values a variable can take) and
k (number of literals per constraint). To analyze its success probability,
we prove a correlation inequality for submodular functions.
1
Introduction
Consider the following extremely simple randomized algorithm for k-SAT: Pick
a variable uniformly at random and call it x. If the formula F contains the
unit clause (x), set x to 1. If it contains (¯x), set it to 0. It if contains neither,
set x uniformly at random (and if it contains both unit clauses, give up). This
algorithm has been proposed and analyzed by Paturi, Pudl´ak, and Zane [4] and
is called ppz.
The idea behind analyzing its success probability can be illustrated nicely
if we assume, for the moment, that F has a unique satisfying assignment α
setting all variables to 1. Switching a variable it from 1 to 0 makes the formula
unsatisfied. Therefore, there is a clause Cx = (x∨¯y1∨· · ·∨¯yk−1). With probability
1/k, the algorithm picks and sets y1, . . . , yk−1 before picking x. Supposed they
yj have been set correctly (i.e., to 1), the clause Cx is now reduced to (x),
and therefore x is also set correctly. Intuitively, this shows that on average, the
algorithm has to guess (1 − 1/k)n variables correctly and can infer the correct
values of the remaining n/k variables. This increases the success probability of
the algorithm from 2−n (simple stupid guessing) to 2−n(1−1/k).
In this paper we generalize the sketched algorithm to general constraint satis-
faction problems, short CSPs. These are a generalization of boolean satisfiability
to problems involving more than two truth values. A set of n variables x1, . . . , xn
is given, each of which can take a value from [d] := {1, . . . , d}. Each assignment
to the n variables can be represented as an element of [d]n. A literal is an ex-
pression of the form (xi 6= c) for some c ∈ [d]. A CSP formula consists of a
conjunction (AND) of constraints, where a constraint is a disjunction (OR) of
literals. We speak of (d, k)-CSP formula if each constraint consists of at most
2
k literals. Finally, (d, k)-CSP is the problem of deciding whether a given (d, k)-
CSP formula has a satisfying assignment. Note that (2, k)-CSP is the same as
k-SAT. Also (d, k)-CSP is well-known to be NP-complete, unless d = 1, k = 1, or
d = k = 2. We can manipulate a CSP formula F by permanently substituting a
value c for a variable x. This means we remove all satisfied constraints, i.e., those
containing a literal (x 6= c′) for some c′ 6= c, and from the remaining constraints
remove the literal (x 6= c), if present. We denote the resulting formula by F [x7→c].
It is obvious how to generalize the algorithm to (d, k)-CSP problems. Again
we process the variables in a random order. When picking x, we collect all unit
constraints of the form (x 6= c) and call the value c forbidden. Values in [d]
which are not forbidden are called allowed, and we set x to a value that we
choose uniformly at random from all allowed values. How can one analyze the
success probability? Let us demonstrate this for d = k = 3. Suppose F has
exactly one satisfying assignment α = (1, . . . , 1). Since changing the value of a
variable x from 1 to 2 or to 3 makes F unsatisfied, we find critical constraints
(x 6= 2 ∨ y 6= 1 ∨ z 6= 1)
(x 6= 3 ∨ u 6= 1 ∨ v 6= 1)
If all variables y, z, u, v are picked before x, then there is only one allowed value
for x left, namely 1, and with probability 1, the algorithm picks the correct
values. If y, z come before x, but at least one of u or v come after x, then it is
possible that the values 1 and 3 are allowed, and the algorithm picks the correct
value with probability 1/2. In theory, we could list all possible cases and compute
their probability. But here comes the difficulty: The probability of all variables
y, z, u, v being picked before x depends on whether these variables are distinct!
Maybe y = u, or z = v... For general d and k, we get d − 1 critical constraints
C2 := (x 6= 2 ∨ y(2)
1
C3 := (x 6= 3 ∨ y(3)
1
6= 1 ∨ · · · ∨ y(2)
6= 1 ∨ · · · ∨ y(3)
k−1 6= 1)
k−1 6= 1)
. . .
Cd := (x 6= d ∨ y(d)
1
6= 1 ∨ · · · ∨ y(d)
k−1 6= 1) .
(1)
We are interested in the distribution of the number of allowed values for x.
However, the above constraints can intersect in complicated ways, since we have
no guarantee that the variables y(c)
j are distinct. Our main technical contribution
is a sort of correlation lemma showing that in the worst case, the y(c)
are indeed
distinct, and therefore we can focus on that case, which we are able to analyze.
j
Previous Work
Feder and Motwani [1] were the first to generalize the ppz-algorithm to CSP
problems. In their paper, they consider (d, 2)-CSP problem, i.e., each variable
can take on d values, and every constraint has at most two literals. In this
case, the clauses C2, . . . , Cd cannot form complex patterns. Feder and Motwani
3
show that the worst case happens if (i) the variables y(2)
are pairwise
distinct and (ii) the CSP formula has a unique satisfying assignment. However,
their proofs do not directly generalize to higher values of k.
1 , . . . , y(d)
1
Recently, Li, Li, Liu, and Xu [2] analyzed ppz for general CSP problems (i.e.,
d, k ≥ 3). Their analysis is overly pessimistic, though, since they distinguish only
the following two cases, for each variable x: When ppz processes x, then either
(i) all d values are allowed, or (ii) at least one value is forbidden. In case (ii), ppz
chooses one value randomly from at most d − 1 values. Since case (ii) happens
with some reasonable probability, this gives a better success probability than the
trivial d−n. However, the authors ignore the case that two, three, or more values
are forbidden and lump it together with case (ii). Therefore, their analysis does
not capture the full power of ppz.
Our Contribution
Our contribution is to show that "everything works as expected", i.e., that in
the worst case all variables y(c)
in (1) are distinct and the formula has a unique
satisfying assignment. For this case, we can compute (or at least, bound from
below) the success probability of the algorithm.
j
Theorem 1.1. For d, k ≥ 1, define
G(d, k) :=
d−1
Xj=0
log2(1 + j)(cid:18)d − 1
j (cid:19)Z 1
0
(1 − rk−1)j(rk−1)d−1−jdr .
Then there is a randomized algorithm running in polynomial time which, given a
(d, k)-CSP formula over n variables, returns a satisfying assignment with prob-
ability at least 2−nG(d,k).
The algorithm we analyze in this paper is not novel. It is a straightforward
generalization of the ppz algorithm to CSP problems with more than two truth
values. However, its analysis is significantly more difficult than for d = 2 (and
also more difficult than for large d and k = 2, the case Feder and Motwani [1]
investigated).
Comparison
We compare the success probability of Schoning's random walk algorithm with
that of ppz. For ppz, we state the bound given by Li, Li, Liu, and Xu [2] and by
this paper. All bounds are approximate and ignore polynomial factors.
1.334−n
(d, k) Schoning [5] Li, Li, Liu, and Xu [2] this paper
1.588−n
(2, 3)
2.077−n
(3, 3)
3.672−n
(5, 4)
4.33−n
(6, 4)
1.588−n
2.62−n
4.73
5.73−n
3.75−n
4.5−n
2−n
4
For small values of d, in particular for the boolean case d = 2, Schoning's
random walk algorithm is much faster than ppz, but ppz overtakes Schoning
already for moderately large values of d and thus is, to our knowledge, the
currently fastest algorithm for (d, k)-CSP.
2 The Algorithm
The algorithm itself is simple. It processes the variables x1, . . . , xn according
to some random permutation π. When the algorithm processes the variable x,
it collects all unit constraints of the form (x 6= c) and calls c forbidden. A
truth value c that is not forbidden is called allowed. If the formula is satisfiable
when the algorithm processes x, there is obviously at least one allowed value.
The algorithm chooses uniformly at random an allowed value c and sets x to
c, reducing the formula. Then it proceeds to the next variable. For technical
reasons, we think of the permutation π as part of the input to the algorithm,
and sampling π uniformly at random from all n! permutations before calling
the algorithm. The algorithm is described formally in Algorithm 1. To analyze
x := xπ(i)
S(x, π) := {c ∈ [d] (x 6= c) 6∈ F }
if S(x, π) = ∅ then
return failure
Algorithm 1 ppz(F : a (d, k)-CSP formula over variables V := {x1, . . . , xn}, π:
a permutation of V )
1: α := the empty assignment
2: for i = 1, . . . , n do
3:
4:
5:
6:
7:
8:
9:
10:
11: end for
12: if α satisfies F then
13:
14: else
15:
16: end if
end if
b ←u.a.r. S(x, π)
α := α ∪ [x 7→ b]
F := F [x7→b]
return α
return failure
the success probability of the algorithm, we can assume that F is satisfiable,
i.e. the set sat(F ) of satisfying assignments is nonempty. This is because if F
is unsatisfiable, the algorithm always correctly returns failure. For a fixed
satisfying assignment, we will bound the probability
Pr[ppz(F, π) returns α] ,
(2)
5
where the probability is over the choice of π and over the randomness used by
ppz. The overall success probability is given by
Pr[ppz(F, π) is successful] = Xα∈satV (F )
Pr[ppz(F, π) returns α] .
(3)
In the next section, we will bound (2) from below. The bound depends on the
level of isolatedness of α: If α has many satisfying neighbors, its probability to be
returned by ppz decreases. However, the existence of many satisfying assignments
will in turn increase the sum in (3). In the end, it turns out that the worst case
happens if F has a unique satisfying assignment. Observe that for the ppz-
algorithm in the boolean case [4], the unique satisfiable case is also the worst
case, whereas for the improved version ppsz [3], it is not, or at least not known
to be.
3 Analyzing the Success Probability
3.1 Preliminaries
In this section, fix a satisfying assignment α. For simplicity, assume that α =
(1, . . . , 1), i.e. it sets every variable to 1. What is the probability that ppz returns
α? For a permutation π and a variable x, let β be the partial truth assignment
obtained by restricting α to the variables that come before x in π, and define
S(x, π, α) := {c ∈ [d] (x 6= c) 6∈ F [β]} .
In words, we process the variables according to π and set them according to α,
but stop before processing x. We check which truth values are not forbidden for
x by a unit constraint, and collect theses truth values in the set S(x, π, α). Let
us give an example:
Example. Let d = 3, k = 2, and α = (1, . . . , 1). We consider
F = (x 6= 2 ∨ y 6= 1) ∧ (x 6= 3 ∧ z 6= 1) .
For π = (x, y, z), no value is forbidden when processing x, thus S(x, π, α) =
{1, 2, 3}. For π′ = (y, x, z), then we consider the partial assignment that sets y
to 1, obtaining
F [y7→1] = (x 6= 2) ∧ (x 6= 3 ∨ z 6= 1) ,
and S(x, π′, α) = {1, 3}. Last, for π′′ = (y, z, x), then we set y and z to 1,
obtaining
F [y7→1,z7→1] = (x 6= 2) ∧ (x 6= 3) ,
thus S(x, π′′, α) = {1}.
(cid:3)
Observe that S(x, π, α) is non-empty, since α(x) ∈ S(x, π, α), i.e. the value
α assigns to x is always allowed. What has to happen in order for the algorithm
6
to return α? In every step of ppz, the value b selected in Line 8 for variable x
must be α(x). Assume now that this was the case in each of the first i steps of
the algorithm, i.e., the variables xπ(1), . . . , xπ(i) have been set to their respective
values under α. Let x = xπ(i+1) be the variable processed in step i + 1. The
set S(x, π, α) coincides with the set S(x, π) of the algorithm, and therefore x is
set to α(x) with probability 1/S(x, π, α). Since this holds in every step of the
algorithm, we conclude that for a fixed permutation π,
Pr[ppz(F, π) returns α] = Yx∈V
1
S(x, π, α)
.
For π being chosen uniformly at random, we obtain
Pr[ppz(F, π) returns α] = Eπ" n
Yx∈V
1
S(x, π, α)# .
The expectation of a product is an uncomfortable term if the factors are not
independent. The usual trick in this context is to apply Jensen's inequality,
hoping that we do not lose too much.
Lemma 3.1 (Jensen's Inequality). Let X be a random variable and f : R →
R a convex function. Then E[f (X)] ≥ f (E[X]), provided both expectations exist.
We apply Jensen's inequality with the convex function being f : x 7→ 2−x
and the random variable being X =Px∈V log2 S(x, π, α). With this notation,
f (X) =Qn
S(x,π,α) , the expectation of which we want to bound from below.
x∈V
1
E"Yx∈V
1
S(x, π, α)# = Eh2− Px∈V log2 S(x,π,α)i
≥ 2E[− Px∈V log2 S(x,π,α)]
= 2− Px∈V E[log2 S(x,π,α)] .
(4)
Proposition 3.2. Pr[ppz(F, π) returns α] ≥ 2− Px∈V E[log2 S(x,π,α)].
Example: The boolean case. In the boolean case, the set S(x, π, α) is either
{1} or {0, 1}, and thus the logarithm is either 0 or 1. Therefore, the term
E[log2 S(x, π, α)] is the probability that the value of x is not determined by a
unit clause, and thus has to be guessed.
So far the calculations are exactly as in the boolean ppz. This will not stay
that way for long. In the boolean case, there are only two cases: Either the value
of x is determined by a unit clause (in which we call x forced), or it is not. For
d ≥ 3, there are more cases: The set of potential values for x can be the full range
[d], it can be just the singleton {1}, but it can also be anything in between, and
even if the algorithm cannot determine the value of x by looking at unit clauses,
it will still be happy if at least, say, d/2 values are forbidden by unit clauses.
7
3.2 Analyzing E[log2 S(x, π, α)]
In this section we prove an upper bound on E[log2 S(x, π, α)]. We assume
without loss of generality that α = (1, . . . , 1). There are d truth assignments
α1, . . . , αd agreeing with α on the variables V \ {x}: For a value c ∈ [d] we define
αc := α[x 7→ c], i.e., we change the value it assignment to x to c, but keep
all other variables fixed. Clearly, α1 = α. The number of assignments among
α1, . . . , αd that satisfy F is called the looseness of α at x, denoted by
ℓ(α, x) .
Since α1 = α satisfies F , the looseness of α at x is at least 1, and since there
are d possible values for x, the looseness is at most d. Thus 1 ≤ ℓ(α, x) ≤ d.
If α is the unique satisfying assignment, then ℓ(α, x) = 1 for every x. Note
that α being unique is sufficient, but not necessary: Suppose α = (1, . . . , 1) and
α′ = (2, 2, 1, 1, . . . , 1) are the only two satisfying assignments. Then ℓ(α, x) =
ℓ(α′, x) = 1 for every variable x.
Why are we considering the looseness ℓ of α at x? Suppose without loss of
generality that the assignments α1, . . . , αℓ satisfy F , whereas αℓ+1, . . . , αd do
not. The set S(x, π, α) is a random object depending on π, but one thing is sure:
for all c = 1, . . . , ℓ(α, x) : c ∈ S(x, π, α) .
For ℓ(α, x) < c ≤ d, what is the probability that c ∈ S(x, π, α)? Since αc does
not satisfy F , there must be a constraint in F that is satisfied by α but not by
αc. Since α and αc disagree on x only, that constraint must be of the following
form:
(x 6= c ∨ y2 6= 1 ∨ y3 6= 1 ∨ · · · ∨ yk 6= 1) .
(5)
For some k − 1 variables y2, . . . , yk. We do not rule out constraints with fewer
than k − 1 literals, but we capture this by not insisting on the yj in (5) being
distinct. In any case, if the variables y2, . . . , yk come before x in the permutation
π, then c 6∈ S(x, π, α): This is because after setting to 1 the variables that come
before x, the constraint in (5) has been reduced to (x 6= c). Note that y2, . . . , yk
coming before x is sufficient for c 6∈ S(x, π, α), but not necessary, since there
could be multiple constraints of the form (5). With probability at least 1/k, all
variables y2, . . . , yk come before x, and we conclude:
Proposition 3.3. If αc does not satisfy F , then Pr[c ∈ S(x, c, α)] ≤ 1 − 1/k.
This proposition is nice, but not yet useful on its own. We can use it to finish the
analysis of the running time, however we will end up with a suboptimal estimate.
3.3 A suboptimal analysis of ppz
The function t 7→ log2(t) is concave. We apply Jensen's inequality to conclude
that
E[log2 S(x, π, α)] ≤ log2 (E[S(x, π, α)]) = log2 n
Xc=1
Pr[c ∈ S(x, π, α)]! (6)
8
We apply what we have learned above: For c = 1, . . . , ℓ(α, x), it always holds
that c ∈ S(x, π, α), and for c = ℓ(α, x) + 1, . . . , d, we have computed that Pr[c ∈
S(x, π, α)] ≤ 1 − 1/k. Therefore
E[log2 S(x, π, α)] ≤ log2(cid:18)ℓ(α, x) + (d − ℓ(α, x))(cid:18)1 −
1
k(cid:19)(cid:19) .
The unique case. If α is the unique satisfying assignment, then ℓ(α, x) = 1 for
every variable x in our CSP formula F , and the above term becomes
log2(cid:18)1 +
(d − 1)(k − 1)
k
(cid:19) = log2(cid:18) d(k − 1) + 1
k
(cid:19) .
We plug this into the bound of Proposition 3.2:
Pr[ppz returns α] ≥ 2− Pn
i=1 E[log2 S(xi,π,α)]
≥ 2−n log2( d(k−1)+1
k
)
=(cid:18) d(k − 1) + 1
k
(cid:19)−n
.
k (cid:17)n
The success probability of Schoning's algorithm for (d, k)-CSP problems is(cid:16) d(k−1)
and we see that even for the unique case, our analysis of ppz does not yield any-
thing better than Schoning. Discouraged by this failure, we do not continue this
suboptimal analysis for the non-unique case.
,
3.4 Detour: Jensen's Inequality Here, There, and Everywhere
The main culprit behind the poor performance of our analysis is Jensen's inequal-
ity in (6). To improve our analysis, we refrain from applying Jensen's inequality
there and instead try to analyze the term E[log2 S(x, π, α)] directly. However,
recall that we have used Jensen's inequality before, in (4). Is it safe to apply it
there? How can we tell when applying it makes sense and when it definitely does
not? To discuss this issue, we restate the two applications of Jensen's inequality:
Eh2− Px∈V log2 S(x,π,α)i ≥ 2E[− Px∈V log2 S(x,π,α)]
E[log2 S(x, π, α)] ≤ log2 (E[S(x, π, α)])
(7)
(8)
Formally, Jensen's inequality states that for a random variable X and a convex
function f , it holds that
E[f (X)] ≥ f (E[X]) ,
(9)
and by multiplying (9) by −1 one obtains a similar inequality for concave func-
tions. As a rule of thumb, Jensen's inequality is pretty tight if X is very concen-
trated around its expectation: In the most extreme case, X is a constant, and (9)
holds with equality. On the other extreme, suppose X is a random variable tak-
ing on values −m and m, each with probability 1/2, and let f : t 7→ t2, which is a
convex function. The left-hand side of (9) evaluates to E[f (X)] = E[X 2] = m2,
whereas the right-hand side evaluates to f (E[X]) = f (0) = 0, and Jensen's in-
equality is very loose indeed. What random variables are we dealing with in (7)
and (8)? These are
9
log2 S(x, π, α)
and
X := Xx∈V
Y := S(x, π, α) ,
and the corresponding functions are f : t 7→ 2−t, which is convex, and g :
t 7→ log2 t, which is concave. In both cases, the underlying probability space is
the set of all permutations of V , endowed with the uniform distribution. We
see that Y is not concentrated at all: Suppose x comes first in π: If our CSP
formula F contains no unit constraints, then S(x, π, α) = d, i.e., no truth value
is forbidden by a unit constraints. On the other hand, if x comes last in π,
then S(x, π, α) = ℓ(α, x). Either case happens with probability 1/n, which is
not very small. Thus, the random variable S(x, π, α) does not seem to be very
concentrated.
Contrary to Y , the random variable X can be very concentrated, in fact for
certain CSP formulas it can be a constant: Suppose d = 2, i.e., the boolean case.
Here X simply counts the number of non-forced variables. Consider the 2-CNF
formula
∧n/2
i=1 (xi ∨ yi) ∧ (xi ∨ ¯yi) ∧ (¯xi ∨ yi) .
(10)
This formula has n variables, and α = (1, . . . , 1) is the unique satisfying assign-
ment. Observe that if xi comes before yi in π, then S(xi, π, α) = {0, 1} and
S(yi, π, α) = {1}. If yi comes before xi, then S(xi, π, α) = {1} and S(yi, π, α) =
{0, 1}. Hence X ≡ n/2 is a constant. Readers who balk at the idea of supplying
a 2-CNF formula as an example for an exponential-time algorithm may try to
generalize (10) for values of k ≥ 3.
3.5 A Better Analysis
After this interlude on Jensen's inequality, let us try to bound E[log2 S(x, π, α)]
directly. In this context, x is some variable, α is a satisfying assignment, for sim-
plicity α = (1, . . . , 1), and π is a permutation of the variables sampled uniformly
at random. Again think of the d truth assignments α1, . . . , αd obtained by setting
αc := α[x 7→ c] for c = 1, . . . , d. Among them, ℓ := ℓ(α, x) satisfy the formula F .
We assume without loss of generality that those are α1, . . . , αℓ. Thus, for each
ℓ < c ≤ d, there is a constraint Cc satisfied by α but not by αc. Let us write
10
down these constraints:
Cℓ+1 := (x 6= ℓ + 1 ∨ y(ℓ+1)
Cℓ+2 := (x 6= ℓ + 2 ∨ y(ℓ+2)
1
1
6= 1 ∨ · · · ∨ y(ℓ+1)
k−1
6= 1 ∨ · · · ∨ y(ℓ+2)
k−1
6= 1)
6= 1)
. . .
(11)
Cd := (x 6= d
∨ y(d)
1
6= 1 ∨ · · · ∨ y(d)
k−1 6= 1)
We define binary random variables Y (c)
follows:
j
for 1 ≤ j ≤ k − 1 and ℓ + 1 ≤ c ≤ d as
Y (c)
j
:=(cid:26) 1 if y(c)
j
0 otherwise .
comes after x in the permutation π ,
1 ∨ · · · ∨ Y (c)
We define Y (c) := Y (c)
variables Y (1), . . . , Y (ℓ) that are constant 1. Finally, we define Y := Pd
Observe that Y (c) = 0 if and only if all variables yc
the permutation, in which case c 6∈ S(x, π, α). Therefore,
k−1. For convenience we also introduce random
c=1 Y (c).
k−1 come before x in
1, . . . , yc
S(x, π, α) ≤ Y
(12)
The variables Y (1), . . . , Y (ℓ) are constant 1, whereas each of the Y (c+1), . . . , Y (d)
is 0 with probability at least 1/k. Since 1 ≤ ℓ ≤ d, the random variable Y can
take values from 1 to d. We want to bound
E[log2 S(x, α, π)] ≤ E[log2(Y )] = E"log2 ℓ +
d
Xc=ℓ+1
Y (c)!# .
(13)
For this, we must bound the probability Pr[Y = j] for j = 1, . . . , d. This is
difficult, since the Y (c) are not independent: For example, conditioning on x
coming very early in π increases the expectation of each Y (c), and conditioning
on x coming late decreases it. We use a standard trick, also used by Paturi,
Pud´ak, Saks and Zane [3] to overcome these dependencies: Instead of viewing π
as a permutation of V , we think of it as a function V → [0, 1] where for each
x ∈ V , its value π(x) is chosen uniformly at random from [0, 1]. With probability
1, all values π(x) are distinct and therefore give rise to a permutation. The trick
is that for x, y, and z being three distinct variables, the events "y comes before
x" and "z comes before x" are independent when conditioning on π(x) = r:
Pr[π(y) < π(x) π(x) = r] = r
Pr[π(z) < π(x) π(x) = r] = r
Pr[π(x) < π(x) and π(z) < π(x) π(x) = r] = r2
11
Compare this to the unconditional probabilities:
Pr[π(y) < π(x)] =
Pr[π(z) < π(x) π(x) = r] =
Pr[π(x) < π(x) and π(z) < π(x) π(x) = r] =
1
2
1
2
1
3
j
j
j
is 1 if and only if the boolean variable y(c)
We want to compute E[Y (c) π(x) = r]. We know that E[Y (c)
since Y (c)
comes after x. Since we
are dealing with constraints of size at most k, there are, for each ℓ + 1 ≤ c ≤ d,
at most k − 1 distinct variables y(c)
k−1, and the probability that all come
before x, conditioned on π(x) = r, is at least rk−1. Therefore
π(x) = r] = 1−r,
1 , . . . , y(c)
E[Y (c)] ≤ 1 − rk−1 .
Still, a variable y(c)
j might occur in several constraints among Cℓ+1, . . . , Cd, and
therefore the Y c are not independent. The main technical tool of our analysis
is a lemma stating that the worst case is achieved exactly if they in fact are
independent, i.e., if all variables y(c)
for c = ℓ + 1, . . . , d and k = 1, . . . , k − 1 are
distinct.
Lemma 3.4 (Independence is Worst Case). Let r, k, ℓ and Y (c) be de-
fined as above. Let Z (ℓ+1), . . . , Z (d) be independent binary random variables with
E[Zi] = 1 − rk−1. Then
j
E"log2 ℓ +
d
Xc=ℓ+1
Y (c)! π(x) = r# ≤ E"log2 ℓ +
d
Xc=ℓ+1
Z (c)!# .
Before we prove the lemma in the next section, we first finish the analysis of the
algorithm. We apply a somewhat peculiar estimate: Let a ≥ 1 and b ≥ 0 be
integers. Then log2(a + b) ≤ log2(a · (b + 1)) = log2(a) + log2(b + 1). Applying
c=ℓ+1 Z (c) and combining it with the lemma and
this with a := ℓ and b := Pd
with (13), we obtain
E[log2 S(x, α, π) π(x) = r] ≤ log2(ℓ) + E"log2 1 +
d
Xc=ℓ+1
Z (c)!# . (14)
This estimate looks wasteful, but consider the case where F has a unique sat-
isfying assignment α: There, ℓ(α, x) = 1 for every variable x, and (14) holds
with equality. In addition to Z (ℓ+1), . . . , Z (d), we introduce ℓ − 1 new indepen-
dent binary random variables Z (2), . . . , Z (ℓ), each with expectation 1 − rk−1, and
define
g(d, k, r) := E"log2 1 +
Z (c)!# .
d
Xc=2
12
The only difference between the expectation in (14) and here is that here, we
sum over c = 2, . . . , d, whereas in (14) we sum only overc = ℓ + 1, . . . , d. We get
the following version of (14):
E[log2 S(x, α, π) (cid:12)(cid:12) π(x) = r] ≤ log2(ℓ) + g(d, k, r) .
We want to get rid of the condition π(x) = r. This is done by integrating (15)
for r from 0 to 1.
(15)
E[log2 S(x, α, π)] ≤ log2(ℓ) +Z 1
0
g(d, k, r)dr =: log2(ℓ) + G(d, k) .
(16)
This G(d, k) is indeed the same G(d, k) as in Theorem 1.1, and below we will do
a detailed calculation showing this.
Lemma 3.5 (Lemma 1 in Feder, Motwani [1]). Let F be a satisfiable CSP
formula over variable set V . Then
Xα∈satV (F ) Yx∈V
1
ℓ(α, x)
≥ 1 .
(17)
This lemma is a quantitative version of the intuitive statement that if a set
S ⊆ [d]n is small, then there must be rather isolated points in S. We now put
everything together:
Pr[ppsz(F, π) is successful] = Xα∈satV (F )
≥ Xα∈satV (F )
Pr[ppsz(F, π) returns α]
2− Px∈V
E[log2 S(x,α,π)] ,
where the inequality follows from (4). Together with (16), we see that
2− Px∈V
Xα∈satV (F )
E[log2 S(x,α,π)] ≥ Xα∈satV (F )
2− Px∈V (log2(ℓ(α,x))+G(d,k))
2− Px∈V log2(ℓ(α,x))
= 2−nG(d,k) Xα∈satV (F )
= 2−nG(d,k) Xα∈satV (F ) Yx∈V
≥ 2−nG(d,k) ,
1
ℓ(α, x)
with expectation 1 − rk−1 each. For 0 ≤ j ≤ d − 1, it holds that
where the last inequality follows from Lemma 3.5. To prove Theorem 1.1, we eval-
0 g(d, k, r)dr, where g(d, k, r) =
uate the term G(d, k). Recall that G(d, k) = R 1
Ehlog2(cid:16)1 +Pd
Pr" d
Xc=2
c=2 Z (c)(cid:17)i, and Z (2), . . . , Z (d) are independent binary variables
Z (c) = j# =(cid:18)d − 1
j (cid:19)(1 − rk−1)j(rk−1)(d − 1 − j) .
(18)
13
By the definition of expectation, it holds that
g(d, k, r) =
d−1
Xj=0
log2(1 + j) Pr" d
Xc=2
Z (c) = j# .
Combining this with (18) and integrating over r from 0 to 1 yields the expressions
Theorem 1.1. This finishes the proof.
4 A Correlation Inequality
The goal of this section to prove Lemma 3.4. We will prove a more general
statement.
Definition 4.1. A function f : {0, 1}n → R is called monotonically increasing,
or simply monotone, if for all x, y ∈ {0, 1}n it holds that
x ≤ y ⇒ f (x) ≤ f (y) ,
(19)
where x ≤ y is understood pointwise, i.e., xi ≤ yi for all 1 ≤ i ≤ n.
For example, the functions ∧ and ∨, seen as functions from {0, 1}n to R, are
monotone, whereas the parity function ⊕ is not.
Definition 4.2. A function f : {0, 1}n → R is called submodular if for all
x, y ∈ {0, 1}, it holds that
f (x) + f (y) ≥ f (x ∧ y) + f (x ∨ z) ,
(20)
where ∨ and ∧ are understood pointwise, i.e. (x1, . . . , xn) ∨ (y1, . . . , yn) = (x1 ∨
y1, . . . , xn ∨ yn).
Example. The OR-function f : (x1, . . . , xn) 7→ x1 ∨ · · · ∨ xn is monotone and
submodular: It is pretty clear that it is monotone, so let us try to show sub-
modularity. There are two cases: First, suppose at least one of x and y is 0, say
y = 0. Then the left-hand side of (20) evaluates to f (x), and the right-hand side
to f (0) + f (x) = f (x). If neither x = 0 nor y = 0, then the left-hand side is 2,
and the right-hand side is obviously at most 2.
Example. The AND-function g : (x1, . . . , xn) 7→ x1 ∧ · · · ∧ xn is monotone, but
not submodular. It is clearly monotone, so let us show that it is not submodu-
lar. Consider n = 2. Set x = (0, 1) and y = (1, 0). Then f (x) + f (y) = 0, but
f (x ∧ y) + f (x ∨ y) = f (0, 0) + f (1, 1) = 1.
We define the notion of glued restrictions of functions. Let A, B be two ar-
bitrary sets, and let f : An → B be a function. We define a new function f ′ by
14
f
g
f
X1X2X3X4X5X6X7
Fig. 1. A 7-ary function f and a gluing restriction g.
X1X2X3X4X5X6X7
"gluing together" two input coordinates of f . Formally, for 1 ≤ i ≤ j ≤ n, we
define the function
f ′ : (a1, . . . , an) 7→ f (a1, . . . , aj−1, ai, aj+1, . . . , an) .
The function f ′ can be viewed as a restriction of f to inputs (a1, . . . , an) for
which ai = aj. Thus, f ′ can be seen as a function An−1 → B. We prefer, however,
to define it as a function An → B that simply ignores the jth coordinate of its
input. We say f ′ is obtained from f by a gluing step. A function g : An → B is a
glued restriction of f if it can be obtained from f by a sequence of gluing steps.
See Figure 1 for an intuition.
Consider a function f : {0, 1}n → R and think of feeding f with random input
bits. Formally, let X1, . . . , Xn be n independent binary random variables, each
with expectation p. We are interested in the term E[f (X1, . . . , Xn)]. In a sec-
ond scenario, we introduce dependencies between the Xi by gluing some of them
together: For example, instead of choosing X1, . . . , Xn independently, we use the
same bit for X1, X2, and Xn, thus computing E[f (X1, X1, X3, X4, . . . , Xn−1, X1)]
instead of E[f (X1, . . . , Xn)]. With the terminology introduced above, we want to
compare E[f (X1, . . . , Xn)] to E[g(X1, . . . , Xn)], where g is a glued restriction of
f . For general functions f , we cannot say anything about how E[f (X1, . . . , Xn)]
compares to E[g(X1, . . . , Xn)]. However, if f is submodular, we can.
To get an intuition, consider the boolean lattice {0, 1}n with 0 at the bottom
and 1 at the top. In that lattice, x ∧ y is below x and y, and x ∨ y is above
them. Thus, in some sense, the points x and y lie between x ∧ y and x ∨ y.
See Figure 2 for an illustration. On the left-hand side of (20), we evaluate f at
points that lie more to the middle of the lattice, whereas on the right-hand side
we evaluate f at points that lie more to the bottom or top of it. The random
vector (X1, . . . , Xn) tends to lie around the pnth level of the lattice, whereas
(X1, X1, X3, X4, . . . , Xn−1, X1) is less concentrated and more often visits the
15
1
x ∨ y
y
x
x ∧ y
0
Fig. 2. The boolean lattice with four points x, y, x ∧ y and x ∨ y.
extremes of the lattice. In the light of (20), we expect that biasing points towards
the extremes will decrease E[f ]. The following lemma formalizes this intuition.
Lemma 4.3. Let f : {0, 1}n → R be a submodular function and g be a glued
restriction of it. Let X1, . . . , Xn be independent binary random variables, each
with expectation p. Then E[f (X1, . . . , Xn)] ≥ E[g(X1, . . . , Xn)].
Proof. It is easy to see that applying a gluing step to a submodular function
results in a submodular function: After all, a gluing step simply means restricting
the function to a subset of its domain. Therefore, it suffices to prove the lemma
for a function g that has been obtained from f by a single gluing step. Without
loss of generality, we can assume that Xn−1 and Xn have been glued together.
We have to show that
E[f (X1, . . . , Xn)] ≥ E[f (X1, . . . , Xn−1, Xn−1)] .
It suffices to show this inequality for every fixed (n − 2)-tuple of values for
(X1, . . . , Xn−2). Formally, for b1, . . . , bn−2 ∈ {0, 1}, let
g : (x, y) 7→ f (b1, . . . , bn−2, x, y) .
The function g is also submodular. Let X, Y be two independent binary ran-
dom variables, each with expectation p. We have to show that E[g(X, Y )] ≥
E[g(X, X)]. This is not difficult:
E[g(X, Y )] = (1 − p)2 · g(0, 0) + p(1 − p) · g(1, 0) +
+(1 − p)p · g(0, 1) + p2 · g(1, 1)
= (1 − p)2 · g(0, 0) + p(1 − p) · (g(1, 0) + g(0, 1)) + p2 · g(1, 1)
≥ (1 − p)2 · g(0, 0) + p(1 − p) · (g(0, 0) + g(1, 1)) + p2 · g(1, 1)
= ((1 − p)2 + p(1 − p)) · g(0, 0) + (p(1 − p) + p2) · g(1, 1)
= (1 − p) · g(0, 0) + p · g(1, 1) = E[g(X, X)] ,
16
where the inequality comes from the submodularity of g.
⊓⊔
Lemma 4.4. Let I ⊆ R be an interval, and let f : {0, 1}n → I be monotone
and submodular, and h : I → R be non-decreasing and concave. Then h ◦ f :
{0, 1}n → R is also monotone and submodular.
Proof. It is clear that h ◦ f , being the composition of two monotone functions, is
again monotone. To show submodularity, consider x, y ∈ {0, 1}n. Without loss
of generality, f (x) ≤ f (y). Using monotonicity, we see that
f (x ∧ y) ≤ f (x) ≤ f (y) ≤ f (x ∨ y) .
Claim. If s ≤ t are in I, and a ≥ b ≥ 0 are such that s − a ∈ I and t + b ∈ I,
then h(s) + h(t) ≥ h(s − a) + h(t + b).
See Figure 3 for an illustration. To prove the claim, compare the line from
h
s − a
s
t
t + b
Fig. 3. A monotone concave function f and two line segments.
(s, h(s)) to (t, h(t)) to the line from (s − a, h(s − a)) to (t + b, h(t + b)). The
midpoints of those lines have the coordinates
(cid:18) s + t
2
,
h(s) + h(t)
2
(cid:19) and (cid:18) s − a + t + b
2
,
h(s − a) + h(t + b)
2
(cid:19) ,
respectively. Since a ≥ b, the first midpoint lies to the right of the second mid-
point. Since both lines have positive slope (by monotonicity of h) and the first
line lies above the second, we conclude that also the first midpoint lies above
the second. Therefore (h(s − a) + h(t + b))/2 ≤ (h(s) + h(t))/2, as claimed.
We apply the above claim with s = f (x), t = f (y), a = f (x) − f (x ∧ y) and
b = f (x ∨ y) − f (y). Note that s, t, s − a, t + b ∈ I and a, b ≥ 0. To apply the
claim we need that a ≥ b, i.e.,
f (x) − f (x ∧ y) ≥ f (x ∨ y) − f (y) ,
17
which follows from submodularity. The claim implies that h(s) + h(t) ≥ h(s −
a) + h(t + b), which with these particular values of s,t,a, and b yields h(f (x)) +
⊓⊔
h(f (y)) ≥ h(f (x ∧ y)) + h(f (x ∨ y)).
Proof (Proof of Lemma 3.4). We define (d − ℓ)(k − 1) random variables Z (c)
for
1 ≤ j ≤ k − 1 and ℓ < c ≤ d. These random variables are all independent and
each has expectation 1 − r. We define the function f : {0, 1}(d−ℓ)(k−1) by
j
f (x(ℓ+1)
1
, . . . , x(d)
k−1) = log2 ℓ +
d
Xc=ℓ+1
OR(x(c)
1 ∨ · · · ∨ x(c)
k−1)! .
(21)
This function is clearly monotone. We claim that it is submodular: The OR-
function is submodular, and it is easy to check that a sum of submodular func-
tions is again submodular. Finally, the function t 7→ log2(ℓ + t) is concave.
We apply Lemma 4.4 with the interval I = [0, ∞), the submodular function
k−1), which has domain I, and the concave function
t 7→ log2(ℓ + t). Thus f is submodular and monotone. To prove Lemma 3.4, we
have to show that
1 ∨ · · · ∨ x(c)
c=ℓ+1 OR(x(c)
Pd
E"log2 ℓ +
d
Xc=ℓ+1
Y (c)! π(x) = r# ≤ E"log2 ℓ +
d
Xc=ℓ+1
Z (c)!# ,
(22)
where the Z (c) are independent binary random variables with expectation 1 −
rk−1 and Y (c) := OR(Y (c)
, . . . , Y (c)
k−1), with
1
Y (c)
j
:=(cid:26) 1 if y(c)
j
0 otherwise .
comes after x in the permutation π ,
The left-hand side of (22) thus reads as
E[f (Y (ℓ+1)
1
, . . . , Y (d)
k−1 π(x) = r]
for f as defined in (21). Since the Z (c) are independent binary random variables
with expectation 1 − rk−1, their distribution is identical to the distribution of
OR(Z (c)
k−1), and the right-hand side of (22) is equal to
1 , . . . , Z (c)
E[f (Z (ℓ+1)
1
, . . . , Z (d)
k−1] .
We have to show that
E[f (Y (ℓ+1)
1
, . . . , Y (d)
k−1 π(x) = r] ≤ E[f (Z (ℓ+1)
1
, . . . , Z (d)
k−1]
(23)
Conditioned on π(x) = r, the distribution of each Y (c)
are "glued together", since the underlying variables y(c)
but some Y (c)
formula need not be distinct. We can, however, assemble the Y (c)
is identical to that of Z (c)
,
j of our CSP
into groups
j
j
j
j
18
according to their underlying variables y(c)
such that (i) random variables from
the same group have the same underlying y(c)
j and thus are identical, (ii) random
variables from different groups are independent. Thus, f (Y (ℓ+1)
k−1 is a
glued restriction of f (Z (ℓ+1)
k−1 or rather can be coupled with a glued
restriction thereof, and thus by Lemma 4.3, the expectation of the former is at
⊓⊔
most the expectation of the latter. Therefore (23) holds.
, . . . , Y (d)
1
j
, . . . , Z (d)
1
References
1. T. Feder and R. Motwani. Worst-case time bounds for coloring and satisfiability
problems. J. Algorithms, 45(2):192 -- 201, 2002.
2. L. Li, X. Li, T. Liu, and K. Xu. From k-SAT to k-CSP: Two generalized algorithms.
CoRR, abs/0801.3147, 2008.
3. R. Paturi, P. Pudl´ak, M. E. Saks, and F. Zane. An improved exponential-time
algorithm for k-SAT. J. ACM, 52(3):337 -- 364, 2005.
4. R. Paturi, P. Pudl´ak, and F. Zane. Satisfiability coding lemma. Chicago J. Theoret.
Comput. Sci., pages Article 11, 19 pp. (electronic), 1999.
5. U. Schoning. A probabilistic algorithm for k-SAT and constraint satisfaction prob-
lems. In FOCS '99: Proceedings of the 40th Annual Symposium on Foundations of
Computer Science, page 410, Washington, DC, USA, 1999. IEEE Computer Society.
|
1806.04310 | 1 | 1806 | 2018-06-12T03:03:13 | MISSION: Ultra Large-Scale Feature Selection using Count-Sketches | [
"cs.DS",
"cs.LG",
"stat.ML"
] | Feature selection is an important challenge in machine learning. It plays a crucial role in the explainability of machine-driven decisions that are rapidly permeating throughout modern society. Unfortunately, the explosion in the size and dimensionality of real-world datasets poses a severe challenge to standard feature selection algorithms. Today, it is not uncommon for datasets to have billions of dimensions. At such scale, even storing the feature vector is impossible, causing most existing feature selection methods to fail. Workarounds like feature hashing, a standard approach to large-scale machine learning, helps with the computational feasibility, but at the cost of losing the interpretability of features. In this paper, we present MISSION, a novel framework for ultra large-scale feature selection that performs stochastic gradient descent while maintaining an efficient representation of the features in memory using a Count-Sketch data structure. MISSION retains the simplicity of feature hashing without sacrificing the interpretability of the features while using only O(log^2(p)) working memory. We demonstrate that MISSION accurately and efficiently performs feature selection on real-world, large-scale datasets with billions of dimensions. | cs.DS | cs | MISSION: Ultra Large-Scale Feature Selection using Count-Sketches
Amirali Aghazadeh * 1 Ryan Spring * 2 Daniel LeJeune 3 Gautam Dasarathy 3 Anshumali Shrivastava 2
Richard G. Baraniuk 3
8
1
0
2
n
u
J
2
1
]
S
D
.
s
c
[
1
v
0
1
3
4
0
.
6
0
8
1
:
v
i
X
r
a
Abstract
Feature selection is an important challenge in ma-
chine learning. It plays a crucial role in the ex-
plainability of machine-driven decisions that are
rapidly permeating throughout modern society.
Unfortunately, the explosion in the size and di-
mensionality of real-world datasets poses a se-
vere challenge to standard feature selection algo-
rithms. Today, it is not uncommon for datasets to
have billions of dimensions. At such scale, even
storing the feature vector is impossible, causing
most existing feature selection methods to fail.
Workarounds like feature hashing, a standard ap-
proach to large-scale machine learning, helps with
the computational feasibility, but at the cost of los-
ing the interpretability of features. In this paper,
we present MISSION, a novel framework for ultra
large-scale feature selection that performs stochas-
tic gradient descent while maintaining an efficient
representation of the features in memory using a
Count-Sketch data structure. MISSION retains the
simplicity of feature hashing without sacrificing
the interpretability of the features while using only
O(log2 p) working memory. We demonstrate that
MISSION accurately and efficiently performs fea-
ture selection on real-world, large-scale datasets
with billions of dimensions.
1. Introduction
Feature selection is an important step in extracting inter-
pretable patterns from data. It has numerous applications in
a wide range of areas, including natural-language process-
ing, genomics, and chemistry. Suppose that there are n or-
dered pairs (Xi, yi)i∈[n], where Xi ∈ Rp are p-dimensional
*These authors contributed equally and are listed alphabeti-
cally 1Department of Electrical Engineering, Stanford University,
Stanford, California 2Department of Computer Science, Rice Uni-
versity, Houston, Texas 3Department of Electrical and Computer
Engineering, Rice University, Houston, Texas. Correspondence to:
Anshumali Shrivastava <[email protected]>.
feature vectors, and yi ∈ R are scalar outputs. Feature selec-
tion aims to identify a small subset of features (coordinates
of the p-dimensional feature vector) that best models the
relationship between the data Xi and the output yi.
A significant complication that is common in modern engi-
neering and scientific applications is that the feature space
p is ultra high-dimensional. For example, Weinberger intro-
duced a dataset with 16 trillion (p = 1013) unique features
(Weinberger et al., 2009). A 16 trillion dimensional fea-
ture vector (of double 8 bytes) requires 128 terabytes of
working memory. Problems from modern genetics are even
more challenging. A particularly useful way to represent
a long DNA sequence is by a feature vector that counts
the occurrence frequency of all length-K sub-strings called
K-mers. This representation plays an important role in
large-scale regression problems in computational biology
(Wood & Salzberg, 2014; Bray et al., 2015; Vervier et al.,
2016; Aghazadeh et al., 2016). Typically, K is chosen to
be larger than 12, and these strings are composed of all
possible combinations of 16 characters ({A, T, C, G} in ad-
dition to 12 wild card characters). In this case, the feature
vector dimension is p = 1612 = 248. A vector of size 248
single-precision variables requires approximately 1 petabyte
of space!
For ultra large-scale feature selection problems, it is impos-
sible to run standard explicit regularization-based methods
like (cid:96)1 regularization (Shalev-Shwartz & Tewari, 2011; Tan
et al., 2014) or to select hyperparameters with a constrained
amount of memory (Langford et al., 2009). This is not sur-
prising, because these methods are not scalable in terms
of memory and computational time (Duchi et al., 2008).
Another important operational concern is that most datasets
represent features in the form of strings or tokens. For exam-
ple, with DNA or n-gram datasets, features are represented
by strings of characters. Even in click-through data (McMa-
han et al., 2013), features are indexed by textual tokens.
Observe that mapping each of these strings to a vector com-
ponent requires maintaining a dictionary whose size equals
the length of the feature vector. As a result, one does not
even have the capability to create a numerical exact vector
representation of the features.
Typically, when faced with such large machine learning
MISSION: Feature Selection via Sketching
is that
tasks, the practitioner chooses to do feature hashing (Wein-
berger et al., 2009). Consider a 3-gram string "abc". With
feature hashing, one uses a lossy, random hash function
h : strings → {0, 1, 2, . . . , R} to map "abc" to a feature
number h(abc) in the range {0, 1, 2, . . . , R}. This is ex-
tremely convenient because it enables one to avoid creating
a large look-up dictionary. Furthermore, this serves as a
dimensionality reduction technique, reducing the problem
dimension to R. Unfortunately, this convenience comes at a
cost. Given that useful dimensionality reduction is strictly
surjective (i.e., R < p), we lose the identity of the original
features. This is not a viable option if one cares about both
feature selection and interpretability.
One reason to remain hopeful
in such high-
dimensional problems, the data vectors Xi are extremely
sparse (Wood & Salzberg, 2014). For instance, the DNA
sequence of an organism contains only a small fraction (at
most the length of the DNA sequence) of p = 1612 features.
The situation is similar whether we are predicting click-
through rates of users on a website or if we seek n-gram
representations of text documents (Mikolov et al., 2013). In
practice, ultra high-dimensional data is almost always ultra-
sparse. Thus, loading a sparse data vector into memory is
usually not a concern. The problem arises in the intermedi-
ate stages of traditional methods, where dense iterates need
to be tracked in the main memory. One popular approach is
to use greedy thresholding methods (Maleki, 2009; Mikolov
et al., 2013; Jain et al., 2014; 2017) combined with stochas-
tic gradient descent (SGD) to prevent the feature vector β
from becoming too dense and blowing up in memory. In
these methods, the intermediate iterates are regularized at
each step, and a full gradient update is never stored nor
computed (since this is memory and computation intensive).
However, it is well known that greedy thresholding can be
myopic and can result in poor convergence. We clearly ob-
serve this phenomenon in our evaluations. See Section 5 for
details.
In this paper we tackle the ultra large-scale feature selec-
tion problem, i.e., feature selection with billions or more
dimensions. We propose a novel feature selection algorithm
called MISSION, a Memory-efficient, Iterative Sketching al-
gorithm for Sparse feature selectION. MISSION, that takes
on all the concerns outlined above. MISSION matches the
accuracy performance of existing large-scale machine learn-
ing frameworks like Vowpal Wabbit (VW) (Agarwal et al.,
2014) on real-world datasets. However, in contrast to VW,
MISSION can perform feature selection exceptionally well.
Furthermore, MISSION significantly surpasses the perfor-
mance of classical algorithms such as Iterative Hard Thresh-
olding (IHT), which is currently the popular feature selec-
tion alternative concerning the problem sizes we consider.
Contributions: In this work, we show that the two-decade
Figure 1. Schematic of the MISSION algorithm. MISSION itera-
tively adds the stochastic gradient term gi ∈ Rp into a Count-
Sketch and queries back the top-k heavy hitters from the Count-
Sketch. The Count-Sketch requires O(log2p) memory to store a
sketch of the O(p)-dimensional feature vector β.
old Count-Sketch data structure (Charikar et al., 2002) from
the streaming algorithms literature is ideally suited for ul-
tra large-scale feature selection. The Count-Sketch data
structure enables us to retain the convenience of feature
hashing along with the identity of important features. More-
over, Count-Sketch can accumulate gradients updates over
several iterations because of linear aggregation. This aggre-
gation eliminates the problem of myopia associated with
existing greedy thresholding approaches.
In particular, we force the parameters (or feature vector)
to reside in a memory-efficient Count-Sketch data struc-
ture (Charikar et al., 2002). SGD gradient updates are eas-
ily applied to the Count-Sketch. Instead of moving in the
gradient direction and then greedily projecting into a sub-
space defined by the regularizer (e.g., in the case of LASSO-
based methods), MISSION adds the gradient directly into the
Count-Sketch data structure, where it aggregates with all the
past updates. See Fig. 1 for the schematic. At any point of
time in the iteration, this data structure stores a compressed,
randomized, and noisy sketch of the sum of all the gradient
updates, while preserving the information of the heavy-
hitters-the coordinates that accumulate the highest amount
of energy. In order to find an estimate of the feature vec-
tor, MISSION queries the Count-Sketch. The Count-Sketch
is used in conjunction with a top-k heap, which explicitly
stores the features with the heaviest weights. Only the fea-
tures in the top-k heap are considered active, and the rest
are set to zero. However, a representation for every weight
is stored, in compressed form, inside the Count-Sketch.
We demonstrate that MISSION surpasses the sparse recov-
ery performance of classical algorithms such as Iterative
Hard Thresholding (IHT), which is the only other method
we could run at our scale. In addition, experiments suggest
that the memory requirements of MISSION scale well with
the dimensionality p of the problem. MISSION matches the
accuracy of existing large-scale machine learning frame-
works like Vowpal Wabbit (VW) on real-world, large-scale
datasets. Moreover, MISSION achieves comparable or even
MISSION: Feature Selection via Sketching
better accuracy while using significantly fewer features.
2. Review: Streaming Setting and the
Count-Sketch Algorithm
In the streaming setting, we are given a very high-
dimensional vector β ∈ Rp that is too costly to store in
memory. We see only a very long sequence of updates
over time. The only information available at time t is of
the form (i, ∆), which means that coordinate i is incre-
mented (or decremented) by the amount ∆. We are given a
limited amount of storage, on the order of O(log p), which
means that we can never store the entire sequence of updates.
Sketching algorithms aim to estimate the value of current
item i, after any number of updates using only O(log p)
memory. Accurate estimation of heavy coordinates is desir-
able.
Count-Sketch is a popular algorithm for estimation in the
streaming setting. Count-Sketch keeps a matrix of counters
(or bins) S of size d × w ∼ O(log p), where d and w are
chosen based on the accuracy guarantees. The algorithm
uses d random hash functions hj j ∈ {1, 2, ..., d} to map
the vector's components to bins w. hj : {1, 2, ..., p} →
{1, 2, ..., w} Every component i of the vector is hashed
to d different bins. In particular, for any row j of sketch
S, component i is hashed into bin S(j, hj(i)). In addition
to hj, Count-Sketch uses d random sign functions to map
the components of the vectors randomly to {+1, −1}. i.e.,
si : {1, 2, ..., D} → {+1,−1} A picture of this sketch
data structure with three hash functions in shown inside
Fig. 1.
The Count-Sketch supports two operations: UPDATE(item i,
increment ∆) and QUERY(item i). The UPDATE operation
updates the sketch with any observed increment. More for-
mally, for an increment ∆ to an item i, the sketch is updated
by adding sj(i)∆ to the cell S(j, hj(i)) ∀j ∈ {1, 2, ..., d}.
The QUERY operation returns an estimate for component i,
the median of all the d different associated counters.
It has been shown that, for any sequence of streaming up-
dates (addition or subtraction) to the vector β, Count-Sketch
provides an unbiased estimate of any component i, (cid:98)βi such
that the following holds with high probability,
βi − β2 ≤ (cid:98)βi ≤ βi + β2.
(1)
It can be shown that the Eq. (1) is sufficient to achieve
near-optimal guarantees for sparse recovery with the given
space budget. Furthermore, these guarantees also meet
the best compressed sensing lower bounds in terms of the
number of counters (or measurements) needed for sparse
recovery (Indyk, 2013).
3. Problem Formulation
Consider the feature selection problem in the ultra high-
dimensional setting: We are given the dataset (Xi, yi) for
i ∈ [n] = {1, 2, . . . , n}, where Xi ∈ Rp and yi ∈ R denote
the ith measured and response variables. We are interested
in finding the k-sparse (k non-zero entries) feature vector
(or regressor) β ∈ Rp from the optimization problem
(cid:107)y − Xβ(cid:107)2,
min
(cid:107)β(cid:107)0=k
(2)
where X = {X1, X2, . . . , Xn} and y = [y1, y1, . . . , yn]
denote the data matrix and response vector and the (cid:96)0-norm
(cid:107)β(cid:107)0 counts the number of non-zero entries in β.
We are interested in solving the feature selection problem
for ultra high-dimensional datasets where the number of
features p is so large that a dense vector (or matrix) of size
p cannot be stored explicitly in memory.
3.1. Hard Thresholding Algorithms
Among the menagerie of feature selection algorithms, the
class of hard thresholding algorithms have the smallest mem-
ory footprint: Hard thresholding algorithms retain only the
top-k values and indices of the entire feature vector us-
ing O(klog(p)) memory (Jain et al., 2014; Blumensath &
Davies, 2009). The iterative hard thresholding (IHT) algo-
rithm generates the following iterates for the ith variable in
an stochastic gradient descent (SGD) framework
βt+1 ← Hk(βt − 2λ(cid:0)yi − Xiβt(cid:1)T
Xi)
(3)
The sparsity of the feature vector βt, enforced by the hard
thresholding operator Hk, alleviates the need to store a
vector of size O(p) in the memory in order to keep track of
the changes of the features over the iterates.
Unfortunately, because it only retains the top-k elements
of β, the hard thresholding procedure greedily discards the
information of the non top-k coordinates from the previous
iteration. In particular, it clips off coordinates that might add
to the support set in later iterations. This drastically affects
the performance of hard thresholding algorithms, especially
in real-world scenarios where the design matrix X is not ran-
dom, normalized, or well-conditioned. In this regime, the
gradient terms corresponding to the true support typically
arrive in lagging order and are prematurely clipped in early
iterations by Hk. The effect of these lagging gradients is
present even in the SGD framework, because the gradients
are quite noisy, and only a small fraction of the energy of
the true gradient is expressed in each iteration. It is not
difficult to see that these small energy, high noise signals
can easily cause the greedy hard thresholding operator to
make sub-optimal or incorrect decisions. Ideally, we want
to accumulate the gradients to get enough confidence in
MISSION: Feature Selection via Sketching
Algorithm 1 MISSION
(cid:16)
(cid:17)
Initialize: β0 = 0, S (Count-Sketch), λ (Learning Rate)
while not stopping criteria do
2 (yi − Xiβt)
T Xi
Find the gradient update gi = λ
Add the gradient update to the sketch gi → S
Get the top-k heavy-hitters from the sketch βt+1 ← S
end while
Return: The top-k heavy-hitters from the Count-Sketch
signal and to average out any noise. However, accumulating
gradients will make the gradient vector dense, blowing up
the memory requirements. This aforementioned problem
is in fact symptomatic of all other thresholding variants in-
cluding the Iterative algorithm with inversion (ITI) (Maleki,
2009) and the Partial hard thresholding (PHT) algorithm
(Jain et al., 2017).
4. The MISSION Algorithm
We now describe the MISSION algorithm. First, we initialize
the Count-Sketch S and the feature vector βt=0 with zeros
entries. The Count-Sketch hashes a p-dimensional vector
into O(log2p) buckets (Recall Fig. 1). We discuss this
particular choice for the size of the Count-Sketch and the
memory-accuracy trade offs of MISSION in Sections 5.3
and 6.1.
At iteration t, MISSION selects a random row Xi from the
data matrix X and computes the stochastic gradient update
term using the learning rate λ. gi = 2λ (yi − Xiβt)
T Xi
the usual gradient update that minimizes the uncon-
i.e.
strained quadratic loss (cid:107)y − Xβ(cid:107)2
2. The data vector Xi
and the corresponding stochastic gradient term are sparse.
We then add the non-zero entries of the stochastic gradient
term {gij : ∀j gij > 0} to the Count-Sketch S. Next, MIS-
SION queries the top-k values of the sketch to form βt+1.
We repeat the same procedure until convergence. MISSION
returns the top-k values of the Count-Sketch as the final
output of the algorithm. The MISSION algorithm is detailed
in Alg. 1. MISSION easily extends to other loss functions
such as the hinge loss and logistic loss.
MISSION is Different from Greedy Thresholding: De-
note the gradient vector update at any iteration t as ut. It is
not difficult to see that starting with an all-zero vector β0, at
any point of time t, the Count-Sketch state is equivalent to
i=1 ut. In other words, the sketch
aggregates the compressed aggregated vector. Thus, even
if an individual SGD update is noisy and contains small
signal energy, thresholding the Count-Sketch is based on
the average update over time. This averaging produces a
robust signal that cancels out the noise. We can therefore
expect MISSION to be superior over thresholding.
the sketch of the vector(cid:80)t
In the supplementary materials, we present initial theoretical
results on the convergence of MISSION. Our results show
that, under certain assumptions, the full-gradient-descent
version of MISSION converges geometrically to the true
parameter β ∈ Rp up to some additive constants. The
exploration of these assumptions and the extension to the
SGD version of MISSION are exciting avenues for future
work.
Feature Selection with the Ease of Feature Hashing: As
argued earlier, the features are usually represented with
strings, and we do not have the capability to map each string
to a unique index in a vector without spending O(p) memory.
Feature hashing is convenient, because we can directly ac-
cess every feature using hashes. We can use any lossy hash
function for strings. MISSION only needs a few independent
hash functions (3 in our Count-Sketch implementation) to
access any component. The top-k estimation is done effi-
ciently using a heap data structure of size k. Overall, we
only access the data using efficient hash functions, which
can be easily implemented in large-scale systems.
5. Simulations
We designed a set of simulations to evaluate MISSION in
a controlled setting.
In contrast to the ultra large-scale,
real-world experiments of Section 6, in the section the data
matrices are drawn from a random Gaussian distribution
and the ground truth features are known.
5.1. Phase Transition
We first demonstrate the advantage of MISSION over greedy
thresholding in feature selection. For this experiment, we
modify MISSION slightly to find the root of the algorithmic
advantage of MISSION: we replace the Count-Sketch with
an "identity" sketch, or a sketch with a single hash func-
tion, h(i) = i. In doing so, we eliminate the complexity
that Count-Sketch adds to the algorithm, so that the main
difference between MISSION and IHT is that MISSION ac-
cumulates the gradients. To improve stability, we scale the
non top-k elements of S by a factor γ ∈ (0, 1) that begins
very near 1 and is gradually decreased until the algorithm
converges. Note: it is also possible to do this scaling in the
Count-Sketch version of MISSION efficiently by exploiting
the linearity of the sketch.
Fig. 2 illustrates the empirical phase transition curves for
sparse recovery using MISSION and the hard thresholding
algorithms. The phase transition curves show the points
where the algorithm successfully recovers the features in
> 50% of the random trails. MISSION shows a better phase
transition curve compared to IHT by a considerable gap.
MISSION: Feature Selection via Sketching
Table 1. Comparison of MISSION against hard thresholding algorithms in subset selection under adversarial effects. We first report the
percentage of instances in which the algorithms accurately find the solution (ACC) with no attenuation (α = 1) over 100 random trials.
We then report the mean of the maximum level of attenuation α applied to the columns of design X before the algorithms fail to recover
the support of β (over the trials that all algorithms can find the solution with α = 1).
(n, k)
(100, 2)
(100, 3)
(100, 4)
(200, 5)
(200, 6)
(200, 7)
ACCα=1
100%
100%
100%
100%
100%
100%
MISSION
α
2.68 ± 0.37
2.52 ± 0.36
2.53 ± 0.23
4.07 ± 0.36
4.17 ± 0.24
4.07 ± 0.11
ACCα=1
100%
92%
72%
99%
97%
83%
α
IHT
1.49 ± 0.33
1.36 ± 0.46
1.92 ± 0.91
2.34 ± 1.12
2.64 ± 1.14
1.64 ± 1.01
ACCα=1
91%
70%
37%
37%
23%
14%
α
ITI
1.33 ± 0.23
1.15 ± 0.20
1.03 ± 0.09
1.15 ± 0.22
1.11 ± 0.12
1.11 ± 0.12
ACCα=1
64%
42%
39%
83%
73%
75%
PHT
α
2.42 ± 0.87
2.05 ± 0.93
2.13 ± 1.07
2.75 ± 1.30
2.26 ± 1.33
3.39 ± 1.36
columns (coordinates) with smaller energy typically lag and
are thresholded by IHT.
To this end, we first construct a random Gaussian data matrix
X ∈ R900×1000, pick a sparse vector β that is supported on
an index set I, and then attenuate the energy of the columns
of X supported by the indices in I by an attenuation fac-
tor of α = {1, 1.25, 1.5, 1.75, 2, . . . , 5}. Note that α = 1
implies that no attenuation is applied to the matrix. In Ta-
ble 1, we report the maximum attenuation level applied to a
column of data matrix X before the algorithms fail to fully
recover the support set I from y = βX. We observe that
MISSION is consistently and up to three times more robust
against adversarial attenuation of the columns of the data
matrix in various design settings.
The robustness of MISSION to the attenuation of the
columns of X in sparse recovery task suggests that the
Count-Sketch data structure enables gradient-based opti-
mization methods such as IHT to store a footprint (or sketch)
of all the gradients from the previous iterations and deliver
them back when they become prominent.
5.3. Logarithmic Scaling of the Count-Sketch Memory
in MISSION
In this section we demonstrate that the memory require-
ments of MISSION grows polylogarithmically in the dimen-
sion of the problem p. We conduct a feature selection ex-
periment with a data matrix X ∈ R100×p whose entries are
drawn from i.i.d. random Gaussian distributions with zero
mean and unit variance. We run MISSION and IHT to re-
cover the feature vector β from the output vector y = Xβ,
where the feature vector β is a k = 5-sparse vector with
random support. We repeat the same experiment 1000 times
with different realizations for the sparse feature vector β
and report the results in Fig. 3. The left plot illustrates the
feature selection accuracy of the algorithms as the dimen-
sion of the problem p grows. The right plot illustrates the
minimum memory requirements of the algorithms to recover
the features with 100% accuracy.
Figure 2. Empirical phase transition in recovering a binary feature
vector β in p = 1000-dimensional space with a Gaussian data
matrix X. We illustrate the empirical 50% probability of success
curves averaged over T = 20 trials. MISSION outperforms the
thresholding algorithms by a large margin.
5.2. Lagging Gradient: Superiority of Count-Sketches
over Greedy Thresholding
A major problem with the IHT algorithm, especially in large-
scale SGD settings, is with thresholding the coordinates with
small gradients in the earlier iterations. IHT misses these
coordinates, since they become prominent only after the
gradients accumulate with the progression of the algorithm.
The problem is amplified with noisy gradient updates such
as SGD, which is unavoidable for large datasets.
This phenomenon occurs frequently in sparse recovery prob-
lems. For example, when the coordinates that correspond to
the columns of the data matrix with smaller energy lag in
the iterations of gradient descent algorithm, IHT thresholds
these lagging-gradient coordinates in first few iterations,
and they never show up again in the support. In contrast,
MISSION retains a footprint of the gradients of all the pre-
vious iterations in the Count-Sketch. When the total sum
of the gradient of a coordinate becomes prominent, the co-
ordinate joins the support after querying the top-k heavy
hitters from the Count-Sketch. We illustrate this phenom-
ena in sparse recovery using synthetic experiments. We re-
cover sparse vector β from its random linear measurements
y = Xβ, where the energy of X is imbalanced across its
columns. In this case, the gradients corresponding to the
0.00.20.40.60.81.0n/p0.00.20.40.60.81.0s/nMISSIONIHTITIPHT MISSION: Feature Selection via Sketching
6.1. Large-scale Feature Extraction
Datasets: We used four datasets in the experiments: 1)
KDD2012, 2) RCV1, 3) Webspam–Trigram, 4) DNA 2. The
statistics of these datasets are summarized in Table 2.
Table 2. Feature extraction dataset statistics.
Dataset
KDD 2012
RCV1
Webspam
DNA (Tiny)
Dim (p)
54,686,452
47,236
16,609,143
14,890,408
Train Size (n) Test Size
119,705,032
20,242
280,000
1,590,000
29,934,073
677,399
70,000
207,468
The DNA metagenomics dataset is a multi-class classifica-
tion task where the model must classify 15 different bacteria
species using DNA K-mers. We sub-sampled the first 15
species from the original dataset containing 193 species. We
use all of the species in the DNA Metagenomics dataset for
the large-scale experiments (See Section 6.2). Following
standard procedures, each bacterial species is associated
with a reference genome. Fragments are sampled from the
reference genome until each nucleotide is covered c times
on average. The fragments are then divided into K-mer
sub-strings. We used fragments of length 200 and K-mers
of length 12. Each model was trained and tested with mean
coverage c = {0.1, 1} respectively. For more details, see
(Vervier et al., 2016). The feature extraction task is to find
the DNA K-mers that best represent each bacteria class.
We implemented the following approaches to compare and
contrast against our approach: For all methods, we used the
logistic loss for binary classification and the cross-entropy
loss for multi-class classification.
MISSION: As described in Section 4.
Iterative Hard Thresholding (IHT): An algorithm where,
after each gradient update, a hard threshold is applied to the
features. Only the top-k features are kept active, while the
rest are set to zero. Since the features are strings or integers,
we used a sorted heap to store and manipulate the top-k ele-
ments. This was the only algorithm we could successfully
run over the large datasets on our single machine.
Batch IHT: A modification to IHT that uses mini-batches
such that the gradient sparsity is the same as the number
of elements in the count-sketch. We accumulate features
and then sort and prune to find the top-k features. This
accumulate, sort, prune process is repeated several times
during training. Note: This setup requires significantly
more memory than MISSION, because it explicitly stores
the feature strings. The memory cost of maintaining a set
of string features can be orders of magnitude more than the
flat array used by MISSION. See Bloom Filters (Broder &
Mitzenmacher, 2004) and related literature. This setup is
not scalable to large-scale datasets.
2http://projects.cbio.mines-paristech.fr/
largescalemetagenomics/
Figure 3. Feature selection accuracy and memory requirements of
MISSION and Hard Thresholding. The memory requirements of
MISSION grows polylogarithmcially ∼ O(log2(p)) (dotted line
illustrates quadratic fit) in p. With only a logarithmic factor more
memory, MISSION has significant advantage over Hard Threshold-
ing in terms of feature selection accuracy.
The plots reveal an interesting phenomenon. The size of
the Count-Sketch in MISSION scales only polylogarithmi-
cally with the dimension of the problem. This is surprising
since the aggregate gradient in a classical SGD framework
becomes typically dense in early iterations and thus requires
a memory of order O(p). MISSION, however, stores only
the essential information of the features in the sketch using
a poly-logarithmic sketch size. Note that IHT sacrifices
accuracy to achieve a small memory footprint. At every
iteration IHT eliminates all the information except for the
top-k features. We observe that, using only a logarithmic
factor more memory, MISSION has a significant advantage
over IHT in recovering the ground truth features.
6. Experiments
All experiments were performed on a single machine, 2x
Intel Xeon E5-2660 v4 processors (28 cores / 56 threads)
with 512 GB of memory. The code 1 for training and running
our randomized-hashing approach is available online. We
designed the experiments to answer these questions:
1. Does MISSION outperform IHT in terms of classifica-
tion accuracy? In particular, how much does myopic
thresholding affect IHT in practice?
2. How well does MISSION match the speed and accuracy
of feature hashing (FH)?
3. How does changing the number of top-k features affect
the accuracy and behaviour of the different methods?
4. What is the effect of changing the memory size of
the Count-Sketch data structure on the classification
accuracy of MISSION in read-world datasets?
5. Does MISSION scale well in comparison to the differ-
ent methods on the ultra large-scale datasets (> 350
GB in size)?
1https://github.com/rdspring1/MISSION
23456log10(p)020406080100Average Accuracy23456log10(p)0100200300400500600MemeoryMISSIONHard Tresholding MISSION: Feature Selection via Sketching
Feature Hashing (FH): A standard machine learning algo-
rithm for dimensionality reduction that reduces the memory
cost associated with large datasets. FH is not a feature se-
lection algorithm and cannot identify important features.
(Agarwal et al., 2014)
Experimental Settings: The MISSION and IHT algorithms
searched for the same number of top-k features. To ensure
fair comparisons, the size of the Count-Sketch and the fea-
ture vector allocated for the FH model were equal. The
size of the MISSION and FH models were set to the near-
est power of 2 greater than the number of features in the
dataset. For all the experiments, the Count-Sketch data
structure used 3 hash functions, and the model weights were
divided equally among the hash arrays. For example, with
the (Tiny) DNA metagenomics dataset, we allocated 24 bits
or 16,777,216 weights for the FH model. Given 3 hash
functions and 15 classes, roughly 372,827 elements were
allocated for each class in the Count-Sketch.
MISSION, IHT, FH Comparison: Fig. 4 shows that MIS-
SION surpasses IHT in classification accuracy in all four
datasets, regardless of the number of features. In addition,
MISSION closely matches FH, which is significant because
FH is allowed to model a much larger set of features than
MISSION or IHT. MISSION is 2–4× slower than FH, which
is expected given that MISSION has the extra overhead of
using a heap to track the top-k features.
MISSION's accuracy rapidly rises with respect to the num-
ber of top-k features, while IHT's accuracy plateaus and
then grows slowly to match MISSION. This observation
corroborates our insight that the greedy nature of IHT hurts
performance. When the number of top-k elements is small,
the capacity of IHT is limited, so it picks the first set of
features that provides good performance, ignoring the rest.
On the other hand, MISSION decouples the memory from
the top-k ranking, which is based on the aggregated gradi-
ents in the compressed sketch. By the linear property of the
count-sketch, this ensures that the heavier entries occur in
the top-k features with high probability.
Count-Sketch Memory Trade-Off: Fig. 5 shows how
MISSION's accuracy degrades gracefully, as the size of the
Count-Sketch decreases. In this experiment, MISSION only
used the top 500K features for classifying the Tiny DNA
metagenomics dataset. When the top-k to Count-Sketch
ratio is 1, then 500K weights were allocated for each class
and hash array in the Count-Sketch data structure. The
Batch IHT baseline was given 8,388,608 memory elements
per class, enabling it to accumulate a significant number of
features before thresholding to find the top-k features. This
experiment shows that MISSION immediately outperforms
IHT and Batch IHT, once the top-k to Count-Sketch ratio is
1:1. Thus, MISSION provides a unique memory-accuracy
knob at any given value of top-k.
6.2. Ultra Large-Scale Feature Selection
Here we demonstrate that MISSION can extract features
from three large-scale datasets: Criteo 1TB, Splice-Site, and
DNA Metagenomics.
Table 3. Ultra Large-Scale dataset statistics
Dataset
Criteo
Splice-Site
DNA
Dim (p) Train Size (n)
1M
4,195,197,692
11.7M 50,000,000
17.3M 13,792,260
Test Size
178,274,637
4,627,840
354,285
Criteo 1TB: The Criteo 1TB 3 dataset represents 24 days of
click-through logs-23 days (training) + 1 day (testing). The
task for this dataset is click-through rate (CTR) prediction-
How likely is a user to click an ad? The dataset contains over
4 billion (training) and 175 million (testing) examples (2.5
TB of disk space). The performance metric is Area Under
the ROC Curve (AUC). The VW baseline 4 achieved 0.7570
AUC score. MISSION and IHT scored close to the VW
baseline with 0.751 AUC using only the top 250K features.
Table 4. Criteo 1TB. Top-K Features: 250K
Metric MISSION
AUC
0.751
IHT
0.752
VW
0.757
Splice-Site: The task for this dataset is to distinguish be-
tween true and fake splice sites using the local context
around the splice site in-question. The dataset is highly
skewed (few positive, many negative values), and so the
performance metric is average precision (AP). Average pre-
cision is the precision score averaged over all recall scores
ranging from 0 to 1. The dataset contains over 50 million
(training) and 4.6 million (testing) examples (3.2 TB of disk
space). All the methods were trained for a single epoch
with a learning rate of 0.5. MISSION, Batch IHT, and SGD
IHT tracked the top 16,384 features. FH, MISSION, and
Batch IHT used 786,432 extra memory elements. MISSION
significantly outperforms Batch IHT and SGD IHT by 2.3%.
Also, unlike in Fig. 5, the extra memory did not help Batch
IHT, since it performed the same as SGD IHT. MISSION
(17.5 hours) is 15% slower than FH (15 hours) in wall-clock
running time.
Table 5. Splice-Site: Top-k features: 16,384. Memory elements:
786,432. MISSION outperforms Batch IHT and SGD IHT.
Metric
AP
FH
0.522
MISSION Batch IHT SGD IHT
0.510
0.498
0.498
DNA Metagenomics: This experiment evaluates MIS-
SION's performance on a medium-sized metagenomics
dataset. The parameters from the Tiny (15 species) dataset
in Section 6.1 are shared with this experiment, except the
3https://www.kaggle.com/c/criteo-display-ad-challenge
4https://github.com/rambler-digital-solutions/
criteo-1tb-benchmark
MISSION: Feature Selection via Sketching
Figure 4. Feature selection on the KDD-2012, RCV1, Webspam, and DNA Metagenomic (Tiny) datasets. FH is not a feature selection
baseline. It is marked by a dashed black line, indicating that its performance is invariant to the number of top-k features.
Figure 5. Count-Sketch Memory/Accuracy Trade-off using the
DNA Metagenomics (Tiny) dataset. The x-axis is the ratio of
Count-Sketch memory per class and hash-array to the # top-k
features per class. Memory Elements for Batch IHT: 8,388,608.
number of species is increased to 193. The size of a sample
batch with mean coverage c = 1 increased from 7 GB (Tiny)
to 68 GB (Medium). Each round (mean coverage c = 0.25)
contains 3.45 million examples and about 16.93 million
unique non-zero features (p). MISSION and IHT tracked
the top 2.5 million features per class. The FH baseline used
231 weights, about 11.1 million weights per class, and we
allocated the same amount of space for the Count-Sketch.
Each model was trained on a dataset with coverage c = 5.
Fig. 6 shows the evolution of classification accuracy over
time for MISSION, IHT, and the FH baseline. After 5 epochs,
MISSION closely matches the FH baseline. Note: MISSION
converges faster than IHT such that MISSION is 1–4 rounds
ahead of IHT, with the gap gradually increasing over time.
On average, the running time of MISSION is 1–2× slower
than IHT. However, this experiment demonstrates that since
MISSION converges faster, it actually needs less time to
reach a certain accuracy level. Therefore, MISSION is effec-
tively faster and more accurate than IHT.
7. Implementation Details and Discussion
Scalability and Parallelism: IHT finds the top-k features
after each gradient update, which requires sorting the fea-
tures based on their weights before thresholding. The speed
of the sorting process is improved by using a heap data
structure, but it is still costly per update. MISSION also uses
Table 6. Ultra Large-Scale Feature Selection for the DNA Metage-
nomics (Medium) Dataset (193 species), Mean Coverage c = 5.
a heap to store its top-k elements, but it achieves the same
accuracy as IHT with far fewer top-k elements because of
the Count-Sketch. (Recall Section 4)
Another suggested improvement for the top-k heap is to
use lazy updates. Updating the weight of a feature does not
change its position in the heap very often, but still requires
an O(log n) operation. With lazy updates, the heap is up-
dated only if it the change is significant. xt − x0 ≥ , i.e.
the new weight at time t exceeds the original value by some
threshold. This tweak significantly reduces the number of
heap updates at the cost of slightly distorting the heap.
8. Conclusion and Future Work
In this paper, we presented MISSION, a new framework for
ultra large-scale feature selection that performs hard thresh-
olding and SGD while maintaining an efficient, approximate
representation for all features using a Count-Sketch data
structure. MISSION retains the simplicity of feature hashing
without sacrificing the interpretability of the features.
Interaction features are important for scientific discovery
with DNA Metagenomics (Basu et al., 2018). Tradition-
ally, the polynomial kernel trick enabled machine learning
algorithms to explore this feature space implicitly without
the exponential increase in dimensionality. However, this
exponential cost is unavoidable with feature extraction. Go-
ing forward, we are interested in leveraging our MISSION
framework to explore pairwise or higher interaction features.
0.650.680.700.730.750.780.8002000004000006000008000001000000AUC#FeaturesKDD 2012MISSIONIHTFH0.500.550.600.650.700.750.800.850.900.951.000200040006000800010000AUC#FeaturesRCV10.900.910.920.930.940.950.960.970.980.991.00010002000300040005000AUC#FeaturesWebspam -Trigram00.10.20.30.40.50.60.70.80.9101000000200000030000004000000Accuracy#FeaturesDNA -Tiny (15 Species) -14,890,408 Features0.550.600.650.700.750.800246810AccuracyTopK : CMS Ratio (%)DNA -Tiny (15 Species) -Top-K: 500KMissionIHTBatch IHT00.10.20.30.40.50.60.70.80.9105101520AccuracyEpochsDNA -Medium (193 Species) -Top-K: 2.5MMissionIHTFH MISSION: Feature Selection via Sketching
Acknowledgements
AAA, DL, GD, and RB were supported by the DOD Van-
nevar Bush Faculty Fellowship grant N00014-18-1-2047,
NSF grant CCF-1527501, ARO grant W911NF-15-1-0316,
AFOSR grant FA9550-14-1-0088, ONR grant N00014-17-1-
2551, DARPA REVEAL grant HR0011-16-C-0028, and an
ONR BRC grant for Randomized Numerical Linear Algebra.
RS and AS were supported by NSF-1652131, AFOSR-YIP
FA9550-18-1-0152, and ONR BRC grant for Randomized
Numerical Linear Algebra. The authors would also like
to thank NVIDIA and Amazon for gifting computing re-
sources.
References
Agarwal, A., Chapelle, O., Dud´ık, M., and Langford, J. A
reliable effective terascale linear learning system. Journal
of Machine Learning Research, 15(1):1111–1133, 2014.
Aghazadeh, A., Lin, A. Y., Sheikh, M. A., Chen, A. L.,
Atkins, L. M., Johnson, C. L., Petrosino, J. F., Drezek,
R. A., and Baraniuk, R. G. Universal microbial diagnos-
tics using random dna probes. Science advances, 2(9):
e1600025, 2016.
Basu, S., Kumbier, K., Brown, J. B., and Yu, B. Iterative
random forests to discover predictive and stable high-
order interactions. Proceedings of the National Academy
of Sciences, 2018. ISSN 0027-8424. doi: 10.1073/pnas.
1711236115. URL http://www.pnas.org/content/early/2018/
01/17/1711236115.
Blumensath, T. and Davies, M. E. Iterative hard thresholding
for compressed sensing. Applied and Computational
Harmonic Analysis, 27(3):265–274, 2009.
Bray, N., Pimentel, H., Melsted, P., and Pachter, L.
Near-optimal RNA-Seq quantification. arXiv preprint
arXiv:1505.02710, 2015.
Broder, A. and Mitzenmacher, M. Network applications
of bloom filters: A survey. Internet mathematics, 1(4):
485–509, 2004.
Charikar, M., Chen, K., and Farach-Colton, M. Finding
frequent items in data streams. In International Collo-
quium on Automata, Languages, and Programming, pp.
693–703. Springer, 2002.
Duchi, J., Shalev-Shwartz, S., Singer, Y., and Chandra, T.
Efficient projections onto the (cid:96)1-ball for learning in high
In Proceedings of the 25th International
dimensions.
Conference on Machine Learning, pp. 272–279. ACM,
2008.
Indyk, P. Sketching via hashing: From heavy hitters to
compressed sensing to sparse fourier transform. In Pro-
ceedings of the 32nd ACM SIGMOD-SIGACT-SIGAI Sym-
posium on Principles of Database Systems, pp. 87–90.
ACM, 2013.
Jain, P., Tewari, A., and Kar, P. On iterative hard thresh-
olding methods for high-dimensional m-estimation. In
Advances in Neural Information Processing Systems, pp.
685–693, 2014.
Jain, P., Tewari, A., and Dhillon, I. S. Partial hard thresh-
olding. IEEE Transactions on Information Theory, 63(5):
3029–3038, 2017.
Langford, J., Li, L., and Zhang, T. Sparse online learning
via truncated gradient. Journal of Machine Learning
Research, 10(Mar):777–801, 2009.
Maleki, A. Coherence analysis of iterative thresholding
algorithms. In Communication, Control, and Computing,
2009. Allerton 2009. 47th Annual Allerton Conference
on, pp. 236–243. IEEE, 2009.
McMahan, H. B., Holt, G., Sculley, D., Young, M., Ebner,
D., Grady, J., Nie, L., Phillips, T., Davydov, E., Golovin,
D., et al. Ad click prediction: A view from the trenches.
In Proceedings of the 19th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining,
pp. 1222–1230. ACM, 2013.
Mikolov, T., Chen, K., Corrado, G., and Dean, J. Efficient
estimation of word representations in vector space. arXiv
preprint arXiv:1301.3781, 2013.
Shalev-Shwartz, S. and Tewari, A. Stochastic methods for
(cid:96)1-regularized loss minimization. Journal of Machine
Learning Research, 12(6):1865–1892, 2011.
Tan, M., Tsang, I. W., and Wang, L. Towards ultrahigh
dimensional feature selection for big data. Journal of
Machine Learning Research, 15(1):1371–1429, 2014.
Vervier, K., Mah´e, P., Tournoud, M., Veyrieras, J.-B., and
Vert, J.-P. Large-scale machine learning for metage-
nomics sequence classification. Bioinformatics, 32(7):
1023–1032, 2016.
Weinberger, K., Dasgupta, A., Langford, J., Smola, A., and
Attenberg, J. Feature hashing for large scale multitask
learning. In Proceedings of the 26th Annual International
Conference on Machine Learning, pp. 1113–1120. ACM,
2009.
Wood, D. E. and Salzberg, S. L. Kraken: Ultrafast metage-
nomic sequence classification using exact alignments.
Genome Biology, 15(3):1, 2014.
MISSION: Feature Selection via Sketching
with probability at least 1 − 3δ
(cid:13)(cid:13)βt+1 − β
∗(cid:13)(cid:13)2 ≤ 2ρ(cid:13)(cid:13)βt − β
∗(cid:13)(cid:13)2 + 2
(cid:114)
+ 2ε1 + 3ε2,
σ2
w(1 + µ)k log p
n
(6)
provided that n > Ck log p, m > Ck log2 p, and that
Assumption 1 holds.
Notice that since ρ < 0.5, the above theorem guarantees
geometric convergence. This implies that the overall error
is of the order of the additive constants ε1 and ε2.
Before we prove this theorem, we will collect some lemmas
that will help us prove our result.
Lemma 2. Suppose X ∈ Rn×p has i.i.d N (0, 1
n ) en-
tries. Then for constants ρ, δ2 > 0, there exists a constant
C2(δ) > 0 such that if n ≥ C2k log p such that for any pair
of unit-norm k−sparse vectors β1, β2 ∈ Sp−1, the following
holds with probability at least 1 − δ2.
(cid:104)Xβ1, Xβ2(cid:105) − (cid:104)β1, β2(cid:105) ≤ ρ.
(7)
Proof. Note that E[(cid:104)Xβ1, Xβ2(cid:105)] = (cid:104)β1, β2(cid:105). For a fixed
pair of β1, β2, the proof follows from a standard Chernoff
bound argument after observing that (cid:104)Xβ1, Xβ2(cid:105) can be
written as a sum of products of independent Gaussian ran-
dom variables. The rest of the proof follows from a standard
covering argument, which gives the requirement on n.
Lemma 3. Suppose X has i.i.d entries drawn according to
N (0, n−1), and w ∼ N (0, σ2
wIn) is drawn independently
of X. Then, for any constant δ3 > 0, there are constants
C3, µ > 0 such that for all unit norm k−sparse β ∈ Sp−1,
the following holds with probability at least 1 − δ3:
(cid:114)
9. Appenix
In this appendix, we present some preliminary results on the
convergence of MISSION. For the sake of exposition, we
will consider the full-gradient descent version of MISSION,
and we will prove that the iterates converge geometrically
upto a small additive error. In order to establish this proof,
we make an assumption (Assumption 1) about the hashing
scheme; see Section 9.1 for more on this.
We begin by establishing some notation. We will assume
that the data satisfies the following linear model:
∗
+ w,
y = Xβ
(4)
where y ∈ Rn is the vector of observation, X ∈ Rn×p is the
data matrix, w ∈ Rn is the noise vector, and β∗ ∈ Rp is the
unknown k−sparse regression vector. We will let ψ and ϕ
respectively denote the hashing and the (top-k) heavy-hitters
operation. We will let βt denote the output of MISSION in
step t. In general, we will let the vector h ∈ Rm denote
the hash table. Finally, as before, we will let Hk denote the
projection operation onto the set of all k−sparse vectors.
We will make the following assumption about the hashing
mechanism:
Assumption 1. For any h ∈ Rm, there exists an βh ∈ Rp
such that the following hold
1. ψ(βh) = h, that is, the hash table contents can be set
to h by hashing the vector βh.
2. (cid:107)βh − Hk(βh)(cid:107)2 ≤ ε1
This assumption requires the hashing algorithm to be such
that there exists a nearly sparse vector that can reproduce
any state of the hash table exactly. This is reasonable since
the hash table is a near optimal "code" for sparse vectors in
Rp. See Section 9.1 for more on this.
We will next state a straightforward lemma about the sketch-
ing procedure
Lemma 1. There exist constants ε2, C1 > 0 such that
provided that the size m of the hash table satisfies m ≥
C1k log2 p, the following holds for any β ∈ Rp with proba-
bility at least 1 − δ1:
(cid:107)ϕ(ψ(β)) − Hk(β)(cid:107)2 ≤ ε2
(5)
This lemma follows directly from the definition of the Count-
Sketch, and we will not prove here.
We next state the main theorem that we will show.
Theorem 1. For any δ ∈(cid:0)0, 1
(cid:1) and ρ ∈ (0, 0.5), there is
a constant C > 0 such that the following statement holds
3
n
σ2
(8)
w(1 + µ)k log p
(cid:104)β, X T w(cid:105) ≤
provided n ≥ C3k log p.
Proof. Notice that for a fixed β, (cid:104)β, X T w(cid:105) = (cid:104)Xβ, w(cid:105)
n (cid:107)w(cid:107)2 (cid:104)β, w2(cid:105), where w2 ∼
has the same distribution as 1√
N (0, In) is independent of w. Now, we can use concentra-
tion inequalities of chi-squared random variables to show
that there is a constant C(cid:48)
2 ≥ σ2
3 > 0
−C(cid:48)
3n.
(9)
w(1 + µ1)n
P(cid:104)(cid:107)w(cid:107)2
P(cid:104)(cid:104)β, w2(cid:105)2 ≥ 1 + µ2
(cid:105) ≤ e
(cid:105) ≤ e
Similarly, from chi-squared concentration, there is a con-
stant C(cid:48)(cid:48)
3 > 0
−C(cid:48)(cid:48)
3
(10)
MISSION: Feature Selection via Sketching
Now, with a standard covering argument, we know that there
is a constant C(cid:48)(cid:48)(cid:48)
3 k log p,
the following holds for at least 1 − δ3 for any k−sparse β:
3 > 0 such that provided n > C(cid:48)(cid:48)(cid:48)
(cid:104)β, AT w(cid:105) = (cid:104)Aβ, w(cid:105)
(cid:114)
≤
σ2
w(1 + µ)nk log p
n
.
Proof of Theorem 1 If we let ht denote the contents of
the hash table at round t, notice that we have the following:
xt+1 = ϕ(ht+1). The (full gradient descent version of the)
MISSION algorithm proceeds by updating the hash table
with hashes of the gradient updates. Therefore, we have the
following relationship:
ht+1 = ht + ψ(cid:0)ηX T X(β
∗ − βt) + X T w(cid:1) ,
(11)
where βt is the output of the algorithm at round t. Notice
that βt = ϕ(ht). According to Assumption 1, we know that
there exists a vector βt such that ψ( βt) = ht. We will use
this observation next. Notice that the output of round t + 1
maybe written as follows:
βt+1 = ϕ(cid:0)ht + ψ(cid:0)ηX T X(β
(cid:16) βt + ηX T X(β
(cid:16)
= ϕ
∗ − βt) + X T w(cid:1)(cid:1)
(cid:17)(cid:17)
∗ − βt) + X T w
ψ
.
Now, we will estimate how close the output of the algorithm
gets to β∗ in round t + 1 in terms of how close the algorithm
got in round t. Notice that
(cid:13)(cid:13)βt+1 − β
(cid:13)(cid:13)(cid:13)ϕ
(cid:16)
≤(cid:13)(cid:13)(cid:13)Hk
∗(cid:13)(cid:13)2
(cid:16) βt + ηX T X(β
(cid:16) βt + ηX T X(β
=
ψ
∗ − βt) + X T w
∗ − βt) + X T w
∗(cid:13)(cid:13)(cid:13)2
(cid:17)(cid:17) − β
∗(cid:13)(cid:13)(cid:13)2
(cid:17) − β
+ ε2,
(12)
which follows from Lemma 1. We will next consider the
first term from above. For notational ease, we will set
γt+1 (cid:44) βt + ηX T X(β∗ − βt) + X T w. Observe that Hk is
an orthogonal projection operator, and that β∗ is k−sparse,
therefore we have that
(cid:0)γt+1(cid:1) − γt+1(cid:13)(cid:13)2
2 ≤(cid:13)(cid:13)γt+1 − β
∗(cid:13)(cid:13)2
(13)
Adding and subtracting β∗ on the left side and cancelling
out the common terms, we have the following.
2 .
(cid:13)(cid:13)Hk
(cid:13)(cid:13)Hk(γt+1) − β
∗(cid:13)(cid:13)2
2
≤ 2(cid:104)Hk(γt+1) − β
= 2(cid:104)Hk(γt+1) − β
= 2(cid:104)Hk(γt+1) − β
∗
∗
∗
∗(cid:105)
, γt+1 − β
, βt + ηX T X(β
, βt + ηX T X(β
∗ − βt) + X T w − β
∗ − βt) + X T w − β
∗(cid:105)
∗(cid:105)
(a)≤ 2(cid:104)Hk(γt+1) − β
≤ 2(cid:104)Hk(γt+1) − β
+ 2(cid:104)Hk(γt+1) − β
∗
, βt − βt(cid:105)
, βt + ηX T X(β
, βt + ηX T X(β
∗
∗
+ 2
(cid:13)(cid:13)Hk(γt+1) − β
∗(cid:13)(cid:13)2
+ 2(cid:13)(cid:13)Hk(γt+1) − β
+ 2(cid:13)(cid:13)Hk(γt+1) − β
∗
∗ − βt) + X T w − β
∗ − βt) + X T w − β
(cid:13)(cid:13)(cid:13)ϕ(ψ( βt)) − βt(cid:13)(cid:13)(cid:13)2
∗(cid:13)(cid:13)2
(cid:16)(cid:13)(cid:13)(cid:13)Hk( βt) − βt(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)Hk( βt) − ϕ(ψ( βt))
∗(cid:13)(cid:13)2 (ε1 + ε2) ,
∗ − βt) + X T w − β
(14)
(cid:17)
+
, βt + ηX T X(β
∗(cid:105)
∗(cid:105)
∗(cid:105)
(b)≤ 2(cid:104)Hk(γt+1) − β
, βt + ηX T X(β
∗ − βt) + X T w − β
where (a) follows form the Cauchy-Schwarz inequality and
from the definition of βt, (b) follows from Assumption 1
and Lemma 1. We will now turn our attention to the first
inner-product in (14). With some rearrangement of terms,
one can see that
(cid:104)Hk(γt+1) − β
∗
= (cid:104)Hk(γt+1) − β
∗
X(βt − β
∗(cid:105) − η(cid:104)X(cid:0)Hk(γt+1) − β
∗(cid:13)(cid:13)2
(cid:13)(cid:13)βt − β
∗(cid:13)(cid:13)2
(a)≤ ρ(cid:13)(cid:13)Hk(γt+1) − β
(cid:13)(cid:13)βt − β
(cid:13)(cid:13)Hk(γt+1) − β
∗(cid:13)(cid:13)2
∗(cid:13)(cid:13)2
(cid:114)
+(cid:13)(cid:13)Hk(γt+1) − β
∗(cid:13)(cid:13)2
, βt − β
)(cid:105) + η(cid:104)Hk(γt+1) − β
∗
+ (cid:104)Hk(γt+1) − β
w(1 + µ)k log p
∗(cid:1) ,
, X T w(cid:105)
, X T w(cid:105)
(b)≤ ρ
∗(cid:105)
σ2
∗
∗
n
(15)
where (a) follows from Lemma 2 and setting η = 1. (b)
follows from Lemma 3.
Putting (14) and (15), we get
(cid:0)γt+1(cid:1) − γt(cid:13)(cid:13)2 ≤ 2ρ(cid:13)(cid:13)βt − β
∗(cid:13)(cid:13)2
(cid:13)(cid:13)Hk
(cid:114)
+ 2
σ2
w(1 + µ)k log p
n
+ 2 (ε1 + ε2) .
(16)
Putting this together with (12) gives us the desired result.
9.1. On Assumption 1
In the full-gradient version of the MISSION algorithm, one
might modify the algorithm explicitly to ensure that As-
sumption 1. Towards this end, one would simply ensure that
the gradients vector is attenuated on all but its top k entries
at each step.
MISSION: Feature Selection via Sketching
It is not hard to see that this clean-up step will ensure that
Assumption 1 holds and the rest of the proof simply goes
through. In MISSION as presented in the manuscript, we
employ stochastic gradient descent (SGD). While the above
proof needs to be modified for it to be applicable to this
case, our simulations suggest that this clean-up step is un-
necessary here. We suspect that this is due to random can-
cellations that are introduced by the SGD. This is indeed an
exciting avenue for future work.
|
1609.07650 | 3 | 1609 | 2016-12-18T13:05:20 | Popularity in the generalized Hospital Residents Setting | [
"cs.DS"
] | We consider the problem of computing popular matchings in a bipartite graph G = (R U H, E) where R and H denote a set of residents and a set of hospitals respectively. Each hospital h has a positive capacity denoting the number of residents that can be matched to h. The residents and the hospitals specify strict preferences over each other. This is the well-studied Hospital Residents (HR) problem which is a generalization of the Stable Marriage (SM) problem. The goal is to assign residents to hospitals optimally while respecting the capacities of the hospitals. Stability is a well-accepted notion of optimality in such problems. However, motivated by the need for larger cardinality matchings, alternative notions of optimality like popularity have been investigated in the SM setting. In this paper, we consider a generalized HR setting -- namely the Laminar Classified Stable Matchings (LCSM+) problem. Here, additionally, hospitals can specify classifications over residents in their preference lists and classes have upper quotas. We show the following new results: We define a notion of popularity and give a structural characterization of popular matchings for the LCSM+ problem. Assume n = |R| + |H| and m = |E|. We give an O(mn) time algorithm for computing a maximum cardinality popular matching in an LCSM+ instance. We give an O(mn^2) time algorithm for computing a matching that is popular amongst the maximum cardinality matchings in an LCSM+ instance. | cs.DS | cs |
Popularity in the generalized Hospital Residents setting
Meghana Nasre and Amit Rawat
Indian Institute of Technology Madras, India
Abstract. We consider the problem of computing popular matchings in a bipartite graph G = (R ∪
H, E) where R and H denote a set of residents and a set of hospitals respectively. Each hospital h has
a positive capacity denoting the number of residents that can be matched to h. The residents and the
hospitals specify strict preferences over each other. This is the well-studied Hospital Residents (HR)
problem which is a generalization of the Stable Marriage (SM) problem. The goal is to assign residents to
hospitals optimally while respecting the capacities of the hospitals. Stability is a well-accepted notion
of optimality in such problems. However, motivated by the need for larger cardinality matchings,
alternative notions of optimality like popularity have been investigated in the SM setting. In this paper,
we consider a generalized HR setting -- namely the Laminar Classified Stable Matchings (LCSM+)
problem. Here, additionally, hospitals can specify classifications over residents in their preference lists
and classes have upper quotas. We show the following new results: We define a notion of popularity
and give a structural characterization of popular matchings for the LCSM+ problem. Assume n =
R+H and m = E. We give an O(mn) time algorithm for computing a maximum cardinality popular
matching in an LCSM+ instance. We give an O(mn2) time algorithm for computing a matching that
is popular amongst the maximum cardinality matchings in an LCSM+ instance.
1
Introduction
Consider an academic institution where students credit an elective course from a set of available courses.
Every student and every course rank a subset of elements from the other set in a strict order of preference.
Each course has a quota denoting the maximum number of students it can accommodate. The goal is to
allocate to every student at most one course respecting the preferences. This is the well-studied Hospital
Residents problem [7]. We consider its generalization where, in addition, a course can classify students -- for
example, the students may be classified as under-graduates and post-graduates and department-wise and so
on. Depending on the classifications, a student may belong to multiple classes. Apart from the total quota,
each course now has a quota for every class. An allocation, in this setting, has to additionally respect the
class quotas. This is the Classified Stable Matching problem introduced by Huang [10].
Stability is a de-facto notion of optimality in settings where both set of participants have preferences.
Informally, an allocation of students to courses is stable if no unallocated student-course pair has incentive
to deviate from the allocation. Stability is appealing for several reasons -- stable allocations are guaranteed
to exist, they are efficiently computable and all stable allocations leave the same set of students unallo-
cated [9]. However, it is known [13] that the cardinality of a stable allocation can be half the size of the
largest sized allocation possible. Furthermore, in applications like student-course allocation, leaving a large
number of students unallocated is undesirable. Thus, it is interesting to consider notions of optimality which
respect preferences but possibly compromise stability in the favor of cardinality. Kavitha and Huang [11,13]
investigated this in the Stable Marriage (SM) setting where they considered popularity as an alternative to
stability. At a high level, an allocation of students to courses is popular if no majority wishes to deviate from
the allocation. Here, we consider popularity in the context of two-sided preferences and one-sided capacities
with classifications.
We formally define our problem now -- we use the familiar hospital residents notation. Let G = (R∪H, E)
be a bipartite graph where R ∪ H = n and E = m. Here R denotes the set of residents, H denotes the set
of hospitals and every hospital h ∈ H has an upper quota q(h) denoting the maximum number of residents
h can occupy. A pair (r, h) ∈ E denotes that r and h are mutually acceptable to each other. Each resident
(resp. hospital) has a strict ordering of a subset of the hospitals (resp. residents) that are acceptable to him
or her (resp. it). This ordering is called the preference list of a vertex. An assignment (or a matching) M in
G is a subset of E such that every resident is assigned to at most one hospital and a hospital h is assigned
at most q(h) residents. Let M (r) (resp. M (h)) denote the hospital (resp. the set of residents) which are
assigned to r (resp. h) in M . A hospital h is under-subscribed if M (h) < q(h). A matching M is stable if no
unassigned pair (r, h) wishes to deviate from M . The goal is to compute a stable matching in G. We denote
it by HR+ throughout the paper 1. The celebrated deferred acceptance algorithm by Gale and Shapley [7]
proves that every instance of the HR+ problem admits a stable matching.
k ) denoting the maximum number of residents that can be matched to h in C h
A generalization of the HR+ problem is the Laminar Classified Stable Matching (LCSM) problem in-
troduced by Huang [10]. An instance of the LCSM+ problem is an instance of the HR+ problem where
additionally, each hospital h is allowed to specify a classification over the set of residents in its preference
list. A class C h
k of a hospital h is a subset of residents in its preference list and has an associated upper
quota q(C h
k . (In the LCSM
problem [10], classes can have lower quotas as well.) We assume that the classes of a hospital form a laminar
set. That is, for any two classes C h
k = ∅), or one is con-
tained inside the other (C h
j ). Huang suitably modified the classical definition of stability
to account for the presence of these classifications. He showed that every instance of the LCSM+ problem
admits a stable matching which can be computed in O(mn) time [10]. A restriction of the LCSM+ problem,
denoted by Partition Classified Stable Matching (PCSM+), is where the classes of every hospital partition
the residents in its preference list.
k , either the two classes are disjoint (C h
j and C h
j ⊂ C h
k or C h
k ⊂ C h
j ∩ C h
Motivated by the need to output larger cardinality matchings, we consider computing popular matchings
in the LCSM+ problem. The notion of popularity uses votes to compare two matchings. Before we can define
voting in the LCSM+ setting, it is useful to discuss voting in the context of the SM problem.
Voting in the SM setting: Let G = (R ∪ H, E) be an instance of the SM problem and let M and M ′ be
any two matchings in G. A vertex u ∈ R ∪ H (where each hospital h has q(h) = 1) prefers M over M ′ and
therefore votes for M over M ′ if either (i) u is matched in M and unmatched in M ′ or (ii) u is matched
in both M and M ′ and prefers M (u) over M ′(u). A matching M is more popular than M ′ if the number
of votes that M gets as compared to M ′ is greater than the number of votes that M ′ gets as compared to
M . A matching M is popular if there does not exist any matching that is more popular than M . In the
SM setting it is known that a stable matching is popular, however it was shown to be minimum cardinality
popular matching [11]. Huang and Kavitha [11,13] gave efficient algorithms for computing a max-cardinality
popular matching and a popular matching amongst max-cardinality matchings in an SM instance.
Voting in the capacitated setting: To extend voting in the capacitated setting, we assign a hospital h
as many votes as its upper quota q(h). This models the scenario in which hospitals with larger capacity
get a larger share of votes. For the HR+ problem, a hospital h compares the most preferred resident in
M (h) \ M ′(h) to the most preferred resident in M ′(h) \ M (h) (and votes for M or M ′ as far as those two
residents are concerned) and so on. For this voting scheme, we can obtain analogous results for computing
popular matchings in the HR+ problem via the standard technique of cloning (that is, creating q(h) copies
of a hospital h and appropriately modifying preference lists of the residents and hospitals 2). However,
our interest is in the LCSM+ problem, for which we are not aware of any reduction to the SM problem.
Furthermore, we show that the straightforward voting scheme as defined in the HR+ does not suffice for
the LCSM+ problem. Therefore, we define a voting scheme for a hospital which takes into consideration the
classifications as well as ensures that every stable matching in the LCSM+ instance is popular. We show the
following results:
-- We define a notion of popularity for the LCSM+ problem. Since our definition ensures that stable
matchings are popular -- this guarantees the existence of popular matchings in the LCSM+ problem.
-- We give a characterization of popular matchings for the LCSM+ problem, which is a natural extension
of the characterization of popular matchings in SM setting [11].
1
2
We use HR+ instead of HR for consistency with other problems discussed in the paper.
For every hospital in the cloned graph, its preference list is the same as in the original instance. For every hospital h, fix an ordering
of its clones. The preference list of a resident r in the cloned instance is obtained by replacing the occurrence of h by the fixed
ordering of its clones. We refer the reader to [4,14] for details.
2
-- We obtain the following algorithmic results. An O(m + n) (resp. O(mn)) time algorithm for computing a
maximum cardinality popular matching in a PCSM+ (resp. LCSM+) instance. An O(mn) (resp. O(mn2))
time algorithm for computing a popular matching amongst maximum cardinality matchings in a PCSM+
(resp. LCSM+) instance.
Very recently, independent of our work, two different groups [4,12] have considered popular matchings in
the one-to-many setting. Brandl and Kavitha [4] have considered computing popular matchings in the HR+
problem. In their work as well as ours, a hospital h is assigned as many votes as its capacity to compare
two matchings M and M ′. In contrast, by the definition of popularity in [4], a hospital h chooses the most
adversarial ordering of residents in M (h) \ M ′(h) and M ′(h) \ M (h) for comparing M and M ′. However, it
is interesting to note that in an HR+ instance the same matching is output by both our algorithms. On the
other hand, we remark that the model considered in our paper is a more general one than the one considered
in [4]. Kamiyama [12] has generalized our work and the results in [4] using a matroid based approach.
We finally remark that one can consider voting schemes where a hospital is given a single vote instead
of capacity many votes. In one such scheme, a hospital compares the set of residents in M (h) and M ′(h) in
lexicographic order and votes accordingly. However, when such a voting is used, it is possible to construct
instances where a stable matching is not popular. The techniques in this paper use the fact that stable
matchings are popular, therefore it is unclear if our techniques will apply for such voting schemes.
Related Work: The notion of popularity was introduced by Gardenfors [8] in the context of stable match-
ings. In [1] Abraham et al. studied popularity in the one-sided preference list model. As mentioned earlier,
our work is inspired by a series of papers where popularity is considered as an alternative to stability in the
stable marriage setting by Huang, Kavitha and Cseh [5,11,13]. Bir´o et al. [3] give several practical scenarios
where stability may be compromised in the favor of size. The PCSM+ problem is a special case of the Stu-
dent Project Allocation (SPA) problem studied by Abraham et al. [2]. They gave a linear time algorithm
to compute a stable matching in an instance of the SPA problem. In this paper, we use the algorithms of
Abraham et al. [2] and Huang [10] for computing stable matchings in the PCSM+ and LCSM+ problems.
Both these algorithms follow the standard deferred acceptance algorithm of Gale and Shapley with problem
specific modifications. We refer the reader to [2] and [10] for details.
Organization: In Section 2 we define the notion of popularity, in Section 3 we present the structural charac-
terization of popular matchings. In Section 4 we describe our algorithms to compute a maximum cardinality
popular matching, and a popular matching amongst maximum cardinality matchings. We conclude with a
short discussion about popular matchings in the LCSM problem.
2 Stability and popularity in the LCSM+ problem
Consider an instance G = (R ∪ H, E) of the LCSM+ problem. As done in [10], assume that for every h ∈ H
there is a class C h
∗ ) = q(h). For a hospital h,
let T (h) denote the tree of classes corresponding to h where C h
∗ is the root of T (h). The leaf classes in T (h)
denote the most refined classifications for a resident whereas as we move up in the tree from a leaf node to
the root, the classifications gets coarser.
∗ containing all the residents in the preference list of h and q(C h
To define stable matchings in the LCSM problem, Huang introduced the notion of a blocking group w.r.t. a
matching. Later, Fleiner and Kamiyama [6] defined a blocking pair which is equivalent to a blocking group of
Huang. We use the definition of stability from [6] which we recall below. A set S = {r1, . . . , rl} is feasible for
a hospital h if S ≤ q(h) and for every class C h
j ).
A matching M in G is feasible if every resident is matched to at most one hospital, and M (h) is feasible for
every hospital h ∈ H. A pair (r, h) /∈ M blocks M iff both the conditions below hold:
j of h (including the root class C h
∗ ), we have C h
j ∩ S ≤ q(C h
-- r is unmatched in M , or r prefers h over M (r), and
-- either the set M (h) ∪ {r} is feasible for h, or there exists a resident r′ ∈ M (h), such that h prefers r over
r′, and (M (h) \ {r′}) ∪ {r} is feasible for h.
A feasible matching M in G is stable if M does not admit any blocking pair.
3
2.1 Popularity
To define popularity, we need to specify how a hospital compares two sets M (h) and M ′(h) in an LCSM+
setting, where M and M ′ are two feasible matchings in the instance.
Illustrative example Consider the following LCSM+ instance where R = {r1, . . . , r4} and H = {h1, . . . , h3}
and the preference lists of the residents and hospitals are as given in Figure 1(a) and (b) respectively. The
preferences can be read as follows: resident r1 has h1 as his top choice hospital. Resident r2 has h2 as
its top choice hospital followed by h1 which is his second choice hospital and so on. For h ∈ {h2, h3} we
have q(h) = 1 and both these hospitals have a single class C h
∗ containing all the residents in the prefer-
ence list of h and q(C h
∗ ) = q(h). For hospital h1 we have q(h1) = 2 and the classes provided by h1 are
C h1
2 ) = 1 and
∗ ) = 2. We remark that the example in Figure 1 is also a PCSM+ instance. Figure 1(c) shows the tree
q(C h1
T (h1).
∗ = {r1, r2, r3, r4} with quotas as follows: q(C h1
1 ) = q(C h1
1 = {r1, r2}, C h1
2 = {r3, r4}, C h1
r1 : h1
r2 : h2, h1, h3
r3 : h1, h2
r4 : h1
(a)
h1 : r2, r3, r4, r1
h2 : r3, r2
h3 : r2
r1, r2, r3, r4
C h1
∗
r1, r2
C h1
1
r3, r4
C h1
2
(b)
(c)
Fig. 1. (a) Resident preferences, (b) Hospital preferences, (c) T (h1). The matchings M = {(r1, h1), (r2, h2), (r3, h1)},
M ′ = {(r2, h1), (r3, h2), (r4, h1)}, and M ′′ = {(r1, h1), (r2, h3), (r3, h2), (r4, h1)} are all feasible in the instance.
Consider the two feasible matchings M and M ′ defined in Fig. 1. Note that M is stable in the instance
whereas the edge (r3, h1) blocks M ′. While comparing M and M ′, the vote for every vertex u in the instance
except h1 is clear -- u compares M (u) with M ′(u) and votes accordingly. In order for h1 to vote between
M and M ′, the hospital compares between M (h1) = {r1, r3} and M ′(h1) = {r2, r4}. A straightforward way
is to compare r3 with r2 (the most preferred resident in M (h1) to the most preferred resident in M ′(h1))
and then compare r1 with r4 (second most preferred resident in M (h1) to second most preferred resident
in M ′(h1)). Thus, both the votes of h1 are in favor of M ′ when compared with M . Such a comparison has
two issues -- (i) it ignores the classifications given by h1, and (ii) the number of votes that M ′ gets when
compared with M is more than the number of votes that M gets as compared to M ′. Therefore M ′ is more
popular than M which implies that M (a stable matching) is not popular.
We propose a comparison scheme for hospitals which addresses both the issues. In the above example,
we note that r1 ∈ M (h) has a corresponding resident r2 ∈ M ′(h) to be compared to in one of the most
refined classes C h1
(see Figure 1(c)). Thus, we compare r1 with r2. The resident r3 ∈ M (h) is compared to
1
r4 ∈ M (h) another leaf class C h1
2 . According to this comparison, h1 is indifferent between M and M ′ and
M ′ is no longer more popular than M . Note that, although in the example, both the comparisons happen
in a leaf class, this may not be the case in a general instance. Finally, we note that the matching M ′′ is a
popular matching in the instance and is strictly larger in size than the stable matching M .
We formalize the above observations in the rest of the section. To take into account the classifications, for
a hospital h and the matchings M and M ′, we set up a correspondence between residents in M (h) \ M ′(h)
and the residents in M ′(h) \ M (h). That is, we define:
corr : M (h) ⊕ M ′(h) → M (h) ⊕ M ′(h) ∪ {⊥}
For a resident r ∈ M (h) ⊕ M ′(h) we denote by corr(r) the corresponding resident to which r gets compared
when the hospital h casts its votes. We let corr(r) = ⊥ if r does not have a corresponding resident to be
compared to from the other matching. The pseudo-code for the algorithm to compute the corr function is
given below.
4
Algorithm 1 Correspondence between residents of M (h) and M ′(h)
1: procedure Find-Correspondence(h, M, M ′)
2:
3:
4:
5:
6:
7:
8:
let T (h) be the classification tree associated with h
set corr(r) = ⊥ for each r ∈ M (h) ⊕ M ′(h)
Y = M (h) \ M ′(h); Y ′ = M ′(h) \ M (h)
while Y 6= ∅ and Y ′ 6= ∅ do
Xj = C h
j = C h
X ′
Let C h
f be one of the most refined classes for which Xf 6= ∅ and X ′
for k = 1, . . . , min(Xf , X ′
for each class C h
j ∩ Y
j ∩ Y ′
j in T (h) do
f ) do
f 6= ∅.
let r be the k-th most preferred resident in Xf
let r′ be the k-th most preferred resident in X ′
f
set corr(r) = r′, and corr(r′) = r
Y = Y \ {r}; Y ′ = Y ′ \ {r′}
9:
10:
11:
12:
13:
14:
The algorithm begins by setting corr for every r ∈ M (h) ⊕ M ′(h) to ⊥. The algorithm maintains two
sets of residents Y = M (h) \ M ′(h) and Y ′ = M ′(h) \ M (h) for whom corr needs to be set. As long as
the sets Y and Y ′ are both non-empty, the algorithm repeatedly computes for every class C h
j (including
j ∩ Y ′. The algorithm then chooses one of the most
the root class C h
f are both non-empty. Finally, residents in Xf and X ′
refined classes, say C h
f
are sorted according to the preference ordering of h and the corr of the k-th most preferred resident in Xf
is set to the k-th most preferred resident in X ′
f in T (h), for whom Xf and X ′
∗ ) the sets Xj = C h
f , where k = 1, . . . , min{Xf , X ′
j ∩ Y and X ′
j = C h
f }.
For r ∈ R, and any feasible matching M in G, if r is unmatched in M then, M (r) = ⊥. A vertex prefers
any of its neighbours over ⊥. For a vertex u ∈ R ∪ H, let x, y ∈ N (u) ∪ {⊥}, where N (u) denotes the
neighbours of u in G.
voteu(x, y) = +1
= −1
= 0
if u prefers x over y
if u prefers y over x
if x = y
Using the above notation, the vote of a resident is easy to define -- a resident r prefers M ′ over M iff the
term Vr > 0, where Vr = voter(M ′(r), M (r)).
Recall that a hospital h uses q(h) votes to compare M and M ′. Let q1(h) = M (h) ∩ M ′(h) (number of
common residents assigned to h in M and M ′) and q2(h) = q(h) − max{M (h), M ′(h)} (number of unfilled
positions of h in both M and M ′). Our voting scheme ensures that q1(h) + q2(h) votes of h remain unused
when comparing M and M ′. A hospital h prefers M ′ over M iff the term Vh > 0, where Vh is defined as
follows:
Vh = (M ′(h) − M (h)) + X
r∈M ′(h)\M(h)
voteh(r, corr(r))
&&
corr(r)6=⊥
The first term in the definition of Vh counts the votes of h w.r.t. the residents from either M or M ′ that
did not find correspondence. The second term counts the votes of h w.r.t. the residents each of which has
a corresponding resident from the other matching. We note that in the SM setting, corr(r) will simply be
M (h). Thus, our definition of votes in the presence of capacities is a natural generalization of the voting
scheme in the SM problem.
Let us define the term ∆(M ′, M ) as the difference between the votes that M ′ gets over M and the votes
that M gets over M ′.
∆(M ′, M ) = X
r∈R
Vr + X
h∈H
Vh
5
Definition 1. A matching M is popular in G iff for every feasible matching M ′, we have ∆(M ′, M ) ≤ 0.
2.2 Decomposing M ⊕ M ′
Here, we present a simple algorithm which allows us to decompose edges of components of M ⊕ M ′ in an
instance into alternating paths and cycles. Consider the graph G = (R ∪ H, M ⊕ M ′), for any two feasible
matchings M and M ′ in G. We note that the degree of every resident in G is at most 2 and the degree of
every hospital in G is at most 2 · q(h). Consider any connected component C of G and let e be any edge in
C. We observe that it is possible to construct a unique maximal M alternating path or cycle ρ containing e
using the following simple procedure. Initially ρ contains only the edge e.
1. Let r ∈ R be one of the end points of the path ρ, and assume that (r, M (r)) ∈ ρ. We grow ρ by adding
the edge (r, M ′(r)). Similarly if an edge from M ′ is incident on r in ρ, we grow the path by adding the
edge (r, M (r)) if it exists.
2. Let h ∈ H be one of the end points of the path ρ, and assume that (r, h) ∈ M \ M ′ belongs to ρ. We
extend ρ by adding (corr(r), h) if corr(r) is not equal to ⊥. A similar step is performed if the last edge
on ρ is (r, h) ∈ M ′ \ M .
3. We stop the procedure when we complete a cycle (ensuring that the two adjacent residents of a hospital
are corr for each other according to the hospital), or the path can no longer be extended. Otherwise we
go to Step 1 or Step 2 as applicable and repeat.
The above procedure gives us a unique decomposition of a connected component in G into alternating paths
and cycles. Note that a hospital may appear multiple times in a single path or a cycle and also can belong
to more than one alternating paths and cycles. Figure 2 gives an example of the decomposition of the two
feasible matchings in the instance in Figure 1.
h1
h2
r1
r2
r3
r4
r1
r2
r3
r4
h1
h2
h1
(a)
(b)
Fig. 2. M and M ′ are feasible matchings in the example as defined in Fig. 1. (a) G = (R ∪ H, M ⊕ M ′); bold edges
belong to M , dashed edges belong to M ′. (b) shows the decomposition of the edges of the component of G into a
single path.
Let YM⊕M ′ denote the collection of alternating paths and alternating cycles obtained by decomposing
every component of G.
We now state a useful property about any alternating path or cycle in YM⊕M ′ .
Lemma 1. If ρ is an alternating path or an alternating cycle in YM⊕M ′ , then M ⊕ ρ is a feasible matching
in G.
Proof. Let hr′, h, ri be any sub-path of ρ, where r′ = corr(r), and (r, h) ∈ M . We prove that (M (h) \ {r}) ∪
{r′} is feasible for h. Let C h
j ) be the unique leaf class of T (h) containing r (resp. r′). See Figure 3.
We consider the following two cases:
i (resp. C h
6
C h
k
C h
∗
C h
t
.., r′, ..
C h
j
.., r, ..
C h
i
Fig. 3. The classification tree T(h) for a hospital h.
-- r and r′ belong to the same leaf class in T (h), i.e. C h
(M (h) \ {r}) ∪ {r′} is feasible for h.
-- r and r′ belong to different leaf classes of T (h), i.e. C h
i
i = C h
j . In this case, it is easy to note that
6= C h
violate the upper quota only for those classes of T (h) which contain r′ but do not contain r. Let C h
the least common ancestor of C h
path from C h
and r /∈ C h
that r′ did not find a corresponding resident in the set (M (h) \ M ′(h)) ∩ C h
because M ′ is feasible. Thus, (M (h) ∩ C h
(M (h) \ {r}) ∪ {r′} is feasible for h.
j . Observe that (M (h) \ {r}) ∪ {r′} can
k be
t which lies in the
t ). As r′ = corr(r)
t ). The first inequality is due to the fact
t . The second inequality is
t . Therefore
t ) ∪ {r′} does not violate the upper quota for C h
i and C h
j excluding the class C h
j in T (h). It suffices to look at any class C h
t ) ∪ {r′} ≤ q(C h
k and show that (M (h) ∩ C h
k to C h
t , we claim that M (h) ∩ C h
t < M ′(h) ∩ C h
t ≤ q(C h
We note that the hospital h may occur multiple times on ρ. Let M (h)ρ denote the set of residents matched
to h restricted to ρ. To complete the proof of the Lemma, we need to prove that (M (h) \ M (h)ρ) ∪ M ′(h)ρ
⊓⊔
is feasible for h. The arguments for this follow from the arguments given above.
As was done in [13], it is convenient to label the edges of M ′ \ M and use these labels to compute ∆(M ′, M ).
Let (r, h) ∈ M ′ \ M ; the label on (r, h) is a tuple:
(voter(h, M (r)), voteh(r, corr(r)))
Note that since we are labeling edges of M ′ \ M , both entries of the tuple come from the set {−1, 1}. With
these definitions in place, we are ready to give the structural characterization of popular matchings in an
LCSM+ instance.
3 Structural characterization of popular matchings
Let G = (R ∪ H, E) be an LCSM+ instance and let M and M ′ be two feasible matchings in G. Using
the corr function, we obtain a correspondence of residents in M (h) ⊕ M ′(h) for every hospital h in G. Let
G = (R ∪ H, M ⊕ M ′) and let YM⊕M ′ denote the collection of alternating paths and cycles obtained by
decomposing every component of G. Finally, we label the edges of M ′ \ M using appropriate votes. The goal
of these steps is to is to rewrite the term ∆(M ′, M ) as a sum of labels on edges.
We note that the only vertices for whom their vote does not get captured on the edges of M ′ \ M are
vertices that are matched in M but not matched in M ′. Let U denote the multi-set of vertices that are end
points of paths in YM⊕M ′ such that there is no M ′ edge incident on them. Note that the same hospital can
belong to multiple alternating paths and cycles in YM⊕M ′ , therefore we need a multi-set. All vertices in U
7
prefer M over M ′ and hence we add a −1 while capturing their vote in ∆(M ′, M ). We can write ∆(M ′, M )
as:
∆(M ′, M ) = X
x∈U
−1 + X
ρ∈YM ⊕M ′
X
(r,h)∈(M ′∩ρ)
{voter(h, M (r)) + voteh(r, corr(r))}
We now delete the edges labeled (−1, −1) from all paths and cycles ρ in YM⊕M ′ . This simply breaks paths
and cycles into one or more paths. Let this new collection of paths and cycles be denoted by YM⊕M ′ . Let
U denote the multi-set of vertices that are end points of paths in YM⊕M ′ such that there is no M ′ edge
incident on them. We rewrite ∆(M ′, M ) as:
∆(M ′, M ) = X
x∈ U
−1 + X
ρ∈ YM ⊕M ′
X
(r,h)∈(M ′∩ρ)
{voter(h, M (r)) + voteh(r, corr(r))}
Theorem below characterizes a popular matching.
Theorem 1. A feasible matching M in G is popular iff for any feasible matching M ′ in G, the set YM⊕M ′
does not contain any of the following:
1. An alternating cycle with a (1, 1) edge,
2. An alternating path which has a (1, 1) edge and starts with an unmatched resident in M or a hospital
which is under-subscribed in M .
3. An alternating path which has both its ends matched in M and has two or more (1, 1) edges.
Proof. We show that if M is a feasible matching such that for any M ′ the set YM⊕M ′ does not contain (1),
(2), (3) as in Theorem 1, then M is popular in G.
Assume for the sake of contradiction that M satisfies the conditions of Theorem 1, and yet M is not
popular. Therefore there exists a feasible matching M ∗ such that ∆(M ∗, M ) > 0. Consider the set YM ∗⊕M .
Recall that this set is a collection of paths and cycles and the edges of M ∗ \ M are labeled. Let ρ be any path
or cycle in YM ∗⊕M and let ∆(M ∗, M )ρ denote the difference between the votes of M ∗ and M when restricted
to the residents and hospitals in ρ. Since ∆(M ∗, M ) > 0, there exists a ρ such that ∆(M ∗, M )ρ > 0. Note
that ρ is present in YM ∗⊕M ; using the presence of ρ we establish the existence of a ρ′ ∈ YM ∗⊕M of the form
(1), (2) or (3) which contradicts our assumption. We consider three cases depending on the structure of ρ.
1. ρ is an alternating cycle or ρ is an alternating path which starts and ends in an M edge:
Since ρ ∈ YM ∗⊕M , and ∆(M ∗, M )ρ > 0, it implies that there are more edges in ρ labeled (1, 1) than the
number of edges labeled (−1, −1). We now delete the edges labeled (−1, −1) from ρ; this breaks ρ in to
multiple alternating paths. Note that each of these paths (say ρ′) start and end with an M edge and are
also present in YM ∗⊕M . Furthermore, since ρ contained more number of edges labeled (1, 1) than the
number of edges labeled (−1, −1), it is clear that there exists at least one ρ′ which has two edges labeled
(1, 1). This is a path of type (3) from the theorem statement and therefore contradicts our assumption
that M satisfied the conditions of the theorem.
2. ρ is an alternating path which starts or ends in an M ∗ edge:
The proof is similar to the previous case except that when we delete from ρ the edges labeled (−1, −1)
we get paths ρ′ ∈ YM ∗⊕M which are paths of type (2) or type (3) from the theorem statement. This
contradicts the assumption that M satisfied the conditions of the theorem.
This completes the proof of one direction of the Theorem. To prove the other direction, we prove the
contrapositive of the statement. That is, if for any feasible matching M ′, YM⊕M ′ contains (1), (2) or (3),
then M is not popular in G. We first assume that ρ ∈ YM⊕M ′ satisfying (1), (2), or (3) is also present in
YM⊕M ′ . Under this condition, it is possible to get a more popular matching than M by the following three
cases.
8
-- Let M2 = M ⊕ ρ be a matching in G; by Lemma 1 we know that M2 is feasible in G. Comparing M2 to
M yields two more votes for M2. Hence, M2 is more popular than M .
-- If ρ is an alternating path in YM⊕M ′ , which has both its endpoints matched in M , and contains more
than one edge labeled (1, 1). Then similar to the case above M2 = M ⊕ ρ is more popular than M .
-- If ρ is an alternating path in YM⊕M ′ , which has exactly one of its endpoints matched in M , and contains
an edge labeled (1, 1), then again M2 = M ⊕ ρ is more popular than M .
Now let us assume that ρ ∈ YM⊕M ′ is not present in YM⊕M ′ . In such a case, ρ is contained in a larger path
or a cycle ρ′ ∈ YM⊕M ′ obtained by combining ρ with other paths in YM⊕M ′ and adding the deleted (−1, −1)
edges. Using the larger path or cycle ρ′ we can construct a matching that is more popular than M . Note
that we need to use paths or cycles in YM⊕M ′ to obtain another matching, since we have to ensure that the
⊓⊔
matching obtained is indeed feasible in the instance and the correspondences are maintained.
We now prove that every stable matching in an LCSM+ instance is popular.
Theorem 2. Every stable matching in an LCSM+ instance G is popular.
Proof. Let M be a stable matching in G. For any feasible matching M ′ in G consider the set YM⊕M ′ . To
prove that M is stable, it suffices to show that there does not exist a path or cycle ρ ∈ YM⊕M ′ such that
an edge of ρ is labeled (1, 1). For the sake of contradiction, assume that ρ is such a path or cycle, which has
an edge (r′, h) ∈ M ′ \ M labeled (1, 1). Let r = corr(r′), where (r, h) ∈ M ∩ ρ. From the proof of Lemma 1
we observe that (M (h) \ {r}) ∪ {r′} is feasible for h, therefore the edge (r′, h) blocks M contradicting the
⊓⊔
stability of M .
4 Popular matchings in LCSM+ problem
In this section we present efficient algorithms for computing (i) a maximum cardinality popular matching, and
(ii) a matching that is popular amongst all the maximum cardinality matchings in a given LCSM+ instance.
Our algorithms are inspired by the reductions of Kavitha and Cseh [5] where they work with a stable marriage
instance. We describe a general reduction from an LCSM+ instance G to another LCSM+ instance Gs. Here
s = 2, . . . , R. The algorithms for the two problems are obtained by choosing an appropriate value of s.
The graph Gs: Let G = (R ∪ H, E) be the input LCSM+ instance. The graph Gs = (Rs ∪ Hs, Es) is
constructed as follows: Corresponding to every resident r ∈ R, we have s copies of r, call them r0, . . . , rs−1
in Rs. The hospitals in H and their capacities remain unchanged; however we have additional dummy
hospitals each of capacity 1. Corresponding to every resident r ∈ R, we have (s − 1) dummy hospitals
d0
r, . . . , ds−2
in Hs. Thus,
r
Rs = { r0, . . . , rs−1 ∀r ∈ R}; Hs = H ∪ { d0
r, . . . , ds−2
r
∀r ∈ R}
We use the term level-i resident for a resident ri ∈ Rs for 0 ≤ i ≤ s − 1. The preference lists corresponding
to s different residents of r in Gs are:
-- For a level-0 resident r0, its preference list in Gs is the preference list of r in G, followed by the dummy
hospital d0
r.
of r in G, followed by di
r.
-- For a level-i resident ri, where 1 ≤ i ≤ s − 2, its preference list in Gs is di−1
followed by preference list
r
-- For a level-(s − 1) resident rs−1, its preference list in Gs is the dummy hospital ds−2
followed by the
r
preference list of r in G.
The preference lists of hospitals in Gs are as follows.
-- The preference list for a dummy hospital di
-- For h ∈ H, its preference list in Gs, has level-(s − 1) residents followed by level-(s − 2) residents, so on
r is ri followed by ri+1.
upto the level-0 residents in the same order as in h's preference list in G.
9
Finally, we need to specify the classifications of the hospitals in Gs. For every class C h
i
we have a corresponding class ¯C h
¯C h
(I1) Each di
(I2) The above invariant implies that for every r ∈ R at most one of {r0, . . . , rs−1} is assigned to a non-dummy
i . Let Ms be a stable matching in Gs. Then Ms satisfies the following properties:
r ∈ Hs for 0 ≤ i ≤ s − 2, is matched to one of {ri, ri+1} in Ms.
{r0, . . . , rs−1} in Gs, such that q( ¯C h
in the instance G,
i ). We note that
i = Sr∈C h
i = s · C h
i ) = q(C h
i
hospital in Ms.
(I3) For a resident r ∈ R, if ri is matched to a non-dummy hospital in Ms, then for all 0 ≤ j ≤ i − 1,
. This also implies that in Ms all
Ms(rj ) = dj
residents r0, . . . , rs−2 are matched and only rs−1 can be left unmatched in Ms.
r. Furthermore, for all i + 1 ≤ p ≤ s − 1, Ms(rp) = dp−1
r
These invariants allow us to naturally map the stable matching Ms to a feasible matching M in G. We
define a function map(Ms) as follows.
M = map(Ms) = {(r, h) : h ∈ H and (ri, h) ∈ Ms for exactly one of 0 ≤ i ≤ s − 1}
We outline an algorithm that computes a feasible matching in an LCSM+ instance G. Given G and s,
construct the graph Gs from G. Compute a stable matching Ms in Gs. If G is an LCSM+ instance we
use the algorithm of Huang [10] to compute a stable matching in G. If G is a PCSM+ instance, it is easy
to observe that Gs is also a PCSM+ instance. In that case, we use the algorithm of Abraham et al.[2] to
compute a stable matching. (The SPA instance is different from a PCSM+ instance, however, there is a
easy reduction from the PCSM+ instance to SPA, we give the reduction (refer Appendix A.1) for the sake
of completeness.). We output M = map(Ms) whose feasibility is guaranteed by the invariants mentioned
earlier. The complexity of our algorithm depends on s and the time required to compute a stable matching
in the problem instance.
In the rest of the paper, we denote by M the matching obtained as map(Ms) where Ms is a stable
matching in Gs. For any resident ri ∈ R, we define
map−1(ri, Ms) = rji
i
= rs−1
i
where 0 ≤ ji ≤ s − 1 and Ms(rji
otherwise.
i ) is a non-dummy hospital
Recall by Invariant (I3), exactly one of the level copy of ri in Gs is matched to a non-dummy hospital in Ms.
For any feasible matching M ′ in G consider the set YM⊕M ′ -- recall that this is a collection of M alternating
paths and cycles in G. For any path or cycle ρ in YM⊕M ′ , let us denote by ρs = map−1(ρ, Ms) the path or
cycle in Gs obtained by replacing every resident r in ρ by map−1(r, Ms). Recall that if a resident r is present
j for i = 0, . . . , s − 1. The map−1
in the class C h
function maps a resident r in G to a unique level-i copy in Gs. Using Lemma 1 and these observations we
get the following corollary.
j defined by a hospital h in G, then in the graph Gs, ri ∈ ¯C h
Corollary 1. Let ρ be an alternating path or an alternating cycle in YM⊕M ′ , then Ms ⊕ ρs is a feasible
matching in Gs, where ρs = map−1(ρ, Ms).
The following technical lemma is useful in proving the properties of the matchings produced by our algo-
rithms.
Lemma 2. Let ρ be an alternating path or an alternating cycle in YM⊕M ′ , and ρs = map−1(ρ, Ms).
1. There cannot be any edge labeled (1, 1) in ρs.
2. Let hrja
b i be a sub-path of ρs, where h = Ms(rjb
a , h, rjb
(1, 1), where j′
a < ja.
b ). Then, the edge (rj′
a
a , h) /∈ ρs cannot be labeled
a , h, rjb
Proof. Let hrja
(Corollary 1), the set (Ms(h) \ {rjb
the edge (rja
(rj′
(Ms(h) \ {rjb
b i be a sub-path of ρs, where h = Ms(rjb
b ) (Figure 4). As Ms ⊕ ρs is feasible in Gs
a , h) is labeled (1, 1),
a , h) blocks Ms contradicting its stability. This proves (1). To prove (2), assume that the edge
k ) in Gs, hence
a } is feasible for h in Gs. Now since (rja
a belong to the same class (say ¯C h
a } is feasible for h. Thus the edge (rj′
a , h) /∈ ρs is labeled (1, 1). The residents rja
a , h) blocks Ms contradicting its stability.
a and rj′
b }) ∪ {rja
a
a
a
b }) ∪ {rj′
a
10
h
rja
a
rjb
b
j′
a
a
r
Fig. 4. The edges (rja
a , h) and (rjb
b , h) belong to ρ, while the edge (r
j′
a , h) does not belong to ρ.
a
4.1 Maximum cardinality popular matching
Let G = (R∪H, E) be an instance of the LCSM+ problem where we are interested in computing a maximum
cardinality popular matching. We use our generic reduction with the value of the parameter s = 2. Since G2
is linear in the size of G, and a stable matching in an LCSM+ instance can be computed in O(mn) time [10],
we obtain an O(mn) time algorithm to compute a maximum cardinality popular matching in G. In case
G is a PCSM+ instance, we use the linear time algorithm in [2] for computing a stable matching to get a
linear time algorithm for our problem. The proof of correctness involves two things -- we first show that M
is popular in G. We then argue that it is the largest size popular matching in G. We state the main theorem
of this section below.
Theorem 3. Let M = map(M2) where M2 is a stable matching in G2. Then M is a maximum cardinality
popular matching in G.
We break down the proof of Theorem 3 in two parts. Lemma 3 shows that the assignment M satisfies
all the conditions of Theorem 1. Lemma 5 shows that the matching output is indeed the largest size pop-
ular matching in the instance. Let M ′ be any assignment in G. Recall the definition of YM⊕M ′ -- this set
contains M alternating paths and M alternating cycles in G and the edge labels on the M ′ edges belong to
{(−1, 1), (1, −1), (1, 1)}.
Lemma 3. Let M = map(M2) where M2 is a stable matching in G2 and let M ′ be any feasible assignment
in G. Consider the set of alternating paths and alternating cycles YM⊕M ′ . Then, the following hold:
1. An alternating cycle C in YM⊕M ′ , does not contain any edge labeled (1, 1).
2. An alternating path P in YM⊕M ′ that starts or ends with an edge in M ′, does not contain any edge
labeled (1, 1).
3. An alternating path P in YM⊕M ′ which starts and ends with an edge in M , contains at most one edge
labeled (1, 1).
Proof. We first prove the parts (1) and (2). Recall that M = map(M2) where M2 is a stable matching in
G2. Assume that ρ = hu0, v1, u1, . . . , vk, uki where for each i = 0, . . . , k, vi = M (ui) (in case ui is a hospital,
vi ∈ M (ui)). In case ρ is a cycle, all subscripts follow mod k arithmetic. The existence of ρ in YM⊕M ′ implies
that there is an associated M2 alternating path or an M2 alternating cycle ρ2 = map−1(ρ, M2) in G2.
Now assume for the sake of contradiction that ρ contains an edge e = (ra, hb) /∈ M labeled (1, 1) for some
a = 0, . . . , k, and b = 0, . . . , k. We observe the following about preferences of ra and hb in G.
(O1) ra prefers hb over ha = M (ra).
(O2) hb prefers ra over rb ∈ M (hb), where rb = corr(ra).
11
Using the presence of an edge labeled (1, 1) in ρ, we will contradict the stability of M2 in G2. Consider
b ), we observe that e′ /∈ M2. We consider the four cases that
a , hb) in G2. Since hb = M2(rjb
the edge e′ = (rja
can arise depending on the values of ja and jb.
1. ja = jb = 0
2. ja = jb = 1
3. ja = 1 and jb = 0
4. ja = 0 and jb = 1
Recall observation (O1), and the fact that the residents do not change their preferences in G2 w.r.t.
a prefers hb over
a ). Using (O2) and the fact that a hospital h in G2 prefers level-1 residents over level-0 residents,
b , which implies
the hospitals originally in G. This implies in all the four cases above, the resident rja
ha = M2(rja
we can conclude the following. For the cases (1), (2) and (3), hospital hb prefers rja
that the pair (rja
a , hb) is labeled (1, 1), and thus forms a blocking pair w.r.t. M2 (using Lemma 2(1)).
a over rjb
We now consider the three different cases for ρ depending on whether ρ is a path or a cycle. When ρ is
a path, we break down its proof in two cases -- (i) ρ starts or ends with a resident unmatched in M . (ii) ρ
starts or ends with an under-subscribed hospital. In each of the different possibilities for ρ, we show that the
stability of M2 can be contradicted even in case (4), i.e. when ja = 0 and jb = 1.
0 , h1, rj1
1 , . . . , hk−1, rjk−1
-- ρ = hr0, h1, r1, . . . , hk−1, rk−1i is an alternating path that starts or ends with a resident which is
k−1 i and for t = 0, . . . , k − 1,
unmatched in M . Here ρ2 = map−1(ρ, M2) = hrj0
jt ∈ {0, 1}.
Using invariants (I1), (I2), and (I3), we conclude that a resident r remains unmatched in M2 when its
level-0 copy is matched to the dummy hospital dr, and the level-1 copy is unmatched in M2. Therefore,
the first resident on the path ρ2 is a level-1 resident. Furthermore, the second resident on the path r1
has to be a level-1 resident. Otherwise, as r1
0 is unmatched in M2 and h1 prefers a level-1 resident over a
level-0 resident, the edge (r1
0, h1) will be labeled (1, 1), and thus forms a blocking pair w.r.t. M2 (using
Lemma 2(1)).
We consider an edge e ∈ ρ such that b = a + 1. In case (4), we observe that as j0 = j1 = 1, ja = 0, and
a < b, there exists an index x in ρ2 such that there is a transition from a level-1 resident to a level-0
resident. That is, (r0
We enumerate the possible labels for the edge ex = (rx−1, hx) in G.
x−1, hx) /∈ M2 both belong to ρ2.
x, hx) ∈ M2 and (r1
• If ex is labeled (1, 1) or (1, −1), then the edge (r1
x−1, hx) is labeled (1, 1), and thus blocks M2 (using
Lemma 2(1)).
• If ex is labeled (−1, 1), then the edge (r0
x−1, hx) is labeled (1, 1), and thus blocks M2 (using Lemma 2(2)).
-- ρ = hh0, r1, h1, . . . , rk−1, hk−1i is an alternating path that starts or ends with an under-subscribed
1 , h1, . . . , rjk−1
k−1 , hk−1i and for t = 1, . . . , k − 1, jt ∈ {0, 1}.
1, h0) is labeled (1, 1), as h0 is unmatched in M , and r0
hospital. Here ρ2 = map−1(ρ, M2) = hh0, rj1
Observe that if j1 = 1, then (r0
1 prefers h0
to dr1 (dr1 = M2(r0
1) using invariants (I1), (I2), and (I3)), contradicting the stability of Ms (using
Lemma 2(1)). Thus, it must be the case that j0 = 0. Note that the edge (r1, h0) can not be labeled (1, 1)
in G, as h0 being under-subscribed prefers being matched to r1, and residents do not change their votes,
and thus the edge (r0
We consider an edge e ∈ ρ such that a = b + 1. In case (4), we observe that as j1 = 0, jb = 1, and a > b,
there exists an index x in ρ2 such that there is a transition from a level-0 resident to a level-1 resident.
That is, (r0
x+1, hx) /∈ M2 both belong to ρ2. Using an argument similar to in the case
above, we can show that either the edge (r1
x+1, hx) is labeled (1, 1), and therefore
forms a blocking pair w.r.t. M2.
1, h0) is labeled (1, 1), contradicting the stability of M2 (using Lemma 2(1)).
x+1, hx) or the edge (r0
x, hx) ∈ M2 and (r1
-- ρ = hr0, h0, r1, h1, . . . , rk, hk, r0i is an alternating cycle. Here ρ2 = map−1(ρ, Ms) = hrj0
0 , h0, rj1
1 , h1, . . . , rjk
k , hk, rj0
0 i
and for t = 1, . . . , k − 1, jt ∈ {0, 1}.
We consider an edge e ∈ ρ such that a = b + 1. As ja = 0 and jb = 1, and b < a, this is a transition
from a level-1 resident to a level-0 resident in the cycle ρ2. To complete the cycle ρ2 there must exist an
index x such that there is a transition from a level-0 resident to a level-1 resident. That is, (r0
x, hx) ∈ M2
and (r1
x+1, hx) /∈ M2 both belong to ρ2. Using an argument similar to as in the first case, we can show
that either the edge (r1
x+1, hx) is labeled (1, 1), and therefore forms a blocking pair
w.r.t. M2.
x+1, hx) or the edge (r0
12
We now prove part (3) of the lemma. Consider P = hr0, h0, . . . , rk−1, hk−1i where for each i = 0, . . . , k−1,
M (ri) = hi. The existence of P in YM⊕M ′ implies that there exists an M2 alternating path P2 in G2. Here
P2 = hrj0
k−1 , hk−1i, and for t = 0, . . . , k − 1, jt ∈ {0, 1}.
0 , h0, . . . , rjk−1
For the sake of contradiction assume that P contains at least two edges, e1 = (rx, hx−1), e2 = (ry, hy−1)
for some x, y = 1, . . . , k − 1, w.l.o.g. x 6= y, x < y and e1, e2 are labeled (1, 1). We observe the following about
preferences of rx, ry and hx−1, hy−1 in G.
(O1) rx prefers hx−1 over hx = M (rx).
ry prefers hy−1 over hy = M (ry).
(O2) hx−1 prefers rx over rx−1 ∈ M (hx−1).
hy−1 prefers ry over ry−1 ∈ M (hy−1).
Using the presence of the edges e1 and e2 labeled (1, 1) in P , we will contradict the stability of M2
) and
y , hy−1) in G2, and since hx−1 = M2(rjx−1
x , hx−1) and e′
2 = (rjy
1 = (rjx
x
2 /∈ M2.
1, e′
1, and consider the four cases that can arise depending on the values of jx
in G2. Consider the edges e′
hy−1 = M2(rjy−1
), note that e′
y
We first consider the edge e′
and jx−1.
1. jx−1 = jx = 0
2. jx−1 = jx = 1
3. jx−1 = 0 and jx = 1
4. jx−1 = 1 and jx = 0
hospitals originally in G. This implies that in all the four cases above, the resident rjx
hx = M2(rjx
we can conclude the following. For the cases (1), (2) and (3), hospital hx−1 prefers rjx
implies that the pair (rjx
Recall observation (O1), and the fact that the residents do not change their preferences in G2 w.r.t the
x prefers hx−1 over
x ). Using (O2) and the fact that a hospital h in G2 prefers level-1 residents over level-0 residents,
x−1 , which
x , hx−1) is labeled (1, 1), which contradicts the stability of Ms (using Lemma 2(1)).
2, we conclude that the first three cases do not arise. There is only
one case left to consider, when jx−1 = 1, jx = 0 and jy−1 = 1, jy = 0. As x 6= y, x < y, and jx = 0, jy−1 = 1,
there exists an index ℓ in P2 such that there is a transition from a level-0 resident to a level-1 resident. That
is, (r0
With a similar analysis for the edge e′
ℓ , hℓ) ∈ M2 and (r1
We enumerate the possible labels for the edge eℓ = (rℓ+1, hℓ) in G.
ℓ+1, hℓ) /∈ M2 both belong to P2.
x over rjx−1
-- If eℓ is labeled (1, 1) or (1, −1), then the edge (r1
ℓ+1, hℓ) is labeled (1, 1), which contradicts the stability
-- If eℓ is labeled (−1, 1), then the edge (r0
ℓ+1, hℓ) is labeled (1, 1), which contradicts the stability of Ms
of Ms (using Lemma 2(1)).
(using Lemma 2(2)).
This completes the proof.
⊓⊔
Lemma 4. There is no augmenting path with respect to M in YM⊕M ′ .
1 , h2, . . . , hk−1, rjk−1
k−1 , hki in G2, and for t = 0, . . . , k − 1, jt ∈ {0, 1}.
Proof. Let P = hr0, h1, r1, h2, . . . , hk−1, rk−1, hki be an augmenting path where for each i = 1, . . . , k −
1, M (ri) = hi. The existence of P in YM⊕M ′
implies that there exists an M2 augmenting path P2 =
hrj0
0 , h1, rj1
Using invariants (I1), (I2), and (I3), we conclude that a resident r remains unmatched in M2 when its
level-0 copy is matched to the dummy vertex dr, and the level-1 copy is unmatched in M2. Therefore the first
resident on the path P2 is a level-1 resident. The second resident on the path r1 has to be a level-1 resident,
otherwise the edge (r1
0, h1) will be labeled (1, 1), and thus contradict the stability of M2 (using Lemma 2(1)).
This is because r1
0 prefers being matched to h1 than being unmatched in M2, and h1 prefers level-1 resident
over a level-0 resident. Observe that jk−1 = 0, else the pair (r0
k−1 is matched to
dr (by invariants (I1) and (I2)), which is at the end of its preference list, and hk is unmatched in M ′.
k−1, hk) is labeled (1, 1), as r0
Therefore the path P2 is of the form hr1
k−1, hki. As j0 = j1 = 1 and jk−1 = 0,
there exists an index x in P2 such that there is a transition from a level-1 resident to a level-0 resident. That
is, (r0
x, hx) ∈ M2 and (r1
We enumerate the possible labels for the edge ex = (rx−1, hx) in G.
x−1, hx) /∈ M2 both belong to P2.
1, h2, . . . , hk−1, r0
0, h1, r1
13
-- If ex is labeled (1, 1) or (1, −1), then the edge (r1
x−1, hx) is labeled (1, 1), which contradicts the stability
of Ms (using Lemma 2(1)).
-- If ex is labeled (−1, 1), then the edge (r0
x−1, hx) is labeled (1, 1), which contradicts the stability of Ms
(using Lemma 2(2)).
This contradicts our assumption that P is augmenting with respect to M in YM⊕M ′ .
⊓⊔
Lemma 5. There exists no popular matching M ∗ in G such that M ∗ > M .
Proof. For contradiction, assume that such an assignment M ∗ exists in G. Consider the set YM⊕M ∗ ; recall
that this set contains alternating paths and cycles possibly containing edges labeled (−1, −1). Since M ∗ >
M there must exist an augmenting path P in YM⊕M ∗ . We first claim that the path P must contain at least
one edge labeled (−1, −1). If not, then the path P is also contained in YM⊕M ∗ . However, by Lemma 4 there
is no augmenting path with respect to M in YM⊕M ′ for any feasible matching M ′ in G.
We now remove all edges from P which are labeled (−1, −1). This breaks the path into sub-paths say
P1, P2, . . . , Pt for some t ≥ 1, where P1 and Pt have one endpoint unmatched in M . Consider the path P1;
since P1 does not contain any (−1, −1) edge this implies that P1 ∈ YM⊕M ∗ . Without loss of generality,
assume that P1 starts with a resident r which is unmatched in M . Thus using Lemma 3(2), P1 does not
contain any edge labeled (1, 1). Let us denote by ∆(M ∗, M )P1 the difference between votes of M ∗ and M
restricted to vertices of path P1. It is clear that ∆(M ∗, M )P1 < 0. Also, for each i = 2, . . . , t − 1, the
alternating paths Pi have both of their endpoints matched in M . Thus we have ∆(M ∗, M )Pi ≤ 0 as there
can be at most one (1, 1) edge (by Lemma 3(3)) in these paths, but the endpoints prefer M , as they are
matched in M but not in M ′. If Pt exists, then a argument similar as given for P1, we have ∆(M ∗, M )Pt < 0.
Using these observations, we conclude that M is more popular than M ∗, a contradiction to the assumption
that M ∗ and M are both popular.
Thus, for any given matching M ∗ such that M ∗ > M , we know that M is more popular than such a
matching. This completes the proof of the lemma, and shows that the matching M = map(M2) is a maximum
⊓⊔
cardinality popular matching in G.
4.2 Popular matching amongst maximum cardinality matchings
In this section we give an efficient algorithm for computing a matching which is popular amongst the set of
maximum cardinality matchings. The matching M that we output cannot be beaten in terms of votes by any
feasible maximum cardinality matching. Our algorithm uses the generic reduction with a value of s = R
= n1 (say). Thus, Rn1 = n2
1). Furthermore, En1 = O(mn1) where m = E.
Thus the running time of the generic algorithm presented earlier with s = n1 for an LCSM+ instance is
O(mn · n1) = O(mn2) and for a PCSM+ instance is O(mn1) = O(mn).
1, and Hn1 = H + O(n2
To prove correctness, we show that the matching output by our algorithm is (i) maximum cardinality
and (ii) popular amongst all maximum cardinality feasible matchings. Let M = map(Mn1) and M ∗ be any
maximum cardinality feasible matching in G. Consider the set YM⊕M ∗ , and let ρ be an alternating path or
an alternating cycle in YM⊕M ∗ . Let ρn1 = map−1(ρ, Mn1) denote the associated alternating path or cycle in
Gn1 . We observe that every hospital on the path ρn1 is a non-dummy hospital since ρn1 was obtained using
the inverse-map of ρ. We observe two useful properties about such a path or cycle ρn1 in Gn1 . We show that
if for a hospital h ∈ ρn1, the level of the unmatched resident incident on h is greater than the level of the
matched resident incident on h, then such a level change is gradual, and the associated edge in ρ has the
label (−1, −1). Lemma 6, gives a proof of these.
Lemma 6. Let ρn1 be an alternating path or an alternating cycle in Gn1 and let h be a hospital which has
degree two in ρn1 . Let hrja
b ) = h. If ja > jb , we claim the
following:
b i be the sub-path containing h where M (rjb
a , h, rjb
1. ja = jb + 1.
2. The associated edge (ra, h) ∈ ρ is labeled (−1, −1).
14
Proof. We first prove that ja = jb + 1. For contradiction, assume that ja > jb + 1. Observe that h prefers all
the level-ja residents over any level-jb resident. We consider the edge e = (ra, h) in the graph G. We claim
that the label for the edge e cannot be (1, 1) or (1, −1), otherwise the edge (rja
a , h) is labeled (1, 1) in Gn1 as
the residents do not change their votes. Similarly, we claim that the label for the edge e cannot be (−1, 1)
or (−1, −1), as rja−1
) (by
invariant (I3)), and prefers h to dja−1
, and h prefers all the level-(ja − 1) residents over any level-jb resident.
In this case the edge (rja −1
is matched in Mn1 to the last dummy on its preference list, dja−1
, h) is labeled (1, 1) in Gn1 , and thus blocks Mn1 (by Lemma 2(2)).
ra = Mn1 (rja−1
ra
a
a
a
To prove part (b), we assume ja = jb + 1. We enumerate the possible labels for the edge e = (ra, h) in G.
-- If e is labeled (1, 1) or (1, −1), then the edge (rja
a , h) is labeled (1, 1), as rja
prefers any level-ja resident over a level-jb resident. Thus, the edge (rja
-- If e is labeled (−1, 1), then the edge (rjb
a , h) is labeled (1, 1), as rjb
prefers rjb
a over rjb
b according to its preference list. Thus, the edge (rjb
a prefers h over Mn1(rja
a ), and h
a , h) blocks Mn1 (by Lemma 2(1)).
a ), and h
a , h) blocks Mn1 (by Lemma 2(2)).
ra = Mn1(rjb
a prefers h over djb
Thus, the only possible label for the edge (ra, h) is (−1, −1).
⊓⊔
We use Lemma 7 to prove that M is a maximum cardinality matching in G.
Lemma 7. Let M ∗ be any feasible maximum cardinality matching in G. Then there is no augmenting path
with respect to M in YM⊕M ∗ .
i ).
Proof. For the sake of contradiction assume that the path P = hr0, h1, r1, . . . , hk−1, rk−1, hki is an aug-
menting path where for each i = 1, . . . , (k − 1), M (ri) = hi. Here r0 is unmatched in M , and hk is
under-subscribed in M . The existence of P in YM⊕M ∗ implies that there exists an Mn1 augmenting path
Pn1 = map−1(P, Mn1 ) = hrj0
k−1 , hki in Gn1 , and for t = 0, . . . , k −1, jt ∈ {0, . . . , n1 −1},
where hi = Mn1(rji
1 , . . . , hk−1, rjk−1
0 , h1, rj1
Since r0 is unmatched in M , by invariant (I3), it implies that for 0 ≤ i ≤ n1 − 2, Mn1(ri
r0 , and
is unmatched in Mn1 . This implies that the first resident in the path Pn1 is a level-(n1 − 1) resident
. The second resident on the path Pn1 also has to be a level-(n1 − 1) resident. If not, then the edge
, h1) is labeled (1, 1) since h1 prefers rn1−1
is
k−1. If not, then the
rk−1 ) on its preference list
rn1−1
0
rn1−1
0
(rn1−1
unmatched in Mn1. The last resident in the path Pn1 is a level-0 resident i.e. rjk−1
edge (r0
k−1 is matched to the last dummy hospital (d0
(by invariant (I3)), and hk is under-subscribed in Mn1.
to any resident at a level lower than n1 − 1 and rn1−1
k−1, hk) is labeled (1, 1), as r0
k−1 = r0
0) = di
0
0
0
Thus, in the path Pn1 , the first two residents are level-(n1 − 1), while the last resident is level-0. Recall
that the path Pn1 was obtained as an inverse-map of the path P in G. Since the path P contains at most n1
residents (possibly all of the residents in G), the path Pn1 also contains at most n1 residents. From Lemma 6
we observe that the difference in the levels of two residents in a sub-path of Pn1 can be at most one. Thus,
it must be the case that residents at all the levels n1 − 1 to 0 are present in Pn1 . However, since there are
two residents at level-(n1 − 1) (first two residents) and one resident at level-0 (last resident), it is clear that
residents at all levels from n1 − 1 to 0 cannot be accommodated in a path containing at most n1 residents.
This contradicts the existence of such a path Pn1 in Gn1 which implies that the assumed augmenting
path P with respect to M cannot exist. This proves that M = map(Mn1) is a max-cardinality matching in
⊓⊔
G.
We can now conclude that the set YM⊕M ∗ is a set of alternating (and not augmenting) paths and
alternating cycles. It remains to show that M is popular amongst all maximum cardinality feasible matchings
in G. Let M ∗ be any feasible maximum cardinality matching in G. In Lemma 8 we show that if there is an
edge (r, h) ∈ M ∗ \ M labeled (1, 1) in ρ, then in ρn1 , for the hospital h, the level of its unmatched neighbour
(resident) is lower than the level of its matched neighbour (resident).
Lemma 8. If an edge (ra, h) ∈ ρ is labeled (1, 1), then in ρn1 for the sub-path hrja
h, we have ja < jb.
a , h, rjb
b i where Mn1(rjb
b ) =
15
Proof. Let an edge e = (ra, h) be labeled (1, 1) in ρ. We observe the following about preferences of ra and h
in G.
(O1) ra prefers h over M (ra).
(O2) h prefers ra over rb ∈ M (h), where rb = corr(ra).
Consider the edge e′ = (rja
a , h) in Gn1 , as h = Mn1(rjb
b ) it implies e′ /∈ Mn1. Consider the three cases
that can arise depending on the values of ja and jb.
1. ja = jb
2. ja > jb
3. ja < jb
Recall observation (O1), and the fact that the residents do not change their preferences in Gn1 w.r.t. the
hospitals originally in G. This implies in all the three cases above, the resident rja
a ).
Using (O2) and the fact that a hospital h in Gn1 prefers level-p residents over level-q residents, when p > q,
we can conclude the following. For the cases (1) and (2), hospital h prefers rja
b , which implies that
the pair (rja
a , h) is labeled (1, 1), which is a blocking pair for Mn1 (using Lemma 2(1)). This contradicts the
⊓⊔
stability of Mn1. We therefore conclude that ja < jb.
a prefers h over Mn1(rja
a over rjb
Lemma 9 shows that in an alternating path in YM⊕M ∗ with exactly one endpoint unmatched in M or an
alternating cycle, the number of edges labeled (1, 1) cannot exceed the number of edges labeled (−1, −1).
Lemma 9. Let ρ be an alternating path or an alternating cycle in YM⊕M ∗ . Then the number of edges labeled
(1, 1) in ρ is at most the number of edges labeled (−1, −1).
Proof. Depending on the nature of ρ we have three different cases.
-- ρ is an alternating path which starts with an unmatched resident in M .
-- ρ is an alternating path which starts with a hospital which is under-subscribed in M .
-- ρ is an alternating cycle.
The proof idea is similar in all the three cases. In each of the above, we consider ρn1 = map−1(ρ, Mn1). For
every edge labeled (1, 1) in ρ we show a change (increase / decrease) in the level of the residents which are
neighbours of a particular hospital. We show that each such change must be complemented with another
change (decrease / increase resp.) in the level of the residents which are neighbours to some other hospital.
Finally, we show that the second type of change translates to a (−1, −1) edge in ρ.
Let ρ be an alternating cycle. Consider a hospital hi ∈ ρ for which there is an edge (r, h) labeled (1, 1)
incident on it in ρ. Consider the associated hospital hi ∈ ρn1 . W.l.o.g. let hrji−1
i i be a sub-path of ρ
when traversing ρn1 in counter-clock-wise direction. By Lemma 8, we know that the level of the unmatched
resident incident on hi is lower than the level of the matched resident incident on hi. Thus there is an
increase in level of residents when at hi (while traversing ρn1 in counter-clockwise direction). This is true
for any hk ∈ ρn1 where the associated hospital in ρ has a (1, 1) edge incident on it. We now recall from
Lemma 6(2) that whenever a hospital hk ∈ ρn1 has a level decrease, the associated edge in ρ is labeled
(−1, −1). Furthermore the decrease in levels at a hospital is gradual. Thus, it must be the case that the
number of (1, 1) edges in ρ is at most the number of (−1, −1) edges in ρ.
i−1 , hi, rji
In case ρ is an alternating path starting at an unmatched resident, we show that in the path ρn1 the
first two residents are level-(n1 − 1) residents (see Claim 1 below for a proof). Furthermore, consider the
first edge (r, h) ∈ ρ that is labeled (1, 1). The associated hospital hi has a increase in the level of its two
neighbouring residents. However, since ρn1 started with two level-(n1 − 1) residents (which is the highest
level possible). Therefore, there must have been some hospital hk preceding hi in ρn1 which has a decrease
in the levels of the two neighbours. Using these facts it is easy to prove the following:
Number of (1, 1) edges in ρ ≤ Number of increases in ρn1 ≤ Number of decreases in ρn1 ≤ Number of
(−1, −1) edges in ρ.
This completes the proof in case ρ is a path starting at an unmatched resident.
Finally, we are left with the case when ρ is an alternating path which starts with a hospital hi which is
under-subscribed in M . We show that in the associated path ρn1 , the first resident is a level-0 resident (see
16
Claim 2 below for a proof). Note that in this case since the path starts at a hospital, whenever we have an
edge labeled (1, 1) in ρ, the associated hospital hi in ρn1 has an increase in the levels of the two neighbouring
residents. Now first resident in the path is at the lowest possible level, it must be the case that there is a
hospital hk preceding hi in ρ for which there is a decrease in the level of the neighbouring residents. Now
using arguments similar to those in the case of path starting at an unmatched resident, we conclude that
the number of (1, 1) edges in ρ is at most the number of (−1, 1) edges in ρ.
Claim 1: ρ = hr0, h1, r1, . . .i starts with an unmatched resident. As r0 is unmatched in M , by invariant
(I3), it implies that for 0 ≤ i ≤ n1 − 2, Mn1(ri
is unmatched in Mn1. Therefore the first
resident rj0
0 on the path ρn1 is a level-(n1 − 1) resident, that is j0 = n1 − 1. Furthermore, the second resident
on the path r1 has to be a level-(n1 − 1) resident. If not, then as rn1−1
is unmatched in Mn1 and h1 prefers
a level-(n1 − 1) resident to a level-v resident (v < n1 − 1), the edge (rn1−1
, h1) will be labeled (1, 1), and
thus blocks Mn1 contradicting its stability (using Lemma 2(1)).
Claim 2: ρ = hh0, r1, h1, . . .i starts with an under-subscribed hospital. The first resident rj1
has to be a level-0 resident, that is j1 = 0. If not, i.e. if j1 = 1, then d0
Mn1, and prefers h1 to d0
matched to r0
(using Lemma 2(2)).
1 on the path ρn1
r0 in
r0. The hospital h1 on the other hand is under-subscribed in Mn1 and prefers being
1, h1) is labeled (1, 1), and blocks Mn1 contradicting its stability
⊓⊔
1 in Mn1. Thus, the edge (r0
r0 = Mn1 (r0
1) (by invariant (I3)) d0
0
0) = di
r, and rn1−1
0
0
Thus, we get the following theorem:
Theorem 4. Let M = map(Mn1) where Mn1 is a stable matching in Gn1 . Then M is a popular matching
amongst all maximum cardinality matchings in G.
Discussion: A natural question is to consider popular matchings in LCSM instances. An LCSM instance
need not admit a stable matching. However we claim that restricted to LCSM instances which admit a stable
matching, our results hold without any modification. To obtain the result, we claim that Lemma 1 holds in
the presence of lower quotas on classes. Additionally, if the given LCSM instance G admits a stable matching,
the graph Gs for s = 1, . . . , n1 also admits a stable matching. We thank Prajakta Nimbhorkar for pointing
this to us.
Acknowledgement: We thank the anonymous reviewers whose comments have improved the presentation.
References
1. D. J. Abraham, R. W. Irving, T. Kavitha, and K. Mehlhorn. Popular Matchings. SIAM Journal on Computing,
37(4):1030 -- 1045, 2007.
2. D. J. Abraham, R. W. Irving, and D. F. Manlove. Two Algorithms for the Student-Project Allocation Problem.
J. of Discrete Algorithms, 5(1):73 -- 90, 2007.
3. P. Bir´o, D. Manlove, and S. Mittal. Size Versus Stability in the Marriage Problem. Theoretical Computer Science,
411(16-18):1828 -- 1841, 2010.
4. F. Brandl and T. Kavitha. Popular Matchings with Multiple Partners. CoRR, abs/1609.07531, 2016.
5. ´A. Cseh and T. Kavitha. Popular Edges and Dominant Matchings. In Proceedings of the Eighteenth Conference
on Integer Programming and Combinatorial Optimization, pages 138 -- 151, 2016.
6. T. Fleiner and N. Kamiyama. A Matroid Approach to Stable Matchings with Lower Quotas. In Proceedings of
the Twenty-third Annual ACM-SIAM Symposium on Discrete Algorithms, pages 135 -- 142, 2012.
7. D. Gale and L. Shapley. College Admissions and the Stability of Marriage. American Mathematical Monthly,
69:9 -- 14, 1962.
8. P. Gardenfors. Match Making: assignments based on bilateral preferences. Behavioural Sciences, 20:166 -- 173,
1975.
9. D. Gusfield and R. W. Irving. The Stable Marriage Problem: Structure and Algorithms. MIT Press, 1989.
10. C.-C. Huang. Classified Stable Matching. In Proceedings of the Twenty-First Annual ACM-SIAM Symposium
on Discrete Algorithms, pages 1235 -- 1253, 2010.
11. C.-C. Huang and T. Kavitha. Popular Matchings in the Stable Marriage Problem.
In Proceedings of 38th
International Colloquium on Automata, Languages and Programming , pages 666 -- 677, 2011.
17
12. N. Kamiyama. Popular Matchings with Two-Sided Preference Lists and Matroid Constraints. Technical Report
MI 2016-13, 2016.
13. T. Kavitha. A Size-Popularity Tradeoff in the Stable Marriage Problem. SIAM Journal on Computing, 43(1):52 --
71, 2014.
14. Z. Kir´aly. Better and Simpler Approximation Algorithms for the Stable Marriage Problem. Algorithmica,
60(1):3 -- 20, 2011.
A Appendix
A.1 Reduction from a PCSM+ instance to an SPA instance
An instance of SPA [2] consists of students, projects and lecturers. Each lecturer has an upper bound on
the maximum number of students that he/she is willing to advise. Each project has an upper bound on the
number of students it can accommodate. Each project is owned by exactly one lecturer. Each student has a
preference ordering over a subset of the projects, and each lecturer has a preference over the students.
We detail on the reduction from PCSM+ instance to an SPA instance here. For a resident r in the PCSM+
instance, a corresponding student sr is introduced in the SPA instance. For each hospital h, a lecturer lh with
capacity q(h) is added in the SPA instance. For each class C h
j in the classification provided by a hospital h,
a project pj is associated with the lecturer lh, and the upper-bound of pj is equal to q(C h
j ). The preference
list of lh is obtained from its corresponding hospital h. If the resident r is the k-th most preferred resident
in the preference list of h, then the student sr is the k-th most preferred student in the preference list of lh.
Similarly, the preference list of a student sr is created from its corresponding resident r. Let C h
j be the class
that the resident r appears in the classification provided by the k-th most preferred hospital in its preference
list, then pj is the k-th most preferred project in the preference list for sr. As the classifications associated
with every hospital in the PCSM+ instance are a partition over its preference list, there is no ambiguity in
describing the preference of the students.
18
|
1512.08555 | 1 | 1512 | 2015-12-28T23:15:12 | Maximium Priority Matchings | [
"cs.DS"
] | Let $G=(V,E)$ be an undirected graph with $n$ vertices and $m$ edges, in which each vertex $u$ is assigned an integer priority in $[1,n]$, with 1 being the "highest" priority. Let $M$ be a matching of $G$. We define the priority score of $M$ to be an $n$-ary integer in which the $i$-th most-significant digit is the number of vertices with priority $i$ that are incident to an edge in $M$. We describe a variation of the augmenting path method (Edmonds' algorithm) that finds a matching with maximum priority score in $O(mn)$ time. | cs.DS | cs | Maximium Priority Matchings
Jonathan Turner
wucse-2015-06
Abstract
Let G = (V, E) be an undirected graph with n vertices and m edges,
in which each vertex u is assigned an integer priority in [1, n], with 1
being the "highest" priority. Let M be a matching of G. We define
the priority score of M to be an n-ary integer in which the i-th most-
significant digit is the number of vertices with priority i that are incident
to an edge in M . We describe a variation of the augmenting path
method (Edmonds' algorithm) that finds a matching with maximum
priority score in O(mn) time.
2
1
5
1
:
v
i
X
r
a
1 Introduction
A matching in an undirected graph is a subset of its edges, no two of which
share a common endpoint.
In the maximum size matching problem, the
objective is to find a matching with the largest possible number of edges.
Edmonds showed how to solve the problem for general graphs [1] and more
efficient implementations of his method are described in [2] and [5]. Hopcroft
and Karp described a simpler algorithm for the case of bipartite graphs [4].
A matching is said to match a vertex u, if one of its edges is incident
to u. Given a graph in which vertices are assigned integer priorities ρ(u), a
maximum priority matching is one that maximizes the number of matched
vertices in the highest priority class, then maximizes the number of matched
vertices in the next priority class (without reducing the number matched in
the highest priority class), and so forth. We define a matching's priority
1
Figure 1: Example showing the priority score for a matching
score to be the n-ary number in which the i-th most-significant digit is the
number of matched vertices with priority i. Figure 1 shows an example of
a graph with a matching whose priority score is 2111000100. Adding the
edge ac yields a matching with a score of 2211100100. A maximum priority
matching is a matching that has a priority score with maximum value. This
version of the matching problem arises as a subproblem in an approximation
algorithm for an np-complete scheduling problem for crossbar switches used
in internet routers [7].
In this paper, we show how Edmonds' algorithm for the ordinary match-
ing problem can be extended to solve priority matching problems. In section
2, we focus on the special case of two priorities (ρ(u) ≤ 2 for all u). This
problem is of independent interest and provides a useful introduction to the
methods used in the more general case, which is discussed in section 3.
2 Two priority case
The two priority case can be phrased more simply by defining a set S consist-
ing of all vertices with priority 1. Our objective is then to find a maximum
2
a d f b c e g h j i 1 1 2 2 3 4 4 5 6 8 priority score: 2111000100 size matching that also matches the largest possible number of vertices in S.
In Edmonds' algorithm for maximum size matching, augmenting paths
are used to convert a matching to a larger one. Given a matching in a graph,
an augmenting path is a simple path in which the edges alternate between
matching edges and non-matching edges, with both endpoints unmatched.
By reversing the status of the matching and non-matching edges on such a
path, we can obtain a new matching with one more edge than the original.
Thus, so long as we can find an augmenting path, we can expand a given
matching. Edmonds showed that if a larger matching exists, the graph must
contain an augmenting path.
For the priority matching problem, we must adjust the definition of an
augmenting path. Let M be a matching of a graph G that does not match
all the vertices in the given set S. An augmenting path in G with respect
to M is a path p = u0, . . . , ut in which every other edge belongs to M , u0 is
unmatched, u0 ∈ S and if ut is matched, ut (cid:54)∈ S.
Figure 2 shows examples of two such paths. Observe that in both cases,
if one replaces the path edges in M with the path edges not in M , we get a
new matching that matches at least one more vertex in S. Also, note that
when we expand a matching in this way, all previously matched vertices in
S remain matched.
Figure 2: Augmenting paths for two priority case (shaded vertices are in S)
Define the count of a two-priority matching to be the number of vertices in
S that are incident to edges in the matching. Now consider two matchings M
and M(cid:48), where M has a smaller count than M(cid:48). Let N be the graph defined
by the edges that are in M or in M(cid:48), but not both. Note that N consists
of a collection of disjoint paths and cycles with alternate edges in M and
M(cid:48). Since M(cid:48) has a larger count than M , N must contain at least one path
p in which M(cid:48) matches more elements of S than does M . Since all interior
vertices of a path are matched by both M and M(cid:48), the path endpoints must
3
account for the difference in the matchings' counts. That means that at
least one endpoint must be in S and unmatched by M . If the other endpoint
is matched, it cannot also be in S. Hence, p satisfies the condtion for an
augmenting path with respect to M . Thus, for any matching that does not
have the largest possible count, there exists an augmenting path that can be
used to obtain a new matching with a larger count.
Observe that given a matching with the maximum possible count, but
less than maximum size, the original augmenting path method can be used to
obtain a matching with the same count, but one more more matching edge.
Hence one can easily convert a maximum count matching to one that has
both maximum count and size. Such a matching also maximizes the number
of matched vertices that are not in S, hence it satisfies the definition of a
maximum priority matching for the 2 priority case.
To complete the description of the augmenting path method, we still need
an algorithm to find an augmenting path. Our presentation is an adaptation
of that given in [6] for the maximum size matching problem. We start with
the special case of bipartite graphs.
The algorithm finds an augmenting path by building a collection of trees
rooted at unmatched vertices in S. Vertices that have not yet been added
to a tree are called unreached, while a vertex u in a tree is called odd or
even depending on the length of the tree path from u to the root of the
tree. Initially, the unmatched vertices in S are the only tree vertices. The
algorithm also maintains a list of eligible edges that initially contains all
edges incident to tree roots. It then repeats the following step until it either
finds an augmenting path or runs out of eligible edges.
Select an eligible edge e = {u, v} for which u is even, remove it from
the eligible list, then apply the applicable case from those listed below.
• If v is unreached and matched, let {v, w} be the matching edge
incident to v. Extend the tree containing u by making v a child
of u and w a child of v. If w is not in S, then the path from w
to the root of its tree is an augmenting path; otherwise, add all
non-matching edges incident to w to the eligible list.
4
Figure 3: Example of bipartite algorithm at start of augmenting path search
and after three steps (even vertices labelled +, odd vertices − and arrows
point to parents in trees)
• If v is unreached and unmatched then the path consisting of e
plus the tree path from u to its tree root is an augmenting path.
• If v is even then the path formed by combining e with the tree
path from u to the root of its tree and the tree path from v to the
root of its tree is an augmenting path. (Note that u and v are in
different trees, since the graph is bipartite.)
• If v is odd, ignore e and proceed to the next eligible edge.
Figure 3 illustrates the operation of the algorithm. Note that if edge bf is
selected next, the algorithm will ignore it, if fi is selected next, it will find
the odd-length augmenting path acfi, and if edge jk is selected next, it will
find the even-length augmenting path ghjkn.
To establish the correctness of the algorithm, we show that if it halts
without finding an augmenting path, then the graph must not contain one.
We start by noting a few properties of the algorithm.
1. Every tree root is unmatched and in S and each tree has exactly one
unmatched vertex.
5
a b h c d f g i j e k m n p + eligible: ab, ac, gh + a b h c d f g i j e k m n p + eligible: bf, fi, jk, jm + - + + + - - 2. Every matching edge either has both endpoints unreached, or one odd
and one even. In the latter case, the even endpoint is the child of the
odd endpoint in some tree.
3. Every edge joining two vertices in the same tree joins an odd vertex to
an even vertex.
4. If the algorithm fails to find an augmenting path, then when it fails,
every edge with an endpoint in some tree has at least one odd endpoint.
5. If the algorithm fails to find an augmenting path, then when it fails,
every even vertex is in S.
Properties 3 and 4 imply that if the algorithm fails to find an augmenting,
any path u0, u1, . . . , ut with u1 a child of u0 that alternates between matching
and non-matching edges, must also alternate between even and odd vertices.
So in particular, if the algorithm fails to find an augmenting path, but the
graph contains an augmenting path p = u0, . . . , ut with u0 unmatched and
in S, all vertices in p must alternate between even and odd.
If t is odd,
this implies that ut is odd and unmatched, but this contradicts the fact that
every odd vertex is matched. If t is even, then ut is even and matched, but
this contradicts property 5, since the matched endpoint of an augmenting
path cannot be in S.
Before proceeding to the case of general graphs, we show that for the
special case where S consists of all vertices of maximum degree, the algorithm
finds a matching that covers all vertices in S. Assume, to the contrary, that
the algorithm fails to find an augmenting path when there is some unmatched
vertex in S. Consider the collection of trees at the time the algorithm halts
and recall that by property 5, all the even vertices must be in S. Since every
tree has one more even vertex than it has odd vertices, some even vertex must
have an edge that connects it to an unreached vertex, but this contradicts
property 4. Hence, the algorithm matches all vertices in S.
Next, we show how to find augmenting paths in general graphs. As
with ordinary matchings, the key issue is handling odd-length cycles, known
as blossoms. Edmonds showed how to extend the ordinary augmenting path
search to recognize blossoms and shrink each blossom down to a single vertex,
6
Figure 4: Example of general graph with blossoms
producing a new graph which has an augmenting path if and only if the
original graph does. This is illustrated in Figure 4, which shows a blossom
B1 containing vertices c, d, e, f and g and a blossom B2 containing vertices j, k
and m. In the "shrunken graph" on the right, we have two augmenting paths
abB1h and abB1B2ih. The corresponding paths in the original graph are
abcdegfh and abcfgkmjih. There is a straightforward procedure to obtain an
augmenting path in the underlying unshrunken graph, given an augmenting
path in the current shrunken graph.
We say than a vertex is internal if it is contained in some blossom, other-
wise it is external. The base of a blossom is the unique vertex in the blossom
that has no incident matching edge with the other endpoint in the blossom.
So, in Figure 4, the base of B1 is c and the base of B2 is j.
Our algorithm modifies the blossom-shrinking procedure to accommodate
our modified augmenting paths. As in the bipartite case, the algorithm finds
an augmenting path by building a collection of trees rooted at unmatched
vertices in S. For convenience, we let β(u) denote the largest blossom con-
taining a vertex u (if u is external, β(u) = u). Initially, all edges incident
7
a d i f g h j e k m + + + b c + + + + - - - - - a d i f g h j e k m + + + b c + + + + - - - - - B1 B2 to tree roots are on the list of eligible edges. The algorithm then repeats
the following step until it either finds an augmenting path in the current
shrunken graph or runs out of eligible edges.
Select an eligible edge e = {u, v} for which β(u) is even and remove it
from the list of eligible edges.
• If v is unreached and matched, let {v, w} be the matching edge
incident to v; extend the tree by making v a child of u and w a
child of v.
If w is not in S, then the path from w to the root
of its tree in the current shrunken graph is an augmenting path;
otherwise, add all non-matching edges incident to w to the eligible
list.
• If v is unreached and unmatched, then the path consisting of e
plus the tree path from β(u) to the root of its tree in the current
shrunken graph is an augmenting path.
• If β(v) is even and in a different tree than β(u), then the path
formed by combining e with the tree path from β(u) to the root
of its tree in the current shrunken graph and the tree path from
β(v) to the root of its tree is an augmenting path.
• If β(v) is even and in the same tree as β(u), then the cycle formed
by combining e with the tree paths in the current shrunken graph
from β(u) and β(v) to their nearest common ancestor forms a
blossom. If some odd vertex x in the cycle is not in S, then there
is an augmenting path in the current shrunken graph from x to
the root of its tree that starts with the matching edge incident to
x, proceeds around the new blossom to its base and then up the
tree to the root. If there is no such vertex, add all non-tree edges
incident to odd vertices in the blossom cycle to the eligible list
and shrink the blossom.
• If β(v) is odd, ignore e and proceed to the next iteration
Figure 4 can be used to illustrate the operation of the algorithm (note that
the shaded vertices are in S). When the algorithm reaches the state shown
8
on the left side of the figure, the eligible edges are hf, dg, gk, eg and jm.
The algorithm ignores edges hf, dg and gk and forms one blossom when
processing edge eg and the other when processing jm. At this point, edges hf
and gk are again eligible. Processing either edge leads to the discovery of an
augmenting path. Also, observe that if d were not a member of S, then when
edge eg was processed, the algorithm would have found the augmenting path
abcfged. We defer the correctness proof of the algorithm to the next section.
3 Unlimited priorities
We can find a maximum priority matching for the general case using a gen-
eralization of the algorithm for the two priority case. We first maximize the
number of priority 1 vertices that are matched, then the number of prior-
ity 2 vertices and so forth. At each step, we find an augmenting path that
increases the number of matched priority i vertices without decreasing the
number of priority j vertices, for all j < i.
Define the i-score of a matching M as the n-ary integer with i digits
in which the j-th most significant digit is the number of priority j vertices
that are matched by M . For example, in Figure 1, the 2-score is 21 and the
5-score is 21110. Given a matching M with a maximum (i − 1)-score, an
i-augmenting path is a path p = u0, . . . , ut in which edges alternate between
matching edges and non-matching edges, u0 is unmatched, ρ(u0) = i, and if
ut is matched, ρ(ut) > i. Observe that because M has a maximum (i − 1)-
score, if ut is unmatched, it cannot have priority less than i. Consequently,
if we exchange the non-matching edges and matching edges in p, we obtain
a new matching with the same (i − 1)-score and a larger i-score than M .
To justify the use of augmenting paths, we must show that if a matching
M has a maximum (i− 1)-score but not a maximum i-score, then there must
be an i-augmenting path for M . Let M be such a matching and let M(cid:48) be a
matching with a larger i-score. Let N be the graph consisting of edges that
are in M or M(cid:48) but not both. N consists of a collection of disjoint paths
and even-length cycles. Since M(cid:48) has a larger i-count than M , there must
be some path p in N , in which M(cid:48) has a larger i-count than M . If p is an
odd-length path, its endpoints must be unmatched by M and at least one of
9
its endpoints must have priority i, making p an augmenting path for M . If
p is an even length path, the endpoint that is unmatched by M must have
priority i and the endpoint matched by M must have priority > i. Hence, p
is an augmenting path in this case, as well. We summarize this argument in
the following theorem.
Theorem 1 Let G = (V, E) be an undirected graph with priorities ρ(u)
and let M be a matching with a maximum (i − 1)-score. G contains an i-
augmenting path with respect to M if and only if M does not have a maximum
i-score.
To find a matching with maximum overall priority score, we initialize i = 1,
then repeat the following step until i > n.
Search for an i-augmenting path; if one is found, augment the matching
by reversing the status of the path edges, otherwise increment i.
The heart of the method is the algorithm used to find an i-augmenting path.
At the start of each path search, all unmatched priority i vertices are tree
roots, and all edges incident to these vertices are in the eligible list. The
algorithm then searches for an i-augmenting path by repeating the following
step until it either finds a path or runs out of eligible edges.
Select an eligible edge e = {u, v} for which β(u) is even and remove it
from the list of eligible edges.
• If v is unreached and matched, let {v, w} be the matching edge
incident to v; extend the tree by making v a child of u and w a
child of v. If ρ(w) > i then the path in the current shrunken graph
from w to the root of its tree is an i-augmenting path; otherwise,
add all non-matching edges incident to w to the eligible list and
continue.
• If v is unreached and unmatched, then the path consisting of e
plus the tree path in the current shrunken graph from β(u) to the
root of its tree is an i-augmenting path.
10
• If β(v) is even and in a different tree than β(u), then the path
formed by combining e with the tree path from β(u) to the root
of its tree in the current shrunken graph and the tree path from
β(v) to the root of its tree is an i-augmenting path.
• If β(v) is even and in the same tree as β(u), then the cycle formed
by combining e with the tree paths in the current shrunken graph
from β(u) and β(v) to their nearest common ancestor forms a
blossom. If some odd vertex x on the blossom cycle has ρ(x) > i,
then there is an i-augmenting path in the current shrunken graph
from x to the root of the tree that starts with the matching edge
incident to x, continues around the blossom cycle to its base and
then up the tree to the root. If there is no such vertex, add all
non-tree edges incident to odd vertices on the blossom cycle to
the eligible list and shrink the blossom.
• If β(v) is odd, ignore e and proceed to the next iteration.
Once again, to establish the correctness of the algorithm, we need to show
that if it halts without finding an i-augmenting path, then the graph must
not contain one. Note the following properties of the algorithm.
1. Every tree root is unmatched and has priority i and each tree has
exactly one unmatched vertex.
2. Every matching edge either has both endpoints unreached, or one odd
and one even. In the latter case, the even endpoint is the child of the
odd endpoint in some tree.
3. For every internal vertex x, β(x) is even, and if x is unmatched, then
β(x) is unmatched.
4. If the algorithm fails to find an augmenting path, then when it fails,
any edge that has endpoints that are both even or internal is contained
within some blossom.
5. If the algorithm fails to find an augmenting path, then when it fails,
every vertex x that is even or internal has ρ(x) ≤ i.
11
Now, suppose that the original graph contains an augmenting path p =
u0, . . . , ut, but the algorithm halts without finding a path. If both endpoints
of p are unmatched, then the endpoints cannot be in the same tree (by
property 1). In this case, let {uk, uk+1} be an edge in p with uk and uk+1 in
different trees. By property 5, at least one of uk and uk+1 must be odd and
external. Assume, without loss of generality, that uk is odd and external.
Since {uk−1, uk} is a matching edge, k is even. If just one endpoint (u0) of
p is unmatched, then t is even, ρ(ut) > i and hence ut odd and external (by
property 4). Thus, in both cases, p contains an odd external vertex uk with
k even.
Let j be the smallest even integer for which uj is odd and external.
Since uj is external, there must some vertex in {u0, . . . , uj−1} that is odd
(If not, all vertices in {u0, . . . , uj−1} must be contained in
and external.
a common blossom (property 5), and since u0 is unmatched, the blossom
must be also (property 3), hence must include the matching edge {uj−1, uj},
contradicting the fact that uj is external.) Let i be the largest integer <
j for which ui is odd and external and note that i must be odd. This
implies that all vertices in {ui+1, . . . , uj−1} are in a common blossom and
that blossom is incident to two matching edges {ui, ui+1} and {uj−1, uj}.
This contradiction implies the correctness of the algorithm. (This argument
was adapted from [6].)
Each augmenting path search can be implemented to run in O(m log n)
time using the method described in [2] to represent the current shrunken
graph. This can be reduced to O(m) time using the data structure described
in [3].
Theorem 2 The augmenting path algorithm for priority matching computes
a matching with maximum priority score. It can be implemented to run in
O(mn) time.
We close by noting that the maximum priority matching is also a max-
imum size matching. If it were not, we could find an ordinary augmenting
path in the graph that would match two more vertices, giving it higher pri-
ority score.
12
4 Closing remarks
The maximum size matching problem can be solved in O(mn1/2) time using
the algorithms described in [4] for the bipartite case and [5] for the general
case. It is possible that one or both of these algorithms could be adapted to
handle maximum priority matching.
It might also be interesting to consider a weighted version of the problem.
In the maximum weight matching problem, we seek a matching that maxi-
mizes the sum of the edge weights of the edges in the matching. While one
cannot simultaneously maximize the weight and priority score of a matching,
one could conceivably maximize the weight of a matching with a specified
minimum priority score, or maximize the priority score of a matching with a
specified minimum weight. Alternatively, one might associate weights with
vertices and find matchings that maximize the weight of matched vertices.
References
[1] Edmonds, Jack. "Paths, trees and flowers," Canadian Journal of Math-
ematics, 1965, pp. 449 -- 467.
[2] Gabow, Harold N. "An efficient implementation of Edmonds' algorithm
for maximum matching on graphs," Journal of the Association for Com-
puting Machinery, 1976, pp. 221 -- 234.
[3] Gabow, Harold N and Robert E. Tarjan. "A linear time algorithm for a
special case of disjoint set union," ACM Symposium on the Theory of
Computing (STOC), 1983, pp. 246-251.
[4] Hopcroft, John E. and Richard M Karp. "An O(n5/2 algorithm for maxi-
mum matching in bipartite graphs," SIAM Journal on Computing, 1973,
pp 225 -- 231.
[5] Micali, Silvio. and V. V. Vazirani. "An O((cid:112)V · E) algorithm for
finding maximum matchings in general graphs," IEEE Symposium on
the Foundations of Computer Science (FOCS), 1980, pp. 17-27.
13
[6] Tarjan, Robert E. Data structures and network algorithms. Society for
Industrial and Applied Mathematics, 1983.
[7] Turner, Jonathan S. "The bounded edge coloring problem and offline
crossbar scheduling," Washington University Computer Science and En-
gineering Department technical report, wucs-2015-07, 2015.
14
|
1611.06589 | 2 | 1611 | 2018-02-25T19:16:36 | Fair Division via Social Comparison | [
"cs.DS",
"cs.AI",
"cs.GT",
"math.CO"
] | In the classical cake cutting problem, a resource must be divided among agents with different utilities so that each agent believes they have received a fair share of the resource relative to the other agents. We introduce a variant of the problem in which we model an underlying social network on the agents with a graph, and agents only evaluate their shares relative to their neighbors' in the network. This formulation captures many situations in which it is unrealistic to assume a global view, and also exposes interesting phenomena in the original problem.
Specifically, we say an allocation is locally envy-free if no agent envies a neighbor's allocation and locally proportional if each agent values her own allocation as much as the average value of her neighbor's allocations, with the former implying the latter. While global envy-freeness implies local envy-freeness, global proportionality does not imply local proportionality, or vice versa. A general result is that for any two distinct graphs on the same set of nodes and an allocation, there exists a set of valuation functions such that the allocation is locally proportional on one but not the other.
We fully characterize the set of graphs for which an oblivious single-cutter protocol-- a protocol that uses a single agent to cut the cake into pieces --admits a bounded protocol with $O(n^2)$ query complexity for locally envy-free allocations in the Robertson-Webb model. We also consider the price of envy-freeness, which compares the total utility of an optimal allocation to the best utility of an allocation that is envy-free. We show that a lower bound of $\Omega(\sqrt{n})$ on the price of envy-freeness for global allocations in fact holds for local envy-freeness in any connected undirected graph. Thus, sparse graphs surprisingly do not provide more flexibility with respect to the quality of envy-free allocations. | cs.DS | cs |
Fair Division via Social Comparison
Rediet Abebe, Jon Kleinberg, and David C. Parkes
Abstract
In the classical problem of cake cutting (also known as fair division), a resource must be
divided among agents with different utilities so that each agent believes they have received a
fair share of the resource relative to the other agents. We introduce a variant of the problem
in which there is a graph on the agents modeling an underlying social network, and agents
only evaluate their shares relative to their neighbors' in the network. This formulation
captures many situations in which it is unrealistic to assume a global view by the agents,
and we also find that it exposes interesting phenomena in the original problem.
Specifically, we say that an allocation is locally envy-free if no agent envies a neighbor's
allocation and locally proportional if each agent values her own allocation as much as the
average value of her neighbor's allocations, with the former implying the latter. While
global envy-freeness implies local envy-freeness, global proportionality does not imply local
proportionaity, or vice versa. A general result is that for any two distinct graphs on the
same set of nodes and an allocation, there exists a set of valuation functions such that the
allocation is locally proportional on one but not the other.
We fully characterize the set of graphs for which an oblivious single-cutter protocol- a
protocol that uses a single agent to cut the cake into pieces -admits a bounded protocol
for locally envy-free allocations in the Robertson-Webb model, and we give a protocol with
O(n2) query complexity. We also consider the price of envy-freeness, which compares the
total utility of an optimal allocation to the best utility of an allocation that is envy-free. We
show that a lower bound of Ω(√n) on the price of envy-freeness for global allocations [10]
in fact holds for local envy-freeness in any connected undirected graph. In this sense, sparse
graphs surprisingly do not provide more flexibility with respect to the quality of envy-free
allocations.
1
Introduction
The fair allocation of resources is a fundamental problem for interacting collections of agents. A
central issue in fair allocation is the process by which each agent compares her allotment to those
of others'. While theoretical models have tended to focus on global comparisons - in which an
agent makes comparisons to the full population - a rich line of empirical work with its origins
in the social sciences has suggested that in practice, individuals often focus their comparisons
on their social network neighbors. This literature, known as social comparison theory, dates
back to work of Festinger [15], and has been explored extensively by economists and sociologists
since; for example, see Akerlof [2] and Burt [9]. The primary argument is that in many contexts,
an individual's view of their subjective well-being is based on a comparison with peers, defined
through an underlying social network structure, rather than through comparison with the overall
population [20].
In this work, we find that the perspective of social comparison theory motivates a rich set of
novel theoretical questions in classical resource allocation problems. In particular, we consider the
cake cutting problem, which refers to the challenge of allocating a single divisible, continuous,
1
good in a fair and efficient manner. The "cake" is intended to stand for a good over which
different agents have different preferences for difference pieces. This problem has a wide range
of applications including international border settlements, divorce and inheritance settlements,
and allocating shared computational resources.
Agent preferences are modeled through functions that map subintervals of the [0, 1] interval,
which represents the entire cake, to real numbers according to the value the agent assigns to that
piece. We normalize these valuations so that each agent's value for the whole cake is 1. The
entire cake is to be allocated, but agents need not receive a single, continguous interval (and
valuations are additive across pieces).
Following our goal of understanding the properties of local comparisons to network neighbors,
we study cake cutting in a setting where there is an underlying network on the agents, and fairness
considerations are defined locally to an agent's neighbors' in the network. Given a graph G and
a cake [0, 1] to be allocated, we define a locally proportional allocation to be one where each
agent values her allocation at least as much as the average value from the allocations given to
her neighbors in G. We define a locally envy-free allocation to be one where no agent envies the
allocation of any neighbor in G. Analogous to graphical games [16], it seems quite plausible that
agents may care about fairness within a local part of the population rather than with respect to
the population as a whole.
As in the global case, it is straightforward to see that a locally envy-free allocation for G is
also locally proportional for G. It is also clear that if H is a subgraph of G on the same node
set (i.e. if it contains a subset of the edges of G), then a locally envy-free allocation for G must
also be locally envy-free for H, since the constraints defining local envy-freeness for G are a
superset of the constraints defining local envy-freeness for H. For local proportionality, however,
the constraints for different graphs G and H (even when one is a subgraph of the other) can
operate quite differently, and so as a first question we ask:
Problem 1 For graphs G and H, what is the relationship between the set of locally proportional
allocations for G and the set of locally proportional allocations for H?
Network topology plays a crucial role in our results. Note that if the network under consider-
ation is the complete graph Kn, then the local definitions coincide with their global analogues; in
this sense, the local formulations contain the standard definitions as special cases. At the other
extreme, if the network is the empty graph In, then any allocation satisfies local envy-freeness
and proportionality. In light of this, we can pose the following problem:
Problem 2 Are there non-trivial classes of graphs for which we can give efficient protocols for
locally envy-free or at least locally proportional allocations?
We believe that, in addition to capturing real-world contexts, posing the cake cutting problem
on a network will give further insight into the structure of the original problem.
We also consider the effect of fairness on welfare as measured through the price of fairness [10].
The price of fairness is defined as the worst case ratio over all inputs between the social welfare
of the optimal allocation (the allocation maximizing the sum of agent valuations) and the social
welfare of the optimal fair allocation- the envy-free or proportional allocation that maximizes
the sum of agent valuations. We will refer to this ratio in the case of envy-free allocations as
the price of envy-freeness, and the ratio in the case of proportional allocations as the price of
proportionality. When there is an underlying graph G governing the comparisons, these ratios
become the price of local envy-freeness and the price of local proportionality for G. We pose the
following question:
2
Problem 3 How do the achievable lower and upper bounds on the price of local envy-freeness
and local proportionality depend on the structure of the graph G?
Caragiannis et al. [10] give an Ω(√n) lower bound for the price of global envy-freeness and
proportionality, a matching upper bound of O(√n) for the price of global proportionality, and a
loose upper bound of n − 1/2 for the price of global envy-freeness.
Overview of Results. With respect to Problem 1, we show that in fact the set of locally
proportional allocations do not satisfy any natural containment relations. In particular, for any
two distinct connected graphs G and H on the same set of nodes, there exists a set of valuations
for the agents and an allocation that is locally proportional for G but not for H. Note that this
includes the case where G is the complete graph, and so global proportionality does not imply
local proportionality on any other connected graph H.
For Problem 2, we start from the structure of the classical Cut-and-Choose solution for two
agents: one agent divides the cake and the other selects a piece. This solution does not provide
any guarantees for global allocations with more than two agents, but with other underlying
graphs G it turns out to have a natural and non-trivial generalization. Specifically, we fully
characterize the family of graphs G for which a locally envy-free allocation can be produced a
by a protocol in which a single designated node performs all the cuts at the outset, based only
on its valuation function.
Finally, for Problem 3, we start from the Ω(√n) lower bound on the price of global envy-
freeness, which provides a lower bound for the price of local envy-freeness in complete graphs.
We show that this Ω(√n) lower bound on the price of local envy-freeness holds in any connected,
undirected graph G. We consider this to be surprising, because one might think that sparse
graphs provide more flexibility with respect to the quality of envy-free allocations. The known
upper bound for the price of global envy-freeness serves as a loose upper bound for the price of
local proportionality and the price of local envy-freeness.
In outline, Section 2 defines local envy-freeness and local proportionality, and shows that
the former implies the latter, and that both are implied by global envy-freeness. This section
also establishes a lack of equivalence between global proportionality and either of the two local
fairness concepts. Section 3 considers the notion of protocols with a single cutter as described
above, and it characterizes the family of graphs (formally, cones of directed acyclic graphs and
their subgraphs) for which such a protocol can yield a locally envy-free allocation. Section 4
turns to the price of fairness. We conclude by giving a list of open directions that we hope
Problems 1, 2, and 3 will inspire.
1.1 Background on the Cake Cutting Problem
Cake cutting algorithms can be traced back to the Hebrew Bible. In the Book of Genesis, when
Abraham and Lot decided to separate, they were presented with the challenge of dividing up the
land. Abraham suggested to mark the place on which they stand as the cutting point and asks
Lot to pick a side. Lot chose the side that is well-watered and Abraham went in the opposite
direction. This is precisely the Cut-and-Choose Protocol from above, which has been shown to
give envy-free allocations for two agents.
The cake cutting problem has a wide range of applications including international border
settlements, divorce and inheritance settlements, and allocating shared computational resources.
The formal study of the cake cutting problem was initiated in the 1940s due to Banach et al. [7,
17, 26]. Later, Steinhaus observed that the Cut-and-Choose protocol could be extended to three
players, and asked whether it can be generalized for any number of agents [7]. This was resolved
3
affirmatively for proportional allocations by Banach et al. [26] and initiated an interesting line
of research for envy-free allocations; see Brams and Taylor [7]. See Procaccia [22, 23] for a recent
survey from a computer science perspective.
A central problem has been finding envy-free protocols. In 1988, Garfunkel [5] even called
this among the most important problems in 20th century mathematics. An envy-free protocol
for any number of agents was proposed by Brams and Taylor [6], but may need an unbounded
number of queries even for four agents. They thus posed the question whether there are bounded,
envy-free protocols for n ≥ 4. Aziz and Mackenize [3] recently provided a bounded protocol for
n = 4. The case for n > 4 remained open until recently when Aziz and Mackenize announced a
discrete and bounded protocol [4]. Although bounded, their solution has a very high multiple-
exponential query complexity and it remains an open question whether there are more efficient
protocols.
The query model was formalized by Roberton and Webb [24]. The focus in this model has
been to minimize the number of cut and eval queries. The cut query, formally cut(x, y, α), asks
an agent for a cut point x such that the subinterval [x, y] has value α to that agent and the
eval query, formally eval(x, y) asks the agent for her ver valuation of the subinterval [x, y]. It is
possible to encode all (decentralized) cake cutting protocols using this model. Note that these
queries do not result in an allocation themselves but rather provide information about the agents'
valuations in order to compute an allocation. The query complexity of a cake cutting problem
is thus the worst case number of queries required in order to output an allocation satisfying the
desired fairness criteria.
Procaccia [21] gives a lower bound of Ω(n2) for envy-free cake cutting, which shows a gap
between this and the Θ(n log n) bound known for the proportional cake cutting problem [19].
Special cases of the problem have also received attention. Kurokawa et al. [18] for example
establish that an algorithm that computes an envy-free allocation for n agents with piecewise
uniform functions using f (n) queries would also be able to compute an envy-free allocation with
general valuation functions using at most f (n) queries. Considering only contiguous allocations,
Stromquist [27] shows that for any n ≥ 3, there are no finite envy-free cake cutting algorithms,
even with an unbounded number of queries. A positive result has been achieved by considering
approximately envy-free allocations [22]. An allocation is said to be ǫ-envy-free if each agent
values the allocation of any other agent to be at most ǫ more than their own. This relaxation
provides a simple algorithm for any n using n⌈1/ǫ⌉ cut queries, as shown in [23]. Alon [1] and
Dubins and Spanier [13] also study allocations under alternate notions of fairness.
Due to the difficult nature of the envy-free cake cutting problem, researchers have imposed
restrictions on different aspects in order to give useful protocols and gain insight into the problem.
A few examples are: restricting valuation functions to be only piecewise constant or uniform [18],
relaxing envy-freeness to approximate envy-freeness [23], considering partial allocations that
simultaneously satisfy envy-freeness and proportionality [11], and limiting allocations to be con-
tiguous pieces [27].
Incentive compatibility is not a standard consideration in the cake cutting literature. One
important exception is the work of Chen et al. [11], who give a polynomial time algorithm outside
of the Robertson-Webb model for finding proportional and envy-free allocations for piecewise
uniform valuation functions and any n, while also achieving strategyproofness.
To the best of our knowledge, there is no preexisting work on the cake cutting problem in
which fairness is determined via comparisons defined by an underlying graph. In perhaps the
most closely related paper, Chevaleyre et al. [12] analyze how the network topology of negotiation
affects the convergence to an envy-free allocation for multiple indivisible goods. In their setting,
agents are only able to negotiate with, and also envy agents that are in their neighborhood.
They ask under what conditions a sequence of negotiations can lead to a state where there is
4
no envy. There are a number of differences between this work and ours: first, they consider
indivisible goods, which leads to a different set of questions; and second, in their setting the
network constrains not just the comparisons that are made in determining fairness, but also the
allowable interactions in the division protocol.
2 Relating Global and Local Properties
Let N = {1, 2,··· , n} denote the set of agents. The cake is represented using the interval [0, 1]
and a piece of cake is a finite union of non-overlapping (interior disjoint) subintervals of [0, 1].
Allocated pieces are a finite union of subintervals. Each agent i has a valuation function Vi
that maps subintervals to values in R. Given subinterval [x, y] ⊆ [0, 1], we write Vi(x, y) instead
of Vi([x, y]) for simplicity. We assume that valuation functions are additive, non-atomic, and
non-negative. Non-atomicity gives us Vi(x, x) = 0 for all x ∈ [0, 1], so we can ignore boundaries
when defining cut-points. We normalize valuations so that Vi(0, 1) = 1 for each agent i.
Definition 1 (Allocation) An allocation is a partition of the [0, 1] interval into n pieces {A1, A2, . . . , An}
such that ∪iAi = [0, 1] and the pieces are pairwise disjoint. Each agent i is assigned the corre-
sponding piece Ai.
As is standard, this ensures that the entire cake is allocated. If we remove this constraint,
then we can have trivial solutions that satisfy fairness, such as assigning each agent nothing in
the case of envy-freeness. This assumption is a natural one to make since the valuation functions
are assumed to be non-negative and additive and thus satisfy free-disposal. A feasible allocation
is one where no subinterval is assigned to more than one agent.
In order to avoid direct revelation of valuation functions, which may be cumbersome, cake
cutting procedures are typically given as protocols that interact with the agents in order to
output a feasible allocation. The Robertson-Webb query model, which is typically used for cake
cutting protocols is defined with the following two types of queries:
• evali(x, y); this asks agent i for the valuation Vi(x, y).
• cuti(x, y, α): given y, α ∈ [0, 1], this asks agent i to pick x ∈ [0, 1] such that Vi(x, y) = α.
These queries are used to gather information regarding the valuations of the agents and need
not directly determine an allocation. Rather, a cake cutting protocol can use other steps for
determining allocations.
Definition 2 (Query complexity) The query complexity of a cake cutting protocol is the
worst case number of queries that the protocol requires to output an allocation over all possi-
ble valuation functions.
The query complexity of a cake cutting problem is the minimum query complexity over all
known protocols for computing the desired allocation.
2.1 Global and Local Fairness
Given a set of agents and an allocation A = (A1, A2,··· , An), we formally define two global
fairness criteria:
Definition 3 (Proportional, Envy-free) An allocation A is proportional if Vi(Ai) ≥ 1/n,
for all i ∈ N , and is envy-free if Vi(Ai) ≥ Vi(Aj ), for all i, j ∈ N .
5
Suppose we are given a directed graph G = (V, E), where the nodes correspond to agents
and edges signify relations between the agents. In particular, we assume that given a directed
edge (i, j), agent i can view agent j's allocation. Agent i's neighborhood is the set of all nodes
to which it has directed edges (i, j), and we denote this set of nodes by Ni. We define i's degree
to be di = Ni. We define local analogues for fairness concepts:
Definition 4 (Local proportional, local envy free) Given a graph G, an allocation A is
for all i and j ∈ Ni and locally envy-free if
locally proportional if Vi(Ai) ≥
Vi(Ai) ≥ Vi(Aj ).
Pj∈Ni Vi(Aj )
Ni
In a locally proportional allocation, each agent assigns as much value to her allocation as the
average value she has for a neighbors' allocation. In a locally envy-free allocation, each agent
values her allocation at least as much as her neighbors' allocation.
When G = Kn, the complete graph on n vertices, these local fairness definitions coincide with
their global analogues. Whereas, if G = In, the empty graph on n nodes, then any allocation
is trivially locally envy-free. So, the graph topology plays a significant role in computing locally
fair allocations.
Lemma 2.1 A locally envy-free allocation A on some graph G is also locally envy-free on all
subgraphs G′ ⊆ G.
Proof We want to show that given a node u and v ∈ Nu, u does not envy v's allocation in G′.
This follows from the fact that A is a locally envy-free allocation on G, and if (u, v) is an edge
in G′, then it is also an edge in G.
One consequence of this lemma is that local envy-freeness is implied by global envy-freeness.
Since globally envy-free allocations exist for all sets of agent valuations [1], a locally envy-free
allocation exists for every graph G and every set of agent valuations.
Lemma 2.2 If an allocation A is locally envy-free on a graph G, then it is also locally propor-
tional on the same graph.
Proof If an allocation A = (A1, A2, . . . , An) is locally envy-free, then for any i ∈ V , Vi(Ai) ≥
Vi(Aj ),∀j ∈ Ni. Therefore, Vi(Ai) ≥ (Pj∈Ni Vi(Aj))/Ni.
Therefore, locally proportional allocations also exist. By considering G = Kn, we also recover
that global envy-freeness implies global proportionality. While global envy-freeness implies local
envy-freeness, global proportionality does not necessarily imply local proportionality, or vice
versa, the former of which violates intuition. We provide a counter example.
Example 1 Let n = 4 and G = C4, the cycle graph on 4 nodes, where the nodes are labeled
clockwise. Assume agents 2, 3, and 4 have the uniform valuation function Vi(x, y) = y − x for
any subinterval (x, y) ⊆ [0, 1]. Let agent 1 have the piecewise uniform valuation function where
V1 (0, 1/4) = 1/2 and V1 (3/4, 1) = 1/2, and no value for the remaining subinterval. It is easy to
verify that the following allocation is locally proportional on K4,
A = ([0, 1/8), [1/8, 3/8), [3/8, 5/8), [5/8, 1]) .
In particular, Vi(Ai) = 1/4 for i ∈ {1, 2, 3} and V4(A4) = 3/8. This allocation is however not
locally proportional on C4 since V1(A1 ∪ A2 ∪ A4) = 1, but V1(0, 1/8) = 1/4 < 1/3. It is also not
locally envy-free since V1(A4) > V1(A1).
6
Global
Envy-freeness
Local
Envy-freeness
Global
Proportionality
Local
Proportionality
Figure 1: Relationship Between Fairness Concepts
We prove a stronger result regarding any pair of distinct graphs. Note by Ni(H) we mean
agent i's neighborhood set in graph H.
Theorem 2.3 Given any pair of distinct, connected graphs G, H on the same set of nodes, there
exists a valuation profile of the agents and an allocation A such that A is locally proportional on
G but not on H.
Proof First, consider the case where H is a strict subgraph of G. Pick a node i such that
Ni(H) < Ni(G). Let Ni(G) = {i1, i2,··· , ik} and Ni(H) = {i1, i2,··· , iℓ} for some ℓ < k.
Assume that all other nodes besides i have a uniform valuation function over the entire cake.
Then, the allocation Aj = ((j − 1)/n, j/n) is locally proportional from the perspective of every
other agent j ∈ N on both H and G. Now, define i's valuation function to be the piecewise
uniform valuation function where, Vi ((i − 1)/n, i/n) = 1/(Ni(G) + 1) and Vi(Ai1 ∪ Ai2 ··· ∪
Aiℓ ) = 1 − 1/(Ni(G) + 1). Agent i's valuation for the allocation of nodes {iℓ+1, iℓ+2,··· , ik} as
well as V (G)\Ni(G) is 0. This allocation A is therefore locally proportional on G. For A to be
locally proportional on H, we need Vi(Ai) ≥ 1/Ni(H). However, since Ni(H) < Ni(G), and
Vi(Ai ∪ Ai1 ∪ Ai2 ∪ ··· ∪ Aiℓ ) = 1, we only have that Vi(Ai) < 1/(Ni(H) + 1).
Now, suppose H * G. Then, there exists an edge (i, j) in the edge-set of H that is not in the
edge-set of G. Assume that all nodes k 6= i have a uniform valuation over the entire cake. Suppose
further that we have the allocation where each k is assigned the piece Ak = ((k − 1)/n, k/n).
As above, this allocation is locally proportional from the perspective of each agent k on both G
and H. Define i's valuation function to be Vi ((j − 1)/n, j/n) = 1 and 0 on the remainder of
the cake. Then, Vi(Ai) = 0 and Vi(Ak) = 0 for all k 6= j. Since j /∈ Ni(G), this allocation is
locally proportional on G. However, it is not locally proportional on H since Vi (∪ℓ∈NiAℓ) = 1,
but Vi(Ai) = 0.
3 Envy-Free Network Allocations
In this section, we consider the question of finding efficient protocols for computing locally envy-
free allocations. We assume that graphs G are directed, unless specified otherwise. When we
mean the component of a directed graph, we will instead take the graph obtained by replacing each
directed edge with an undirected one, and a component in the directed graph is the corresponding
subgraph to the connected component in the undirected analogue. We will use strongly connected
component when we mean to take directed reachability into account.
Lemma 3.1 Suppose we have a bounded protocol for computing locally envy-free allocations on
G. The same protocol can be used to compute locally envy-free allocations on the following two
7
classes of graphs: (i) H = G ∪ G′ where G and G′ are disjoint components, and (ii) when H is
a graph with a directed cut such that every edge across the cut goes from a node in G to a node
in H\G = G′.
Proof In both instances, we simply apply the protocol on G and allocate agents in G′ the empty
allocation. This is a locally envy-free allocation on H, since no two agents in G envy one another
by the assumption on the protocol, and no agent envies the allocation of another agent in G′.
A few consequences of this lemma are that: given a graph G with more than one connected
component, we can reduce the search for a bounded protocol on G to any one of the connected
components. Furthermore, if G is a directed acyclic graph (DAG), then there exists at least one
node with no incoming edges. Therefore, the allocation where such a node gets the entire cake-
or where it is divided among a set of such nodes -is locally envy-free.
3.1 Directed Acyclic Graphs and Their Cones
We consider a conceptually useful class of graphs for which we can give a protocol with query
complexity of O(n2).
Definition Given a graph G = (V, E), we say that G′ is a cone of G if it is the join of G and a
single node c, which we call the apex. That is, G′ has node set V ∪ {c}, and edge set consisting
of the edges of G, together with undirected edges (u, c) for all u ∈ V . We denote the cone G′ of
G by G ⋆ c.
We consider cones of DAGs. These are the class of graphs where there is a single node c that
lies on all cycles. We now show how to compute a locally envy-free allocation on any graph that
is the cone of a DAG.
c
2
3
n-1
n
Figure 2: Cone of a Directed Acyclic Graph
Protocol 1: Cone of DAGs
1: Agent c cuts the cake into n pieces that she values equally.
2: Topologically sort and label the nodes N\{c} such that for every edge (i, j), i ≤ j.
3: Nodes N\{c} pick a piece they prefer most in increasing order of their index.
4: Agent c takes the remaining piece.
Theorem 3.2 Given a graph G that is a cone of a DAG, Protocol 1 computes a locally envy-free
allocation on G using a bounded number of queries.
Proof We first show that the allocation is locally envy-free. First, there is no envy between agent
c and any other agent since agent c cuts the cake into n pieces she values equally. Therefore, her
8
valuation for all the allocated pieces is 1/n. Each of the other agents picks a piece before c, and
so are able to pick a piece that they value at least as much as the remaining piece that agent c
is assigned. Finally, given any directed edge (i, j) such that i, j 6= c, note that Vi(Ai) ≥ Vi(Aj ),
since if such an edge exists, then i < j and thus i selects a piece before j.
To count the number of queries, the first step requires n − 1 cut queries by agent c. Then,
each agent i must perform n − i + 2 eval queries to determine the piece for which they have the
highest value. Therefore, the protocol above uses (n2 + 3n − 4)/2 queries.
The importance of cones of DAGs can be seen in the following result, which shows that they
emerge naturally as the characterization of graphs on which a particular fundamental kind of
protocol succeeds.
Definition 5 An oblivious single-cutter protocol is one in which a single agent i first divides up
the cake into a set of pieces P1, P2, . . . , Pt (potentially t > n), and then all remaining operations
consist of agents choosing from among these pieces.
The classical Cut-and-Choose protocol is an oblivious single-cutter protocol that works for all
sets of valuation functions on the complete, two-node graph K2. Protocol 1 is an oblivious single-
cutter protocol which works for any graph that is the cone of a DAG. We show that subgraphs
of cones of DAGs are in fact precisely the graphs on which oblivious single-cutter protocols are
guaranteed to produce a locally envy-free allocation.
Theorem 3.3 If G is a graph for which an oblivious single-cutter protocol produces a locally
envy-free allocation for all sets of valuation functions, then G is a subgraph of the cone of a
DAG.
Proof Suppose, by way of contradiction, that G is not a subgraph of the cone of a DAG, but that
there is an oblivious single-cutter protocol on G, in which a node i starts by dividing the cake
into pieces using only knowledge of her own valuation function. Since G is not a subgraph of the
cone of a DAG, the graph Gi = G\{i} is not acyclic, so there is a cycle C = (c1, c2,··· , cm, c1)
in Gv.
Let i have a valuation function such that she produces a partition of the cake into pieces
P1, P2, . . . , Pt. Because agent i produces these pieces without knowledge of the valuation func-
tions of the other agents, we can imagine that we adversarially choose the valuations of the other
agents after these pieces have been produced. In particular, consider the valuation functions in
which each node cj on the cycle C values piece Pr (for r < t) at 2 · 3−r, and the last piece Pt
with the remaining value. These valuations have the property that for each r, the piece Pr is
more valuable than the union of all pieces Pr+1 ∪ Pr+2 ∪ ··· ∪ Pt.
After the protocol is run, each agent cj on C will get a subset of the pieces produced by i.
Let s be the minimum index of any piece allocated to an agent cj on C. Then, the agent cj−1
who has a directed edge to cj will have a union of pieces that she values less than she would value
Ps, and hence envies cj. This contradicts the assumption that the protocol produces a locally
envy-free allocation on G.
We highlight an important connection between computing locally envy-free allocations on
graphs and what is known in the literature as irrevocable advantage. Given a partial allocation,
an agent i is said to have irrevocable advantage over agent j (or dominate j) if agent i remains
unenvious of agent j's allocation even if the entire remaining piece of the cake (the residue) is
added to agent j's allocation.
With the additional guarantee that each agent dominates some number of other agents, this
concept is often used to extend partial globally envy-free allocations to complete ones. For
9
instance, it is a key concept in the Aziz-Mackenize protocol for K4. Their protocol can be
decomposed to three subprotocols: Core, Permutation, and Post-Double Domination Protocols,
in order. The Core Protocol computes partial envy-free allocation where each agent dominates
at least two other agents, while the Post Double Domination Protocol extends this to a complete
allocation. We will use Protocol 1 to show that given a partial envy-free allocation on Kn where
each agent dominates at least n − 2 other agents, we can extend the allocation to a complete
one, thereby generalizing the Post Double Domination Protocol [3] for any n. This is presented
in Appendix A.
4 Price of Fairness
Finally, we consider the efficiency of allocation from the perspective of local fairness. We follow
the approach introduced by Caragiannis et al. [10] of studying the price of envy-freeness, and
for this we begin with the following definitions. Recall that for an allocation A into pieces
{A1, A2, . . . , An} for the n agents, we use Vi(Ai) to denote agent i's valuation for its piece.
Definition 6 (Optimality) An allocation A, is said to be optimal if Pi Vi(Ai) ≥ Pi Vi(Bi)
for any allocation B. We denote this optimal allocation by A∗.
We define the optimal locally envy-free (resp. optimal locally proportional ) allocations, de-
), analogously by imposing the constraint that A and B be locally
noted by ALEF∗
envy-free (resp. locally proportional) and maximizing sum of the values across all agents.
(resp. ALP∗
Definition 7 (Price of Local Envy-Freeness, Proportionality) Given an instance of a cake
cutting problem on a graph G, the price of local envy-freeness is the ratio,
Pi Vi(A∗i )
Pi Vi (cid:0)ALEF∗
i
,
(cid:1)
where the sum is over all agents i ∈ N. We likewise define the price of proportionality by taking
the denominator to be Pi Vi (cid:0)ALP∗
(cid:1).
i
We are measuring the degradation in efficiency when considering allocations that maximize
the welfare in both instances under the given constraints. To quantify the loss of efficiency,
we are interested in giving a tight lower and upper bound. More specifically, given a graph G
and a fairness concept in consideration, say local envy-freeness, we seek to find an input (i.e.,
a valuation profile) for which the price of local envy-freeness is high. This corresponds to a
lower bound on the price of fairness. On the other hand, the upper bound will be given via an
argument that shows, for any valuation profile, the price of fairness cannot exceed that stated.
The main result on global envy-freeness, due to Caragiannis et al. [10] is an Ω(√n) lower
bound on the price of (global) fairness: there exist valuation functions for which the ratio is
Ω(√n). (Very little is know about the upper bound for the price of envy-freeness: an upper
bound of n is immediate, and the best known upper bound is n − 1/2 [10].)
These existing results are for the standard model in which each agent can envy every other
agent. Using our graph-theoretic formulation, we can study the price of local fairness. As defined
in Section 2, this is the ratio of the total welfare of the optimal allocation to the maximum total
welfare of any allocation that is locally envy-free.
The numerator of this ratio- based on the optimal allocation -is independent of G, while
the denominator is a maximum over a set of allocations that is constrained by G. Now, if we
imagine reducing the set of edges in G, the set of allocations eligible for the maximum in the
10
denominator becomes less constrained; consequently, we would expect that the price of fairness
may become significantly smaller as G becomes sparser. Is this in fact the case? We show that
it is not. Our main result is that the lower bound for global envy-freeness also applies to local
envy-freeness on any connected undirected graph.
Theorem 4.1 For any connected undirected graph G, there exist valuation functions for which
the price of local envy-freeness on G is Ω(√n).
To prove this theorem, we start by adapting a set of valuation functions that Caragiannis
et al [10] used in their lower bound for global envy-freeness. To argue about their effect on
allocations in an arbitrary graph G, we need to reason about the paths connecting agents in G
to others with different valuation functions. This, in turn, requires a delicate graph-theoretic
definition and argument: we introduce a structure that we term a (k, ε)-linked partition; we show
that if G contains this structure, then we can carry out the lower bound argument in G; and
finally we show that every connected undirected graph contains a (k, ε)-linked partition.
Definition For a connected graph G = (V, E), a natural number k ≥ 1, and a real number
0 < ε ≤ 1, we define a (k, ε)-linked partition as follows. It consists of a set L ⊆ V of size k, and
a partition of S = V − L into sets {Si : i ∈ L} each of size at least (εn/k) − 1, such that for
each j ∈ Si, there is an i-j path in S ∪ {i}. That is, each j ∈ Si can reach i ∈ L without passing
through any other nodes of L.
We next show that if a connected undirected graph G has such a structure, with appropriate
values of k and ε, then we obtain a lower bound on the price of local envy-freeness on G.
Lemma 4.2 If a connected undirected graph G has a (k, ε)-linked partition with k = ⌊√n⌋ and
ε, a constant, then there exist valuation functions on the nodes of G for which the price of local
envy-freeness is Ω(√n).
Proof Suppose G has a (k, ε)-linked partition consisting of L and {Si : i ∈ L}, where k = √n.
Thus, each Si has size at least ε√n − 1. (We will assume for the sake of exposition that √n is
an integer, although it is straightforward to slightly modify the argument if it is not.)
We now use a valuation function adapted from the construction of Caragiannis et al [10], who
considered the price of global envy-freeness. We partition the full resource to be allocated, the
interval [0, 1], into √n disjoint intervals I1, . . . , I√n. We will give each i ∈ L a valuation vi that
places all value on distinct interval Ii, and each j ∈ S a valuation that is uniform on [0, 1]. The
optimal allocation for this set of valuations has total welfare of √n, which is achieved by giving
each i ∈ L the entire interval where it places value.
Now let us consider any envy-free allocation A = {Ai : i ∈ V }. Let µi be a real number
denoting the Lebesgue measure of the set Ai assigned to node i. If j ∈ S, then j's valuation
for its set, vj (Aj) is equal to µj. If i ∈ L, then i's valuation vi(Ai) is √n times the measure of
Ai ∩ Ii; hence vi(Ai) ≤ µi√n.
For each i ∈ L, each j ∈ Si has a path Pj to i entirely through nodes of S; let the nodes
on this path, beginning at i, be Pj = i, j1, j2, . . . , jd = j. The immediate neighbor j1 of i on Pj
must satisfy µj1 ≥ µi, since j1 is in S and hence has a uniform valuation on intervals. For each
successive jt on P , we must have µjt = µjt−1 , since jt and jt−1 have the same valuation on all
sets, and the allocation is locally envy-free. Thus, by induction we have µjt ≥ µi for all t, and
hence µj ≥ µi.
We can now derive a set of inequalities that establishes the lower bound. First we have,
X
j∈Si
vj(Aj ) = X
j∈Si
µj ≥ X
j∈Si
µi ≥ µi(ε√n − 1).
11
Let us assume n is large enough that ε√n − 1 ≥ ε√n/2, so we have,
vj(Aj) ≥ µi(ε√n/2).
X
j∈Si
Since vi(Ai) ≤ µi√n for i ∈ L, we have,
X
j∈Si
vj(Aj ) ≥ εvi(Ai)/2.
(1)
Thus, the total welfare of the allocation is,
X
h∈V
vj(Aj )
vj(Aj )]
vi(Ai) + X
j∈S
[vi(Ai) + X
j∈Si
vh(Ah) = X
i∈L
= X
i∈L
≤ X
i∈L
= (2ε−1 + 1) X
i∈L
= (2ε−1 + 1) X
j∈S
vj(Aj )]
[(2ε−1 + 1) X
j∈Si
X
j∈Si
vj(Aj ) ≤ (2ε−1 + 1),
vj(Aj)
where the first inequality is by (1) and the second since
X
j∈S
vj (Aj) = X
j∈S
µj ≤ 1,
because all agents in S get disjoint intervals. Since (2ε−1 + 1) is a constant, while the optimal
allocation has total welfare √n, this implies an Ω(√n) lower bound for the price of envy-freeness
on G.
Finally, we establish that every connected undirected graph G has a (k, ε)-linked partition
for appropriate values of k and ε. We begin by showing that it is enough to find a structure
satisfying a slightly more relaxed definition, in which the set L can have more than k elements,
and we do not need to include all the nodes of G. Specifically, we have the following definition:
Definition For a connected graph G = (V, E), a natural number k ≥ 1, and a real number
0 < ε ≤ 1, we define a (k, ε)-linked subpartition as follows. It consists of a set L ⊆ V of size
ℓ ≥ k, together with disjoint subsets S1, S2, . . . , Sℓ ⊆ S = V − L, each of size of size at least
(εn/k) − 1, such that for each j ∈ Si, there is an i-j path in S ∪ {i}.
The following lemma says that it is sufficient to find a (k, ε)-linked subpartition.
Lemma 4.3 If a connected undirected graph G contains a (k, ε)-linked subpartition, then it con-
tains a (k, ε)-linked partition.
Proof We start with a (k, ε)-linked subpartition of G, with disjoint sets L of size ℓ ≥ k, and
S1, S2, . . . , Sℓ ⊆ S = V − L. First, for every node v 6∈ L∪ S, we assign it to a subset Si as follows:
we find the shortest path from v to any node in L; suppose it is to i ∈ L. We add v to Si. Note
12
that this preserves the property that all Si are disjoint, and v has a path to i that does not meet
any other node of L, since if h ∈ L were to lie on this path, it would be closer to v than i is.
At this point, every node of G belongs to L∪ S. We now must remove nodes from L to reduce
its size to exactly k while preserving the properties of a (k, ε)-linked partition. To do this, we
choose a node i ∈ L arbitrary, remove i from L, and remove the set Si from the collection of
subsets. We then assign each node in Si ∪{i} to an existing subset Sh exactly as in the previous
paragraph. After this process, the size of L has been reduced by 1, and we still have a partition
of V − L into subsets Si with the desired properties. Continuing in this way, we can reduce the
size of L to exactly k, at which point we have a (k, ε)-linked partition.
Finally, we prove the following graph-theoretic result, which together with Lemma 4.2 estab-
lishes Theorem 4.1.
Theorem 4.4 For every k ≥ 2 and with ε = 1/2, every connected undirected graph has a (k, ε)-
linked subpartition.
Proof It is enough to find the required structure on a spanning tree T of G, since if the paths
required by the definition exist in T , then they also exist in G. Thus, it is sufficient to prove the
result for an arbitrary tree T .
We root T at an arbitrary node, and let X be the set of leaves of T . If X ≥ k, then we
can choose any k leaves of T and partition the remaining nodes of T arbitrarily into sets of size
(n − k)/k to satisfy the definition. Otherwise, X < k. In this case, we begin by including all
nodes of X in L.
Now, we process the nodes of T , working upward from the leaves, so that when we get to a
node v in T , we have already processed all descendents of v. Each node is processed once, and
at that point we decide whether to add it to L, and if not which set Si to place it in, given the
current set L. For a node v, we say that w is downward-reachable from v if w is a descendent of
v, and if the v-w path in T does not contain any internal nodes belonging to L.
When we process a node v, we do one of two things:
(i) We label v with the name of a node in L that is downward-reachable from v; or
(ii) We place v in L.
Let b = (εn/k) − 1. We perform action (i) if there is any w ∈ L that is downward-reachable
from v, such that there are not yet b nodes labeled with w. In this case, we label v arbitrarily
with one such w. Since v and all its descendents are now processed, v will continue to have a
path to w that does not pass through any other nodes of L. Otherwise suppose there is no such
w; that is, all w ∈ L that are downward-reachable from v have b nodes labeled with w. In this
case, we perform action (ii). Note that at this point, every w ∈ L that is a descendent of v has
a set Sw of exactly b nodes, and these nodes can all reach w without passing through any other
node of L.
Our procedure comes to an end when we process the root node v∗. There are three cases to
consider, the first two of which are straightforward.
First, if we place v∗ into L, then T − v∗ is partitioned into L and sets {Sw : w ∈ L} such that
Sw = b for each w. Thus, if we remove v∗ from L, we have a (k, ε)-linked subpartition, since
the sets Sw are disjoint and of size at least b, and
L = (n − 1)/((εn/k) − 1) ≥ (n/(εn/k)) = k/ε > k.
Otherwise, v∗ is labeled with some downward-reachable u ∈ L. Our second case, which is
also straightforward, is that after this labeling of the root, all sets Sw for w ∈ L have size exactly
b, then we have a (k, ε)-linked subpartition.
13
L1 + X
w∈L1
Sw < k + kb = k(b + 1) = kn/(2k) = n/2.
If not, then we are in the third case: v∗ is labeled with some downward-reachable u ∈ L, and
after this labeling there still exist downward-reachable nodes w that we have placed in L that
do not have associated sets Sw of size b. We therefore need to prune our set L to a smaller set
that has Sw ≥ b for each w ∈ L. The goal is to show that the smaller L we end up with still
has enough elements; if that holds, then we have a (k, ε)-linked subpartition.
To show this, we proceed as follows. We say that w ∈ L is active if Sw < b. We first observe
that any active w must be downward-reachable from the root v∗. Indeed, if w is active and not
downward-reachable from the root, then there is a v ∈ L such that w is a descendent of v. But
in the step when we placed v in L, it was not possible to label v with w, and hence we must have
had Sw = b at that point.
Next we claim that there are < k active w ∈ L. To prove this, for each active w, we associate
w with a leaf that is a descendent of w. (This can be w itself if w is a leaf.) Observe that the
same leaf x cannot be associated with two distinct active w, w′, for then on the path from v∗ to
x, one of w or w′ would be closer to v∗, and the other would not be downward-reachable from
v∗. Given that we can associate a distinct leaf to each active w, and there are < k leaves, there
are < k active w ∈ L.
inactive nodes of L and L1 be the active nodes of L. We have
We say that a node w ∈ L is inactive if it is not active; that is, if Sw = b. Let L0 be the
L0 + X
w∈L0
Sw + L1 + X
w∈L1
Sw = n.
We know that L1 < k and Sw < b for each w ∈ L1; hence, using the fact that ε = 1/2, we
have
It follows that L0 + Pw∈L0 Sw > n/2. But since Sw = b for each w ∈ L0, we have
n/2 < L0 + X
w∈L0
Sw = L0(1 + b) = L0n/(2k),
from which it follows that L0 > k. We now conclude the construction by declaring L to be L0;
since the sets Sw for w ∈ L0 are all pairwise disjoint and each has size at least b, we have the
desired (k, ε)-linked subpartition.
5 Conclusion and Future Work
We have introduced a new line of inquiry for the envy-free and proportional cake cutting problems
by considering local notions of fairness. We show interesting relations between these local fairness
concepts and their global analogues. Besides introducing this new model, our main contribution
has been to fully classify the class of graphs for which there is an oblivious single-cutter protocol
for computing locally envy-free allocations. Furthermore, we quantify the degredation in welfare
resulting from adding the local envy-freeness constraint on the allocations; in particular, we show
that the known Ω(√n) lower-bound for the (global) price of envy-freeness continues to hold even
for sparse graphs.
It is of interest to give efficient protocols for computing locally envy-free allocations on rich
classes of graphs without the single-cutter constraint. Since local envy-freeness is a stronger
condition than local proportionality, the same problem can also be considered for locally propor-
tional allocations. Finally, whether there is a similar lower bound of Ω(√n) for the price of local
proportionality is an open question. Currently, the upper bound for both local fairness concepts
is the loose n − 1/2 bound, and giving tighter bounds is another direction.
14
Acknowledgements
The first author would like to thank Felix Fischer for the introduction to the cake cutting problem.
We would also like to thank Rahmtin Rotabi for insights for the proof of Theorem 4.4 as well
as Barabra Grosz, Radhika Nagpal, Ariel Procaccia, Ofra Amir, and Adam Brauer for helpful
discussions and references.
References
[1] Noga Alon. Splitting necklaces. Advances in Mathematics, 63(3):247–53, 1987.
[2] George Akerlof. Social distance and social decisions. Econometrica, 65(2):1005-1028, 1997.
[3] Haris Aziz and Simon Mackenzie. A discrete and bounded envy-free cake cutting protocol
for four agents. CoRR, abs/1508.05143, August 2015.
[4] Haris Aziz and Simon Mackenzie. A discrete and bounded envy-free cake cutting protocol
for any agents. CoRR, abs/1604.03655, April 2016.
[5] Joseph Blatt, Stephanie Stewart, and Sol Garfunkel. More equal than others: Weighted
voting ; zero sum games: Games of conflict. Technical report, Annenberg/CPB Project,
1986.
[6] Steven Brams and Alan Taylor. An envy-free cake division protocol. American Mathematical
Monthly, 102(1):9–18, 1995.
[7] Steven Brams and Alan Taylor. Fair Division: From Cake Cutting to Dispute Resolution.
Cambridge University Press, Cambridge, 1996.
[8] Simina Brazei, Ariel Procaccia, and Jie Zhang. Externalities in cake cutting. Proceedings
of the 23rd International Joint Conference on Artificial Intelligence, IJCAI'12, pages 55–61,
Beiging, China, 2013.
[9] Ronald Burt.The shadow of other people: socialization and social comparison in marketing.
In S. Wuyt, M. Dekimpe, E. Gijsbrechts, and R. Peters, editors, The Connected Customer.
Taylor and Francis, 2010.
[10] Ioannis Caragiannis, Christis Kaklamanis, Panagiotis Kanellopoulos, and Maria Kyropoulou.
The efficiency of fair division. Theory of Computing Systems, 50, May 2012.
[11] Yiling Chen, John Lai, David Parkes, and Ariel Procaccia. Truth, justice, and cake cutting.
Games and Economic Behavior, 77, January 2013.
[12] Yann Chevaleyre, Ulle Endriss, and Nicolas Maudet. Allocating goods on a graph to elimi-
nate envy. In In Proceedings of AAAI-07, pages 700–705, 2007.
[13] Lester Dubins and E.H. Spanier. how to cut a cake fairly. American Mathematical Monthly,
68(1):1–17, 1961.
[14] S. Evan and A. Paz. A note on cake cutting. Discrete Applied Mathematics, 7(3):285–96,
1984.
[15] Leon Festinger. A theory of social comparison process. Human Relations,7(2), 1954.
15
[16] Michael Kearns. Graphical games. In Noam Nisan, Tim Roughgarden, ´Eva Tardos, and
Vijay Vazirani, editors, Algorithmic Game Theory, chapter 7. Cambridge University Press,
2007.
[17] B. Knaster. Sur le probleme du partage pragmatique de h. steinhaus. Annales de la Societe
Polonaise de Mathematique, 19:228–230, 1946.
[18] David Kurokawa, John Lai, and Ariel Procaccia. How to cut a cake before the party ends.
In AAAI Conference on Artificial Intelligence, 2013.
[19] Malik Magdon-Ismail, Costas Busch, and Mukkai Krishnamoorthy. Cake-cutting is not a
piece of cake. In Proc. 20th Annual Symposium on Theoretical Aspects of Computer Science,
pages 596–607, 2003.
[20] Michael McBridge.Relative-income effects on subjective well-being in the cross section.Jour-
nal of Economic Behavior and Organization, 45(3): 251-278, 2001.
[21] Ariel Procaccia. Thou shalt covet thy neighbor's cake. In Proceedings of the 21st Interna-
tional Joint Conference on Artifical Intelligence, IJCAI'09, pages 239–244, San Francisco,
CA, USA, 2009. Morgan Kaufmann Publishers Inc.
[22] Ariel Procaccia. Cake cutting. Communications of the ACM, 56(7), 2013.
[23] Ariel Procaccia. Cake cutting algorithms. In F. Brandt, V. Conitzer, U. Endriss, J. Lang, and
A.D. Procaccia, editors, Handbook of Computational Social Choice. Cambridge University
Press, 2016.
[24] Jack Robertson and William Webb. Cake Cutting Algorithms: Be Fair If You Can. A.K.
Peters Ltd., London, 1998.
[25] Erel Segal-Halevi, Avinatan Hassidim, and Yonatan Aumann. Waste makes haste: bounded
time algorithms for envy-free cake cutting with free disposal. CoRR, abs/1511.02599, De-
cember 2015.
[26] Hugo Steinhaus. The problem of fair division. Econometrica, 16:101–104, 1948.
[27] Walter Stromquist. Envy-free cake divisions cannot be found by finite protocols. The
Electronic Journal of Combinatorics, 15(R11), 2008.
16
A Taking Advantage of Irrevocable Advantage
Given a partial envy-free allocation, an agent i is said to dominate an agent j, if i remains
envy-free of j even if the entire residue (the remaining subinterval of the cake) is allocated to j.
We can thus define:
Definition 8 A domination graph on n is a graph where V is the set of agents and there is a
directed edge (i, j) if i has irrevocable advantage over j.
Constraints on the number of agents each agent must dominate at a certain stage in the
protocol can be used to extend partial envy-free allocations to complete ones. One salient example
is the Aziz-Mackenzie protocol for K4, where they obtain a partial envy-free allocation using what
they call the Core Protocol and the Permutation Protocol such that each agent is guaranteed to
dominate at least two other agents. They then use the Post-Double Domination Protocol to
extend this to a complete envy-free allocation. We generalize this protocol to any n. That is,
given a partial envy-free allocation such that each agent dominates n − 2 other agents, we can
apply Protocol 1 to extend the allocation to a complete allocation. This provides an alternate
proof to a recent paper by Segal-Halevi et al. [25]. We first define a special class of graphs.
Definition 9 A pseudoforest is a graph where each vertex has at most one outgoing edge.
Each component of a pseudoforest is a subgraph of a cone of a DAG; since each node has at
most one outgoing edge, there are at most n edges. If there are fewer than n, then it is a DAG.
If there are exactly n, then there exists a cycle. Find this cycle and remove an edge e = (u, v)
from the cycle. The resulting graph will be a DAG, and we can therefore Protocol 1 by setting
u = c. This makes the protocol key to extending partial envy-free allocations to complete ones
under particular domination criteria.
Lemma A.1 We can extend a partial globally envy-free allocation in which each agent dominates
at least n − 2 other agents to a complete, envy-free allocation.
Proof Suppose we have a partial globally envy-free allocation (P1, P2,··· , Pn), with residue R.
If each agent dominates at least n−2 other agents, then the complement of the domination graph,
denoted by Gc, is a pseudoforest. We apply Protocol 1 on Gc using residue R and denote this
allocation by (R1, R2,··· , Rn). We want to show that (P1∪R1, P2∪R2,··· , Pn∪Rn) is a globally
envy-free allocation. Suppose it is not. Then, there exists i, j such that Vi(Pi∪Ri) < Vi(Pj ∪Rj ),
but this is only possible if either Vi(Pi) < Vi(Pj ) or Vi(Ri) < Vi(Rj ).
The assumption that each agent dominates at least n − 2 agents is necessary. In particular,
suppose that there exists one agent that dominates only n− 3 other agents, such as in Example 2.
1
2
3
n-2
n-1
n
Figure 3: Counterexample to the extension lemma.
17
Example 2 Suppose that each agent i 6= 2 dominates every other agent but agents i + 1 mod n
and that agent 2 dominates agents {4, 5,··· , n}. The domination graph is given in Figure 3.
The complement of the domination graph is the cycle graph (1, 2, 3,··· , n, 1) plus the edge (2, 1).
It therefore consists of more than one simple cycle, and hence a direct application of Protocol
1 to the complement of the domination graph will not extend a partial allocation to a complete
allocation.
18
|
1102.1273 | 1 | 1102 | 2011-02-07T11:09:34 | One to Rule Them All: a General Randomized Algorithm for Buffer Management with Bounded Delay | [
"cs.DS"
] | We give a memoryless scale-invariant randomized algorithm for the Buffer Management with Bounded Delay problem that is e/(e-1)-competitive against an adaptive adversary, together with better performance guarantees for many restricted variants, including the s-bounded instances. In particular, our algorithm attains the optimum competitive ratio of 4/3 on 2-bounded instances.
Both the algorithm and its analysis are applicable to a more general problem, called Collecting Items, in which only the relative order between packets' deadlines is known. Our algorithm is the optimal randomized memoryless algorithm against adaptive adversary for that problem in a strong sense.
While some of provided upper bounds were already known, in general, they were attained by several different algorithms. | cs.DS | cs |
One to Rule Them All:
A General Randomized Algorithm for Buffer
Management with Bounded Delay
Lukasz Jez∗
August 28, 2018
Abstract
We give a memoryless scale-invariant randomized algorithm Mix-R
for buffer management with bounded delay that is e/(e − 1)-competitive
against an adaptive adversary, together with better performance guar-
antees for many restricted variants, including the s-bounded instances.
In particular, Mix-R attains the optimum competitive ratio of 4/3 on
2-bounded instances.
Both Mix-R and its analysis are applicable to a more general problem,
called Collecting Items, in which only the relative order between packets'
deadlines is known. Mix-R is the optimal memoryless randomized algo-
rithm against adaptive adversary for that problem in a strong sense.
While some of provided upper bounds were already known, in general,
they were attained by several different algorithms.
1
Introduction
In this paper, we consider the problem of buffer management with bounded
delay, introduced by Kesselman et al. [16]. This problem models the behavior of
a single network switch responsible for scheduling packet transmissions along an
outgoing link as follows. We assume that time is divided into unit-length steps.
At the beginning of a time step, any number of packets may arrive at a switch
and be stored in its buffer. Each packet has a positive weight, corresponding
to the packets priority, and a deadline, which specifies the latest time when the
packet can be transmitted. Only one packet from the buffer can be transmitted
in a single step. A packet is removed from the buffer upon transmission or
expiration, i.e., reaching its deadline. The goal is to maximize the gain, defined
as the total weight of the packets transmitted.
We note that buffer management with bounded delay is equivalent to a schedul-
ing problem in which packets are represented as jobs of unit length, with given
release times, deadlines and weights; release times and deadlines are restricted
to integer values. In this setting, the goal is to maximize the total weight of
jobs completed before their respective deadlines.
∗University of Wroc law. Work supported by MNiSW grant N N206 490638, 2010 -- 2011.
1
As the process of managing packet queue is inherently a real-time task, we
model it as an online problem. This means that the algorithm, when deciding
which packets to transmit, has to base its decision solely on the packets which
have already arrived at a switch, without the knowledge of the future.
1.1 Competitive Analysis
To measure the performance of an online algorithm, we use the standard notion
of competitive analysis [6], which, roughly speaking, compares the gain of the
algorithm to the gain of the optimal solution on the same instance. For any
algorithm Alg, we denote its gain on instance I by GAlg(I). The optimal
offline algorithm is denoted by Opt. We say that a deterministic algorithm
Alg is R-competitive if on any instance I it holds that GAlg(I) ≥ 1
R · GOpt(I).
When analyzing the performance of an online algorithm Alg, we view the
process as a game between Alg and an adversary. The adversary controls what
packets are injected into the buffer and chooses which of them to send. The
goal is then to show that the adversary's gain is at most R times Alg's gain.
If the algorithm is randomized, we consider its expected gain, E[GAlg(I)],
where the expectation is taken over all possible random choices made by Alg.
However, in the randomized case, the power of the adversary has to be further
specified. Following Ben-David et al. [3], we distinguish between an oblivious
and an adaptive-online adversary, which from now on we will call adaptive, for
short. An oblivious adversary has to construct the whole instance in advance.
This instance may depend on Alg but not on the random bits used by Alg dur-
ing the computation. The expected gain of Alg is compared to the gain of the
optimal offline solution on I. In contrast, in case of an adaptive adversary, the
choice of packets to be injected into the buffer may depend on the algorithm's
behavior up to the given time step. This adversary must also provide an an-
swering entity Adv, which creates a solution in parallel to Alg. This solution
may not be changed afterwards. We say that Alg is R-competitive against an
adaptive adversary if for any adaptively created instance I and any answering
algorithm Adv, it holds that E[GAlg(I)] ≥ 1
R · E[GAdv(I)]. We note that Adv
is (wlog) deterministic, but as Alg is randomized, so is the instance I.
In the literature on online algorithms [6], the definition of the competitive
ratio sometimes allows an additive constant, i.e., a deterministic algorithm is
then called R-competitive if there exists a constant α ≥ 0 such that for any
instance I it holds that GAlg(I) ≥ 1
R · GOpt(I) − α. An analogous definition
applies to the randomized case. For our algorithm Mix-R the bound holds for
α = 0, which is the best possible.
1.2 Basic Definitions
We denote a packet with weight w and relative deadline d by (w, d), where
the relative deadline of a packet is, at any time, the number of steps after
which it expires. The packet's absolute deadline, on the other hand, is the exact
point in time at which the packet expires. a packet that is in the buffer, i.e.,
has already been released and has neither expired nor been transmitted by an
algorithm, is called pending for the algorithm. The lifespan of a packet is its
relative deadline value upon injection, or in other words the difference between
its absolute deadline and release time.
2
The goal is to maximize the weighted throughput, i.e., the total weight of
transmitted packets. We assume that time is slotted in the following way. We
distinguish between points in time and time intervals, called steps. In step t,
corresponding to the interval (t, t + 1), Adv and the algorithm choose, inde-
pendently, a packet from their buffers and transmit it. The packet transmitted
by the algorithm (Adv) is immediately removed from the buffer and no longer
pending. Afterwards, at time t + 1, the relative deadlines of all remaining pack-
ets are decremented by 1, and the packets whose relative deadlines reach 0
expire and are removed from both Adv's and the algorithm's buffers. Next, the
adversary injects any set of packets. At this point, we proceed to step t + 1.
To no surprise, all known algorithms are scale-invariant, which means that
they make the same decisions if all the weights of packets in an instance are
scaled by a positive constant. a class of further restricted algorithms is of special
interest for their simplicity. An algorithm is memoryless if in every step its
decision depends only on the set of packets pending at that step. An algorithm
that is both memoryless and scale-invariant is called memoryless scale-invariant.
1.3 Previous and Related Work, Restricted Variants
The currently best, 1.828-competitive, deterministic algorithm for general in-
stances was given by Englert and Westermann [10]. Their algorithm is scale-
invariant, but it is not memoryless. However, in the same article Englert and
Westermann provide another, 1.893-competitive, deterministic algorithm that is
memoryless scale-invariant. The best known randomized algorithm is the 1.582-
competitive memoryless scale-invariant RMix, proposed by Chin et al. [7]. For
reasons explained in Section 2.1 the original analysis by Chin et al. is only appli-
cable in the oblivious adversary model. However, a refined analysis shows that
the algorithm remains 1.582-competitive in the adaptive adversary model [14].
Consider a (memoryless scale-invariant) greedy algorithm that always trans-
mits the heaviest pending packet. It is not hard to observe that it is 2-competitive,
and actually no better than that. But for a few years no better deterministic
algorithm for the general case was known. This naturally led to a study of
many restricted variants. Below we present some of them, together with known
results. The most relevant bounds known are summarized in Table 1. Note that
the majority of algorithms are memoryless scale-invariant.
For a general overview of techniques and results on buffer management, see
the surveys by Azar [2], Epstein and Van Stee [11] and Goldwasser [12].
Uniform Sequences An instance is s-uniform if the lifespan of each packet is
exactly s. Such instances have been considered for two reasons. Firstly, there is
a certain connection between them and the FIFO model of buffer management,
also considered by Kesselmann et al. [16]. Secondly, the 2-uniform instances are
among the most elementary restrictions that do not make the problem trivial.
However, analyzing these sequences is not easy: while a simple deterministic
1.414-competitive algorithm for 2-uniform instances [18] is known to be optimal
among memoryless scale-invariant algorithms [7], for unrestricted algorithms
a sophisticated analysis shows the optimum competitive ratio is 1.377 [9].
Bounded Sequences An instance is s-bounded if the lifespan of each packet is
at most s; therefore every s-uniform instances is also s-bounded. This class of in-
3
general
s-bounded
2-bounded
upper
lower
upper
lower
upper
lower
deterministic
(rand.) adaptive
(rand.) oblivious
1.828 [10], 1.893∗ [10]
1.618
1.582∗ [14]
1.333
1.582∗ [7]
1.25
s + o( 1
s )∗ [7]
2 − 2
1.618
1.618∗ [16]
1.618 [1, 8, 13]
1/(cid:0)1 − (1 −
1.333
1
s )s(cid:1)∗
1.333∗ [5]
1.333 [5]
1/(cid:0)1 − (1 − 1
1.25
s )s(cid:1)∗
1.25∗ [7]
1.25 [8]
Table 1: Comparison of known and new results. The results of this paper are
shown in boldface; a reference next to such entry means that this particular
bound was already known. The results without citations are implied by other
entries of the table. An asterisk denotes that the algorithm attaining the bound
is memoryless scale-invariant.
stances is important, because the strongest lower bounds on the competitive ra-
tio known for the problem employ 2-bounded instances. These are φ ≈ 1.618 for
deterministic algorithms [1, 8, 13], 1.25 for randomized algorithms in the obliv-
ious adversary model [8], and 4/3 in the adaptive adversary model [5]. For 2-
bounded instances algorithms matching these bounds are known [16, 7, 5]. A φ-
competitive deterministic algorithm is also known for 3-bounded instances [7],
but in general the best algorithms for s-bounded instances are only known to
be 2 − 2/s + o(1/s)-competitive [7].
Similarly Ordered Sequences An instance is similarly ordered or has agree-
able deadlines if for every two packets i and j their spanning intervals are not
properly contained in one another, i.e., if ri < rj implies di ≤ dj. Note that
every 2-bounded instance is similarly ordered, as is every s-uniform instance,
for any s. An optimal deterministic φ-competitive algorithm [17] and a ran-
domized 4/3-competitive algorithm for the oblivious adversary model [15] are
known. With the exception of 3-bounded instances, this is the most general
class of instances for which a φ-competitive deterministic algorithm is known.
Other restrictions Among other possible restrictions, let us mention one
for which our algorithm provides some bounds. Motivated by certain trans-
mission protocols, which usually specify only several priorities for packets, one
might bound the number of different packet weights. In fact, Kesselmann et al.
considered deterministic algorithms for instances with only two distinct packet
weights [16].
Generalization: Collecting Weighted Items from a Dynamic Queue
Bienkowski et al. [4] studied a generalization of buffer management with bounded
delay, in which the algorithm knows only the relative order between packets'
deadlines rather than their exact values; after Bienkowski et al. we dub the
generalized problem Collecting Items. Their paper focuses on deterministic al-
gorithms but it does provide certain lower bounds for memoryless algorithms,
matched by our algorithm. See Appendix A for details.
4
1.4 Our Contribution
We consider randomized algorithms against an adaptive adversary, motivated
by the following observation. In reality, traffic through a switch is not at all
independent of the packet scheduling algorithm. For example, lost packets are
typically resent, and throughput through a node affects the choice of routes for
data streams in a network. These phenomena can be captured by the adaptive
adversary model but not by the oblivious one. The adaptive adversary model
is also of its own theoretical interest and has been studied in numerous other
settings [6].
The main contribution of this paper is a simple memoryless scale-invariant
algorithm Mix-R, which may be viewed as RMix, proposed by Chin et al. [7],
with a different probability distribution over pending packets. The competitive
ratio of Mix-R is at most e/(e − 1) on the one hand, but on the other it is
provably better than that for many restricted variants of the problem. Some of
the upper bounds we provide were known before (cf. Table 1), but in general
they were achieved by several different algorithms.
versary, where N is the maximum, over steps, number of packets that have
Specifically, Mix-R is 1/(cid:0)1 − (1 − 1
positive probability of transmission in the step. Note that 1/(cid:0)1 − (1 − 1
N )N(cid:1)-competitive against adaptive ad-
N )N(cid:1)
tends to e/(e − 1) from below. The number N can be bounded a priori in
certain restricted variants of the problem, thus giving better bounds for them,
as we discuss in detail in Section 2.4. For now let us mention that N ≤ s in
s-bounded instances and instances with at most s different packet weights. The
particular upper bound of 4/3 that we obtain for 2-bounded instances is tight
in the adaptive adversary model [5].
As is the case with RMix, both Mix-R and its analysis rely only on the rel-
ative order between the packets' deadlines. Therefore our upper bound(s) apply
to the Collecting Items problem [4]. In fact, Mix-R is the optimum randomized
memoryless algorithm for that problem in a strong sense, cf. Appendix A.
2 General Upper Bound
2.1 Analysis technique
In our analysis, we follow the paradigm of modifying the adversary's buffer,
introduced by Li et al. [17]. Namely, we assume that in each step the algorithm
and the adversary have precisely the same pending packets in their buffers. Once
they both transmit a packet, we modify the adversary's buffer judiciously to
make it identical with that of the algorithm. This amortized analysis technique
leads to a streamlined and intuitive proof.
When modifying the buffer, we may have to let the adversary transmit an-
other packet, inject an extra packet to his buffer, or upgrade one of the packets in
its buffer by increasing its weight or deadline. We will ensure that these changes
will be advantageous to the adversary in the following sense: for any adversary
strategy Adv, starting with the current step and buffer content, there is an
adversary strategy Adv that continues computation with the modified buffer,
such that the total gain of Adv from the current step on (inclusive), on any
instance, is at least as large as that of Adv.
5
To prove R-competitiveness, we show that in each step the expected amor-
tized gain of the adversary is at most R times the expected gain of the algorithm,
where the former is the total weight of the packets that Adv eventually trans-
mitted in this step. Both expected values are taken over random choices of the
algorithm.
We are going to assume that Adv never transmits a packet a if there is
another pending packet b such that transmitting b is always advantageous to
Adv. Formally, we introduce a dominance relation among the pending packets
and assume that Adv never transmits a dominated packet.
We say that a packet a = (wa, da) is dominated by a packet b = (wb, db) at
time t if at time t both a and b are pending, wa ≤ wb and da ≥ db. If one of
these inequalities is strict, we say that a is strictly dominated by b. We say that
packet a is (strictly) dominated whenever there exists a packet b that (strictly)
dominates it. Then the following fact can be shown by a standard exchange
argument.
Fact 1. For any adversary strategy Adv, there is a strategy Adv with the
following properties:
1. the gain of Adv on every sequence is at least the gain of Adv,
2. in every step t, Adv does not transmit a strictly dominated packet at
time t.
Proof. Adv can be transformed into Adv iteratively: take the minimum t0 such
that Adv first violates the second property in step t0, and transform Adv into
an algorithm Adv′ with gain no smaller than that of Adv, which satisfies the
second property up to step t0, possibly violating it in further steps.
Let t0 be the first step in which the second property is violated. Let y =
(w, d) be the packet transmitted by Adv and x = (w′, d′) be the packet that
dominates y; then w′ ≥ w and d′ ≤ d. Let Adv′ transmit the same packets as
Adv up to step t0 − 1, but in step t0 let it transmit x, and in the remaining
steps let it try to transmit the same packets as Adv. It is impossible in one case
only: when Adv transmits x in some step t. But then d ≥ d′ > t, so let Adv′
transmit y, still pending at t. Clearly, the gain of Adv′ is at least as large as
the gain of Adv.
Let us stress that Fact 1 holds for the adaptive adversary model. Now
we give an example of another simplifying assumption, often assumed in the
oblivious adversary model, which seems to break down in the adaptive adversary
model. In the oblivious adversary model the instance is fixed in advance by the
adversary, so Adv may precompute the optimum schedule to the instance and
follow it. Moreover, by standard exchange argument for the fixed set of packets
to be transmitted, Adv may always send the packet with the smallest deadline
from that set -- this is usually called the earliest deadline first (EDF) property
or order. This assumption not only simplifies analyses of algorithms but is often
crucial for them to yields desired bounds [7, 9, 17, 15].
In the adaptive adversary model, however, the following phenomenon occurs:
as the instance I is randomized, Adv does not know for sure which packets it
will transmit in the future. Consequently, deprived of that knowledge, it cannot
ensure any specific order of packet transmissions.
6
2.2 The Algorithm
We describe the algorithm's behavior in a single step.
Algorithm 1 Mix-R (single step)
do nothing and proceed to the next step
1: if there are no pending packets then
2:
3: end if
4: m ← 0
5: n ← 0
6: r ← 1
7: H0 ← pending packets
8: h0 = (w0, d0) ← heaviest packet from H0
9: while Hm 6= ∅ do
10:
⊲ counts packets that are not strictly dominated
⊲ counts packets with positive probability assigned
⊲ unassigned probability
11:
12:
13:
14:
15:
16:
m ← m + 1
hm = (wm, dm) ← heaviest not strictly dominated packet from Hm−1
pm−1 ← min{1 − wm
wm−1
r ← r − pm−1
if r > 0 then
n ← n + 1
, r}
end if
Hm ← {x ∈ Hm−1 x is not dominated by hm}
17:
18: end while
19: pm ← r
20: transmit h chosen from h1, . . . , hn with probability distribution p1, . . . , pn
21: proceed to the next step
We introduce the packet h0 to shorten Mix-R's pseudocode by making it
possible to set the value of p1 in the first iteration of the loop. The packet itself
is chosen in such a way that p0 = 0, to make it clear that it is not considered
for transmission (unless h0 = h1). The while loop itself could be terminated as
soon as r = 0, because afterwards Mix-R does not assign positive probability
to any packet. However, letting it construct the whole sequence h1, h2, . . . hm
such that Hm = ∅ simplifies our analysis. Before proceeding with the analysis,
we note a few facts about Mix-R.
Fact 2. The sequence of packets h0, h1, . . . , hm selected by Mix-R satisfies
w0 = w1 > w2 > · · · > wm ,
d1 > d2 > · · · > dm .
Furthermore, every pending packet is dominated by one of h1, . . . , hm.
Fact 3. The numbers p1, p2, . . . , pm form a probability distribution such that
pi ≤ 1 −
wi+1
wi
for all i < m .
Furthermore, the bound is tight for i < n, while pi = 0 for i > n, i.e.,
pi =(1 − wi+1
0,
wi
for i < n
for i > n
,
7
(1)
(2)
Theorem 4. Mix-R is 1/(cid:0)1 − (1 − 1
versary, where N is the maximum, over steps, number of packets that are as-
signed positive probability in a step.
N )N(cid:1)-competitive against an adaptive ad-
Proof. For a given step, we describe the changes to Adv's scheduling decisions
and modifications to its buffer that make it the same as Mix-R's buffer. Then,
to prove our claim, we will show that
E [GAdv] ≤ w1 ,
E [GMix-R] ≥ w1(cid:18)1 − (1 −
1
n
)n(cid:19) ,
(3)
(4)
where n is the number of packets assigned positive probability in the step. The
theorem follows by summation over all steps.
Recall that, by Fact 1, Adv (wlog) sends a packet that is not strictly dom-
inated. By Fact 2, the packets h1, h2, . . . hm dominate all pending packets, so
the one sent by Adv, say p is (wlog) one of h1, h2, . . . hm: if p is dominated by
hi, but not strictly dominated, then p has the same weight and deadline as hi.
We begin by describing modifications to Adv's buffer and estimate Adv's
amortized gain. To this end we need to fix the packet sent by Mix-R, so let us
assume it is hf = (wf , df ). Assume that Adv transmits a packet hz = (wz, dz).
We will denote the adversary's amortized gain given the latter assumption by
G(z)
Adv. We consider two cases.
Case 1: df ≤ dz. Then wf ≤ wz, since hz is not dominated. After both Adv
and Mix-R transmit their packets, we replace hf in the buffer of Adv
by a copy of hz. This way their buffers remain the same afterwards, and
the change is advantageous to Adv: this is essentially an upgrade of the
packet hf in its buffer, as both df ≤ dz and wf ≤ wz hold.
Case 2: df > dz. After both Adv and Mix-R transmit their packets, we let
Adv additionally transmit hf , and we inject a copy of hz into its buffer,
both of which are clearly advantageous to Adv. This makes the buffers
of Adv and Mix-R identical afterwards.
We start by proving (3), the bound on the adversary's expected amortized
gain. Note that Adv always gains wz, and if dz < df (z > f ), it additionally
gains wf . Thus, when Adv transmits hz, its expected amortized gain is
(5)
(6)
EhG(z)
Advi = wz +Xi<z
piwi .
As the adversary's expected amortized gain satisfies
E [GAdv] ≤ max
1≤i≤mnEhG(i)
Advio ,
to establish (3), we will prove that
max
1≤i≤mnEhG(i)
Advio ≤ G(1)
Adv = w1 .
The equality in (6) follows trivially from (5). To see that the inequality in (6)
holds as well, observe that, by (5), for all j < m,
EhG(i)
Advi − EhG(i+1)
Adv i = wi − wi+1 − piwi ≥ 0 ,
(7)
8
where the inequality follows from (1).
Now we turn to (4), the bound on the expected gain of Mix-R in a single
step. Obviously,
n
E [GMix-R] =
piwi .
(8)
Xi=1
By (2), piwi = wi − wi+1 for all i < n. Also, pn = 1 −Pi<n pi, by Fact 3.
Making corresponding substitutions in (8) yields
E [GMix-R] = n−1
Xi=1
= w1 − wn
(wi − wi+1)! + 1 −
Xi=1
pi .
n−1
pi! wn
n−1
Xi=1
As (2) implies wi = wi−1(1 − pi−1) for all i ≤ n, we can express wn as
n−1
wn = w1
Yi=1
(1 − pi) .
Substituting (10) for wn in (9), we obtain
(9)
(10)
(11)
Note that
pi! .
E [GMix-R] = w1 1 −
n−1
n−1
(1 − pi)
Xi=1
Yi=1
(1 − pi) + n−1
pi! = n − 1 ,
Xi=1
n−1
Xi=1
and therefore the inequality between arithmetic and geometric means yields
n−1
(1 − pi)
Yi=1
n−1
Xi=1
pi ≤ (1 −
1
n
)n .
Plugging (12) into (11) yields
E [GMix-R] ≥ w1(cid:18)1 − (1 −
1
n
)n(cid:19) ,
which proves (4), and together with (3), the whole theorem.
2.3 Rationale behind the probability distribution
Recall that the upper bound on the competitive ratio of Mix-R is
max1≤z≤m{EhG(z)
Advi}
E [GMix-R]
,
irrespective of the choice of p1, . . . , pm.
(12)
(13)
9
The particular probability distribution used in Mix-R is chosen to (heuristi-
cally) minimize above ratio by maximizing E [GMix-R], while keeping (6) satisfied,
i.e., keeping E [GAdv] ≤ G(1)
Adv = w1.
The first goal is trivially achieved by setting p1 ← 1. This however makes
ferred to p2, p3, . . . in the following way. To keep E [GMix-R] as large as possible,
p2 is greedily set to its maximum, if there is any unassigned probability left, p3
Advi > w1 for all z > 1. Therefore, some of the probability mass is trans-
Advi does not depend on pi for i ≥ z,
Advi can be equalized with w1 sequentially, with z increasing, un-
Adv i
Advi with EhG(j−1)
EhG(z)
is set to its maximum, and so on. As EhG(z)
the values EhG(z)
til there is no unassigned probability left. Equalizing EhG(j)
Advi cannot be equalized, they are only smaller than w1.
EhG(z)
consists in setting pj−1 ← 1 − wj
, as shown in (7). The same inequality shows
wj−1
what is intuitively clear: once there is no probability left and further values
The lower bound for the Collecting Items problem [4], presented in Ap-
pendix A, proves that this heuristic does minimize (13).
2.4
Implications for Restricted Variants
We have already mentioned that for s-bounded instances or those with at most
s different packet weights, N ≤ m ≤ s in Theorem 4, which trivially follows
from Fact 2. Thus for either kind of instances Mix-R is 1/(cid:0)1 − (1 − 1
competitive. In particular, on 2-bounded instances Mix-R coincides with the
previously known optimal 4/3-competitive algorithm Rand [5] for the adaptive
adversary model.
s )s(cid:1)-
Sometimes it may be possible to give more sophisticated bounds on N , and
consequently on the competitive ratio for particular variant of the problem, as
we now explain. The reason for considering only the packets h0, h1, . . . , hm is
clear: by Fact 1 and Fact 2, Adv (wlog) transmits one of them. Therefore,
Mix-R tries to mimic Adv's behavior by adopting a probability distribution
over these packets (recall that in the analysis the packets pending for Mix-R
and Adv are exactly the same) that keeps the maximum, over Adv's choices,
expected amortized gain of Adv and its own expected gain as close as possible
(cf. Section 2.3). Now, if for whatever reason we know that Adv is going to
transmit a packet from some set S, then H0 can be initialized to S rather than
all pending packets, and Theorem 4 will still hold. And as the upper bound
guaranteed by Theorem 4 depends on N , it might improve if the cardinality of
S is small.
While it seems unlikely that bounds for any restricted variant other than
s-bounded instances or instances with at most s different packet weights can
be obtained this way, there is one interesting example that shows it is possi-
ble. For similarly ordered instances (aka instances with agreeable deadlines)
and oblivious adversary one can always find such set S of cardinality at most
2 [15, Lemma 2.7]; while not explicitly stated, this fact was proved before by
Li et al. [17]. Roughly, the set S contains the earliest-deadline and the heaviest
packet from any optimal provisional schedule. The latter is the optimal schedule
under the assumption that no further packets are ever injected, and as such can
be found in any step.
10
3 Conclusion and Open Problems
While Mix-R is very simple to analyze, it subsumes almost all previously known
randomized algorithms for packet scheduling and provides new bounds for sev-
eral restricted variants of the problem. One notable exception is the optimum
algorithm against oblivious adversary for 2-bounded instances [7]. This exposes
that the strength of our analysis, i.e., applicability to adaptive adversary model,
is most likely a weakness at the same time. The strongest lower bounds on com-
petitive ratio for oblivious and adaptive adversary differ. And as both are tight
for 2-bounded instances, it seems impossible to obtain an upper bound smaller
than 4/3 on the competitive ratio of Mix-R for any non-trivial restriction of
the problem in the oblivious adversary model.
In both the algorithm and its analysis it is the respective order of packets'
deadlines rather than their exact values that matter. Therefore, our results
are also applicable to the Collecting Items problem [4], briefly described in
Section 1.3. As mentioned in Section 1.4, Mix-R is the optimum randomized
memoryless algorithm for Collecting Items, cf. Appendix A.
Therefore, to beat either the general bound of e/(e − 1), or any of the
1/(cid:0)1 − (1 − 1
s )s(cid:1) bounds for s-bounded instances for buffer management with
bounded delay, one either needs to consider algorithms that are not memory-
less scale-invariant, or better utilize the knowledge of exact deadlines -- in the
analysis at least, if not in the algorithm itself.
Last but not least, let us remark again that Mix-R and its analysis might au-
tomatically provide better bounds for further restricted variants of the problem,
provided that some insight allows to confine the adversary's choice of packets
for transmission in a step, while knowing which packets are pending for it -- one
such example is the algorithm for similarly ordered instances (aka instances with
agreeable deadlines) [15], as we discussed in Section 2.4.
References
[1] N. Andelman, Y. Mansour, and A. Zhu. Competitive queueing policies for
qos switches. In Proc. of the 14th ACM-SIAM Symp. on Discrete Algo-
rithms (SODA), pages 761 -- 770, 2003.
[2] Y. Azar. Online packet switching. In Proc. of the 2nd Workshop on Ap-
proximation and Online Algorithms (WAOA), pages 1 -- 5, 2004.
[3] S. Ben-David, A. Borodin, R. M. Karp, G. Tardos, and A. Wigderson. On
the power of randomization in online algorithms. Algorithmica, 11(1):2 -- 14,
1994. Also appeared in Proc. of the 22nd STOC, pages 379 -- 386, 1990.
[4] M. Bienkowski, M. Chrobak, C. Durr, M. Hurand, A. Jez, L. Jez, and
G. Stachowiak. Collecting weighted items from a dynamic queue. In Proc. of
the 20th ACM-SIAM Symp. on Discrete Algorithms (SODA), pages 1126 --
1135, 2009.
[5] M. Bienkowski, M. Chrobak, and L. Jez. Randomized algorithms for buffer
management with 2-bounded delay. In Proc. of the 6th Workshop on Ap-
proximation and Online Algorithms (WAOA), pages 92 -- 104, 2008.
11
[6] A. Borodin and R. El-Yaniv. Online Computation and Competitive Anal-
ysis. Cambridge University Press, 1998.
[7] F. Y. L. Chin, M. Chrobak, S. P. Y. Fung, W. Jawor, J. Sgall, and T. Tich´y.
Online competitive algorithms for maximizing weighted throughput of unit
jobs. Journal of Discrete Algorithms, 4:255 -- 276, 2006.
[8] F. Y. L. Chin and S. P. Y. Fung. Online scheduling for partial job values:
Does timesharing or randomization help? Algorithmica, 37:149 -- 164, 2003.
[9] M. Chrobak, W. Jawor, J. Sgall, and T. Tich´y. Improved online algorithms
for buffer management in QoS switches. ACM Transactions on Algorithms,
3(4), 2007. Also appeared in Proc. of the 12th ESA, pages 204 -- 215, 2004.
[10] M. Englert and M. Westermann. Considering suppressed packets improves
buffer management in QoS switches. In Proc. of the 18th ACM-SIAM Symp.
on Discrete Algorithms (SODA), pages 209 -- 218, 2007.
[11] L. Epstein and R. van Stee. Buffer management problems. Sigact News,
35:58 -- 66, 2004.
[12] M. Goldwasser. A survey of buffer management policies for packet switches.
SIGACT News, 41(1):100 -- 128, 2010.
[13] B. Hajek. On the competitiveness of online scheduling of unit-length pack-
ets with hard deadlines in slotted time. In Conference in Information Sci-
ences and Systems, pages 434 -- 438, 2001.
[14] L. Jez. Randomised buffer management with bounded delay against adap-
tive adversary. CoRR, abs/0907.2050, 2009.
[15] L. Jez. Randomized algorithm for agreeable deadlines packet scheduling.
In Proc. of the 27th Symp. on Theoretical Aspects of Computer Science
(STACS), pages 489 -- 500, 2010.
[16] A. Kesselman, Z. Lotker, Y. Mansour, B. Patt-Shamir, B. Schieber, and
M. Sviridenko. Buffer overflow management in QoS switches. SIAM Journal
on Computing, 33(3):563 -- 583, 2004. Also appeared in Proc. of the 33rd
STOC, pages 520 -- 529, 2001.
[17] F. Li, J. Sethuraman, and C. Stein. An optimal online algorithm for packet
scheduling with agreeable deadlines. In Proc. of the 16th ACM-SIAM Symp.
on Discrete Algorithms (SODA), pages 801 -- 802, 2005.
[18] A. Zhu. Analysis of queueing policies in QoS switches. Journal of Algo-
rithms, 53(2):137 -- 168, 2004.
12
A Lower Bound for Collecting Items
In this section, for completeness, we evoke the lower bound on the Collecting
Items problem [4]. As the proof is omitted in the original article due to space
constraints, and the original theorem statement therein is not parametrized by
N , we restate the theorem.
Theorem 5 (Theorem 6.3 of [4]). For every randomized memoryless algorithm
for the Collecting Items problem, there is an adaptive adversary's strategy using
at most N different packet weights such that the algorithm's competitive ratio
against the strategy is at least 1/(cid:0)1 − (1 − 1
has at most N packets in its queue.
N )N(cid:1), and at every step the algorithm
Below we present the original proof from [4].
Proof. Fix some online memoryless randomized algorithm A, and consider the
following scheme. Let a > 1 be a constant, which we specify later, and let
n = N − 1 At the beginning, the adversary inserts items a0, a1, . . . , an into the
queue, in this order. (To simplify notation, in this proof we identify items with
their weights.) In our construction we maintain the invariant that in each step,
the list of items pending for A is equal to a0, a1, . . . , an. Since A is memoryless,
in each step it uses the same probability distribution (qj)n
j=0, where qj is the
i=0 qi = 1, as without loss of
probability of collecting item aj. Moreover, Pn
generality the algorithm always makes a move.
We consider n+1 strategies for an adversary, numbered 0, 1, . . . , n. The k-th
in each step collect ak, delete items a0, a1, . . . , ak, and
strategy is as follows:
then issue new copies of these items. Additionally, if A collected aj for some
j > k, then the adversary issues a new copy of aj as well. This way, in each step
exactly one copy of each aj is pending for A, while the adversary accumulates
in its pending set copies of the items aj, for j > k, that were collected by A.
This step is repeated T ≫ n times, and after the last step both the adversary
and the algorithm collect all their pending items. Since T ≫ n, we only need
to focus on the expected amortized profits (defined below) in a single step.
We look at the gains of A and the adversary in a single step. If the adversary
chooses strategy k, then it gains ak. Additionally, at the end it collects the item
collected by the algorithm if this item is greater than ak. Thus, its amortized
expected gain in a single step is ak + Pi>k qiai. The expected gain of A is
Pi qiai.
j=0 of the algorithm, the adversary
chooses a strategy k which maximizes the competitive ratio. Thus, the compet-
itive ratio of A is is at least
For any probability distribution (qj)n
R = max
k ( ak +Pj>k qj aj
Pj qjaj
for any coefficients v0, . . . , vn ≥ 0 such thatPk vk = 1. Let M = an+1−n(a−1).
For k = 0, 1, ..., n, we choose
) ≥ Xk
vk
ak +Pj>k qjaj
Pj qjaj
,
(14)
vk = ( 1
M an−k(a − 1),
1
M (a − n(a − 1), )
if k < n ,
if k = n .
13
The choice of these values may seem somewhat mysterious, but it's in fact quite
simple -- it is obtained by considering A's distributions where qj = 1 for some j
(and thus when A is deterministic), assuming that the resulting lower bounds
on the right-hand side of (14) are equal, and solving the resulting system of
equations.
For these values of vk we obtain
M vkak +
n
Xk=0
M R
n
Xj=0
qjaj ≥
=
n
n−1
Xk=0
Xk=0
M vkak + M vnan +
M vk
qjaj
M vkXj>k
Xj=0
n
qjajXk<j
Xj=0
n
n
= n(a − 1)an + [a − n(a − 1)]an +
qj(aj − 1)an+1
= an+1 + an+1
= an+1 + an+1
qjaj − an+1
qj
Xj=0
qjaj − an+1
n
n
Xj=0
Xj=0
= an+1
n
Xj=0
qjaj .
Therefore, R ≥ an+1/M . This bound is maximized for a = 1 + 1/n, in which
case we get
n(cid:1)n+1
R ≥ (cid:0)1 + 1
n(cid:1)n+1
(cid:0)1 + 1
− 1
completing the proof.
1
1 −(cid:0)1 − 1
N(cid:1)N ,
N −1(cid:17)N
= (cid:16)1 + 1
N −1(cid:17)N
(cid:16)1 + 1
− 1
=
14
|
1110.1079 | 1 | 1110 | 2011-10-05T19:27:57 | A Near-Optimal Sublinear-Time Algorithm for Approximating the Minimum Vertex Cover Size | [
"cs.DS"
] | We give a nearly optimal sublinear-time algorithm for approximating the size of a minimum vertex cover in a graph G. The algorithm may query the degree deg(v) of any vertex v of its choice, and for each 1 <= i <= deg(v), it may ask for the i-th neighbor of v. Letting VC_opt(G) denote the minimum size of vertex cover in G, the algorithm outputs, with high constant success probability, an estimate VC_estimate(G) such that VC_opt(G) <= VC_estimate(G) <= 2 * VC_opt(G) + epsilon*n, where epsilon is a given additive approximation parameter. We refer to such an estimate as a (2,epsilon)-estimate. The query complexity and running time of the algorithm are ~O(avg_deg * poly(1/epsilon)), where avg_deg denotes the average vertex degree in the graph. The best previously known sublinear algorithm, of Yoshida et al. (STOC 2009), has query complexity and running time O(d^4/epsilon^2), where d is the maximum degree in the graph. Given the lower bound of Omega(avg_deg) (for constant epsilon) for obtaining such an estimate (with any constant multiplicative factor) due to Parnas and Ron (TCS 2007), our result is nearly optimal.
In the case that the graph is dense, that is, the number of edges is Theta(n^2), we consider another model, in which the algorithm may ask, for any pair of vertices u and v, whether there is an edge between u and v. We show how to adapt the algorithm that uses neighbor queries to this model and obtain an algorithm that outputs a (2,epsilon)-estimate of the size of a minimum vertex cover whose query complexity and running time are ~O(n) * poly(1/epsilon). | cs.DS | cs | A Near-Optimal Sublinear-Time Algorithm
for Approximating the Minimum Vertex Cover Size
Krzysztof Onak∗
Dana Ron †
Michal Rosen ‡
Ronitt Rubinfeld§
October 29, 2018
Abstract
We give a nearly optimal sublinear-time algorithm for approximating the size of a minimum vertex
cover in a graph G. The algorithm may query the degree deg(v) of any vertex v of its choice, and for
each 1 ≤ i ≤ deg(v), it may ask for the ith neighbor of v. Letting VCopt(G) denote the minimum
size of vertex cover in G, the algorithm outputs, with high constant success probability, an estimate
cVC(G) such that VCopt(G) ≤ cVC(G) ≤ 2VCopt(G) + ǫn, where ǫ is a given additive approximation
parameter. We refer to such an estimate as a (2, ǫ)-estimate. The query complexity and running time
of the algorithm are O( ¯d · poly(1/ǫ)), where ¯d denotes the average vertex degree in the graph. The
best previously known sublinear algorithm, of Yoshida et al. (STOC 2009), has query complexity and
running time O(d4/ǫ2), where d is the maximum degree in the graph. Given the lower bound of Ω( ¯d)
(for constant ǫ) for obtaining such an estimate (with any constant multiplicative factor) due to Parnas and
Ron (TCS 2007), our result is nearly optimal.
In the case that the graph is dense, that is, the number of edges is Θ(n2), we consider another model,
in which the algorithm may ask, for any pair of vertices u and v, whether there is an edge between
u and v. We show how to adapt the algorithm that uses neighbor queries to this model and obtain an
algorithm that outputs a (2, ǫ)-estimate of the size of a minimum vertex cover whose query complexity
and running time are O(n) · poly(1/ǫ).
1
1
0
2
t
c
O
5
]
S
D
.
s
c
[
1
v
9
7
0
1
.
0
1
1
1
:
v
i
X
r
a
∗School of Computer Science, Carnegie Mellon University, E-mail: [email protected]. Research supported by a Simons
Postdoctoral Fellowship and the NSF grant 0728645.
†School of Electrical Engineering, Tel Aviv University. E-mail: [email protected]. Research supported by the Israel
Science Foundation grant number 246/08.
‡Blavatnik School of Computer Science, Tel Aviv University, E-mail: [email protected]
§CSAIL, MIT, Cambridge MA 02139 and the Blavatnik School of Computer Science, Tel Aviv University. E-mail:
[email protected]. Research supported by NSF awards CCF-1065125 and CCF-0728645, Marie Curie Reintegration
grant PIRG03-GA-2008-231077 and the Israel Science Foundation grant nos. 1147/09 and 1675/09.
1 Introduction
Computing the size of a minimum vertex cover in a graph is a classic NP-hard problem. However, one can
approximate the optimal value of the solution to within a multiplicative factor of two, via a neat and simple
algorithm whose running time is linear in the size of the graph (this algorithm was independently discovered
by Gavril and Yanakakis, see e.g. [PS98]).
A natural question is whether it is possible to obtain a good approximation for the size of the optimal
vertex cover in time that is sublinear in the size of the graph G. Since achieving a pure multiplicative
0 ≤ ǫ < 1, where VCopt(G) denotes the minimum size of a vertex cover in G. We refer to such an estimate
approximation is easily seen to require linear time, we focus on algorithms that compute an estimate cVC(G)
such that with high constant probability, VCopt(G) ≤ cVC(G) ≤ α · VCopt(G) + ǫn, for α ≥ 1 and
cVC(G) as an (α, ǫ)-estimate of VCopt(G). Observe that in the special case when the vertex cover is very
large, namely VCopt(G) = Θ(n) (which happens for example when the maximum degree and the average
degree are of the same order), then an (α, ǫ)-estimate yields an (α + O(ǫ))-multiplicative approximation.
Since an algorithm with complexity sublinear in the size of the graph cannot even read the entire graph,
it must have query access to the graph. In this work we consider two standard models of queries. In the first
model, the algorithm may query the degree deg(v) of any vertex v of its choice, and it may also query the
ith neighbor of v (where the the order on the neighbors is arbitrary). In the second model, more appropriate
when the graph is stored as an adjacency matrix, the algorithm can check in a single query whether there
is an edge between two vertices v and w chosen by the algorithm. We focus on the first model, but we
eventually show that our algorithm can be modified to work in the second model as well.
Previous work. The aforementioned question was first posed by Parnas and Ron [PR07], who showed how
to obtain a (2, ǫ)-estimate (for any given additive approximation parameter ǫ) in time dO(log d/ǫ3), where d
is the maximum degree in the graph. The dependence on the maximum degree d can actually be replaced
by a dependence on ¯d/ǫ, where ¯d is the average degree in the graph [PR07]. The upper bound of dO(log d/ǫ3)
was significantly improved in a sequence of papers [MR09, NO08, YYI09], where the best result due to
Yoshida, Yamamoto, and Ito [YYI09] (who analyze an algorithm proposed by Nguyen and Onak [NO08]) is
an upper bound of O(d4/ǫ2). Their analysis can also easily be adapted to give an upper bound of O( ¯d4/ǫ4)
for graphs with bounded average vertex degree ¯d.
On the negative side, it was also proved in [PR07] that at least a linear dependence on the average
degree, ¯d, is necessary. Namely, Ω( ¯d) queries are necessary for obtaining an (α, ǫ)-estimate for any α ≥ 1
and ǫ < 1/4, provided that ¯d = O(n/α), and in particular this is true for α = 2. We also mention that
obtaining a (2 − γ, ǫ)-estimate for any constant γ requires a number of queries that grows at least as the
square root of the number of vertices [PR07, due to Trevisan].
Our Result.
In this work we describe and analyze an algorithm that computes a (2, ǫ)-estimate of
VCopt(G) in time O( ¯d) · poly(1/ǫ). Note that since the graph contains ¯dn/2 edges, our running time
is sublinear for all values of ¯d. In particular, for graphs of constant average degree, the running time is inde-
pendent of the number of nodes and edges in the graph, whereas for general graphs it is bounded by at most
the square root of the number of edges. In view of the aforementioned lower bound of Ω( ¯d), our algorithm
is optimal in terms of the dependence on the average degree up to a polylogarithmic factor. Since our algo-
rithm builds on previous work, and in particular on the algorithm proposed and analyzed in [NO08, YYI09],
1
we describe the latter algorithm first.1 We refer to this algorithm as Approx-VC-I.
The Algorithm Approx-VC-I. Recall that the size of a minimum vertex cover is lower-bounded by the
size of any (maximal) matching in the graph, and is upper-bounded by twice the size of any maximal match-
ing. This is indeed the basis of the aforementioned factor-two approximation algorithm, which runs in
linear-time. To estimate the size of an arbitrary such maximal matching, the algorithm follows the sampling
paradigm of Parnas and Ron [PR07]. That is, the algorithm Approx-VC-I selects, uniformly, independently
and at random, Θ(d2/ǫ2) edges. For each edge selected, it calls a maximal matching oracle, which we
describe momentarily, where the oracle's answers indicate whether or not the edge is in the maximal match-
ing M, for some arbitrary maximal matching M (that is not a function of the queries to the oracle). The
algorithm then outputs an estimate of the size of the maximal matching M (and hence of a minimum vertex
cover) based on the fraction of sampled edges for which the maximal matching oracle returned a positive
answer. The number of sampled edges ensures that with high constant probability, the additive error of the
estimate is O((ǫ/d)m) ≤ ǫn, where m is the number of edges in the graph.
The main idea of the algorithm follows the idea suggested in [NO08] which is to simulate the answers
of the standard sequential greedy algorithm. The greedy algorithm supposes a fixed ranking (ordering) of
the edges in G, which uniquely determines a maximal matching as follows: proceeding along the edges
according to the order determined by the ranking, add to the matching each edge that does not share an end-
point with any edge previously placed in the matching. The maximal matching oracle essentially emulates
this procedure while selecting a random ranking "on the fly", but is able to achieve great savings in running
time by noting that to determine whether an edge is placed in the matching, it is only necessary to know
whether or not adjacent edges that are ranked lower than the current edge have been placed in the matching.
Namely, given an edge (u, v), it considers all edges that share an endpoint with (u, v) and whose (randomly
assigned) ranking is lower than that of (u, v). If there are no such edges, then the oracle returns TRUE.
Otherwise it performs recursive calls to these edges, where the order of the calls is according to their ranking.
If any recursive call is answered TRUE, then the answer on (u, v) is FALSE, while if all answers (on the
incident edges with a lower rank) is answered FALSE, then the answer on (u, v) is TRUE.
Though the correctness of the above algorithm follows directly from the correctness of the greedy al-
gorithm, the query and runtime analysis are more difficult. The analysis of [NO08] is based on a counting
argument that shows that it is unlikely that there is a long path of recursive calls with a monotone decreas-
ing set of ranks. Their bound gives a runtime that is independent of the size of the graph, but exponential
in the degree d. However, using that the algorithm recurses according to the smallest ranked neighbor,
[YYI09] give an ingenious analysis that bounds by O(d) the total number of expected recursive calls when
selecting an edge uniformly at random, and when selecting a ranking uniformly at random. This is what
allows [YYI09] to obtain an algorithm whose query complexity and running time are O(d4/ǫ2).
Our Algorithm.
In what follows we describe an algorithm that has almost linear dependence on the max-
imum degree d. The transformation to an algorithm whose complexity depends on the average degree ¯d is
done on a high level along the lines described in [PR07]. We first depart from Approx-VC-I by performing
the following variation. Rather than sampling edges and approximating the size of a maximal matching by
1Yoshida et al. [YYI09] actually analyze an algorithm for approximating the size of a maximal independent set. They then apply
it to the line graph of a given graph G, so as to obtain an estimate of the size of a maximal matching, and hence of a minimum
vertex cover (with a multiplicative cost of 2 in the quality of the estimate). For the sake of simplicity, we describe their algorithm
directly for a maximal matching (minimum vertex cover).
2
calling the maximal matching oracle on the sampled edges, we sample vertices (as in [PR07]), and we call
a vertex cover oracle on each selected vertex v. The vertex cover oracle calls the maximal matching oracle
on the edges incident to v according to the order induced by their ranking (where the ranking is selected
randomly). Once some edge returns TRUE, the vertex cover oracle returns TRUE, and if all incident edges
return FALSE, the vertex cover oracle returns FALSE. By performing this variation we can take a sample of
vertices that has size Θ(1/ǫ2) rather than2 Θ(d2/ǫ2).
Unfortunately, the analysis of [YYI09] is no longer applicable as is. Recall that their analysis bounds
the expected number of recursive calls to the maximal matching oracle, for a random ranking, and for a
randomly selected edge. In contrast, we select a random vertex and call the maximal matching oracle on
its (at most d) incident edges. Nonetheless, we are able to adapt the analysis of [YYI09] and give a bound
of O(d) on the expected number of recursive calls to the maximal matching oracle, when selecting a vertex
uniformly at random.3
As a direct corollary of the above we can get an algorithm whose query complexity and running time
grow quadratically with d. Namely, whenever the maximal matching oracle is called on a new edge (u, v),
the algorithm needs to perform recursive calls on the edges incident to u and v, in an order determined by
their ranking. To this end it can query the O(d) neighbors of u and v, assign them (random) rankings, and
continue in a manner consistent with the assigned rankings.
To reduce the complexity of the algorithm further, we show a method that for most of the edges that
we visit, allows us to query only a small subset of adjacent edges. Ideally, we would like to make only
k queries when k recursive calls are made. One of the problems that we encounter here is that if we do
not query all adjacent edges, then for some edge (u, v), we could visit a different edge incident to u and a
different edge incident to v and make conflicting decisions about the ranking of (u, v) from the point of view
of these edges. This could result in an inconsistent execution of the algorithm with results that are hard to
predict. Instead, we devise a probabilistic procedure, that, together with appropriate data structures, allows
us to perform queries almost "only when needed" (we elaborate on this in the next paragraph). By this we
mean that we perform queries only on a number of edges that is a poly(log(d/ǫ)) factor larger than the total
number of recursive calls made to the maximal matching oracle. We next discuss our general approach.
As in previous work, we implement the random ranking by assigning numbers to edges independently,
uniformly at random from (0, 1] (or, more precisely, from an appropriate discretization of (0, 1]). For each
vertex we keep a copy of a data structure that is responsible for generating and assigning random numbers to
incident edges. For each vertex, we can ask the corresponding data structure for the incident edge with the
ith lowest number. How does the data structure work? Conceptually, the edges attached to each vertex are
grouped into "layers", where the edges in the first layer have random numbers that are at most 1/d, the edges
in layer i > 1 have random numbers in the range 2i−1/d to 2i/d. The algorithm randomly chooses edges
to be in a layer for each vertex, one layer at a time, starting with the lowest layer. Each successive layer is
processed only as needed by the algorithm. If the algorithm decides that an edge is in the current layer, then
it picks a random number for the edge uniformly from the range associated with the layer. In particular, it
is possible to ensure that the final random number comes from the uniform distribution on (0, 1]. In order
to make sure that the same decision is made at both endpoints of an edge (u, v), the data structures for u
and v communicate whenever they want to assign a specific random number to the edge. The algorithm
2We note that it is actually possible to save one factor of d without changing the algorithm Approx-VC-I by slightly refining
the probabilistic analysis. This would reduce the complexity of Approx-VC-I to cubic in d.
3To be more precise, we first give a bound that depends on the ratio between the maximum and minimum degrees as well as on
the average degree, and later we show how to obtain a dependence on d (at an extra cost of 1/ǫ) by slightly modifying the input
graph.
3
works in such a way so that vertices need query their incident edges only when a communication regarding
the specific edge occurs. Our final algorithm is obtained by minimizing the amount of communication
between different data structures, and therefore, making them discover not many more edges than necessary
for recursive calls in the graph exploration.
Other Related Work. For some restricted classes of graphs it is possible to obtain a (1, ǫ)-estimate of
the size of the minimum vertex cover in time that is a function of only ǫ. Elek shows that this is the case
for graphs of subexponential growth [Ele10]. For minor-free graphs, one obtains this result by applying the
generic reduction of Parnas and Ron [PR07] to local distributed algorithm of Czygrinow, Ha´n´ckowiak, and
Wawrzyniak [CHW08]. Both of these results are generalized by Hassidim et al. [HKNO09] to any class of
hyperfinite graphs. In particular, for planar graphs, they give an algorithm that computes a (1, ǫ)-estimate
in 2poly(1/ǫ) time. While the running time must be exponential in 1/ǫ, unless there exists a randomized
subexponential algorithm for SAT, it remains a neat open question whether the query complexity can be
reduced to polynomial in 1/ǫ.
For bipartite graphs, a (1, ǫn)-estimate can be computed in dO(1/ǫ2) time. This follows from the relation
between the maximum matching size and the minimum vertex size captured by Konig's theorem and fast
approximation algorithms for the maximum matching size [NO08, YYI09].
Ideas similar to those discussed in this paper are used to construct sublinear time estimations of other
parameters of sparse combinatorial objects, such as maximum matching, set cover, constraint satisfaction
[NO08, YYI09, Yos11]. In the related setting of property testing, sublinear time algorithms are given for
testing any class of graphs with a fixed excluded minor and any property of graphs with a fixed excluded
minor [CSS09, BSS08, Ele10, HKNO09, NS11].
There are also other works on sublinear algorithms for various other graph measures such as the mini-
mum weight spanning tree [CRT05, CS09, CEF+05], the average degree [Fei06, GR08], and the number of
stars [GRS10].
2 The Oracle-Based Algorithm
Let G = (V, E) be an undirected graph with n vertices and m edges, where we allow G to contain parallel
edges and self-loops. Let d denote the maximum degree in the graph, and let ¯d denote the average degree.
Consider a ranking π : E → [m] of the edges in G = (V, E). As noted in the introduction, such a ranking
determines a maximal matching Mπ(G). Given Mπ(G), we define a vertex cover Cπ(G) as the set of all
endpoints of edges in Mπ(G). Therefore, VCopt ≤ Cπ(G) ≤ 2VCopt, where VCopt is the minimum size
of a vertex cover in G. We assume without loss of generality that there are no isolated vertices in G, since
such vertices need not belong to any vertex cover. We shall use the shorthand Mπ and Cπ for Mπ(G) and
Cπ(G), respectively, when G is clear from the context.
Assume we have an oracle VOπ for a vertex cover based on a ranking π of the edges, where VOπ(v) =
TRUE if v ∈ Cπ(G), VOπ(v) = FALSE otherwise. The next lemma follows by applying an additive Chernoff
bound.
Lemma 2.1 For any fixed choice of π, let C = Cπ(G). Suppose that we uniformly and independently select
s = Θ( 1
ǫ2 ) vertices v from V . Let t be a random variable equal to the number of selected vertices that
4
belong to C. With high constant probability,
C − ǫn ≤
t
s
· n ≤ C + ǫn .
Algorithm 1, provided below, implements an oracle VOπ, that given a vertex v, decides whether v ∈ Cπ.
This oracle uses another oracle, MOπ (described in Algorithm 2) that given an edge e, decides whether
e ∈ Mπ. Both oracles can determine π(e) for any edge e of their choice. The oracle MOπ essentially
emulates the greedy algorithm for finding a maximal matching (based on the ranking π). We assume that
once the oracle for the maximal matching decides whether an edge e belongs to Mπ or not, it records this
information in a data structure that allows to retrieve it later. By Lemma 2.1, if we perform Θ(1/ǫ2) calls to
VOπ, we can get an estimate of the size of the vertex cover Cπ up to an additive error of (ǫ/2)n, and hence
we can obtain a (2, ǫ)-estimate (as defined in the introduction) of the size of a minimum vertex cover in G.
Hence our focus is on upper bounding the query complexity and running time of the resulting approximation
algorithm when π is selected uniformly at random.
Algorithm 1: An oracle VOπ(v) for a vertex cover based on a ranking π of the edges. Given a vertex v, the
oracle returns TRUE if v ∈ Cπ and it returns FALSE otherwise.
1
Let e1, . . . , et be the edges incident to the vertex v in order of increasing rank (that is,
π(ei+1) > π(ei)).
for i = 1, . . . , t do
if MOπ(ei) = TRUE then
return TRUE
return FALSE
2
3
4
5
1
2
3
4
5
6
7
8
9
Algorithm 2: An oracle MOπ(e) for a maximal matching based on a ranking π of the edges. Given an edge
e, the oracle returns TRUE if e ∈ Mπ and it returns FALSE otherwise.
if MOπ(e) has already been computed then
return the computed answer.
Let e1, . . . , et be the edges that share an endpoint with e, in order of increasing rank (that is,
π(ei+1) > π(ei)).
i ← 1.
while π(ei) < π(e) do
if MOπ(ei) = TRUE then
return FALSE
else
i ← i + 1.
10
return TRUE
We start (in Section 3) by bounding the expected number of calls made to the maximal-matching oracle
MOπ in the course of the execution of a call to the vertex-cover oracle VOπ. This bound depends on
the average degree in the graph and on the ratio between the maximum degree and the minimum degree.
A straightforward implementation of the oracles would give us an upper bound on the complexity of the
5
In Section 4 we describe a
algorithm that is a factor of d larger than our final near-optimal algorithm.
sophisticated method of simulating the behavior of the oracle MOπ for randomly selected ranking π, which
is selected "on the fly". Using this method we obtain an algorithm with only a polylogarithmic overhead (as
a function of d) over the number of recursive calls. Thus, for graphs that are close to being regular, we get
an algorithm whose complexity is O(d/ǫ2). In Section 5 we address the issue of variable degrees, and in
particular, show how to get a nearly-linear dependence on the average degree.
3 Bounding the Expected Number of Calls to the Maximal-Matching Oracle
For a ranking π of the edges of a graph G and a vertex v ∈ V , let N (π, v) = NG(π, v) denote the number
of different edges e such that a call MOπ(e) was made to the maximal matching oracle in the course of the
computation of VOπ(v). Let Π denote the set of all rankings π over the edges of G. Our goal is to bound
the expected value of N (π, v) (taken over a uniformly selected ranking π and vertex v). We next state our
first main theorem.
Theorem 3.1 Let G be a graph with m edges and average degree ¯d, and let the ratio between the maximum
degree and the minimum degree in G be denoted by ρ. The average value of N (π, v) taken over all rankings
π and vertices v is O(ρ · ¯d). That is:
1
m!
·
1
n
·Xπ∈ΠXv∈V
N (π, v) = O(ρ · ¯d) .
(1)
If the graph is (close to) regular, then the bound we get in Theorem 3.1 is O( ¯d) = O(d). However, for
graphs with varying degrees the bound can be Θ(d2). As noted previously, we later show how to deal with
variable degree graphs without having to pay a quadratic cost in the maximum degree.
As noted in the introduction, our analysis builds on the work of Yoshida et al. [YYI09]. While our
analysis does not reduce to theirs4, it uses many of their ideas. We start by making a very simple but useful
observation about the maximal matching oracle MOπ (Algorithm 2), which follows immediately from the
definition of the oracle.
Observation 3.2 For any edge e, consider the execution of MOπ on e. If in the course of this execution,
a recursive call is made to MOπ on another edge e′, then necessarily π(e′) < π(e). Therefore, for any
consecutive sequence of (recursive) calls to edges eℓ, . . . , e1, π(eℓ) > π(eℓ−1) > . . . > π(e1).
In order to prove Theorem 3.1 we introduce more notation. For any edge e ∈ E, we arbitrarily label its
endpoints by va(e) and vb(e) (where if e is a self-loop then va(e) = vb(e), and if e and e′ are parallel edges,
then va(e) = va(e′) and vb(e) = vb(e′)). For a ranking π and an index k, let πk denote the edge e such that
π(e) = k.
We say that an edge e is visited if a call is made on e either in the course of an oracle computation of
VOπ(va(e)) or VOπ(vb(e)) (that is, as a non-recursive call), or in the course of an oracle computation of
4Indeed, we initially tried to find such a reduction. The main difficulty we encountered is that the vertex cover oracle, when
executed on a vertex v, performs calls to the maximal matching oracle on the edges incident to v until it gets a positive answer (or
all the incident edges return a negative answer). While the analysis of [YYI09] gives us an upper bound on the expected number
of recursive calls for a given edge, it is not clear how to use such a bound without incurring an additional multiplicative cost that
depends on the degree of the vertices.
6
MOπ(e′) for an edge e′ that shares an endpoint with e (as a recursive call). For a vertex v and an edge e, let
G(v, e) equal 1 if e is visited in the course of the execution of VOπ(v). Using the notation
X π(v, e) = X π
just introduced, we have that
X π(v, e) .
(2)
N (π, v) = Xe∈E
Observation 3.3 Let e = (v, u). If X π(v, e) = 1, then for each edge e′ that shares the endpoint v with e
and for which π(e′) < π(e) we have that MOπ(e′) = FALSE.
To verify Observation 3.3, assume, contrary to the claim, that there exists an edge e′ as described in the
observation and MOπ(e′) = TRUE. We first note that by Observation 3.2, the edge e cannot be visited in
the course of an execution of MOπ on any edge e′′ = (v, w) such that π(e′′) < π(e) (and in particular this
is true for e′′ = e′). Since VOπ(v) performs calls to the edges incident to v in order of increasing rank,
if MOπ(e′) = TRUE, then VOπ(v) returns TRUE without making a call to MOπ(e). This contradicts the
premise of the observation that X π(v, e) = 1.
The next notation is central to our analysis. For k ∈ [m] and a fixed edge e:
Xk(e)
def
= Xπ∈Π(cid:16)X π(va(πk), e) + X π(vb(πk), e)(cid:17) .
(3)
That is, Xk(e) is the total number of calls made to the maximal matching oracle on the edge e when summing
over all rankings π, and performing an oracle call to the vertex-cover oracle from one of the endpoints of
πk. Observe that
where deg(v) denotes the degree of v in the graph, and for simplicity of the presenation we count each
self-loop as contributing 2 to the degree of the vertex. We next give an upper bound on Xk(e).
Lemma 3.4 For every edge e and every k ∈ [m]:
Xk(e) ≤ 2(m − 1)! + (k − 1) · (m − 2)! · d .
In order to prove Lemma 3.4, we establish the following lemma.
Lemma 3.5 For every edge e and every k ∈ [m − 1]:
Xk+1(e) − Xk(e) ≤ (m − 2)! · d .
Before proving Lemma 3.5, we show how Lemma 3.4 easily follows from it.
Proof of Lemma 3.4: We prove the lemma by induction on k. For the base case, k = 1,
X1(e) =Xπ (cid:16)X π(va(π1), e) + X π(vb(π1), e)(cid:17) .
(5)
(6)
(7)
By the definition of the vertex-cover oracle, when starting from either va(π1) or from vb(π1), only a single
call is made to the maximal matching oracle. This call is on the edge π1, which returns TRUE without making
7
mXk=1
Xk(e) = Xπ∈ΠXv∈V
deg(v) · X π(v, e)
(4)
any further calls, because all edges (that share an endpoint with π1) have a larger rank. This implies that if
e = π1, then X π(va(π1), e) = X π(vb(π1), e) = 1, and otherwise X π(va(π1), e) = X π(vb(π1), e) = 0.
For any fixed e, the number of rankings π such that e = π1 is simply (m − 1)! and so X1(e) = 2(m − 1)!,
as required.
We now turn to the induction step. Assuming the induction hypothesis holds for k − 1 ≥ 1, we prove it
for k > 1. This follows directly from Lemma 3.5 (and the induction hypothesis):
Xk(e) ≤ Xk−1(e) + (m − 2)! · d
≤ 2(m − 1)! + (k − 2) · (m − 2)! · d + (m − 2)! · d
= 2(m − 1)! + (k − 1) · (m − 2)! · d ,
(8)
(9)
(10)
and the lemma is established.
Proof of Lemma 3.5: Throughout the proof we fix k and e. For a ranking π, let π′ be defined as follows:
π′
k+1 = πk, π′
j = πj for every j /∈ {k, k + 1}.
k = πk+1 and π′
Observation 3.6 If π and π′ are as defined above, then for each edge e where π(e) < k (and therefore,
π′(e) < k): MOπ(e) = MOπ′
(e).
Observation 3.6 is true due to the fact that if π(e) < k then by the definition of π′, we have that π′(e) = π(e).
Since in a recursive call we only go to an edge with a lower rank (see Observation 3.2), we get that the
execution of MOπ(e) is equivalent to the execution of MOπ′
(e).
We shall use the notation Πk for those rankings π in which πk and πk+1 share a common endpoint. Note
that if π ∈ Πk, then π′ ∈ Πk as well (and if π /∈ Πk, then π′ /∈ Πk). For two edges e = (v1, v2) and
e′ = (v2, v3) (which share a common endpoint v2), we let vc(e, e′) = vc(e′, e) = v2 ('c' for 'common')
and vd(e, e′) = v1, vd(e′, e) = v3 ('d' for 'different'). If e and e′ are parallel edges, then we let vd(e, e′) =
vd(e′, e) be va(e) = va(e′) and vc(e, e′) = vc(e′, e) be vb(e) = vb(e′). If e is a self-loop on a vertex v1 that
is also an endpoint of e′ (so that v2 = v1), then vd(e, e′) = vc(e, e′) = v1.
For any edge e and for 1 ≤ k ≤ m − 1,
Xk+1(e) − Xk(e)
= Xπ (cid:16)X π(va(πk+1), e) + X π(vb(πk+1), e)(cid:17) −Xπ (cid:16)X π(va(πk), e) + X π(vb(πk), e)(cid:17)
= Xπ∈Πk(cid:16)X π(va(πk+1), e) + X π(vb(πk+1), e)(cid:17) − Xπ∈Πk(cid:16)X π(va(πk), e) + X π(vb(πk), e)(cid:17)
+ Xπ /∈Πk(cid:16)X π(va(πk+1), e) + X π(vb(πk+1), e)(cid:17) − Xπ /∈Πk(cid:16)X π(va(πk), e) + X π(vb(πk), e)(cid:17) (12)
= Xπ∈Πk
+ Xπ∈Πk
+ Xπ /∈Πk(cid:16)X π(va(πk+1), e) + X π(vb(πk+1), e)(cid:17) − Xπ /∈Πk(cid:16)X π(va(πk), e) + X π(vb(πk), e)(cid:17).(15)
X π(vc(πk+1, πk), e) − Xπ∈Πk
X π(vd(πk+1, πk), e) − Xπ∈Πk
X π(vd(πk, πk+1), e)
X π(vc(πk, πk+1), e)
(11)
(13)
(14)
8
By the definition of vc(·, ·), for every π ∈ Πk we have that vc(πk+1, πk) = vc(πk, πk+1) and so
X π(vc(πk+1, πk), e) = X π(vc(πk, πk+1), e) ,
(16)
implying that the expression in Equation (13) evaluates to 0. Since π′ ∈ Πk if and only if π ∈ Πk, we get
(17)
(18)
(19)
(20)
and
X π′
(vd(π′
k+1, π′
X π(vd(πk+1, πk), e) = Xπ′∈Πk
that Xπ∈Πk
Xπ /∈Πk(cid:16)X π(va(πk+1), e) + X π(vb(πk+1), e)] = Xπ′ /∈Πk(cid:16)X π′
= Xπ /∈Πk(cid:16)X π′
k), e) = Xπ∈Πk
X π′
(vd(π′
k+1, π′
k), e) ,
(va(π′
k+1), e) + X π′
(vb(π′
(va(π′
k+1), e) + X π′
(vb(π′
k+1), e)(cid:17)
k+1), e)(cid:17) .
Therefore,
Xk+1(e) − Xk(e) = Xπ∈Πk
k), e) − Xπ∈Πk
X π′
(vd(π′
k+1, π′
X π(vd(πk, πk+1), e)
(va(π′
k+1), e) + X π′
+ Xπ /∈Πk(cid:16)X π′
k+1), e)(cid:17)
− Xπ /∈Πk(cid:16)X π(va(πk), e) + X π(vb(πk), e)(cid:17) .
(vb(π′
The next useful observation is that for every π /∈ Πk (and for every e and j ∈ {a, b}),
X π′
(vj(π′
k+1), e) = X π(vj(πk), e) .
This follows by combining the fact that vj(π′
k+1) = vj(πk) with Observations 3.2 and 3.6.
By combining Equation (19) with Equation (20) we obtain that
Xk+1(e) − Xk(e) = Xπ∈Πk
X π′
(vd(π′
k+1, π′
k), e) − Xπ∈Πk
X π(vd(πk, πk+1), e) .
(21)
Therefore, we need only consider executions in which the underlying rankings π and π′ belong to Πk,
k) = vd(πk, πk+1). We shall use the shorthand
k+1) = vd(πk+1, πk). For an
and the execution starts from the vertex v1(π′) = vd(π′
notation v2(π′) = vc(π′
illustration, see Figure 1. We shall make use of the following simple observation.
k) = vc(πk, πk+1), and v3(π′) = vd(π′
k+1, π′
k+1, π′
k, π′
Observation 3.7 Let e be a self-loop. For any vertex v and ranking π, if in the course of the execution of
VOπ(v) a call is made to MOπ(e), then MOπ(e) = TRUE.
Observation 3.7 is true since if a call is made to MOπ(e) where e is a self-loop, i.e., e = (v, v) for some
vertex v, then from Observation 3.3 we know that all other edges incident to v with ranks smaller than π(e)
return FALSE. Therefore, by the definition of MOπ we get that MOπ(e) = TRUE.
We would like to understand when X π′(v1(π′), e) = 1 while X π(v1(π′), e) = 0. We consider three
possible cases (for an illustration see Figure 2) :
9
(a)
v1(π
′)
π
′
k+1
= πk
π
′
k
= πk+1
v2(π
′)
π
′
k+1
= πk
v3(π
′)
(b)
π
′
k
= πk+1
v1(π
′)
v2(π
′)
(c)
v1(π
′)
v2(π
′)
π
′
k+1
= πk
(d)
v1(π
′)
π
′
k+1
= πk
v2(π
′)
π
′
k
= πk+1
π
′
k
= πk+1
π
′
k+1
= πk
(e)
v1(π
′)
v2(π
′)
π
′
k
= πk+1
π
′
k+1
= πk
(f)
v1(π
′)
π
′
k
= πk+1
Figure 1: An illustration for the various cases in which π ∈ Πk (i.e., π′
need to compare the executions of VOπ(v1(π′)) and VOπ′
We refer to the different cases (a) -- (f) in the analysis.
(v1(π′)) (where v1(π′) = vd(π′
k and π′
k+1 share at least one endpoint) and we
k) = vd(πk, πk+1)).
k+1, π′
1. e = (v1(π′), v2(π′)) (so that π′(e) = k + 1 and π(e) = k). In this case, if X π′(v1(π′), e) = 1,
then X π(v1(π′), e) = 1. To verify this, note that if X π′
(v1(π′), e) = 1 then by Observation 3.3,
MOπ′
(e′) = FALSE for each edge e′ where v1(π′) is one of its endpoints and π′(e′) < k + 1. By
applying Observation 3.6 we get that for each edge e′ such that π(e′) < k we have that MOπ(e′) =
MOπ′
(e′). Therefore, for each edge e′ such that π(e′) < k and v1(π′) is one of its endpoints we have
that MOπ(e′) = MOπ′
We note that if π′
that MOπ′
the course of the execution of VOπ′
k is a self-loop (see cases (c) and (f) in Figure 1), then by Observation 3.7 we have
k+1 = e will not be visited in
k) = TRUE. By the definition of VOπ′ this implies that π′
(e′) = FALSE. Hence X π(v1(π′), e) = 1.
(v1(π′), e) is necessarily 0.
(v1(π′)), so that X π′
(π′
2. e = (v2(π′), v3(π′)), (so that π(e) = k + 1 and π′(e) = k). In this case it is possible (though not
necessary) that X π′(v1(π′), e) = 1 and X π(v1(π′), e) = 0.
3. e /∈ {(v1(π′), v2(π′)), (v2(π′), v3(π′))}. In this case it is also possible (though not necessary) that
(v1(π′), e) = 1 and X π(v1(π′), e) = 0.
X π′
Out of all cases illustrated in Figure 1, this is possible only in cases (a) and (b). We next explain why
it is not possible in all other cases.
• Case (c). If VOπ′
(v1(π′)) visits e before it visits π′
(v1(π′)) visits π′
k, then so does VOπ(v1(π′)) (from Observa-
k first, but since it is a self-loop, from Observation 3.7
tion 3.6). Otherwise, VOπ′
we have that MOπ′
• Case (d). If VOπ′
k) = TRUE. By the definition of VOπ′ we get that X π′(v1(π′), e) = 0.
(π′
(v1(π′)) visits e before it visits π′
(v1(π′)) visits π′
servation 3.6). Otherwise, if VOπ′
sive calls without visiting π′
VOπ′
(v1(π′)) will visit π′
k+1 and π′
k+1, then so does VOπ(v1(π′)) (from Ob-
k+1 and e in the same sequence of recur-
k, then so does VOπ(v1(π′)). If there is no such sequence, then
k is a self-loop, from Observation 3.7 we have that
k. Since π′
10
(π′
k) = TRUE, implying that MOπ′
MOπ′
k+1) = f alse. Therefore, the sequence of recur-
sive calls that visits e in the execution of VOπ′
(v1(π′)), starts from an edge incident to v1(π′)
whose rank is greater than k + 1, and the same sequence of calls is made in the execution of
VOπ(v1(π′)).
(π′
• Case (e). Since the edges are parallel, if there is a sequence of recursive calls that visits e in
(v1(π′)), then there is such a sequence in the execution of VOπ(v1(π′)),
k while the second includes πk
the execution of VOπ′
where the only difference is that the first sequence includes π′
(which are parallel edges).
• Case (f). If VOπ′
(v1(π′)) visits e in a sequence of recursive calls that starts with an edge having
k is a
k), then it returns TRUE, causing the
(v1(π′)) to terminate without visiting any additional edges (so that e cannot
rank smaller than k, then from Observation 3.6 so will VOπ(v1(π′)). Otherwise, since π′
self-loop, by Observation 3.7, if a call is made to MOπ′
execution of VOπ′
be visited in a sequence of recursive calls that starts with an edge having rank at least k).
(π′
1.
2.
3.
π
′
k+1
= πk
v1(π
′)
e
π
′
k
= πk+1
v2(π
′)
v3(π
′)
π
′
k+1
= πk
v1(π
′)
v2(π
′)
π
′
k
= πk+1
e
v3(π
′)
π
′
k+1
= πk
v1(π
′)
v2(π
′)
π
′
k
= πk+1
v3(π
′)
e
Figure 2: An illustration for the three possible (sub-)cases when π′ ∈ Πk: 1. e = (v1(π′), v2(π′)); 2. e =
(v2(π′), v3(π′)); 3. e /∈ {(v1(π′), v2(π′)), (v2(π′), v3(π′))}. This illustration corresponds to Case (a) in Figure 1 (i.e.,
no self-loops and no parallel edges).
For a fixed edge e we shall use the following notation for the sets of rankings that correspond to the last
two cases described above. Specifically:
• Let Πe,1 = Πe,1
k
denote the set of all rankings π′ ∈ Πk where e = (v2(π′), v3(π′)) and
X π′
(v1(π′), e) = 1. (Here we shall make the worst case assumption that X π(v1(π′), e) = 0).
• Let Π¬e = Π¬e
k denote the set of all rankings π′ ∈ Πk where e /∈ {(v1(π′), v2(π′)), (v2(π′), v3(π′))}
and X π′(v1(π′), e) = 1 while X π(v1(π′), e) = 0.
Thus, Xk+1(e) − Xk(e) ≤ Πe,1 + Π¬e. In order to upper bound Πe,1 + Π¬e, we consider another set
of rankings:
11
• Let Πe,0 = Πe,0
k
(v1(π′), e) = 0.
X π′
denote the set of all rankings π′ ∈ Πk such that e = (v2(π′), v3(π′)) and
By the definition of Πe,1 and Πe,0, we have that
Πe,1 + Πe,0 ≤ (m − 2)! · d .
(22)
This is true since each ranking π′ ∈ Πe,1 ∪ Πe,0 is determined by first setting π′(e) = k, then selecting
another edge incident to the endpoint v2(π′) of e (if e is a self-loop then v2(π′) = v1(π′)) and giving it rank
k + 1 (where there are at most deg(v2(π′)) − 1 ≤ d − 1 such edges), and finally selecting one of the possible
(m − 2)! rankings for the remaining edges. We next prove that Π¬e ≤ Πe,0, from which Lemma 3.5
follows.
To this end we prove the next claim.
Claim 3.8 There is an injection from Π¬e to Πe,0.
The proof of Claim 3.8 is very similar to a proof of a corresponding claim in [YYI09], but due to our need to
extend the proof to a graph with self-loops and parallel edges, and also due to several additional differences,
we include it here for completeness.
Proof: We start by making the following observations:
Observation 3.9 If π′ ∈ Π¬e and we are in Case (a) as illustrated in Figure 1, then in the course of the
execution of VOπ′
k and e
at the end. That is, there is a sequence of recursive calls corresponding to a path of edges (eℓ, eℓ−1 . . . e1)
such that eℓ = π′
(v1(π′)) there is a consecutive sequence of recursive calls that includes π′
k+1, π′
k+1, eℓ−1 = π′
k and e1 = e.
To verify Observation 3.9, note that since π′ ∈ Π¬e we know that X π′
0. The only difference between the execution of VOπ′
call MOπ′
that VOπ′
calls MOπ′
the other, since by Observation 3.2, the ranks can only decrease in a sequence of recursive calls.
(π′
(v1(π′)) and VOπ(v1(π′)) will create different sequences of recursive calls is when VOπ′
(π′
(v1(π′), e) = 1 and X π(v1(π′), e) =
k+1) can
k) = MOπ(πk+1). Thus, the only way
(v1(π′))
k). Furthermore, these two calls have to be one after
k+1) = MOπ(πk) cannot call MOπ(π′
(v1(π′)) and VOπ(v1(π′)) is that MOπ′
k+1) and then MOπ′
k+1) calls MOπ′
k) but MOπ(π′
(π′
(π′
(π′
Observation 3.10 If π′ ∈ Π¬e and we are in Case (b) as illustrated in Figure 1, then in the course of
the execution of VOπ′
k, and
ends with e (so that, in particular, it does not include π′
k+1). That is, there is a sequence of recursive calls
corresponding to a path of edges (eℓ−1 . . . e1) such that eℓ−1 = π′
(v1(π′)) there is a consecutive sequence of recursive calls that starts with π′
k and e1 = e.
To verify Observation 3.10, note that since π′ ∈ Π¬e we know that X π′
(v1(π′), e) = 1 and X π(v1(π′), e) =
0. The execution of VOπ′
(v1(π′)) cannot visit e in the course of a sequence of recursive calls starting
from an edge incident to v1(π′) where the edge has ranking smaller k. Otherwise, from Observation 3.6
we would get that VOπ(v1(π′)) also visits e which contradicts the premise that π′ ∈ Π¬e. We also
know that VOπ′
k+1. If it would have, then since it is a self-loop, from Observa-
tion 3.7, MOπ′
(v1(π′)) to terminate without visiting e, which contradicts
X π′
(v1(π′)) cannot visit π′
k+1) = TRUE, causing VOπ′
(π′
(v1(π′), e) = 1.
12
k+1, eℓ−1 = π1
We shall now prove Claim 3.8. Let π1 be a ranking in Π¬e (so that π1(e) /∈ {k, k + 1}). By the
definition of Π¬e and by Observations 3.9 and 3.10, we have the following. In Case (a), the execution of
VOπ1
(v1(π1)) induces a sequence of (recursive) calls to the maximal matching oracle, where this sequence
corresponds to a path P = (eℓ, . . . , e1) such that eℓ = π1
k, and e1 = e. In Case (b), the
execution of VOπ1
(v1(π1)) induces a sequence of (recursive) calls to the maximal matching oracle, where
this sequence corresponds to a path P ′ = (eℓ−1, . . . , e1) such that eℓ−1 = π1
k, and e1 = e. Since in Case
(b) P is also a path in the graph, we may refer to the path P in both cases (and take into account, if needed,
that in Case (b) eℓ = π1
k−1 is a self-loop and is not part of the sequence of recursive calls that reaches e).
While we do not know the rankings of the edges eℓ−2, . . . e1, we know from Observation 3.2 that they are
in monotonically decreasing order, and that they are all smaller than k. We also know that the path does
not include any parallel edges. This is true since if et and et−1 are adjacent edges in the path P and they
are parallel edges, then from Observation 3.11 π′(et−1) < π′(et). But since they are parallel, they have
the same endpoints, therefore, by the definition of VOπ′ and of MOπ′, the vertex/edge from which the call
MOπ′
k+1 in Case (b),
the the only edge along the path P that might be a self-loop is e. Otherwise, from Observation 3.7, the
self-loop will return true, and thus path P will not visit e.
(et−1). Furthermore, with the exception of π′
(et) was made, would have called MOπ′
σ(ℓ), . . . π1
We can write P as P = (π1
σ(1)) where σ(i) = π1(ei), so that σ(ℓ) = k + 1 and σ(ℓ − 1) = k.
We next define a mapping ϕ between rankings, such that ϕ(π1) = π0, where we shall show that π0 ∈ Πe,0,
and that ϕ is one-to-one. The ranking π0 is defined as follows by "rotating" the ranks of the edges on P (and
leaving the ranks of all other edges as in π1). Namely, π0(e2) = k + 1, π0(e1) = k, and π0(ej) = σ(j − 2)
for every 3 ≤ j ≤ ℓ. For an illustration, see Table 1. We first verify that ϕ is a projection from Π¬e to Πe,0.
eℓ
eℓ−1
Rank in π1 σ(ℓ) = k + 1 σ(ℓ − 1) = k
Rank in π0 σ(ℓ − 2)
σ(ℓ − 3)
e2
e3
. . .
. . . σ(3) σ(2)
. . . σ(1) σ(ℓ) = k + 1 σ(ℓ − 1) = k
e1 = e
σ(1)
Table 1: Ranking of P = (eℓ, . . . , e1) in π1 and in π0
Namely, we need to show that:
• π0 ∈ Πk, i.e., π0
k+1 and π0
k share an endpoint v2(π0), and e = (v2(π0), v3(π0)).
• X π0
(v1(π0), e) = 0 (that is, the execution of VOπ0
(v1(π0)) does not create a call to MOπ0
(e)). In
other words, (the execution of) VOπ0
(v1(π0)) does not visit e.
The first item directly follows from the definition of π0. We thus turn to the second item. Recall that by
our notational convention, v1(π0) = vd(π0
k) = vd(e2, e1) (i.e, it is the endpoint that e2 does not share
with e1) so that it is the common endpoint of e2 and e3, i.e., vc(e2, e3). Since
k+1, π0
π0(e3) = σ(1) < σ(ℓ) = k + 1 = π0(e2) ,
(23)
the execution of VOπ0
VOπ0
(v1(π0)), the call to MOπ0
(v1(π0)) will visit e3 before visiting e2. Since π0(e) = k, during the execution of
(e3) will not cause a recursive call to MOπ0
(e).
Observe that in the execution of VOπ1
(e3) creates a recursive call on e2
(since e2 follows e3 on the path P ). Therefore, it must be the case that MOπ1
(e′)=FALSE for every e′ that
has a common endpoint with e3 and such that π1(e′) < σ(2). By the definition of ϕ, all edges that are not
(v1(π1)), the call to MOπ1
13
on the path P have the same ranks in π0 and in π1. Therefore, all edges with rank lower than σ(1) have
the same rank in π1 and in π0. It follows that for every e′ that has a common endpoint with e3 and such
that π1(e′) < σ(2), MOπ0
(v1(π0))
returns TRUE without visiting e1 = e, as required.
(e′) = FALSE. We can conclude that MOπ0
(e3) = TRUE and so VOπ0
ℓ2 , e2
ℓ2−1 . . . e2
ℓ1, e1
ℓ1−1 . . . e1
1) and P 2 = (e2
It remains to show that ϕ is an injection from Π¬e to Πe,0. Assume, contrary to the claim, that ϕ is not
an injection. That is, there are two different rankings π1 6= π2 ∈ Π¬e where ϕ(π1) = ϕ(π2). Let P 1 =
(e1
1) be the paths that correspond to the sequence of recursive calls
to the maximal matching oracle, in the executions of VOπ1
(v1(π2)) respectively, where
ℓ2−1) = k (recall that if π1 corresponds to
e1
1 = e2
ℓ1) = π2(e2
Case (b), then e1
is a self-loop and is not actually part of the sequence of recursive calls that reaches e, and an
ℓ1
s−1 . . . e2
analogous statement holds for π2). Let s be the largest index such that (e1
1).
We denote this common subsequence by (es, es−1 . . . e1). Observe that s ≥ 2. This is true since: (1) By the
definitions of the paths, e1
2) = k +1,
it holds that e1
1 = e, and (2) given that ϕ(π1) = ϕ(π2) = π0 and π0(e1
ℓ2) = k + 1 and π1(e1
(v1(π1)) and VOπ2
ℓ1−1) = π2(e2
1 = e, π1(e1
2) = π0(e2
s, e1
s−1 . . . e1
1) = (e2
s, e2
1 = e2
2 = e2
2.
By the definitions of ϕ and s we have that π1(ei) = π2(ei) for each i ∈ [s − 2]. Thus, σ1(i) = σ2(i)
for each i ∈ [s − 2], where we shall sometimes use the shorthand σ(i) for this common value. For an
illustration, see Table 2.
π0(e1)
π0(e2)
π0(e3)
...
π0(es−1) σ1(s − 3)
π0(es)
σ1(s − 2)
Rank from ϕ(π1) Rank from ϕ(π2)
σ1(ℓ1 − 1) = k
σ1(ℓ1) = k + 1
σ1(1)
...
σ2(ℓ2 − 1) = k
σ2(ℓ2) = k + 1
σ2(1)
...
σ2(s − 3)
σ2(s − 2)
Table 2: Ranks of edges e1
1 = e2
1 . . . e1
s−2 = e2
s−2 are equal in π1 and π2
.
The next observation will be useful.
Observation 3.11 For every edge e′, if π1(e′) < min{σ1(s − 1), σ2(s − 1)} or π2(e′) < min{σ1(s −
1), σ2(s − 1)}, then π1(e′) = π2(e′). Therefore, MOπ1
(e′) for e′ such that π1(e′) = π2(e′) <
min{σ1(s − 1), σ2(s − 1)}.
(e′) = MOπ2
We consider two cases:
1. P 2 is a suffix of P 1 or P 1 is a suffix of P 2. Without loss of generality, assume that P 2 is a suffix of
P 1, so that s = ℓ2.
2. Otherwise (neither path is a suffix of the other), assume without loss of generality that σ1(s − 1) <
σ2(s − 1).
In both cases, since e1
π0(e1
s+1) = π2(e1
s+1 is not on P 2, ϕ, when applied to π2 does not change the ranking of e1
s+1. That is,
s+1). Since (by the definition of ϕ) π0(e1
s+1) = σ1(s − 1), we get that
π2(e1
s+1) = σ1(s − 1) = π1(e1
s−1) .
(24)
14
In the first case (where P 2 is a suffix of P 1), we have that σ2(s − 1) = k, while σ1(s − 1) < k, and so
σ1(s − 1) < σ2(s − 1) (= π2(e2
s−1)) .
(25)
In the second case, this inequality was made as an explicit assumption.
(a)
v1(π
′)
π
′
k+1
= πk
π
′
k
= πk+1
v2(π
′)
π
′
k+1
= πk
v3(π
′)
(b)
π
′
k
= πk+1
v1(π
′)
v2(π
′)
(c)
v1(π
′)
v2(π
′)
π
′
k+1
= πk
(d)
v1(π
′)
π
′
k+1
= πk
v2(π
′)
π
′
k
= πk+1
π
′
k
= πk+1
π
′
k+1
= πk
(e)
v1(π
′)
v2(π
′)
π
′
k
= πk+1
π
′
k+1
= πk
(f)
v1(π
′)
π
′
k
= πk+1
Figure 3: An illustration for the proof of Claim 3.8.
s+1 = π1
(e2
s) visits e1
(e1
We thus have that the execution of MOπ2
k+1 is a self-loop, from Observation 3.7, MOπ2
s+1 before visiting e2
s+1). If we are in Case (b) and P 1 = (π1
s−1. We would like to under-
stand what occurs in the call to MOπ2
k+1, π1
k, e), i.e., s = k, then,
s = e1
since e1
s)
s−1 = e, but this stands in contradiction to the definition of P 2. If we are in
returns FALSE without visiting e2
Case (a), then since the path P 1 corresponds to a sequence of recursive calls to the maximal-matching oracle,
s+1 and such that π1(e′) < σ1(s) = π1(e1
we have that for every edge e′ that shares an end-point with e1
s),
the call to MOπ1
(e′) returns FALSE. Combining this with Observation 3.11, we get that for every edge
e′ that shares an end-point with e1
(e′) returns FALSE.
By Equation (24) we get that MOπ2
s) returns FALSE without
s−1, but this stands in contradiction to the definition of P 2.
visiting e2
Having established Claim 3.8, the proof of Lemma 3.5 is completed.
s+1 and such that π2(e′) < σ1(s), the call to MOπ2
s+1) returns TRUE. Hence, MOπ2
(e1
s+1) = TRUE. Hence MOπ2
(e2
(e1
s = e1
(e2
(Claim 3.8)
We are now ready to prove Theorem 3.1.
Proof of Theorem 3.1: Recall that d denotes the maximum degree, ¯d denotes the average degree and ρ
denotes the ratio between the maximum degree and the minimum degree, which is denoted by dmin (where
the latter is at least 1 since we assumed without loss of generality that there are no isolated vertices). By
combining Equations (2) and (4) and applying Lemma 3.4 (as well as recalling that we counted each self-
15
loop as contributing 2 to the degree of a vertex), we get that:
1
m!
·
1
n
·Xπ∈ΠXv∈V
N (π, v)
≤
≤
1
m!
1
m!
·
·
1
n
1
n
= O(cid:18) m
n
·
·
·
1
1
Xk(e)
mXk=1
· m ·(cid:18)m · 2(m − 1)! +
2dminXe∈E
dmin(cid:19) = O(ρ · ¯d) ,
2dmin
d
m · m − 1
2
· (m − 2)! · d(cid:19)
(26)
(27)
(28)
and we obtain the bound claimed.
4 Limiting the Exploration of Neighbor Sets
The analysis in the previous section suffices to show an algorithm whose query complexity and running time
are a factor of d larger than the number of oracle calls that it makes. The factor of d in this expression is
due to querying all edges that are incident to the endpoints of each edge for which a call to the maximal
matching oracle is made (where, as we explain momentarily, a random ranking can be selected in an online
fashion).
This section is devoted to a method for selecting incident edges of low rank efficiently without querying
entire neighborhoods of relevant vertices. By applying the method, we reduce the query complexity and the
running time by a factor of almost d. The main challenges here are to ensure that the ranks of encountered
edges indeed come from the uniform distribution over all permutations and that the same decision with
respect to a rank of an edge is made at both endpoints of the edges.
Replacing a random ranking with random numbers. The oracle construction described as Algorithm 1
and Algorithm 2 uses a random ranking π : E → [m] of edges. We start by replacing a random ranking
of edges with random real numbers in (0, 1] selected uniformly and independently for every edge e ∈ E,
yielding a vector σ : E → (0, 1] which we use in the same way as the ranking π. Since the probability that
two edges are assigned the same real number is 0, whenever the oracle compares the ranks of two edges e
and e′, it can check whether σ(e) < σ(e′), instead of whether π(e) < π(e′), effectively yielding a random
ranking of edges. Since each σ(e) is independent, this small conceptual shift allows one to generate σ(e) at
random in an easier manner and to simplify the analysis. Though it is not possible to generate and store real
numbers in (0, 1], we later introduce a proper discretization.
4.1 A Data Structure for Accessing Neighbors
The oracle described as Algorithms 1 and 2 always collects all edges around the vertex or edge being
considered and sorts them to explore them recursively in increasing order of their random numbers. In this
section we introduce a data structure that is responsible for generating the random numbers and providing
edges for further exploration in the desired order.
For every vertex v ∈ V , we have a copy neighbors[v] of the data structure. (In fact, a copy for a
given vertex is created when it is accessed for the very first time.) From the point of view of the exploration
16
algorithm, the data structure exposes only one operation: lowest(k), where k is a positive integer. The
operation neighbors[v].lowest(k) lists edges incident to v in order of the random numbers assigned to
them, omitting all appearances of parallel edges or self-loops except the first one, which has been assigned
the lowest number. For each positive k, the operation returns a pair hw, ri, where w is a vertex and r is a
number in (0, 1] ∪ {∞}. If r 6= ∞, then (v, w) is the edge with the kth lowest number in the above order,
and r is the number assigned to it. Otherwise, the list is shorter than k and r = ∞ indicates the query
concerned a non-existing edge. We present the implementation of the data structure in Section 4.4.
We rewrite Algorithms 1 and 2 to use the data structure, and present them as the oracle VOσ(v) in
Algorithm 3 and the oracle MOσ(e) in Algorithm 4, respectively.
Algorithm 3: An oracle VOσ(v) for a vertex cover based on the input from the data structures neighbors,
which assigns edges e random numbers σ(e) (online). Given a vertex v, the oracle returns TRUE if v belongs
to the corresponding vertex cover and it returns FALSE otherwise.
1
2
3
4
5
6
7
8
i:=1
hw, ri := neighbors[v].lowest(i)
while r 6= ∞ do
if MOσ((v, w)) = TRUE then
return TRUE
i := i + 1
hw, ri := neighbors[v].lowest(i)
return FALSE
Algorithm 4: An oracle MOσ((u, v)) for a maximal matching based on the input from the data structures
neighbors, which assigns edges e random numbers σ(e) (online). Given an edge (u, v), the oracle returns
TRUE if (u, v) belongs to the corresponding matching and it returns FALSE, otherwise.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
if MOσ((u, v)) has already been computed then
return the computed answer
k1 := 1 and k2 := 1
hw1, r1i := neighbors[u].lowest(k1)
hw2, r2i := neighbors[v].lowest(k2)
while w1 6= v or w2 6= u do
if r1 < r2 then
if MOσ((u, w1)) = TRUE then return FALSE
k1 := k1 + 1
hw1, r1i :=neighbors[u].lowest(k1)
else
if MOσ((v, w2)) = TRUE then return FALSE
k2 := k2 + 1
hw2, r2i := neighbors[v].lowest(k2)
15
return TRUE
17
Claim 4.1 Let σ be an injection from E to (0, 1]. Let π : E → [E] be the corresponding ranking defined
in such a way that for every edge e, σ(e) is the π(e)th lowest number in the set {σ(e′) : e′ ∈ E}.
For every vertex v, the answer returned by VOσ(v) (Algorithm 3) is the same as the answer returned
by VOπ(v) (Algorithm 1) provided the operation lowest works as specified and gives answers consistent
with σ.
This
that
the
easy
to
verify
It
is
is
claim holds when
true because when there are no parallel edges,
Proof:
parallel
to
edges.
neighbors[v].lowest(1), . . . , neighbors[v].lowest(k) simply returns the first k edges inci-
dent to v in order of increasing rank. Furthermore, when called on an edge (u, v), Algorithm 4 effectively
merges the two corresponding lists of adjacent edges (i.e., those incident to u and those incident to v) to
obtain a single list sorted according to rank, and makes recursive calls in the order dictated by the list.
are
there
a sequence of calls
no
It remains to verify that the same is true when there are parallel edges. For a fixed choice of σ and
the induced ranking π consider the two trees of recursive calls when calling VOπ(v) (Algorithm 1) and
VOσ(v) (Algorithm 3), where the former calls the oracle Mπ (Algorithm 2), and the latter calls the oracle
Mσ (Algorithm 4). When we refer to an edge in in the tree we actually mean an occurence of an edge in G
on a path of recursive calls.
These trees are both rooted at v, and with each edge there is an associated rank (number) and an asso-
ciated answer computed by the corresponding maximal matching oracle. Recall that each path of recursive
calls from the root to a leaf passes through edges with decreasing ranks (numbers). Furthermore, in both
trees, if an edge (u, v) in the tree is associated with the answer FALSE, then there must be an edge (u, w) (or
(v, w)) adjacent to it in the tree with lower rank (a "child" of this edge) that is associated with the answer
TRUE, and it is the highest ranking child that (u, v) has. If (u, v) is associated with the answer TRUE, then
all the children of (u, v) in the tree are associated with the answer FALSE. It will actually be convenient to
consider the full recursion trees without the "memoization" rule that we employ (which says that once an
answer is determined for an edge it is not computed again). This in particular implies that for each edge that
is the last edge on a path of recursive calls, the answer associated with it must be TRUE.
By the definition of VOσ(v) and the operation lowest, the tree corresponding to VOσ(v) contains
only edges that have minimal ranking among each set of parallel edges that connect a pair of vertices. We
claim that this tree is a "pruned" version of the tree that corresponds to VOπ(v), in the sense that all subtrees
containing non-minimally ranked parallel edges are removed, and the answers associated with the remaining
edges (and hence with the root v) are exactly the same.
Let T π(v) denote the tree of recursive calls for VOπ(v), and let h be the height of T π(v). Starting
from ℓ = h and going up the tree, we show that we can remove all non-minimally ranked parallel edges in
level ℓ of T π(v) without altering the answer for their parent edges. For ℓ = h, we claim that there are no
non-minimally ranked parallel edges in the last level of T π(v), so that no pruning needs to be performed.
To verify this, assume in contradiction that e is a non-minimally ranked parallel edge between vertices u
and w where e is at the end of a recursive path of length h in T π(v). Since e is not minimally ranked, there
should be a "sibling" of e in the tree which correspond to the minimally ranked edge e′ between u and w.
Since π(e′) < π(e), it must be the case that the answer associated with e′, that is, Mπ(e′), is FALSE. But e′
also belongs to level h, so that e′ is the last edge on a path of recursive calls, and hence cannot be answered
FALSE.
Assuming we have performed the pruning successfully for all levels ℓ < ℓ′ ≤ h, we show that we can
perform it for level ℓ. Consider a non-minimally ranked parallel edge e between vertices u and v in level ℓ
of T π(v). As argued above, there is a "sibling" of e in the tree which correspond to the minimally ranked
18
edge e′ between u and w. Since π(e′) < π(e), it must be the case that the answer associated with e′, that
is, Mπ(e′), is FALSE. This implies that e′ has a child e′′ in the tree resulting from pruning all non-minimal
parallel edges from levels ℓ′ > ℓ, such that Mπ(e′′) = TRUE. But since π(e′′) < π(e′) < π(e), and e′′ is
also adjacent to e, we get that Mπ(e) is FALSE as well. Hence, it is possible to prune e from the tree without
altering the answer obtained for its parent.
4.2 Implementing lowest: The High-Level Idea
The pseudo-code for the procedure lowest as well as the data structure that it uses, are given in full detail
in Subsection 4.4. Here we give a high-level description. For the sake of simplicity of the presentation, in
this description we assume that there are no parallel edges.
Roughly speaking, the procedure lowest for a vertex v is implemented in "batches". Namely, con-
sidering intervals of (0, 1] of the form (2−i, 2−i+1] (for i ∈ [d⋆], where d⋆ = ⌈log d⌉, as well as the
interval (0, 2−d⋆ ]), the procedure does the following. It first decides which edges incident to v should be
assigned a value in the current interval (2−i, 2−i+1]. In this stage each edge is identified with its label (in
{1, . . . , deg(v)}). The procedure then determines the identity of the other endpoint of each of these edges
by performing a neighbor query, and it assigns the edge a value σ((v, w)), selected uniformly at random
from the interval. This assignment is performed unless a certain constraint is discovered due to information
held in neighbors[w], as we explain subsequently. Once σ((v, w)) is determined, the other endpoint of
the edge, w, is "notified". That is, the data structure neighbors[w] is updated with this new information.
The procedure "opens" a new interval (2−i+1, 2−i+2] if the index k it is called with is such that the number
of neighbors w of v whose identity has been revealed and such that σ((v, w)) ≤ 2−i+1 is strictly less than k.
Thus, the procedure performs queries and assigns valued to edges "on demand", but it does so for "batches"
of edges. More precise details follow.
The data structure neighbors maintains two values for each vertex v: lb, and next lb (where the
latter is always twice the former). When a vertex is first encountered, lb is set to 0 and next lb is set
to 2−d⋆. Second, the data structure maintains a dictionary assigned number, which holds, for those
vertices w that are known to be neighbors of v, the value σ((v, w)) that was assigned to the edge between
them (initially, the dictionary is empty). The subset of indices in {1, . . . , deg(v)} that correspond to edges
for which the other endpoint has not been revealed (and do not yet have an associated value), are considered
unassigned. Third, the data structure maintains a list of pairs hw, ri, where w is a (known) neighbor of v
and r = σ((v, w)). This list is sorted in ascending order of r's, and it contains exactly those w for which
the corresponding r is at most lb.
If a call is made to neighbors[v].lowest(k) with k > deg(v) then it returns5 hv, ∞i. Otherwise, the
procedure does the following until the length of sorted is at least k. It first considers those edges (v, w)
incident to v that were already assigned a value r and this value belongs to the interval (lb, next lb] (that
is, assigned number[w] ∈ (lb, next lb]). The setting of the value r for each such edge (v, w) was
performed previously in the course of call to neighbors[w].lowest(·). Let the corresponding subset of
pairs hw, ri be denoted S.
The procedure next selects a subset T of {1, . . . , deg(v)} containing the labels of those (additional)
edges that it will (tentatively) assign a value in (lb,next lb]. Putting aside for now the issue of time-
efficiency (which we return to later), this can be done by flipping a coin with bias next lb−lb
indepen-
1−lb
5Recall that we assume that there are no parallel edges, or else hv, ∞i is returned if k exceeds the "effective" degree of v, that
is, counting parallel edges as a single edge.
19
dently for each edge label in the subset of unassigned edge labels. For each t ∈ T , the procedure now
performs a neighbor query to obtain the tth neighbor of v. Denoting this neighbor by w, let lb' denote
the lower bound lb held by w, that is, in the data structure neighbors[w]. If lb′ ≤ lb, so that the
lower bound constraint imposed by w is no larger than that imposed by v, then the following operations are
performed.
First, a random number r in the interval (lb,next lb] is selected uniformly at random, and
assigned number[w] is set to r.
In addition, assigned number[v] is set to r in the data struc-
ture neighbors[w] (so that w is "notified" of the revealed edge (v, w) as well as the assignment
r = σ((v, w))). Finally, the pair hw, ri is added to S.
If lb′ > lb, which means that lb′ ≥ next lb (given the way the intervals are defined), then the lower
bound constraint imposed by the end point w of the edge (v, w) does not allow the edge to be assigned a
value in the interval (lb,next lb], and so effectively its selection to T is retracted. Note that since the
decision whether an edge label is added to T is done independently for the different edges, the end effect (of
not assigning (v, w) a value in (lb,next lb]) is exactly the same as the one we would get if we had the
knowledge in advance (before selecting T ), that the corresponding edge label t should not be selected.
After going over all labels t in T , the resulting set S of pairs hw, ri is sorted in ascending order of r's,
and it is appended to the end of the list sorted. Thus, sorted now includes all pairs hw, ri such that w is
a neighbor of v, the value assigned to this edge is r, and r ≤ next lb. The variables lb and next lb are
then updated so that lb is set to next lb and next lb is set to 2 · next lb. Once the length of sorted
is at least k, the procedure returns sorted[k]. In Subsection 4.4 we formally establish that the distribution
of random numbers the data structures neighbors[v] provide access to via the operation lowest(k) is
the same as assigning independently at random a number from the range (0, 1] to each edge.
4.3 Generating Random Numbers
In this subsection we describe a random process that generates random numbers σ(e) for edges e ∈ E. The
procedure lowest applies this process in the course of its executions. In the remainder of this section, I
denotes the length of an arbitrary real interval I. We do not distinguish open and closed intervals here. For
instance, (0, 1) = [0, 1] = (0, 1] = [0, 1) = 1.
Let d be an upper bound on the maximum vertex degree. We set d⋆ = ⌈log d⌉. For every edge e, the
number σ(e) should be selected independently, uniformly at random from the range (0, 1]. We partition this
range into d⋆ + 1 intervals. We set
Ii =((2−i, 2−i+1]
(0, 2−d⋆ ]
for i ∈ [d⋆],
for i = d⋆ + 1.
Algorithm 5: A Process for Selecting a Random Number Assigned to an Edge
1
2
3
for i ← d∗ + 1 downto 2 do
Ii
with probability
terminate)
P1≤j≤i Ij : return a number selected from Ii uniformly at random (and
return a number selected from I1 uniformly at random
We describe our process as Algorithm 5. The process first selects one of the intervals Ii, and then selects
a number uniformly at random from this interval. The selection of the interval is conducted as follows. We
20
consider the intervals in reverse order, from Id⋆+1 to I1. For a considered interval, we decide that the
number belongs to this interval with probability equal to the length of the interval over the sum of lengths of
all the remaining intervals. The process selects each interval with probability proportional to its length, and
since the lengths of all intervals sum up to 1, the number that the process returns is uniformly distributed on
the entire interval (0, 1].
Corollary 4.2 Algorithm 5 selects a random number from the uniform distribution on (0, 1].
Note that by simulating a few iterations of the loop in the above process, one can decide that the number
assigned to a given edge is either a specific number less than or equal to 2−i, or that it is greater than 2−i
without specifying it further, for some i. Later, whenever more information about the number is needed,
we may continue with consecutive iterations of the loop. As we see later, we use the process in our data
structures neighbors[v] to lower the query and time complexity of the resulting vertex cover algorithm.
4.4 Data Structures
We now describe the data structures neighbors[v]. Each data structure neighbors[v] simulates the
random process described in Section 4.3 for all edges incident to v in the course of the executions of
neighbors[v].lowest. The data structure simultaneously makes a single iteration of the loop in Al-
gorithm 5 for all incident edges. It may be the case that for some edge (v, w), the random number has
already been specified. In this case, the result of the iteration for this (v, w) is discarded. It may also be
the case that this iteration of the loop has already been taken care of by neighbors[w], the data structure
for the other endpoint of the edge. The data structures communicate to make sure that a second execution
of a given iteration does not overrule the first. The data structures are designed to minimize the amount of
necessary communication. Note that if a data structure does not have to communicate with a data structure
at the other endpoint of a given neighbor, it does not even have to know the neighbor it is connected to with
a given edge, which can be used to save a single query. By using this approach, we eventually save a factor
of nearly d in the query complexity.
Each data structure neighbors[v] supports the following operations:
neighbors[v].lowest(k): As already mentioned, this is the only operation that is directly used by the
oracles (Algorithm 3 and Algorithm 4). It returns a pair hw, ri, where (v, w) is the edge with the kth
lowest random number assigned to edges incident to v, omitting a second and furher appearances for
parallel edges, and r is the random value assigned to (v, w). If r = ∞, then k is greater than the
length of such a defined list.
neighbors[v].lower bound(): The operation returns the current lower bound the data structure im-
poses on the edges that are incident to v and have not been assigned a specific random number yet.
The set of possible values returned by the procedure is {0} ∪ {2i : −d⋆ ≤ i ≤ 0}. Let ℓv be the num-
ber returned by the operation. It implies that the data structure simultaneously simulated the random
process described in Section 4.3 for incident edges until it made sure that the random numbers that
have not been fixed belong to (ℓv, 1].
Furthermore, let (v, w) be an edge in the graph. Let ℓv and ℓw be the numbers returned by the
operation for neighbors[v] and neighbors[w], respectively. If no specific random number has
been assigned to (v, w), then we know that the random number will eventually be selected uniformly
at random from (max{ℓv, ℓw}, 1].
21
neighbors[v].set value(w, r): It is used to notify the data structure neighbors[v] that the ran-
dom value assigned to (v, w) has been set to r. This operation is used when the data structure
neighbors[w] assigns a specific random number to (v, w). Before assigning r, the data struc-
ture neighbors[w] has to make sure that r > neighbors[v].lower bound(), i.e., it has not
been decided by the data structure neighbors[v] that the random number assigned to v is greater
than r.
To implement the above operations, each data structure neighbors[v] maintains the following infor-
mation:
lb: The variable specifies the lower bound on the incident edges that were not assigned a random number
yet. This is the value returned by the operation neighbors[v].lower bound(). This is also the
value at which the simulation of the process generating random number for edges incident to v has
stopped.
next lb: If specific random numbers assigned to more edges are necessary, the next considered range of
random numbers will be (lb, next lb], and next lb will become the new lower bound for the
edges that have not been assigned any random number. This variable is redundant, because its value
is implied by the value of lb, but using it simplifies the pseudocode.
assigned number: This is a dictionary that maps neighbors w of v to numbers in (0, 1]. Initially, the dic-
tionary is empty. If assigned number[w] = NONE, i.e., there is no mapping for w, then no specific
random number has been assigned to any of the edges (v, w) yet. Otherwise, assigned number[w]
is the lowest random number that has been assigned to any parallel edge (v, w).
sorted: This is a list consisting of pairs hw, ri, where w is a neighbor of v and r is the number assigned
to the edge (v, w). It is sorted in ascending order of r's, and it contains exactly those w for which
the edge (v, w) (with the lowest assigned random number) has an assigned random number less than
or equal to lb. For all neighbors w that do not appear on the list, the lowest number assigned to any
edge (v, w) is greater than lb.
We give pseudocode for all data structure operations as Algorithms 6, 7, 8, and 9. We postpone all issues
related to an efficient implementation of the data structure to Section 4.6. Three of them are straightforward,
and we only elaborate on the operation neighbors[v].lowest(k) (see Algorithm 9).
Algorithm 6: The procedure for initializing neighbors[v]
1
2
3
4
lb := 0
next lb := 2−d⋆
assigned number := {empty map}
sorted := {empty list}
Algorithm 7: The procedure neighbors[v].set value(w, r)
1
assigned number[w] := r
22
Algorithm 8: The procedure neighbors[v].lower bound()
1
return lb
As long as not sufficiently many lowest random numbers assigned to edges incident to v have been
determined, the operation lowest simulates the next iteration of the loop in the random process that we
use for generating random numbers. Let I be the interval (lb, next lb]. The operation wants to determine
all random numbers assigned to edges incident to v that lay in I. First, in Line 2, it determines the random
numbers in I that have already been assigned by the other endpoints of corresponding edges. In Line 3, the
operation simulates an iteration of the loop of the random process for all edges incident to v to determine
a subset of them that will have numbers in I (unless it has already been decided for a given edge that its
random number is not in I). In the loop in Line 4, the operation considers each of these edges. Let (v, w)
be one of them, where w is its other endpoint, queried by the operation. In Line 6, the operation generates
a prospective random number r ∈ I for the edge. First, the operation makes sure that this iteration of the
has not been simulated by the other endpoint (the condition in Step 7). If this is the case, the operation
considers two further cases. If r is lower than the lowest number assigned to any parallel edge (v, w) so
far, the procedure updates the appropriate data structures with this information (Steps 8 -- 11). If no random
number has ever been assigned to any edge (v, w), the procedure assigns it and updates the data structures
appropriately (Step 12 -- 15). When the operation finishes going over the list of all potentially selected edges
and eventually determines all incident edges with new lowest random numbers, it sorts them in order of
their random number and appends them in this order to the list sorted. Finally, when sufficiently many
edges with lowest numbers have been determined, the operation returns the identity of the edge with the kth
smallest number.
Lemma 4.3 The lists of incident edges that the data structures neighbors[v] provide access to via the
operation lowest(k) are distributed in the same way as when each edge is assigned independently at
random a number from the range (0, 1].
Proof: We know from Corollary 4.2 that the random process generates a random number from the distribu-
tion (0, 1]. Each data structure neighbors[v] simulates consecutive iterations of the loop in this process
for all edges incident to v. Consider a group of parallel edges (v, w). For each of these edges, the ran-
dom process is simulated by both neighbors[v] and neighbors[w]. We have to show that until the
lowest number assigned to the edges in this group is determined (which happens when it is added to the
list sorted), then for each edge the decision made in the first simulation matters. Why is this the case?
Recall that the random process considers intervals Id⋆+1, Id⋆, . . . , I1 as the sources of the random number
in this order. As long as both neighbors[v] and neighbors[w] reject a given interval their decisions
are the same, so the first decision is in effect. Now suppose without loss of generality that neighbors[w]
simulates a consecutive iteration of the loop in the random process and decides to use Ii as the source of the
random number for a given edge (v, w) in Step 9 of the operation lowest. If neighbors[v] has already
simulated this iteration (the condition verified in Step 7), the operation does not proceed. Otherwise, the ran-
dom number assigned to the edge is considered for a new minimum random number assigned to this group
of parallel edges. Note that since the operation keeps simulating iterations even after a random number is
assigned, it could be the case for a specific copy of (v, w) that a new, higher random number is considered,
but it is ignored, because it is higher than the first decision, which is the only one that has impact on the list
that the operation lowest provides access to.
23
Algorithm 9: The procedure neighbors[v].lowest(k)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
while length(sorted) < k and lb < 1 do
S := set of pairs hw, ri such that assigned number[w] = r and r ∈ (lb, next lb]
T := subset of {1, . . . , deg(v)} with each number included independently
with probability next lb−lb
1−lb
foreach t ∈ T do
w := tth neighbor of v
r := a number selected uniformly at random from (lb, next lb]
if neighbors[w].lower bound() ≤ lb then
if ∃hw, r′i ∈ S s.t. r < r′ then
assigned number[w] := r
neighbors[w].set value(v, r)
replace hw, r′i with hw, ri in S
if assigned number[w] = NONE then
assigned number[w] := r
neighbors[w].set value(v, r)
S := S ∪ {hw, ri}
Sort S in ascending order of their r, and append at the end of sorted
lb := next lb
next lb := 2 · next lb
if length(sorted) < k then return hv, ∞i
else return sorted[k]
The correctness of the data structure follows from the fact that it extends the list sorted by always
adding all edges with random numbers in a consecutive interval, and it always takes into consideration
decisions already made by data structures for the other endpoints for these intervals.
4.5 Query Complexity
We now show that the number of queries that the algorithm makes is not much higher than the number of
recursive calls in the graph exploration procedures. The following simple lemma easily follows from the
Chernoff bound and will help us analyze the behavior of the algorithm.
Lemma 4.4 Let X1, . . . , Xs be independent random Bernoulli variables such that each Xi equals 1 with
probability p. It holds:
• For any δ ∈ (0, 1/2),
with probability at least 1 − δ.
Xi ≤ 6 · ln(1/δ) · max{1, ps}.
Xi
24
• For any δ ∈ (0, 1/2), if ps > 8 ln(1/δ), then
with probability at least 1 − δ.
Xi ≥
ps
2
.
Xi
Proof: Let us first prove the first claim.
Otherwise, there exist independent Bernoulli random variables Yi, 1 ≤ i ≤ s such that for each i,
If 6 · ln(1/δ) · max{1, ps} ≥ s, the claim follows trivially.
Pr[Yi = 1] = 3 · ln(1/δ) · max{1/s, p} > p
since from the definition of δ: 3 · ln(1/δ) > 1. Therefore Pr[Xi = 1] < Pr[Yi = 1]. By this fact and by the
Chernoff bound,
Pr[X Xi > 2E[X Yi]] ≤ Pr[X Yi > 2E[X Yi]]
≤ exp(− ln(1/δ) · max{1, ps})
≤ exp(− ln(1/δ)) ≤ δ.
The second claim also directly follows from the Chernoff bound:
Pr[X Xi < ps/2] ≤ exp(−(1/2)2 · ps/2) ≤ δ.
Definition 4.5 Denote Ji = Sd⋆+1
(0, 1
d ]. We expect that the number of incident edges to v with random numbers in Ji to be deg(v) · Ji.
j=i Ij, where 1 ≤ i ≤ d⋆ + 1. For example: J1 = (0, 1] and Jd⋆+1 =
We now define a property of vertices that is useful in our analysis. Intuitively, we say that a vertex is "usual"
if the numbers of incident edges with random numbers in specific subranges of (0, 1] are close to their
expectations.
Definition 4.6 Let α > 0. We say that a vertex v is α-usual if the random numbers assigned to edges
incident to v have the following properties for all i ∈ {1, . . . , d⋆ + 1}:
• Upper bound: The number of incident edges with random numbers in Ji is
at most max{α, α · deg(v) · Ji}.
• Lower bound: If deg(v) · Ji ≥ α, then the number of edges with random numbers in Ji is
at least deg(v) · Ji/2.
We now basically want to show that the relevant vertices are α-usual, and later on we will use it to prove
a lower bound.
We define an additional quantity that is useful later in bounding the total running time of the algorithm.
25
Definition 4.7 For an execution of Step 9 of Algorithm 9 where the number of neighbors is k and p ∈ [0, 1]
is the probability of selecting each of them, we say that the toll for running it is kp.
We now prove a bound on the query complexity of the algorithm and other quantities, which are useful
later to bound the running time. We start by introducing the main Lemma (Lemma 4.8), followed by proving
Lemma 4.9 which will help us prove Lemma 4.8.
Lemma 4.8 Consider an algorithm A that queries the input graph only via the oracle described as Algo-
rithm 1. Let t ≥ 1 be the expected resulting number of calls in A to the oracles described as Algorithm 1
and Algorithm 2. Let d be an upper bound on the maximum degree of the input graph.
Suppose now that we run this algorithm replacing calls to Algorithm 1 with calls to Algorithm 3. The
following events hold all at the same time with probability 1 − 1/20:
1. The total number of calls to Algorithms 3 and 4 is O(t)
2. The operation lowest in data structures neighbors[v] runs at most O(t) times.
3. The query complexity of A is O(t · log2(dt)).
4. The total toll for running Step 9 of Algorithm 9 is O(t · log(dt)).
Before proving Lemma 4.8 we establish the following Lemma:
Lemma 4.9 Assume the conditions of Lemma 4.8. Let t′ = 100t, δ = 1/(40000t(d + 1)(d⋆ + 1)), and
α = 8 · ln(1/δ). The following three events happen with probability less than 1
100 for each:
1. The total number of calls to Algorithm 3 and Algorithm 4 is bounded by t′.
2. The first 2t′ vertices for which the operation lowest is called are α-usual.
3. For the first 2t′ vertices v for which the operation lowest is called, the size of the set T generated
in the jth execution of Step 9 of the operation is bounded by α · max{1, deg(v) · 2j−d⋆}.
Proof: For every group of parallel edges, the operation lowest lists only the edge with the lowest number.
For the purpose of this analysis we assume that the operation lists in fact all occurences of a given parallel
edge. The final complexity is only reduced because of the fact that some unnecessary calls are omitted.
1. Let us bound the probability that one of the above events does not occur. By Markov's inequality the
probability that the first event does not occur is bounded by 1
100 .
2. We shall now prove that the first 2t′ vertices for which the operation lowest is called are α-usual.
The total number of vertices that have an incident edge for which the process generating random
numbers is simulated in the above calls is bounded by 2t′ · (d + 1). The property of being α-usual is
a function of only random numbers assigned to incident edges.
incident edges to v with random numbers in Ji. From Lemma 4.4 we get that:
j=1 Xj where p = Pr[Xj = 1] = Ji, s = deg(v), i.e. X is the number of all
For Ji let X = Ps
Pr(cid:2)Xi
Xi > α · max{1, Jideg(v)}(cid:3) = Pr(cid:2)Xi
≤ Pr[Xi
Xi > 8 · ln(1/δ) · max{1, ps}(cid:3)
Xi > 6 · ln(1/δ) · max{1, ps}] < δ
26
Also, from Lemma 4.4 we get that:
Pr(cid:2)Xi
Xi <
deg(v) · Ji
2
(cid:3) = Pr(cid:2)Xi
Xi <
ps
2(cid:3) < δ
i.e. v is not α-usual because of Ji with probability less than 2δ. From union bound on all all i ∈
[d⋆ + 1] we get that vertex v is not α-usual with probability less than 2δ(d⋆ + 1).
Using the union bound again, this time over the vertices incident to edges for which the random
process is run, the probability that any of them is not α-usual is bounded by
2t′ · (d + 1) · 2δ(d⋆ + 1) = 400tδ(d + 1)(d⋆ + 1) =
1
100
.
3. We need to prove that for the first 2t′ vertices v for which the operation lowest is called,
the size of the set T generated in the jth execution of Step 9 of the operation is bounded by
α · max{1, deg(v) · 2j−d⋆}.
Let v be one of the first 2t′ vertices for which the operation neighbors[v].lowest is called. Ob-
serve that in the jth iteration of the loop while, (next lb−lb)/(1−lb) is at most 2j−d⋆. Therefore,
it follows from Lemma 4.4 that for each j ∈ {1, . . . , d⋆ + 1}, the size of the set T in Algorithm 9
selected in the jth execution of Step 9 is bounded by α · max{1, deg(v) · 2j−d⋆} with probability
1 − δ. By the union bound over all j and the first 2t′ vertices, the probability that the third event does
not occur is bounded by
2t′(d⋆ + 1)δ = 200t(d⋆ + 1) · 1/(40000t(d + 1)(d⋆ + 1) <
1
100
Summarizing, the probability that at least one of the three events does not occur is bounded by
3
100
<
1
20
Let us now prove Lemma 4.8 assuming that the events in Lemma 4.9 occur.
Proof of Lemma 4.8:
1. We need to prove that the total number of calls to Algorithms 3 and 4 is O(t). This follows directly
from Lemma 4.9, we proved it there for t′ = O(t).
2. We need to show that the operation lowest in data structures neighbors[v] runs at most O(t)
times.
The total number of vertices v for which the operation neighbors[v].lowest is called is bounded
by 2t′, because a call to one of the oracles (Algorithms 3 and 4) requires calling the operation
lowest for at most two vertices.
It follows from the implementation of the oracles that the op-
eration neighbors[v].lowest is executed at most 3t′ = O(t) times if the number of oracle calls
is bounded by t′ (which was proved in Lemma 4.9). This is true because in Algorithm 3 we call
neighbors[v].lowest once and in Algorithm 4 we call neighbors[v].lowest twice.
27
3. We will now show that the query complexity of A is O(t · log2(dt)).
For each vertex v, denote kv ∈ [0, deg(v)] the number of times we call neighbors[v].lowest(k)
on v. We assume that if the operation is not executed for a given vertex, then kv = 0. It holds that:
kv ≤ 3t′
Xv∈V
We now attempt
to bound the query complexity necessary to execute the operation
neighbors[v].lowest for a given v such that kv > 0. Note that the expected number of edges with
random numbers in a given Ji is deg(v)/2i−1. Recall that from Lemma 4.9 we know that the first
2t′ vertices for which the operation lowest is called are α-usual. From the lower bound of α-usual
(Definition 4.5) we get that If deg(v) · Ji ≥ α, then the number of edges with random numbers in
Ji is at least
deg(v) · Ji/2.
Therefore, if
then the number of edges with random numbers in Ji is at least
deg(v) · Ji = deg(v)/2i−1 ≥ max{2α, 2kv }
max{2α, 2kv}
2
= max{α, kv} ≥ kv
i.e. if i is such that deg(v)/2i−1 is at least max{2α, 2kv }, then at least kv edges incident to v have
random numbers in Ji. This also holds for i such that deg(v)/2i−1 ≥ 2αkv. Let iv be the largest
integer i such that 2i ≤ deg(v)
(remember i = d⋆+1, d⋆ · · · ). Since iv is the maximum i that satisfies
αkv
this, then
2iv +1 >
deg(v)
αkv
⇒ 2iv >
deg(v)
2αkv
⇒ 2−iv <
2αkv
deg(v)
The body of the loop while in Algorithm 9 is executed at most d⋆ + 2 − iv times for v (remember
we start from i = d⋆+1), independently of how many times the operation is executed for v, because
all relevant edges incident to v are discovered during these iterations. From Lemma 4.9 we know
that the size of the set T in Algorithm 9 selected in the jth execution of this loop is bounded by
α · max{1, deg(v) · 2j−d⋆}. Furthermore, the sum of sizes of all sets T generated for v is bounded by
d⋆+2−ivXj=1
α · max{1, deg(v) · 2j−d⋆} ≤ α(d⋆ + 1) + 2α · deg(v) · 22−iv
≤ α(d⋆ + 1) + 16α2kv.
This also bounds the number of neighbor queries for v. Since these are the only neighbor queries in
the algorithm, by summing over all v with kv ≥ 0, the total number of neighbor queries is bounded
by
16α2kv ≤ 200αt(d⋆ + 1) + 16α2 · 300t = O(αt(d⋆ + α)) = O(t · log2(dt)).
2t′ · α(d⋆ + 1) +Xv∈V
(Recall t′ = 200t and thatPv∈V kv ≤ 3t′). Note that degree queries appear only in Step 9 of the
operation neighbors[v].lowest with one query to discover the size of the set from which a subset
is selected. The number of degree queries is in this case bounded by the total number of executions of
Step 9, which is at most O(t · log d). Summarizing, the total query complexity is O(t · log2(dt)).
28
4. Finally, we need to prove that the total toll for running Step 9 of Algorithm 9 is O(t · log(dt)). Recall
that the toll is defined as kp where k is the number of neighbors and p is the probability to selecting
each of them in an execution of Step 9 of Algorithm 9. Using arguments as above, the toll for running
Step 9 in the operation neighbors[v].lowest for a given v is bounded by
d⋆+2−ivXj=1
deg(v) · 2j−d⋆ ≤ 2 · deg(v) · 22−iv ≤ 8 · deg(v) ·
2αkv
deg(v)
= 16αkv
By summing over all vertices v, we obtain a bound on the total toll:
16αkv ≤ 4800αt = O(t · log(dt)).
Xv∈V
4.6 Efficient Implementation
We have already introduced techniques that can be used to show an approximation algorithm whose query
complexity has near-linear dependence on the maximum degree d. Unfortunately, a straightforward imple-
mentation of the algorithm results in a running time with approximately quadratic dependence on d. The
goal of this section is to remove a factor of approximately d from the running time of the algorithm. Our
main problem is how to efficiently simulate Step 9 in the operation lowest. Note that Step 9 is sampling
from a binomial distribution.
First, in Lemma 4.11, we prove that there is an algorithm that can simulate a binomial distribution which
runs in efficient time. Finally, in Theorem 4.13, we will show how to use it in our algorithms and how to
bound the running time by O(t · log3(dt)).
We start by defining the binomial distribution.
Definition 4.10 We write B(k, p), where k is a positive integer and p ∈ [0, 1], to denote the binomial
i=1 Xi, where each Xi, 1 ≤ i ≤ k,
is an independent random variable that equals 1 with probability p, and 0 with probability 1 − p.
distribution with success probability p on {0, 1, . . . , k} distributed asPk
is(cid:0)k
q(cid:1)pq(1 − p)k−q. We now show how to efficiently sample from this distribution.
It is well known that the probability that a value drawn from the binomial distribution B(k, p) equals q
Lemma 4.11 Let a, b, k, and Q be positive integers, where a ≤ b and Q > 1, that can be represented in
the standard binary form, using a constant number of machine words. There is an algorithm that takes a,
b, k, and Q as parameters, runs in O(max{ka/b, 1} · log Q) time, and outputs an integer selected from a
distribution D on {0, 1, . . . , k} such that the total variation distance between D and B(k, a/b) is bounded
by 1/Q.
Proof: If a = b, then the algorithm can return the trivial answer in O(1) time, so we can safely assume for
the rest of the proof that a < b. Let p = a/b and let qi = (cid:0)k
i(cid:1)pi(1 − p)k−i be the probability of drawing
i from B(k, p). Let s = min{6 · ln(2Q) · max{1, ka/b}, k}. For each i ≤ s, we compute a real number
q′
i ∈ [0, 1] such that qi − q′
i are
i = 1 (details about how to compute those q′
i=0 q′
i ≤ 1/2(k + 1)Q andPs
29
given in Lemma 4.12). Then we select the output of the algorithm from the distribution given by q′
write D to denote this distribution.
i's. We
Let us bound the total variation distance between this distribution and B(k, p). It suffices to show that
for every subsets S of {0, . . . , k}, the probability of selecting an integer from S in B(k, p) is not greater by
more than 1/Q, compared to the probability of selecting an integer in S from D. Consider an arbitrary such
set S. Let S1 be the subset of S consisting of numbers at most s. Let S2 be the subset of S consisting of
integers greater than s. We have
Recall that Xi = 1 with probability p. If s = k then
Xi∈S
q′
i ≥Xi∈S1
q′
i ≥ Xi∈S1(cid:18)qi −
1
2(k + 1)Q(cid:19) ≥Xi∈S1
Xi > k(cid:3) = 0
Pr(cid:2) kXi=1
Xi > s(cid:3) =(cid:2) kXi=1
qi −
1
2Q
.
(29)
If s = 6 · ln(2Q) · max{1, ka/b} then we define δ = 1
2Q , and from Lemma 4.4 we have that
Pr(cid:2) kXi=1
Hence,
In other words: the probability that a number greater than s is being selected from B(k, p) (i.e. s Xi's are
1) is bounded by 1
2Q . Therefore,
> 6 · ln(
1
δ
1
2Q
) · max{1, pk}(cid:3) < δ
Pr(cid:2) kXi=1
> s(cid:3) <
Xi∈S2
qi <
+Xi∈S2
qi −
Xi∈S
[qi − q′
1
2Q
1
2Q
1
2Q
i] ≤
1
Q
,
≥ Xi∈S
From 29 and 30 we get:
Xi∈S
q′
i ≥Xi∈S1
qi −
Therefore,
(30)
qi! −
1
Q
,
which proves our Lemma. Next, in Lemma 4.12 we will also show that the running time of the algorithm is
O(s) = O(max{k a
b , 1} log(Q)).
We now describe how to compute values q′
i that are approximation to qi.
Lemma 4.12 Recall: a < b, p = a/b and qi =(cid:0)k
Let s = min{6 · ln(2Q) · max{1, ka/b}, k}. For each i ≤ s, we can compute a real number q′
that qi − q′
i = 1. The total running time is O(max{ka/b, 1} · log Q).
i=0 q′
i(cid:1)pi(1 − p)k−i (probability of drawing i from B(k, p)).
i ∈ [0, 1] such
i ≤ 1/2(k + 1)Q andPs
30
Proof: Observe that for 1 ≤ i ≤ k:
qi = qi−1 ·
k + 1 − i
i
·
p
1 − p
Let ti = qi
q0
for 0 ≤ i ≤ s. It holds that for 1 ≤ i ≤ s:
ti = ti−1 ·
k + 1 − i
i
·
p
1 − p
= ti−1 ·
k + 1 − i
i
·
a
b − a
Note that for 0 ≤ i ≤ s:
(31)
(32)
(33)
(34)
=
tiPj≤s tj
qi
q0
1
q0Pj≤s qj
≥ qi
t′
i ≥ ti −
max0≤j≤s tj
4(k + 1)2Q
Suppose now that instead of ti, we use t′
definition of t′
i we get:
i ≥ 0, 0 ≤ i ≤ s, such that ti − t′
i ≤ max0≤j≤s tj
4(k+1)2Q . Then from the
Also:
t′
j ≤ s ·
Xj≤s
We have
(max0≤j≤s tj)
4(k + 1)2Q +Pj≤s tj
≤
max0≤j≤s tj
4(k + 1)Q +Pj≤s tj
≤(cid:0)1 +
1
4(k + 1)Q(cid:1) ·Xj≤s
tj
(From 32 and that max0≤j≤s tj ≤Pj≤s tj) ≥
( Since 1 +
1
4(k+1)2Q ≥ 1) ≥
t′
iPj≤s t′
j
≥
(1 +
ti − max0≤j≤s tj
4(k+1)2Q
1
1
4(k+1)Q )Pj≤s tj
qi ·Pj≤s tj
4(k+1)Q ) ·Pj≤s tj
4(k + 1)Q(cid:1) −
qi
4(k+1)Q )
1
−
1
1
1
−
(1 +
(1 +
≥ qi(cid:0)1 −
≥ qi −
4(k + 1)2Q
1
4(k + 1)2Q
1
4(k + 1)Q
4(k + 1)2Q
−
(1 +
Pj≤s tj
4(k+1)2Q
1
4(k+1)Q ) ·Pj≤s tj
≥ qi −
1
2(k + 1)Q
.
(35)
So eventually we get that
Also, note thatPs
i=0
qi −
t′
i
j=0 t′
j
Ps
= 1.
≤
1
2(k + 1)Q
t′
iPj≤s t′
j
Therefore, in our distribution D, we will define q′
It remains to show how we obtain t′
i with the desired properties. For this purpose, we use floating-point
arithmetic. Each positive number that we obtain during the computation is stored as a pair hS, Ei repre-
senting S · 2E. We require that 2α ≤ S < 2α+1 and E ≤ β, for some α and β to be set later. If we can
Pj≤s t′
.
j
i = t′
i
31
perform all standard operations on these integers in O(1) time, then we can perform the operations on the
represented positive real numbers in O(1) time as well. We call S a significand and E an exponent.
In particular, to multiply two numbers hS1, E1i and hS2, E2i it suffices to multiply S1 and S2, truncate
the least significant bits of the product, and set the new exponent accordingly. If these two numbers are
multiplicative (1 ± δ1)- and (1 ± δ2)-approximations to some quantities X1 and X2, respectively, then the
product of S1 and S2 in our arithmetic is a multiplicative (1 ± (δ1 + δ2 + δ1δ2 + 2−α))-approximation to
X1X2. If δ1 < 1, then the product is a (1 ± (δ1 + 2δ2 + 2−α))-approximation.
For each i of interest, one can easily compute a multiplicative (1±C ·2−α)-approximation for k+1−i
· a
b−a
in our arithmetic, where C > 1 is a constant. We make the assumption that 3Ck2α ≤ 1, which we satisfy
later by setting a sufficiently large α. Hence we use Equation 31 to obtain a sequence of multiplicative
(1 ± 3Ck2−α)-approximations t′
i, which
i, we no longer require that Si ≥ 2α and we modify their
is represented as a pair hSi, Eii. For all other t′
representation hSi, Eii so that Ei is the same as in the representation of the maximum t′
i. In the process we
i or even all non-zero bits. Assuming again that 3Ck2−α < 1,
may lose least significant bits of the some t′
the maximum additive error ti − t′
i for ti, where 0 ≤ i ≤ s. At the end, we find the maximum t′
i
i we get for each i for the modified representation is bounded by
t′
j ≤ 3Ck2−α · ti + 2 · 2−α · max
tj ≤ (3Ck + 2) · 2−α · max
tj,
3Ck2−α · ti + 2−α · max
j
j
j
where the first error term comes from the multiplicative error we obtain approximating each ti and the
second error term comes from making all exponents in the representation match the exponent of the largest
t′
i. Finally, we set α = ⌈log((3Ck+2)·4(k+1)2Q)⌉. This meets the previous assumption that 3Ck2−α < 1
and the guarantee on the error we may make on each t′
i is as desired. Note that since k and Q can be
represented using a constant number of words, so can integers of size at most 2α+1. To bound β, observe
that every k+1−i
a
b−a lies in the range [1/kb, kb], which implies that all ti lie in [1/kbk, kbk], and the
maximum absolute value of an exponent we need is of order O(k log(kb)), which can be stored using a
constant number of machine words.
·
i
sentation of t′
the algorithm is O(s).
To generate a random number from D, we consider only the significands Si in the final modified repre-
j. The total running time of
i's, and select each i with probability Si/Pj<s Sj = t′
We are ready to prove that the entire algorithm can be implemented efficiently. We use the algorithm of
i/Pj<s t′
Lemma 4.11 for efficiently simulating Step 9 in the operation lowest.
Theorem 4.13 Consider an algorithm A that queries the input graph only via the oracle described as
Algorithm 1. Let t ≥ 1 be a bound on the expected resulting number of calls in A to the oracles described
as Algorithm 1 and Algorithm 2, and such that t fits into a constant number of machine words using the
standard binary representation. Let d be an upper bound on the maximum degree of the input graph.
Suppose that calls to Algorithm 1 are replaced with calls to Algorithm 3. The oracles described as
Algorithm 3 and Algorithm 4 can be implemented in such a way that with probability 4/5 all of the following
events hold:
• The number of queries to the graph is O(t · log2(dt)).
• The total time necessary to compute the answers for the queries to the oracles is O(t · log3(dt)).
• The distribution of the answers that the oracle gives is D such that for some other distribution D′ over
5 · D′ is the distribution of answers of the oracle described
5 · D + 1
answers, the convex combination 4
as Algorithm 1.
32
Proof: Let a⋆ = d · O(t), where O(t) is the bound from Lemma 4.8 on the number of vertices for which
the operation lowest is called. If the event specified in Lemma 4.8 occurs, then a⋆ is an upper bound
on the number of edges for which the process for generating random numbers is simulated. Let b⋆ =
O(t) · (d⋆ + 1) = O(t log d), where O(t) is the same bound as above. Then b⋆ bounds the number of times
Step 9 in Algorithm 9 is run, provided the event specified in Lemma 4.8 occurs. Let Q = 20b⋆.
Let c⋆ = max{d⋆, ⌈log(20a2
⋆)⌉}. Since it is impossible to generate and store real numbers, we assign
to edges uniform random numbers from the set {i/2c⋆ : 1 ≤ i ≤ 2c⋆}, instead of the set (0, 1]. This can be
seen as selecting a random number from (0, 1] and then rounding it up to the next multiplicity of 1/2c⋆. In
particular, for every i ∈ {1, . . . , 2c⋆}, all numbers in ((i − 1)/2c⋆ , i/2c⋆ ] become i/2c⋆. Observe also that
each range Ij is a union of some number of sets ((i − 1)/2c⋆ , i/2c⋆ ], because c⋆ ≥ d⋆. This means that there
is no need to modify the process for generating random numbers, except for selecting a random i/2c⋆ in a
specific Ij, instead of an arbitrary real number from Ij. Observe also that as long we do not select the same
number i/2c⋆ twice, the entire exploration procedure behaves in the same way as in the idealized algorithm,
since the ordering of numbers remains the same.
Note that due to the assumption in the lemma statement, t can be represented in the standard binary form,
using a constant number of machine words. This is also the case for d, because of the standard assumption
that we can address all neighbors of all vertices in neighbor queries. This implies that Q = O(t log d)
also has this property. Finally, the probabilities next lb−lb
can easily be expressed using fractions a/b,
where a and b are of order O(d), and therefore, fit into a constant number of machine words as well. This
implies that we can use the algorithm of Lemma 4.11. Instead of directly simulating Step 9, we proceed as
follows. First, we run the algorithm of Lemma 4.11 with the error parameter Q to select a number t of edges
in T . Then we select a random subset of edges of size t. This can be done in O(t log d) time.
1−lb
We show that the algorithms and data structures can be implemented in such a way that the main claim
of the theorem holds, provided the following events occur:
• the events described in the statement of Lemma 4.8,
• the rounded numbers assigned to the first a⋆ edges for which the process for generating random
numbers is simulated are different,
• the first b⋆ simulations of the algorithm described by Lemma 4.11 do not result in selecting a random
number from the part on which the output distribution of the algorithm and the binomial distribution
differ.
The first of the events does not happen with probability at most 1/10. This follows from Lemma 4.8.
Consider the second event. The probability that two random numbers i/2c⋆ are identical is bounded by
1/2c⋆ ≤ 1/(20a2
⋆). Consider the first a⋆ edges for which the process generating random numbers is run.
The expected number of pairs of the edges that have the same random number is bounded by a2
⋆) =
1/20. By Markov's inequality, the probability that two of the edges have the same random number assigned
is bounded by 1/20. Finally, the probability that the last event does not occur is bounded by 1/20 as well
via the union bound. Summarizing, the events occur with probability at least 4/5.
⋆ ·1/(20a2
We now bound the running time, provided the above events occur. We assume that we use a standard
data structure (say, balanced binary search trees) to maintain collections of items. The time necessary for
each operation in these data structures is of order at most the logarithm of the maximum collection size.
For instance, we keep a collection of data structures neighbors[v] for v that appear in our algorithm. We
create neighbors[v] for a given v only when it is accessed for the first time. Observe that the number of
33
v for which we have to create neighbors[v] is bounded by the query complexity O(t log2(dt)), because
of how we access vertices. Therefore, accessing each neighbors[v] requires at most O(τ ) time, where
we write τ to denote the logarithm of the bound on the query complexity. That is, τ = O(log t + log log d).
The time necessary to run Algorithm 3 is bounded by O(τ ), which we need to locate the data structure
neighbors[v] for a given v, plus O(1) time per each call to Algorithm 4 (we do not include the cost of
running Algorithm 4 or the operation lowest here; they are analyzed later). The amount of computation
in Algorithm 3 without the resulting calls to other procedures is bounded by O(t · τ ).
Consider now Algorithm 4. In every run, we first spend O(log t) time to check if we have already com-
puted the answer for a given edge. Then locating the data structures neighbors[u] and neighbors[v]
for the endpoints u and v costs at most O(τ ). The running time of the reminder of the algorithm requires
time proportional to the number of recursive calls. Therefore, the total amount of time spent executing
Algorithm 4 (without calls to other procedures) is bounded by O(t · τ ).
We now bound the running time necessary to execute all operations of data structures neighbors. The
initialization of neighbors[v] (Algorithm 6) for a given v can be done O(1) time plus O(τ ) time necessary
for inserting the data structure into the collection of all neighbors[v]. Overall, since at most O(t log2(dt))
data structures are created, the total time necessary to initialize the data structures neighbors[v] is O(t ·
log2(dt) · τ ). Setting a value for some edge in Algorithm 7 takes at most O(log d) time to insert the value
into the mapping. This operation is run at most once for every neighbor query, so the total amount of
computation in this procedure is O(t · log2(dt) · log d). So far, the total computation time is bounded by
O(t log3(dt)).
Clearly, running the operation described by Algorithm 8 takes O(1) time, so overall the total amount
of computation in all executions of Algorithm 8 is not greater than some constant times the total amount
of computation in the operation lowest (Algorithm 9). Hence it suffices to bound the total amount of
computation in Algorithm 9, which we do next.
Recall that Algorithm 9 is run at most O(t) times. Therefore all operations in the loop while are run
at most O(t log d) times. The total size of sets S in Step 2 is bounded by the query complexity, and dis-
covering each element of S costs at most O(log d) time, if the data structure assigned number is properly
implemented, using augmented balanced binary search trees. Therefore the total cost of running Step 2 is
at most O(t · log d + t · log2(dt) · log d) = O(t · log2(dt) · log d). In Step 3, we use the algorithm of
Lemma 4.11. The total toll for running the algorithm is O(t · log(dt)). Therefore, the total time necessary
to simulate all executions of Step 2 is bounded by O((t · log d + t · log(dt)) · log Q) = O(t · log2(dt)). The
total number of executions of the body of the loop foreach in Step 4 is bounded by the query complexity
O(t · log2(dt)) times 2. The time required to execute the body of the loop is dominated by the following
two kinds of operations. One kind is querying and modifying the data structure assigned number[w]
and the data structure for S. With a proper implementation (say, augmented balanced binary search trees)
these operations take at most O(log d) time each. The other kind of operation is locating neighbors[w]
for the discovered neighbor w, which takes most O(τ ) time. The total computation time for all executions
of the loop foreach is therefore bounded by O(t · log3(dt)).
Finally sorting S never takes more than O(S log d) time, because S ≤ d, and each element of S can
be added at the end of the list sorted in amortized O(1) time if the list is implemented using extendable
arrays. This amounts to O(t · log2(dt) · log d) in all executions of Step 11. At the end of the operation, the
requested kth adjacent edge can be returned in O(1) time.
Summarizing, the computation of the answers of the oracles takes at most O(t · log3(dt)) time, if all
the desired events occur, which happens with probability at least 4/5. Note that when these events occur,
34
then also despite rounding random numbers assigned to edges, the implementation does not diverge from
the behavior of the idealized oracle.
5 The Near-Optimal Algorithms
Theorem 3.1 gives a bound on the expected number of recursive calls to oracles, sufficient to compute an
answer when the vertex cover oracle is called for a random vertex. The expected number of calls is O(ρ · ¯d),
where ρ is the ratio between the maximum degree d and the minimum degree dmin, and ¯d is the average
degree. (Recall that we assume without loss of generality that dmin ≥ 1. For isolated vertices, the oracle
answers that they are not in the vertex cover in O(1) time, and therefore, it suffices to focus on the subgraph
consisting of non-isolated vertices.)
A straightforward application of Theorem 3.1 gives a bound of O(d2) for graphs with maximum degree
bounded by d. We show a bound of O(d/ǫ) for a modified graph, which is preferable if 1/ǫ < d, and we also
show how to use the modified graph to obtain an estimate for the minimum vertex cover size in the original
input graph. We combine the obtained bound with Theorem 4.13 to get a fast and query-efficient algorithm.
Next we show how to obtain an efficient algorithm for the case when only the average degree of the
input graph is bounded. Finally, we show how to adapt the algorithm to the dense graph case, when only
vertex-pair queries are allowed.
5.1 Bounded Maximum Degree
As we have mentioned above, we can assume that 1/ǫ < d. We transform our graph into one with large
minimum degree, so that the ratio of maximum to minimum degree is small. For a given graph G = (V, E)
V ′ and E′ are defined as follows. The set V ′ contains a "shadow" vertex v′ for each vertex v ∈ V , and E′
contains ⌊ǫd⌋ parallel edges between v and v′, and 8d parallel self-loops for v′.
with maximum degree d, consider a graph eG = (eV ,eE), such that eV = V ∪ V ′ and eE = E ∪ E′ where
For a random ranking π over eE, for the output vertex cover Cπ(eG) on the new graph eG, we are interested
in bounding the size of Cπ(eG) ∩ V as compared to VCopt(G) (the size of a minimum vertex cover of G).
Since Cπ(eG) ∩ V is a vertex cover of G, we have that Cπ(eG) ∩ V ≥ VCopt(G), and so we focus on an
upper bound for Cπ(eG) ∩ V .
the construction of Cπ(eG) ∩ V , we have
Consider an arbitrary ranking π of E. Observe that for each v ∈ V , the matching Mπ(eG) either includes
rank of self-loops incident to v′ is lower than the lowest rank of edges (v, v′), then Mπ(eG) contains one
most ǫd/8d = ǫ/8. Therefore, the expected number of edges in Mπ(eG) ∩ F is upper bounded by ǫn/8.
probability 1 − 1/20, Mπ(eG) ∩ F ≤ ǫn/4.
Cπ(eG) ∩ V ≤ 2Mπ(eG) ∩ E + Mπ(eG) ∩ F ≤ 2VCopt(G) + Mπ(eG) ∩ F .
Without loss of generality, we can assume that ǫn > 72, since otherwise we can read the entire input with
only O(1/ǫ2) queries and compute a maximal matching in it. It follows from the Chernoff bound that with
Let F be the set of all parallel edges connecting each v with the corresponding v′. By the properties of
a parallel edge between v and v′ or it includes a self-loop incident to v′. For every v′ ∈ V ′, if the lowest
of the self-loops, and does not contain any parallel edge (v, v′).
If the ranking π is selected uniformly
at random, the above inequality on ranks does not hold for each vertex independently with probability at
35
in O(1) time, using O(1) queries to G. Therefore, we can simulate an execution of the vertex-cover and the
Observe that given query access to G, we can provide query access to eG (in particular, the edges in E′
that are incident to each v ∈ V can be indexed starting from deg(v) + 1). Every query to eG can be answered
maximal-matching oracles on eG.
Note that the expected number of recursive calls to the maximal matching oracle is bounded for a random
vertex v ∈ V by O(d/ǫ), because the maximum degree and the minimum degree are within a factor of
O(1/ǫ). Also note that since V = V /2, this expectation for a random vertex v ∈ V is at most twice as
much, i.e., it is still O(d/ǫ).
For any ranking π of edges in eE, if we sample O(1/ǫ2) vertices from V with replacement, then the
fraction of those in Cπ(eG) ∩ V is within an additive ǫ/8 of Cπ(eG) ∩ V /V with probability at least
1 − 1/20. Let µ be this fraction of vertices. Therefore, we have that
VCopt(G) − ǫn/4 ≤ µ · n ≤ 2VCopt(G) + ǫn/2
with probability at least 1 − 1/10. Thus (µ + ǫ/4) · n is the desired (2, ǫn)-estimate. The expected number
of calls to the vertex cover and maximal matching oracles is bounded by O(d/ǫ3). Note that without loss
of generality, ǫ ≥ 1/4n, because any additive approximation to within an additive factor smaller than 1/4
yields in fact the exact value. Therefore the expected number of calls to the oracles is bounded by O(n4),
which can be represented with a constant number of machine words in the standard binary representation,
using the usual assumption that we can address all vertices of the input graph. By applying now Theo-
rem 4.13, we obtain an implementation of the algorithm. It runs in O(d/ǫ3 · log3(d/ǫ)) time and makes
O(d/ǫ3 · log2(d/ǫ)) queries. Moreover, the probability that the implementation diverges from the ideal
algorithm is bounded by 1/5. Therefore, the implementation outputs a (2, ǫn)-estimate with probability
1 − 1/10 − 1/5 ≥ 2/3.
Corollary 5.1 There is an algorithm that makes O( d
log3 d
ǫ ) time, and with probability 2/3, outputs a (2, ǫn)-estimate to the minimum vertex cover size.
ǫ3 · log3 d
ǫ ) neighbor and degree queries, runs in O( d
ǫ3 ·
5.2 Bounded Average Degree
In this section, we assume an upper bound ¯d on the average graph degree and show an efficient algorithm
in this case.6 To do this, we will transform the graph into a new graph for which the ratio of the maximum
degree to the minimum degree is small.
Our first transformation is to automatically add high degree vertices to the cover, and continue by finding
a cover for the graph that is induced by the remaining vertices. Given a graph G = (V, E) with average
degree ¯d, let L denote the subset of vertices in G whose degree is greater than 8 ¯d/ǫ. Hence, L ≤ ǫn/8.
Let E(L) denote the subset of edges in G that are incident to vertices in L, and let G = (V , E) be defined
by V = V \ L and E = E \ E(L), so that the maximum degree in G is at most 8 ¯d/ǫ. For any maximal
matching M in G we have that
VCopt(G) ≤ 2M + L ≤ 2VCopt(G) +
ǫ
8
n .
Thus, the first modification we make to the oracles is that if the vertex-cover oracle is called on a vertex v
such that the degree of v is greater than (4/ǫ) ¯d, then it immediately returns TRUE.
6As shown in [PR07], we don't actually need to know ¯d for this purpose, but it suffices to get a bound that is not much higher
than (4/ǫ) ¯d, and such that the number of vertices with a larger degree is O(ǫn), where such a bound can be obtained efficiently.
36
The remaining problem is that when we remove the high degree vertices, there are still edges incident
to vertices with degree at most (4/ǫ) ¯d whose other endpoint is a high degree vertex, and this is not known
until the appropriate neighbor query is performed. We deal with this by adding shadow vertices to replace
the removed high degree vertices. At the same time, we increase the minimum degree as in the previous
subsection. We now create a graph eG = (V ∪eV , E ∪ eE) as follows. For every v ∈ V , we add toeV a vertex
i , where 1 ≤ i ≤ ⌈degG(v)/ ¯d⌉ and degG(v) is the degree of v in G. Each of these new
v′ and vertices v′′
vertices has 32 ¯d/ǫ parallel self-loops. Moreover, we add ¯d parallel edges between v and v′. Finally, partition
the edges incident to v in G into ⌈degG(v)/ ¯d⌉ groups, each of size at most ¯d. The first group corresponds
to the first ¯d edges on the neighborhood list of v, the second group corresponds to the next ¯d edges, and so
on. Let Ev,i ⊂ E be the set of edges in the i-th group. For every i of interest, we add Ev,i ∩ E(L) parallel
edges between v and v′′
i . We add these edges so that we
i and Ev,i \ E(L) parallel self-loops incident to v′′
are later able to simulate every query to eG using a constant number of queries to G.
Let us bound the total number of vertices ineV . The number of vertices v′ is V . The number of vertices
v′′
i is bounded by
Xv∈V(cid:24) degG(v)
¯d
(cid:25) ≤Xv∈V(cid:18) degG(v)
¯d
+ 1(cid:19) ≤
¯d · V
¯d
+ V = 2V ,
i . We have
We now repeat an argument from the previous section that despite the additional edges and vertices,
because eV has been created by removing vertices with highest degrees in G, and the average degree of
vertices in eV in G cannot be greater than ¯d, the initial average degree. This shows that eV ≤ 3V .
Cπ(eG) ∩eV is likely to be a good approximation to VCopt(G) for a random ranking π. First, Cπ(eG) ∩eV
is still a vertex cover for G, so VCopt(G) ≤ Cπ(eG) ∩eV . Let F be the set of edges connecting all v with
the corresponding v′ and v′′
Cπ(eG) ∩ V ≤ 2Mπ(eG) ∩ E + Mπ(eG) ∩ F ≤ 2VCopt(G) + Mπ(eG) ∩ F .
of the parallel edges connecting this vertex to the corresponding vertex in V , then one of the self-loops is
selected for the maximal matching as opposed to the parallel edges. The inequality on ranks does not hold
Observe that if for some of the vertices in eV , the lowest rank of self-loops is lower than the lowest rank
with probability at most ¯d/(32 ¯d/ǫ) = ǫ/32 independently for each vertex in eV . It therefore follows from
the Chernoff bound that the number of edges in Mπ(eG) ∩ F is not bounded by ǫeV /16 with probability at
most exp(−ǫeV /32), which is less than 1/20 if eV > 100/ǫ, and we can assume that this is the case. (To
circumvent the case of eV ≤ 100/ǫ, we can modify the algorithm as follows. If a sampled vertex belongs
to a connected component in V of size at most 100/ǫ, then we can read its connected component in eG and
deterministically find a maximal matching that uses only edges in E and self-loops in eE. This all takes at
most O( ¯d/ǫ2) time, which as we see later, we are allowed to spend per each sampled vertex.) Therefore, we
have
with probability at least 1 − 1/20.
Cπ(eG) ∩ V ≤ 2VCopt(G) + ǫV /4,
Observe that given query access to G, we can efficiently provide query access to eG. Degrees of vertices
in V are the same as in G. For associated vertices ineV it is easy to compute their degree in O(1) time, using
the degree of the corresponding vertex in V . To answer neighbor queries for vertices v in V , except for the
fixed connections to v′, it suffices to notice that if the corresponding edge in G is connected to a vertex in L,
this connection is replaced by a connection to an appropriate vertex v′′
i . Otherwise, the edge remains in E.
37
For vertices v′′
i some number of connections can either be a connection to the corresponding v or a self-loop.
This can be checked in O(1) time with a single query to the neighborhood list of v. All the other edges are
fixed. Therefore, we can simulate an execution of the vertex-cover and the maximal-matching oracles on
which we assume we can do.
random vertex v ∈ V is at most twice as much, i.e., it is still O( ¯d/ǫ2).
The expected number of recursive calls to the maximal matching oracle is bounded by O( ¯d/ǫ2) for a
eG. Answering every query to eG requires O(1) time and O(1) queries to G. Sampling vertices uniformly
at random from in V ∪eV is more involved, but in our algorithm, we only need to sample vertices from V ,
random vertex v ∈ V ∪ eV , because the maximum degree and the minimum degree are within a factor of
O(1/ǫ) and the maximum degree is bounded by O( ¯d/ǫ). Note that since 3V ≥ eV , this expectation for a
For any ranking π of edges in eG, if we sample O(1/ǫ2) vertices from V with replacement, then the
fraction of those for which the oracle answers TRUE is within an additive error ǫ/8 of the total fraction of
vertices for which the oracle answers TRUE. with probability 1 − 1/20. Let µ be the fraction of sampled
vertices. We have
VCopt(G) − ǫn/8 ≤ µ · n ≤ 2VCopt(G) + ǫn/4 + ǫn/8 + ǫn/8
with probability 1 − 1/10. Then (µ + ǫ/8)n is the desired (2, ǫn)-estimate. The expected number of
calls to the vertex cover and maximal matching oracles is bounded by O(d/ǫ4). As before, without loss of
generality, this quantity can be bounded by O(n5), which fits into a constant number of machine words. By
applying now Theorem 4.13, we obtain an implementation of the algorithm. It runs in O(d/ǫ3 · log3(d/ǫ))
time and makes O(d/ǫ3 · log2(d/ǫ)) queries. Moreover, the probability that the implementation diverges
from the ideal algorithm is bounded by 1/5. Therefore, the implementation outputs a (2, ǫn)-estimate with
probability at least 1 − 1/5 − 1/10 ≥ 2/3.
Corollary 5.2 There is an algorithm that makes O(
log3 ¯d
ǫ ) time, and with probability 2/3, outputs a (2, ǫn)-estimate to the minimum vertex cover size.
ǫ4 · log2 ¯d
¯d
ǫ ) neighbor and degree queries, runs in O(
¯d
ǫ4 ·
5.3 Adapting the Algorithm to the Vertex-Pair Query Model
The focus of this paper was on designing a sublinear-time algorithm whose access to the graph is via degree
queries and neighbor queries. In other words, we assumed that the graph was represented by adjacency
lists (of known lengths). When a graph is dense (i.e., when the number of edges is Θ(n2)), then a natural
alternative representation is by an adjacency matrix. This representation supports queries of the form: "Is
there an edge between vertex u and vertex v?", which we refer to as vertex-pair queries.
We next show how to adapt the algorithm described in the previous section to an algorithm that per-
forms vertex-pair queries. The query complexity and running time of the algorithm are (with high constant
probability) O(n/ǫ4), which is linear in the average degree for dense graphs. As in the previous section, the
algorithm outputs (with high constant probability) a (2, ǫ)-estimate of the size of the minimum vertex cover.
We recall that the linear lower bound in the average degree [PR07] also holds for the case that the average
degree is Θ(n) and when vertex-pair queries are allowed.
Given a graph G = (V, E), let eG = (eV ,eE) be a supergraph of G that is defined as follows. For
every vertex v ∈ V whose degree in G is less than7 n, there exists a vertex v′ ∈ eV , where there are
7 If there are no self-loops in the original graph, then the bound is n − 1.
38
n − degG(v) parallel edges between v and v′, and there are (8/ǫ)n self-loops incident to v′. As shown in
2VCopt(G) + (ǫ/4)n.
Subsection 5.1, with high probability over the choice of a ranking π over eG, we have that Cπ(eG) ∩ V ≤
Note that we can emulate neighbor queries to eG given access to vertex-pair queries in G as follows. Let
degree and the minimum degree in eG is at most 1/ǫ, and the maximum and average degrees are O(n/ǫ), we
the vertices in G be {1, . . . , n}. When the jth neighbor of vertex i is queried, then the answer to the query
is j when (i, j) ∈ E, and it is i′ (the new auxiliary vertex adjacent to i) when (i, j) /∈ E. The degree of
every vertex in V is n, so there is no need to perform degree queries. Since the ratio between the maximum
obtain an algorithm whose complexity is O(n/ǫ4), as claimed.
References
[BSS08]
Itai Benjamini, Oded Schramm, and Asaf Shapira. Every minor-closed property of sparse
graphs is testable. In STOC, pages 393 -- 402, 2008.
[CEF+05] Artut Czumaj, Funda Ergun, Lance Fortnow, Avner Magen, Ilan Newman, Ronitt Rubinfeld,
and Christian Sohler. Approximating the weight of the euclidean minimum spanning tree in
sublinear time. SIAM Journal on Computing, 35(1):91 -- 109, 2005.
[CHW08] Andrzej Czygrinow, Michal Ha´n´ckowiak, and Wojciech Wawrzyniak. Fast distributed approx-
imations in planar graphs. In DISC, pages 78 -- 92, 2008.
[CRT05]
Bernard Chazelle, Ronitt Rubinfeld, and Luca Trevisan. Approximating the minimum spanning
tree weight in sublinear time. SIAM J. Comput., 34(6):1370 -- 1379, 2005.
[CS09]
Artur Czumaj and Christian Sohler. Estimating the weight of metric minimum spanning trees
in sublinear time. SIAM Journal on Computing, 39(3):904 -- 922, 2009.
[CSS09]
Artur Czumaj, Asaf Shapira, and Christian Sohler. Testing hereditary properties of nonexpand-
ing bounded-degree graphs. SIAM J. Comput., 38(6):2499 -- 2510, 2009.
[Ele10]
[Fei06]
[GR08]
G´abor Elek. Parameter testing in bounded degree graphs of subexponential growth. Random
Struct. Algorithms, 37(2):248 -- 270, 2010.
Uriel Feige. On sums of independent random variables with unbounded variance, and estimat-
ing the average degree in a graph. SIAM Journal on Computing, 35(4):964 -- 984, 2006.
Oded Goldreich and Dana Ron. Approximating average parameters of graphs. Random Struc-
tures and Algorithms, 32(4):473 -- 493, 2008.
[GRS10] Mira Gonen, Dana Ron, and Yuval Shavitt. Counting stars and other small subgraphs in sub-
linear time. In SODA, pages 99 -- 116, 2010.
[HKNO09] Avinatan Hassidim, Jonathan A. Kelner, Huy N. Nguyen, and Krzysztof Onak. Local graph
partitions for approximation and testing. In FOCS, pages 22 -- 31, 2009.
[MR09]
Sharon Marko and Dana Ron. Approximating the distance to properties in bounded-degree and
general sparse graphs. ACM Transactions on Algorithms, 5(2), 2009.
39
[NO08]
[NS11]
[PR07]
[PS98]
[Yos11]
[YYI09]
Huy N. Nguyen and Krzysztof Onak. Constant-time approximation algorithms via local im-
provements. In FOCS, pages 327 -- 336, 2008.
Ilan Newman and Christian Sohler. Every property of hyperfinite graphs is testable. In STOC,
2011. To appear.
Michal Parnas and Dana Ron. Approximating the minimum vertex cover in sublinear time and
a connection to distributed algorithms. Theor. Comput. Sci., 381(1-3):183 -- 196, 2007.
Christos Papadimitriou and Kenneth Steiglitz. Combinatorial Optimization: Algorithms and
Complexity. Dover publications, 1998.
Yuichi Yoshida. Optimal constant-time approximation algorithms and (unconditional) inap-
proximability results for every bounded-degree CSP. In STOC, 2011.
Yuichi Yoshida, Masaki Yamamoto, and Hiro Ito. An improved constant-time approximation
algorithm for maximum matchings. In STOC, pages 225 -- 234, 2009.
40
|
1008.3546 | 1 | 1008 | 2010-08-20T17:58:00 | Homogeneous and Non Homogeneous Algorithms | [
"cs.DS",
"cs.CC"
] | Motivated by recent best case analyses for some sorting algorithms and based on the type of complexity we partition the algorithms into two classes: homogeneous and non homogeneous algorithms. Although both classes contain algorithms with worst and best cases, homogeneous algorithms behave uniformly on all instances. This partition clarifies in a completely mathematical way the previously mentioned terms and reveals that in classifying an algorithm as homogeneous or not best case analysis is equally important with worst case analysis. | cs.DS | cs | Homogeneous and Non Homogeneous Algorithms
Paparrizos K. Ioannis
Department of Informatics
Aristotle University of Thessaloniki
54124 Thessaloniki, Greece
E-mail: [email protected]
Abstract
Motivated by recent best case analyses for some sorting algorithms and based on the type of
complexity we partition the algorithms into two classes: homogeneous and non homogeneous
algorithms. Although both classes contain algorithms with worst and best cases, homogeneous
algorithms behave uniformly on all instances. This partition clarifies in a completely
mathematical way the previously mentioned terms and reveals that in classifying an algorithm
as homogeneous or not best case analysis is equally important with worst case analysis.
Keywords: Algorithm analysis, Algorithm complexity, Algorithm classification.
1. Introduction
In the 70s and 80s a lot of discussion was going on regarding the right use of the
asymptotic symbols O, Θ and Ω used to analyze algorithms and compare their
theoretical efficiency. Some researchers use these symbols to denote the rate of
growth of functions and others to denote sets of functions; see relevant comments in
[Brassard (1985)], [Gurevich (1986)] and [Knuth (1976)]. Following the approach of
using the asymptotic symbols as sets of functions we partition the class of algorithms
into two non empty subclasses: homogeneous and non homogeneous algorithms. Both
classes are wide. They contain iterative and recursive algorithms. Although both
classes contain algorithms with worst and best cases, homogeneous algorithms
behave uniformly on all instances of the problem being solved. The partition clarifies
in a completely mathematical way the terms of algorithm, worst and best case
complexity, the only difference between them being the sets of instances they referred
to.
This classification of algorithms was triggered by recent theoretical result concerning
best case analysis of some heapsort algorithms [Bollobas et.al (1996)], [Ding et.al
(1992)], [Dutton (1993)], [Edelkamp (2000)], [Edelkamp (2002)], [Fleischer (1994)]
[Schaffer et.al (1993)] and [Wang et.al (2007)]. Also, computational results indicate
that best case analysis might have practical value too, see for example [Edelkamp
(2002)] and [Wang et.al (2007)]. Our results indicate that in order to classify an
algorithm as homogeneous or not the complexity of the exact, up to a set of functions
defined by the asymptotic symbol Θ, best case and worst case must be computed.
When the classification is accomplished the analysis of the complexity of the
algorithm is complete, indicating, from a theoretical point of view, that best case
analysis is equally important with the worst case analysis.
The term inhomogeneity has been used by Nadel B.A. [Nadel (1993)] who
characterises the imprecision of an analysis of an algorithm in terms of the difference
ΔC = cw - cb between the worst and best case complexity, where C is a proper measure
of complexity. In particular, for the sorting problem C is the number of comparisons.
Using various combinations of disorder parameters, Nadel [Nadel (1993)] partitions
the set of instances in big, medium, small, tiny and singleton subclasses and computes
the inhomogeneity in each subclass. Other relevant results for other problems are
presented in [Haralick et.al (1980)],[Nadel (1986)],[Nadel (1990)] and [Nudel
(1983)]. Our approach is different in the sense that the set of algorithms is partitioned
and not the set of instances of the problem.
In the next section we formally describe the two classes of algorithms. Some details
regarding the algorithm classification are presented in section 3. Recursive and divide
and conquer homogeneous and non homogeneous algorithms are discussed and some
side results are also presented in the last section.
2. Description of the two classes.
We derive our results using the Random Access Machine (RAM) model in which
every elementary operation such as addition, subtraction, multiplication and division
of two numbers, comparison of two numbers, reading and writing a number in the
memory, calling a function, etc., is executed in constant time. It is well known that all
constant functions belong to the set Θ(1). Recall that Θ(g(n)) denotes a set of
functions defined as follows:
Definition 1. Given a function g(n) we denote by Θ(g(n)) the set of functions t(n) for
which there exists constants a > 0 and b > 0 and a positive integer n0 such that
b g(n) ≤ t (n) ≤ a g(n)
(1)
for every n ≥ n0.
All functions used in this paper denote time and therefore they are positive. The
argument n denotes the dimension of the problem and, hence, it is a positive integer.
The sets of functions O(g(n)) and Ω(g(n)) are similarly defined. Simply, in the
definition of O(g(n)) the left inequality of (1) is missing while in the definition of
Ω(g(n)) the right. Observe that Θ(g(n)) is strictly contained in the sets O(g(n)) and
Ω(g(n)). As a result the assumption that the basic operations are executed in Θ(1)
time (instead of O(1) or Ω(1) time) provides a more precise algorithm analysis.
It is well known that the symbol Θ considered as a binary relation between functions
is reflexive, symmetric and transitive and therefore it partitions the set of functions
into disjoined classes. In other words, if f(n) and g(n) are two different functions, then
either Θ(f(n)) = Θ(g(n)) or Θ(f(n)) ∩ Θ(g(n)) = Ø. In particular the following two
results are well known.
Theorem 1. If f(n) ∈ Θ(g(n)), then Θ(f(n)) = Θ(g(n)).
Theorem 2. The sets Θ(1) and Θ(n) are disjoint.
Given a computational problem we denote the set of instances of dimension n by I(n).
Consider now an algorithm A solving the problem under consideration. The time
taken by algorithm A to solve instance i of dimension n is denoted by tA(i, n). In
algorithm analysis we try to describe in a nice way the set of time functions
Ѕ = { tA(i, n) : i∈ I(n) }
One way to do this is via the sets of functions defined by the asymptotic symbols O,
Θ, Ω. We are completely satisfied if we can determine a function g(n) such that
(2)
S ⊆ Θ(g(n))
Once again, observe that we use the set Θ(g(n)) which is strictly contained in the sets
O(g(n)) and Ω(g(n)) and therefore the description of set S is more precise. This
preference though leads us naturally to the following definition.
Definition 2. An algorithm is homogeneous if there exists a function g(n) such that
relation (2) holds. Otherwise, the algorithm is non homogeneous.
Theorem 3. The class of algorithms is partitioned into two non empty and disjoined
subclasses, the subclasses of homogeneous and non homogeneous algorithms.
Proof. Let U be the class of all algorithms, H the class of homogeneous and NH the
class of non homogeneous algorithms. It is obvious from Definition 2 that
H ∩ NH = Ø and H ∪ NH = U.
It remains to show that H ≠ Ø and NH ≠ Ø. This proof is done by providing a simple
algorithm for each class.
Firstly, consider the problem of finding the smallest among n given numbers stored as
elements of an array T. The algorithm min whose pseudocode is the following
a←T(1)
for j = 2, 3,…, n
if T(j) < a, a←T(j),
solves this problem and is homogeneous. Indeed assuming that an element of an array
is reached in constant time Θ(1) in the computational model of constant times, it is
easy to conclude that
tmin(i, n) ∈ Θ(n)
for every instance i ∈ I(n). Hence, algorithm min is homogeneous and H ≠ Ø.
Secondly, consider the following problem. Given an array T of n elements sorted in
increasing order, i.e.
T(j) ≤ T(j+1) for i = 1, 2,…, n-1,
and a number x, sort all elements of T and the number x in increasing order.
This problem is solved by the algorithm insert whose pseudocode is the following:
j←n, T(n+1)←x
while (j ≥ 1) and (T(j) > T(j+1))
temp←T(j), T(j)←T(j+1), T(j+1)←temp
j←j-1
Denote by ib the instance T = [1 2 3 … n-1 n] and x = n+1. When algorithm insert is
applied on instance ib, the while loop is executed once and hence,
(3)
tinsert(ib, n) ∈ Θ(1).
Denote now by iw the instance T = [1 2 3 … n-1 n] and x = 0. When algorithm insert
is applied on instance iw, the while loop is executed Θ(n) times and therefore
(4)
tinsert(iw, n) ∈ Θ(n).
This is so because the first two assignments of the pseudo code insert are executed in
Θ(1) time and each execution of the while loop takes Θ(1) time.
We show now that there is no function g(n) such that relation (2) holds. This in turn
shows that algorithm insert is non homogeneous. Suppose on the contrary that such a
function g(n) does exist. By relation (2) we conclude that
tinsert(ib, n) ∈ Θ(g(n)) and tinsert(iw, n) ∈ Θ(g(n))
By Theorem 1 and relations (3) and (4) we conclude that
Θ(tinsert(ib, n)) = Θ(1) and Θ(tinsert(iw, n)) = Θ(n)
Combining Theorem 1 and relations (5) we conclude that
Θ(tinsert(ib, n)) = Θ(tinsert(iw, n)) = Θ(g(n))
(7)
(5)
(6)
Finally, from relations (6) and (7) we conclude that Θ(1) = Θ(n), which contradicts
Theorem 2. This completes the proof of the Theorem.
In the proof of Theorem 3 we used two simple algorithms to show that the classes of
homogeneous and non homogeneous algorithms are non empty. In fact both classes
are wide and include recursive and iterative algorithms. The class of non
homogeneous algorithms includes plenty of iterative algorithms. The great majority
of recursive and divide and conquer algorithms are homogeneous. Among the
exceptions is the well known recursive sorting algorithm quick sort [Hoare (1962)]
and Euclid’s algorithm for computing the greatest common divisor of two numbers.
3. Algorithm classification
The instances ib and iw used in Theorem 3 are the well known best and worst cases
respectively. We call ib minimum time instance and iw maximum time instance. More
precisely, we give the following definition.
Definition 3. An instance i is a minimum (maximum) time instance for an algorithm
A, if the total number of elementary operations executed when algorithm A is applied
on it is the minimum (maximum) possible.
The analysis so far and particularly algorithm min used in the proof of Theorem 3
might mislead someone to conclude that homogeneous algorithms do not contain
minimum and maximum time instances. This is not correct. A striking example of an
iterative homogeneous algorithm containing minimum and maximum time instances
is the well known Floyd’s classical algorithm [Floyd (1964)] for building an initial
heap. A heap is a data structure introduced in [Williams (1964)] to develop an
efficient general iterative sorting algorithm known today as heapsort. A recursive
homogeneous algorithm containing worst and best cases is the well known algorithm
in [Blum et.al (1973)], which computes order statistics in linear time.
Some algorithms are obviously homogeneous. If this is not clear for a new algorithm
with unknown complexity, using Definition 3 we can set
Sb = { ib: ib ∈ I(n) is a minimum time instance},
Sw = { iw: iw ∈ I(n) is a maximum time instance}.
In the worst (best) case analysis of an algorithm we try to determine a set Θ(g(n))
(Θ(f(n))) containing the set Sw (Sb) and say that the worst (best) case complexity of
the algorithm is Θ(g(n)) (Θf(n))). Observe the similarities among the worst and best
case complexities of a non homogeneous algorithm and the complexity of a
homogeneous algorithm. In particular, the only difference is the set of instances on
which they are referred to. Therefore, all these complexities should be described by
sets of the form Θ(g(n)).
It is now of interest to determine the complexity of a non homogeneous algorithm, i.e,
to find a set of functions including set S. Since a set of the form Θ(g(n)) does not
exist, we generalize Definition 1 as follows.
Definition 4. Given two (proper) functions f(n) and g(n) we denote by Θ(f(n), g(n))
the set of functions t(n) for which there exist constants a > 0 and b > 0 and a positive
integer n0 such that
bf(n) ≤ t(n) ≤ ag(n)
for n ≥ n0.
It is easy to see that Θ(f(n), g(n)) = Ω(f(n)) ∩ O(g(n)). It is also easy to see that the
sets Θ(0, ∞) = Ω(0) = Ο(∞) include always set S. However, in order to be as precise
as possible, we are always looking for a minimal set containing set S. In the case of
non homogeneous algorithms we are seeking the minimal set Θ(f(n), g(n)) containing
set S. Obviously, the set Θ(f(n), g(n)) is minimal if there exist worst and best case
instances iw and ib such that t(iw, n) ∈ Θ(g(n)) and t(ib, n) ∈ Θ(f(n)), respectively.
Recall that the set Θ(1, n) describing the complexity of algorithm insert in Theorem 3
is minimal. Observe also that the classification of an algorithm as homogeneous or
not is not possible unless the set Θ(f(n), g(n)) describing its complexity is minimal.
As the set Θ(f(n), g(n)) is described by best and worst case complexities, both
complexities are equally important from the theoretical point of view.
i
,
t ( n )
4. Additional results and discussion
We mentioned earlier that homogeneous algorithms contain worst and best cases.
Hence, the average complexity of a homogeneous algorithm is easily defined.
Clearly, the mean time of the algorithm on a random instance is
= ∑
I ( n ) t ( i , n )
∈
I ( n )
where I(n) denotes the number of elements of set I(n). If the complexity of the
homogeneous algorithm is Θ(g(n)), it is natural to expect that t(n) ∈ Θ(g(n)). Indeed,
this is the case.
Theorem 4. The average complexity of a homogeneous algorithm of complexity
Θ(g(n)), is also Θ(g(n)).
Proof. Let t(n) be the expected time to solve a random instance. Then
∑
∑
t ( i , n )
( g ( n ))
∈ Θ
i
i
I ( n )
I ( n )
∈
I ( n )
I ( n )
I ( n )
(g ( n ))
Θ
I ( n )
= Θ
(g ( n ))
t ( n )
=
∈
=
and the proof is complete.
Observe that this result is independent of the distribution of the instances.
So far we focused our attention on iterative algorithms. Recursive algorithm can be
homogeneous and non homogeneous too. But how recursive homogeneous and non
homogeneous algorithms look like? A recursive or divide and conquer algorithm
makes a fixed number of calls to itself. Therefore, if each call is made on a problem
with fixed dimension, the algorithm is homogeneous provided the work required to
solve all sub-problems dominates the remaining work. On the contrary, if the
dimensions of the sub problems on which calls are made are not fixed and depend on
the instance, the algorithm might very well be non homogeneous. Recall that this is
the case for the algorithm quicksort [Hoare (1962)]. A recursive or divide and
conquer algorithm can be non homogeneous if the number of calls to sub problems is
not fixed and depends on the instance. This is the case for Euclid’s algorithm
computing the greatest common divisor.
5. Acknowledge
We thank an anonymous referee for useful suggestions and for bringing to our
attention the reference [Nadel (1993)]
6. References
1. Blum M., Floyd R., Pratt V., Rivest R. and Tarjan R. Time bounds for selection,
Journal of Computer and System Sciences 7(4), 448-461, 1973
2. Bollobas B., Fenner T. I. and Frieze A. M. On best case of heapsort, Journal of
Algorithms, 20, 205-217, 1996
3. Brassard G. Crusade for a better notation, ACM Sigact News, 17(1), 60-64, 1985
4. Ding Y. and Weiss M.A. Best case lower bounds for Heapsort, Computing 49, 1-
9, 1992
5. Dutton R. Weak-heapsort, BIT 33, 372-381, 1993
6. Edelkamp S. and Wegener I. On the performance of weak heasort, STACS,
Lecture Notes in Computer Science, Springer – Verlag, 254-266, 2000
7. Edelkamp S. and Stiegeler P. Implementing heapsort with nlogn - 0.9n and
quicksort with nlogn + 0.2n comparisons, ACM Journal of experimental
algorithmics (JEA) 7(1), 1-20,2002
8. Fleischer R. A tied lower bound for the worst case of bottom – up heapsort,
Algorithmica 11, 104 – 115, 1994
9. Floyd R. Algorithm 245: treesort 3, Communication of the ACM, 7, 701, 1964
10. Gurevich Y. What does O(n) mean?, ACM Sigact News, 17(4), 61-63, 1986
11. Hoare A. Quicksort, Computer Journal, 5, 10-15, 1962
12. Knuth D. Big omicron and big theta and big omega, ACM Sigact News, 8(2),
18-23, 1976
13. Schaffer R. and Sedgwick R. The analysis of heapsort, Journal of algorithms 15,
76 – 100, 1993
14. Wang X. D. and Wu Y. J. An improved heapsort algorithm with nlogn –
0.788928n comparisons in the worst case, Journal of Computer Science and
Technology, 22, 6, 898 – 903, 2007
15. Williams J.W.J. Algorithm 232: heapsort, Communication of the ACM, 6, 347-
348, 1964
16. Nadel B.A., Precision Complexity Analysis: A Case Study using
Insertion Sort, Information Sciences, 73, pp. 139-189, 1993.
17. Haralick R.M. and Elliot G.L., Increase tree search efficiency for constraint
satisfaction problems, Artificial Intell. 14:263-313 (1980).
18. Nadel B.A., The consistent labeling problem and its algorithms: Towards exact-
case complexities and theory-based heuristics, Ph.D. dissertation Dept. Computer
Science, Rutgers Univ., New Brunswick, NJ, May 1986.
19. Nadel B.A., Representation selection for constrain satisfaction: A case study
using n-queens, IEEE Expert 5(3):16-23 (June 1990).
20. Nadel B.A., The complexity of constraint satisfaction in Prolog, in Proc. 8th Nat.
Conf. Artificial Intell. (AAAI`90), Boston, MA, Aug. 1990, pp. 33-39. An
expanded version is available as Technical Report CSC-89-004, 1989, Dept.
Computer Science, Wayne State Univ. Detroit, MI.
21. Nudel B.A., Sovling the general consistent labeling (or constraint satisfaction)
problem: Two algorithms and their expected complexities, in Proc. 3rd Nat. Conf.
Artificial Intell. (AAAI`83), Washington, DC, Aug. 1983, pp. 292-296.
|
1509.00930 | 1 | 1509 | 2015-09-03T03:27:55 | Testing Properties of Functions on Finite Groups | [
"cs.DS"
] | We study testing properties of functions on finite groups. First we consider functions of the form $f:G \to \mathbb{C}$, where $G$ is a finite group. We show that conjugate invariance, homomorphism, and the property of being proportional to an irreducible character is testable with a constant number of queries to $f$, where a character is a crucial notion in representation theory. Our proof relies on representation theory and harmonic analysis on finite groups. Next we consider functions of the form $f: G \to M_d(\mathbb{C})$, where $d$ is a fixed constant and $M_d(\mathbb{C})$ is the family of $d$ by $d$ matrices with each element in $\mathbb{C}$. For a function $g:G \to M_d(\mathbb{C})$, we show that the unitary isomorphism to $g$ is testable with a constant number of queries to $f$, where we say that $f$ and $g$ are unitary isomorphic if there exists a unitary matrix $U$ such that $f(x) = Ug(x)U^{-1}$ for any $x \in G$. | cs.DS | cs |
Testing Properties of Functions on Finite Groups
Kenta Oono
Yuichi Yoshida∗
Preferred Networks, Inc.
National Institute of Informatics, and
[email protected]
Preferred Infrastructure, Inc.
[email protected]
September 4, 2015
Abstract
We study testing properties of functions on finite groups. First we consider functions of the
form f : G → C, where G is a finite group. We show that conjugate invariance, homomorphism,
and the property of being proportional to an irreducible character is testable with a constant
number of queries to f , where a character is a crucial notion in representation theory. Our
proof relies on representation theory and harmonic analysis on finite groups. Next we consider
functions of the form f : G → Md(C), where d is a fixed constant and Md(C) is the family of d
by d matrices with each element in C. For a function g : G → Md(C), we show that the unitary
isomorphism to g is testable with a constant number of queries to f , where we say that f and
g are unitary isomorphic if there exists a unitary matrix U such that f (x) = U g(x)U −1 for any
x ∈ G.
1
Introduction
In property testing [27, 14], we want to decide whether the input function f satisfies a predetermined
property P or “far” from it. More specifically, an algorithm is called a tester for a property P if,
given a query access to the input function f and a parameter ǫ > 0, it accepts with probability at
least 2/3 when f satisfies P, and rejects with probability at least 2/3 when f is ǫ-far from P. If
a tester accepts with probability one when f satisfies P, then it is called a one-sided error tester.
The definition of ǫ-farness depends on the model, but for the case of Boolean functions, we say that
a function f : Fn
2 → {−1, 1} is ǫ-far from P if the distance dist(f, g) := Prx[f (x) 6= g(x)] of f and
g is at least ǫ for any function g : Fn
2 → {−1, 1} satisfying P. Here −1 and 1 are corresponding
to true and false, respectively. The efficiency of the tester is measured by the number of queries
to f , called the query complexity. We say that a property P is constant-query testable if there is a
tester with query complexity depending only on ǫ (and not on n at all).
The study of testing properties of Boolean functions, or more generally, functions on finite fields
was initiated by Rubinfeld and Sudan [27], and then subsequently many properties have been shown
to be constant-query testable [14, 11, 16, 13]. To incorporate the algebraic structure of the finite
field, Kaufman and Sudan [20] asked to study affine-invariant properties, that is, properties P such
that, if f : Fn
p → {0, 1} satisfies P, then f ◦ A also satisfies P for any bijective affine transformation
∗Supported by JSPS Grant-in-Aid for Young Scientists (B) (No. 26730009), MEXT Grant-in-Aid for Scientific
Research on Innovative Areas (No. 24106001), and JST, ERATO, Kawarabayashi Large Graph Project.
1
A. A lot of progress has been made on the study of the constant-query testability of affine-invariant
properties [10, 17], and finally (almost) complete characterizations of constant-query testability were
achieved [9, 33]. For further details on function property testing, refer to [8, 25] for surveys.
Besides finite fields, functions over finite groups such as the cyclic group and the permuta-
tion group are also objects that naturally appear in various contexts, e.g., circuit complexity [1],
computational learning [32], and machine learning [18]. Despite its importance, there are only a
few works on testing properties on functions over finite groups [5, 14], and extending this line of
research is the main focus of the present paper. More specifically, we consider testing properties of
functions f of the form f : G → D, and more generally, f : G → D(d), where G is a finite group,
D = {z ∈ C z ≤ 1} is the unit disk, and D(d) = {A ∈ Md(C) kAkF ≤ 1} is the set of d by d
matrices with Frobenius norm at most one. Note that D(1) = D. We regard d as a constant. The
reason that we use D and D(d) is that they are maximal sets closed under multiplication. Below,
we get into the details of these two settings.
Testing properties of functions of the form f : G → D: We define the distance between two
2kf − gk2, where kfk2 :=pEx∈G f (x)2 is the L2 norm.
functions f, g : G → D as dist(f, g) = 1
Note that dist(f, g) is always in [0, 1]. We say that a function f : G → D is ǫ-far from a property P
if dist(f, g) ≥ ǫ for any function g : G → D satisfying P. We note that, for {−1, 1}-valued functions
f, g : G → {−1, 1}, we have Pr[f (x) 6= g(x)] = ǫ if and only if dist(f, g) = √ǫ holds. Hence, we have
a quadratic gap between our definition and the standard definition using the Hamming distance.
However, we adopt the L2 norm as it is more friendly with our analysis.
We first show the following:
• Conjugate invariance, that is, f (yxy−1) = f (x) for any x, y ∈ G, is one-sided error testable
with O(1/ǫ2) queries.
• Homomorphism, that is, f (x)f (y) = f (xy) for any x, y ∈ G, is one-sided error testable with
O(1/ǫ2 log(1/ǫ)) queries.
When G = Fn
of property testing [4, 6, 14, 28]. Indeed in this case, a function f : Fn
and only if f (x) = χS(x) := (−1)Pi∈S xi for some S ⊆ {1, . . . , n}.
We show the constant-query testability of conjugate invariance by a simple combinatorial argument.
2 , then homomorphism is often called linearity and intensively studied in the area
2 → D is homomorphism if
The case G = Sn, the permutation group of order n, is easy to understand. Let f : Sn → D
be a function on the permutation group. If f is conjugate invariant, then the value of f (π) only
depends on the cycle pattern of π. If f is homomorphism, then f is the all-zero function, the all-one
function, or the function that returns the sign of the input permutation.
We note that Ben-Or et al. [5] studied the constant-query testability of homomorphism from a
finite group to another finite group. The query complexity of their algorithm is O(1/ǫ), where the
distance is measured by the Hamming distance. Their algorithm and analysis by a combinatorial
argument extends to our setting, in which the range is D and the distance is measured by the
L2 norm, with query complexity O(1/ǫ2). In this sense, our result on homomorphism is not new.
Nevertheless, we prove it again using harmonic analysis over finite groups. By doing so, we can
generalize it for testing other properties.
To describe our next result, we need to introduce the basic of representation theory. A rep-
resentation of a group G is a homomorphism ϕ of the form ϕ : G → Md(C) for some integer d.
In particular, we study the family of irreducible representations, where any representation can be
2
described as the direct sum of irreducible representations. We mention that irreducible representa-
tions ϕ : G → Md(C) can be chosen as unitary, that is, ϕ(x) is unitary for all x ∈ G. The character
of a representation ϕ is the function χϕ(x) = tr(ϕ(x)), where tr(·) denotes the trace of a ma-
trix. The character carries the essential information about the representation in a more condensed
form and is intensively studied in character theory. The character of an irreducible representation
is called an irreducible character. For example, every irreducible character of Fn
2 is of the form
χS(x) = (−1)Pi∈S xi for some S ⊆ {1, . . . , n}, which is often called a character in Fourier analysis
of Boolean functions. If G is abelian, that is, commutative, then representations of G always map
to one-dimensional matrices, and hence a representation and its corresponding character coincide.
However, this is not the case when G is not abelian, which makes the analysis more involved. We
show the following:
• The property of being proportional to an irreducible character, that is, f = cχϕ for some
c ∈ C and irreducible representation ϕ, is testable with O(1/ǫ8 log2(1/ǫ)) queries.
When G = Fn
The reason that we do not consider irreducible characters themselves is that irreducible characters
may take values outside of D. In particular, χϕ(1) = d holds for the identity element 1 ∈ G and
the dimension d of ϕ.
2 , then irreducible characters coincide with linear functions, that is, χS for some
S ⊆ {1, . . . , n}. Hence, testing the property of being proportional to an irreducible character can
be seen as another generalization of linearity testing.
The form of irreducible characters is quite complicated in general. When G = Sn, however, its
combinatorial interpretation via Young Tableau is well studied [12, 21, 29] (though still complicated
to state here).
Testing properties of functions of the form f : G → D(d): Since the representations of
a group G are matrix-valued, it is natural to consider testing properties of functions of the form
f : G → D(d). For two functions f, g : G → D(d), we define dist(f, g) = 1
F .
Note that this is indeed a metric and matches the previous definition of distance when d = 1.
2qEx kf (x) − g(x)k2
Let U (d) be the set of d by d unitary matrices with each element in C. We show the following:
• Unitary equivalence to g : G → D(d), that is, f = U gU −1 for some unitary matrix U ∈ U (d),
is testable with (d3/2/ǫ)O(d2) queries.
Here g is a parameter of the problem and not a part of the input. Unitary equivalence is an
important notion when studying representations since the irreducibility of a representation ϕ is
preserved by the transformation ϕ 7→ U ϕU −1 for a unitary matrix U .
Our tester samples unitary matrices from the Haar measure, a fundamental tool in the repre-
sentation theory of Lie Groups, and then checks whether g becomes close to f by applying these
unitary matrices.
Arguably the simplest property of matrix-valued functions is again homomorphism. However,
homomorphism is known to be constant-query testable by a combinatorial argument [5], and the
harmonic analysis does not facilitate the analysis. Therefore we do not study homomorphism of
matrix-valued functions in this paper.
3
Related work: There are a number of works on testing whether a function on a finite group is a
homomorphism. Blum et al. [14] gave a tester (the BLR tester ) for homomorphism of functions on
a finite group. However, the number of queries depends on the number of generators of the group,
which may depend on the size of the group in general. Ben-Or et al. [5] gave another algorithm
without this dependency. Bellare et al. [4] gave a Fourier-analytic proof of the BLR tester when
the domain and the range are Fn
2 and F2, respectively. Our tester for homomorphism can be seen
as a generalization of their analysis to general groups. There has been an interest in improving
various parameters of homomorphism testing results, due to their applications in the construction
of probabilistically checkable proof (PCP) [3]. Bellare et al. [4] gave an almost tight connection
between the distance to homomorphism and the rejection probability of the BLR tester. Ben-
Sasson et al. [6] and Shpilka and Wigderson [28] reduced the number of random bits required by
the test as it affects the efficiency of the proof system and in turn the hardness of approximation
results that one can achieve using the proof system. Rubinfeld [26] studied properties of a function
on a finite group that are defined by functional equations and gave a sufficient condition of constant-
query testability.
Using the Lp norm for p ≥ 1 as a distance measure in property testing is recently systematically
studied by Berman et al. [7]. One of their motivations is exploring the connection of property
testing with learning theory and approximation theory. For this purpose, the Lp norm is more
favorable than the Hamming distance because, in learning theory and approximation theory, we
typically measure errors in the Lp norm for p = 1 or 2. Indeed, several lower bounds and upper
bounds for property testing in the Lp norm were shown using this connection. See [7] for more
details.
Representation theory is one of the most important areas in modern mathematics. Representa-
tion theory itself is intensively studied and it is also used as an analytical tool in harmonic analysis,
invariant theories, and modular theory. Representation theory have been used in various problems
of theoretical computer science such as constructing pseudorandom objects [2, 19], circuit com-
plexity [1], communication complexity [24], computational learning [32], machine learning [18], and
quantum property testing [23].
In Section 2, we introduce representation theory, harmonic analysis on finite
Organization:
groups, and the Haar measure in more detail. We discuss the testability of conjugacy invariance in
Section 3. We show that homomorphism and the property of a being proportional to an irreducible
character are constant-query testable in Sections 4 and 5, respectively. Section 6 is devoted to
testing unitary equivalence.
2 Preliminaries
For an integer n ≥ 1, [n] denotes the set {1, 2, . . . , n}. Let δij be Kronecker’s delta, that is, δij = 1
if i = j and δij = 0 if i 6= j. For a matrix M , we denote its (i, j)-th element by Mij. We write the
real and imginary part of a complex number z as ℜz and ℑz, respectively (hence z = ℜz +√−1ℑz).
For a complex number z, z denotes its conjugate.
We frequently use the following lemma.
Lemma 2.1. Let f : G → D be a function for some finite group G. For any ǫ > 0, with probability
at least 1 − δ, we can compute an estimate z of Ex∈G[f (x)] such that z − Ex∈G[f (x)] ≤ ǫ . The
number of queries to f is O(1/ǫ2 log 1/δ).
4
O(1/ǫ2 log 1/δ).
Proof. Let c = Ex∈G[f (x)]. To estimate c, we sample x1, . . . , xs uniformly at random from G,
sPi∈[s] f (xx). Clearly, the query complexity is
where s = Θ(1/ǫ2 log 1/δ). Then, we outputec := 1
We now show thatec is indeed a good approximation to c. Since ℜ(f (x)) ≤ 1, from Hoeffding’s
inequality, we have ℜ(c) − ℜ(ec) ≤ ǫ/2 with probability at least 1 − δ/2 by choosing the hidden
constant in s large enough. Similarly, we have ℑ(c)−ℑ(ec) ≤ ǫ/2 with probability at least 1− δ/2.
By the union bound, we have c −ec ≤ ℜ(c) − ℜ(ec) + ℑ(c) − ℑ(ec) ≤ ǫ with probability at least
1 − δ.
2.1 Representation theory
We introduce basic notions and facts in representation theory. See, e.g., [30] for more details.
For a vector space V over a field F, GLF(V ) denotes the set of invertible linear transformations.
We only consider the case F = C in this paper, and hence we omit the subscript for simplicity.
A representation of G is a pair (ϕ, V ) of a finite-dimensional vector space V and a homomor-
phism ϕ : G → GL(V ), that is, ϕ(xy) = ϕ(x)ϕ(y) for every x, y ∈ G and ϕ(1) is the identity
transformation for the identity element 1 ∈ G. For a representation (ϕ, V ), V is called the repre-
sentation space of it. When V is clear from the context, we simply call ϕ a representation. The
dimension of a representation (ϕ, V ) is the dimension of V . When V is a finite-dimensional vector
space, then we say that (ϕ, V ) is finite-dimensional representation. In our argument, we only need
finite-dimensional representations.
We describe the decomposition of a representation into irreducible representations, which is a
fundamental tool used in representation theory. For a representation (ϕ, V ) and a subspace W of
V , we say that W is G-invariant if ϕ(G)W ⊆ W . If W is a G-invariant space, then we can regard
the range of ϕ as GL(W ), and hence we obtain a representation (ϕ, W ). Note that {0} and V are
G-invariant from the definition. A representation (ϕ, V ) is called irreducible if {0} and V are the
only G-invariant spaces. Note that a one-dimensional representation is always irreducible. When G
is abelian, then we have the converse from Schur’s Lemma, that is, any irreducible representation
is one-dimensional. When G is non-abelian, however, an irreducible representation might have
dimension more than one. This fact makes the analysis of algorithms for functions on a non-abelian
group more involved.
Two representations (ϕ, V ) and (ψ, W ) of G are equivalent if there exists an invertible linear
transformation T : V → W such that, for every x ∈ G, it holds that ψ(x)◦T = T ◦ϕ(x) . We identify
representations. It is known that there is a one-to-one correspondence between conjugacy classes
equivalent representations, and we denote by bG the family of equivalence classes of irreducible
of G and bG.
representations as a complete system of representatives of bG, and we identify it with bG. Since
G is finite, so is bG. For ϕ ∈ bG, we denote the dimension of its representation space by dϕ. In
A representation (ϕ, V ) is unitary if, for all x ∈ G, ϕ(x) is a unitary transformation. For any
representation of G, there is an equivalent unitary representation. Hence, we can take unitary
what follows, we fix a basis of the vector space of each representation (ϕ, V ), and we regard it as
a homomorphism from G to Mdϕ(C), where dϕ is the dimension of V .
5
2.2 Fourier analysis on non-abelian finite groups
We regard the space of C-valued functions of G as an inner product space by defining hf, gi =
Ex∈G[f (x)g(x)] for f, g : G → C. The following fact is known.
Lemma 2.2 ([30]). For a finite group G, the set(cid:8)pdϕϕij ϕ ∈ bG, i, j ∈ [dϕ](cid:9) forms an orthonor-
mal basis of the space of C-valued functions of G.
Hence, we can decompose f : G → C as
f (x) =Xϕ∈ bG
dϕ Xi,j∈[dϕ]
hf, ϕijiϕij(x) = Xϕ∈ bG
dϕ Xi,j∈[dϕ]bf (ϕ)ijϕij(x),
where bf (ϕ) ∈ Mdϕ(C) is defined as bf (ϕ) = Ex∈G[f (x)ϕ(x)] and called the Fourier coefficient of
ϕ. This decomposition is called the Fourier expansion of f . Note that Fourier coefficients are
matrix-valued functions. The following is well known.
Lemma 2.3 ([30]). Let f, g : G → C be functions. Then, we have
hf, gi =Xϕ
2 =Xϕ
kfk2
dϕ Xi,j∈[dϕ]bf (ϕ)ijbg(ϕ)ij,
dϕ Xi,j∈[dϕ]
bf (ϕ)ij2.
(Plancherel’s identity)
(Parseval’s identity)
2.3 Class functions and characters
For a representation ϕ : G → Mdϕ(C), the character χϕ : G → C of ϕ is defined as χϕ(x) = tr(ϕ(x))
for x ∈ G. We say that a function f : G → C is conjugate invariant if f (x) = f (yxy−1) for all
x, y ∈ G. A conjugate invariant function is sometimes called a class function. It is not hard to
check that characters are conjugate invariant. Indeed, the following fact is known.
Lemma 2.4. For a finite group G, the set (cid:8)χϕ ϕ ∈ bG(cid:9) forms an orthonormal basis of the space
Note that if a representation is one-dimensional, its character is identical to the original repre-
of C-valued class functions of G.
sentation, hence is a homomorphism. This is not the case in general.
The following lemma says that Fourier coefficients of a class function are always diagonal.
Lemma 2.5. For any class function f : G → C, it holds that bf (ϕ) = hf,χϕi
In order to prove Lemma 2.5, we need the following two auxiliary lemmas:
dϕ
Idϕ.
Lemma 2.6. For a function f : G → C and an irreducible representation ϕ, we have
Proof.
hf, χϕi = tr(bf (ϕ)).
ϕkk(x) =Xρ
hf, χϕi = E
dρXi,j bf (ρ)ijρij(x)Xk
xXρ
=Xk bf (ϕ)kk = tr(bf (ϕ)).
6
dρXi,j bf (ρ)ijXk
hρij, ϕkki
Lemma 2.7. Let f : G → C be a function and g : G → C be a class function. Then,
Proof. Since g is a class function, we can represent g(x) =Pϕhg, χϕiχϕ(x). Now we have
hf, gi =Xϕ
tr(bf (ϕ))tr(bg(ϕ)).
[ϕij(x)ϕ′
j ′j ′(x)]
E
x
j ′j ′i
ϕi E
x
[ϕij(x)χϕ′ (x)]
dϕXi,j bf (ϕ)ijhg, χ′
hf, gi =Xϕ,ϕ′
dϕXi,j bf (ϕ)ijXi′ bg(ϕ′)i′i′Xj ′
=Xϕ,ϕ′
dϕXi,j bf (ϕ)ijXi′ bg(ϕ′)i′i′Xj ′ hϕij, ϕ′
=Xϕ,ϕ′
=Xϕ Xi bf (ϕ)iiXi′ bg(ϕ)i′i′
=Xϕ
tr(bf (ϕ))tr(bg(ϕ)).
2 =Xϕ (cid:12)(cid:12)(cid:12)tr(bf (ϕ))(cid:12)(cid:12)(cid:12)
=Xϕ (cid:12)(cid:12)(cid:12)Xi∈[dϕ]bf (ϕ)ii(cid:12)(cid:12)(cid:12)
≤Xϕ
dϕ Xi,j∈[dϕ](cid:12)(cid:12)(cid:12)bf (ϕ)ij(cid:12)(cid:12)(cid:12)
≤Xϕ
dϕ Xi∈[dϕ](cid:12)(cid:12)(cid:12)bf (ϕ)ii(cid:12)(cid:12)(cid:12)
= kfk2
2.
kfk2
2
2
2
2
Proof of Lemma 2.5.
(from Lemma 2.6)
(from Lemma 2.7)
(by Cauchy-Schwarz)
(by Parseval’s identity)
Therefore, the equality holds for both inequalities in the formula above.
In particular, bf (ϕ) is
proportional to the identity matrix cIdϕ for some c ∈ C. By Lemma 2.6, hf, χϕi = tr(bf (ϕ)) = cdϕ.
Hence, c = hf, χϕi/dϕ, and we have the lemma.
2.4
Introduction to the Haar measure
In this section, we introduce Haar measure briefly. See, e.g., a textbook [22] for more details.
A topological group is a group equipped with a topology and whose group operations are contin-
uous in its topology. A finite group is a topological group if it is endowed with a discrete topology.
Any subgroup of Md(C) is a topological group, in which we identify Md(C) with Cd2
and introduce
the topology induced by Cd2
.
We call a measure µ on a topological group G left invariant (resp., right invariant) if µ(xS) =
µ(S) (resp., µ(S) = µ(Sx)) for any x ∈ G and Borel set S ⊆ G. Similarly, We call µ invariant
under taking inverse if µ(S−1) = µ(S) for any Borel set S ⊆ G where S−1 = {x−1 x ∈ S}.
For any compact topological group G, there exists a measure on G which is left invariant, right
invariant, and invariant under taking inverse. Such a measure is unique up to scalar multiplication
and called the Haar measure on G. For example, the Haar measure µ of a finite group G is (a
scalar multiplication of) the counting measure, that is, µ(S) = S/G for subset S ⊆ G.
Hence, the Haar measure of U (d) exists.
We regard U (d) as a closed subgroup of Md(C) and regard it as a compact topological group.
7
Algorithm 1 (Tester for conjugate invariance)
1: for s := O(1/ǫ2) times do
2:
Sample x and y ∈ G uniformly at random.
if f (x) 6= f (yxy−1) then reject.
3:
4: Accept.
3 Conjugate Invariance
In this section, we first show that conjugate-invariance is constant-query testable.
Theorem 3.1. Conjugate invariance is one-sided error testable with O(1/ǫ2) queries.
Then, we show the following lemma, which simplifies testing properties that imply conjugate
invariance.
Lemma 3.2. Let P be a property such that every f satisfying P is a class function. Suppose that
there is a tester A for P with query complexity q(ǫ) if the input is restricted to be a class function.
Then there is a tester A′ for P with query complexity O(1/ǫ2 + q(ǫ/2) log q(ǫ/2)). Moreover, if A
is a one-sided error tester, then A′ is also a one-sided error tester.
3.1 Proof of Theorem 3.1
Our algorithm for testing conjugate invariance is described in Algorithm 1. It is easy to see that
the query complexity of Algorithm 1 is O(1/ǫ2) and the tester always accepts when f is conjugate
invariant. Thus, it suffices to show that Algorithm 1 rejects with probability at least 2/3 when f
is ǫ-far from class functions.
It is well known that conjugacy classes of G form a partition of G. We define G♯ as the set of
conjugacy classes of G. Also, for an element x ∈ G, we define x♯ as the unique conjugacy class x
belongs to.
For y ∈ x♯, we define Nx,y = {z ∈ G zxz−1 = y}. Since we have Nx,y ∩ Nx,y′ = ∅ for y, y′ ∈ x♯
with y 6= y′, it holds that G =Fy∈x♯ Nx,y. Therefore, the following lemma guarantees that uniform
sampling from a conjugacy class is executed by uniformly sampling from the whole group.
Lemma 3.3. The number of elements in Nx,y depends only on the conjugacy class to which y
belongs.
−1 = y′ hold. We construct
Proof. For y, y′ ∈ x♯, fix z0, z′
mappings Φy,y′ : Nx,y → Nx,y′ by z 7→ z′
0. By a
direct calculation, we can check that Φy,y′ ◦ Φy′,y = idNx,y , and Φy′,y ◦ Φy,y′ = idNx,y′ . Therefore
Nx,y = Nx,y′ holds.
0z−1z0 and Φy′,y : Nx,y′ → Nx,y by z 7→ z0z−1z′
0 ∈ G so that z0xz−1
0 = y and z′
0xz′
0
Fix a function f : G → D. For a conjugacy class C ∈ G♯ and z ∈ C, define pC(z) := ♯{x ∈
C f (x) = z}/C as the probability that f (x) = z if we sample x ∈ C uniformly at random. We
define pC := maxz∈C pC(z) and zC := arg maxz∈C pC(z). Then, we define ef as ef (x) = zx♯. Note
that ef is a class function such that ef (x) ∈ D for any x ∈ G.
Lemma 3.4.
Pr
x,y∈G
[f (x) 6= f (yxy−1)] ≥ dist(f,ef )2.
8
Proof. Since x − y ≤ 2 for any x, y ∈ D, we have
1
4
dist(f,ef )2 =
E
x f (x) − ef (x)2 ≤ Pr
x
[f (x) 6= ef (x)] =
1
G XC∈G♯
C(1 − pC).
By Lemma 3.3, if we fix x ∈ G and sample y ∈ G uniformly at random, then yxy−1 forms a
uniform distribution over elements in x♯. Thus,
Pr
x,y∈G
1
[f (x) 6= f (yxy−1)] ≥
G XC∈G♯
1
G XC∈G♯
CZ pC(z)(1 − pC)dz =
CZ pC(z)(1 − pC(z))dz
G XC∈G♯
1
≥
C(1 − pC) ≥ dist(f,ef )2.
Lemma 3.5. If f is ǫ-far from being conjugate invariant, then Algorithm 1 rejects with probability
at least 2/3.
Proof. Since ef : G → D is a class function, we have dist(f,ef ) ≥ ǫ. Hence, the probability we reject
at Line 3 in each trial is at least ǫ2 by Lemma 3.4. Hence the tester rejects with probability 2/3
by choosing the hidden constant in s large enough.
We establish Theorem 3.1 by Lemma 3.5.
3.2 Proof of Lemma 3.2
The following lemma shows that we can obtain a query access to a class function that is close to f .
Lemma 3.6. Let f : G → D be a function that is ǫ-close to a class function. There exists a class
function f ′ : G → D with the following property.
• For any x ∈ G, with O(log 1/δ) queries to f , we can correctly compute f ′(x) or find a witness
that f is not a class function with probability at least 1 − δ. Moreover, if f itself is a class
function, then we can always compute f ′(x) correctly.
• dist(f ′, f ) ≤ 3ǫ. In particular, f ′ = f when f itself is a class function.
Proof. For a conjugacy class C, let z∗
We define f ∗ : G → D as f ∗(x) = z∗
We define f ′ : G → D as follows:
x♯. Note that f ∗ is the class function closest to f .
C ∈ D be the unique value that minimizesPx∈C f (x) − z∗
C2.
f ′(x) =(z∗
if px♯ ≤ 1
2 ,
otherwise.
x♯
zx♯
We first show the first claim. Our algorithm for computing f ′(x) is as follows. Given x ∈ G,
) for each
j ) for some i 6= j, then we reject f and output the pair as the
we pick y1, . . . , ys ∈ G for s := O(log 1/δ) uniformly at random, and compute f (yixy−1
i ∈ [s]. If f (yixy−1
witness that f is not a class function. If all of them are the same, we output the value as f ′(x).
) 6= f (yjxy−1
i
i
9
i
i
at least 1 − δ, we have f (yixy−1
) 6= f (yjxy−1
probability at least 1 − δ, the majority of {f (yixy−1
at least 1 − δ, either we output zC or reject.
algorithm always outputs f (x) as f ′(x).
Now we analyze the correctness of the algorithm above. If px♯ ≤ 1/2, then with probability
j ) for some i 6= j, and we reject. If px♯ > 1/2, with
)}i∈[s] is equal to zC. Hence, with probability
Moreover, if f itself is a class function, then we have f ′(x) = f (x) for any x ∈ G, and our
We turn to the second claim. For two functions g, h : G → C and a conjugacy class C,
define distC(g, h) :=pPx∈C g(x) − h(x)2. We will show that, for each conjugacy class C ∈ G♯,
distC(f, f ′) ≤ 3distC(f, f ∗), which implies dist(f, f ′) ≤ 3ǫ.
If C satisfies pC ≤ 1/2, we have nothing to show. Thus suppose pC > 1/2. Then, we have
distC(f, f ′) ≤ distC (f, f ∗) + distC(f ∗, f ′) =sXx∈C
C − zC2 ≤Px∈C f (x) − f ∗(x)2. By pC > 1/2, we have Cz∗
Since a pC-fraction of values has moved from zC to z∗
C when constructing f ∗ from f , we have
C − zC2 ≤ 2Px∈C f (x) −
pCCz∗
f ∗(x)2. Combining this with the previous inequality, we have distC(f, f ′) ≤ (1 +√2)distC(f, f ∗) ≤
3distC (f, f ∗).
f (x) − f ∗(x)2 +qCz∗
C − zC2.
Proof of Lemma 3.2. We first apply the ǫ/6-tester for conjugate invariance (Algorithm 1). If the
tester rejects, we immediately reject f as it implies that f does not satisfy P . Otherwise, using
Lemma 3.6, we construct a query access to a class function f ′ with δ = O(1/q(ǫ/2)). Then we
apply the tester A to f ′ with the error parameter ǫ/2. The query complexity is clearly as stated.
Suppose that f satisfies the property P . Then, we never reject when testing conjugate invari-
ance. Also f ′(x) = f (x) holds for every x ∈ G and it follows that f ′ satisfies the property P .
Hence, the tester A accepts f ′ with probability at least 2/3. Moreover if A is a one-sided error
tester, then A accepts f ′ with probability one.
Suppose that f is ǫ-far from the property P . If f is ǫ/6-far from conjugate invariance, then we
reject f with probability at least 2/3. Thus assume that f is ǫ/6-close to conjugate invariance. In
this case f ′ is a class function that is ǫ/2-close to f . Hence f ′ is still ǫ/2-far from the property P .
Then the tester A on f ′ should reject with probability 2/3.
4 Testing Homomorphism
In this section, we show the following:
Theorem 4.1. Homomorphism is one-sided error testable with O(1/ǫ2 log(1/ǫ)) queries.
We note that, if f : G → C is a homomorphism, then it is a one-dimensional representation
and hence an irreducible representation. First we observe that homomorphism implies conjugate
invariance.
Lemma 4.2. If f : G → C is a homomorphism, then f is conjugate invariant.
Proof. Since f is a homomorphism, we have for any x, y ∈ G, f (yxy−1) = f (y)f (x)f (y−1) =
f (y)f (y−1)f (x) = f (yy−1)f (x) = f (1)f (x). By setting y = 1, we have f (x) = f (1)f (x), which
means f (x) = 0 or f (1) = 1.
10
Algorithm 2 (Tester for homomorphism)
Input: A class function f : G → D.
1: for s = O(1/ǫ2) times do
2:
Sample x, y ∈ G uniformly at random.
if f (x)f (y) 6= f (xy) then reject.
3:
4: Accept.
If f (x) = 0 for all x ∈ G, then f is clearly conjugate invariant. If f (x) 6= 0 for some x ∈ G, then
f (1) = 1. In this case, we have f (yxy−1) = f (1)f (x) = f (x) and f is again conjugate invariant.
From Lemmas 3.2 and 4.2, to test homomorphism, it suffices to show that homomorphism is
one-sided error testable with O(1/ǫ2) queries when the input function is a class function. Our
tester is given in Algorithm 2. It is clear that the query complexity is O(1/ǫ2). We next see that
Algorithm 2 always accepts homomorphisms:
Lemma 4.3. If a class function f : G → D is a homomorphism, then Algorithm 2 always accepts.
Proof. We always accept because f (x)f (y) = f (xy) for any x, y ∈ G
Now we turn to the case that f is ǫ-far from homomorphisms. To show that Pr[f (x)f (y) 6=
f (xy)] is much smaller than 1, we analyze the term f (x)f (y)f (xy).
Lemma 4.4. For any function f : G → C, we have
E
x,y
[f (x)f (y)f (xy)] =Xϕ
dϕ Xi,j,k∈[dϕ]bf (ϕ)ijbf (ϕ)jkbf (ϕ)ik.
Proof. The left hand side is equal to
Xϕ,ϕ′,ϕ′′
dϕdϕ′dϕ′′ Xi,j∈[dϕ] Xi′,j ′∈[dϕ′ ] Xi′′,j ′′∈[dϕ′′ ]bf (ϕ)ijbf (ϕ′)i′j ′bf (ϕ′′)i′′j ′′ E
x,y
Now we analyze the expectation in (1).
[ϕij(x)ϕ′
i′j ′(y)ϕ′′
i′′j ′′(xy)].
(1)
[ϕij(x)ϕ′
i′j ′(y)ϕ′′
i′′j ′′(xy)] = E
E
x,y
= Xk′′∈[dϕ′′ ]
hϕij, ϕ′′
i′′k′′ihϕ′
i′j ′ϕ′′
x,yhϕij(x)ϕ′
k′′j ′′i =( 1
d2
ϕ
0
i′j ′(y) Xk′′∈[dϕ′′ ]
ϕ′′
i′′k′′(x)ϕ′′
k′′j ′′(y)i
if ϕ = ϕ′ = ϕ′′, i = i′′, j = i′ = k′′, and j′ = j′′,
otherwise.
Corollary 4.5. For any class function f : G → C, we have
Hence (1) =Pϕ dϕPi,j,j ′∈[dϕ] bf (ϕ)ijbf (ϕ)jj ′bf (ϕ)ij ′.
[f (x)f (y)f (xy)] =Xϕ
E
x,y
dϕ Xi∈[dϕ]bf (ϕ)iibf (ϕ)ii2.
11
Proof. If f is a class function, then bf (ϕ)ij = 0 for any ϕ ∈ bG and i 6= j ∈ [dϕ] by Lemma 2.5.
Hence, we have the corollary from Lemma 4.4.
The following lemma completes the proof of Theorem 4.1.
Lemma 4.6. If a class function f : G → D is ǫ-far from homomorphism, then Algorithm 2 rejects
with probability at least 2/3.
Proof. From Corollary 4.5, we have
ℜ E
x,y
≤ max
dϕ Xi∈[dϕ]bf (ϕ)iibf (ϕ)ii2
[f (x)f (y)f (x + y)] = ℜXϕ
dϕ Xi∈[dϕ]
ϕ∈ bG,i∈[dϕ]ℜbf (ϕ)ii ·Xϕ
bf (ϕ)ii2
ϕ∈ bG,i∈[dϕ]ℜbf (ϕ)ii.
ϕ∈ bG,i∈[dϕ]ℜbf (ϕ)ii · kfk2
2 ≤ max
= max
For any ϕ with dimension more than one, bf (ϕ)ii ≤ 1/2 and hence ℜbf (ϕ)ii ≤ 1/2 (see Lemma 2.5).
Now consider a one-dimensional irreducible representation ϕ. Since f is ǫ-far from homomorphism,
we have
ǫ ≤ dist(f, ϕ) =
2 + kϕk2
2 − 2ℜhf, ϕi ≤
1
2qkfk2
1
2q2 − 2ℜbf (ϕ).
2 ≤ 1 and kϕk2
2 = 1 as ϕ is a (non-zero) homomorphism. Hence, ℜbf (ϕ) ≤ 1 − 2ǫ2.
Note that kfk2
We have shown that ℜ Ex,y[f (x)f (y)f (x + y)] ≤ 1 − 2ǫ2. Since f (x)f (y)f (x + y) ≤ 1, at least
an Ω(ǫ2)-fraction of pairs (x, y) satisfy f (x)f (y)f (x + y) 6= 1. Hence we have Prx,y[f (x)f (y) =
f (x+ y)] ≤ 1− Ω(ǫ2). By choosing the hidden constant in s large enough, we reject with probability
at least 2/3.
5 Testing the Property of Being Proportional to an Irreducible
Character
In this section, we show the following:
Theorem 5.1. The property of being proportional to an irreducible character is testable with
O(1/ǫ8 log2(1/ǫ)) queries.
As any character is a class function, by Lemma 3.2, it suffices to give a tester with query
complexity O(1/ǫ8 log(1/ǫ)) that works when the input function is a class function. The following
fact is crucial for our algorithm.
Lemma 5.2 ([31]). For a function f : G → C, the following are equivalent.
1. f (x) = f (1)eχϕ(x) for some irreducible representation ϕ, where eχϕ = χϕ/dϕ.
2. f (x)f (y) = f (1) Ez∈G[f (yzxz−1)] for any x, y ∈ G.
12
2 obtained by applying Lemma 2.1 with the error parameter
Algorithm 3 (Tester for being proportional to an irreducible character)
Input: A function f : G → D.
1: Let e1 be the estimation to kfk2
2: if e1 < ǫ2/2 then accept.
3: for each i = 1 to s := O(1/ǫ4) do
4:
ǫ2/100 and the confidence parameter 1/100.
5:
Sample x, y ∈ G uniformly at random.
Let ei
parameter ǫ2/10 and the confidence parameter 1/100s.
Let ei
if ei
7:
8: accept.
6 = f (x)f (y) − f (1)ei
6 > ǫ4/100 then reject.
52.
6:
5 be the estimation to Ez[f (yzxz−1)] obtained by applying Lemma 2.1 with the error
As we can freely change the value of f (1) by multiplying a constant, the second condition is a
necesary and sufficient condition of being proportional to an irreducible character.
The most simple test based on Lemma 5.2 is checking whether f (x)f (y) ≈ f (1) Ez[f (yzxz−1)]
(by estimating the latter by sampling z ∈ G a constant number of times). However, we were unable
to handle the term Ex,y,z[f (x)f (y)f (yzxz−1)f (1)] that naturally arises when analyzing this test.
Instead, we estimate f (x)f (y) − f (1) Ez[f (xyzxz−1)] and check whether it is small. The detail is
given in Algorithm 3. It is clear that the query complexity of Algorithm 3 is O(1/ǫ8 log 1/ǫ).
Lemma 5.3. If a function f : G → C is proportional to an irreducible character, then Algorithm 3
accepts with probability at least 2/3.
Proof. By the union bound, all the estimations succeed with probability at least 2/3. Below we
assume this indeed happens.
Recall that f (x)f (y) − f (1) Ez[f (yzxz−1)] = 0 for any x, y ∈ G by Lemma 5.2. Then for each
6 = f (x)f (y) − f (1)ei
52 ≤ ǫ4/100 holds for every i ∈ [s].
52 = f (1) Ez[f (yzxz−1)] − f (1)ei
i, ei
Hence, we accept with probability at least 2/3.
Now we turn to the case that f is ǫ-far from being proportional to an irreducible character. We
need the following auxiliary lemma.
Lemma 5.4. For any function f : G → C, we have
E
x,yh(cid:12)(cid:12)f (x)f (y) − f (1) E
z
[f (yzxz−1)](cid:12)(cid:12)2i ≥ kfk2
2 min
ϕ kf − f (1)eχϕk2
2.
13
Proof. We have
dϕ Xi,j∈[dϕ]bf (ϕ)ijϕij(yzxz−1)i
ϕik(y)ϕkl(z)ϕlm(x)ϕmj (z−1)i
zhϕkl(z)ϕjm(z)i
ϕik(y)ϕlm(x) E
ϕik(y)ϕlm(x)
δkjδlm
dϕ
E
= E
zhXϕ
zhf (yzxz−1)i = E
zhXϕ
dϕ Xi,j∈[dϕ]bf (ϕ)ij Xk,l,m∈[dϕ]
dϕ Xi,j∈[dϕ]bf (ϕ)ij Xk,l,m∈[dϕ]
=Xϕ
dϕ Xi,j∈[dϕ]bf (ϕ)ij Xk,l,m∈[dϕ]
=Xϕ
dϕ Xi,j∈[dϕ]bf (ϕ)ijϕij(y) Xk∈[dϕ]
=Xϕ
ϕkk(x)
dϕ
=Xϕ
dϕeχϕ(x) Xi,j∈[dϕ]bf (ϕ)ijϕij(y)
In the third equality, we used the fact that ϕmj(z−1) = ϕjm(z). This follows from the fact that
ϕ(z−1)ϕ(z) = ϕ(1) = I and ϕ(z) is unitary.
Therefore,
E
It follows that
f (x)f (y) − f (1) E
z
dϕeχϕ Xi,j∈[dϕ]bf (ϕ)ijϕij(y)
dϕ Xi,j∈[dϕ]bf (ϕ)ij ϕij(y) − f (1)Xϕ
[f (yzxz−1)] = f (x)Xϕ
dϕ(f (x) − f (1)eχϕ(x)) Xi,j∈[dϕ]bf (ϕ)ij ϕij(y)
=Xϕ
x,yh(cid:12)(cid:12)Xϕ
[f (yzxz−1)](cid:12)(cid:12)2i = E
x,yh(cid:12)(cid:12)f (x)f (y) − f (1) E
=Xϕ,ϕ′
2 Xi,j∈[dϕ](cid:12)(cid:12)(cid:12)bf (ϕ)ij(cid:12)(cid:12)(cid:12)
=Xϕ
≥kfk2
In the third equality, we used the fact that Ey(cid:2)ϕij(y)ϕ′
x(cid:2)(f (x) − f (1)eχϕ(x))(f (x) − f (1)eχϕ′(x))(cid:3) Xi,j∈[dϕ] Xi′,j ′∈[dϕ′ ]bf (ϕ)ijbf (ϕ′)i′j ′ E
dϕ Xi,j∈[dϕ](cid:12)(cid:12)(cid:12)bf (ϕ)ij(cid:12)(cid:12)(cid:12)
dϕkf − f (1)eχϕk2
ϕ kf − f (1)eχϕk2
dϕ(f (x) − f (1)eχϕ(x)) Xi,j∈[dϕ]bf (ϕ)ij ϕ(y)ij(cid:12)(cid:12)2i
y(cid:2)ϕij(y)ϕ′
2Xϕ
ϕ kf − f (1)eχϕk2
i′j ′(y)(cid:3) is equal to 1/dϕ if ϕ = ϕ′, i = i′, and
j = j′, and is equal to zero otherwise.
i′j ′(y)(cid:3)
2
≥ min
dϕdϕ′ E
2 min
2
2
z
Lemma 5.5. If a function f : G → C with f (1) = 1 is ǫ-far from being proportional to an
irreducible character, then Algorithm 3 rejects with probability at least 2/3.
Proof. By the union bound, with probability at least 5/6, all the estimations succeed. Below we
assume it indeed happens.
From Lemma 5.4, when f is ǫ-far, the expectation of ei
6 is at least
E
x,yh(cid:12)(cid:12)f (x)f (y)− f (1)ei
5(cid:12)(cid:12)2i ≥ (1−
2ǫ2
10
)kfk2
2 min
ϕ kf − f (1)eχϕk2
2−
14
ǫ4
100 ≥ (1−
ǫ2
5
)
ǫ2
4 ·kfk2
2−
ǫ4
100 ≥
ǫ4
25
.
Algorithm 4 (Tester for unitary equivalence)
Input: Functions f, g : G → D(d)
1: for s :=(cid:0)d3/2/ǫ(cid:1)Θ(d2) times do
2:
3:
Sample U ∈ U (d) with respect to the (normalized) Haar measure of U (d).
Let e be the estimation of dist(f, U gU ∗) obtained by applying Lemma 2.1 with the error
parameter ǫ2/100 and the confidence parameter 1/6s.
if e < ǫ2/10 then accept.
4:
5: Reject.
6 > ǫ4/100]. Then, we have 2· p + ǫ4/100·
We also note that ei
(1− p) ≥ ǫ4/25, and it follows that p ≥ ǫ4/100. By choosing the hidden constant in s large enough,
we reject f with probability at least 2/3.
6 ≤ (1 + ǫ2/100)2 ≤ 2. Let p = Prx,y[ei
6 Testing Unitary Equivalence
In this section, we prove the following:
Theorem 6.1. The unitary equivalence to g : G → D(d) is testable with(cid:0)d3/2/ǫ(cid:1)O(d2) queries.
Our algorithm is described in Algorithm 4. We use the Haar measure on U (d) to sample
unitary matrices. We do not need the detailed definition of the Haar measure, and we only have to
understand that it defines a probability distribution on U (d). See Section 2.4 for a brief introduction
to the Haar measure.
The basic idea of our algorithm and analysis is the following. Suppose that functions f, g :
G → D(d) are unitary equivalent, that is, f = U0gU ∗
0 for some unitary matrix U0 ∈ U (d). Then,
by sampling a sufficient number of unitary matrices from the Haar measure, we get a unitary
matrix U that is sufficiently close to U0 in the sense that the Frobenius norm of U − U0 is small
(Lemma 6.3). Then, we can show that the Frobenius norm of f (x) − U g(x)U ∗ is also small for
any x ∈ G (Lemma 6.5). On the other hand, if f and g are ǫ-far from being unitary equivalent,
then the average Frobenius norm of f (x) − U g(x)U ∗ over x ∈ G is large for any unitary matrix
U . Hence, by checking whether there is a unitary matrix U (in the sample) such that the average
Frobenius norm is small, we can distinguish the case that f and g are unitary equivalent from the
case that f and g are ǫ-far from being unitary equivalent.
Let U ∈ U (d) be a random matrix sampled with respect to the Haar measure. We diagonalize
U as U = W ΛW ∗ where W ∈ U (d) and Λ = diag(λ1, . . . λd). By the unitarity of U , the absolute
value of each eigenvalue of U is 1. Therefore we can write λi = exp(√−1θi) for some θi ∈ [−π, π),
which we call the phase of λi. We use the following proposition, which is Weyl’s integral formula
applied to U (d).
ZdQi>j λi −
Proposition 6.2 ([15]). The distribution µ of the phases θ = (θ1, . . . , θd) is dµ(θ) = 1
λj2dθ, where Zd := (2π)dd! is a normalization constant and dθ is a standard Euclid measure.
For ǫ > 0, we write BU (d)(ǫ) = {U ∈ U (d) kU − IdkF ≤ ǫ}. We need the following auxiliary
lemma, which says that the set of a sufficiently large number of randomly chosen unitary matrices
forms an “ǫ-net” of unitary matrices with respect to the Frobenius norm.
15
Lemma 6.3. Let U0 ∈ U (d) and U be a random matrix sampled with respect to the Haar measure
of U (d). For sufficiently small ǫ > 0, the probability Pr[kU − U0kF ≤ ǫ] ≥ δ6.3(ǫ, d), where
δ6.3(ǫ, d) =(cid:0) ǫ
d3/2(cid:1)O(d2).
i=1 λi − 12 = kθk2
Proof. Since the Haar measure is invariant under left multiplication, we can assume U0 = Id without
loss of generality. Hence, we want to bound Pr[U ∈ BU (d)(ǫ)]. Let λ1, . . . , λd be eigenvalues of U and
θ1, . . . , θd be corresponding phases. By the conjugate invariance of Frobenius norm, U ∈ BU (d)(ǫ)
iffPd
i=1 λi − 12 ≤ ǫ2.
Suppose kθk2 ≤ 3ǫ/4. Since θi’s are sufficiently small, we can expand as λi = 1+√−1θi +O(θ2
i ).
ThenPd
2 maxi θi < ǫ2. It implies the probability is
bounded below by Pr[U ∈ BU (d)(ǫ)] ≥RBd(3ǫ/4) dµ where Bd(r) is the ball in Rd of radius r centered
Let ǫ = ǫ
d3/2 , θ0 = [0, ǫ, . . . , (d − 1)ǫ], and θ1 = [ǫ/3, 4ǫ/3, . . . , (d − 2/3)ǫ]. Note that kθ0k2
2 ≤
ǫ2/3 and kθ1k2
2 = ǫ2(6d2 − d + 2)/18d3 < ǫ2/2. Therefore, Bd(3ǫ/4) contains the d-dimensional
hypercube V = {θ ∈ Rd θ0 ≤ θ ≤ θ1}, where we write x ≤ y if xi ≤ yi for each i. Note that if
θ ∈ V , θi − θj ≥ ǫ/3 for any i 6= j. Therefore, we have
2 + O(Pd
i ) = kθk2
2 +kθk2
at the origin.
i=1 θ3
ZBd(3ǫ/4)
dµ ≥ZV
1
dµ ≥
ZdZV Yi>j
ZdZV Yi>j(cid:16) ǫ
3(cid:17)2
1
dθ =
λi − λj2dθ ∼
3(cid:17)d(d−1)ZV
Zd(cid:16) ǫ
1
≥
(By the fact that ǫ is small enough)
1
ZdZV Yi>j
dθ =
θi − θj2dθ
Zd(cid:16) ǫ
3(cid:17)d2
1
.
Let Skew(d) be the set of d-dimensional skew-Hermitian matrices, i.e., Skew(d) = {X ∈ Md(C)
X ∗ = −X}. Although the following lemma is an almost immediate consequence of the fact that
Skew(d) is the Lie algebra of U (d), we prove it for completeness.
Lemma 6.4. If ǫ > 0 is small enough, then the image of the exponential map exp : Skew(d) → U (d)
contains BU (d)(ǫ). Furthermore, if U ∈ BU (d)(ǫ), then we can choose X ∈ Skew(d) such that
exp(X) = U and kXkF ≤ 2ǫ.
Proof. We diagonalize U as W ΛW ∗, where W ∈ U (d) and Λ = diag(λ1, . . . , λd). Let λi =
exp(√−1θi) for θi ∈ R and µi = 1 − λi. For z ∈ C, we define log(1 + z) =P∞
i=1(−1)n+1zn/n. This
Taylor expansion converges for z ≤ 1. Note that log λi = √−1(θi + 2πmi) for some mi ∈ Z.
Since kU − IdkF ≤ ǫ, we can define X = log U . Note that U = exp(X) as z = exp(log z). We
have the following formulas:
X = log U = log W ΛW ∗ = W (log Λ)W ∗,
log Λ = diag(log λ1, . . . , log λd),
log λi = √−1(θi + 2πmi) = −√−1(θi + 2πmi) = − log λi,
(log Λ)∗ = diag(log λ1, . . . , log λd) = −diag(log λ1, . . . , log λd) = − log Λ.
Combining these formulas, X ∗ = W (log Λ)∗W ∗ = −W (log Λ)W ∗ = −X, that is, X ∈ Skew(d).
16
i=1 µi2 + O(µ3
i ) ≤ 2ǫ2.
F = k log Λk2
i ) and we can conclude kXk2
Next, we estimate the Frobenius norm of X. The condition kU − IdkF ≤ ǫ impliesPd
ǫ2. It implies µi is small enough. So we can expand log λi = log(1 + µi) = µi + O(µ2
fore log λi2 = µi2 + O(µ3
Pd
i=1 log λi2 =Pd
Lemma 6.5. Let A ∈ D(d) be a matrix and U, V ∈ U (d) be unitary matrices. If kU − V kF ≤ ǫ for
sufficiently small ǫ > 0, then kU AU ∗ − V AV ∗kF ≤ 3ǫ.
Proof. Let U0 = V ∗U . By a direct calculation, kU − V kF = kU0 − IdkF and kU AU ∗ − V AV ∗kF =
kU0AU ∗
0 − AkF . Therefore, we can assume that V = Id. By Lemma 6.4, there exists X ∈ Skew(d)
such that kXkF ≤ 1 and U = exp(2ǫX). Note that U ∗ = exp(−2ǫX). Then,
i=1 µi2 ≤
i ). There-
F =
F = kW (log Λ)W ∗k2
U ∗AU − A = exp(−2ǫX)A exp(2ǫX) − A = (Id − 2ǫX + O(ǫ2)J)A(Id + 2ǫX + O(ǫ2)J)
= 2ǫ(AX − XA) + O(ǫ2)J,
where O(ǫ2)J denotes a matrix with each entry having absolute value O(ǫ2). We evaluate the
Frobenius norm of the commutator as
kAX − XAkF ≤ kAXkF + kXAkF = 2kXkF ≤ 2,
where we use the triangle inequality of the Frobenius norm, and the assumption kAkF ≤ 1. There-
fore,
kU AU ∗ − Ak2
F ≤ 4ǫ2kAX − XAk2
F + O(ǫ3) ≤ 8ǫ2 + O(ǫ3) < 9ǫ2.
Now, we prove Theorem 6.1.
Proof of Theorem 6.1. We assume that the all the estimations have succeeded, which happens with
probability at least 5/6 by the union bound.
Suppose that f and g are unitary equivalent and U0 ∈ U (d) be the unitary matrix with f (x) =
U0g(x)U ∗
0 for any x ∈ G. From Lemma 6.3, we sample U such that kU0 − UkF ≤ ǫ/10 with
F ] ≤ 9ǫ2/100. For such
probability at least 5/6. From Lemma 6.5, we have Ex∈G[kf (x)− U g(x)U ∗k2
U , we obtain the estimation e satisfies e < ǫ2/10. By the union bound, we accept with probability
at least 2/3.
Suppose that f and g are ǫ-far from being unitary equivalent. For every U we sample, we have
F ] ≥ 4ǫ2. Hence, the estimation e satisfies e > ǫ2/10 and we accept. To
Ex∈G[kf (x) − U g(x)U ∗k2
summarize, we reject with probability at least 5/6 > 2/3.
Acknowledgements
We thank Mitsuru Kusumoto and anonymous referees for comments that greatly improved the
manuscript.
17
References
[1] E. Allender, J. Jiao, M. Mahajan, and V. Vinay. Non-commutative arithmetic circuits: depth
reduction and size lower bounds. Theoretical Computer Science, 209(1-2):47–86, 1998.
[2] N. Alon and S. Lovett. Almost k-wise vs. k-wise independent permutations, and uniformity
for general group actions. Theory of Computing, 9(1):559–577, 2013.
[3] S. Arora, C. Lund, R. Motwani, M. Sudan, and M. Szegedy. Proof verification and the hardness
of approximation problems. Journal of the ACM, 45(3):501–555, 1998.
[4] M. Bellare, D. Coppersmith, J. Hastad, M. Kiwi, and M. Sudan. Linearity testing in charac-
teristic two. IEEE Transactions on Information Theory, 42(6):1781–1795, 1996.
[5] M. Ben-Or, D. Coppersmith, M. Luby, and R. Rubinfeld. Non-abelian homomorphism testing,
and distributions close to their self-convolutions. Random Structures & Algorithms, 32(1):49–
70, 2007.
[6] E. Ben-Sasson, M. Sudan, S. Vadhan, and A. Wigderson. Randomness-efficient low degree tests
and short PCPs via epsilon-biased sets. In Proceedings of the 35th annual ACM symposium
on Theory of computing (STOC), pages 612–621, 2003.
[7] P. Berman, S. Raskhodnikova, and G. Yaroslavtsev. Lp-testing. In Proceedings of the 46th
Annual ACM Symposium on Theory of Computing (STOC), pages 164–173, 2014.
[8] A. Bhattacharyya. Guest column: on testing affine-invariant properties over finite fields. ACM
SIGACT News, 44(4):53–72, 2013.
[9] A. Bhattacharyya, E. Fischer, H. Hatami, P. Hatami, and S. Lovett. Every locally characterized
affine-invariant property is testable. In Proceedings of the 45th Annual ACM Symposium on
Theory of Computing (STOC), pages 429–436, 2013.
[10] A. Bhattacharyya, E. Fischer, and S. Lovett. Testing low complexity affine-invariant properties.
Proceedings of the 24th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages
1337–1355, 2012.
[11] A. Bhattacharyya, S. Kopparty, G. Schoenebeck, M. Sudan, and D. Zuckerman. Optimal test-
ing of Reed-Muller codes. In Proceedings of the 51st Annual IEEE Symposium on Foundations
of Computer Science (FOCS), pages 488–497, 2010.
[12] P. Biane. Characters of symmetric groups and free cumulants. Asymptotic Combinatorics with
Applications to Mathematical Physics, 1815:185–200, 2003.
[13] E. Blais, A. Weinstein, and Y. Yoshida. Partially symmetric functions are efficiently
isomorphism-testable. In Proceedings of the 53rd Annual IEEE Symposium on Foundations of
Computer Science (FOCS), pages 551–560, 2012.
[14] M. Blum, M. Luby, and R. Rubinfeld. Self-testing/correcting with applications to numerical
problems. Journal of Computer and System Sciences, 47(3):549–595, 1993.
18
[15] T. Brocker and T. Dieck. Representations of Compact Lie Groups. Graduate Texts in Mathe-
matics. Springer, 1985.
[16] B. Green. A Szemer´edi-type regularity lemma in abelian groups, with applications. Geometric
and Functional Analysis, 15(2):340–376, 2005.
[17] H. Hatami and S. Lovett. Estimating the distance from testable affine-invariant properties.
In Proceedings of the 54th Annual IEEE Symposium on Foundations of Computer Science
(FOCS), pages 237–242, 2013.
[18] J. Huang, C. Guestrin, and L. Guibas. Fourier theoretic probabilistic inference over permuta-
tions. The Journal of Machine Learning Research, 10:997–1070, 2009.
[19] M. Kassabov. Symmetric groups and expander graphs. Inventiones Mathematicae, 170(2):327–
354, 2007.
[20] T. Kaufman and M. Sudan. Algebraic property testing: the role of invariance. In Proceedings
of the 40th Annual ACM Symposium on Theory of Computing (STOC), pages 403–412, 2008.
[21] S. V. Kerov. Talk at IHP conference. 2000.
[22] A. Kirillov. Introduction to Lie groups and Lie algebras. Cambridge University Press, 2008.
[23] R. O’Donnell and J. Wright. Quantum spectrum testing. In Proceedings of the 47th ACM
Symposium on Theory of Computing (STOC), 2015.
[24] R. Raz and B. Spieker. On the “log rank”-conjecture in communication complexity. Combi-
natorica, 15(4):567–588, 1995.
[25] D. Ron. Algorithmic and analysis techniques in property testing. Foundations and Trends R(cid:13)
in Theoretical Computer Science, 5(2):73–205, 2009.
[26] R. Rubinfeld. On the robustness of functional equations. SIAM Journal on Computing,
28(6):1972–1997, 2006.
[27] R. Rubinfeld and M. Sudan. Robust characterizations of polynomials with applications to
program testing. SIAM Journal on Computing, 25(2):252–271, 1996.
[28] A. Shpilka and A. Wigderson. Derandomizing homomorphism testing in general groups. SIAM
Journal on Computing, 36(4):1215–1230, 2006.
[29] R. P. Stanley.
Irreducible symmetric group characters of rectangular shape. S´eminaire
Lotharingien de Combinatoire, 50, 2003.
[30] B. Steinberg. Representation Theory of Finite Groups: An Introductory Approach. Springer,
2011.
[31] H. Weyl. Quantenmechanik und gruppentheorie. Zeitschrift fur Physik, 46(1-2):1–46, 1927.
[32] K. Wimmer. Agnostically learning under permutation invariant distributions. In Proceedings
of the 51st Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages
113–122, 2010.
19
[33] Y. Yoshida. A characterization of locally testable affine-invariant properties via decomposi-
tion theorems. In Proceedings of the 46th Annual ACM Symposium on Theory of Computing
(STOC), pages 154–163, 2014.
20
|
1411.0960 | 2 | 1411 | 2015-01-14T12:22:14 | Fully Dynamic Bin Packing Revisited | [
"cs.DS"
] | We consider the fully dynamic bin packing problem, where items arrive and depart in an online fashion and repacking of previously packed items is allowed. The goal is, of course, to minimize both the number of bins used as well as the amount of repacking. A recently introduced way of measuring the repacking costs at each timestep is the migration factor, defined as the total size of repacked items divided by the size of an arriving or departing item. Concerning the trade-off between number of bins and migration factor, if we wish to achieve an asymptotic competitive ration of $1 + \epsilon$ for the number of bins, a relatively simple argument proves a lower bound of $\Omega(\frac{1}{\epsilon})$ for the migration factor. We establish a nearly matching upper bound of $O(\frac{1}{\epsilon}^4 \log \frac{1}{\epsilon})$ using a new dynamic rounding technique and new ideas to handle small items in a dynamic setting such that no amortization is needed. The running time of our algorithm is polynomial in the number of items $n$ and in $\frac{1}{\epsilon}$. The previous best trade-off was for an asymptotic competitive ratio of $\frac{5}{4}$ for the bins (rather than $1+\epsilon$) and needed an amortized number of $O(\log n)$ repackings (while in our scheme the number of repackings is independent of $n$ and non-amortized). | cs.DS | cs |
Fully Dynamic Bin Packing Revisited∗
Sebastian Berndt1, Klaus Jansen2 and Kim-Manuel Klein2
1Institute for Theoretical Computer Science, Universitat zu Lubeck,
2Department of Computer Science, Christian-Albrechts-University to
[email protected]
Kiel, {kj,kmk}@informatik.uni-kiel.de
We consider the fully dynamic bin packing problem, where items arrive and depart
in an online fashion and repacking of previously packed items is allowed. The goal
is, of course, to minimize both the number of bins used as well as the amount
of repacking. A recently introduced way of measuring the repacking costs at each
timestep is the migration factor, defined as the total size of repacked items divided by
the size of an arriving or departing item. Concerning the trade-off between number
of bins and migration factor, if we wish to achieve an asymptotic competitive ration
of 1 + ǫ for the number of bins, a relatively simple argument proves a lower bound
of Ω(1/ǫ) for the migration factor. We establish a nearly matching upper bound of
O(1/ǫ4 log 1/ǫ) using a new dynamic rounding technique and new ideas to handle small
items in a dynamic setting such that no amortization is needed. The running time
of our algorithm is polynomial in the number of items n and in 1/ǫ. The previous
best trade-off was for an asymptotic competitive ratio of 5/4 for the bins (rather than
1 + ǫ) and needed an amortized number of O(log n) repackings (while in our scheme
the number of repackings is independent of n and non-amortized).
1 Introduction
For the classical bin packing problem, we are given a set I of items with a size function s : I →
(0, 1] and need to pack them into as few unit sized bins as possible. In practice, the complete
instance is often not known in advance, which has lead to the definition of a variety of online
versions of the bin packing problem. First, in the classical online bin packing [Ull71], items
arrive over time and have to be packed on arrival. Second, in dynamic bin packing [CGJ83],
items may also depart over time. This dynamic bin packing model is often used for instance in
• the placement and movement of virtual machines onto different servers for cloud computing
• the development of guaranteed quality of service channels over certain multi-frequency
[BB10, BKB07, SKZ08, VAN08, JJH+08, JJH+09],
time division multiple access systems [PSC+00],
∗Supported by DFG Project, Entwicklung und Analyse von effizienten polynomiellen Approximationsschemata
fur Scheduling- und verwandte Optimierungsprobleme, Ja 612/14-1.
1
[Sto13, SZ13],
ters [DKL14, LTC14].
• the placement of processes, which require different resources, onto physical host machines
• the resource allocation in a cloud network where the cost depends upon different parame-
Third and fourth, we may allow already packed items to be slightly rearranged, leading to online
bin packing with repacking (known as relaxed online bin packing) [GPT00] and dynamic bin
packing with repacking (known as fully dynamic bin packing) [IL98]. See Figure 1 for a short
overview on the different models.
Deletion Repacking
Name
Online Bin Packing
Relaxed Online Bin Packing
Dynamic Bin Packing
Fully Dynamic Bin Packing
Figure 1: Overview of online models
The amount of repacking can be measured in different ways. We can either count the total
number of moved items at each timestep or the sum of the sizes of the moved items at each
timestep. If one wants to count the number of moved items, one typically counts a group of
tiny items as a single move. A shifting move [GPT00] thus involves either a single large item
or a bundle of small items in the same bin of total size s with 1/10 ≤ s ≤ 1/5. Such a bundle
may consists of up to Ω(n) (very small) items.
If an algorithm measures the repacking by
shifting moves, a new tiny item may lead to a large amount of repacking. In order to guarantee
that a tiny item i with size s(i) only leads to a small amount of repacking, one may allow to
repack items whose size adds up to at most β · s(i). The term β is called the migration factor
[SSS09]. Note that shifting moves and migration factor are incomparable in the sense that a
small migration factor does not imply a small number of shifting moves and vice versa.
In order to measure the quality of an online algorithm, we compare the costs incurred by an
online algorithm with the costs incurred by an optimal offline algorithm. An online algorithm
receives as input a sequence of items I = (i1, i2, i3, . . .) and decides at each timestep t, where to
place the item it without knowing future items it+1, it+2, . . .. We denote by I(t) = (i1, i2, . . . , it)
the instance containing the first t items of the instance I and by opt(I(t)) the minimal number
of bins needed to pack all items in I(t). Note that the packings corresponding to opt(I(t))
and opt(I(t + 1)) may differ significantly, as those packings do not need to be consistent. For
an online algorithm A, we denote by A(I(t)) the number of bins generated by the algorithm
on the input sequence I(t). Note that A must make its decision online, while opt(I(t)) is the
optimal value of the offline instance. The quality of an algorithm for the online bin packing
problem is typically measured by its asymptotic competitive ratio. An online algorithm A
is called an asymptotic α-competitive algorithm, if there is a function f ∈ o(opt) such that
A(I(t)) ≤ α opt(I(t)) + f (I(t)) for all instances I and all t ≤ I. The minimum α such
that A is an asymptotic α-competitive algorithm is called the asymptotic competitive ratio
of A, denoted by ron
∞ (A) = min{α A is an asymptotic
α-competitive algorithm}. The online algorithm A thus has a double disadvantage: It does
not know future items and we compare its quality to the optimal offline algorithm which may
produce arbitrary different packings at time t and time t + 1. In order to remedy this situation,
one may also compare the solution generated by A to a non-repacking optimal offline algorithm.
This non-repacking optimal offline algorithm knows the complete instance, but is not allowed
to repack.
∞ (A), i. e., the ratio is defined as ron
2
✗
✗
✗
✓
✓
✗
✓
✓
In this work, we present new results in fully dynamic bin packing where we measure the
quality of an algorithm against a repacking optimal offline algorithm and achieve a asymptotic
competitive ratio of 1 + ǫ. The amount of repacking is bounded by O(1/ǫ4 log(1/ǫ)). While
we measure the amount of repacking in terms of the migration factor, we also prove that our
algorithm uses at most O(1/ǫ4 log(1/ǫ)) shifting moves. Our algorithm runs in time polynomial
in the instance size and in 1/ǫ.
1.1 Previous Results on Online Variants of Bin Packing
Online Bin Packing
The classical version of online bin packing problem was introduced by Ullman [Ull71]. In this
classical model items arrive over time and have to be packed at their arrival, while one is not
allowed to repack already packed items. Ullman gave the very first online algorithm FirstFit
for the problem and proved that it its absolute competitive ratio is at most 2. The next
algorithm NextFit was given by Johnson [Joh74], who proved that its absolute competitive
is also at most 2. The analysis of the FirstFit algorithm was refined by Johnson, Demers,
Ullman, Garey and Graham [JDU+74a], who proved that its asymptotic competitive ratio is
at most 17/10. A revised version of FirstFit, called Revised FirstFit was shown to have
asymptotic competitive ratio of at most 5/3 by Yao [Yao80]. A series of developments of so called
harmonic algorithms for this problem was started by Lee and Lee [LL85] and the best known
algorithm of this class which has asymptotic competitive ratio at most 1.58889 was given by
Seiden [Sei02]. The lower bound on the absolute approximation ratio of 3/2 also holds for the
asymptotic competitive ratio as shown by Yao [Yao80]. This lower bound was first improved
independently by Brown [Bro79] and Liang [Lia80] to 1.53635 and subsequently to 1.54014 by
van Vliet [Vli92] and finally to 1.54037 by Balogh, B´ek´esi and Galambos [BBG10].
Relaxed Online Bin Packing Model
In contrast to the classical online bin packing problem, Gambosi, Postiglione and Talamo
[GPT00] considered the online case where one is allowed to repack items. They called this
model the relaxed online bin packing model and proved that the lower bound on the competi-
tive ratio in the classical online bin packing model can be beaten. They presented an algorithm
that uses 3 shifting moves and has an asymptotic competitive ratio of at most 3/2, and an al-
gorithm that uses at most 7 shifting moves and has an asymptotic competitive ratio of 4/3. In
another work, Ivkovi´c and Lloyd [IL97] gave an algorithm that uses O(log n) amortized shifting
moves and achieves an asymptotic competitive ratio of 1 + ǫ. In this amortized setting, shifting
moves can be saved up for later use and the algorithm may repack the whole instance sometimes.
Epstein and Levin [EL09] used the measure of the migration factor to give an algorithm that
has an asymptotic competitive ratio of 1 + ǫ and a migration factor of 2O((1/ǫ) log2(1/ǫ)). This
result was improved by Jansen and Klein [JK13] who achieved polynomial migration. Their
algorithm uses a migration factor of O(1/ǫ4) to achieve an asymptotic competitive ratio of 1 + ǫ.
Concerning lower bounds on the migration factor, Epstein and Levin [EL09] showed that no
optimal solution can be maintained while having a constant migration factor (independent of
1/ǫ). Furthermore, Balogh, B´ek´esi, Galambos and Reinelt [BBGR08] proved that a lower bound
on the asymptotic competitive ratio of 1.3877 holds, if the amount of repacking is measured by
the number of items and one is only allowed to repack a constant number of items.
3
Dynamic Bin Packing
An extension to the classical online bin packing model was given by Coffman, Garey and Johnson
[CGJ83], called the dynamic bin packing model. In addition to the insertion of items, items
also depart over time. No repacking is allowed in this model. It is easily seen that no algorithm
can achieve a constant asymptotic competitive ratio in this setting. In order to measure the
performance of an online algorithm A in this case, they compared the maximum number of
bins used by A with the maximum number of bins used by an optimal offline algorithm, i. e.,
an algorithm A in this dynamic model is called an asymptotic α-competitive algorithm, if there
is a function f ∈ o(max-opt), where max-opt(I) = maxt opt(I(t)) such that maxt A(I(t)) ≤
α· maxt opt(I(t)) + f (I) for all instances I. The minimum of all such α is called the asymptotic
competitive ratio of A. Coffman, Garey and Johnson modified the FirstFit algorithm and
proved that its asymptotic competitive ratio is at most 2.897. Furthermore, they showed a
lower bound of 2.5 on the asymptotic competitive ratio when the performance of the algorithm
is compared to a repacking optimal offline algorith, i. e., maxt opt(I(t)).
In the case that the performance of the algorithm is compared to an optimal non-repacking
offline algorithm, Coffman, Garey and Johnson showed a lower bound of 2.388. This lower
bound on the non-repacking optimum was later improved by Chan, Lam and Wong [CLW08]
to 2.428 and even further in a later work by Chan, Wong and Yung [CWY09] to 2.5.
Fully Dynamic Bin Packing
We consider the dynamic bin packing when repacking of already packed items is allowed. This
model was first investigated by Ivkovi´c and Lloyd [IL98] and is called fully dynamic bin packing.
In this model, items arrive and depart in an online fashion and limited repacking is allowed.
The quality of an algorithm is measured by the asymptotic competitive ratio as defined in the
classical online model (no maximum is taken as in the dynamic bin packing model). Ivkovi´c and
Lloyd developed an algorithm that uses amortized O(log n) many shifting moves (see definition
above) to achieve an asymptotic competitive ratio of 5/4.
Related Results on the Migration Factor
Since the introduction of the migration factor, several problems were considered in this model
and different robust algorithms for these problems have been developed. Following the termi-
nology of Sanders, Sivadasan and Skutella [SSS09] we sometimes use the term (online) approx-
imation ratio instead of competitive ratio. Hence, we also use the term asymptotic polynomial
time approximation scheme (APTAS) and asymptotic fully polynomial time approximation
scheme (AFPTAS) in the context of online algorithms. If the migration factor of an algorithm
A only depends upon the approximation ratio ǫ and not on the size of the instance, we say that
A is an robust algorithm.
In the case of online bin packing, Epstein and Levin [EL09] developed the first robust APTAS
for the problem using a migration factor of 2O((1/ǫ2) log(1/ǫ)). They also proved that there is no
online algorithm for this problem that has a constant migration factor and that maintains an
optimal solution. The APTAS by Epstein and Levin was later improved by Jansen and Klein
[JK13], who developed a robust AFPTAS for the problem with migration factor O(1/ǫ4). In their
paper, they developed new linear program (LP)/integer linear program (ILP) techniques, which
we make use of to obtain polynomial migration. It was shown by Epstein and Levin [EL13] that
their APTAS for bin packing can be generalized to packing d-dimensional cubes into unit cubes.
Sanders, Sivadasan and Skutella [SSS09] developed a robust polynomial time approximation
scheme (PTAS) for the scheduling problem on identical machines with a migration factor of
4
2O((1/ǫ) log2(1/ǫ)). Skutella and Verschae [SV10] studied the problem of maximizing the minimum
load given n jobs and m identical machines. They also considered a dynamic setting, where jobs
may also depart. They showed that there is no robust PTAS for this machine covering problem
with constant migration. The main reason for the nonexistence is due to very small jobs. By
using an amortized migration factor, they developed a PTAS for the problem with amortized
migration of 2O((1/ǫ) log2(1/ǫ)).
1.2 Our Contributions
Main Result
In this work, we investigate the fully dynamic bin packing model. We measure the amount of
repacking by the migration factor; but our algorithm uses a bounded number of shifting moves
as well. Since the work of Ivkovi´c and Lloyd from 1998 [IL98], no progress was made on the fully
dynamic bin packing problem concerning the asymptotic competitive ratio of 5/4. It was also
unclear whether the number of shifting moves (respectively migration factor) must depend on
the number of packed items n. In this paper we give positive answers for both of these concerns.
We develop an algorithm that provides at each time step t an approximation guarantee of
(1 + ǫ) opt(I(t)) + O(1/ǫ log(1/ǫ)). The algorithm uses a migration factor of O(1/ǫ4 · log(1/ǫ)) by
repacking at most O(1/ǫ3 · log(1/ǫ)) bins. Hence, the generated solution can be arbitrarily close
to the optimum solution, and for every fixed ǫ the provided migration factor is constant (it
does not depend on the number of packed items). The running time is polynomial in n and
1/ǫ. In case that no deletions are used, the algorithm has a migration factor of O(1/ǫ3 · log(1/ǫ)),
which beats the best known migration factor of O(1/ǫ4) by Jansen and Klein [JK13]. Since the
number of repacked bins is bounded, so is the number of shifting moves as it requires at most
O(1/ǫ) shifting moves to repack a single bin. Furthermore, we prove that there is no asymptotic
approximation scheme for the online bin packing problem with a migration factor of o(1/ǫ) even
in the case that no items depart (and even if P = NP).
Technical Contributions
We use the following techniques to achieve our results:
• In order to obtain a lower bound on the migration factor in Section 2, we construct a
series of instances that provably need a migration factor of Ω(1/ǫ) in order to have an
asymptotic approximation ratio of 1 + ǫ.
• In Section 3, we show how to handle large items in a fully dynamic setting. The fully
dynamic setting involves more difficulties in the rounding procedure, in contrast to the
setting where large items may not depart, treated in [JK13]. A simple adaption of the
dynamic techniques developed in [JK13] does not work (see introduction of Section 3).
We modify the offline rounding technique by Karmarkar and Karp [KK82] such that a
feasible rounding structure can be maintained when items are inserted or removed. This
way, we can make use of the LP-techniques developed in Jansen and Klein [JK13].
• In Section 4, we explain how to deal with small items in a dynamic setting. In contrast
to the setting where departure of items is not allowed, the fully dynamic setting provides
major challenges in the treatment of small items. An approach is thus developed where
small items of similar size are packed near each other. We describe how this structure
can be maintained as new items arrive or depart. Note that the algorithm of Ivkovi´c and
Lloyd [IL98] relies on the ability to manipulate up to Ω(n) very small items in constant
time. See also their updated work for a thorough discussion of this issue [IL09].
5
2M
{z
}
2M (c+1)
{z
}
• In order to unify the different approaches for small and large items, in Section 4.2, we
develop an advanced structure for the packing. We give novel techniques and ideas to
manage this mixed setting of small and large items. The advanced structure makes use
of a potential function, which bounds the number of bins that need to be reserved for
incoming items.
2 Lower Bound
We start by showing that there is no robust (asymptotic) approximation scheme for bin packing
with migration factor of o(1/ǫ), even if P = NP. This improves the lower bound given by
Epstein and Levin [EL09], which states that no algorithm for bin packing, that maintains an
optimal solution can have a constant migration factor. Previously it was not clear whether
there exists a robust approximation algorithm for bin packing with sublinear migration factor
or even a constant migration factor.
Theorem 1. For a fixed migration factor γ > 0, there is no robust approximation algorithm
for bin packing with asymptotic approximation ratio better than 1 + 1
6⌈γ⌉+5 .
Proof. Let A be an approximation algorithm with migration factor γ > 0 and c = ⌈γ⌉. We will
now construct an instance such that the asymptotic approximation ratio of A with migration
factor c is at least 1 + 1
6c+5 . The instance contains only two types of items: An A-item has size
a = 3/2
3c+2 and an B-item has size b = 1/2 − a/3. For a M ∈ N, let
IM = [(b, Insert), (b, Insert), . . . , (b, Insert)
]
, (a, Insert), (a, Insert), . . . , (a, Insert)
be the instance consisting of 2M insertions of B-items, followed by 2M (c + 1) insertions of
A-items. Denote by r(t) the approximation ratio of the algorithm at time t ∈ N. The approxi-
mation ratio of the algorithm is thus r = maxt{r(t)}.
The insertion of the B-items produces a packing with β1 bins containing a single B-item and
β2 bins containing two B-items. These are the only possible packings and hence β1 + 2β2 = 2M .
The optimal solution is reached if β1 = 0, β2 = M . We thus have an approximation ratio of
r(2M ) =: r1 =
β1 + β2
M
=
2M − β2
M
,
which is strictly monotonically decreasing in β2.
The A-items, which are inserted afterwards, may either be put into bins which only contain
A-items or into bins which contain only one B-item. The choice of a, b implies 2 · b + a > 1
which shows that no A-item can be put into a bin containing two B-items. Denote by α the
number of bins containing only A-items. The existing B-items may not be moved as the choice
of a, b implies b > c· a > γ · a. At most 1/2+a/3
a = c + 1 items of type A may be put into the bins
containing only one B-item. Note that this also implies that a bin which contains one B-item
and c + 1 items of type A is filled completely. The optimal packing thus consists of 2M of those
bins and the approximation ratio of the solution is given by
r(2M (c + 2)) =: r2 =
β1 + β2 + α
2M
=
2M − 2β2 + β2 + α
2M
=
2M − β2 + α
2M
.
There are at most β1 · (c + 1) items of type A which can be put into bins containing only one
B-item. The remaining (2M − β1)(c + 1) items of type A therefore need to be put into bins
6
containing only A-items. We can thus conclude α ≥ (2M − β1)(c + 1)a = (2M − 2M + 2β2)(c +
1)a = 2β2(c + 1)a. As noted above, 1/2+a/3
a = c + 1 and thus (c + 1)a = 1/2 + a/3. Hence the
approximation ratio is at least
β1 + β2 + α
r2 =
2M + β2(−1 + 1 + 2a/3)
2M
≥
2M
2M − β2 + 2β2(1/2 + a/3)
=
2M
=
2M + β2 · 2a/3
2M
,
which is strictly monotonically increasing in β2.
As r ≥ max{r1, r2}, a lower bound on the approximation ratio is thus given if r1 = r2 by
M = 2M +β·2a/3
a/3+1 . The lower bound
for a certain β. Solving this equation leads to β = M
2M −β
2M
is thus given as
2M − β
M
r ≥
= 2 −
1
a/3 + 1
= 1 +
1
6c + 5
by the choice of a. Note that this lower bound is independent from M . Hence, r is also a lower
bound on the asymptotic approximation ratio of any algorithm as the instance size grows with
M .
We obtain the following corollary:
Corollary 1. There is no robust/dynamic (asymptotic) approximation scheme for bin packing
with a migration factor γ ≤ 1/6(1/ǫ − 11) = Θ(1/ǫ).
3 Dynamic Rounding
In the first subsection we present a general rounding structure.
The goal of this section is to give a robust AFPTAS for the case that only large items arrive
and depart.
In the second
subsection we give operations on how the rounding can be modified such that the general
structure is preserved. We give the final algorithm in Section 3.3, which is performed, when
large items arrive or depart. Finally, the correctness is proved by using the LP/ILP techniques
developed in [JK13].
In [JK13], the last two authors developed a dynamic rounding technique based on an offline
rounding technique from Fernandez de la Vega and Lueker [FdlVL81]. However, a simple
adaption of these techniques does not work in the dynamic case where items may also depart.
In the case of the offline rounding by Fernandez de la Vega and Lueker, items are sorted and
then collected in groups of the same cardinality. As a new item arrives in an online fashion, this
structure can be maintained by inserting the new item to its corresponding group. By shifting
the largest item of each group to the left, the cardinality of each group (except for the first
one) can be maintained. However, shifting items to the right whenever an item departs leads
to difficulties in the LP/ILP techniques. As the rounding for a group may increase, patterns
of the existing LP/ILP solution might become infeasible. We overcome these difficulties by
developing a new dynamic rounding structure and operations based on the offline rounding
technique by Karmarkar and Karp [KK82]. We felt that the dynamic rounding technique based
on Karmarkar and Karp is easier to analyze since the structure can essentially be maintained
by shifting items.
A bin packing instance consists of a set of items I = {i1, i2, . . . , in} with size function s :
I → [0, 1] ∩ Q. A feasible solution is a partition B1, . . . , Bk of I such that Pi∈Bj s(i) ≤ 1 for
j = 1, . . . , k. We call a partition B1, . . . , Bk a packing and a single set Bj is called a bin. The
7
goal is to find a solution with a minimal number of bins. If the item i is packed into the bin
Bj, we write B(i) = j. The smallest value of k ∈ N such that a packing with k bins exists is
denoted by opt(I, s) or if the size function is clear by opt(I). A trivial lower bound is given
by the value size(I, s) =Pi∈I s(i).
3.1 Rounding
To obtain an LP formulation of fixed (independent of I) dimension, we use a rounding technique
based on the offline AFPTAS by Karmarkar and Karp [KK82]. In order to use the technique
for our dynamic setting, we give a more general rounding. This generalized rounding has a
certain structure that is maintained throughout the algorithm and guarantees an approximate
solution for the original instance. First, we divide the set of items into small ones and large
ones. An item i is called small if s(i) < ǫ/14, otherwise it is called large. Instance I is partitioned
accordingly into a set of large items IL and a set of small items IS. We treat small items and
large items differently. Small items can be packed using an algorithm presented in Section 4.1
while large items will be assigned using an ILP. In this section we discuss how to handle large
items.
First, we characterize the set of large items more precisely by their sizes. We say that two
large items i, i′ are in the same size category if there is a ℓ ∈ N such that s(i) ∈ (2−(ℓ+1), 2−ℓ] and
s(i′) ∈ (2−(ℓ+1), 2−ℓ]. Denote the set of all size categories by W . As every large item has size at
least ǫ/14, the number of size categories is bounded by log(1/ǫ) + 5. Next, items of the same size
category are characterized by their block, which is either A or B and their position r ∈ N in this
block. Therefore, we partition the set of large items into a set of groups G ⊆ W × {A, B} × N.
A group g ∈ G consists of a triple (ℓ, X, r) with size category ℓ ∈ W , block X ∈ {A, B} and
position r ∈ N. The rounding function is defined as a function R : IL 7→ G that maps each
large item i ∈ IL to a group g ∈ G. By g[R] we denote the set of items being mapped to the
group g, i. e., g[R] = {i ∈ IL R(i) = g}.
Let q(ℓ, X) be the maximal r ∈ N such that (ℓ, X, r)[R] > 0. If (ℓ, X1, r1) and (ℓ, X2, r2)
are two different groups, we say that (ℓ, X1, r1) is left of (ℓ, X2, r1), if X1 = A and X2 = B or
X1 = X2 and r1 < r2. We say that (ℓ, X1, r1) is right of (ℓ, X2, r2) if it is not left of it.
(ℓ, A, 0)
. . .
(ℓ, A, q(ℓ, A))
(ℓ, B, 0)
. . .
(ℓ, B, q(ℓ, B))
s ∈ [2−(ℓ+1), 2−ℓ)
Figure 2: Grouping in (ℓ, A,·) and (ℓ, B,·)
Given an instance (I, s) and a rounding function R, we define the rounded size function sR
by rounding the size of every large item i ∈ g[R] up to the size of the largest item in its group,
hence sR(i) = max {s(i′) R(i′) = R(i)}. We denote by opt(I, sR) the value of an optimal
solution of the rounded instance (I, sR).
Depending on a parameter k, we define the following properties for a rounding function R.
(a) For each i ∈ (ℓ, X, r)[R] we have 2−(ℓ+1) < s(i) ≤ 2−ℓ.
(b) For each i ∈ (ℓ, X, r)[R] and each i′ ∈ (ℓ, X, r′)[R] and r < r′, we have s(i) ≥ s(i′).
(c) For each ℓ ∈ W and 1 ≤ r ≤ q(ℓ, A) we have (ℓ, A, r)[R] = 2ℓk and (ℓ, A, 0)[R] ≤ 2ℓk.
(d) For each ℓ ∈ W and each 0 ≤ r ≤ q(ℓ, B) − 1 we have (ℓ, B, r)[R] = 2ℓ(k − 1) and
Property (a) guarantees that the items are categorized correctly according to their sizes. Prop-
erty (b) guarantees that items of the same size category are sorted by their size and properties
(c) and (d) define the number of items in each group.
furthermore (ℓ, B, q(ℓ, B))[R] ≤ 2ℓ(k − 1).
8
size(IL)·ǫ
Lemma 1. For k = j
above by O(1/ǫ log(1/ǫ)) assuming that size(IL) > 8/ǫ · (⌈log(1/ǫ)⌉ + 5).
Proof. Using the definition of k and the assumption, we show 2 size(IL)
have
2(⌊log(1/ǫ)⌋+5)k the number of non-empty groups in G is bounded from
k−1 ≤ 8/ǫ(⌊log(1/ǫ)⌋ + 5). We
2 size(IL)
k − 1
=
2 size(IL)
size(IL)·ǫ
2(⌊log(1/ǫ)⌋+5)k − 1 ≤
j
2 size(IL)
2 size(IL)
size(IL)·ǫ
=
2(⌊log(1/ǫ)⌋+5) − 2
2 size(IL) · 2(⌊log(1/ǫ)⌋ + 5)
size(IL) · ǫ − 4(⌊log(1/ǫ)⌋ + 5)
=
size(IL)·ǫ−4(⌊log(1/ǫ)⌋+5)
2(⌊log(1/ǫ)⌋+5)
As size(IL) > 8/ǫ · (⌈log(1/ǫ)⌉ + 5), we have ǫ/2 size(IL) > 4(⌊log(1/ǫ)⌋ + 5). We can thus bound:
2 size(IL) · 2(⌊log(1/ǫ)⌋ + 5)
size(IL) · ǫ − 4(⌊log(1/ǫ)⌋ + 5) ≤
2 size(IL) · 2(⌊log(1/ǫ)⌋ + 5)
=
2 size(IL) · 2(⌊log(1/ǫ)⌋ + 5)
size(IL) · ǫ − ǫ/2 size(IL) + 1
=
4(⌊log(1/ǫ)⌋ + 5)
8(⌊log(1/ǫ)⌋ + 5)
=
ǫ/2
ǫ
size(IL) · ǫ/2
Note that property (c) and property (d) imply I(ℓ) ≥ (q(ℓ, A) + q(ℓ, B)− 2)2ℓ(k− 1) . Hence
property (a) implies that size(I(ℓ), s) ≥ I(ℓ)2−ℓ+1 ≥ (q(ℓ, A) + q(ℓ, B) − 2)(k − 1)/2 and
therefore q(ℓ, A) + q(ℓ, B) ≤ 2 size(I(ℓ))/(k − 1) + 2. We can now bound the total number of
used groups by
+ 2(cid:19)
k − 1
2
size(IL)
Xℓ∈W
= 2W +
q(ℓ, A) + q(ℓ, B) ≤ Xℓ∈W(cid:18) 2 size(I(ℓ))
k − 1
size(I(ℓ)) = 2W +
2
k − 1 Xℓ∈W
(⌊log(1/ǫ)⌋ + 5) ≤
8
ǫ
≤ 2W +
2 · (log(1/ǫ) + 5) +
(8/ǫ + 2)(log(1/ǫ) + 5) ∈ O(1/ǫ log(1/ǫ))
(log(1/ǫ) + 5) =
8
ǫ
The total number of used groups is therefore bounded by O(1/ǫ log(1/ǫ)).
The following lemma shows that the rounding function does in fact yield a (1+ǫ)-approximation.
Lemma 2. Given an instance (I, s) with items greater than ǫ/14 and a rounding function R
fulfilling properties (a) to (d), then opt(I, sR) ≤ (1 + ǫ)OPT (I, s).
Proof. As (I, s) only contains large items, IL = I. Define for every ℓ the instances Jℓ =
Sq(ℓ,A)
r=2 (ℓ, A, r)[R]∪Sq(ℓ,B)
(ℓ, B, r)[R], J =Sℓ∈W Jℓ and K =Sℓ∈W (ℓ, A, 0)[R]∪ (ℓ, A, 1)[R]. We
will now prove, that the error generated by this rounding is bounded by ǫ. As each solution to
J∪K yields a solution to J and a solution to K, we get opt(J∪K, sR) ≤ opt(J, sR)+opt(K, sR).
For i ∈ (ℓ, A, 0)[R] ∪ (ℓ, A, 1)[R], we have s(i) ≤ max {s(i′) i′ ∈ (ℓ, A, 0)[R]} ≤ 2−ℓ because of
property (a). We can therefore pack at least 2ℓ items from (ℓ, A, 0)[R]∪ (ℓ, A, 1)[R] into a single
bin. Hence, we get with property (c):
r=0
opt((ℓ, A, 0)[R] ∪ (ℓ, A, 1)[R]), sR)
≤ ((ℓ, A, 0)[R] + (ℓ, A, 1)[R]) · 2−ℓ
= 2k
9
We can therefore bound opt(K, sR) as follows:
opt((ℓ, A, 0)[R] ∪ (ℓ, A, 1)[R]), sR)
2k
opt(K, sR) ≤ Xℓ∈W
≤ Xℓ∈W
≤ 2(⌊log(1/ǫ)⌋ + 5)k
= 2⌊
size(I)ǫ
2(⌊log(1/ǫ)⌋ + 5)⌋ · (⌊log(1/ǫ)⌋ + 5)
2(⌊log(1/ǫ)⌋ + 5) · (⌊log(1/ǫ)⌋ + 5)
size(I)ǫ
≤ 2
= ǫ size(I)
≤ ǫ opt(I, s)
Using property (b) for each item in ((ℓ, X, r + 1)[R]), sR) we find a unique larger item in
(ℓ, X, r)[R]. Therefore we have for every item in the rounded instance (J, sR) an item with
larger size in instance (I, s) and hence
The optimal value of the rounded solution can be bounded by
opt(J, sR) ≤ opt(I, s).
opt(I, sR) ≤ opt(J, sR) + opt(K, sR) ≤ (1 + ǫ) opt(I, s).
We therefore have a rounding function, which generates only O(1/ǫ log(1/ǫ)) different item
sizes and the generated error is bounded by ǫ.
3.2 Rounding Operations
Let us consider the case where large items arrive and depart in an online fashion. Formally this
is described by a sequence of pairs (i1, A1), . . . , (in, An) where Ai ∈ {Insert, Delete}. At each
time t ∈ {1, . . . , n} we need to pack the item it into the corresponding packing of i1, . . . , it−1 if
Ai = Insert or remove the item it from the corresponding packing of i1, . . . , it−1 if Ai = Delete.
We will denote the instance i1, . . . , it at time t by I(t) and the corresponding packing by Bt.
We will also round our items and denote the rounding function at time t by Rt. The large items
of I(t) are denoted by IL(t). At time t we are allowed to repack several items with a total
size of β · s(it) but we intend to keep the migration factor β as small as possible. The term
repack(t) = Pi,Bt−1(i)6=Bt(i) s(i) denotes the sum of the items which are moved at time t, the
migration factor β of an algorithm is then defined as maxt {repack(t)/s(it)}. As the value of size
will also change over the time, we define the value κ(t) as
κ(t) =
size(IL(t)) · ǫ
2(⌊log(1/ǫ)⌋ + 5)
.
As shown in Lemma 1, we will make use of the value k(t) := ⌊κ(t)⌋.
We present operations that modify the current rounding Rt and packing Bt with its corre-
sponding LP/ILP solutions to give a solution for the new instance I(t + 1). At every time t the
rounding Rt maintains properties (a) to (d). Therefore the rounding provides an asymptotic
10
approximation ratio of 1 + ǫ (Lemma 2) while maintaining only O(1/ǫ log(1/ǫ)) many groups
(Lemma 1). We will now present a way how to adapt this rounding to a dynamic setting, where
items arrive or depart online.
Our rounding Rt is manipulated by different operations, called the insert, delete, shiftA and
shiftB operation. Some ideas behind the operations are inspired by Epstein and Levin [EL09].
The insert operation is performed whenever a large item arrives and the delete operation is
performed whenever a large item departs. The shiftA/shiftB operations are used to modify the
number of groups that are contained in the A and B block. As we often need to filter the largest
items of a group g belonging to a rounding R, we denote this item by λ(g, R).
• shift: A shift operation takes two groups (ℓ, X1, r1) and (ℓ, X2, r2), where (ℓ, X1, r1) is left
of (ℓ, X2, r2), and a rounding function R and produces a new rounding function R′ and
packing B′ by shifting the largest item from (ℓ, X2, r2) to (ℓ, X2, r2 − 1) and so on until
(ℓ, X1, r1) is reached.
– For all groups g left of (ℓ, X1, r1) or right of (ℓ, X2, r2) set g[R′] = g[R].
– As we move an items out of (ℓ, X2, r2), set
(ℓ, X2, r2)[R′] = (ℓ, X2, r2)[R] \ λ((ℓ, X2, r2), R).
– As we move an item into (ℓ, X1, r1), set
(ℓ, X1, r1)[R′] = (ℓ, X1, r1)[R] ∪ λ(right(ℓ, X1, r1), R).
Whenever a shift-operation on (ℓ, X1, r1) and (ℓ, X2, r2) is performed, the LP solution x
and the corresponding ILP solution y is updated to x′ and y′. Let Ci be a configuration
containing λ((ℓ, X2, r2), R) with xi ≥ 1. Let Cj = Ci \ s(λ((ℓ, X2, r2), R)) be the configu-
i = xi − 1, y′
ration without λ((ℓ, X2, r2), R). Set x′
i = yi − 1.
In order to add the new item in (ℓ, X1, r1), set x′
h = yh + 1 for the index
h with Ch = {1 : s(λ((ℓ, X1, r1), R))}. The remaining configurations do not change.
h = xh + 1 and y′
j = yj + 1 and x′
j = xj + 1, y′
. . .
(ℓ, X1, r1)
. . .
(ℓ, A, q(ℓ, A))
(ℓ, B, 0)
. . .
(ℓ, X2, r2)
. . .
Figure 3: shift with parameters (ℓ, X1, r1) and (ℓ, X2, r2)
• Insert: To insert item it, find the corresponding group (ℓ, X, r) with
– s(it) ∈ [ℓ, 2ℓ),
– min {s(i) i ∈ (ℓ, X, r − 1)} > s(it) and
– s(λ((ℓ, X, r + 1), R)) ≤ s(it).
We will then insert it into (ℓ, X, r) and get the rounding R′ by shifting the largest element
of (ℓ, X, r) to (ℓ, X, r − 1) and the largest item of (ℓ, X, r − 1) to (ℓ, X, r − 2) and so on
until (ℓ, A, 0) is reached. Formally, set R∗(it) = (ℓ, X, r) and R∗(ij) = R(ij) for j 6= t.
The rounding function R′ is then obtained by applying the shift operation on R∗ i.e. the
new rounding is R′ = shift((ℓ, A, 0), (ℓ, X, r), R∗).
In order to pack the new item, let i be the index with Ci = {1 : s(λ((ℓ, X, r), R′))}, as it
is rounded to the largest size in (ℓ, X, r)[R] after the shift. Place item it into a new bin
by setting B′(it) = maxj B(ij) + 1 and x′
i = xi + 1 and y′
i = yi + 1.
11
If (ℓ, A, 0)[R′] = 2ℓ·k+1, we have to create a new rounding group (ℓ, A,−1). Additionally
we shift the largest item in (ℓ, A, 0)[R′] to the new group (ℓ, A,−1)[R′]. The final rounding
R′′ is then obtained by setting (ℓ, A, r)[R′′] = (ℓ, A, r−1)[R′] i.e. incrementing the number
of each rounding group by 1. Note that the largest item in (ℓ, A, 0)[R′] is already packed
into a bin of its own due to the shift operation. Hence, no change in the packing or the
LP/ILP solution is needed. The insert operation thus yields a new packing B′ (or B′′)
which uses two more bins than the packing B.
(ℓ, A, 0)
. . .
(ℓ, X, r)
. . .
(ℓ, X, q(ℓ, X))
i
Figure 4: Insert i into (ℓ, X,·)
• Delete: To delete item it from the group (ℓ, X, r) with R(it) = (ℓ, X, r), we remove it from
this group and move the largest item from (ℓ, X, r + 1) into (ℓ, X, r) and the largest item
from (ℓ, X, r + 2) into (ℓ, X, r + 1) and so on until (ℓ, B, q(ℓ, B)). Formally the rounding
R′ is described by the expression shift((ℓ, X, r), (ℓ, B, q(ℓ, B)), R∗) where
g[R∗] =((ℓ, X, r)[R] \ {it} g = (ℓ, X, r)
g[R]
else
.
As a single shift operation is used, the delete operation yields a new packing B′ which
uses one more bin than the packing B.
For the LP/ ILP solution let Ci be a configuration containing λ((ℓ, B, q(ℓ, B)), R) with
xi ≥ 1. Let Cj = Ci s(λ((ℓ, B, q(ℓ, B)), R)) be the configuration without the item
λ((ℓ, B, q(ℓ, B)), R). Set x′
i = yi − 1. Set
B′(ij) = B(ij) for all j 6= t in order to remove the item it from the packing.
i = xi − 1, y′
j = yj + 1 and x′
j = xj + 1, y′
(ℓ, X, 0)
. . .
(ℓ, X, r)
. . .
(ℓ, B, q(ℓ, B))
i
Figure 5: Delete i from (ℓ, X,·)
To control the number of groups in A and B we introduce operations shiftA and shiftB that
increase or decrease the number of groups in A respectively B. An operation shiftA increases
the number of groups in A by 1 and decreases the number of groups in B by 1. Operations
shiftB is doing the inverse of shiftA.
• shiftA: In order to move a group from B to A we will perform exactly 2ℓ times the
operation shift((ℓ, B, 0), (ℓ, B, q(ℓ, B)), R) to receive the rounding R∗. Instead of opening
a new bin for each of those 2ℓ items in every shift operation, we rather open one bin
containing all items. Since every item in the corresponding size category has size ≤ 2−ℓ,
the items fit into a single bin. The group (ℓ, B, 0) has now the same size as the groups in
(ℓ, A,·). We transfer (ℓ, B, 0) to block A. Hence we define for the final rounding R′ that
12
(ℓ, A, r)[R′] = (ℓ, A, r)[R∗] for r = 0, . . . , q(ℓ, A) and (ℓ, A, q(ℓ, A) + 1)[R′] = (ℓ, B, 0)[R∗]
as well as (ℓ, B, r)[R′] = (ℓ, B, r + 1)[R∗] for r = 0, . . . , q(ℓ, B) − 1. The resulting packing
B′ hence uses one more bin than the packing B.
2ℓ
2ℓ
2ℓ
2ℓ
(ℓ, B, 0)
. . .
(ℓ, B, r)
. . .
(ℓ, B, q(ℓ, B))
Figure 6: shiftA
• shiftB: In order to move a group from A to B we will perform exactly 2ℓ times the operation
shift((ℓ, A, 0), (ℓ, A, q(ℓ, A)), R) to receive the rounding R∗. As before in shiftA, we open a
single bin containing all of the 2ℓ items. The group (ℓ, A, q(ℓ, A)) has now the same size as
the groups in (ℓ, B,·). We transfer (ℓ, A, q(ℓ, A)) to block B. Similar to shiftA we define
for the final rounding R′ that (ℓ, A, r)[R′] = (ℓ, A, r)[R∗] for r = 0, . . . , q(ℓ, A) − 1 and
(ℓ, B, 0)[R′] = (ℓ, A, q(ℓ, A))[R∗] as well as (ℓ, B, r + 1)[R′] = (ℓ, B, r)[R∗]. The resulting
packing B′ hence uses one more bin than the packing B.
Lemma 3. Let R be a rounding function fulfilling properties (a) to (d). Applying one of
the operations insert, delete, shiftA or shiftB on R results in a rounding function R′ fulfilling
properties (a) to (d).
Proof. Property (a) is always fulfilled as no item is moved between different size categories and
the insert operation inserts an item into its appropriate size category.
As the order of items never changes and the insert operation inserts an item into the appro-
priate place, property (b) also holds.
For properties (c) and (d) we first note that the operation shift(g, g′, R) increases the number
of items in g by 1 and decreases the number of items in g′ by 1. The insert operation consists of
adding a new item to a group g followed by a shift((ℓ, A, 0), g, R) operation. Hence the number of
items in every group except for (ℓ, A, 0) (which is increased by 1) remains the same. The delete
operation consists of removing an item from a group g followed by a shift(g, (ℓ, B, q(ℓ, B)), R)
operation. Therefore the number of items in all groups except for (ℓ, B, q(ℓ, B)) (which is
decreased by 1) remains the same. As the number of items in (ℓ, A, 0) and (ℓ, B, q(ℓ, B)) are
treated seperately and may be smaller than 2ℓ · k respectively 2ℓ · (k − 1), the properties (c)
and (d) are always fulfilled for the insert and the delete operation. Concerning the shiftA
operation we increase the number of items in a group (ℓ, B, 0) by 2ℓ. Therefore it now contains
2ℓ(k − 1) + 2ℓ = 2ℓ · k items, which equals the number of items in groups of block A. As this
group is now moved to block A, the properties (c) and (d) are fulfilled. Symmetrically the
shiftB operation decreases the number of items in a group (ℓ, A, q(ℓ, A)) by 2ℓ. Therefore the
number of items in the group is now 2ℓ · k − 2ℓ = 2ℓ · (k − 1), which equals the number of items
in the groups of block B. As this group is now moved to block B, the properties (c) and (d)
are fulfilled.
According to Lemma 1 the rounded instance (I, sR) has O(1/ǫ log(1/ǫ)) different item sizes
(given a suitable k). Using the LP formulation of Eisemann [Eis57], the resulting LP called
LP (I, sR) has m = O(1/ǫ log(1/ǫ)) constraints. We say a packing B corresponds to a rounding
R and an integral solution y of the ILP if all items in (I, sR) are packed by B according to y.
13
Lemma 4. Applying any of the operations insert, delete, shiftA or shiftB on a rounding function
R and ILP solution y with corresponding packing B defines a new rounding function R′ and a
new integral solution y′. Solution y′ is a feasible solution of LP (I, sR′
).
) changes in comparison to the LP
Proof. We have to analyze how the LP for instance (I, sR′
for instance (I, sR).
Shift Operation: A single shift(g1, g2, R) operation moves one item from each group g be-
tween g1 and g2 into g and one item out of g. As no item is moved out of g1 and no item is
moved into g2, the number of items in g1 is increased by 1 and the number of items in g2 is
decreased by 1. The right hand side of the LP (I, sR) is defined by the cardinalities g[R] of
the rounding groups g in R. As only the cardinalities of g1 and g2 change by ±1 the right hand
side changes accordingly to ±1 in the corresponding components of y. The moved item from
g2 is removed from the configuration and a new configuration containing the new item of g1 is
added. The LP and ILP solutions x and y are being modified such that λ(g2, R) is removed
from its configuration and a new configuration is added such that the enhanced right hand side
of g1 is covered. Since the largest item λ(g, R) of every group g between g1 and g2 is shifted
(i) of item i ∈ g[R] is defined by sR′
to its left group, the size sR′
(i) = s(ι(g, R)), where ι(g, R)
is the second largest item of g[R]. Therefore each item in (I, sR′
) is rounded to a smaller or
equal value as s(ι(g, R)) ≤ s(λ(g, R)). All configurations of (I, sR) can thus be transformed
into feasible configurations of (I, sR′
Insert Operation: The insert operation consists of inserting the new item into its correspond-
ing group g followed by a shift operation.
Inserting the new item into g increases the right
hand side of the LP by 1. To cover the increased right hand side, we add a new configuration
{1 : sR′
(i)} containing only the new item. In order to reflect the change in the LP solution, the
new item is added into an additional bin. The remaining changes are due to the shift operation
already treated above.
Delete Operation: The delete operation consists of removing an item i from its corresponding
group g followed by a shift operation. Removing the new item from g decreases the right hand
side of the LP by 1. The current LP and ILP solutions x and y do not need to be changed to
cover the new right hand side. The remaining changes are due to the shift operation already
treated above.
shiftA/shiftB Operation: As the shiftA and shiftB operations consist only of repeated use
of the shift operation, the correspondence between the packing and the LP/ILP solution follow
simply by induction.
).
3.3 Algorithm for Dynamic Bin Packing
parameter k = ⌊κ⌋ = j
size(IL)·ǫ
We will use the operations from the previous section to obtain a dynamic algorithm for bin
packing with respect to large items. The operations insert and delete are designed to process
the input depending of whether an item is to be inserted or removed. Keep in mind that the
2(⌊log(1/ǫ)⌋+5)k changes over time as size(IL) may increase or decrease.
In order to fulfill the properties (c) and (d), we need to adapt the number of items per group
whenever k changes. The shiftA and shiftB operations are thus designed to manage the dynamic
number of items in the groups as k changes. Note that a group in the A-block with parameter
k has by definition the same number of items as a group in the B-block with parameter k − 1
assuming they are in the same size category. If k increases, the former A block is treated as
the new B block in order to fulfill the properties (c) and (d) while a new empty A block is
introduced. To be able to rename the blocks, the B block needs to be empty. Accordingly the
A block needs to be empty if k decreases in order to treat the old B block as new A block.
14
Hence we need to make sure that there are no groups in the B-block if k increases and vice
versa, that there are no groups in the A-block if k decreases.
i
i+1
We denote the number of all groups in the A-blocks at time t by A(t) and the number of
groups in B-blocks at time t by B(t). To make sure that the B-block (respectively the A-block)
A(t)+B(t) needs to correlate to the fractional
is empty when k increases (decreases) the ratio
digits of κ(t) at time t denoted by ∆(t). Hence we partition the interval [0, 1) into exactly
A(t)
A(t)
A(t)+B(t)(cid:17). We will make sure that ∆(t) ∈ Ji iff
A(t) + B(t) smaller intervals Ji = h
A(t)+B(t) ,
A(t)+B(t) ∈ Ji. Note that the term A(t)
A(t)+B(t) is 0 if the A-block is empty and the term is 1 if
the B-block is empty. This way, we can make sure that as soon as k(t) increases, the number
of B-blocks is close to 0 and as soon as k(t) decreases, the number of A-blocks is close to 0.
Therefore, the A, B-block can be renamed whenever k(t) changes. The algorithm uses shiftA
and shiftB operations to adjust the number of A- and B-blocks. Recall that a shiftA operation
reduces the number of groups in the B-block by 1 and increases the number of groups in the
A-block by 1 (shiftB works vice versa). Let d be the number of shiftA/shiftB operations that
need to be performed to adjust
A(t)
A(t)+B(t) .
J0
J1
. . .
Jj
. . .
k(t − 1)
k(t − 1) + 1
∆(t − 1)
k(t − 1) + 2
B(t − 1)
2%
. . .
. . .
. . .
k(t) + 1
k
k(t − 1) + 2
A(t − 1)
98%
(a) Before Insert
Jj
k
J1
J0
k(t)
k
∆(t)
k(t − 1) + 1
B(t)
99%
(b) After Insert
k(t) − 1
k(t − 1)
k
A(t)
1%
Figure 7: Comparison of the situation before and after an Insert Operation
In the following algorithm we make use of an algorithm called improve, which was developed
in [JK13] to reduce the number of used bins. Using improve(x) on a packing B with approx-
imation guarantee maxi B(i) ≤ (1 + ¯ǫ) opt +C for some ¯ǫ = O(ǫ) and some additive term C
yields a new packing B′ with approximation guarantee maxi B(i) ≤ (1 + ¯ǫ) opt +C − x. We
use the operations in combination with the improve algorithm to obtain a fixed approximation
guarantee.
15
Algorithm 1 (AFPTAS for large items).
Algorithm: Insertion
if SIZE(I(t)) < (m + 2)(1/δ + 2) or SIZE(I(t)) < 8(1/δ + 1) then
use offline Bin Packing
else
A(t)
improve(2); insert(i);
// Shifting to the correct interval
Let Ji be the interval containing ∆(t);
Let Jj be the interval containing
Set d = i − j;
if k(t) > k(t − 1) then // Modulo A(t) + B(t) when k increases
// Shifting d groups from B to A
for p := 0 to d − 1 do
Rename(A, B);
if i+p = A(t) + B(t) then
d = d + (A(t) + B(t));
A(t)+B(t) ;
improve(1); shiftA;
Algorithm: Deletion
if SIZE(I(t)) < (m + 2)(1/δ + 2) or SIZE(I(t)) < 8(1/δ + 1) then
use offline Bin Packing
else
// Departing item i
improve(4); delete(i);
ReduceComponents;
//
// Shifting to the correct interval
Let Ji be the interval containing ∆(t);
Let Jj be the interval containing
Set d = i − j;
if k(t) < k(t − 1) then // Modulo A(t) + B(t) when k decreases
// Shifting d groups from A to B
for p := 0 to d − 1 do
Rename(A,B);
d = d - (A(t)+B(t));
if i-p = 0 then
A(t)+B(t) ;
A(t)
improve(3); shiftB;
Note that as exactly d groups are shifted from A to B (or B to A) we have by definition that
∆(t) ∈h
A(t)
A(t)+B(t) , A(t)+1
A(t)+B(t)(cid:17) at the end of the algorithm. Note that d can be bounded by 11.
Lemma 5. At most 11 groups are shifted from A to B (or B to A) in Algorithm 1.
ǫ
Proof. Since the value size(I(t − 1)) − size(I(t)) changes at most by 1 we can bound κ(t −
1) − κ(t) by
log(1/ǫ)+5 to obtain the change in the fractional part. By Lemma
1 the number of intervals (=the number of groups) is bounded by ( 8
ǫ + 2)(log(1/ǫ) + 5). Using
A(t−1)+B(t−1) ) and the fact that the number of groups A(t−1)+B(t−1)
∆(t−1) ∈ [
2(⌊log(1/ǫ)⌋+5) ≤
A(t−1)+B(t−1) ,
A(t−1)+1
A(t−1)
ǫ
16
increases or decreases at most by 1, we can give a bound for the parameter d in both cases by
D
ǫ
d ≤
(cid:18)(
interval length
8
ǫ
log(1/ǫ) + 5
) · (
+ 1 = D · #intervals + 1 ≤
+ 2) · (log(1/ǫ) + 5)(cid:19) + 1 =
8 + 2ǫ + 1 < 11
Hence, the number of shiftA/shiftB operations is bounded by 11.
Lemma 6. Every rounding function Rt produced by Algorithm 1 fulfills properties (a) to (d)
with parameter k(t) =j
size(IL)·ǫ
2(⌊log(1/ǫ)⌋+5)k.
Proof. Since Algorithm 1 uses only the operations insert, delete, shiftA and shiftB, the proper-
ties (a)to(d) are always fulfilled by Lemma 3 and the LP/ILP solutions x, y correspond to the
rounding function by Lemma 4.
Furthermore, the algorithm is designed such that whenever k increases the B-block is empty
and the A-block is renamed to be the new B-block. Whenever k decreases the A-block is empty
and the B-block is renamed to be the new A-block. Therefore the number of items in the groups
is dynamically adapted to match with the parameter k.
3.4 Large items
In this section we prove that Algorithm 1 is a dynamic robust AFPTAS for the bin packing
problem if all items have size at least ǫ/14. The treatment of small items is described in Section
4 and the general case is described in Section 4.2.
We will prove that the migration between packings Bt and Bt+1 is bounded by O(1/ǫ3 log(1/ǫ))
and that we can guarantee an asymptotic approximation ratio such that max Bt(i) ≤ (1 +
2∆) opt(I(t), s) + poly(1/∆) for a parameter ∆ = O(ǫ) and for every t ∈ N. The Algorithm
improve was developed in [JK13] to improve the objective value of an LP with integral solution
y and corresponding fractional solution x. For a vector z ∈ Rn let V (z) be the set of all integral
vectors v = (v1, . . . vn)T such that 0 ≤ vi ≤ zi.
Let x be an approximate solution of the LP min{kxk1 Ax ≥ b, x ≥ 0} with m inequalities
and let kxk1 ≤ (1 + δ) lin and kxk1 ≥ 2α(1/δ + 1), where lin denotes the fractional optimum of
the LP and α ∈ N is part of the input of the algorithm (see Jansen and Klein [JK13]). Let y be
an approximate integer solution of the LP with kyk1 ≤ lin +2C for some value C ≥ δ lin and
with kyk1 ≥ (m + 2)(1/δ + 2). Suppose that both x and y have only ≤ C non-zero components.
For every component i we suppose that yi ≥ xi. Furthermore we are given indices a1, . . . , aK ,
such that the non-zero components yaj are sorted in non-decreasing order, i. e., ya1 ≤ . . . ≤ yaK .
Algorithm 2 (improve).
1. Set xvar := 2 α(1/δ+1)
kxk x, xf ix := x − xvar and bvar = b − A(xf ix)
2. Compute an approximate solution x of the LP min{kxk1 Ax ≥ bvar, x ≥ 0} with ratio
(1 + δ/2)
4. Choose the largest ℓ such that the sum of the smallest components y1, . . . , yℓ is bounded by
3. If (cid:13)(cid:13)(cid:13)xf ix + x(cid:13)(cid:13)(cid:13)1 ≥ kxk1 then set x′ = x, y = y and goto step 9
P1≤i≤ℓ yai ≤ (m + 2)(1/δ + 2)
17
5. For all i set ¯xf ix
i =(0
xf ix
i
if i = aj, j ≤ ℓ
else
and ¯yi =(0
yi
if i = aj, j ≤ ℓ
else
6. Set ¯x = x + xℓ where xℓ is a vector consisting of the components xa1, . . . , xaℓ . Reduce the
number of non-zero components to at most m + 1.
7. x′ = ¯xf ix + ¯x
8. For all non-zero components i set yi = max{⌈x′
9. If possible choose d ∈ V (y− x′) such that kdk1 = α(1/δ + 1) otherwise choose d ∈ V (y− x′)
i⌉, ¯yi}
such that kdk1 < α(1/δ + 1) is maximal.
10. Return y′ = y − d
In the following we prove that the algorithm improve applied to the bin packing ILP actually
generates a new improved packing B′ from the packing B with corresponding LP and ILP
solutions x′ and y′. We therefore use Theorem 2 and Corollary 2 that were proven in [JK13].
Theorem 2. Let x be a solution of the LP with kxk1 ≤ (1 + δ) lin and furthermore kxk1 ≥
2α(1/δ + 1). Let y be an integral solution of the LP with ky′k1 ≤ lin +2C for some value
C ≥ δ lin and with kyk1 ≥ (m + 2)(1/δ + 2). Solutions x and y have the same number of non-
zero components and for each component we have xi ≤ yi. The Algorithm improve(α) then
returns a fractional solution x′ with kx′k1 ≤ (1+δ) lin −α and an integral solution y′′ where one
of the two properties hold: ky′k1 = kyk1 − α or ky′k1 = kx′k1 + C. Both, x′ and y′ have at most
C non-zero components and the distance between y′ and y is bounded by ky′ − yk1 = O( m+α
).
Corollary 2. Let kxk1 = (1 + δ′) lin for some δ′ ≥ δ and kxk1 ≥ 2α(1/δ + 1) and let kyk1 ≤
lin +2C for some C ≥ δ′ lin and kyk1 ≥ (m + 2)(1/δ + 2). Solutions x and y have the same
number of non-zero components and for each component we have xi ≤ yi. Then Algorithm
improve(α) returns a fractional solution x′ with kx′k1 ≤ kxk1−α = (1+δ′) lin −α and integral
solution y′ where one of the two properties hold: ky′k1 = kyk1−α or ky′k1 = kxk1−α+C. Both,
x′ and y′ have at most C non-zero components and the distance between y′ and y is bounded by
ky′ − yk1 ∈ O( m+α
).
δ
δ
Let ∆ = ǫ + δ + ǫδ and C = ∆ opt(I, s) + m.
Theorem 3. Given a rounding function R and an LP defined for (I, sR), let x be a fractional so-
lution of the LP with kxk1 ≤ (1+∆) opt(I, s), kxk1 ≥ 2α(1/δ +1) and kxk1 = (1+δ′) lin(I, sR)
for some δ′ > 0. Let y be an integral solution of the LP with kyk1 ≥ (m + 2)(1/δ + 2) and
corresponding packing B such that maxi B(i) = kyk1 ≤ (1 + 2∆) opt(I, s) + m. Suppose x
and y have the same number ≤ C of non-zero components and for all components i we have
yi ≥ xi. Then Algorithm improve(α) on x and y returns a new fractional solution x′ with
kx′k1 ≤ (1 + ∆) opt(I, s)− α and also a new integral solution y′ with corresponding packing B′
such that
i
max
B′(i) =(cid:13)(cid:13)y′(cid:13)(cid:13)1 ≤ (1 + 2∆) opt(I, s) + m − α.
i ≤ y′
Further, both solutions x′ and y′ have the same number ≤ C of non-zero components and for
each component we have x′
i. The number of changed bins from the packing B to the packing
B′ is bounded by O( m
δ ).
18
Proof. To use Theorem 2 and Corollary 2 we have to prove that certain conditions follow
from the requisites of Theorem 3. We have maxi B(i) = kyk1 ≤ (1 + 2∆) opt(I, s) + m by
condition. Since opt(I, s) ≤ opt(I, sR) we obtain for the integral solution y that kyk1 ≤
2∆ opt(I, s) + m + opt(I, sR) ≤ 2∆ opt(I, s) + m + lin(I, sR) + m. Hence by definition of C
we get kyk1 ≤ lin(I, sR) + 2C. This is one requirement to use Theorem 2 or Corollary 2. We
distinguish the cases where δ′ ≤ δ and δ′ > δ and look at them separately.
Case 1: δ′ ≤ δ. For the parameter C we give a lower bound by the inequality C >
∆ opt(I, s) = (δ + ǫ + δǫ) opt(I, s). Lemma 2 shows that opt(I, sR) ≤ (1 + ǫ) opt(I, s) and
therefore yields
δ + ǫ + δǫ
opt(I, sR) =
1 + ǫ
= (1 + δ) opt(I, sR) −
1 + ǫ
≥ δ opt(I, sR) ≥ δLIN (I, sR)
(1 + δ)(1 + ǫ) − 1
1
1 + ǫ
opt(I, sR)
opt(I, sR)
and hence C > δ lin(I, sR). We can therefore use Theorem 2.
m
Algorithm improve returns by Theorem 2 a x′ with kx′k1 ≤ (1 + δ) lin(I, sR) − α ≤ (1 +
δ) opt(I, sR)− α and an integral solution y′ with ky′k1 ≤ kx′k1 + C or ky′k1 ≤ kyk1 − α. Using
that opt(I, sR) ≤ (1 + ǫ) opt(I, s) we can conclude kx′k1 ≤ (1 + δ)(1 + ǫ) opt(I, s) − α =
(1 + ∆) opt(I, s) − α. In the case where ky′k1 ≤ kx′k1 + C we can bound the number of bins of
the new packing B′ by maxi B′(i) = ky′k1 ≤ kx′k1 + C ≤ (1 + 2∆) opt(I, s) + m− α. In the case
that ky′k1 ≤ kyk1 − α we obtain maxi B′(i) = ky′k1 ≤ kyk1 − α ≤ (1 + 2∆) opt(I, s) + m − α.
Furthermore we know by Theorem 2 that x′ and y′ have at most C non-zero components.
Case 2: δ′ > δ. First we prove that C is bounded from below. Since kxk1 = (1 +
δ′) lin(I, sR) ≤ (1+∆) opt(I, s) ≤ (1+∆) opt(I, sR) ≤ (1+∆) opt(I, sR) ≤ (1+∆)(lin(I, sR)+
2 ) ≤ lin(I, sR) + C we obtain that C ≥ δ′ lin(I, sR), which is a requirement to use Corol-
lary 2. By using Algorithm improve on solutions x with kxk1 = (1 + δ′) lin(I, sR) and y
with kyk1 ≤ lin(I, sR) + 2C we obtain by Corollary 2 a fractional solution x′ with kx′k1 ≤
kxk1 − α ≤ (1 + ∆) opt(I, s) − α and an integral solution y′ with either ky′k1 ≤ kyk1 − α or
ky′k1 ≤ kxk1 + C − α. So for the new packing B′ we can guarantee that maxi B′(i) = ky′k1 ≤
kyk1 − α = maxi B(i) − α ≤ (1 + 2∆) opt(I, s) + m − α if ky′k1 ≤ kyk1 − α.
In the case
that ky′k1 ≤ kxk1 + C − α, we can guarantee that maxi B′(i) = ky′k1 ≤ kxk1 + C − α ≤
(1 + ∆) opt(I, s) + C − α ≤ (1 + 2∆) opt(I, s) + m − α. Furthermore we know by Corollary 2
that x′ and y′ have at most C non-zero components.
Theorem 2 as well as Corollary 2 state that the distance ky′ − yk1 is bounded by O(m/δ).
Since y corresponds directly to the packing B and the new integral solution y′ corresponds to
the new packing B′, we know that only O(m/δ) bins of B need to be changed to obtain packing
B′.
In order to prove correctness of Algorithm 1, we will make use of the auxiliary Algorithm
3 (ReduceComponents). Due to a delete-operation, the value of the optimal solution opt(I, s)
might decrease. Since the number of non-zero components has to be bounded by C = ∆ opt(I, s)+
m, the number of non-zero components might have to be adjusted down. The following algo-
rithm describes how a fractional solution x′ and an integral solution y′ with reduced number
of non-zero components can be computed such that ky − y′k1 is bounded. The idea behind the
algorithm is also used in the Improve algorithm. The smallest m+2 components are reduced to
m+1 components using a standard technique presented for example in [BM98]. Arbitrary many
components of x′ can thus be reduced to m + 1 components without making the approximation
guarantee worse.
19
i =
⌈xi + x′
i⌉
0
yi
xi + xi
0
xi
if i = bj for some j ≤ m
if i = aj for some j ≤ m + 1
else
if i = bj for some j ≤ m
if i = aj for some j ≤ m + 1
else
4. For all i set x′
and yi =
Algorithm 3 (ReduceComponents).
1. Choose the smallest non-zero components ya1, . . . , yam+2.
2. If P1≤i≤m+2 yai ≥ (1/∆ + 2)(m + 2) then return x = x′ and y = y′
3. Reduce the components xa1, . . . , xam+2 to m+1 components xb1, . . . , xbm+1 withPm+2
Pm+1
j=1 xbj .
j=1 xaj =
5. If possible choose d ∈ V (y − x′) such that kdk1 = m + 1 otherwise choose d ∈ V (y − x′)
such that kdk1 < m + 1 is maximal.
6. Return y′ = y − d
The following theorem shows that the algorithm above yields a new fractional solution x′ and
a new integral solution y′ with a reduced number of non-zero components.
i ≤ y′
Theorem 4. Let x be a fractional solution of the LP with kxk1 ≤ (1 + ∆) opt(I, s). Let y
be an integral solution of the LP with kyk1 ≤ (1 + 2∆) opt(I, s) + m. Suppose x and y have
the same number ≤ C + 1 of non-zero components and for all components i we have yi ≥ xi.
Using the Algorithm ReduceComponents on x and y returns a new fractional solution x′ with
kx′k1 ≤ (1 + ∆) opt(I, s) and a new integral solution y′ with ky′k1 ≤ (1 + 2∆) opt(I, s) + m.
Further, both solutions x′ and y′ have the same number of non-zero components and for each
component we have x′
i. The number of non-zero components can now be bounded by ≤ C.
Furthermore, we have that ky − y′k1 ≤ 2 · (1/∆ + 3)(m + 2).
Proof. Case 1: P1≤i≤m+2 yai ≥ (1/∆ + 2)(m + 2). We will show that in this case, x and y
already have ≤ C non-zero components. In this case the algorithm returns x′ = x and y′ = y.
SinceP1≤i≤m+2 yai ≥ (1/∆+2)(m+2) the components ya1, . . . , yam+2 have an average size of at
least (1/∆ + 2) and since ya1, . . . , yam+2 are the smallest components, all components of y have
average size at least (1/∆ + 2). The size kyk1 is bounded by (1 + 2∆) opt(I, s) + m. Hence the
number of non-zero components can be bounded by (1+2∆) opt(I,s)+m
≤ ∆ opt(I, s) + ∆m ≤ C.
Case 2: P1≤i≤m+1 yai < (1/∆ + 2)(m + 2). We have to prove different properties for the new
fractional solution x′ and the new integral solution y′.
Number of non-zero components: The only change in the number of non-zero components
is in step 3 of the algorithm, where the number of non-zero components is reduced by 1. As
x, y have at most C + 1 non-zero components, x′, y′ have at most C non-zero components. In
step 4 of the algorithm, y is defined such that yi ≥ x′
i. In step 5 of the algorithm d is chosen
such that yi − d ≥ x′
Distance between y and y′: The only steps where components of y changes are in step 4
and 5. The distance between y and y is bounded by the sum of the components that are set
j=1 ⌈xbj⌉ ≤
j=1 yaj < (1/∆ + 2)(m + 2), we
j=1 yaj and the sum of the increase of the increased components Pm+1
i. Hence we obtain that y′
i = yi − d ≥ x′
i.
1/∆+2
to 0, i. e., Pm+2
Pm+1
j=1 xbj + m + 1 = Pm+2
j=1 xaj + m + 1. As Pm+2
j=1 xaj ≤ Pm+2
20
i − x′
ǫ3 · log(1/ǫ)) for the
obtain that the distance between y and y is bounded by 2· (1/∆ + 2)(m + 2) + m + 1. Using that
kdk1 ≤ m + 1, the distance between y and y′ is bounded by ky′ − yk1 < 2 · (1/∆ + 3)(m + 2).
Approximation guarantee: The fractional solution x is modified by condition of step 3
such that the sum of the components does not change. Hence kx′k1 = kxk1 ≤ (1 + ∆) opt(I, s).
Case 2a: kdk1 < m + 1. Since d is chosen maximally we have for every non-zero component
that y′
i < 1. Since there are at most C = ∆ opt(I, s) + m non-zero components we obtain
that ky′k1 ≤ kx′k1 + C ≤ (1 + 2∆) opt(I, s) + m. Case 2b: kdk1 = m + 1. By definition of y
we have kyk1 ≤ kyk1 +Pm+1
j=1 xaj ≤ kyk1 + m + 1. We obtain for y′ that
ky′k1 = kyk1 − kdk1 ≤ kyk1 + m + 1 − (m + 1) = kyk1 ≤ (1 + 2∆) opt(I, s) + m.
Theorem 5. Algorithm 1 is an AFPTAS with migration factor at most O( 1
fully dynamic bin packing problem with respect to large items.
Proof. Set δ = ǫ. Then ∆ = 2ǫ + ǫ2 = O(ǫ). We assume in the following that ∆ ≤ 1 (which
holds for ǫ ≤ √2 − 1).
We prove by induction that four properties hold for any packing Bt and the corresponding
LP solutions. Let x be a fractional solution of the LP defined by the instance (It, sRt) and y
be an integral solution of this LP. The properties (2) to (4) are necessary to apply Theorem 3
and property (1) provides the wished approximation ratio for the bin packing problem.
j=1 ⌈ xbj + xbj⌉ −Pm+2
(1) maxi Bt(i) = kyk1 ≤ (1 + 2∆) opt(I(t), s) + m (the number of bins is bounded)
(2) kxk1 ≤ (1 + ∆) opt(I(t), s)
(3) for every configuration i we have xi ≤ yi
(4) x and y have the same number of non-zero components and that number is bounded by
∆ opt(I(t), s) + m
To apply Theorem 3 we furthermore need a guaranteed minimal size for kxk1 and kyk1. Accord-
ing to Theorem 3 the integral solution y needs kyk1 ≥ (m+2)(1/δ+2) and kxk1 ≥ 8(1/δ+1) as we
set α ≤ 4. By condition of the while-loop the call of improve is made iff SIZE(It, s) ≥ 8(1/δ+1)
and SIZE(It, s) ≥ (m + 2)(1/δ + 2). Since kyk1 ≥ kxk1 ≥ SIZE(It, s) the requirements for the
minimum size are fulfilled. As long as the instance is smaller than 8(1/δ + 1) or (m + 2)(1/δ + 2)
an offline algorithm for bin packing is used. Note that there is an offline algorithm which fulfills
properties (1) to (4) as shown by Jansen and Klein [JK13].
Now let Bt be a packing with SIZE(It, s) ≥ 8(1/δ + 1) and SIZE(It, s) ≥ (m + 2)(1/δ + 2)
for instance It with solutions x and y of the LP defined by (I(t), sRt ). Suppose by induction
that the properties (1) to (4) hold for the instance It. We have to prove that these properties
also hold for the instance I(t + 1) and the corresponding solutions x′′ and y′′. The packing
Bt+1 is created by the repeated use of an call of improve for x and y followed by an operation
(insert, delete, shiftA or shiftB). We will prove that the properties (1) to (4) hold after a call of
improve followed by an operation.
improve: Let x′ be the resulting fractional solution of Theorem 3, let y′ be the resulting integral
solution of Theorem 3 and let B′
t be the corresponding packing. Properties (1) to (4) are fulfilled
for x, y and Bt by induction hypothesis. Hence all conditions are fulfilled to use Theorem 3.
By Theorem 3 the properties (1) to (4) are still fulfilled for x′, y′ and B′
t and moreover we get
kx′k1 ≤ (1 + ∆) opt(I(t), s) − α and ky′k1 = maxi B′
t(i) ≤ (1 + 2∆) opt(I(t), s) + m − α for
chosen parameter α. Let x′′ and y′′ be the fractional and integral solution after an operation
is applied to x′ and y′. We have to prove that the properties (1) to (4) are also fulfilled for x′′
and y′′.
21
operations: First we take a look at how the operations modify kx′k1 and ky′k1 = maxi B′
t(i).
By construction of the insertion operation, kx′k1 and ky′k are increased at most by 2. By
construction of the delete operation, kx′k1 and ky′k1 are increased by 1. By construction of the
shiftA and shiftB operation, kx′k1 and ky′k1 are increased by 1. An improve(2) call followed
by an insertion operation therefore yields ky′′k = ky′k1 + 2 = (1 + 2∆) opt(I(t), s) + m− 2 + 2 =
(1 + 2∆) opt(I(t + 1), s) + m since opt(I(t), s) ≤ opt(I(t + 1), s). An improve(4) call followed
by a delete operation yields ky′′k = ky′k1 +1 = (1+2∆) opt(I(t), s)+m−3 ≤ (1+2∆) opt(I(t+
1), s)+(1+2∆)+m−3 ≤ (1+2∆) opt(I(t+1), s) since opt(I(t), s) ≤ opt(I(t+1), s)+1 (an item
is removed) and ∆ ≤ 1. In the same way we obtain that ky′′k1 ≤ ky′k1 + 1 ≤ (1 + 2∆) opt(I(t +
1), s) + m for an improve(1)/improve(3) call followed by a shiftA/shiftB operation. This
concludes the proof that property (1) is fulfilled for I(t + 1). The proof that property (2) holds
is analog since kx′k1 increases in the same way as ky′k1 and kx′k1 ≤ (1 + ∆) opt(I(t), s) − α.
For property (3) note that in the operations a configuration xi of the fractional solution is
increased by 1 if and only if a configuration yi is increased by 1. Therefore the property that
i retains from x′ and y′. By Theorem 3 the number of non-zero
for all configurations x′′
components of x′ and y′ is bounded by ∆ opt(I(t), s) + m ≤ ∆ opt(I(t + 1), s) + m in case of
an insert operation. If an item is removed, the number of non-zero components of x′ and y′ is
bounded by ∆ opt(I(t), s)+m ≤ ∆ opt(I(t+1), s)+m+1 = C +1. By Theorem 4 the algorithm
ReduceComponents guarantees that there are at most C = ∆ opt(I(t + 1), s) + m non-zero
components. By construction of the shift-operation, x′′ and y′′ might have two additional non-
zero components. But since these are being reduced by Algorithm 1 (note that we increased the
number of components being reduced in step 6 by 2 to- see [JK13] for details), the LP solutions
x′′ and y′′ have at most ∆ opt(I(t + 1), s) + m non-zero components which proves property (4).
Algorithm 1 therefore has an asymptotic approximation ratio of 1 + ǫ.
i ≤ y′′
ǫ3 log(1/ǫ)).
We still need to examine the migration factor of Algorithm 1. In the case that the offline
algorithm is used, the size of the instance is smaller than 8(1/δ + 1) = O(1/ǫ) or smaller than
(m + 2)(1/δ + 2) = O( 1
ǫ2 log(1/ǫ)). Hence the migration factor in that case is bounded by
O( 1
If the instance is bigger the call of improve repacks at most O(m/ǫ) bins by
Theorem 3. Since every large arriving item has size > ǫ/14 and m = O( 1
ǫ log(1/ǫ)) we obtain
a migration factor of O( 1
ǫ3 log(1/ǫ)) for the Algorithm improve. Since the migration factor
of each operation is also bounded by O( 1
ǫ2 log(1/ǫ)), we obtain an overall migration factor of
O( 1
ǫ3 log(1/ǫ)).
The main complexity of Algorithm 1 lies in the use of Algorithm improve. As described
by Jansen and Klein [JK13] the running time of improve is bounded by O(M (1/ǫ log(1/ǫ)) ·
1/ǫ3 log(1/ǫ)), where M (n) is the time needed to solve a system of n linear equations. By using
heap structures to store the items, each operation can be performed in time O(1/ǫ log(1/ǫ) ·
log(ǫ2 · n(t))) at time t, where n(t) denotes the number of items in the instance at time t.
As the number of non-zero components is bounded by O(ǫ · n(t)), the total running time of
the algorithm is bounded by O(M (1/ǫ log(1/ǫ)) · 1/ǫ3 log(1/ǫ) + 1/ǫ log(1/ǫ) log(ǫ2 · n(t)) + ǫn(t)).
The best known running time for the dynamic bin packing problem without removals was
O(M (1/ǫ2) · 1/ǫ4 + ǫn(t) + 1
ǫ2 log(ǫ2n(t))) and is due to Jansen and Klein [JK13]. As this is
polynomial in n(t) and in 1/ǫ we can conclude that Algorithm 1 is an AFPTAS.
If no deletions are present, we can use a simple FirstFit algorithm (as described by Jansen
and Klein [JK13]) to pack the small items into the bins. This does not change the migration
factor or the running time of the algorithm and we obtain a robust AFPTAS with O( 1
ǫ3 ·log(1/ǫ))
migration for the case that no items is removed. This improves the best known migration factor
of O( 1
ǫ4 ) [JK13].
22
4 Handling Small Items
In this section we present methods for dealing with arbitrary small items in a dynamic online
setting. First, we present a robust AFPTAS with migration factor of O(1/ǫ) for the case that
only small items arrive and depart. In Section 4.3 we generalize these techniques to a setting
where small items arrive into a packing where large items are already packed and can not be
rearranged. Finally we state the AFPTAS for the general fully dynamic bin packing problem.
In a robust setting without departing items, small items can easily be treated by packing them
greedily via the classical FirstFit algorithm of Johnson et al. [JDU+74b] (see Epstein and Levin
[EL09] or Jansen and Klein [JK13]). However, in a setting where items may also depart, small
items need to be treated much more carefully. We show that the FirstFit algorithm does not
work in this dynamic setting.
Lemma 7. Using the FirstFit algorithm to pack small items may lead to an arbitrarily bad
approximation.
Proof. Suppose, that there is an algorithm A with migration factor c which uses FirstFit on
items with size < ǫ/14. We will now construct an instance where A yields an arbitrary bad
approximation ratio. Let b = ǫ/14 − δ and a = ǫ/14c − ((δ+cδ)/c) for a small δ such that (1−b)/a is
integral. Note that ac < b by definition. Furthermore, let M ∈ N be an arbitrary integer and
consider the instance
with
IM = [A, A, . . . , A
, B, B, . . . , B
]
M
{z
}
M
{z
}
A = (b, Insert), (a, Insert), (a, Insert), . . . , (a, Insert)
B = (a, Delete), (a, Delete), . . . , (a, Delete)
(1−b)/a
{z
}
.
(1−b)/a
{z
}
After the insertion of all items, there are M bins containing an item of size b and 1−b/a items
of size a (see Figure 8a). As ac < b, the deletion of the items of size a can not move the items
of size b. The remaining M bins thus only contain a single item of size b (see Figure 8b), while
⌈M · b⌉ bins would be sufficient to pack all of the remaining items. The approximation ratio
is thus at least M/M ·b = 1/b ≈ 1
ǫ and thus grows as ǫ shrinks. In order to avoid this problem,
we design an algorithm which groups items of similar size together. Using such a mechanism
would therefore put the second item of size b into the first bin by shifting out an appropriate
number of items of size a and so on. Our algorithms achieves this grouping of small items by
enumerating the bins and maintaining the property, that larger small items are always left of
smaller small items.
4.1 Only Small Items
We consider a setting where only small items exist, i. e., items with a size less than ǫ/14. First,
we divide the set of small items into different size intervals Sj where Sj =h ǫ
Let b1, . . . , bm be the used bins of our packing. We say a size category Sj is bigger than a size
category Sk if j < k, i. e., the item sizes contained in Sj are larger (note that a size category
2j(cid:17) for j ≥ 1.
2j+1 , ǫ
23
a
b
b
(a) A single bin after the insertion
(b) A single bin after the deletion
Figure 8: Construction in the proof of Lemma 7
Sj with large index j is called small). We say a bin bi is filled completely if it has less than ǫ
2j
remaining space, where Sj is the biggest size category appearing in bi. Furthermore we label
bins bi as normal or as buffer bins and partition all bins b1, . . . , bm into queues Q1, . . . , Qd for
Q ≤ m. A queue is a subsequence of bins bi, bi+1 . . . , bi+c where bins bi, . . . , bi+c−1 are normal
bins and bin bi+c is a buffer bin. We denote the i-th queue by Qi and the number of bins in Qi
by Qi. The buffer bin of queue Qi is denoted by bbi.
erties are always fulfilled. For the sake of simplicity, we assume that 1/ǫ is integral.
(1) For every item i ∈ bd with size s(i) ∈ Sj for some j, d ∈ N, there is no item i′ ∈ bd′ with
size s(i′) ∈ Sj ′ such that d′ > d and j′ > j. This means: Items are ordered from left to
right by their size intervals.
We will maintain a special form for the packing of small items such that the following prop-
(2) Every normal bin is filled completely.
(3) The length of each queue is at least 1/ǫ and at most 2/ǫ except for the last queue Qd.
Note that property (1) implies that all items in the same size interval Sj are packed into bins
bx, bx+1, . . . , bx+c for constants x and c. Items in the next smaller size category Sj+1 are then
packed into bins bx+c, bx+c+1, . . . and so on. We denote by bS(ℓ) the last bin in which an item
of size interval Sℓ appears. We denote by S>ℓ the set of smaller size categories Sℓ′ with ℓ′ > ℓ.
Note that items in size category S>ℓ are smaller than items in size category Sℓ.
. . .
. . .
. . .
. . .
b1
Q1
bQ1−1 bb1
bQ1+1
Q2
bQ2−1 bb2
bQd−1+1
Qd
bQd−1 bbd
Figure 9: Distribution of bins with small items into queues
The following lemma guarantees that a packing that fulfills properties (1) to (3) is close to
the optimum solution.
Lemma 8. If properties (1) to (3) hold, then at most (1 + O(ǫ)) opt(I, s) + 2 bins are used in
the packing for every ǫ ≤ 1/3.
Proof. Let C be the number of used bins in our packing. By property (2) we know that all
normal bins have less than ǫ/14 free space. Property (3) implies that there are at most ǫ · C + 1
buffer bins and hence possibly empty. The number of normal bins is thus at least (1− ǫ)· C − 1.
Therefore we can bound the total size of all items by ≥ (1 − ǫ/14) · ((1 − ǫ) · C − 1). As
opt(I, s) ≥ SIZE(I, s) ≥ (1− ǫ/14)· ((1 − ǫ)· C − 1) and
(1−ǫ/14)(1−ǫ) ≤ 1 + 2ǫ for ǫ ≤ 1/3 we get
C ≤ (1 + 2ǫ) opt(I, s) + 2.
1
24
We will now describe the operations that are applied whenever a small item has to be inserted
or removed from the packing. The operations are designed such that properties (1) to (3) are
never violated and hence a good approximation ratio can be guaranteed by Lemma 8 at every
step of the algorithm. The operations are applied recursively such that some items from each
size interval are shifted from left to right (insert) or right to left (delete). The recursion halts if
the first buffer bin is reached. Therefore, the free space in the buffer bins will change over time.
Since the recursion always halts at the buffer bin, the algorithm is applied on a single queue
Qk.
The following Insert/Delete operation is defined for a whole set J = {i1, . . . , in} of items.
If an item i of size interval Sℓ has to be inserted or deleted, the algorithm is called with
Insert({i}, bS(ℓ), Qk) respectively Delete({i}, bx, Qk), where bx is the bin containing item i and
2j(cid:17) is a fixed interval for
Qk is the queue containing bin bS(ℓ) or bx. Recall that Sj = h ǫ
every j ≥ 1 and S≤j =Sj
• Insert(J, bx, Qk):
i=1 Si and S>j =Si>j Si.
Algorithm 4 (Insert or Delete for only small items).
2j+1 , ǫ
– Insert the set of small items J = {i1, . . . , in} with size s(ij) ∈ S≤ℓ into bin bx. (By
Lemma 9 the total size of J is bounded by O(1/ǫ) times the size of the item which
triggered the first Insert operation.)
– Remove just as many items J ′ = {i′
m} of the smaller size interval S>ℓ appear-
ing in bin bx (starting by the smallest) such that the items i1, . . . , in fit into the bin
bx. If there are not enough items of smaller size categories to insert all items from
I, insert the remaining items from I into bin bx+1.
1, . . . , i′
– Let J ′
ℓ′ ⊆ J ′ be the items in the respective size interval Sℓ′ with ℓ′ > ℓ. Put the items
ℓ′ , bS(ℓ′), Qk) for each ℓ′ > ℓ). If the
ℓ′ recursively into bin bS(ℓ′) (i. e., call Insert(J ′
J ′
buffer bin bbk is left of bS(ℓ′) call Insert(J ′
ℓ′ , bbk, Qk) instead.
• Delete(J, bx, Qk):
– Insert as many small items J ′ = {i′
– Remove the set of items J = {i1, . . . , in} with size s(ij) ∈ S≤ℓ from bin bx (By Lemma
9 the total size of J is bounded by O(1/ǫ) times the size of the item which triggered
the first Delete operation.)
m} from bS(ℓ′), where Sℓ′ is the smallest
size interval appearing in bx such that bx is filled completely. If there are not enough
items from the size category Sℓ′, choose items from size category S≥ℓ′+1 in bin bx+1.
ℓ′ ⊆ J ′ be the items in the respective size interval Sℓ′ with ℓ′ > ℓ. Remove items
ℓ′, bS(ℓ′), Qk) for each ℓ′ > ℓ). If the
ℓ′ from bin bS(ℓ′) recursively (i. e., call Delete(J ′
J ′
buffer bin bbk is left of bS(ℓ′), call Delete(J ′
ℓ′, bbk, Qk) instead.
1, . . . , i′
– Let J ′
Using the above operations maintains the property of normal bins to be filled completely.
However, the size of items in buffer bins changes. In the following we describe how to handle
buffer bins that are being emptied or filled completely.
Algorithm 5 (Handle filled or emptied buffer bins).
• Case 1: The buffer bin of Qi is filled completely by an insert operation.
– Label the filled bin as a normal bin and add a new empty buffer bin to the end of Qi.
– If Qi > 2/ǫ, split Qi into two new queues Q′
i + 1. The buffer
i is the newly added buffer bin. Add an empty bin labeled as the buffer bin
i with Q′′
i = Q′
i, Q′′
bin of Q′′
to Q′
i such that Q′
i = Q′′
i .
– Remove the now empty bin.
• Case 2: The buffer bin of Qi is being emptied due to a delete operation.
25
bin of Qi.
– If Qi ≥ Qi+1 and Qi > 1/ǫ, choose the last bin of Qi and label it as new buffer
– If Qi+1 > Qi and Qi+1 > 1/ǫ, choose the first bin of Qi+1 and move the bin to Qi
– If Qi+1 = Qi = 1/ǫ, merge the two queues Qi and Qi+1. As Qi+1 already contains
a buffer bin, there is no need to label another bin as buffer bin for the merged queue.
and label it as buffer bin.
Creating and deleting buffer bins this way guarantees that property (3) is never violated since
queues never exceed the length of 2/ǫ and never fall below 1/ǫ.
i
S0
S1
S2
bx
S2
S3
S2
bx+1
bx+2
. . .
Sj
Sj+1
bbk
S0
i
S1
S2
bx
S2
S3
S2
bx+1
bx+2
. . .
Sj
Sj+1
bbk
(a) Insert({i},bx,Qk) with s(i) ∈ S1
(b) Delete({i},bx,Qk) with s(i) ∈ S1
Figure 10: Example calls of Insert and Delete.
Figure 10a shows an example call of Insert({i},bx,Qk). Item i with s(i) ∈ S1 is put into the
corresponding bin bx into the size interval S1. As bx now contains too many items, some items
from the smallest size interval S2 (marked by the dashed lines) are put into the last bin bx+2
containing items from S2. Those items in turn push items from the smallest size interval S3
into the last bin containing items of this size and so on. This process terminates if either no
items need to be shifted to the next bin or the buffer bin bbk is reached.
It remains to prove that the migration of the operations is bounded and that the properties
are invariant under those operations.
Lemma 9.
(i) Let I be an instance that fulfills properties (1) to (3). Applying operations insert/delete
on I yields an instance I ′ that also fulfills properties (1) to (3).
(ii) The migration factor of a single insert/delete operation is bounded by O(1/ǫ) for all ǫ ≤ 2/7.
Proof. Proof for (i): Suppose the insert/delete operation is applied to a packing which fulfills
properties (1) to (3). By construction of the insert operation, items from a size category Sℓ in bin
bx are shifted to a bin by. The bin by is either bS(ℓ) or the a buffer bin left of bS(ℓ). By definition
by contains items of size category Sℓ. Therefore property (1) is not violated. Symmetrically, by
construction of the delete operation, items from a size category Sℓ in bin bS(ℓ) are shifted to
a bin bx. By definition bx contains items of size category Sℓ and property (1) is therefore not
violated. For property (2): Let bx be a normal bin, where items i1, . . . , in of size category S≤ℓ
are inserted. We have to prove that the free space in bx remains smaller than ǫ/2j, where Sj is
the smallest size category appearing in bin bx. By construction of the insert operation, just as
many items of size categories S>ℓ are shifted out of bin bx such that i1, . . . , in fit into bx. Hence
the remaining free space is less than ǫ
2ℓ and bin bx is filled completely. The same argumentation
holds for the delete operation. Property (3) is always fulfilled by definition of Algorithm 5.
Proof for (ii): According to the insert operation, in every recursion step of the algorithm, it
tries to insert a set of items into a bin bx′, starting with an Insert({i}, bx′ , Qk) operation. Let
insert(S≤ℓ+y, bx) (x ≥ x′) be the size of all items in size categories Sj with j ≤ ℓ + y that
the algorithm tries to insert into bx as a result of an Insert({i}, bx′ , Qk) call. Let pack(bx) be
26
the size of items that are actually packed into bin bx. We have to distinguish between two
cases. In the case that insert(S≤ℓ+y, bx) = pack(bx) there are enough items of smaller size
categories S>ℓ+y that can be shifted out, such that items I fit into bin bx. In the case that
insert(S≤ℓ+y, bx) > pack(bx) there are not enough items of smaller size category that can be
shifted out and the remaining size of insert(S≤ℓ+y, bx) − pack(bx) has to be shifted to the
following bin bx+1. Under the assumption that each insert(S≤ℓ, bx) ≤ 1 for all x and ℓ (which
is shown in the following) all items fit into bx+1. Note that no items from bins left of bx can be
shifted into bx+1 since bx = bS(ℓ+y) is the last bin where items of size category S≤ℓ+y appear.
Hence all items shifted out from bins left of bx are of size categories S≤ℓ+y (property (1)) and
they are inserted into bins left of bx+1. We prove by induction that for each insert(S≤ℓ+y, bx)
the total size of moved items is at most
insert(S≤ℓ+y, bx) ≤ s(i) + 3
y
Xj=1
ǫ
2ℓ+j
The claim holds obviously for insert(S≤ℓ, bx′) since bx′ = bS(ℓ) is the bin where only item i is
inserted.
Insert [S1, S2]
S1
S2
bx
S2
S3
bx+1
(a) Case 1
Insert [S1]
S1
S2
bx
S2
S3
bx+1
. . .
(b) Case 2a
Insert [S1, S2]
S1
S2
bx
S2
bx+1
(c) Case 2b
S2
S3
bx+1
. . .
Figure 11: All cases to consider in Lemma 9
Case 1: insert(S≤ℓ+y, bx) > pack(bx)
In this case, the size of all items that have to be inserted into bx+1 can be bounded by the size
of items that did not fit into bin bx plus the size of items that were removed from bin bx. We
can bound insert(S≤ℓ+¯y, bx+1) where ¯y > y is the largest index Sℓ+¯y appearing in bin bx by
insert(S≤ℓ+y, bx) +
ǫ
2ℓ+y ≤ s(i) + 3
y
Xj=1
ǫ
2ℓ+j + 2
ǫ
2ℓ+y+1 < s(i) + 3
y+1
Xj=1
ǫ
2ℓ+j
Case 2: insert(S≤ℓ+y, bx) = pack(bx)
Suppose that the algorithm tries to insert a set of items I of size categories S≤ℓ+¯y into the bin
bx+1 = bS(ℓ+¯y). The items I can only be shifted from previous bins where items of size category
S≤ℓ+¯y appear. There are only two possibilities remaining. Either all items I are shifted from a
single bin bx (x ≤ x) or from two consecutive bins bx, bx+1 with insert(S≤ℓ+y, bx) > pack(bx).
Note that bx+1 can only receive items from more than one bin if there are two bins bx, bx+1
with insert(S≤ℓ+y, bx) > pack(bx) such that bx+1 = bS(ℓ+¯y) and all items shifted out of bx, bx+1
and into bx+1 are of size category Sℓ+¯y. Hence bins left of bx or right of bx+1 can not shift items
into bx+1.
Case 2a: All items I are shifted from a single bin bx with x ≤ x (note that x < x is possible
since pack(bx) = insert(S≤ℓ+y, bx) can be zero). The total size of items that are shifted out
of bx can be bounded by insert(S≤ℓ+y, bx) + ǫ
2ℓ+y . By induction hypothesis insert(S≤ℓ+y, bx)
ǫ
2ℓ+j . Since all items that are inserted into bx+1 come from bx,
2ℓ+y ≤ s(i) +
the value insert(S≤ℓ+¯y, bx+1) (¯y > y) can be bounded by insert(S≤ℓ+y, bx) + ǫ
is bounded by s(i) + 3Py
j=1
27
ǫ
j=1
j=1
2ℓ+j + ǫ
2ℓ+y < s(i) + 3P¯y
ǫ
2ℓ+j where Sℓ+¯y is the smallest size category inserted into
bx+1. Note that the items I belong to only one size category Sℓ+¯y if x < x since all items that
are in size intervals S<ℓ+¯y are inserted into bin bx+1.
3Py
Case 2b: Items I are shifted from bins bx and bx+1 (x + 1 ≤ x) with insert(S≤ℓ+y, bx) >
pack(bx). In this case, all items I belong to the size category Sℓ+¯y since bx is left of bx. Hence all
items which are inserted into bx+1 are from I, i. e., insert(S≤ℓ+y, bx) = pack(bx) + pack(bx+1)
as all items in I belong to the same size category Sℓ+¯y. We can bound insert(Sℓ+¯y, bx+1) by
the size of items that are shifted out of bx plus the size of items that are shifted out of bx+1.
We obtain
ǫ
2ℓ+y + pack(bx+1) +
ǫ
2ℓ+¯y
insert(S≤ℓ+¯y, bx+1) ≤ pack(bx) +
ǫ
= insert(S≤ℓ+y, bx)) +
ǫ
2ℓ+y +
2ℓ+¯y
ǫ
ǫ
2ℓ+j +
ǫ
2ℓ+y +
2ℓ+¯y
≤ s(i) + 3
≤ s(i) + 3
y
Xj=1
Xj=1
y
ǫ
2ℓ+j + 3
ǫ
2ℓ+¯y ≤ s(i) + 3
¯y
Xj=1
ǫ
2ℓ+j
This yields that insert(S≤ℓ+y, bx) is bounded by s(i)+ 3P¯y
we can bound the migration factor for every bin bx of Qk for any y ∈ N by pack(bx) + ǫ
insert(S≤ℓ+y, bx) + ǫ
ǫ
2ℓ+j for all bins bx in Qk. Now,
2ℓ+y ≤
2ℓ+y . Using the above claim, we get:
j=1
insert(S≤ℓ+y, bx) +
< s(i) + 3
∞
Xj=1
ǫ
2ℓ+y ≤ s(i) + 3
Xj=1
ǫ
2ℓ
∞
ǫ
2ℓ+j = s(i) + 3
y
ǫ
2ℓ+j + 2
Xj=1
1
2j = s(i) + 3 ·
ǫ
2ℓ+y+1
ǫ
2ℓ ≤ 7s(i)
Since there are at most 2/ǫ bins per queue, we can bound the total migration of Insert({i}, bS(ℓ), Qk)
by 7· 2/ǫ ∈ O(1/ǫ). Note also that s(i) ≤ ǫ/14 for every i implies that insert(S≤ℓ, bx) is bounded
by ǫ/2 for all x and ℓ .
Suppose that items i1, . . . , in of size interval Sℓ+y have to be removed from bin bx. In order
to fill the emerging free space, items from the same size category are moved out of bS(ℓ) into
the free space. As the bin bx may already have additional free space, we need to move at most
a size of size(i1, . . . , in) + ǫ/2ℓ+y. Using a symmetric proof as above yields a migration factor of
O( 1
ǫ ).
4.2 Handling small items in the general setting
In the scenario that there are mixed item types (small and large items), we need to be more
careful in the creation and the deletion of buffer bins. To maintain the approximation guarantee,
we have to make sure that as long as there are bins containing only small items, the remaining
free space of all bins can be bounded. Packing small items into empty bins and leaving bins
with large items untouched does not lead to a good approximation guarantee as the free space
of the bins containing only large items is not used. In this section we consider the case where a
sequence of small items is inserted or deleted. We assume that the packing of large items does
not change. Therefore the number of bins containing large items equals a fixed constant Λ(B).
In the previous section, the bins b1, . . . , bm(B) all had a capacity of 1. In order to handle a mixed
28
setting, we will treat a bin bi containing large items as having capacity of c(bi) = 1 − S, where
S is the total size of the large items in bi. The bins containing small items are enumerated
by b1, . . . , bL(B), bL(B)+1, . . . , bm(B) for some L(B) ≤ m(B) where c(b1), . . . , c(bL(B)) < 1 and
c(bL(B)+1) = . . . = c(bm(B)) = 1. Additionally we have a separate set of bins, called the heap
bins, which contain only large items. This set of bins is enumerated by h1, . . . hh(B). Note
that L(B) + h(B) = Λ(B).
In general we may consider only bins bi and hi with capacity
c(bi) ≥ ǫ/14 and c(hi) ≥ ǫ/14 since bins with less capacity are already packed well enough for our
approximation guarantee as shown by Lemma 9. Therefore, full bins are not considered in the
following.
Only large items
Large and small items
Only small items
. . .
. . .
. . .
. . .
. . .
. . .
. . .
h1
h2
hh
b1
Q1
bb1
b
bbℓ = bL
bL+1
Qℓ
Qℓ+1
bbℓ+1
b
bbd = bm
Qd
heap bins
Figure 12: Distribution of bins
As before, we partition the bins b1, . . . , bL(B), bL(B)+1, . . . , bm(B) into several different queues
Q1, . . . , Qℓ(B), Qℓ(B)+1, . . . , Qd(B) such that b1, . . . bL(B) = Q1, . . . Qℓ(B) and bL(B)+1, . . . bm(B) =
Qℓ(B)+1, . . . , Qd(B). If the corresponding packing B is clear from the context, we will simply
write h, L, ℓ, d, m, Λ instead of h(B), L(B), ℓ(B), d(B), m(B), Λ(B). We denote the last bin of
queue Qi by bbi which is a buffer bin. The buffer bin bbℓ is special and will be treated differently
in the insert and delete operation. Note that the bins containing large items b1, . . . , bL(B) are
enumerated first. This guarantees that the free space in the bins containing large items is
used before new empty bins are opened to pack the small items. However, enumerating bins
containing large items first, leads to a problem if according to Algorithm 5 when a buffer bin is
being filled and a new bin has to be inserted right to the filled bin. Instead of inserting a new
empty bin, we insert a heap bin at this position. Since the heap bin contains only large items,
we do not violate the order of the small items (see Figure 12). As the inserted heap bin has
remaining free space (is not filled completely) for small items, it can be used as a buffer bin.
In order to get an idea of how many heap bins we have to reserve for Algorithm 5 where new
bins are inserted or deleted, we define a potential function. As a buffer bin is being filled or
emptied completely the Algorithm 5 is executed and inserts or deletes buffer bins. The potential
function Φ(B) thus bounds the number of buffer bins in Q1, . . . , Qℓ(B) that are about to get
filled or emptied. The potential Φ(B) is defined by
Φ(B) =
ri + ⌈ǫΛ⌉ − ℓ
where the fill ratio ri is defined by ri = s(bbi)
c(bbi) and s(bbi) is the total size of all small items in bbi
. Note that the potential only depends on the queues Q1, . . . , Qℓ(B) and the bins which contain
small and large items. The term ri intends to measure the number of buffer bins that become
full. According to Case 1 of the previous section a new buffer bin is opened when bbi is filled
i=1 ri bounds the number of buffer bins getting filled. The term
ǫΛ in the potential measures the number of bins that need to be inserted due to the length of
a queue exceeding 2/ǫ, as we need to split the queue Qi into two queues of length 1/ǫ according
to Case 1. Each of those queues needs a buffer bin, hence we need to insert a new buffer bin
out of the heap bins. Therefore the potential Φ(B) bounds the number of bins which will be
inserted as new buffer bins according to Case 1.
i. e., ri ≈ 1. Hence the sum Pℓ−1
ℓ−1
Xi=1
29
Just like in the previous section we propose the following properties to bound the approxima-
tion ratio and the migration factor. The first three properties remain the same as in Section
4.1 and the last property gives the desired connection between the potential function and the
heap bins.
(1) For every item i ∈ bd with size s(i) ∈ Sj for some j, d ∈ N, there is no item i′ ∈ bd′ with
size s(i′) ∈ sj ′ such that d′ > d and j′ > j. This means: Items are ordered from left to
right by their size intervals.
(2) Every normal bin of b1, . . . , bm is filled completely
(3) The length of each queue is at least 1/ǫ and at most 2/ǫ except for Qℓ and Qd. The
length of Qℓ and Qd is only limited by 1 ≤ Qℓ,Qd ≤ 1/ǫ. Furthermore, Qℓ+1 = 1 and
1 ≤ Qℓ+2 ≤ 2/ǫ.
(4) The number of heap bins H1, . . . , Hh is exactly h = ⌊Φ(B)⌋
Since bins containing large items are enumerated first, property (1) implies in this setting that
bins with large items are filled before bins that contain no large items. Note also that property
(3) implies that Φ(B) ≥ 0 for arbitrary packings B since ǫΛ ≥ ℓ − 1 + ǫ and thus ⌈ǫΛ⌉ ≥ ℓ. The
following lemma proves that a packing which fulfills properties (1) to (4) provides a solution
that is close to the optimum.
Lemma 10. Let M = m + h be the number of used bins and ǫ ≤ 1/4. If properties (1) to (4)
hold, then at most max{Λ, (1 + O(ǫ)) opt(I, s) + O(1)} bins are used in the packing.
Proof. Case 1: There is no bin containing only small items, i. e., L = m. Hence all items are
packed into M = L + h = Λ bins.
Case 2: There are bins containing only small items, i. e., L < m. Property (3) implies that the
number of queues d is bounded by d ≤ ǫm + 4. Hence the number of buffer bins is bounded by
ǫm+4 and the number of heap bins Φ(B) (property (4)) is bounded by Φ(B) =Pℓ−1
i=1 ri +⌈ǫΛ⌉−
ℓ ≤ ℓ − 1 + ǫΛ + 1 − ℓ = ǫΛ as ri ≤ 1. Since Λ < M , we can bound Φ(B) by Φ(B) < ǫM . The
number of normal bins is thus at least M−(ǫm+5)−(ǫM−1) ≥ M−2ǫM−4 = (1−2ǫ)M−4. By
property (2) every normal bin has less than ǫ/14 free space and the total size S of all items is thus
at least S ≥ (1−ǫ/14)(1−2ǫ)M−4. Since opt(I, s) ≥ S, we have opt(I, s) ≥ (1−ǫ/14(1−2ǫ)M−4.
(1−ǫ/14)(1−2ǫ) ≤ (1 + 5ǫ) for ǫ ≤ 1/4. Therefore we can bound
A simple calculation shows that
the number of used bins by (1 + 5ǫ) opt(I, s) + 4.
1
According to property (4) we have to guarantee, that if the rounded potential ⌊Φ(B)⌋ changes,
the number of heap bins has to be adjusted accordingly. The potential ⌊Φ(B)⌋ might increases
by 1 due to an insert operation. Therefore the number of heap bins has to be incremented.
If the potential ⌊Φ(B)⌋ decreases due to a delete operation, the number of heap bins has to
be decremented. In order to maintain property (4) we have to make sure, that the number of
heap bins can be adjusted whenever ⌊Φ(B)⌋ changes. Therefore we define the fractional part
{Φ(B)} = Φ(B) − ⌊Φ(B)⌋ of Φ(B) and put it in relation to the fill ratio rℓ of bbℓ (the last bin
containing large items) through the following equation:
(1 − rℓ) − {Φ(B)} ≤
s
c(bbℓ)
(Heap Equation)
where s is the biggest size of a small item appearing in bbℓ. The Heap Equation ensures that
the potential Φ(B) is correlated to 1 − rℓ. The values may only differ by the small term s
c(bbℓ) .
30
Note that the Heap Equation can always be fulfilled by shifting items from bbℓ to queue Qℓ+1
or vice versa.
Assuming the Heap Equation holds and the potential ⌊Φ(B)⌋ increases by 1, we can guarantee
that buffer bin bbℓ is nearly empty. Hence the remaining items can be shifted to Qℓ+1 and bbℓ
can be moved to the heap bins. The bin left of bbℓ becomes the new buffer bin of Qℓ. Vice
versa, if ⌊Φ(B)⌋ decreases, we know by the Heap Equation that bbℓ is nearly full, hence we
can label bbℓ as a normal bin and open a new buffer bin from the heap at the end of queue
Qℓ. Our goal is to ensure that the Heap Equation is fulfilled at every step of the algorithm
along with properties (1) to (4). Therefore we enhance the delete and insert operations from
the previous section. Whenever a small item i is inserted or removed, we will perform the
operations described in Algorithm 4 (which can be applied to bins of different capacities) in
the previous section. This will maintain properties (1) to (3). If items are inserted or deleted
from queue Qℓ (the last queue containing large and small items) the recursion does not halt at
bbℓ. Instead the recursion goes further and halts at bbℓ+1. So, when items are inserted into bin
bbℓ according to Algorithm 4 the bin bbℓ is treated as a normal bin. Items are shifted from bbℓ
to queue Qℓ+1 until the Heap Equation is fulfilled. This way we can make sure that the Heap
Equation maintains fulfilled whenever an item is inserted or removed from Qℓ.
Algorithm 6 (Insert or Delete small items for the mixed setting).
Insert(i, bx, Qj):
• Use Algorithm 4 to insert item i into Qj with j < ℓ.
• Let i1, . . . , im be the items that are inserted at the last step of Algorithm 4 into bbj.
• For k = 1, . . . , m do
1. Insert item ik into bin bbj.
2. If bbj is completely filled use Algorithm 5.
3. If the potential ⌊Φ(B)⌋ increases use Algorithm 7 (see below) to adjust the number
4. Decrease the fill ratio rℓ of bbℓ by shifting the smallest items in bbℓ to Qℓ+1 until
of heap bins (property (4)).
(1 − rℓ) ≤ {Φ(B)} to fulfill the Heap Equation.
Delete(i, bx, Qj):
• Use Algorithm 4 to remove item i from bin bx in queue Qj with j < ℓ.
• Let i1, . . . , im be the items that are removed at the last step of Algorithm 4 from bbj.
• For k = 1, . . . , m do
1. If bbj is empty use Algorithm 5.
2. Remove item ik from bin bbj.
3. If the potential ⌊Φ(B)⌋ decreases use Algorithm 7.
4. Increase the fill ratio rℓ of bbℓ by shifting the smallest items in bbℓ to Qℓ+1 until
(1 − rℓ) ≥ {Φ(B)} to fulfill the Heap Equation.
For the correctness of step 4 (the adjustment to rℓ) note the following: In case of the insert
operation, the potential Φ(B) increases and we have Φ(B) ≥ 1 − rℓ. As items are being shifted
from bbℓ to Qℓ+1, the first time that (1 − rℓ) ≤ {Φ(B)} is fulfilled, the Heap Equation is also
c(bbℓ) as an item (which has size at most
fulfilled. Since the fill ratio of bbℓ changes at most by
s) is shifted to Qℓ+1 we know that (1 − rℓ) − {Φ(B)} ≤ s
c(bbℓ) . Correctness of step 4 in the
delete operation follows symmetrically.
s
The potential Φ(B) changes if items are inserted or deleted into queues Q1, . . . , Qℓ−1. Due
to these insert or delete operations it might happen that the potential ⌊Φ(B)⌋ increases or that
a buffer bin is being filled or emptied. The following operation is applied as soon as an item is
inserted or deleted into a buffer bin and the potential ⌊Φ(B)⌋ increases or decreases.
31
Algorithm 7 (Change in the potential).
• Case 1: The potential ⌊Φ(B)⌋ increases by 1.
– According to the Heap Equation the remaining size of small items in bbℓ can be
bounded. Shift all small items from bbℓ to Qℓ+1.
Qℓ is labeled as a buffer bin.
– If Qℓ > 1 then label the now empty buffer bin bbℓ as a heap bin and the last bin in
– If Qℓ only consists of the buffer bin (i. e., Qℓ = 1) shift items from bbℓ−1 to Qℓ+1
until the heap equation is fulfilled. If bbℓ−1 becomes empty remove bbℓ−1 and bbℓ. The
bin left to bbℓ−1 becomes the new buffer bin of Qℓ−1. The queue Qℓ is deleted and
Qℓ−1 becomes the new last queue containing large items.
• Case 2: The potential ⌊Φ(B)⌋ decreases by 1.
– According to the Heap Equation the remaining free space in bbℓ can be bounded. Shift
items from bbℓ+1 to bbℓ such that the buffer bin bbℓ is filled completely.
– Add the new buffer bin from the heap to Qℓ.
– If Qℓ = 1/ǫ label an additional heap bin as a buffer bin to create a new queue Qℓ+1
with Qℓ+1 = 1.
Like in the last section we also have to describe how to handle buffer bins that are being
emptied or filled completely. We apply the same algorithm when a buffer bin is being emp-
tied or filled but have to distinguish now between buffer bins of Q1, . . . , Qℓ and buffer bins of
Qℓ+1, . . . , Qd. Since the buffer bins in Qℓ+1, . . . , Qd all have capacity 1, we will use the same
technique as in the last section. If a buffer bin in Q1, . . . , Qℓ is emptied or filled we will also
use similar technique. But instead of inserting a new empty bin as a new buffer bin, we take
an existing bin out of the heap. And if a buffer bin from Q1, . . . Qℓ is being emptied (it still
contains large items), it is put into the heap. This way we make sure that there are always
sufficiently many bins containing large items which are filled completely.
Lemma 11. Let B be an packing which fulfills the properties (1) to (4) and the Heap Equation.
Applying Algorithm 7 or Algorithm 5 on B during an insert/delete operation yields an packing
B′ which also fulfills properties (1) to (4). The migration to fulfill the Heap Equation is bounded
by O(1/ǫ).
Proof. Analysis of Algorithm 7
Properties (1) and (2) are never violated by the algorithm because the items are only moved by
shift operations. Property (3) is never violated because no queue (except for Qℓ) exceeds 2/ǫ or
falls below 1/ǫ by construction. Algorithm 7 is called during an insert or delete operation. The
Algorithm is executed as items are shifted into or out of buffer bbj such that ⌊Φ(B)⌋ changes.
In the following we prove property (4) for the packing B′ assuming that ⌊Φ(B)⌋ = h(B) holds
by induction. Furthermore we give a bound for the migration to fulfill the heap equation:
• Case 1: The potential ⌊Φ(B)⌋ increases during an insert operation, i. e., it holds ⌊Φ(B′)⌋ =
⌊Φ(B)⌋ + 1. Let item i∗ be the first item that is shifted into a bin bbj such that ⌊Φ(B) +
r∗⌋ = ⌊Φ(B′)⌋, where r∗ is the fill ratio being added to bbj by item i∗. In this situation,
the fractional part changes from {Φ(B)} ≈ 1 to {Φ(B′)} ≈ 0.
– In the case that Qℓ > 1, the buffer bin bbℓ is being emptied and moved to the heap
bins. The bin left of bbℓ becomes the new buffer bin bb′
ℓ of Qℓ. Hence the number of
heap bins increases and we have h(B′) = h(B) + 1 = ⌊Φ(B)⌋ + 1 = ⌊Φ(B′)⌋, which
implies property (4).
To give a bound on the total size of items needed to be shifted out of (or into) bin
ℓ)−{Φ(B′)} by some term
bbℓ to fulfill the heap equation, we bound the term (1− r′
32
ℓ is the fill ratio of bb′
C ≤ O(s(i)/ǫ), where r′
ℓ and s(i) is the size of the arriving or
departing item. If the term (1− r′
ℓ)−{Φ(B′)} can be bounded by C, the fill ratio of
bb′
ℓ has to be adjusted to fulfill the heap equation according to the insert and delete
operation. This can be done be shifting a total size of at most C items out of (or
into) bb′
ℓ.
The bin bb′
ℓ is completely filled by property (3) and therefore has a fill ratio of
ℓ ≥ c(bbℓ)−s
r′
2k is the largest size of a small item appearing
ℓ. Let k′ be the largest size
in bbℓ and Sk is the largest size category appearing in bb′
category appearing in bin bbj. As the bin bb′
ℓ is right of bbj we know k ≤ k′ (property
. Using that {Φ(B′)} ≤ r∗ ≤ 2s(i∗)/ǫ,
(1)) and hence s ≤ 2s(i∗). We get r′
we can bound (1 − r′
ǫ + 2s(i∗)/ǫ = O(s(i∗)/ǫ). Hence the Heap
Equation can be fulfilled by shifting items of total size O(s(i∗)/ǫ) at the end of the
insert operation.
ℓ) − {Φ(B′)} by 4 s(i∗)
ℓ ≥ 1− 4 s(i∗)
c(bbℓ) ≥ 1 − 2 s
ǫ , where s ≤ ǫ
ǫ
– If Qℓ = 1 a set of items in the buffer bin bbℓ−1 is shifted to Qℓ+1 to fulfill the
Heap Equation. Since items are being removed from bbℓ−1 the potential decreases. If
rℓ−1 > {Φ(B′)}, there are enough items which can be shifted out of bbℓ−1 such that
we obtain a new potential Φ(B′′) < Φ(B′)−{Φ(B′)}. Hence ⌊Φ(B′′)⌋ = ⌊Φ(B)⌋ and
the Heap Equation is fulfilled.
Note that the size of items that are shifted out of bbℓ−1 is bounded by r∗ + s =
O(s(i∗)/ǫ), where s is the biggest size of an item appearing in bbℓ−1.
If rℓ−1 ≤ {Φ(B′)} all items are shifted out of bbℓ−1. As the number of queues
decreases, we obtain the new potential Φ(B′′) = Φ(B′) − rℓ−1 + 1 = ⌊Φ(B′)⌋ +
{Φ(B′)}−rℓ−1+1 ≥ ⌊Φ(B′)⌋+1. Hence ⌊Φ(B′′)⌋ = ⌊Φ(B)⌋+2. The buffer bins bbℓ−1
and bbℓ are moved to the heap and thus h(B′′) = h(B) + 2 = ⌊Φ(B)⌋ + 2 = ⌊Φ(B′′)⌋
(property (4)).
Note that if rℓ−1 ≤ {Φ(B′)}, item i∗ is not inserted into bin bbℓ−1 as rℓ−1 ≥ r∗ >
{Φ(B′)}. Therefore the bin bbj is left of bbℓ−1 and we can bound the fill ratio of the
. Using {Φ(B′′)} ≤ r∗ = O(s(i∗)/ǫ) the heap
bin left of bbℓ−1 called r′′
equation can be fulfilled by shifting items of total size O(s(i)/ǫ) at the end of the
insert operation.
ℓ by 1 − 2 s(i∗)
ǫ
• Case 2: The potential ⌊Φ(B)⌋ decreases during a delete operation, i. e., it holds ⌊Φ(B′)⌋ =
⌊Φ(B)⌋ − 1 = ⌊Φ(B) − r∗⌋, where r∗ is the fill ratio being removed from a buffer bin bbj
due to the first shift of an item i∗ that decreases the potential.
According to Algorithm 7, buffer bin bbℓ is being filled completely and a new buffer bin
for Qℓ is inserted from the heap. Hence the number of heap bins decreases and we have
⌊Φ(B′)⌋ = h(B) − 1 = h(B′).
As ⌊Φ(B)⌋ − 1 = Φ(B) − {Φ(B)} − 1 = ⌊Φ(B) − r∗⌋, it holds that {Φ(B)} ≤ r∗ and by
the heap equation the fill ratio of bbℓ is rℓ ≥ r∗ + s, where s is the largest size of a small
item in bbℓ. As above, r∗ and s can be bounded by O( s(i∗)
ǫ ). Hence the total size that is
shifted from Qℓ+1 into bin bbℓ can be bounded by O( s(i∗)
ǫ ).
Furthermore {Φ(B′)} ≥ 1− r∗ (as Φ(B′) = Φ(B)− r∗) and r′
(1 − r′
a total size of at most O(s(i∗)/ǫ) items.
In the case that Qℓ = 1/ǫ a new queue Qℓ+1 is created which consists of a single buffer bin
(inserted from the heap), which does not contain small items, i. e., h(B′′) = h(B′) − 1 =
ℓ = 0, therefore we can bound
ℓ) − {Φ(B′)} by r∗ ≤ O(s(i∗)/ǫ) and the Heap Equation can be fulfilled by shifting
33
h(B)−2, where B′′ is the packing after the insertion of item i∗. Let Φ(B′′) be the potential
after the queue Qℓ+1 is created. Then Φ(B′′) =Pℓ(B ′′)−1
ri +
ǫΛ−ℓ(B′)−1 = Φ(B′)−1, as the buffer bin bbℓ is now counted in the potential, but does not
contain any small items and thus r′′
ℓ = 0. Hence Φ(B′′) = Φ(B′)− 1 = h(B′)− 1 = h(B′′).
ri + ǫΛ − ℓ(B′′) =Pℓ(B ′)−2
i=1
i=1
Analysis of Algorithm 5
Algorithm 5 is executed as an item i∗ is moved into a buffer bin bbj such that bbj is completely
filled or Algorithm 5 is executed if the buffer bin bbj is emptied by moving the last item i∗ out
of the bin. As in the analysis of Algorithm 7, properties (1) and (2) are never violated by the
algorithm because the items are only moved by shift operations. Property (3) is never violated
because no queue (except for Qℓ) exceeds 2/ǫ or falls below 1/ǫ by construction.
It remains to prove property (4) and a bound for the migration to fulfill the heap equation:
• Case 1: An item i∗ is moved into the buffer bin bbj such that bbj is filled completely for
some j < ℓ. According to Algorithm 5 a bin is taken out of the heap and labeled as
the new buffer bin bb′
j = 0 of queue Qj, i. e., the number of heap bins
decreases by 1. Let Φ(B) be the potential before Algorithm 5 is executed and let Φ(B′)
be the potential after Algorithm 5 is executed. The potential changes as follows:
j with fill ratio r′
Φ(B) − Φ(B′) = (rj − r′
j) − (ℓ(B) − ℓ(B′))
Since r′
as the splitting of queue is handled later on).
j = 0 the new potential is Φ(B′) = Φ(B) − rj ≈ Φ(B) − 1 (assuming ℓ(B) = ℓ(B′),
– If ⌊Φ(B′)⌋ = ⌊Φ(B)⌋ − 1 property (4) is fulfilled since the number of heap bins
decreases by h(B′) = h(B) − 1 = ⌊Φ(B)⌋ − 1 = ⌊Φ(B′)⌋. As rj ≥ c(bbj )−s
c(bbj ) , where
s is the biggest size category appearing in bbj and s ≤ 2s(i∗), we obtain for the
fractional part of the potential that {Φ(B)} − {Φ(B′)} ≤ 2 s
. Hence the
Heap Equation can be fulfilled by shifting items of total size O(s(i∗)/ǫ) at the end of
the insert operation as in the above proof.
ǫ ≤ 4 s(i∗)
ǫ
– In the case that ⌊Φ(B′)⌋ = ⌊Φ(B)⌋ = ⌊Φ(B) − rj⌋ we know that the fractional part
changes by {Φ(B′)} = {Φ(B)} − rj. Since the bin bbj is filled completely we know
that rj ≥ c(bbj )−s
c(bbj ) ≈ 1 and hence {Φ(B)} ≥ rj ≈ 1 and {Φ(B′)} ≤ 1 − rj ≈ 0.
According to the Heap Equation, items have to be shifted out of rℓ such that the fill
ratio rℓ changes from rℓ ≤ 1 − rj to rℓ ≈ 1. Therefore we know that as items are
shifted out of bbℓ to fulfill the Heap Equation, the buffer bin bbℓ is being emptied and
moved to the heap (see Algorithm 7). We obtain for the number of heap bins that
h(B′) = h(B) + 1 − 1 = h(B) and hence h(B′) = ⌊Φ(B′)⌋ (property (4)).
, the Heap Equation implies that rℓ ≤ 4 s(i∗)
As {Φ(B)} ≥ rj ≥ 1 − 4 s(i∗)
c(bbℓ) =
O(s(i∗)/ǫ). The buffer bin bbℓ is thus emptied by moving a size of O(s(i∗)/ǫ) items
out of the bin. Let bb′
ℓ be the new buffer bin of Qℓ that was left of bbℓ. The Heap
Equation can be fulfilled by shifting at most O(s(i)/ǫ) out of bb′
ℓ since {Φ(B′)} is
bounded by 1 − rj = O(s(i∗)/ǫ).
ǫ + s
– In the case that Qj > 2/ǫ the queue is split into two queues and an additional
heap bin is inserted, i. e., h(B′′) = h(B′) − 1. As the potential changes by Φ(B′′) =
Φ(B′) + (ℓ(B′) − ℓ(B′′)) = Φ(B′) − 1 we obtain again that h(B′′) = ⌊Φ(B′′)⌋.
ǫ
• Case 2: Algorithm 5 is executed if bin bbj is emptied due to the removal of an item i∗ as
a result of a Delete(i, bx, Qj) call. According to Algorithm 5, the emptied bin is moved
34
to the heap, i. e., the number of heap bins increases by 1. Depending on the length of Qj
and Qj+1, the bin right of bbj or the bin left of bbj is chosen as the new buffer bin bb′
j.
The potential changes by Φ(B′) = Φ(B) + r′
j as in case 1.
– If ⌊Φ(B′)⌋ = ⌊Φ(B)⌋ + 1 property (4) is fulfilled since the number of heap bins
j is the fill ratio of bb′
j, where r′
ǫ ≤ 4 s(i)
j ≥ 1 − 2 s
j is completely filled, the fill ratio is bounded by r′
increases by h(B′) = h(B) + 1.
As bin bb′
ǫ , where s
is the largest size appearing in bb′
j. Since the bin bx has to be left of bbj we know
that s ≤ 2s(i). We obtain for the fractional part of the potential that {Φ(B)} ≥
{Φ(B′)} − 2 s
ǫ . Hence the Heap Equation can be fulfilled by shifting items of
total size O(s(i)/ǫ) at the end of the remove operation.
– In the case that ⌊Φ(B′)⌋ = ⌊Φ(B)⌋ = ⌊Φ(B) + r′
j⌋ we know that the fractional
part changes similar to case 1 by {Φ(B′)} = {Φ(B)} + r′
j. Since the bin bbj is
filled completely we know that rj ≥ c(bbj )−s
c(bbj ) ≈ 1 and hence {Φ(B′)} ≥ rj ≈ 1 and
{Φ(B)} ≤ 1 − rj ≈ 0. According to the Heap Equation items have to be shifted to
bbℓ such that the fill ratio rℓ changes from rℓ ≈ 0 to rℓ ≈ 1. Therefore we know that
as items are shifted into bbℓ to fulfill the Heap Equation, bbℓ is filled completely and
a bin from the heap is labeled as the new buffer bin of Qℓ (see Algorithm 7). We
obtain for the number of heap bins that h(B′) = h(B) − 1 + 1 = h(B) and hence
h(B′) = Φ(B′) (property (4)). The Heap Equation can be fulfilled similarly to case
1 by shifting items of total size O(s(i)/ǫ).
Using the above lemma for, we can finally prove the following central theorem, which states
that the migration of an insert/delete operation is bounded and that properties (1) to (4) are
maintained.
Theorem 6.
(i) Let B be a packing which fulfills properties (1) to (4) and the Heap Equation. Applying
operations insert(i, bx, Qj) or delete(i, bx, Qj) on a packing B yields an instance B′ which
also fulfills properties (1) to (4) and the Heap Equation.
(ii) The migration factor of an insert/delete operation is bounded by O(1/ǫ).
Proof. Suppose a small item i with size s(i) is inserted or deleted from queue Qj. The insert
and delete operation basically consists of application of Algorithm 4 and iterated use of steps
(1) to (3) where Algorithms 5 and 7 are used and items in bbℓ are moved to Qℓ+1 and vice versa.
Let B be the packing before the insert/delete operation and let B′ be the packing after the
operation.
Proof for (i): Now suppose by induction that property (1) to (4) and the Heap Equation is
fulfilled for packing B. We prove that property (4) and the Heap Equation maintain fulfilled
after applying an insert or delete operation on B resulting in the new packing B′. Properties
(1) to (3) hold by conclusion of Lemma 9 and Lemma 11. Since the potential and the number of
heap bins only change as a result of Algorithm 5 or Algorithm 7, property (4) maintains fulfilled
also. By definition of step 4 in the insert operation, items are shifted from bbℓ to Qℓ+1 until
the Heap Equation is fulfilled. By definition of step 4 of the delete operation, the size of small
items in bbℓ is adjusted such that the Heap Equation is fulfilled. Hence the Heap Equation is
always fulfilled after application of Insert(i, bx, Qj) or Delete(i, bx, Qj).
35
Proof for (ii): According to Lemma 9 the migration factor of the usual insert operation is
bounded by O(1/ǫ). By Lemma 11 the migration in Algorithm 5 and Algorithm 7 is also bounded
by O(1/ǫ). It remains to bound the migration for step 4 in the insert/delete operation. Therefore
we have to analyze the total size of items to be shifted out or into bbℓ in order to fulfill the Heap
Equation.
Since the size of all items i1, . . . , ik that are inserted into bbj is bounded by 7s(i) (see Lemma
9) and the capacity of bbj is at least ǫ/14 the potential Φ(B) changes by at most O(s(i)/ǫ). By
Lemma 11 the size of items that needs to be shifted out or into bbℓ as a result of Algorithm 5
or 7 is also bounded by O(s(i)/ǫ). Therefore the size of all items that need to be shifted out or
into bbℓ in step (4) of the insert/delete operation is bounded by O(s(i)/ǫ).
Shifting a size of O(s(i)/ǫ) to Qℓ+1 or vice versa leads to a migration factor of O(1/ǫ2) (Lemma
9). Fortunately we can modify the structure of queues Qℓ+1 and Qℓ+2 such that we obtain a
smaller migration factor. Assuming that Qℓ+1 consists of a single buffer bin, i. e., Qℓ+1 = 1
items can directly be shifted from bbℓ to bbℓ+1 and therefore we obtain a migration factor of
O(1/ǫ). A structure with Qℓ+1 = 1 and 1 ≤ Qℓ+2 ≤ 2/ǫ (see property (3)) can be maintained
by changing Algorithm 5 in the following way:
• If bbℓ+1 is filled completely, move the filled bin to Qℓ+2.
– If Qℓ+2 > 2/ǫ, split Qℓ+2 into two queues.
• If bbℓ+1 is being emptied, remove the bin and label the first bin of Qℓ+2 as bbℓ+1.
– If Qℓ+2 = 0, remove Qℓ+2.
4.3 Handling the General Setting
In the previous section we described how to handle small items in a mixed setting. It remains
to describe how large items are handled in this mixed setting. Algorithm 1 describes how to
handle large items only. However, in a mixed setting, where there are also small items, we have
to make sure that properties (1) to (4) and the Heap Equation maintain fulfilled as a large item
is inserted or deleted. Algorithm 1 changes the configuration of at most O(1/ǫ2 · log 1/ǫ) bins
(Theorem 5). Therefore, the size of large items in a bin b (= 1 − c(b)) changes, as Algorithm
1 may increase or decrease the capacity of a bin. Changing the capacity of a bin may violate
properties (2) to (4) and the Heap Equation. We describe an algorithm to change the packing
of small items such that all properties and the Heap Equation are fulfilled again after Algorithm
1 was applied.
The following algorithm describes how the length of a queue Qj is adjusted if the length Qj
falls below 1/ǫ:
Algorithm 8 (Adjust the queue length).
• Remove all small item IS from bbj and add bbj to the heap.
• Merge Qj with Qj+1. The merged queue is called Qj.
• If Qj > 2/ǫ split queue Qj by adding a heap bin in the middle.
• Insert items IS using Algorithm 6.
The following algorithm describes how the number of heap bins can be adjusted.
36
Algorithm 9 (Adjust number of heap bins).
• Decreasing the number of heap bins by 1.
– Shift small items from Qℓ+1 to bbℓ until bbℓ is filled completely
– Label a heap bin as the new buffer bin of Qℓ
• Increasing the number of heap bins by 1.
– Shift all small items from bbℓ to Qℓ+1
– Label bbℓ as a heap bin
– Label the bin left of bbℓ as new buffer bin of Qℓ
Note that the Heap Equation can be fulfilled in the same way, by shifting items from bbℓ to
Qℓ+1 or vice versa.
Using these algorithms, we obtain our final algorithm for the fully dynamic binpacking prob-
lem.
Algorithm 10 (AFPTAS for the mixed setting).
• If i is large do
1. Use Algorithm 1.
2. Remove all small items IS of bins b with changed capacity.
3. Adjust queue length.
4. Adjust the number of heap bins.
5. Adjust the Heap Equation.
6. Insert all items IS using Algorithm 6.
• If i is small use Algorithm 6
Combining all the results from the current and the previous section, we finally prove the
central result that there is fully dynamic AFPTAS for the binpacking problem with polynomial
migration.
Theorem 7. Algorithm 10 is a fully dynamic AFPTAS for the binpacking problem, that
achieves a migration factor of at most O(1/ǫ4 · log 1/ǫ) by repacking items from at most O(1/ǫ3 ·
log 1/ǫ) bins.
Proof. Approximation guarantee: By definition of the algorithm,
it generates at every
timestep t a packing Bt of instance I(t) such that properties (1) to (4) are fulfilled. According to
Lemma 10, at most max{Λ, (1 +O(ǫ)) opt(I(t), s) +O(1)} bins are used where Λ is the number
of bins containing large items. Since we use Algorithm 1 to pack the large items, Theorem 5
implies that Λ ≤ (1 + O(ǫ)) opt(I(t), s) + O(1/ǫ log 1/ǫ). Hence the number of used bins can be
bounded in any case by (1 + O(ǫ)) opt(I(t), s) + O(1/ǫ log 1/ǫ).
Migration Factor: Note that the Algorithm uses Algorithm 6 or Algorithm 1 to insert and
delete small or large items. The migration factor for Algorithm 6 is bounded by O(1/ǫ) due
to Theorem 6 while the migration factor for Algorithm 1 is bounded by O(1/ǫ3 · log 1/ǫ) due to
Theorem 5.
It remains to bound the migration that is needed to adjust the heap bins, the length of a
queue falling below 1/ǫ and the Heap Equation in case a large item arrives and Algorithm 1 is
applied.
37
Suppose the number of heap bins has to be adjusted by 1. In this case Algorithm 9 shifts
items from Qℓ+1 to bbℓ or vice versa until bbℓ is either filled or emptied. Hence, the size of moved
items is bounded by 1. Since the size of the arriving or departing item is ≥ ǫ/14 the migration
factor is bounded by O(1/ǫ). In the same way, a migration of at most O(1/ǫ) is used to fulfill
the Heap Equation which implies that the migration in step 5 is bounded by O(1/ǫ).
If Qj falls below 1/ǫ, the two queues Qj and Qj+1 are merged by emptying bbj. The removed
items are inserted by Algorithm 6. As their total size is bounded by 1 and the algorithm has a
migration factor of O(1/ǫ), the size of the moved items is bounded by O(1/ǫ). The migration to
merge two queues can thus be bounded by O(1/ǫ2).
Note that the proof of Theorem 5 implies that at most γ = O(1/ǫ2 log 1/ǫ) bins are changed
by Algorithm 1. The total size of the items IS which are removed in step 2 is thus bounded by
γ. Similarly, the length of at most γ queues can fall below 1/ǫ. The migration of step 3 is thus
bounded by γ · 1/ǫ2. As at most γ buffer bins are changed, the change of the potential (and thus
the number of heap bins) is also bounded by γ and the migration in step 4 can be bounded by
γ · 1/ǫ. The migration in step 6 is bounded by s(IS) · 1/ǫ ≤ γ · 1/ǫ as Algorithm 6 has migration
factor 1/ǫ. The total migration of the adjustments is thus bounded by γ · 1/ǫ2 = O(1/ǫ4 log 1/ǫ).
Running Time: The handling of small items can be performed in linear time while the
handling of large items requires O(M (1/ǫ log(1/ǫ)) · 1/ǫ3 log(1/ǫ) + 1/ǫ log(1/ǫ) log(ǫ2 · n(t)) + ǫn(t)),
where M (n) is the time needed to solve a system of n linear equations (see Theorem 5). The
total running time of the algorithm is thus O(M (1/ǫ log(1/ǫ)) · 1/ǫ3 log(1/ǫ) + 1/ǫ log(1/ǫ) log(ǫ2 ·
n(t)) + n(t)).
Acknowledgements
We would like to thank Till Tantau for his valuable comments and suggestions to improve the
presentation of the paper.
References
[BB10] A. Beloglazov and R. Buyya. Energy efficient allocation of virtual machines in cloud
data centers. In 10th IEEE/ACM International Conference on Cluster, Cloud and
Grid Computing, CCGrid 2010, pages 577–578, 2010.
[BBG10] J. Balogh, J. B´ek´esi, and G. Galambos. New lower bounds for certain classes
In Workshop on Approximation and Online Algo-
of bin packing algorithms.
rithms(WAOA), volume 6534 of LNCS, pages 25–36, 2010.
[BBGR08] J. Balogh, J. B´ek´esi, G. Galambos, and G. Reinelt. Lower bound for the online bin
packing problem with restricted repacking. SIAM Journal on Computing, 38(1):398–
410, 2008.
[BKB07] N. Bobroff, A. Kochut, and K.A. Beaty. Dynamic placement of virtual machines
for managing SLA violations. In Integrated Network Management, IM 2007. 10th
IFIP/IEEE International Symposium on Integrated Network Management, pages
119–128, 2007.
[BM98] P.A.
trix
Computer Science, 205(1–2):307–316, 1998.
multiplication
Megiddo.
basic
to
find
Beling
and
N.
solutions.
Using
fast
ma-
Theoretical
38
[Bro79] D.J. Brown. A lower bound for on-line one-dimensional bin packing algorithms.
Technical Report R-864, Coordinated Sci Lab Univ of Illinois Urbana, 1979.
[CGJ83] E.G. Coffman, M.R. Garey, and D.S. Johnson. Dynamic bin packing. SIAM Journal
on Computing, 12(2):227–258, 1983.
[CLW08] J.W. Chan, T. Lam, and P.W.H. Wong. Dynamic bin packing of unit fractions
items. Theoretical Computer Science, 409(3):521–529, 2008.
[CWY09] J.W. Chan, P.W.H. Wong, and F.C.C. Yung. On dynamic bin packing: An im-
proved lower bound and resource augmentation analysis. Algorithmica, 53(2):172–
206, 2009.
[DKL14] Khuzaima D., Shahin K., and Alejandro L. On the online fault-tolerant server
consolidation problem. In 26th ACM Symposium on Parallelism in Algorithms and
Architectures, SPAA ’14, pages 12–21, 2014.
[Eis57] K. Eisemann. The Trim Problem. Management Science, 3(3):279–284, 1957.
[EL09] L. Epstein and A. Levin. A robust APTAS for the classical bin packing problem.
Mathematical Programming, 119(1):33–49, 2009.
[EL13] L. Epstein and A. Levin. Robust approximation schemes for cube packing. SIAM
Journal on Optimization, 23(2):1310–1343, 2013.
[FdlVL81] W. Fernandez de la Vega and G.S. Lueker. Bin packing can be solved within 1 + ǫ
in linear time. Combinatorica, 1(4):349–355, 1981.
[GPT00] G.
Gambosi,
A.
Postiglione,
and M.
Talamo.
Algo-
rithms
SIAM Journal on Computing, 30(5):1532–1551, 2000.
relaxed
online
the
for
bin-packing
model.
[IL97] Z. Ivkovi´c and E.L. Lloyd. Partially dynamic bin packing can be solved within 1 + ǫ
in (amortized) polylogarithmic time. Information Processing Letter, 63(1):45–50,
1997.
[IL98] Z.
for
Journal on Computing, 28(2):574–611, 1998.
and
packing:
Ivkovi´c
bin
Lloyd.
Being
E.L.
(mostly) myopic
helps.
Fully
dynamic
algorithms
SIAM
[IL09] Z. Ivkovi´c and E.L. Lloyd. Fully dynamic bin packing. In Fundamental Problems
in Computing, pages 407–434. Springer, 2009.
[JDU+74a] D.S. Johnson, A. Demers, J.D. Ullman, M.R. Garey, and R.L. Graham. Worst-case
performance bounds for simple one-dimensional packing algorithms. SIAM Journal
on Computing, 3(4):299–325, 1974.
[JDU+74b] D.S. Johnson, A.J. Demers, J.D. Ullman, M.R. Garey, and R.L. Graham. Worst-
case performance bounds for simple one-dimensional packing algorithms. SIAM
Journal on Computing, 3(4):299–325, 1974.
[JJH+08] G. Jung, K.R. Joshi, M.A. Hiltunen, R.D. Schlichting, and C. Pu. Generating
adaptation policies for multi-tier applications in consolidated server environments.
In 2008 International Conference on Autonomic Computing, ICAC 2008, June 2-6,
2008, Chicago, Illinois, USA, pages 23–32, 2008.
39
[JJH+09] G. Jung, K.R. Joshi, M.A. Hiltunen, R.D. Schlichting, and C. Pu. A cost-sensitive
adaptation engine for server consolidation of multitier applications. In Middleware
2009, ACM/IFIP/USENIX, 10th International Middleware Conference, Proceed-
ings, pages 163–183, 2009.
[JK13] K. Jansen and K. Klein. A robust AFPTAS for online bin packing with polynomial
In International Colloquium on Automata, Languages, and Program-
migration.
ming(ICALP), pages 589–600, 2013.
[Joh74] D.S.
Johnson.
Fast
algorithms
for
bin
packing.
Journal of Computer and System Sciences, 8(3):272–314, 1974.
[KK82] N. Karmarkar and R.M. Karp. An efficient approximation scheme for the one-
dimensional bin-packing problem. In 23rd Annual Symposium on Foundations of
Computer Science (FOCS), pages 312–320. IEEE Computer Society, 1982.
[Lia80] F.M. Liang. A lower bound for on-line bin packing. Information processing letters,
10(2):76–79, 1980.
[LL85] C.C. Lee and D. Lee. A simple on-line bin-packing algorithm. Journal of the ACM
(JACM), 32(3):562–572, 1985.
[LTC14] Y. Li, X. Tang, and W. Cai. On dynamic bin packing for resource allocation in the
cloud. In 26th ACM Symposium on Parallelism in Algorithms and Architectures,
SPAA ’14, pages 2–11, 2014.
[PSC+00] J.M. Park, Uday R. Savagaonkar, E.K.P. Chong, H.J. Siegel, and S.D. Jones. Effi-
cient resource allocation for qos channels in mf-tdma satellite systems. In MILCOM
2000. 21st Century Military Communications Conference Proceedings, volume 2,
pages 645–649. IEEE, 2000.
[Sei02] S.S. Seiden. On the online bin packing problem. Journal of the ACM, 49(5):640–671,
2002.
[SKZ08] S. Srikantaiah, A. Kansal, and F. Zhao. Energy aware consolidation for cloud
computing. In Proceedings of the 2008 Conference on Power Aware Computing and
Systems, HotPower’08, pages 10–10, 2008.
[SSS09] P. Sanders, N. Sivadasan, and M. Skutella. Online scheduling with bounded migra-
tion. Mathematics of Operations Research, 34(2):481–498, 2009.
[Sto13] A.L. Stolyar. An infinite server system with general packing constraints. Operations
Research, 61(5):1200–1217, 2013.
[SV10] M. Skutella and J. Verschae. A robust PTAS for machine covering and packing.
In European Symposium on Algorithms(ESA), volume 6346 of LNCS, pages 36–47,
2010.
[SZ13] A.L. Stolyar and Y. Zhong. A large-scale service system with packing con-
In Proceedings of the
on
straints: Minimizing the number of occupied servers.
ACM
Measurement and modeling of computer systems, pages 41–52. ACM, 2013.
SIGMETRICS/international
conference
40
[Ull71] J.D. Ullman. The Performance of a Memory Allocation Algorithm. Technical report.
Princeton University, 1971.
[VAN08] A. Verma, P. Ahuja, and A. Neogi. pmapper: Power and migration cost aware appli-
cation placement in virtualized systems. In Middleware 2008, ACM/IFIP/USENIX
9th International Middleware Conference, Proceedings, pages 243–264, 2008.
[Vli92] A. Vliet. An improved lower bound for on-line bin packing algorithms. Information
Processing Letters, 43(5):277–284, 1992.
[Yao80] A.C. Yao. New algorithms for bin packing. Journal of the ACM (JACM), 27(2):207–
227, 1980.
41
|
1502.02051 | 2 | 1502 | 2015-08-13T12:58:37 | Approximating ATSP by Relaxing Connectivity | [
"cs.DS"
] | The standard LP relaxation of the asymmetric traveling salesman problem has been conjectured to have a constant integrality gap in the metric case. We prove this conjecture when restricted to shortest path metrics of node-weighted digraphs. Our arguments are constructive and give a constant factor approximation algorithm for these metrics. We remark that the considered case is more general than the directed analog of the special case of the symmetric traveling salesman problem for which there were recent improvements on Christofides' algorithm.
The main idea of our approach is to first consider an easier problem obtained by significantly relaxing the general connectivity requirements into local connectivity conditions. For this relaxed problem, it is quite easy to give an algorithm with a guarantee of 3 on node-weighted shortest path metrics. More surprisingly, we then show that any algorithm (irrespective of the metric) for the relaxed problem can be turned into an algorithm for the asymmetric traveling salesman problem by only losing a small constant factor in the performance guarantee. This leaves open the intriguing task of designing a "good" algorithm for the relaxed problem on general metrics. | cs.DS | cs | Approximating ATSP by Relaxing Connectivity
Ola Svensson
EPFL
[email protected]
August 14, 2015
Abstract
The standard LP relaxation of the asymmetric traveling salesman problem has
been conjectured to have a constant integrality gap in the metric case. We prove this
conjecture when restricted to shortest path metrics of node-weighted digraphs. Our
arguments are constructive and give a constant factor approximation algorithm
for these metrics. We remark that the considered case is more general than the
directed analog of the special case of the symmetric traveling salesman problem
for which there were recent improvements on Christofides’ algorithm.
The main idea of our approach is to first consider an easier problem obtained by
significantly relaxing the general connectivity requirements into local connectivity
conditions. For this relaxed problem, it is quite easy to give an algorithm with
a guarantee of 3 on node-weighted shortest path metrics. More surprisingly, we
then show that any algorithm (irrespective of the metric) for the relaxed problem
can be turned into an algorithm for the asymmetric traveling salesman problem by
only losing a small constant factor in the performance guarantee. This leaves open
the intriguing task of designing a “good” algorithm for the relaxed problem on
general metrics.
Keywords: approximation algorithms, asymmetric traveling salesman problem, combinatorial
optimization, linear programming
5
1
0
2
g
u
A
3
1
]
S
D
.
s
c
[
2
v
1
5
0
2
0
.
2
0
5
1
:
v
i
X
r
a
1 Introduction
The traveling salesman problem is one of the most fundamental combinatorial optimiza-
tion problems. Given a set V of n cities and a distance/weight function w : V × V → R+,
it is the problem of finding a tour of minimum total weight that visits each city exactly
once. There are two variants of this general definition: the symmetric traveling salesman
problem (STSP) and the asymmetric traveling salesman problem (ATSP). In the symmet-
ric version we assume w(u, v) = w(v, u) for each pair u, v ∈ V of cities; whereas we make
no such assumption in the more general asymmetric traveling salesman problem.
In both versions, it is common to assume the triangle inequality and we shall do so
in the rest of this paper. Recall that the triangle inequality says that for any triple i, j, k
of cities, we have w(i, j) + w(j, k) (cid:62) w(i, k). In other words, it is not more expensive to
take the direct path compared to a path that makes a detour. Another equivalent view
of the triangle inequality is that, instead of insisting that each city is visited exactly once,
we should find a tour that visits each city at least once. These assumptions are arguably
natural in many, if not most, settings. They are also necessary in the following sense:
any reasonable approximation algorithm (with approximation guarantee O(exp(n))) for
the traveling salesman problem without the triangle inequality would imply P = NP
because it would solve the problem of deciding Hamiltonicity.
Understanding the approximability of the symmetric and the asymmetric traveling
salesman problem (where we have the triangle inequality) turns out to be a much
more interesting and notorious problem. On the one hand, the strongest known
inapproximability results, by Karpinski, Lampis, and Schmied [14], say that it is
NP-hard to approximate STSP within a factor of 123/122 and that it is NP-hard to
approximate ATSP within a factor of 75/74. On the other hand, the current best
approximation algorithms are far from these guarantees, especially in the case of ATSP.
For the symmetric traveling salesman problem, Christofides’ beautiful algorithm
from 1976 sill achieves the best known approximation guarantee of 1.5 [7]. However,
a recent series of papers [11, 16, 17, 19], broke this barrier for the interesting special
case of shortest path metrics of unweighted undirected graphs1. Specifically, Oveis
Gharan, Saberi, and Singh [11] first gave an approximation guarantee of 1.5− ε; Mömke
and Svensson [16] proposed a different approach yielding a 1.461-approximation
guarantee; Mucha [17] gave a tighter analysis of this algorithm; and Sebö and Vygen [19]
significantly developed the approach to give the current best approximation guarantee
of 1.4.
The interest in shortest path metrics has several motivations. It is a natural special
case that seems to capture the difficulty of the problem: it remains APX-hard and the
worst known integrality gap for the Held-Karp relaxation is of this type. Moreover,
it has an attractive graph theoretic formulation: given an unweighted graph, find a
shortest tour that visits each vertex at least once. This is the (unweighted) “graph”
analog of STSP. Indeed, if allow the graph to be edge-weighted, this formulation
is equivalent to STSP on general metrics. Let us also mention that the polynomial
time approximation scheme for the symmetric traveling salesman problem on planar
1The shortest path metric of a graph G = (V, E) is defined as follows: the weight w(u, v) between cities
u, v ∈ V equals the shortest path between u and v in G. If the graph is node-weighted f : V → R+, the
weight/length of an edge {u, v} ∈ E is f (u) + f (v).
2
graphs was first obtained for the special case of unweighted graphs [12], i.e., when
restricted to shortest path metrics of unweighted graphs, and then generalized to the
case of edge-weights [3]. For STSP, it remains a major open problem whether the ideas
in [11, 16, 17, 19] can be applied to general metrics. We further discuss this in Section 6.
The gap in our understanding is much larger for the asymmetric traveling salesman
problem for which it remains a notorious open problem to design an algorithm with any
constant approximation guarantee. This is a particularly intriguing as the standard linear
programming relaxation, often referred to as the Held-Karp relaxation, is only known
to have an integrality gap of at least 2 [6]. There are in general two available approaches
for designing approximation algorithms for ATSP in the literature. The first approach
is due to Frieze, Galbiati, and Maffiolo [9] who gave a log2(n)-approximation algorithm
for ATSP already in 1982. Their basic idea is simple and elegant: a minimum weight
cycle cover has weight at most that of an optimal tour and it will decrease the number
of connected components by a factor of at least 2. Hence, if we repeat the selection of a
minimum weight cycle cover log2(n) times, we get a connected Eulerian graph which
(by shortcutting) is a log2(n)-approximate tour. Although the above analysis is tight
only in the case when almost all cycles in the cycle covers have length 2, it is highly
non-trivial to refine the method to decrease the number of iterations. It was first in 2003
that Bläser [5] managed to give an approximation guarantee of 0.999 log2(n). This was
improved shortly thereafter by Kaplan, Lewenstein, Shafrir and Sviridenko [13] who
further developed this approach to obtain a 4/3 log3(n) ≈ 0.84 log2(n)-approximation
algorithm; and later by Feige and Singh [8] who obtained an approximation guarantee
of 2/3 log2(n).
A second approach was more recently proposed in an influential and beautiful
paper by Asadpour, Goemans, Madry, Oveis Gharan, and Saberi [4] who gave an
O(log n/ log log n)-approximation algorithm for ATSP. Their approach is based on
finding a so-called α-thin spanning tree which is a (unweighted) graph theoretic
problem. Here, the parameter α is proportional to the approximation guarantee so
α = O(log n/ log log n) in [4]. Following their publication, Oveis Gharan and Saberi [10]
gave an efficient algorithm for finding O(1)-thin spanning trees for planar and bounded
genus graphs yielding a constant factor approximation algorithm for ATSP on these
graph classes. Also, in a very recent major progress, Anari and Oveis Gharan [1]
showed the existence of O(polylog log n)-thin spanning trees for general instances.
This implies a O(polylog log n) upper bound on the integrality gap of the Held-Karp
relaxation. Hence, it gives an efficient so-called estimation algorithm for estimating
the optimal value of a tour within a factor O(polylog log n) but, as their arguments are
non-constructive, no approximation algorithm for finding a tour of matching guarantee.
The result in [1] is based on developing and extending several advanced techniques.
Notably, they rely on their extension [2] of the recent proof of the Kadison-Singer
conjecture which was a major breakthrough by Marcus, Spielman, and Srivastava [15].
To summarize, the current best approximation algorithm has a guarantee of
O(log n/ log log n) [4] and the best upper bound on the integrality gap of the Held-Karp
relaxation is O(polylog log n) [1]. These two bounds are far away from the known
inapproximability results [14] and from the lower bound of 2 on the integrality gap of
the Held-Karp relaxation [6]. Moreover, there were no better approximation algorithms
3
known in the case of shortest path metrics of unweighted digraphs for which there was
recent progress in the undirected setting. In particular, it is not clear how to use the two
available approaches mentioned above to get an improved approximation guarantee in
this case: in the cycle cover approach, the main difficulty is to bound the number of
iterations and, in the thin spanning tree approach, ATSP is reduced to an unweighted
graph theoretic problem.
1.1 Our Results and Overview of Approach
We propose a new approach for approximating the asymmetric traveling salesman
problem based on relaxing the global connectivity constraints into local connectivity
conditions. We also use this approach to obtain the following result where we refer to
ATSP on shortest path metrics of node-weighted digraphs as Node-Weighted ATSP.
Theorem 1.1. There is a constant approximation algorithm for Node-Weighted ATSP. Specifi-
cally, for Node-Weighted ATSP, the integrality gap of the Held-Karp relaxation is at most 15
and, for any ε > 0, there is a polynomial time algorithm that finds a tour of weight at most
(27 + ε) OPTHK where OPTHK denotes the optimal value of the Held-Karp relaxation.
As further discussed in Section 6, the constants in the theorem can be slightly
improved by specializing our general approach to the node-weighted case. However,
it remains an interesting open problem to give a tight bound on the integrality gap.
Let us continue with a brief overview of our approach that is not restricted to the
node-weighted version. It is illustrative to consider the following “naive” algorithm
that actually was the starting point of this work:
1. Select a random cycle cover C using the Held-Karp relaxation.
It is well known that one can sample such a cycle cover C of expected weight
equal to the optimal value OPTHK of the Held-Karp relaxation.
2. While there exist more than one component, add the lightest cycle (i.e., the cycle
of smallest weight) that decreases the number of components.
It is clear that the above algorithm always returns a solution to ATSP: we start with
a Eulerian graph2 and the graph stays Eulerian during the execution of the while-
loop which does not terminate until the graph is connected. This gives a tour that
visits each vertex at least once and hence a solution to ATSP (using that we have
the triangle-inequality). However, what is the weight of the obtained tour? First, as
remarked above, we have that the expected weight of the cycle cover is OPTHK. So if
C contains k = C cycles, we would expect that a cycle in C has weight OPTHK /k (at
least on average). Moreover, the number of cycles added in Step 2 is at most k − 1 since
each cycle decreases the number of components by at least one. Thus, if each cycle in
Step 2 has weight at most the average weight OPTHK /k of a cycle in C, we obtain a
2-approximate tour of weight at most OPTHK + k−1
k OPTHK (cid:54) 2 OPTHK.
Unfortunately, it seems hard to find a cycle cover C so that we can always connect it
with light cycles. Instead, what we can do, is to first select a cycle cover C then add light
2Recall that a directed graph is Eulerian if the in-degree equals the out-degree of each vertex.
4
cycles that decreases the number of components as long as possible. When there are no
more light cycles to add, the vertices/cities are partitioned into V1, . . . , Vk connected
components. In order to make progress from this point, we would like to find a “light”
Eulerian set F of edges that crosses the cuts {(Vi, ¯Vi) i = 1, 2, . . . , k}. We could then
hope to add F to our solution and continue from there. It turns out that it is very
important what “light” means in this context. For our arguments to work, we need that
F is selected so that the weight of the edges in each component has weight at most α
times what the linear programming solution “pays” for the vertices in that component.
This is the intuition behind the definitions in Section 3 of Local-Connectivity ATSP and
α-light algorithms for that problem. We also need to be very careful in which way we
add edges from light cycles and how to use the α-light algorithm for Local-Connectivity
ATSP. In Section 5, our algorithm will iteratively solve the Local-Connectivity ATSP
and, in each iteration, it will add a carefully chosen subset of the found edges together
with light cycles.
We remark that in Local-Connectivity ATSP we have relaxed the global connectivity
properties of ATSP into local connectivity conditions that only say that we need to find
a Eulerian set of edges that crosses at most n = V cuts defined by a partitioning of the
vertices. In spite of that, we are able to leverage the intuition above to obtain our main
technical result:
Theorem (Simplified statement of Theorem 5.1). The integrality gap of the Held-Karp
relaxation is at most 5α if there exists an α-light algorithm A for Local-Connectivity ATSP.
Moreover, for any ε > 0, we can find a (9 + ε)α-approximate tour in time polynomial in n, 1/ε,
and in the running time of A.
The proof of the above theorem (Section 5) is based on generalizing and, as alluded to
above, deviating from the above intuition in several ways. First, we start with a carefully
chosen “Eulerian partition” which generalizes the role of the cycle cover C in Step 1
above. Second, both the iterative use of the α-light algorithm for Local-Connectivity
ATSP and the way we add light cycles are done in a careful and dependent manner so as
to be able to bound the total weight of the returned solution. Theorem 1.1 follows from
Theorem 5.1 together with a 3-light algorithm for Node-Weighted Local-Connectivity
ATSP. The 3-light algorithm, described in Section 4, is a rather simple application of
classic theory of flows and circulations. We also remark that it is the only part of the
paper that relies on having shortest path metrics of node-weighted digraphs.
Our work raises several natural questions. Perhaps the most immediate and
intriguing question is whether there is a O(1)-light algorithm for Local-Connectivity
ATSP on general metrics. We further elaborate on this and other related questions in
Section 6.
2 Preliminaries
2.1 Basic Notation
Consider a directed graph G = (V, E). For a subset S ⊆ V, we let δ+(S) = {(u, v) ∈ E
u ∈ S, v (cid:60) S} be the outgoing edges and we let δ−
(S) = {(u, v) ∈ E u (cid:60) S, v ∈ S} be the
5
(cid:48)
E(cid:48)(S) = δ+(S)∩ E
(cid:48)
When considering a function f : U → R, we let f (X) =(cid:80)
(cid:48) ⊆ E of the edges,
incoming edges of the cut defined by S. When considering a subset E
(S)∩ E
and by δ−
(cid:48)
E(cid:48)(S) = δ−
we denote the restrictions to that subset by δ+
.
We also let C(E
) = { G1 = ( V1, E1), G2 = ( V2, E2), . . . , Gk = ( Vk, Ek)} denote the set of
(cid:48)
subgraphs corresponding to the k connected components of the graph (V, E
); the vertex
set V will always be clear from the context. Here connected means that the subgraphs
are connected if we undirect the edges.
x∈X f (x) for X ⊆ U. For
example, if G is edge weighted, i.e., there exists a function w : E → R, then w(E
) denotes
(cid:48) ⊆ E. Similarly, if G is node-weighted, then there
the total weight of the edges in E
exists a function f : V → R and f (S) denotes the total weight of the vertices in S ⊆ V.
When talking about graphs, we shall slightly abuse notation and sometimes write w(G)
instead of w(E) and f (G) instead of f (V) when it is clear from the context that w and f
are functions on the edges and vertices. Finally, our subsets of edges are multisets, i.e., may
contain the same edge several times. The set operators ∪,∩,\ are defined in the natural
way. For example, {e1, e1, e2} ∪ {e1, e2} = {e1, e1, e1, e2, e2},{e1, e1, e2} ∩ {e1, e2} = {e1, e2}, and
{e1, e1, e2} \ {e1, e2} = {e1}. Other sets, such as subsets of vertices, will always be simple
sets without any multiplicities.
(cid:48)
2.2 The (Node-Weighted) Asymmetric Traveling Salesman Problem
It will be convenient to define ATSP using the Eulerian point of view, i.e., we wish to
find a tour that visits each vertex at least once. As already mentioned in the introduction,
this definition is equivalent to that of visiting each city exactly once (in the metric
completion) since we assume the triangle inequality.
ATSP
Given: An edge-weighted (strongly connected) digraph G = (V, E, w : E → R+).
is a multisubset of E that
Find: A connected Eulerian digraph G
(cid:48)
) where E
(cid:48) = (V, E
(cid:48)
(cid:48)
minimizes w(E
).
Similar to the recent progress on STSP, it is natural to consider special cases that are
easier to argue about but at the same time capture the combinatorial structure of the
problem. In particular, we shall consider the Node-Weighted ATSP, where we assume
that there exists a weight function f : V → R+ on the vertices so that w(u, v) = f (u).
(Another equivalent definition, which also applies to undirected graphs, is to let the
weight of an edge (u, v) be f (u) + f (v). This is equivalent to the definition above, if
scaled down by a factor of 2, since the solutions are Eulerian.)
Note that this generalizes ATSP on shortest path metrics of unweighted digraphs:
that is the problem where f is the constant function. As a curiosity, we also note that the
recent progress on STSP when restricted to shortest path metrics of unweighted graphs
is not known to generalize to the node-weighted case. We raise this as an interesting
open problem in Section 6.
6
2.3 Held-Karp Relaxation
The Held-Karp relaxation has a variable xe (cid:62) 0 for every edge in the given edge-
weighted graph G = (V, E, w). The intended solution is that xe should equal the number
of times e is used in the solution. The relaxation LP(G) is now defined as follows:
minimize
subject to
(cid:88)
xew(e)
e∈E
x(δ+(v)) = x(δ−
x(δ+(S)) (cid:62) 1
x (cid:62) 0.
(v))
v ∈ V,
∅ (cid:44) S ⊂ V,
program as(cid:80)
The first set of constraints says that the in-degree should equal the out-degree for each
vertex, i.e., the solution should be Eulerian. The second set of constraints enforces that
the solution is connected and they are sometimes referred to as subtour elimination
constraints. For Node-Weighted ATSP, we can write the objective function of the linear
v∈V f (v) · x(δ+(v)), where f : V → R+ is the weights on the vertices that
Finally, we remark that although the Held-Karp relaxation has exponentially many
constraints, it is well known that we can solve it in polynomial time either by using the
ellipsoid method with a separation oracle or by formulating an equivalent compact
(polynomial size) linear program.
defines the node-weighted metric.
3 ATSP with Local Connectivity
In this section we define a seemingly easier problem than ATSP by relaxing the
connectivity requirements. Consider an optimal solution x
to LP(G). Its value, which
is a lower bound on OPT, can be decomposed into a “lower bound” for each vertex v:
∗
(cid:88)
e∈E
∗
ew(e) =
x
(cid:88)
v∈V
(cid:88)
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
∗
ew(e).
x
e∈δ+(v)
lower bound for v
lbx∗,G(v) =(cid:80)
With this intuition, we let lb : V → R be the lower bound function defined by
∗
ew(e). We simplify notation and write lb instead of lbx∗,G as G will
∗
always be clear from the context and therefore also x
(if the optimal solution to LP(G)
is not unique then make an arbitrary but consistent choice). Note that lb(V) equals the
value of the optimal solution to the Held-Karp relaxation.
e∈δ+(v) x
Perhaps the main difficulty of ATSP is to satisfy the connectivity requirement, i.e.,
to select a Eulerian subset F of edges that connects the whole graph. We shall now relax
this condition to obtain what we call Local-Connectivity ATSP:
7
Local-Connectivity ATSP
Given: An edge-weighted (strongly connected) digraph G = (V, E, w) and a parti-
tioning V1 ∪ V2 ∪ . . . ∪ Vk of the vertices that satisfy: the graph induced by
Vi is strongly connected for i = 1, . . . , k.
Find: A Eulerian multisubset F of E such that
δ+
F (Vi) (cid:62) 1 for i = 1, 2, . . . , k
and
max
G∈C(F)
w( G)
lb( G)
is minimized.
Recall that C(F) denotes the set of connected components of the graph (V, F). We remark
that the restriction that each Vi should induce a strongly connected component is not
necessary but it makes our proofs in Section 4 easier.
We say that an algorithm for Local-Connectivity ATSP is α-light if it is guaranteed
(over all instances) to find a solution F such that
max
G∈C(F)
w( G)
lb( G)
(cid:54) α.
(3.1)
We also say that an algorithm is α-light on an ATSP instance G = (V, E, w) if, for each
partitioning V1 ∪ . . . ∪ Vk of V (such that Vi induces a strongly connected graph), it
returns a solution satisfying (3.1). We remark that we use the α-light terminology to
avoid any ambiguities with the concept of approximation algorithms because an α-light
algorithm does not compare its solution with respect to an optimal solution to the given
instance of Local-Connectivity ATSP.
An α-approximation algorithm for ATSP with respect to the Held-Karp relaxation
is trivially an α-light algorithm for Local-Connectivity ATSP: output the same Eulerian
subset F as the algorithm for ATSP. Since the set F connects the graph we have
max G∈C(F) w( G)/ lb( G) = w(F)/ lb(V) (cid:54) α. Moreover, Local-Connectivity ATSP seems
like a significantly easier problem than ATSP as the Eulerian set of edges only needs to
cross k cuts formed by a partitioning of the vertices. We substantiate this intuition by
proving, in Section 4, that there exists a simple 3-approximation for Local-Connectivity
ATSP on shortest path metrics of node-weighted graphs. We refer to this case as
Node-Weighted Local-Connectivity ATSP. Perhaps more surprisingly, we show in
Section 5 that any α-light algorithm for Local-Connectivity ATSP can be turned into an
algorithm for ATSP with an approximation guarantee of 5α with respect to the same
lower bound (from the Held-Karp relaxation).
Remark 3.1. Our generic reduction from ATSP to Local-Connectivity ATSP (Theorem 5.1)
is robust with respect to the definition of lb and there are many possibilities to define
∗
ew(e)/2. In fact,
in order to get a constant bound on the integrality gap of the Held-Karp relaxation,
our results say that it is enough to find an O(1)-light algorithm for Local-Connectivity
ATSP with respect to some nonnegative lb that only needs to satisfy that lb(V) is at
most the value of the optimal solution to the LP. Even more generally, if lb(V) is at most
the value of an optimal tour then our methods would give a similar approximation
guarantee (but not with respect to the Held-Karp relaxation).
such a lower bound. Another natural example is lb(v) =(cid:80)
e∈δ+(v)∪δ−(v) x
8
V1
V2
A1
A2
(cid:48)
Figure 1: A depiction of the construction of the auxiliary graph G
(in the proof of
Theorem 4.1): edges are subdivided, an auxiliary vertex Ai is added for each partition
Vi, and Ai is “connected” to subdivisions of the edges in δ+(Vi) and δ−
(Vi).
4 Approximating Local-Connectivity ATSP
We give a simple 3-light algorithm for Node-Weighted Local-Connectivity ATSP.
The proof is based on finding an integral circulation that sends flow across the cuts
{(Vi, ¯Vi) : i = 1, 2, . . . , k} and, in addition, satisfies that the outgoing flow of each vertex
(δ+(v))(cid:101) + 1 which in turn, by the assumptions on the metric, implies
v ∈ V is at most (cid:100)x
∗
a 3-light algorithm.
Theorem 4.1. There exists a polynomial time 3-light algorithm for Node-Weighted Local-
Connectivity ATSP.
Proof. Let G = (V, E, w) and V1 ∪ V2 ∪ . . . ∪ Vk be an instance of Local-Connectivity
ATSP where w : E → R+ is a node-weighted metric defined by f : V → R+. Let also x
∗
be an optimal solution to LP(G). We prove the theorem by giving a polynomial time
algorithm that finds a Eulerian multisubset F of E satisfying
∗
(4.1)
F (v) (cid:54) (cid:100)x
(δ+(v))(cid:101) + 1 for v ∈ V.
δ+
F (Vi) (cid:62) 1 for i = 1, . . . , k
and δ+
To see that this is sufficient, note that the Eulerian set F forms a solution to the Local-
F (Vi) (cid:62) 1 for i = 1, . . . , k; and it is 3-light since,
Connectivity ATSP instance because δ+
(cid:80)
for each G = ( V, E) ∈ C(F), we have (using that it is a node-weighted metric)
(cid:80)
(cid:80)
δ+
E(v) f (v)
v∈ V
v∈ V x∗(δ+(v)) f (v)
(δ+(v))(cid:101) + 1) f (v)
(cid:80)
v∈ V((cid:100)x
∗
v∈ V x∗(δ+(v)) f (v)
(δ+(v)) (cid:62) 1 and therefore (cid:100)x
∗
(cid:54)
∗
The last inequality follows from x
(δ+(v)).
We proceed by describing a polynomial time algorithm for finding a Eulerian set F
obtained
satisfying (4.1). We shall do so by finding a circulation in an auxiliary graph G
from G as follows (see also Figure 1):
(δ+(v))(cid:101) + 1 (cid:54) 3x
∗
(cid:48)
w( G)
lb( G)
=
(cid:54) 3.
1. Replace each edge e = (u, v) in G by adding vertices Oe, Ie and edges
(u, Oe), (Oe, Ie), (Ie, v);
every e ∈ δ+(Vi) and (Ie, Ai) for every e ∈ δ−
(Vi).
2. For each partition Vi, i = 1, . . . , k, add an auxiliary vertex Ai and edges (Ai, Oe) for
9
(cid:48)
Recall that a circulation in G
satisfying flow conservation: y(δ+(v)) = y(δ−
(cid:48)
claim follows from the construction of G
circulations.
(cid:48)
Claim 4.2. We can in polynomial time find an integral circulation y in G
is a vector y with a nonnegative value for each edge
(v)) for every vertex v. The following
together with basic properties of flows and
satisfying:
and y(δ+(v)) (cid:54) (cid:100)x
∗
to LP(G) to define a fractional circulation y
y(δ+(Ai)) = 1 for i = 1, . . . , k
∗
(cid:48)
Proof. We use the optimal solution x
in G
that satisfies the above degree bounds. As the vertex-degree bounds are integral, it
follows from basic facts about flows that we can in polynomial time find an integral
circulation y satisfying the same bounds (see e.g. Chapter 11 in [18]). Circulation y
is
defined as follows:
(δ+(v)(cid:101) for v ∈ V.
(cid:48)
(cid:48)
1. for each edge e = (u, v) in G with u, v ∈ Vi:
(cid:48)
y
(u,Oe)
(cid:48)
= y
(Oe,Ie)
x∗(δ+(Vi))
2. for each edge e = (u, v) in G with u ∈ Vi, v ∈ Vj where i (cid:44) j:
= y
= x
(cid:48)
(Ie,v)
∗
(u,v)
1
(cid:32)
1 −
(cid:33)
.
(cid:48)
y
(Oe,Ie)
(cid:48)
y
(Ai,Oe)
(cid:48)
(Ie,Aj)
y
,
∗
= x
(u,v)
∗
x
(u,v)
=
x∗(δ+(Vi))
∗
(u,v)
x
=
x∗(δ+(Vj))
,
,
(cid:48)
y
(u,Oe)
(cid:48)
y
(Ie,v)
∗
= x
(u,v)
∗
= x
(u,v)
(cid:32)
1 −
(cid:32)
1 −
(cid:33)
(cid:33)
x∗(δ+(Vi))
1
1
x∗(δ+(Vj))
,
.
(cid:48)
∗
∗
is defined so that a fraction 1/x
(δ+(Vi)) of the flow crossing the cut
Basically, y
(Vi, V \ Vi) goes through Ai. As x
(cid:48)
(δ+(Vi)) (cid:62) 1 we have that y
is nonnegative. It is also
(cid:48)
immediate from the definition of y
that it satisfies flow conservation and the degree
bounds of the claim: the in- and out-flow of a vertex v ∈ Vi is
(δ+(v));
the in- and out-flow of an auxiliary vertex Ai is 1 by design; and the in- and out-flow
of Oe and Ie for e = (u, v) is (1 − 1/x
e if u, v ∈ Vi for some i = 1, . . . , k and
∗
(δ+(Vi)))x
∗
(cid:48)
x
e otherwise. As mentioned above, the existence of fractional circulation y
implies
that we can also find, in polynomial time, an integral circulation y with the required
(cid:3)
properties.
x∗(δ+(Vi))
1 −
∗
x
(cid:16)
(cid:17)
∗
1
Having found an integral circulation y as in the above claim, we now obtain the
Eulerian subset F of edges. Initially, the set F contains y(Oe,Ie) multiplicities of each edge
e in G. Note that with respect to this edge set, in each partition Vi, either all vertices in
Vi are balanced (each vertex’s in-degree equals its out-degree) or there exist exactly
one vertex u so that δ+
F (u) − δ−
F (v) = 1.
Specifically, let u be the head of the unique edge e such that y(Ie,Ai) = 1 and let v be the
(cid:48)
so that y(Ai,Oe(cid:48) ) = 1. If u = v then all vertices in Vi are balanced.
tail of the unique edge e
F (u) = −1 and one vertex v so that δ+
F (v) − δ−
10
F (v) − δ−
F (u) − δ−
F (u) = −1 and v is so that δ+
F (v) = 1. In that
Otherwise u is so that δ+
case, we add a simple path from u to v to make the in-degrees and out-degrees of these
vertices balanced. As the graph induced by Vi is strongly connected, we can select the
path so that it only visits vertices in Vi. Therefore, we only increase the degree of vertices
in Vi by at most 1. Hence, after repeating this operation for each partition Vi, we have
that F is a Eulerian subset of edges and δF(δ+(v)) (cid:54) y(δ+(v)) + 1 (cid:54) (cid:100)x
(δ+(v))(cid:101) + 1 for all
∗
v ∈ V. Finally, we have δ+
F (Vi) (cid:62) 1 for each i = 1, . . . , k because yAi,Oe
= 1 (and therefore
(cid:62) 1) for one edge e ∈ δ+(Vi). We have thus given a polynomial time algorithm that
yOe,Ie
finds a Eulerian subset F satisfying the properties of (4.1), which, as discussed above,
implies that it is a 3-light algorithm for Node-Weighted Local-Connectivity ATSP. (cid:3)
5 From Local to Global Connectivity
In this section, we prove that if there is an α-light algorithm for Local-Connectivity
ATSP, then there exists an algorithm for ATSP with an approximation guarantee of
O(α). The main theorem can be stated as follows.
Theorem 5.1. Let A be an algorithm for Local-Connectivity ATSP and consider an ATSP
instance G = (V, E, w). If A is α-light on G, there exists a tour of G with value at most 5α lb(V).
Moreover, for any ε > 0, a tour of value at most (9 + ε)α lb(V) can be found in time polynomial
in the number n = V of vertices, in 1/ε, and in the running time of A.
Throughout this section, we let G = (V, E, w) and A be fixed as in the statement
of the theorem. The proof of the theorem is by giving an algorithm that uses A as a
subroutine. We first give the non-polynomial algorithm in Section 5.1 (with the better
guarantee) followed by Section 5.2 where we modify the arguments so that we also
efficiently find a tour (with slightly worse guarantee).
5.1 Existence of a Good Tour
Before describing the (non-polynomial) algorithm, we need to introduce the concept
of Eulerian partition. We say that graphs H1 = (V1, E1), H2 = (V2, E2), . . . , Hk = (Vk, Ek)
form a Eulerian partition of G if the vertex sets V1, . . . , Vk form a partition of V and each
Hi is a connected Eulerian graph where Ei is a multisubset of E. It is an β-light Eulerian
partition if in addition
w(Hi) (cid:54) β · lb(Hi)
for i = 1, . . . , k.
Our goal is to find a 5α-light Eulerian partition that only consists of a single component,
i.e., a 5α-approximate solution to the ATSP instance G with respect to the Held-Karp
relaxation.
The idea of the algorithm is to start with a Eulerian partition and then iteratively
merge/connect these connected components into a single connected component by
adding (cheap) Eulerian subsets of edges. Note that, since we will only add Eulerian
subsets, the algorithm always maintains that the connected components are Eulerian.
that contains the
∗
The state of the algorithm is described by a Eulerian multiset E
multiplicities of the edges that the algorithm has picked.
11
∗
Initialization. The algorithm starts with a 2α-light Eulerian partition H
1
(V
∗
, E
k) that maximizes the lexicographic order of
∗
∗
1), . . . , H
k
∗
= (V
k
, E
∗
1
=
(cid:104)lb(H
∗
1), lb(H
∗
2), . . . , lb(H
k)(cid:105).
∗
(5.1)
∗
1) (cid:62) lb(H
2) (cid:62) ··· (cid:62) lb(H
∗
As the lexicographic order is maximized, the Eulerian partitions are ordered so that
∗
lb(H
k). For simplicity, we assume that these inequalities are strict
∗
(which is w.l.o.g. by breaking ties arbitrarily but consistently). The set E
is initialized
∪ ··· ∪ E
∪ E
∗
∗
so that it contains the edges of the Eulerian partitions, i.e., E
k.
2
a connected subgraph G = ( V, E) of G, let low( G) denote the Eulerian partition H
lowest index i that intersects G3. That is,
During the execution of the algorithm we will also use the following concept. For
∗
i of
∗ = E
∗
1
low( G) = H
∗
min{i:V∗
i
∩ V(cid:44)∅}.
∗
, . . . , H
∗
∗
k. This means that H
i
). Moreover, as the algorithm will only add edges, each H
Note that after initialization, the connected components in C(E
∗
) are exactly the
∗
= low( G) for exactly one component
subgraphs H
G ∈ C(E
∗
1
i will be in at most
one component throughout the execution.
Remark 5.2. The main difference in the polynomial time algorithm is the initialization
since we do not know how to find a 2α-light Eulerian partition that maximizes the
lexicographic order in polynomial time. Indeed, it is consistent with our knowledge
that 2α (even 2) is an upper bound on the integrality gap and, in that case, such an
algorithm would always find a tour.
Remark 5.3. For intuition, let us mention that the reason for starting with a Eulerian
partition that maximizes the lexicographic order is that we will use the following
properties to bound the weight of the total tour:
1. A connected Eulerian subgraph H of G with w(H) (cid:54) 2α lb(H) has lb(H) (cid:54)
lb(low(H)).
2. For any disjoint connected Eulerian subgraphs H1, H2, . . . , H(cid:96) of G with low(Hj) =
∗
i and w(Hj) (cid:54) α lb(Hj) for j = 1, . . . , (cid:96), we have
H
(cid:96)(cid:88)
j=1
lb(Hj) (cid:54) 2 lb(H
∗
i ).
These bounds will be used to bound the weight of the edges added in the merge
procedure. Their proofs are easy and can be found in the analysis (see the proofs of
Claim 5.8 and Claim 5.9).
3Equivalently, it is the set H
∗
i maximizing lb(H
∗
i ) over all sets in the Eulerian partition that intersect G.
12
∗
F ( V) (cid:62) 1
for all ( V, E) ∈ C(E
∗
Merge procedure. The algorithm repeats the following “merge procedure” until C(E
∗
)
contains a single connected component. The components in C(E
) partition the vertex
∗
set and each component is strongly connected as it is Eulerian (since E
is a Eulerian
subset of edges). The algorithm can therefore use A to find a Eulerian multisubset F of
E such that
(i) δ+
); and
(ii) for each G ∈ C(F) we have w( G) (cid:54) α lb( G).
Note that A is guaranteed to find such a set since it is assumed to be an α-light
algorithm for Local-Connectivity ATSP on G. Furthermore, we may assume that no
connected component in C(F) is completely contained in a connected component in
C(E
∗
) (except for the trivial components formed by singletons). Indeed, the edges of
such a component can safely be removed from F and we have a new (smaller) multiset
that satisfies the above conditions. Having selected F, we now proceed to explain the
“update phase”:
U1: Let X = ∅.
U2: Select the component G = ( V, E) ∈ C(E
U3: If there exists a cycle C = (VC, EC) in G of weight w(C) (cid:54) α lb(low( G)) that connects
∗ ∪ F ∪ X), then add EC to X and repeat from
∗∪( E∩F)∪( E∩X).
G to another component in C(E
Step U2.
∗ ∪ F ∪ X) that minimizes lb(low( G)).
by adding the “new” edges in E, i.e., E
∗
U4: Otherwise, update E
∗ ← E
∗
∗
6
∗
, H
9
∗
, H
7
∗
Some comments about the update of E
are in order. We emphasize that we do
not add all edges of F ∪ X to E
∗
. Instead, we only add those new edges that belong
to the component G selected in the final iteration of the update phase. As G is a
connected component in C(E
∗ ∪ F ∪ X), F and X are Eulerian subsets of edges, we have
that E
remains Eulerian after the update. This finishes the description of the merging
procedure and the algorithm (see also the example below).
Example 5.4. In Figure 2, we have that, at the start of a merging step, C(E
∗
) consists of
6 components containing {H
}, and {H
},{H
},{H
}. The blue
∗
∗
1
4
(solid) cycles depict the connected Eulerian components of the edge set F. First, we
set X = ∅ and the algorithm selects the component G in C(E
∗ ∪ F ∪ X) that minimizes
lb(low( G)) or, equivalently, that maximizes min{i : H
i intersects G}. In this example, it
∗
would be the left most of the three components in C(E
∗
4. The
algorithm now tries to connect this component to another component by adding a
∗
cycle with weight at most α lb(H
4). The red (dashed) cycle corresponds to such a cycle
and its edge set is added to X. In the next iteration, the algorithm considers the two
∗ ∪ F ∪ X). The smallest component (with respect to lb(low( G))) is
components in C(E
∗
∗
the one that contains H
8. Now suppose that there is no cycle of weight at
∗
3
most α lb(H
is
updated by adding those edges of F ∪ X that belong to this component (depicted by
the thick cycle).
∗
3) that connects this component to another component. Then the set E
∗ ∪ F) with low( G) = H
∗
5, and H
},{H
},{H
∗
10
, H
, H
∗
3
∗
5
∗
, H
8
∗
2
13
∗
H
9
∗
6
H
∗
10
H
∗
H
7
∗
H
4
∗
H
3
∗
H
2
∗
H
5
∗
H
8
∗
H
1
Figure 2: An illustration of the merge procedure. Blue (solid) cycles depict F and the
red (dashed) cycle depicts X after one iteration of the update phase. The thick cycle
represents the edges that this merge procedure would add to E
∗
.
5.1.1 Analysis
Termination. We show that the algorithm terminates by arguing that the update phase
terminates with fewer connected components and the merge procedure is therefore
repeated at most k (cid:54) n times.
Lemma 5.5. The update phase terminates in polynomial time and decreases the number of
connected components in C(E
∗
Proof. First, observe that each single step of the update phase can be implemented
in polynomial time. The only nontrivial part is Step U3 which can be implemented
as follows: for each edge (u, v) ∈ δ+( V) consider the cycle consisting of (u, v) and a
shortest path from v to u. Moreover, the whole update phase terminates in polynomial
time because each time the if-condition of Step U3 is satisfied, we add a cycle to X that
decreases the number of connected components in C(E
∗ ∪ F ∪ X). The if-condition of
Step U3 can therefore be satisfied at most k (cid:54) n times.
We proceed by proving that at termination the update phase decreases the number
of connected components in C(E
). Consider when the algorithm reaches Step U4. In
∗ ∪ F ∪ X). Note that G (cid:60) C(E
that case it has selected a component G = ( V, E) ∈ C(E
∗
)
because the edge set F crosses each cut defined by the vertex sets of the connected
components in C(E
by adding all the edges
(F ∪ X) ∩ E it decreases the number of components in C(E
(cid:3)
∗
). Therefore when the algorithm updates E
) by at least one.
∗
).
∗
∗
Performance Guarantee. To analyze the performance guarantee we shall split our
analysis into two parts. Note that when one execution of the merge procedure
terminates (Step U4) we add edge set (F ∩ E) ∪ (X ∩ E) to our solution. We shall
analyze the contribution of these two sets F ∩ E and X ∩ E separately. More formally,
suppose that the algorithm does T repetitions of the merge procedure. Let G1 =
( V1, E1), G2 = ( V2, E2), . . . , GT = ( VT, ET), F1, F2, . . . , FT, and X1, X2, . . . , XT denote the
14
+
+
w
w
i=1
t=1
t=1
Ft
Xt
w(H
k(cid:88)
(cid:54)α lb(V) by Lemma 5.6
(cid:54)2α lb(V) by Lemma 5.7
With this notation, we proceed to bound the total weight of the solution by
selected components, the edge set F, and the edge set X, respectively, at the end of each
∗
repetition. To simplify notation, we denote the edges added to E
in the t:th repetition
by Ft = Ft ∩ Et and Xt = Xt ∩ Et.
(cid:16)∪T
(cid:17)
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
(cid:16)∪T
(cid:16)∪T
(cid:17)
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
Here we used that(cid:80)k
∗
i ) (cid:54) 5α lb(V) as claimed in Theorem 5.1.
Lemma 5.6. We have w
Proof. Note that Xt consists of a subset of the cycles added to Xt in Step U3 of the update
phase. Specifically, those cycles contained in the connected component Gt selected
at Step U2 in the last iteration of the update phase during the t:th repetition of the
merge procedure. We can therefore decompose ∪T
Xt into cycles C1 = (V1, E1), C2 =
(V2, E2), . . . , Cc = (Vc, Ec) indexed in the order they were added by the algorithm. When
Cj was selected in Step U3 of the update phase, it satisfied the following two properties:
(i) it connected the component G selected in Step U2 with at least one other component
∗
k is a 2α-light Eulerian partition.
(cid:17) (cid:54) α lb(V).
It remains to prove Lemmas 5.6 and 5.7.
∗
∗
i ) (cid:54) 2α lb(V) since H
1
i=1 w(H
, . . . , H
Xt
t=1
t=1
(cid:48)
G
(cid:48)
such that lb(low( G
)) > lb(low( G)); and
(ii) it had weight w(Cj) at most α lb(low( G)).
In this case, we say that Cj marks low( G).
∗
We claim that at most one cycle in C1, C2, . . . , Cc marks each H
1
∗
k. To see
∗
i (if any). By (i) above, when Cj was
this, consider the first cycle Cj that marks H
(cid:48)
)) > lb(low( G))
added, it connected two components G and G
where low( G) = H
will remain connected
throughout the execution of the algorithm. Therefore, by the definition of low and by
the fact that lb(low( G
appearing later in
the algorithm always has low( G
∗
i and such a cycle
(cid:3)
∗
i . As the algorithm only adds edges, G and G
∗
, H
2
(cid:48)
such that lb(low( G
The bound now follows from that at most one cycle marks each H
∗
∗
i . Hence, no other cycle marks H
i .
)) > lb(low( G)), we have that a component G
has weight at most α lb(H
, . . . , H
) (cid:44) H
(cid:48)(cid:48)
(cid:48)(cid:48)
(cid:48)
(cid:48)
We complete the analysis of the performance guarantee with the following lemma.
∗
i ).
(cid:16)∪T
(cid:17) (cid:54) 2α lb(V).
Ft
t=1
Lemma 5.7. We have w
Proof. Consider the t:th repetition of the merge procedure. The edge set Ft is Eulerian
but not necessarily connected. Let F t denote the set of the Eulerian subgraphs
corresponding to the connected components in C( Ft) where we disregard the trivial
, . . . ,F t
components that only consist of a single vertex. Further, partition F t into F t
where F t
∗
k
i and do not intersect
∗
any of the subgraphs H
1
i contains those Eulerian subgraphs in F t that intersect H
∗
i−1. That is,
∗
, H
2
,F t
1
2
, . . . , H
F t
i
= {H ∈ F t : low(H) = H
}.
∗
i
15
Note that the total weight of Ft, w( Ft), equals w(F t) =(cid:80)k
i=1 w(F t
∗
i ).
∗
i−1
∗
, H
2
, . . . , H
, . . . , H
i , we have lb(H) (cid:54) lb(low(H)) = lb(H
i ). We bound the weight
of F t by considering each F t
i separately. We start by two simple claims that follow
from that each H ∈ F t satisfies w(H) (cid:54) α lb(H) (since A is an α-light algorithm) and
∗
∗
the choice of H
k to maximize the lexicographic order of (5.1). We remark that
1
the proofs of the following claims are the only arguments that use the fact that the
lexicographic order was maximized.
Claim 5.8. For H ∈ F t
∗
i ) together with the fact that w(H) (cid:54) α lb(H) (cid:54) 2α lb(H)
Proof. Inequality lb(H) > lb(H
∗
∗
, . . . , H
would contradict that H
k was chosen to maximize the lexicographic order of (5.1).
1
Indeed, in that case, a 2α-light Eulerian partition of higher lexicographic order would
∗
, H and the remaining vertices (as trivial singleton components) that
be H
1
(cid:3)
do not belong to any of these Eulerian subgraphs.
Claim 5.9. We have lb(F t
= {H1, H2, . . . , H(cid:96)}
Proof. Suppose toward contradiction that lb(F t
∗
and define H
i and
∗
H1, . . . , H(cid:96). Consider the Eulerian partition H
and the remaining vertices
1
(as trivial singleton components) that do not belong to any of these Eulerian subgraphs.
∗
We have lb(H
i ) and therefore the lexicographic value of this Eulerian partition
∗
∗
is larger than the lexicographic value of H
k. This is a contradiction if it is also a
1
∗
2α-light Eulerian partition, i.e., if w(H
)
(cid:54) 2α.
lb(H∗)
∗
). By the facts that w(Hj) (cid:54) α lb(Hj) (since
) > 2α lb(H
A is an α-light algorithm) and that H
∗
∗
, . . . , H
k is a 2α-light Eulerian partition,
1
i ). Let F t
∗
i ) > 2 lb(H
∗
, H
, . . . , H
to be the Eulerian graph obtained by taking the union of the graphs H
Therefore, we must have w(H
i ) (cid:54) 2 lb(H
) > lb(H
, . . . , H
∗
i−1
∗
i ).
∗
∗
∗
i
∗
w(H
) = w(H
∗
i ) +
w(Hj) (cid:54) 2α lb(H
∗
i ) +
α lb(Hj)
and
(cid:96)(cid:88)
j=1
) > 2α lb(H
∗
) imply lb(F t
(cid:96)(cid:88)
j=1
(cid:16)∪T
k(cid:88)
(cid:17)
(cid:88)
Ft
as
t=1
lb(Hj).
j=1 lb(Hj) (cid:54)
(cid:3)
∗
j=1
) (cid:62)
lb(H
(cid:96)(cid:88)
i ) = (cid:80)(cid:96)
(cid:88)
k(cid:88)
These inequalities together with w(H
2 lb(H
∗
i ).
∗
Using the above claim, we can write w
T(cid:88)
k(cid:88)
T(cid:88)
k(cid:88)
w(F t
i ) (cid:54) α
lb(F t
i ) = α
lb(F t
i ) (cid:54) 2α
lb(H
∗
i ).
i
i
i=1
i=1
i=1
i=1
t=1
t=1
(cid:44)∅
t:F t
t:F t
(cid:44)∅
We complete the proof of the lemma by using Claim 5.8 to prove that F t
i is non-empty
for at most one repetition t of the merge procedure. Suppose toward contradiction that
there exist 1 (cid:54) t0 < t1 (cid:54) T so that both F t0
(cid:44) ∅. In the t0:th repetition of
∗
i is contained in the subgraph Gt0 since otherwise no edges
the merge procedure, H
∗
incident to H
i ). Now
consider a Eulerian subgraph H ∈ F t1
. First, we cannot have that H is contained in the
component Gt0 since each (nontrivial) component of F is assumed to not be contained
in any component of C(E
∗
i ).
(cid:44) ∅ and F t1
∗
∗
i would have been added to E
). Second, by Claim 5.8, we have w(H) (cid:54) α lb(H) (cid:54) α lb(H
. Therefore lb(low( Gt0)) (cid:62) lb(H
∗
i
i
i
16
In short, H is a Eulerian subgraph that connects Gt0 to another component and it
has weight at most α lb(low( Gt0)). As H is Eulerian, it can be decomposed into cycles.
One of these cycles, say C, connects Gt0 to another component and
w(C) (cid:54) w(H) (cid:54) α lb(H
∗
i ) (cid:54) α lb(low( Gt0)).
(5.2)
In other words, there exists a cycle C that, in the t0:th repetition of the merge procedure,
satisfied the if-condition of Step U3, which contradicts the fact that Step U4 was reached
(cid:3)
when component Gt0 was selected.
5.2 Polynomial Time Algorithm
In this section we describe how to modify the arguments in Section 5.1 to obtain an
algorithm that runs in time polynomial in the number n of vertices, in 1/ε, and in the
running time of A.
By Lemma 5.5, the update phase can be implemented in polynomial time in n.
Therefore, the merge procedure described in Section 5.1 runs in time polynomial in
n and in the running time of A. The problem is the initialization: as mentioned in
Remark 5.2, it seems difficult to find a polynomial time algorithm for finding a 2α-light
∗
Eulerian partition H
1
, . . . , H
∗
k that maximizes the lexicographic order of
(cid:104)lb(H
∗
2), . . . , lb(H
∗
1), lb(H
k)(cid:105).
∗
We overcome this obstacle by first identifying the properties that we actually use
from selecting the Eulerian partition as above. We then show that we can obtain a
Eulerian partition that satisfies these properties in polynomial time.
As mentioned in the analysis in Section 5.1, the only place where we use that
the Eulerian partition maximizes the lexicographic order of (5.1) is in the proof of
Lemma 5.7. Specifically, it is used in the proofs of Claims 5.8 and 5.9. Instead of
proving these claims, we shall simply concentrate on finding a Eulerian partition that
satisfies a relaxed variant of them (formalized in the lemma below, see Condition (5.3)).
The claimed polynomial time algorithm is then obtained by first proving that a slight
modification of the merge procedure returns a tour of value at most (9α + 2ε) lb(V)
if Condition (5.3) holds, and then we show that a Eulerian partition satisfying this
condition can be found in time polynomial in n and in the running time of A. We start
by describing the modification to the merge procedure.
Modified merge procedure. The only modification to the merge procedure in Sec-
tion 5.1 is that we change the update phase by relaxing the condition of the if-statement
in Step U3 from w(C) (cid:54) α lb(low( G)) to w(C) (cid:54) α(3 lb(low( G)) + ε lb(V)/n). In other
words, Step U3 is replaced by
U3’: If there exists a cycle C = (VC, EC) in G of weight w(C) (cid:54) α(3 lb(low( G)) +ε lb(V)/n)
∗ ∪ F ∪ X), then add EC to X and
that connects G to another component in C(E
repeat from Step U2.
17
Clearly the modified merge procedure still runs in time polynomial in n and in
the running time of A. Moreover, we show that if Condition (5.3) holds then the
returned tour will have weight O(α). Recall from Section 5.1 that Ft denotes the
subset of F and Xt denotes the subset of X that were added in the t:th repetition of
the (modified) merge procedure. Furthermore, we define (as in the previous section)
F t
∗
i and H is a nontrivial component, i.e., H contains more
than one vertex}.
Lemma 5.10. Assume that the algorithm is initialized with a 3α-light Eulerian partition
∗
∗
k so that, in each repetition t of the modified merge procedure, we add a subset Ft
H
1
such that
= {H ∈ C( Ft) : low(H) = H
, . . . , H
∗
, H
2
i
lb(F t
i ) (cid:54) 3 lb(H
∗
i ) +
ε lb(V)
n
for i = 1, 2, . . . , k.
(5.3)
Then the returned tour has weight at most (9 + 2ε)α lb(V).
Let us comment on the above statement before giving its proof. The reason that we
use a 3α-light Eulerian partition (instead of one that is 2α-light) is that it leads to a better
constant when balancing the parameters. We also remark that (5.3) is a relaxation of the
bound of Claim 5.9 from lb(F t
∗
i ) + ε lb(V)/n; and it also
∗
implies a relaxed version of Claim 5.8: from lb(H) (cid:54) lb(H
i )+ε lb(V)/n.
It is because of this relaxed bound that we modified the if-condition of the update
phase (by relaxing it by the same amount) which will be apparent in the proof.
∗
i ) to lb(H) (cid:54) 3 lb(H
i ) to lb(F t
∗
i ) (cid:54) 3 lb(H
i ) < 2 lb(H
Proof. As in the analysis of the performance guarantee in Section 5.1, we can write the
weight of the returned tour as
(cid:16)∪T
(cid:17)
(cid:16)∪T
w
Ft
t=1
+ w
Xt
t=1
+
(cid:17)
k(cid:88)
i=1
w(H
∗
i ).
(cid:16)∪T
(cid:17)
Xt
To bound w
except that the weight of a cycle C that marks H
instead of by α lb(H
update phase). Hence, w
k (cid:54) n.
, we observe that proof of Lemma 5.6 generalizes verbatim
∗
t=1
i )+ε lb(V)/n)
∗
i ) (because of the relaxation of the bound in the if-condition of the
∗
i ) + ε lb(V)/n) (cid:54) (3 + ε)α lb(V) because
∗
i is now bounded by α(3 lb(H
(cid:16)∪T
α(3 lb(H
t=1
i=1
We proceed to bound w
. Using the same arguments as in the proof of
Xt
(cid:17) (cid:54)(cid:80)k
(cid:17)
(cid:16)∪T
(cid:88)
Ft
t=1
i=1
(cid:44)∅
t:F t
i
Lemma 5.7,
w
(cid:16)∪T
Ft
t=1
(cid:17) (cid:54) α
k(cid:88)
lb(F t
i ) (cid:54) α
3 lb(H
∗
i ) + ε lb(V)/n
k(cid:88)
(cid:88)
(cid:16)
i=1
(cid:44)∅
t:F t
i
(cid:17)
where, for the last inequality, we used the assumption of the lemma. Now we apply
exactly the same arguments as in the end of the proof of Lemma 5.7 to prove that F t
i is
non-empty for at most one repetition t of the merge procedure. The only difference, is
that (5.2) should be replaced by
w(C) (cid:54) w(H) (cid:54) α(3 lb(H
∗
i ) + ε lb(V)/n) (cid:54) α(3 lb(low( Gt0)) + ε lb(V)/n)
18
(because (5.3) can be seen as a relaxed version of Claim 5.8). However, as we also
updated the bound in the if-condition, the argument that C would satisfy the if-condition
of Step U3’ is still valid. Hence, we conclude that F t
is non-empty in at most one
repetition and therefore
i
(cid:16)∪T
w
Ft
t=1
(cid:17) (cid:54) α
k(cid:88)
(cid:88)
(cid:16)
i=1
(cid:44)∅
t:F t
i
(cid:17) (cid:54) (3 + ε)α lb(V).
3 lb(H
∗
i ) + ε lb(V)/n
∗
By the above bounds and since H
1
∗
, H
2
have that the weight of the returned tour is
, . . . , H
∗
k is a 3α-light Eulerian partition, we
(cid:16)∪T
(cid:17)
(cid:16)∪T
w
Ft
t=1
+ w
Xt
t=1
+
(cid:17)
k(cid:88)
i=1
w(H
∗
i ) (cid:54) (3 + ε)α lb(V) + (3 + ε)α lb(V) + 3α lb(V)
= (9 + 2ε)α lb(V).
(cid:3)
Finding a good Eulerian partition in polynomial time. By the above lemma, it is
sufficient to find a 3α-light Eulerian partition so that Condition (5.3) holds during the
execution of the modified merge procedure. However, how can we do it in polynomial
time? We do as follows. First, we select the trivial 3α-light Eulerian partition where
each subgraph is only a single vertex. Then we run the modified merge procedure and,
in each repetition, we verify that Condition (5.3) holds. Note that this condition is easy
to verify in time polynomial in n. If it holds until we return a tour, then we know by
Lemma 5.10 that the tour has weight at most (9 + 2ε)α lb(V). If it does not hold during
one repetition, then we will restart the algorithm with a new 3α-light Eulerian partition
that we find using the following lemma. We continue in this manner until the merge
procedure executes without violating Condition (5.3) and therefore it returns a tour of
weight at most (9α + 2ε) lb(V).
Lemma 5.11. Suppose that repetition t of the (modified) merge procedure violates Condition (5.3)
∗
∗
∗
, . . . , H
, H
when run starting from a 3α-light Eulerian partition H
k. Then we can, in time
∗
∗
∗
2
1
, . . . , H
, H
polynomial in n, find a new 3α-light Eulerian partition H
k
2
1
so that
k(cid:88)
j=1
j)2 − k(cid:88)
∗
j=1
lb( H
lb(H
∗
j)2 (cid:62)
ε2
3n2 lb(V)2.
(5.4)
(cid:80)k
Note that the above lemma implies that we will reinitialize (in polynomial time) the
∗
∗
Eulerian partition at most 3n2/ε2 times because any Eulerian partition H
k has
1
∗
i )2 (cid:54) lb(V)2. As each execution of the merge procedure takes time polynomial
i=1 lb(H
in n and in the running time of A, we can therefore find a tour of weight at most
(9 + 2ε)α lb(V) = (9 + ε(cid:48)
)α lb(V) in the time claimed in Theorem 5.1, i.e., polynomial in
n, 1/ε(cid:48)
, and in the running time of A. It remains to prove the lemma.
, . . . , H
19
Proof. Since the t:th repetition of the merge procedure violates Condition (5.3), there is
an 1 (cid:54) i (cid:54) k such that
lb(F t
i ) > 3 lb(H
∗
i ) +
ε
n
lb(V).
∗
∗
∗
1
together with a subset of {H
We shall use this fact to construct a new 3α-light Eulerian partition consisting of
} containing
a new Eulerian subgraph H
those subgraphs that do not intersect H
and finally the vertices (as trivial singleton
components) that do not belong to any of these Eulerian subgraphs. We need to define
. Let I ⊆ {1, 2, . . . , k} be the indices of those Eulerian subgraphs
the Eulerian subgraph H
i . Note that, by definition, we have i ∈ I
∗
of H
and j (cid:62) i for all j ∈ I. We shall construct the graph H
1
iteratively. Initially, we let H
be
the connected Eulerian subgraph obtained by taking the union of F t
∗
i and H
i . This is a
connected Eulerian subgraph as each Eulerian subgraph in F t
∗
∗
i intersects H
i and H
i is a
connected Eulerian subgraph.
k that intersect the vertices in F t
∗
∗
, . . . , H
k
, . . . , H
∗
, H
2
∗
∗
∗
∗
∗
∗
∗
∗
j
∗
j
) (cid:54) 3α lb(H
}j∈I to H
∗
The careful reader can observe that up to now H
. Specifically, we would like to add {H
is defined in the same way as
in the proof of Claim 5.9. However, in order to satisfy (5.4) we shall add more of
the Eulerian subgraphs in {H
}j∈I(cid:48),
∗
where I
) (because we wish to increase the
“potential” in (5.4)) subject to that w(H
) (because the new Eulerian partition
should be 3α-light).
∗
To see that w(H
(cid:48) ⊆ I \ {i} is selected so as to maximize lb(H
∗
) (cid:54) 3α lb(H
) implies that the new Eulerian partition is 3α-light,
}j(cid:60)I, and
recall that the new Eulerian partition consists of H
the vertices that do not belong to any of these Eulerian subgraphs. By the definition of I,
∗
no H
k are disjoint, it follows that the new Eulerian
∗
∗
j) (cid:54) 3α lb(H
partition consists of disjoint subgraphs. Moreover, each H
j)
since the Eulerian partition we started with is 3α-light. Hence, the new Eulerian
partition is 3α-light if w(H
). Inequality (5.5) is thus a sufficient condition
for the new Eulerian partition to be 3α-light. We remark that the condition trivially
∗
i ) > 3 lb(H
i ) + ε lb(V)/n.
holds for I
∗
Claim 5.12. We have w(H
) if
i ) (cid:54) 2
3
∗
(cid:48) = ∅ because lb(F t
(cid:88)
, the Eulerian subgraphs {H
∗
j with j (cid:60) I intersects H
) (cid:54) 3α lb(H
∩ F t
∗
j satisfies w(H
i ) − lb(H
) (cid:54) 3α lb(H
∩ F t
i ).
lb(F t
, . . . , H
. As H
lb(H
(5.5)
∗
1
∗
i
∗
j
∗
j
∗
∗
∗
j∈I(cid:48)
(cid:88)
j∈I(cid:48)
∗
j
\ F t
i ).
lb(H
∗
) = lb(F t
∗
i ) + lb(H
i
\ F t
i ) +
lb(H
20
Proof. We have
∗
w(H
) = w(F t
i ) + w(H
∗
i ) +
(cid:88)
j∈I(cid:48)
w(H
where the inequality follows from that F t
∗
H
1
∗
k is a 3α-light Eulerian partition. Moreover,
, . . . , H
(cid:88)
∗
i ) + 3α
i ) + 3α lb(H
j) (cid:54) α lb(F t
∗
i was selected by the α-light algorithm A and
lb(H
∗
j),
j∈I(cid:48)
Hence, we have, by rearranging terms and using lb(H
w(H
) (cid:54) 3α lb(H
) holds if
∗
∗
j)− lb(H
∗
∗
j
\F t
i ) = lb(H
∗
j
∩F t
i ), that
∗
3α lb(H
i
∩ F t
i ) + 3α
∗
j
∩ F t
i ) (cid:54) 2α lb(F t
i ).
lb(H
(cid:88)
j∈I(cid:48)
The above can be simplified to(cid:88)
∗
j
∩ F t
i ) (cid:54) 2 lb(F t
i )/3 − lb(H
∗
i
∩ F t
i ).
lb(H
j∈I(cid:48)
(cid:3)
(cid:48) ⊆ I \ {i} that satisfies (5.5)
From the above discussion, we wish to find a subset I
and maximizes
∗
) = lb(F t
∗
i ) + lb(H
i
\ F t
i ) +
lb(H
∗
lb(H
i
\ F t
i ),
(cid:88)
j∈I(cid:48)
(cid:48)
where only the last term depends on the selection of I
problem that, for each j ∈ I \ {i}, has an item of size sj = lb(H
∩ F t
∗
j
i ) − lb(H
∗
pj = lb(H
(cid:48)
i
right-hand-side of (5.5). We solve this knapsack problem and obtain I
. We interpret this as a knapsack
i ) and profit
∩ F t
i ), i.e., the
as follows:
i ); the capacity U of the knapsack is 2
3 lb(F t
\ F t
∗
j
1. Find an optimal extreme point solution z
relaxation of the knapsack problem:
∗
to the standard linear programming
maximize
subject to
(cid:88)
(cid:88)
j∈I\{i}
zjpj
zjsj (cid:54) U,
j∈I\{i}
0 (cid:54) zj (cid:54) 1
for all j ∈ I \ {i}.
2. As the above relaxation has only one constraint (apart from the boundary
constraints), the extreme point z
has at most one variable with a fractional value.
We obtain an integral solution (i.e., a packing) by simply dropping the fractionally
packed item. That is, we let I
(cid:48) = {j ∈ I \ {i} : z
= 1}.
∗
∗
j
The running time of the above procedure is dominated by the time it takes to solve the
linear program. This can be done very efficiently by solving the fractional knapsack
problem with the greedy algorithm (or, for the purpose here, use any general polynomial
time algorithm for linear programming). We can therefore obtain I
and the new Eulerian
partition in time polynomial in I (cid:54) n as stated in lemma.
(cid:48)
\ F t
It remains to prove (5.4). Let us first bound the profit of our “knapsack solution” I
Claim 5.13. We have(cid:80)
\ F t
(cid:80)
i ) − lb(H
∗
i ).
j∈I\i lb(H
j∈I(cid:48) lb(H
i ) (cid:62) 1
3
∗
j
∗
j
(cid:48)
.
21
Proof. By definition,(cid:88)
j∈I(cid:48)
∗
j
\ F t
i ) =
lb(H
(cid:88)
(cid:88)
j∈I\{i}
pj (cid:62)
=1
jpj − max
∗
z
j∈I\{i} pj,
j∈I\{i}:z∗
j
where we used that at most one item is fractionally packed in z
maxj∈I\{i} pj = maxj∈I\{i} lb(H
thus sufficient to prove that z
relaxation of the knapsack problem. Indeed, by the optimality of z
. As j (cid:62) i for all j ∈ I,
\ F t
∗
i ) (cid:54) lb(H
i ). To complete the proof of the claim, it is
= 1/3 for all j ∈ I \ {i} is a feasible solution to the LP
(cid:48)
j
, we then have
∗
j
∗
∗
(cid:80)
(cid:80)
j∈I\{i} z
We have that z
j∈I\{i} z
(cid:48)
(cid:48)
jpj = 1
3
\ F t
i ).
is a feasible solution because
j∈I\{i} lb(H
∗
j
lb(H
∗
j
∩ F t
i ) (cid:54) 1
3
lb(F t
i ) (cid:54)
2
3
− lb(H
∩ F t
∗
i )
i
lb(F t
i )
lb(F t
i ) = U,
∗
jpj (cid:62)(cid:80)
(cid:88)
1
3
j∈I\{i}
where the first inequality follows from that the subgraphs {H
second inequality follows from that lb(H
i ) (cid:54) 1/3.
∗
j
}j∈I are disjoint and the
(cid:3)
We finish the proof of the lemma by using the above claim to show the increase
of the “potential” function as stated in (5.4). By the definition of the new Eulerian
partition (it contains {H
}j(cid:60)I), we have that the increase is at least
∗
j
i )/ lb(F t
∗
)2 −(cid:88)
lb(H
j∈I
∗
j)2.
∗
lb(H
Let us concentrate on the first term:
By using lb(H
i )/3, we can further lower bound this expression by
i ) (cid:54) lb(F t
∗
lb(F t
i ) + 1
3
lb(F t
1
3
i ) +
(cid:88)
j∈I
lb(H
∗
j
\ F t
i )
22
By Claim 5.13, we have that the expression inside the parenthesis is at least
∗
)2 =
lb(H
\ F t
i ) +
∗
j
\ F t
i )
lb(H
(cid:62) lb(F t
i )
∗
i ) + lb(H
i
i ) +
∗
j
\ F t
i )
lb(H
(cid:88)
j∈I(cid:48)
\ F t
(cid:88)
j∈I(cid:48)
2
.
lb(F t
∗
i ) + lb(H
i
lb(F t
(cid:88)
lb(F t
i ) + lb(H
∗
i
\ F t
i )+1
3
j∈I\{i}
(cid:62) lb(F t
lb(H
∗
j
i ) + 1
3
j∈I
\ F t
lb(H
i ) − lb(H
∗
(cid:88)
i )
i ) − lb(H
\ F t
∗
∗
i ).
j
= 1
(cid:88)
i ) + 1
3
lb(F t
3
lb(H
j∈I
∗
j).
Finally, as lb(F t
i ) (cid:62) ε lb(V)/n, lb(F t
)2 −(cid:88)
∗
i ) (cid:62) 3 lb(H
i ), and lb(H
)2 − lb(H
∗
∗
∗
∗
j)2 (cid:62) lb(H
j)
i )
(cid:88)
lb(H
lb(H
∗
lb(H
j∈I
∗
j) (cid:54) lb(H
i ) for all j ∈ I, we have
∗
1
3
(cid:62) lb(F t
i )
(cid:62) lb(F t
i )2
3
lb(F t
lb(F t
i )
3
+
(cid:88)
j∈I
i ) + 1
(cid:88)
3
j∈I
j∈I
lb(H
∗
j)
(cid:88)
j∈I
lb(H
∗
j)
− lb(H
(cid:88)
∗
i )
j∈I
j) − lb(H
∗
∗
i )
lb(H
lb(H
∗
j)
(cid:62)
ε2
3n2 lb(V)2
which completes the proof of Lemma 5.11.
(cid:3)
6 Discussion and Open Problems
We gave a new approach for approximating the asymmetric traveling salesman problem.
It is based on relaxing the global connectivity requirements into local connectivity
conditions, which is formalized as Local-Connectivity ATSP. We showed a rather easy 3-
light algorithm for Local-Connectivity ATSP on shortest path metrics of node-weighted
graphs. This yields via our generic reduction a constant factor approximation algorithm
for Node-Weighted ATSP. However, we do not know any O(1)-light algorithm for
Local-Connectivity ATSP on general metrics and, motivated by our generic reduction,
we raise the following intriguing question:
Open Question 6.1. Is there a O(1)-light algorithm for Local-Connectivity ATSP on
general metrics?
We note that there is great flexibility in the exact choice of the lower bound lb as
noted in Remark 3.1. A further generalization of our approach is to interpret it as a
primal-dual approach. Specifically, it might be useful to interpret the lower bound as a
feasible solution of the dual of the Held-Karp relaxation: the lower bound is then not
only defined over the vertices but over all cuts in the graph. We do not know if any of
these generalizations are useful at this point and it may be that there is a nice O(1)-light
algorithm for Local-Connectivity ATSP without changing the definition of lb.
By specializing the generic reduction to Node-Weighted ATSP, it is possible to
improve our bounds slightly for this case. Specifically, one can exploit the fact that a
cycle C always has w(C) (cid:54) lb(C) in these metrics. This allows one to change the bound
in Step U3 of the update phase to be w(C) (cid:54) lb(low( G)) instead of w(C) (cid:54) α lb(low( G)),
which in turn improves the upper bound on the integrality gap of the Held-Karp
relaxation to 4 · α + 1 = 13 (since α = 3 for node-weighted metrics). That said, we do
not see how to make a significant improvement in the guarantee and it would be very
interesting with a tight analysis of the integrality gap of the Held-Karp relaxation for
Node-Weighted ATSP. We believe that such a result would also be very interesting even
if we restrict ourselves to shortest path metrics of unweighted graphs.
23
Finally, let us remark that the recent progress for STSP on shortest path metrics
of unweighted graphs is not known to extend to node-weighted graphs, i.e., Node-
Weighted STSP. Is it possible to give a (1.5 − ε)-approximation algorithm for Node-
Weighted STSP for some constant ε > 0? We think that this is a very natural question
that lies in between the now fairly well understood STSP on shortest path metrics
of unweighted graphs and STSP on general metrics (i.e., edge-weighted instead of
node-weighted graphs).
Acknowledgments
The author is very grateful to László Végh, Johan Håstad, and Hyung-Chan An for
inspiring discussions and valuable comments that influenced this work. We also thank
Jakub Tarnawski and Jens Vygen for useful feedback on the manuscript.
This research is supported by ERC Starting Grant 335288-OptApprox.
References
[1] N. Anari and S. O. Gharan. Effective-resistance-reducing flows, spectrally thin trees, and
asymmetric TSP. CoRR, abs/1411.4613, 2014. 3
[2] N. Anari and S. O. Gharan. The kadison-singer problem for strongly rayleigh measures
and applications to asymmetric TSP. CoRR, abs/1412.1143, 2014. 3
[3] S. Arora, M. Grigni, D. R. Karger, P. N. Klein, and A. Woloszyn. A polynomial-time
approximation scheme for weighted planar graph TSP. In Proceedings of the Ninth Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 1998, pages 33–41, 1998. 3
[4] A. Asadpour, M. X. Goemans, A. Madry, S. O. Gharan, and A. Saberi. An O(log n/ log log
n)-approximation algorithm for the asymmetric traveling salesman problem. In Proceedings
of the Twenty-First Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2010, pages
379–389, 2010. 3
[5] M. Bläser. A new approximation algorithm for the asymmetric TSP with triangle inequality.
ACM Transactions on Algorithms, 4(4), 2008. 3
[6] M. Charikar, M. X. Goemans, and H. J. Karloff. On the integrality ratio for the asymmetric
traveling salesman problem. Math. Oper. Res., 31(2):245–252, 2006. 3
[7] N. Christofides. Worst-case analysis of a new heuristic for the travelling salesman problem.
Technical report, Graduate School of Industrial Administration, CMU, 1976. 2
[8] U. Feige and M. Singh. Improved approximation ratios for traveling salesperson tours and
paths in directed graphs. In Approximation, Randomization, and Combinatorial Optimization.
Algorithms and Techniques, 10th International Workshop, APPROX 2007, and 11th International
Workshop, RANDOM 2007, pages 104–118, 2007. 3
[9] A. M. Frieze, G. Galbiati, and F. Maffioli. On the worst-case performance of some algorithms
for the asymmetric traveling salesman problem. Networks, 12(1):23–39, 1982. 3
[10] S. O. Gharan and A. Saberi. The asymmetric traveling salesman problem on graphs with
bounded genus. In Proceedings of the Twenty-Second Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA 2011, pages 967–975, 2011. 3
24
[11] S. O. Gharan, A. Saberi, and M. Singh. A randomized rounding approach to the traveling
salesman problem. In IEEE 52nd Annual Symposium on Foundations of Computer Science,
FOCS 2011, pages 550–559, 2011. 2, 3
[12] M. Grigni, E. Koutsoupias, and C. H. Papadimitriou. An approximation scheme for planar
graph TSP. In 36th Annual Symposium on Foundations of Computer Science, FOCS 1995, pages
640–645, 1995. 2
[13] H. Kaplan, M. Lewenstein, N. Shafrir, and M. Sviridenko. Approximation algorithms for
asymmetric TSP by decomposing directed regular multigraphs. J. ACM,
52(4):602–626, 2005. 3
[14] M. Karpinski, M. Lampis, and R. Schmied. New inapproximability bounds for TSP. In
Algorithms and Computation - 24th International Symposium, ISAAC 2013, pages 568–578,
2013. 2, 3
[15] A. Marcus, D. A. Spielman, and N. Srivastava. Interlacing families II: Mixed characteristic
polynomials and the kadison-singer problem, 2013. 3
[16] T. Mömke and O. Svensson. Approximating graphic TSP by matchings. In IEEE 52nd
Annual Symposium on Foundations of Computer Science, FOCS 2011, pages 560–569, 2011. 2, 3
In 29th International Symposium on
[17] M. Mucha. 13/9-approximation for graphic TSP.
Theoretical Aspects of Computer Science, STACS 2012, pages 30–41, 2012. 2, 3
[18] A. Schrijver. Combinatorial Optimization - Polyhedra and Efficiency. Springer-Verlag, Berlin,
2003. 10
[19] A. Sebö and J. Vygen. Shorter tours by nicer ears: 7/5-approximation for the graph-
TSP, 3/2 for the path version, and 4/3 for two-edge-connected subgraphs. Combinatorica,
34(5):597–629, 2014. 2, 3
25
|
1212.6327 | 1 | 1212 | 2012-12-27T08:59:40 | Speeding up shortest path algorithms | [
"cs.DS",
"cs.CC"
] | Given an arbitrary, non-negatively weighted, directed graph $G=(V,E)$ we present an algorithm that computes all pairs shortest paths in time $\mathcal{O}(m^* n + m \lg n + nT_\psi(m^*, n))$, where $m^*$ is the number of different edges contained in shortest paths and $T_\psi(m^*, n)$ is a running time of an algorithm to solve a single-source shortest path problem (SSSP). This is a substantial improvement over a trivial $n$ times application of $\psi$ that runs in $\mathcal{O}(nT_\psi(m,n))$. In our algorithm we use $\psi$ as a black box and hence any improvement on $\psi$ results also in improvement of our algorithm.
Furthermore, a combination of our method, Johnson's reweighting technique and topological sorting results in an $\mathcal{O}(m^*n + m \lg n)$ all-pairs shortest path algorithm for arbitrarily-weighted directed acyclic graphs.
In addition, we also point out a connection between the complexity of a certain sorting problem defined on shortest paths and SSSP. | cs.DS | cs |
Speeding up shortest path algorithms
Andrej Brodnik12 and Marko Grgurovic1
1 University of Primorska, Department of Information Science and Technology,
Slovenia,
[email protected], [email protected]
2 University of Ljubljana, Faculty of Computer and Information Science, Slovenia,
Abstract. Given an arbitrary, non-negatively weighted, directed graph
G = (V, E) we present an algorithm that computes all pairs shortest
paths in time O(m∗n + m lg n + nTψ(m∗, n)), where m∗ is the number
of different edges contained in shortest paths and Tψ(m∗, n) is a run-
ning time of an algorithm to solve a single-source shortest path problem
(SSSP). This is a substantial improvement over a trivial n times appli-
cation of ψ that runs in O(nTψ(m, n)). In our algorithm we use ψ as a
black box and hence any improvement on ψ results also in improvement
of our algorithm.
Furthermore, a combination of our method, Johnson's reweighting tech-
nique and topological sorting results in an O(m∗n + m lg n) all-pairs
shortest path algorithm for arbitrarily-weighted directed acyclic graphs.
In addition, we also point out a connection between the complexity of a
certain sorting problem defined on shortest paths and SSSP.
Keywords: all pairs shortest path, single source shortest path
1
Introduction
Let G = (V, E) denote a directed graph where E is the set of edges and V is the
set of vertices of the graph and let ℓ(·) be a function mapping each edge to its
length. Without loss of generality, we assume G is strongly connected. To simplify
notation, we define m = E and n = V . Furthermore, we define d(u, v) for two
vertices u, v ∈ V as the length of the shortest path from u to v. A classic problem
in algorithmic graph theory is to find shortest paths. Two of the most common
variants of the problem are the single-source shortest path (SSSP) problem and
the all-pairs shortest path problem (APSP). In the SSSP variant, we are asked
to find the path with the least total length from a fixed vertex s ∈ V to every
other vertex in the graph. Similarly, the APSP problem asks for the shortest
path between every pair of vertices u, v ∈ V . A common simplification of the
problem constrains the edge length function to be non-negative, i.e. ℓ : E → R+,
which we assume throughout the rest of the paper, except where explicitly stated
otherwise. Additionally, we define ∀(u, v) /∈ E : ℓ(u, v) = ∞.
It is obvious that the APSP problem can be solved by n calls to an SSSP algo-
rithm. Let us denote the SSSP algorithm as ψ. We can quantify the asymptotic
time bound of such an APSP algorithm as O(nTψ(m, n)) and the asymptotic
2
space bound as O(Sψ(m, n)), where Tψ(m, n) is the time required by algorithm
ψ and Sψ(m, n) is the space requirement of the same algorithm. We assume that
the time and space bounds can be written as functions of m and n only, even
though this is not necessarily the case in more "exotic" algorithms that depend
on other parameters of G. Note, that if we are required to store the computed
distance matrix, then we will need at least Θ(n2) additional space. If we account
for this, then the space bound becomes O(Sψ(m, n) + n2).
In this paper we are interested in the following problem: what is the best way
to make use of an SSSP algorithm ψ when solving APSP? There exists some prior
work on a very similar subject in the form of an algorithm named the Hidden
Paths Algorithm [1]. The Hidden Paths Algorithm is essentially a modification of
Dijkstra's algorithm [2] to make it more efficient when solving APSP. Solving the
APSP problem by repeated calls to Dijkstra's algorithm requires O(mn+n2 lg n)
time using Fibonacci heaps [3]. The Hidden Paths Algorithm then reduces the
running time to O(m∗n + n2 lg n). The quantity m∗ represents the number of
edges (u, v) ∈ E such that (u, v) is included in at least one shortest path. In the
Hidden Paths Algorithm this is accomplished by modifying Dijkstra's algorithm,
so that it essentially runs in parallel from all vertex sources in G, and then reusing
the computations performed by other vertices. The idea is simple: we can delay
the inclusion of an edge (u, v) as a candidate for forming shortest paths until
vertex u has found (u, v) to be the shortest path to v. However, the Hidden
Paths Algorithm is limited to Dijkstra's algorithm, since it explicitly sorts the
shortest path lists by path lengths, through the use of a priority queue. As a
related algorithm, we also point out that a different measure U P related to
the number of so-called uniform paths has also been exploited to yield faster
algorithms [4].
In Sections 3, 4 and 5 we show that there is a method for solving APSP which
produces the shortest path lists of individual vertices in sorted order according
to the path lengths. The interesting part is that it can accomplish this without
the use of priority queues of any form and requires only an SSSP algorithm to be
provided. This avoidance of priority queues permits us to state a time complexity
relationship between a sorted variant of APSP and SSSP. Since it is very difficult
to prove meaningful lower bounds for SSSP, we believe this connection might
prove useful.
As a direct application of our approach, we show that an algorithm with a
similar time bound to the Hidden Paths Algorithm can be obtained. Unlike the
Hidden Paths Algorithm, the resulting method is general in that it works for any
SSSP algorithm, effectively providing a speed-up for arbitrary SSSP algorithms.
The proposed method, given an SSSP algorithm ψ, has an asymptotic worst-case
running time of O(m∗n + m lg n + nTψ(m∗, n)) and space O(Sψ(m, n) + n2). We
point out that the m∗n term is dominated by the nTψ(m∗, n) term, but we feel
that stating the complexity in this (redundant) form makes the result clearer to
the reader. For the case of ψ being Dijkstra's algorithm, this is asymptotically
equivalent to the Hidden Paths Algorithm. However, since the algorithm ψ is
arbitrary, we show that the combination of our method, Johnson's reweighting
3
technique [5] and topological sorting gives an O(m∗n + m lg n) APSP algorithm
for arbitrarily-weighted directed acyclic graphs.
2 Preliminaries
Throughout the paper and without loss of generality, we assume that we are not
interested in paths beginning in v and returning back to v. We have previously
defined the edge length function ℓ(·), which we now extend to the case of paths.
Thus, for a path π, we write ℓ(π) to denote its length, which corresponds to the
sum of the length of its edges.
Similar to the way shortest paths are discovered in Dijkstra's algorithm, we
rank shortest paths in nondecreasing order of their lengths. Thus, we call a path π
the k-th shortest path if it is at position k in the length-sorted shortest path list.
The list of paths is typically taken to be from a single source to variable target
vertices. In contrast, we store paths from variable sources to a single target.
By reversing the edge directions we obtain the same lists, but it is conceptually
simpler to consider the modified case. Thus, the k-th shortest path of vertex v
actually represents the k-th shortest incoming path into v. We will now prove
a theorem on the structure of shortest paths, which is the cornerstone of the
proposed algorithm.
Definition 1. (Ordered shortest path list Pv)
Let Pv = (π1, π2, ..., πn−1) denote the shortest path list for each vertex v ∈ V .
Then, let Pv,k denote the k-th element in the list Pv. The shortest path lists are
ordered according to path lengths, thus we have ∀i, j : 0 < i < j < n ⇒ ℓ(πi) ≤
ℓ(πj).
Theorem 1. To determine Pv,k we only need to know every edge {(u, v) ∈ E
∀u ∈ V } and the first k elements of each list Pu, where (u, v) ∈ E.
Proof. We assume that we have found the first k shortest paths for all neighbors
of v, and are now looking for the k-th shortest path into v, which we denote as
πk. There are two possibilities: either πk is simply an edge (u, v), in which case
we already have the relevant information, or it is the concatenation of some path
π and an edge (u, v). The next step is to show that π is already contained in
Pu,i where i ≤ k.
We will prove this by contradiction. Assume the contrary, that π is either
not included in Pu, or is included at position i > k. This would imply the
existence of some path π′ for which ℓ(π′) ≤ ℓ(π) and which is contained in Pu
at position i ≤ k. Then we could simply take πk to be the concatenation of
(u, v) and π′, thereby obtaining a shorter path than the concatenation of (u, v)
and π. However, this is not yet sufficient for a contradiction. Note that we may
obtain a path that is shorter, but connects vertices that have an even shorter
path between them, i.e. the path is not the shortest path between the source s
and target v.
To show that it does contradict our initial assumption, we point out that
Pu contains k shortest paths, therefore it contains shortest paths from k unique
4
sources. In contrast, the list Pv contains at most k − 1 shortest paths. By a
counting argument we have that there must exist a path π′, stored in Pu with
an index i ≤ k, which originates from a source vertex s that is not contained in
Pv, thereby obtaining a contradiction.
⊓⊔
3 The algorithm
Suppose we have an SSSP algorithm ψ and we can call it using ψ(V, E, s) where V
and E correspond to the vertex and edge sets, respectively and s corresponds to
the source vertex. The method we propose works in the fundamental comparison-
addition model and does not assume a specific kind of edge length function,
except the requirement that it is non-negative. However, the algorithm ψ that
is invoked can be arbitrary, so if ψ requires a different model or a specific length
function, then implicitly by using ψ, our algorithm does as well.
First we give a simpler variant of the algorithm, resulting in bounds O(mn +
nTψ(m∗, n)). We limit our interaction with ψ only to execution and reading its
output. To improve the running time we construct a graph G′ = (V ′, E ′) on
which we run ψ. There are two processes involved: the method for solving APSP
which runs on G, and the SSSP algorithm ψ which runs on G′. Let n′ = V ′
and m′ = E ′. We will maintain m′ ≤ m∗ + n and n′ = n + 1 throughout the
execution. There are n − 1 phases of the main algorithm, each composed of three
steps: (1) Prepare the graph G′; (2) Run ψ on G′; and (3) Interpret the results
of ψ.
Although the proposed algorithm effectively works on n−1 new graphs, these
graphs are similar to one another. Thus, we can consider the algorithm to work
only on a single graph G′, with the ability to modify edge lengths and introduce
new edges into G′. Initially we define V ′ = V ∪ {i}, where i is a new vertex
unrelated to the graph G. We create n new edges from i to every vertex v ∈ V ,
i.e. E ′ = Sv∈V {(i, v)}. We set the cost of these edges to some arbitrary value in
the beginning.
Definition 2. (Shortest path list for vertex v, Sv) The shortest path list of some
vertex v ∈ V is denoted by Sv. The length of Sv is at most n + 1 and contains
pairs of the form (a, δ) where a ∈ V ∪ {null} and δ ∈ R+. The first element of
Sv is always (v, 0), the last element plays the role of a sentinel and is always
(null, ∞). For all inner (between the first and the last element) elements (a, δ),
we require that δ = d(a, v). A list with k ≤ n − 1 inner elements:
Sv = (cid:0)(v, 0), (a1, δ1), (a2, δ2), ..., (ak, δk), (null, ∞)(cid:1).
Next we describe the data structures. Each vertex v ∈ V keeps its shortest
path list Sv, which initially contains only two pairs (v, 0) and (null, ∞). For each
edge (u, v) ∈ E, vertex v keeps a pointer p[(u, v)], which points to some element
in the shortest path list Su. Initially, each such pointer p[(u, v)] is set to point
to the first element of Su.
5
Definition 3. (Viable pair for vertex v) A pair (a, δ) is viable for a vertex v ∈ V
if ∀(a′, δ′) ∈ Sv : a 6= a′. Alternatively, if a = null we define the pair as viable.
Definition 4. (Currently best pair for vertex v, (av, δv)) A pair (av, δv) ∈ Sw,
where (w, v) ∈ E is the currently best pair for vertex v if and only if (av, δv) is
viable for v and: ∀(u, v) ∈ E : ∀(a′, δ′) ∈ Su : (a′, δ′) viable for v and δ′+ℓ(u, v) ≥
δv + ℓ(w, v).
We now look at the first step taken in each phase of the algorithm: prepa-
ration of the graph G′. In this step, each vertex v finds the currently best pair
(av, δv). To determine the currently best pair, a vertex v inspects the elements
pointed to by its pointers p[(u, v)] for each (u, v) ∈ E in the following manner:
For each pointer p[(u, v)], vertex v keeps moving the pointer to the next element
in the list Su until it reaches a viable pair, and takes the minimum amongst
these as per Definition 4. We call this process reloading.
Once reloaded we modify the edges in the graph G′. Let (av, δv) ∈ Sw where
(w, v) ∈ E be the currently best pair for vertex v, then we set ℓ(i, v) ← δv +
ℓ(w, v). Now we call ψ(V ′, E ′, i). Suppose the SSSP algorithm returns an array
Π[ ] of length n. Let each element Π[v] be a pair (c, δ) where δ is the length of
the shortest path from i to v, and c is the first vertex encountered on this path.
When determining the first vertex on the path we exclude i, i.e. if the path is
πv = {(i, v)} then Π[v].c = v. The inclusion of the first encountered vertex is a
mere convenience, and can otherwise easily be accomodated by examining the
shortest path tree returned by the algorithm. For each vertex v ∈ V we append
the pair (aΠ[v].c, Π[v].δ) to its shortest path list. Note, that the edges (i, v) ∈ E ′
are essentially shorthands for paths in G. Thus, aΠ[v].c represents the source of
the path in G. We call this process propagation.
After propagation, we modify the graph G′ as follows. For each vertex v ∈ V
such that Π[v].c = v, we check whether the currently best pair (av, δv) ∈ Su
that was selected during the reloading phase is the first element of the list Su. If
it is the first element, then we add the edge (u, v) into the set E ′. This concludes
the description of the algorithm. We formalize the procedure in pseudocode and
obtain Algorithm 1. To see why the algorithm correctly computes the shortest
paths, we prove the following two lemmata.
Lemma 1. For each vertex v ∈ V whose k-th shortest path was found during
the reloading step, ψ(V ′, E ′, i) finds the edge (i, v) to be the shortest path into v.
Proof. For the case when the k-th shortest path depends only on a path at
position j < k in a neighbor's list, the path is already found during the reloading
step. What has to be shown is that this is preserved after the execution of the
SSSP algorithm. Consider a vertex v ∈ V which has already found the k-th
shortest path during the reloading step. This path is represented by the edge
(i, v) of the same length as the k-th shortest path. Now consider the case that
some path, other than the edge (i, v) itself, would be found to be a better path
to v by the SSSP algorithm. Since each of the outgoing edges of i represents
a path in G, this would mean that taking this path and adding the remaining
6
V ′ := V ∪ {i}
E ′ := S∀v∈V {(i, v)}
best[ ] := new array [n] of pairs (a, δ)
solved[ ][ ] := new array [n][n] of boolean values
Initialize solved[ ][ ] to f alse
for all v ∈ V do
for all v ∈ V do
best[v] := (null, ∞)
for all u ∈ V s.t. (u, v) ∈ E do
while solved[v][p[(u, v)].a] do
end while
if p[(u, v)].δ + ℓ(u, v) < best[v].δ then
best[v].a := p[(u, v)].a
best[v].δ := p[(u, v)].δ + ℓ(u, v)
end if
Sv.append( (v, 0) )
end for
for k := 1 to n − 1 do
Algorithm 1 All-pairs shortest path
1: procedure APSP(V, E, ψ)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33: end procedure
end for
Π[ ] := ψ(V ′, E ′, i)
for all v ∈ V do
end for
ℓ(i, v) := best[v].δ
end if
end for
E ′ := E ′ ∪ (u, v)
end for
⊲ Reloading
p[(u, v)].next()
⊲ An end-of-list element is always viable
⊲ Considering only k − 1 neighboring paths
⊲ Propagation
Sv.append( (best[Π[v].c].a, Π[v].δ) )
solved[v][best[Π[v].c].a] := true
if Π[v].c = v and best[v] was the first element of some list Su then
7
edges used to reach v would consistute a shorter path than the k-th shortest
path of v. Let us denote the path obtained by this construction as π′. Clearly
this is a contradiction unless π′ is not the k-th shortest path, i.e. a shorter path
connecting the two vertices is already known.
Without loss of generality, assume that π′ = {(i, u), (u, v)}. However, ℓ(π′)
can only be shorter than ℓ(i, v) if v could not find a viable (non-null) pair in the
list Su, since otherwise a shorter path would have been chosen in the reloading
phase. This means that all vertex sources (the a component of a pair) contained
in the list Su are also contained in the list Sv. Therefore a viable pair for u must
also be a viable pair for v. This concludes the proof by contradiction, since the
⊓⊔
path obtained is indeed the shortest path between the two vertices.
Lemma 2. ψ(V ′, E ′, i) correctly computes the k-th shortest paths for all vertices
v ∈ V given only k − 1 shortest paths for each vertex.
Proof. The case when the k-th path requires only k − 1 neighboring paths to be
known has already been proven by the proof of Lemma 1. We now consider the
case when the k-th path depends on a neighbor's k-th path. If the k-th path of
vertex v requires the k-th path from the list of its neighbor u, then we know the
k-th path of u must be the same as that of v except for the inclusion of the edge
(u, v). The same argument applies to the dependency of vertex u on its neighbor's
list. Thus, the path becomes shorter after each such dependency, eventually
becoming dependent on a path included at position j < k in a neighbor's list
(this includes edges), which has already been found during the reloading step
and is preserved as the shortest path due to Lemma 1.
We now proceed in the same way that we obtained the contradiction in the
proof of Lemma 1, except it is not a contradiction in this case. What follows is
that any path from i to v in G′ which is shorter than ℓ(i, v) must represent a
viable pair for v. It is easy to see, then, that the shortest among these paths is
the k-th shortest path for v in G and also the shortest path from i to v in G′. ⊓⊔
3.1 Time and space complexity
First, we look at the time complexity. The main loop of Algorithm 1 (lines 7 -- 29)
performs n − 1 iterations. The reloading loop (lines 8 -- 20) considers each edge
(u, v) ∈ E which takes m steps. This amounts to O(mn). Since each shortest
path list is of length n+1, each pointer is moved to the next element n times over
the execution of the algorithm. There are m pointers, so this amounts to O(mn).
Algorithm ψ is executed n − 1 times. In total, the running time of Algorithm 1
is O(mn + nTψ(m∗, n)).
The space complexity of Algorithm 1 is as follows. Each vertex keeps track
of its shortest path list, which is of size n + 1 and amounts to Θ(n2) space
over all vertices. Since there are exactly m pointers in total, the space needed
for them is simply O(m). On top of the costs mentioned, we require as much
space as is required by algorithm ψ. In total, the combined space complexity for
Algorithm 1 is O(n2 + Sψ(m∗, n)).
8
3.2 Implications
We will show how to further improve the time complexity of the algorithm in
Section 4, but already at its current stage, the algorithm reveals an interesting
relationship between the complexity of non-negative SSSP and a stricter variant
of APSP.
Definition 5. (Sorted all-pairs shortest path SAPSP)
The problem SAPSP(m, n) is that of finding shortest paths between all pairs of
vertices in a non-negatively weighted graph with m edges and n vertices in the
form of Pv for each v ∈ V (see Definition 1).
Theorem 2. Let TSSSP denote the complexity of the single-source shortest path
problem on non-negatively weighted graphs with m edges and n vertices. Then
the complexity of SAPSP is at most O(nTSSSP ).
Proof. Given an algorithm ψ which solves SSSP, we can construct a solution to
SAPSP in time O(nTψ(m, n)) according to Algorithm 1, since the lists Sv found
by the algorithm are ordered by increasing distance from the source.
⊓⊔
What Theorem 2 says is that when solving APSP, either we can follow in the
footsteps of Dijkstra and visit vertices in increasing distance from the source
without worrying about a sorting bottleneck, or that if such a sorting bottleneck
exists, then it proves a non-trivial lower bound for the single-source case.
4
Improving the time bound
The algorithm presented in the previous section has a running time of O(mn +
nTψ(m∗, n)). We show how to bring this down to O(m∗n + m lg n + nTψ(m∗, n)).
We sort each set of incoming edges Ev = S(u,v)∈E{(u, v)} by edge lengths in
non-decreasing order. By using any off-the-shelf sorting algorithm, this takes
O(m lg n) time.
We only keep pointers p[(u, v)] for the edges which are shortest paths between
u and v, and up to one additional edge per vertex for which we do not know
whether it is part of a shortest path. Since edges are sorted by their lengths, a
vertex v can ignore an edge at position t in the sorted list Ev until the edge at
position t − 1 is either found to be a shortest path, or found not to be a shortest
path. For some edge (u, v) the former case simply corresponds to using the first
element, i.e. u, provided by p[(u, v)] as a shortest path. The latter case on the
other hand, is not using the first element offered by p[(u, v)], i.e. finding it is not
viable during the reloading phase. Whenever one of these two conditions is met,
we include the next edge in the sorted list as a pointer, and either throw away
the previous edge if it was found not to be a shortest path, or keep it otherwise.
This means the total amount of pointers is at most m∗ + n at any given time,
which is O(m∗), since m∗ is at least n. The total amount of time spent by the
algorithm then becomes O(m∗n + m lg n + nTψ(m∗, n)).
9
Theorem 3. Let ψ be an algorithm which solves the single-source shortest path
problem on non-negatively weighted graphs. Then, the all-pairs shortest path
problem on non-negatively weighted graphs can be solved in time O(m∗n+m lg n+
nTψ(m∗, n)) and space O(n2 + Sψ(m∗, n)) where Tψ(m, n) is the time required
by algorithm ψ on a graph with m edges and n nodes and Sψ(m, n) is the space
required by algorithm ψ on the same graph.
Proof. See discussion above and in Section 3.
⊓⊔
5 Directed acyclic graphs
A combination of a few techniques yields an O(m∗n + m lg n) APSP algorithm
for arbitrarily weighted directed acyclic graphs (DAGs). The first step is to
transform the original (possibly negatively-weighted) graph into a non-negatively
weighted graph through Johnson's [5] reweighting technique. Instead of using
Bellman-Ford in the Johnson step, we visit nodes in their topological order,
thus obtaining a non-negatively weighted graph in O(m) time. Next, we use
the improved time bound algorithm as presented in Section 4. For the SSSP
algorithm, we again visit nodes according to their topological order. Note that if
the graph G is a DAG then G′ is also a DAG. The reasoning is simple: the only
new edges introduced in G′ are those from i to each vertex v ∈ V . But since i
has no incoming edges, the acyclic property of the graph is preserved. The time
bounds become O(m) for Johnson's step and O(m∗n + m lg n + nTψ(m∗, n)) for
the APSP algorithm where Tψ(m∗, n) = O(m∗). Thus, the combined asymptotic
running time is O(m∗n + m lg n). The asymptotic space bound is simply Θ(n2).
Theorem 4. All-pairs shortest path on directed acyclic graphs can be solved in
time O(m∗n + m lg n) and Θ(n2) space.
Proof. See discussion above.
⊓⊔
6 Discussion
In this paper we have shown that the "standard" approach to solving APSP via
independent SSSP computations can be improved upon even if we know virtually
nothing about the SSSP algorithm itself. However, we should mention that in
recent years, asymptotically efficient algorithms for APSP have been formulated
in the so-called component hierarchy framework. These algorithms can be seen
as computing either SSSP or APSP. Our algorithm is only capable of speeding
up SSSP hierarchy algorithms, such as Thorup's [6], but not those which reuse
the hierarchy, such as Pettie's [7], Pettie-Ramachandran [8] or Hagerup's [9]
since our SSSP reduction requires modifications to the graph G′. These modi-
fications would require the hierarchy to be recomputed, making the algorithms
prohibitively slow. This raises the following question: is there a way to avoid
10
recomputing the hierarchy at each step, while keeping the number of edges in
the hierarchy O(m∗)?
Further, if there exists an o(mn) algorithm for the arbitrarily-weighted SSSP
problem, then by using Johnson's reweighting technique, our algorithm might
become an attractive solution for that case. For the general case, no such algo-
rithms are known, but for certain types of graphs, there exist algorithms with
an o(mn) asymptotic time bound [10,11].
Furthermore, we can generalize the approach used on DAGs. Namely, in Al-
gorithm 1 we can use an SSSP algorithm ψ that works on a specialized graph G,
as long our constructed graph G′ has these properties. Therefore, our algorithm
can be applied to undirected graphs, integer-weighted graphs, etc., but it cannot
be applied, for example, to planar graphs, since G′ is not necessarily planar.
Finally, we have shown a connection between the sorted all-pairs shortest
path problem and the single-source shortest path problem. If a meaningful lower
bound can be proven for SAPSP, then this would imply a non-trivial lower bound
for SSSP. Alternatively, if SAPSP can be solved in O(mn) time, then this implies
a Dijkstra-like algorithm for APSP, which visits vertices in increasing distance
from the source.
References
1. Karger, D., Koller, D., Phillips, S.J.: Finding the hidden path: time bounds for
all-pairs shortest paths. SIAM Journal on Computing 22(6) (1993) 1199 -- 1217
2. Dijkstra, E.W.: A note on two problems in connexion with graphs. Numerische
Mathematik 1 (1959) 269 -- 271
3. Fredman, M.L., Tarjan, R.E.: Fibonacci heaps and their uses in improved network
optimization algorithms. J. ACM 34(3) (July 1987) 596 -- 615
4. Demetrescu, C., Italiano, G.F.: Experimental analysis of dynamic all pairs shortest
path algorithms. ACM Transactions on Algorithms 2(4) (2006) 578 -- 601
5. Johnson, D.B.: Efficient algorithms for shortest paths in sparse networks. J. ACM
24(1) (January 1977) 1 -- 13
6. Thorup, M.: Undirected single-source shortest paths with positive integer weights
in linear time. J. ACM 46(3) (May 1999) 362 -- 394
7. Pettie, S.: A new approach to all-pairs shortest paths on real-weighted graphs.
Theor. Comput. Sci. 312(1) (January 2004) 47 -- 74
8. Pettie, S., Ramachandran, V.: A shortest path algorithm for real-weighted undi-
rected graphs. SIAM J. Comput. 34(6) (June 2005) 1398 -- 1431
9. Hagerup, T.:
Improved shortest paths on the word RAM.
In: Proceedings of
the 27th International Colloquium on Automata, Languages and Programming.
ICALP '00, London, UK, Springer-Verlag (2000) 61 -- 72
10. Goldberg, A.V.: Scaling algorithms for the shortest paths problem. In: Proceedings
of the fourth annual ACM-SIAM Symposium on Discrete algorithms. SODA '93,
Philadelphia, PA, USA, Society for Industrial and Applied Mathematics (1993)
222 -- 231
11. Gabow, H.N., Tarjan, R.E.: Faster scaling algorithms for network problems. SIAM
J. Comput. 18(5) (October 1989) 1013 -- 1036
|
1503.00049 | 1 | 1503 | 2015-02-28T01:20:55 | Algorithms for Longest Common Abelian Factors | [
"cs.DS"
] | In this paper we consider the problem of computing the longest common abelian factor (LCAF) between two given strings. We present a simple $O(\sigma~ n^2)$ time algorithm, where $n$ is the length of the strings and $\sigma$ is the alphabet size, and a sub-quadratic running time solution for the binary string case, both having linear space requirement. Furthermore, we present a modified algorithm applying some interesting tricks and experimentally show that the resulting algorithm runs faster. | cs.DS | cs |
Algorithms for Longest Common Abelian Factors
Ali Alatabbia, Costas S. Iliopoulosa, Alessio Langiua,b,∗, M. Sohel Rahmana,c
aDepartment of Informatics, King's College London, London, UK
bIAMC-CNR, National Research Council,Trapani, Italy
cA(cid:96)EDA Group, Department of CSE, BUET, Dhaka-1000, Bangladesh
Abstract
In this paper we consider the problem of computing the longest common abelian
factor (LCAF) between two given strings. We present a simple O(σ n2) time
algorithm, where n is the length of the strings and σ is the alphabet size, and a
sub-quadratic running time solution for the binary string case, both having lin-
ear space requirement. Furthermore, we present a modified algorithm applying
some interesting tricks and experimentally show that the resulting algorithm
runs faster.
1. Introduction
Abelian properties concerning words have been investigated since the very
beginning of the study of Formal Languages and Combinatorics on Words.
Abelian powers were first considered in 1961 by Erdos [Erd61] as a natural
generalization of usual powers. In 1966, Parikh [Par66] defined a vector having
length equal to the alphabet cardinality, which reports the number of occur-
rences of each alphabet symbol inside a given string. Later on, the scientific
community started referring to such a vector as the Parikh vector. Clearly, two
strings having the same Parikh vector are permutations of one another and
there is an abelian match between them.
Abelian properties of strings have recently grown tremendous interest among
the Stringology researchers and have become an involving topic of discussion in
the recent issues of the StringMasters meetings. Despite the fact that there are
not so many real life applications where comparing commutative sequence of
objects is relevant, abelian combinatorics has a potential role in filtering the
data in order to find potential occurrences of some approximate matches. For
instance, when one is looking for typing errors in a natural language, it can be
useful to select the abelian matches first and then look for swap of adjacent
or even near appearing letters. The swap errors and the inversion errors are
also very common in the evolutionary process of the genetic code of a living
organism and hence is often interesting from Bioinformatics perspective. Similar
applications can also be found in the context of network communications.
∗Corresponding author
Preprint submitted to Elsevier
August 15, 2018
In this paper, we focus on the problem of finding the Longest Common
Abelian Factor of two given strings. The problem is combinatorially interesting
and analogous to the Longest Common Substring (LCStr) problem for the usual
strings. The LCStr problem is a Historical problem and Dan Gusfield reported
the following in his book [Gus97, Sec. 7.4] regarding the belief of Don Knuth
about the complexity of the problem:
...in 1970 Don Knuth conjectured a linear time algorithm for this
problem would be impossible.
However, contrary to the above conjecture, decades later, a linear time solution
for the LCStr problem was in fact obtained by using the linear construction of
the suffix tree. For Stringology researchers this alone could be the motivation
for considering LCAF from both algorithmic and combinatorics point of view.
However, despite a number of works on abelian matching, to the best of our
knowledge, this problem has never been considered until very recently when it
was posed in the latest issue of the StringMasters, i.e., StringMasters 2013. To
this end, this research work can be seen as a first attempt to solve this problem
with the hope of many more to follow.
In this paper, we first present a simple solution to the problem running in
O(σ n2) time, where σ is the alphabet size (Section 3). Then we present a sub-
quadratic algorithm for the binary string case (Section 4). Both the algorithms
have linear space requirement. Furthermore, we present a modified algorithm
applying some interesting tricks (Section 5) and experimentally show that the
resulting algorithm runs in O(n log n) time (Section 6).
2. Preliminaries
An alphabet Σ of size σ > 0 is a finite set whose elements are called letters.
A string on an alphabet Σ is a finite, possibly empty, sequence of elements of Σ.
The zero-letter sequence is called the empty string, and is denoted by ε. The
length of a string S is defined as the length of the sequence associated with the
string S, and is denoted by S. We denote by S[i] the i-th letter of S, for all
1 ≤ i ≤ S and S = S[1 . .S]. A string w is a factor of a string S if there exist
two strings u and v, possibly empty, such that S = uwv. A factor w of a string
S is proper if w (cid:54)= S. If u = ε (v = ε), then w is a prefix (suffix) of S.
Given a string S over the alphabet Σ = {a1, . . . aσ}, we denote by Saj
the number of aj's in S, for 1 ≤ j ≤ σ. We define the Parikh vector of S as
PS = (Sa1, . . .Saσ ).
In the binary case, we denote Σ = {0, 1}, the number of 0's in S by S0, the
number of 1's in S by S1 and the Parikh vector of S as PS = (S0,S1). We
now focus on binary strings. The general alphabet case will be considered later.
For a given binary string S of length n, we define an n × n matrix MS as
follows. Each row of MS is dedicated to a particular length of factors of S.
So, Row (cid:96) of MS is dedicated to (cid:96)-length factors of S. Each column of MS
is dedicated to a particular starting position of factors of S. So, Column i of
2
MS is dedicated to the position i of S. Hence, MS[(cid:96)][i] is dedicated to the (cid:96)-
length factor that starts at position i of S and it reports the number of 1's of
that factor. Now, MS[(cid:96)][i] = m if and only if the (cid:96)-length factor that starts at
position i of S has a total of m 1's, that is, S[i . . i + (cid:96) − 1]1 = m. We formally
define the matrix MS as follows.
Definition 1. Given a binary string S of length n, MS is an n× n matrix such
that MS[(cid:96)][i] = S[i . . i + (cid:96) − 1]1, for 1 ≤ (cid:96) ≤ n and 1 ≤ i ≤ (n − (cid:96) + 1), and
MS[(cid:96)][i] = 0, otherwise.
In what follows, we will use MS[(cid:96)] to refer to Row (cid:96) of MS. Assume that
we are given two strings A and B on an alphabet Σ. For the sake of ease, we
assume that A = B = n. We want to find the length of a longest common
abelian factor between A and B.
Definition 2. Given two strings A and B over the alphabet Σ, we say that w
is a common abelian factor for A and B if there exist a factor (or substring) u in
A and a factor v in B such that Pw = Pu = Pv. A common abelian factor of the
highest length is called the Longest Common Abelian Factor (LCAF) between
A and B. The length of LCAF is referred to as the LCAF length.
In this paper we study the following problem.
Problem 1 (LCAF Problem). Given two strings A and B over the alphabet
Σ, compute the length of an LCAF and identify some occurrences of an LCAF
between A and B .
Assume that the strings A and B of length n are given. Now, suppose that
the matrices MA and MB for the binary strings A and B have been computed.
Now we have the following easy lemma that will be useful for us later.
Lemma 2. There is a common abelian factor of length (cid:96) between A and B if
and only if there exists p, q such that 1 ≤ p, q ≤ n−(cid:96)+1 and MA[(cid:96)][p] = MB[(cid:96)][q].
Proof. Suppose there exists p, q such that 1 ≤ p, q ≤ n− (cid:96) + 1 and MA[(cid:96)][p] =
MB[(cid:96)][q]. By definition this means A[p . . p + (cid:96) − 1]1 = B[q . . q + (cid:96) − 1]1. So
there is a common abelian factor of length (cid:96) between A and B. The other way
is also obvious by definition.
Clearly, if we have MA and MB we can compute the LCAF by identifying the
highest (cid:96) such that there exists p, q having 1 ≤ p, q ≤ n − (cid:96) + 1 and MA[(cid:96)][p] =
MB[(cid:96)][q]. Then we can say that the LCAF between A and B is either A[p . . p +
(cid:96) − 1] or B[q . . q + (cid:96) − 1] having length (cid:96).
We now generalize the definition of the matrix MS for strings over a fixed
size alphabet Σ = {a1, . . . aσ} by defining an n × n matrix MS of (σ − 1)-
length vectors. MS[(cid:96)][i] = V(cid:96),i, where V(cid:96),i[j] = S[i . . i + (cid:96) − 1]aj , for 1 ≤ (cid:96) ≤ n,
1 ≤ i ≤ (n−(cid:96)+1) and 1 ≤ j < σ, and V(cid:96),i[j] = 0, otherwise. We will refer to the
j-th element of the array V(cid:96),i of the matrix MS by using the notation MS[(cid:96)][i][j].
3
Notice that the last component of a Parikh vector is determined by using the
length of the string and all the other components of the Parikh vector. Now,
MS[(cid:96)][i][j] = m if and only if the (cid:96)-length factor that starts at position i of S has
a total of m aj's, that is S[i . . i + (cid:96)− 1]aj = m. Clearly, we can compute MS[(cid:96)]
using the following steps. Similar to the binary case, the above computation
runs in linear time because we can compute S[i + 1 . . i + 1 + (cid:96) − 1]aj from
S[i . . i + (cid:96) − 1]aj in constant time by simply decrementing the S[i] component
and incrementing the S[i + (cid:96)] one.
3. A Quadratic Algorithm
A simple approach for finding the LCAF length considers computing, for
1 ≤ (cid:96) ≤ n, the Parikh vectors of all the factors of length (cid:96) in both A and B,
i.e., MA[(cid:96)] and MB[(cid:96)]. Then, we check whether MA[(cid:96)] and MB[(cid:96)] have non-
empty intersection. If yes, then (cid:96) could be the LCAF length. So, we return the
highest of such (cid:96). Moreover, if one knows a Parikh vector having the LCAF
length belonging to such intersection, a linear scan of A and B produces one
occurrence of such a factor. The asymptotic time complexity of this approach
is O(σ n2) and it requires O(σ n log n) bits of extra space. The basic steps are
outlined as follows.
1. For (cid:96) = 1 to n do the following
2.
3.
4.
5.
For i = 1 to n − (cid:96) + 1 do the following
compute MA[(cid:96)][i] and MB[(cid:96)][i]
If MA[(cid:96)](cid:84) MB[(cid:96)] (cid:54)= ∅ then
LCAF = (cid:96)
It is well known that, for fixed length (cid:96), one can compute all the Parikh
vectors in linear time and store them in O(σ n log n) bits. Now once MA and MB
are computed, we simply need to apply the idea of Lemma 2. The idea is to check
for all values of (cid:96) whether there exists a pair p, q such that 1 ≤ p, q ≤ n − (cid:96) + 1
and MA[(cid:96)][p] = MB[(cid:96)][q]. Then return the highest value of (cid:96) and corresponding
values of p, q.
In the binary case, a Parikh vector is fully represented by just one arbitrary
chosen component. Hence, the set of Parikh vectors of binary factors is just
a one dimension list of integers that can be stored in O(n log n) bits, since we
have n values in the range [0 . . n]. The intersection can be accomplished in two
steps. First, we sort the MA[(cid:96)] and MB[(cid:96)] rows in O(n) time by putting them in
two lists and using the classic Counting Sort algorithm [CLRS01, Section 8.2].
Then, we check for a non empty intersection with a simple linear scan of the
two lists in linear time by starting in parallel from the beginning of the two lists
and moving forward element by element on the list having the smallest value
among the two examined elements. A further linear scan of MA[(cid:96)] and MB[(cid:96)]
4
will find the indexes p, q of an element of the not empty intersection. This gives
us an O(n2) time algorithm requiring O(n log n) bits of space for computing an
LCAF of two given binary strings.
In the more general case of alphabet greater than two, comparing two Parikh
vectors is no more a constant time operation and checking for empty intersec-
tions is not a trivial task. In fact, sorting the set of vectors requires a full order
to be defined. We can define an order component by component giving more
value to the first component, then to the second one and so on. More formally,
we define x < y, with x, y ∈ Nσ, if there exist 1 ≥ k ≥ σ such that x[k] < y[k]
and, for any i with 1 ≤ i < k, x[i] = y[i]. Notice that comparing two vectors
will take BO(σ) time.
Now, one can sort two list of n vectors of dimension σ − 1, i.e., MA[(cid:96)] and
MB[(cid:96)], in O(σ n) by using n comparisons taking O(σ) each. Therefore, now the
algorithm runs in O(σ n2) time using O(σ n log σ) bits of extra space.
4. A Sub-quadratic Algorithm for the Binary Case
In Section 3, we have presented an O(n2) algorithm to compute the LCAF
between two binary strings and two occurrences of common abelian factors, one
in each string, having LCAF length. In this section, we show how we can achieve
a better running time for the LCAF problem. We will make use of the recent
data structure of Moosa and Rahman [MR10] for indexing an abelian pattern.
The results of Moosa and Rahman [MR10] is presented in the form of following
lemmas with appropriate rephrasing to facilitate our description.
Lemma 3. (Interpolation lemma). If S1 and S2 are two substrings of a string
S on a binary alphabet such that (cid:96) = S1 = S2, i = S11, j = S21, j > i + 1,
then, there exists another substring S3 such that (cid:96) = S3 and i < S31 < j.
Lemma 4. Suppose we are given a string S of length n on a binary alpha-
bet. Suppose that maxOne(S, (cid:96)) and minOne(S, (cid:96)) denote, respectively, the
maximum and minimum number of 1's in any substring of S having length
(cid:96). Then, for all 1 ≤ (cid:96) ≤ n, maxOne(S, (cid:96)) and minOne(S, (cid:96)) can be computed
in O(n2/ log n) time and linear space.
A result similar to Lemma 3 is contained in the paper of Cicalese et al.
[CFL09, Lemma 4], while the result of Lemma 4 has been discovered simulta-
neously and independently by Moosa and Rahman [MR10] and by Burcsi et al.
[BCFL10]. In addition to the above results we further use the following lemma.
Lemma 5. Suppose we are given two binary strings A, B of length n each.
There is a common abelian factor of A and B having length (cid:96) if and only if
maxOne(B, (cid:96)) ≥ minOne(A, (cid:96)) and maxOne(A, (cid:96)) ≥ minOne(B, (cid:96)).
Proof. Assume that minA = minOne(A, (cid:96)), maxA = maxOne(A, (cid:96)), minB =
minOne(B, (cid:96)), maxB = maxOne(B, (cid:96)). Now by Lemma 3, for all minA ≤ kA ≤
5
maxA, we have some (cid:96)-length substrings A(kA) of A such that A(kA)1 = kA.
Similarly, for all minB ≤ kB ≤ maxB, we have some (cid:96)-length factors B(k)
of B such that B(kB)1 = kB. Now, consider the range [minA . . maxA] and
[minB . . maxB]. Clearly, these two ranges overlap if and only if maxB (cid:54)< minA
and maxA (cid:54)< minB. If these two ranges overlap then there exists some k such
that minA ≤ k ≤ maxA and minB ≤ k ≤ maxB. Then we must have some
substring (cid:96)-length factors A(k) and B(k). Hence the result follows.
Let us now focus on devising an algorithm for computing the LCAF given
two binary strings A and B of length n. For all 1 ≤ (cid:96) ≤ n, we compute
maxOne(A, (cid:96)), minOne(A, (cid:96)), maxOne(B, (cid:96)) and minOne(B, (cid:96)) in O(n2/ log n)
time (Lemma 4). Now we try to check the necessary and sufficient condition of
Lemma 5 for all 1 ≤ (cid:96) ≤ n starting from n down to 1. We compute the highest
(cid:96) such that
[minOne(A, (cid:96)) . . maxOne(A, (cid:96))] and [minOne(B, (cid:96)) . . maxOne(B, (cid:96))] overlap.
Suppose that K is the set of values that is contained in the above overlap, that
is K = { k k ∈ [minOne(A, (cid:96)) . . maxOne(A, (cid:96))] and k ∈ [minOne(B, (cid:96)) . .
maxOne(B, (cid:96))] }. Then by Lemma 5, we must have a set S of common abelian
factors of A, B such that for all S ∈ S, S = (cid:96). Since we identify the highest
(cid:96), the length of a longest common factor must be (cid:96), i.e., LCAF length is (cid:96).
Additionally, we have further identified the number of 1's in such longest factors
in the form of the set K. Also, note that for a k ∈ K we must have a factor
S ∈ S such that S1 = k.
Now let us focus on identifying an occurrence of the LCAF. There are a
number of ways to do that. But a straightforward and conceptually easy way is
to run the folklore (cid:96)-window based algorithm in [MR10] on the strings A and B
to find the (cid:96)-length factor with number of 1's equal to a particular value k ∈ K.
The overall running time of the algorithm is deduced as follows. By Lemma 4,
the computation of maxOne(A, (cid:96)), minOne(A, (cid:96)), maxOne(B, (cid:96)) and minOne(B, (cid:96))
can be done in O(n2/ log n) time and linear space. The checking of the condition
of Lemma 5 can be done in constant time for a particular value of (cid:96). Therefore,
in total, it can be done in O(n) time. Finally, the folklore algorithm requires
O(n) time to identify an occurrence (or all of them) of the factors. In total the
running time is O(n2/ log n) and linear space.
5. Towards a Better Time Complexity
In this section we discuss a simple variant of the quadratic algorithm pre-
sented in 3. We recall that the main idea of the quadratic solution is to find the
greatest (cid:96) with MA[(cid:96)](cid:84) MB[(cid:96)] (cid:54)= ∅. The variant we present here is based on the
following two simple observations:
1. One can start considering sets of factors of decreasing lengths;
6
2. When an empty intersection is found between MA[(cid:96)] and MB[(cid:96)], some
rows can possibly be skipped based on the evaluation of the gap between
MA[(cid:96)] and MB[(cid:96)].
skip trick. Assume that MA[(cid:96)] and MB[(cid:96)] have been computed and MA[(cid:96)](cid:84) MB[(cid:96)] =
The first observation is trivial. The second observation is what we call the
∅ have been found. It is easy to see that, for any starting position i and for any
component j (i.e., a letter aj), we have
MA[(cid:96)][i][j] − 1 ≤ MA[(cid:96) − 1][i][j] ≤ MA[(cid:96)][i][j] + 1
Exploiting this property, we keep track, along the computation of MA[(cid:96)] and
MB[(cid:96)], of the minimum and maximum values that appear in Parikh vectors of
factors of length (cid:96). We use four arrays indexed by σ, namely minA, maxA,
minB, maxB. Notice that such arrays do not represent Parikh vectors as they
just contain min and max values component by component. Formally, minA[j] =
min{MA[(cid:96)][i][j]}, for any i = 1, . . . (cid:96) + 1. The others have similar definitions.
We compare, component by component, the range of aj in A and B and
j=1 (minB[j] − maxA[j]), assuming minB[j] ≥
we skip as many Rows as maxσ−1
maxA[j] (swap A and B, otherwise). The modified algorithm is reported in
Algorithm 1.
Note that the tricks employed in our skip trick algorithm are motivated by
the fact that the expected value of the LCAF length of an independent and
identically distributed (i.i.d.) source is exponentially close to n according to
classic Large Deviation results [Ell85]. The same result is classically extended
to an ergodic source and it is meant to be a good approximation for real life
data when the two strings follow the same probability distribution. Based on
this, we have the following conjecture.
Conjecture 6. The expected length of LCAF between two strings A, B drawn
from an i.i.d. source is LCAFavg = n − O(log n), where A = B = n, and the
number of computed Rows in Algorithm 1 is Ø(log n) in average.
Finally, we will make use of one more trick, that is, computing the first vector
of the current row in constant time from the first vector of the previous row,
when we skip some rows, instead of computing the new row from scratch, we
can use the first vector of the row below to compute the first vector of the new
row. When we compute the rows we need, we will just populate the required
two lists and save a copy of the first vector of the computed row as we will need
it along the next iterative steps as shown in Algorithm 2.
For instance, if we know M[(cid:96)] and we jump to M[(cid:96)−3], i.e., we skip M[(cid:96)−1]
and M[(cid:96)− 2], we take M[(cid:96)][1] and compute in constant time M[(cid:96)− 1][1], M[(cid:96)−
2][1], then again compute M[(cid:96) − 3][1]. From M[(cid:96)][1], to compute M[(cid:96) − 1][1],
we have to subtract 1 from the vector M[(cid:96)][1] at index s[(cid:96)], that is the last
character of the factor of length (cid:96) starting at position 1 (i.e., M[(cid:96)][1]). For
example, consider s = aacgcctaatcg, we have M[12][1] = (4a, 4c, 2g, 2t) and
M[11][1] = (4a, 4c, 1g, 2t), i.e., (4a, 4c, 2g, 2t) minus 1g.
7
Algorithm 1 Compute LCAF of x and y using the skip trick.
1: function ComputeLCAF(x, y)
2:
3:
4:
5:
6:
set (cid:96) = n = x
set found = False
compute maxx = MAX(x, (cid:96)), maxy = MAX(y, (cid:96))
compute minx = MIN(x, (cid:96)), miny = MIN(y, (cid:96))
if maxx == maxy then
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
else
found = True
(cid:96) = (cid:96)− SKIP(minx, maxx, miny, maxy)
end if
while (found == False) AND ((cid:96) ≥ 0) do
compute maxx = MAX(x, (cid:96)), minx = MIN(x, (cid:96))
compute maxy = MAX(y, (cid:96)), miny = MIN(y, (cid:96))
compute listx = Mx[(cid:96)], listy = My[(cid:96)]
sort listx, listy
compute listx(cid:84) listy
if listx(cid:84) listy (cid:54)= ∅ then
end if
(cid:96) = (cid:96)− SKIP(minx, maxx, miny, maxy)
found = True
break
end while
return (cid:96)
18:
19:
20:
21:
22:
23:
24: end function
25: function max(s, (cid:96))
26:
27:
28:
29:
30:
count[s[i]]++
int count[σ],max[σ]
for (i = 1; i ≤ (cid:96); i++) do
end for
max = count
for (i = (cid:96); i < s − (cid:96); i++) do
31:
32:
33:
34:
35:
36:
count[s[i − 1]]- -
count[s[i + (cid:96) − 1]]++
if count[s[i + (cid:96) − 1]] ≥ max[s[i + (cid:96) − 1]] then
max[s[i + (cid:96) − 1]] = count[s[i + (cid:96) − 1]]
end if
end for
37:
return max
38:
39: end function
40: function skip(minx, maxx, miny, maxy)
41:
42:
43:
int gap[σ − 1]
for (j = 1; j < σ; j++) do
if maxx[j] >= miny[j] then
gap[j] = minx[j] − maxy[j]
gap[j] = miny[j] − maxx[j]
8
else
44:
45:
46:
47:
48:
49:
50: end function
end if
end for
return max(gap)
6. Experiments
We have conducted some experiments to analyze the behaviour and running
time of our skip trick algorithm in practice. The experiments have been run on
a Windows Server 2008 R2 64-bit Operating System, with Intel(R) Core(TM)
i7 2600 processor @ 3.40GHz having an installed memory (RAM) of 8.00 GB.
Codes were implemented in C# language using Visual Studio 2010.
Figure 1: Plot of the average number of rows computed executing Algorithm 1
on all the strings of length 2, 3, . . . 16 over the binary alphabet.
Our first experiment have been carried out principally to verify our rationale
behind using the skip trick. We experimentally evaluated the expected number
of rows computed in average by using the skip trick of Algorithm 1.
Figure 2: Plot of the average number of rows computed executing Algorithm 1
on both genomic and random datasets over the DNA alphabet.
Figure 1 shows the average number of rows computed executing Algorithm
1 on all the strings of length 2, 3, . . . 16 over the binary alphabet. Naive method
line refers to the number of rows used without the skip trick, but starting from
9
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 2 4 6 8 10 12 14 16 Average Count Length Skip trick Naive method log (n) 0 1 2 3 4 5 6 7 8 9 10 11 10 40 70 100 130 160 190 220 250 280 310 340 370 400 430 460 490 520 550 580 610 640 670 700 730 760 790 820 850 880 910 940 970 1000 Average Count Length log(n) Computed Rows (genomic data) Computed Rows (random data) (cid:96) = n and decreasing (cid:96) by one at each step. Notice that the skip trick line is
always below the log n line.
To this end we have conducted an experiment to evaluate the expected num-
ber of rows computed by our skip trick algorithm. In particular, we have imple-
mented the skip trick algorithm as well as the naive algorithm and have counted
the average number of rows computed by the algorithms on all the strings of
length 2, 3, . . . 16 on binary alphabet. The results are reported in Figure 1. It
shows that the computed rows of x, y, starting from (cid:96) = n to (cid:96) = n− log n, sum
up to O(log n).
√
On the other hand, to reach a conclusion in this aspect we would have to
increase the value of n in our experiment to substantially more than 64; for
n is just above log n. Regrettably, limitation of computing power
n = 64,
prevents us from doing such an experiment. So, we resort to two more (non-
comprehensive) experimental setup as follows to check the practical running
time of the skip trick algorithm.
Figure 3: Plot of the average number of rows computed executing Algorithm 1
on sequences taken from the Homo sapiens genome.
Furthermore, we conduct our experiments on two datasets, real genomic data
and random data. We have taken a sequence (S) from the Homo sapiens genome
(250MB) for the former dataset. The latter dataset is generated randomly on
the DNA alphabet (i.e., Σ = {a, c, g, t}). In particular, Here we have run the
skip trick algorithm on 2 sets of pairs of strings of lengths 10, 20, .., 1000. For the
genomic dataset, these pairs of strings have been created as follows. For each
length (cid:96), (cid:96) ∈ {10, 20, .., 1000} two indexes i, j ∈ [1..x − (cid:96)] have been randomly
selected to get a pair of strings S[i..i + (cid:96) − 1],S[j..i + (cid:96) − 1], each of length (cid:96).
A total of 1000 pairs of strings have been generated in this way for each length
(cid:96) and the skip trick algorithm has been run on these pairs to get the average
results. On the other hand for random dataset, we simply generate the same
number of strings pairs randomly and run the skip trick algorithm on each pair
of strings and get the average results for each length group. In both cases, we
basically count the numbers of computed rows.
10
0 2000 4000 6000 8000 10000 12000 10 50 90 130 170 210 250 290 330 370 410 450 490 530 570 610 650 690 730 770 810 850 890 930 970 Average Count Length Genomic Data First Vector Computed Rows Total n log(n) Figure 4: Plot of the average number of rows computed executing Algorithm 1
on randomly generated sequences over the alphabet Σ = {a, c, g, t}.
Figure 2 shows the average number of rows computed executing Algorithm
1 on both genomic and random datasets over the DNA alphabet (i.e., Σ =
{a, c, g, t}). Notice that the skip trick line is always below the log n line. Figure
2 shows that the computed rows of x, y, starting from (cid:96) = n to (cid:96) = n − log n,
sum up to O(log n).
We experimentally evaluated the computing of the first vector and the ex-
pected number of rows computed in average by employing the first vector trick
(Algorithm 2). We have used the same experiment configuration as the above.
The average number of rows and of the first vector computed executing Algo-
rithm 2 on both genomic and random datasets over the DNA alphabet (i.e.,
Σ = {a, c, g, t}). In both cases, we basically count the numbers of computed
rows and the first vector. The results are illustrated in Figures 3 and 4.
In both cases, The figures report the average count of computed rows (Num-
ber of Rows), the average count of the first vector (First Vector) and the sum-
mation of these two counts (Total). It also shows the n log n curve. Both of
the figures show that the algorithm computed the first vector of the visited
rows in O(n) and the total running time for Algorithm 2 would be O(n log n)
in practice.
Since any row computation takes O(σ n), this suggests an average time
complexity of O(σ n log n), i.e., O(n log n) for a constant alphabet.
7. Conclusion
In this paper we present a simple quadratic running time algorithm for the
LCAF problem and a sub-quadratic running time solution for the binary string
case, both having linear space requirement. Furthermore, we present a variant
of the quadratic solution that is experimentally shown to achieve a better time
complexity of O(n log n).
11
0 2000 4000 6000 8000 10000 12000 10 40 70 100 130 160 190 220 250 280 310 340 370 400 430 460 490 520 550 580 610 640 670 700 730 760 790 820 850 880 910 940 970 1000 Average Count Length Random Data First Vector Computed Rows Total n log(n) Algorithm 2 Compute LCAF of x and y using the first vector trick.
1: function first(s, (cid:96))
2:
3:
4:
5:
int first[σ]
for (i = 1; i < (cid:96); i++) do
first[s[i]]++
end for
return first
6:
7: end function
8: function row(s, (cid:96), first)
9:
10:
11:
12:
int row[σ]
row =first
for (i = 1; i < s − (cid:96); i++) do
row[s[i − 1]]- -
row[s[i + l − 1]]++
13:
end for
14:
return row
15:
16: end function
17: function ComputeLCAF(x, y)
18:
19:
set (cid:96) = n = x
set found = False
compute f irstx = F IRST (x, (cid:96))
compute f irsty = F IRST (y, (cid:96))
while (found == False) AND ((cid:96) ≥ 0) do
compute rowx = ROW (x, (cid:96), f irstx)
compute rowy = ROW (y, (cid:96), f irsty)
compute listx = Mx[(cid:96)], listy = My[(cid:96)]
sort listx, listB
compute listx(cid:84) listy
if listx(cid:84) listy (cid:54)= ∅ then
found = True
break
end if
compute maxx = M AX(x, (cid:96)), minx = M IN (x, (cid:96))
compute maxy = M AX(y, (cid:96)), miny = M IN (y, (cid:96))
(cid:96) = (cid:96)− SKIP(minx, maxx, miny, maxy)
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
end while
return (cid:96)
36:
37: end function
Acknowledgement
We thank Thierry Lecroq and Arnaud Lefebvre for proposing the LCAF
problem and the participants of Stringmasters 2013 meetings for helping us to
get more involved with this topic.
12
References
[BCFL10] Peter Burcsi, Ferdinando Cicalese, Gabriele Fici, and Zsuzsanna
Lipt´ak. On table arrangements, scrabble freaks, and jumbled pattern
matching. In Paolo Boldi and Luisa Gargano, editors, FUN, volume
6099 of Lecture Notes in Computer Science, pages 89 -- 101. Springer,
2010.
[CFL09] Ferdinando Cicalese, Gabriele Fici, and Zsuzsanna Lipt´ak. Searching
for jumbled patterns in strings. In Jan Holub and Jan Zd´arek, editors,
Stringology, pages 105 -- 117. Prague Stringology Club, Department of
Computer Science and Engineering, Faculty of Electrical Engineer-
ing, Czech Technical University in Prague, 2009.
[CLRS01] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and
Clifford Stein. Introduction to Algorithms. MIT Press, Cambridge,
MA, second edition, 2001.
[Ell85] R. S. Ellis. Entropy, Large Deviations, and Statistical Mechanics.
Springer, 1985.
[Erd61] P. Erdos. Some unsolved problems. Magyar Tud. Akad. Mat. Kutato.
Int. Kozl., 6:221 -- 254, 1961.
[Gus97] Dan Gusfield. Algorithms on Strings, Trees, and Sequences - Com-
puter Science and Computational Biology. Cambridge University
Press, 1997.
[MR10] Tanaeem M. Moosa and M. Sohel Rahman. Indexing permutations
for binary strings. Inf. Process. Lett., 110(18-19):795 -- 798, September
2010.
[Par66] R. J. Parikh. On context-free languages. J. Assoc. Comput. Mach.,
13(4):570 -- 581, 1966.
13
|
1503.07192 | 1 | 1503 | 2015-03-24T20:39:43 | Shortest-Path Queries in Planar Graphs on GPU-Accelerated Architectures | [
"cs.DS"
] | We develop an efficient parallel algorithm for answering shortest-path queries in planar graphs and implement it on a multi-node CPU/GPU clusters. The algorithm uses a divide-and-conquer approach for decomposing the input graph into small and roughly equal subgraphs and constructs a distributed data structure containing shortest distances within each of those subgraphs and between their boundary vertices. For a planar graph with $n$ vertices, that data structure needs $O(n)$ storage per processor and allows queries to be answered in $O(n^{1/4})$ time. | cs.DS | cs |
Shortest-Path Queries in Planar Graphs
on GPU-Accelerated Architectures
Guillaume Chapuis and Hristo Djidjev
Los Alamos National Laboratory, Los Alamos, NM 87545, USA
{gchapuis,djidjev}@lanl.gov
Abstract. We develop an efficient parallel algorithm for answering shortest-
path queries in planar graphs and implement it on a multi-node CPU/GPU
clusters. The algorithm uses a divide-and-conquer approach for decom-
posing the input graph into small and roughly equal subgraphs and con-
structs a distributed data structure containing shortest distances within
each of those subgraphs and between their boundary vertices. For a pla-
nar graph with n vertices, that data structure needs O(n) storage per
processor and allows queries to be answered in O(n1/4) time.
Keywords: shortest path problems, graph algorithms, distributed com-
puting, GPU computing, graph partitioning
1
Introduction
Finding shortest paths (SPs) in graphs has applications in transportation, social
network analysis, network routing, and robotics, among others. The problem
asks for a path of shortest length between one or more pairs of vertices. There
are many algorithm for solving SP problems sequentially. Dijkstra's algorithm [2]
finds the distances between a source vertex v and all other vertices of the graph
in O(m log n) time, where n and m are the numbers of the vertices and edges
of the graph, respectively. It can also be used to find efficiently the distance
between a pair of vertices. This algorithm is nearly optimal (within a logarithmic
factor), but has irregular structure, which makes it hard to implement efficiently
in parallel. Floyd-Warshall's algorithm, on the other hand, finds the distances
between all pairs of vertices of the graph in O(n3) time, which is efficient for dense
(m = Θ(n2)) graphs, has a regular structure good for parallel implementation,
but is inefficient for sparse (m = O(n)) graphs such as planar graphs.
In this paper we are considering the query version of the problem. It asks
to construct a data structure that will allow to answer any subsequent distance
query fast. A distance query asks, given an arbitrary pair of vertices v, w, to
compute dist(v, w). This problem has applications in web mapping services such
as MapQuest and Google Maps. There is a tradeoff between the size of the data
structure and the time for answering a query. For instance, Dijkstra's algorithm
gives a trivial solution of the query version of the SP problem with (small)
O(n + m) space (for storing the input graph), but large O(m log n) query time
(for running Dijkstra's algorithm with a source the first query vertex). On the
other end of the spectrum, Floyd-Warshall's algorithm can be used to construct
a (large) O(n2) data structure (the distance matrix) allowing (short) O(1) query
time (retrieving the distance from the data base). However, for very large graphs,
the O(n2) space requirement is impractical. We are interested in an algorithm
that needs significantly less than than O(n2) space, but will answer queries
faster than Disjkstra's algorithm. Our algorithm will use the structure of planar
graphs for increased efficiency, as most road networks are planar or near-planar,
and will also be highly parallelizable, making use of the features available in
modern high-performance clusters and specialized processors such as the GPUs.
The query version for shortest path queries in planar graphs was proposed
in [3] and after that different aspects of the problem were studied by multiple
authors, e.g., [6,1,8,9]. Here we present the first distributed implementation for
solving the problem that is designed to make use of the potential for parallelism
offered by GPUs. Our solution makes use of the fast parallel algorithm for com-
puting shortest paths in planar graphs from [4], resulting in asymptotically faster
and also shown to be efficient in practice.
2 Preliminaries
Given a graph G with a weight wt(e) on each edge e, the length of a path p is
the sum of the weights of the edges of the path. The single-pair shortest path
problem (SPSP ) is, given a pair v, w of vertices of G, to find a path between
v and w, called shortest path (SP), with minimum length. The length of that
path is called distance between v and w and is denoted as dist(v, w). For any
subgraph H of G, the distance between v and w in H is denoted as distH (v, w).
The single-source shortest path problem (SSSP ) is to find SPs from a fixed vertex
v to all other vertices of G. Finally, the all-pairs shortest path problem (APSP )
is to find SPs between all pairs of vertices. There are distance versions of SPSP,
SSSP, and APSP, which are more commonly studied, where the objective is to
compute the corresponding distances instead of SPs. Most distance algorithms
allow the corresponding SPs to be retrieved in additional time proportional to
the number of the edges of the path. In this paper, by SPSP, SSSP, and APSP
we mean the distance versions of these problems.
A k-partition P of G is a set V1, . . . , Vk of subsets of V (G), the set of the
i=1 Vi = V (G). We call the
subgraphs of G induced by Vi components of P. The boundary of the partition
consists consists of all vertices of G that have at least one neighbor in a different
component. We denote by BG(G) or simply by BG the subgraph of G induced by
the boundary vertices. For any C ∈ P, we denote by B(C) the set of all boundary
vertices that are from C. For any planar graph of n vertices and bounded (O(1)
as a function of n) vertex degree one can find in O(n) time a k-partition P with
vertices of G, such that Vi ∩ Vj = ∅ if i (cid:54)= j and (cid:83)k
B(C) = O((cid:112)n/k) for each component C ∈ P.
3 Algorithm overview and analysis
Our algorithm works in two modes: preprocessing mode, during which a data
structure is computed that allows efficient SP queries, and the query mode that
uses that data structure to compute the distance between a query pair of vertices.
We assume that the input is a planar graph G of n vertices and bounded vertex
degree and the cluster has p nodes.
3.1 Preprocessing mode
The preprocessing algorithm (Algorithm 1) has three phases. During the first
phase (line 1), the graph is partitioned and each component is assigned to a
distinct cluster node. During the second phase (lines 2-5), the APSP problem is
solved for each component C independently and in parallel and the computed
distance matrix APSP(C) is stored at the same node. Finally, in the third phase
(lines 6-10), the boundary graph BG is constructed and the APSP is solved for
BG. That computation is done distributedly such that the distances from vertex
v ∈ BG to all other vertices of BG are computed at the node containing v, by
using Dijkstra's algorithm [2]. The computed distance matrix is stored at the
node that has done the computations. Hence, at the end of the algorithm, the
node N (C) contains two matrices: one containing the SP distances in C and the
other containing all SP distances in BG with source a vertex in BG ∩ C.
One can think of BG as a compressed version of G where the non-boundary
vertices are removed, but are implicitly represented in BG by the information
encoded in its edge weights. Note however that the distances APSP(C) (and the
corresponding edge weights of BG) are not distances in G; the reason is that a
shortest path between two vertices v and w from C might pass through vertices
not in C. Hence the following fact is non-trivial.
Lemma 1. [4] For any two vertices v, w ∈ BG the distance between v and w in
BG is equal to the distance between v and w in G.
We will next estimate the time and space (memory) required to run the al-
gorithm. As G is planar and of bounded vertex degree (as a function of n), it
can be divided in O(n) time into k parts so that each part has no more than
(n/k) vertices and O((cid:112)n/k) boundary vertices [5]. We will estimate the require-
ments of each phase. Since the maximum amount of coarse-grained parallelism
of Algorithm 1 is min{p, k}, we assume without loss of generalization that p ≤ k.
Phase 1 requires O(n) running time and O(n) space [5].
The complexity of Phase 2 is dominated by the time for computing dis-
tances in line 3. We assume that we are using the algorithm from [4] that can
be implemented efficiently on a GPU-accelerated architecture and has complex-
ity O(N 9/4). Then Phase 2 requires O((k/p)(n/k)9/4) = O(n9/4/(pk5/4)) time
and kO((n/k)(cid:112)n/k) = O(n3/2/k1/2) total space. The space per processor is
For Phase 3, the number of the vertices of BG is k O((cid:112)n/k) = O(
the number of the edges is k O(((cid:112)n/k)2) = O(n). One execution of line 8 (for one
kO((n/k)2) = O(n2/k).
√
nk) and
component C) takes (k/p)BG∩CE(BG) log(BG) = (k/p)O((cid:112)n/k)O(n log n)
CBG = O((cid:112)n/k
time and O(n) space. The space needed for one iteration of Step 9 is BG ∩
nk) = O(n). Hence Phase 3 requires O((k/p)n3/2/k1/2 log n)
√
= O(n3/2k1/2 log n/p) time and O(nk/p) space per processor.
Summing up the requirements for Phases 1, 2, and 3, we get O(n9/4/(pk5/4)+
n3/2k1/2 log n/p)) time and O(n + n2/(pk) + nk/p) space per processor needed
for Algorithm 1. Assuming space is more important in this case than time (since
nodes have limited memory), we find that k = n1/2 minimizes the function
n2/k + nk. Hence we have the following result.
Lemma 2. With k = (cid:100)n1/2(cid:101) and p ≤ k, Algorithm 1 runs in O(n7/4 log n/p)
time and uses O(n3/2/p) space per processor. With p = k, the time and space
are O(n5/4) and O(n), respectively.
The time bound of Lemma 2 is conservative as it doesn't take into account
our use of fine-grain parallelism due to multi-threading, e.g., by the GPUs.
Algorithm 1 Preprocessing algorithm
Input: A planar graph G
Output: A data structure for efficient shortest path queries in G
1: Construct a k-partition P of G and assign each component C to a distinct node
/∗ Partitioning ∗/
N (C)
/∗ Solve the APSP problem for each component ∗/
2: for all components C ∈ P do in parallel
3:
4:
Solve APSP for C and save the distances in a table APSP(C)
For each pair of boundary vertices v, w ∈ C define edge (v, w), if not already in
G, and assign a weight wt(v, w) = distC (v, w)
5: end for
/∗ Solve the APSP problem for the boundary graph ∗/
defined in the previous step and store it at each node
6: Define a boundary graph BG with vertices all boundary vertices of G and edges as
7: for all components C ∈ P do in parallel
8:
9:
Solve SSSP in BG for each vertex of C ∩ BG
Store the distances from all vertices of C ∩ BG to all vertices of BG in a
table APSPBG (C)
10: end for
3.2 Query mode
The query algorithm (Algorithm 2) is based on the fact that if C1 (cid:54)= C2, then
any path between v1 and v2 should cross both B(C1) and B(C2). Let π be a
shortest path between v1 and v2. Then π can be divided into three parts: from
v1 to a vertex b1 from B(C1), from b1 to a vertex b2 on p from B(C2), and from
b2 to v2. Vertices b1 and b2 minimizing the length of p are found as follows: in
the loop on lines 2-7, for each b2 an optimal b1 and dist(v1, b2) are found; in lines
10-12 an optimal b2 is found.
Algorithm 2 Query algorithm
Input: Vertices v1, v2 of G, a k-partition P of G, tables APSP(C) and APSPBG (C)
for all C ∈ P
Output: dist(v1, v2)
1: Determine components C1 and C2 such that v1 ∈ C1, v2 ∈ C2
2: for all vertices b2 ∈ B(C2) do in parallel
/∗ Compute dist(v1, b2) ∗/
dist(v1, b2) = ∞
for all vertices b1 ∈ B(C1) do
dist(v1, b2) = min{dist(v1, b2), distC1 (v1, b1) + distBG (b1, b2)}
3:
4:
5:
6:
7: end for
8: If N (C1) (cid:54)= N (C2) then transfer the column of SP(C2) corresponding to v2 from
end for
/∗ Now we can compute dist(v1, v2) ∗/
N (C2) to N (C1).
9: dist(v1, v2) = ∞
10: for all vertices b2 ∈ B(C2) do
11:
12: end for
13: If C1 = C2 then dist(v1, v2) = min{dist(v1, v2), distC1 (v1, v2)}, where the distance
dist(v1, v2) = min{dist(v1, v2), dist(v1, b2) + distC2 (b2, v2)}
distC1 (v1, v2) is taken from APSP(C1).
Lemma 3. Algorithm 2 correctly computes dist(v1, v2) and its running time is
O(n1/4) with k = (cid:100)n1/2(cid:101) and p ≥ (cid:100)n1/4(cid:101).
Proof. Let π be a shortest path between v1 and v2, let C1 (cid:54)= C2, and let b1 be
the first vertex along π that is on B(C1) and π1 be the subpath of π from v1 to
b1, let π2 be the last vertex along π that is on B(C2) and π2 be the subpath of π
from b1 to b2, and let π3 be the subpath of π from b2 to v2. Then π1 is entirely in
C1 and hence distC1(v1, b1) = distG(v1, b1) (note, however, that the distances in
APSP(C1) from v1 to other vertices from B(C1) may not be correct). Similarly,
distC2(b2, v2) = distG(b2, v2). Finally, distBG (b1, b2) = distG(b1, b2) by Lemma 1.
Hence lines 5 and 11 use correct values for computing the distances between v1
and b2 and between b2 and v2.
The loop on lines 5-10 takes time B(C1)B(C2)/p = O((cid:112)n/k(cid:112)n/k/p) =
If C1 = C2 (line 13), then a shortest path between v1 and v2 may or may not
leave C1. In the first case lines 1-12 compute correctly dist(v1, v2), in the second
case APSP(C1) contains the correct distance.
O(n/(pk)), for p ≤ min{k, (n/k)1/2}. If k = n1/2 and p = n1/4 (the maximum
value for which the formula applies), that time becomes O(n1/4). The loop in
lines 10-12 takes time O((n/k)1/2) = O(n1/4) for k = n1/2.
Note that using the methodology of [3], a more complex implementation of
Algorithm 2 can reduce the query time to logarithmic. Note also that compu-
tation in lines 2-7 can be overlapped with transferring of data in line 8 thereby
saving time (upto a factor of two).
4
Implementation details
In this section, we describe how the preprocessing and query modes are imple-
mented on a hybrid CPU-GPU cluster. We use a distance matrix to represented
both the input graph G and the output. Such a 2-dimensional matrix contains
in cell (i, j) the value of the distance from vertex i to vertex j. Initially, cell (i, j)
contains wt(i, j) if an edge (i, j) is present in G, or infinity otherwise. These
values are updated as the algorithm progresses. At the end of the algorithm, cell
(i, j) contains dist(i, j).
In phase 1 of the preprocessing mode, we construct a k-partition of G using
the METIS library[7]. Based on that partition, we reorder the vertices of G so
that vertices from the same component have consecutive indices and boundary
vertices of each components have the lowest indices -- see Figure 1 .
Fig. 1: Distance matrix after reorder-
ing of the vertices. Vertices from the
same component are stored contigu-
ously starting with boundary vertices.
Red submatrices are also part of the
boundary distance matrix. Grey sub-
matrices do not generate any compu-
tations in preprocessing mode.
Fig. 2: The distances required to com-
pute dist(v, w), shown in green, are
scattered in three submatrices: two di-
agonal ones, for component I and for
component J, and a non-diagonal sub-
matrix (I, J).
In phase 2, we compute the shortest distances within each of the components.
For k components, this phase gives a total k independent tasks that can be
executed in parallel. Computations at this phase are already balanced across
C1C3C2C1C3C1C3B(C1)B(C1)other C1B(C1)other C1other C1VerticesB(C1)B(C3)other C1B(C3)B(C1)other C3other C1otherC3vw......Submatrix (I, I)Submatrix (I, J)Submatrix (J, J)B(CI)B(CJ)vB(CI)wB(CJ) nodes as components contain roughly the same number of vertices and the APSP
algorithm from [4] ensures the same O(N 9/4) complexity with respect to the
number of nodes.
Finally, phase 3 consists in computing the shortest distances within the
boundary graph using Dijkstra's algorithm. Computations at this phase may
be imbalanced between nodes for two reasons. First, the number of boundary
vertices in two components may differ and, second, the complexity of Dijkstra's
algorithm does not solely depend on the number of vertices in the graph, but
also on the number of edges, which may vary even more than the number of
vertices between two components' boundary graphs.
In the query mode, we are interested in finding dist(v, w), where v and w are
from components I and J, respectively. The required values for that computation
are scattered in three submatrices, as illustarted in Figure 2. For such a query,
assuming k = p, node i, holding the required values from diagonal submatrix
I and non-diagonal submatrix (I, J), will be in charge of the computations.
Required values from diagonal submatrix J are held by node j and need to be
transfered to node i.
5 Experimental evaluation
In this section we describe experiments designed to test our algorithm and its
implementation. Specifically, we are going to test the strong scaling properties by
running our code on a fixed graph size and a varying number p of cluster nodes
and number k of components. All computations are run on a 300 node cluster.
Each cluster node is comprised of 2 x Eight-Core Intel Xeon model E5-2670 @
2.6 GHz and two GPGPU Nvidia Tesla M2090 cards connected to PCIe-2.0 x16
slots. In order to make full use of the available GPUs, each node is assigned at
least two graph components so that the two associated diagonal submatrices can
be computed simultaneously on the two GPUs.
For the strong-scaling experiment, the graph size is fixed to 256k vertices.
Preprocessing and queries are run with increasing numbers of nodes ranging from
4 to 64. Each node handles 2 components (one per available GPU); therefore the
number of components k ranges from 8 to 128.
Figure 3 shows the run times for the preprocessing mode. For low numbers
of nodes and thus low values of k, preprocessing time is dominated by step 2 -
the computation of the shortest distances within each component - since lower k
values means larger components. For higher numbers of nodes and thus higher
values of k, preprocessing time becomes dominated by step 3 - the computation
of the boundary graph - as more components mean higher numbers of incident
edges and thus larger boundary graphs. Note that while the figure seems to show
supralinear speedup, that is not the case (and similarly for the memory usage).
The reason is that, with increasing the number of processors p, the number k of
parts is increased too (as it is tied to p in this implementation) and hence the
complexity of the algorithm is also reduced.
Fig. 3: Preprocessing run times for a
fixed graph size of 256k vertices and in-
creasing number of nodes.
Fig. 4: Peak memories and run times for
10k queries for a fixed graph size of
256k vertices and increasing number of
parts/processors.
Figure 4 shows the query times and peak memory usage per node. The run
times are given for 10, 000 queries from random sources to random targets. Note
that in the query mode only fine-grain (node-level) parallelism is used, while
multiple nodes are still needed for distributed storage and, optionally, to handle
multiple queries in parallel (not implemented in the current version). For the
n -- or 512
memory usage, the optimal value for k, theoretically expected to be
for this instance -- is not reached in this experiment since k only goes up to 128.
We can however see that peak memory usage per node is still dropping with
increasing values of k up to 128. The query times in the figure vary from about 2
milliseconds per query for k = 8 to 0.25 milliseconds for k = 128. Compared with
the Boost library implementation of Dijkstra's algorithm, our implementation
answers queries on the largest instances about 1000 times faster.
√
6 Conclusion
√
We developed and implemented a distributed algorithm for shortest path queries
in planar graphs with good scalability. It allows answering SP queries in O(n1/4)
n) processors with O(n) space per processor and O(n5/4)
time by using O(
preprocessing time. Our implementation on 300 node CPU-GPU cluster has
preprocessing time of less than 10 seconds using 32 or more nodes and 0.025
milliseconds per query using two nodes. Interesting tasks for future research is
implementing a version allowing parallel queries and reducing the query time of
the implementation to O(log n) by using properties of graph planarity.
References
1. Chen, D.Z., Xu, J.: Shortest path queries in planar graphs. pp. 469 -- 478. STOC '00,
ACM, New York, NY, USA (2000)
2. Dijkstra, E.W.: A note on two problems in connexion with graphs. Numerische
mathematik 1(1), 269 -- 271 (1959)
481632640200400600800100012001400run time step 3 (s)run time step 2 (s)run time init +step 1 (s)# of nodesrun time (s)48163264051015202505101520253035peak memory (GB)run time queries (s)# of nodesRun time (s)Peak memory (GB) 3. Djidjev, H.: Efficient algorithms for shortest path queries in planar digraphs. In:
Graph-Theoretic Concepts in Computer Science, Lecture Notes in Computer Sci-
ence, vol. 1197, pp. 151 -- 165 (1997)
4. Djidjev, H., Thulasidasan, S., Chapuis, G., Andonov, R., Lavenier, D.: Efficient
multi-GPU computation of all-pairs shortest paths. In: IPDPS'2014. pp. 360 -- 369
(2014)
5. Frederickson, G.N.: Planar graph decomposition and all pairs shortest paths. J.
ACM 38(1), 162 -- 204 (1991)
6. Hutchinson, D., Maheshwari, A., Zeh, N.: An external memory data structure for
shortest path queries (extended abstract). pp. 51 -- 60. COCOON'99, Springer-Verlag,
Berlin, Heidelberg (1999)
7. Karypis, G., Kumar, V.: Multilevel k-way partitioning scheme for irregular graphs.
Journal of Parallel and Distributed computing 48(1), 96 -- 129 (1998)
8. Kowalik, L.: Shortest path queries in planar graphs in constant time. In: STOC03.
pp. 143 -- 148. ACM (2003)
9. Mozes, S., Sommer, C.: Exact distance oracles for planar graphs. pp. 209 -- 222. SODA
'12, SIAM (2012)
|
1607.08337 | 2 | 1607 | 2017-02-06T13:00:15 | Efficient Algorithms for Constructing Very Sparse Spanners and Emulators | [
"cs.DS"
] | Miller et al. \cite{MPVX15} devised a distributed\footnote{They actually showed a PRAM algorithm. The distributed algorithm with these properties is implicit in \cite{MPVX15}.} algorithm in the CONGEST model, that given a parameter $k = 1,2,\ldots$, constructs an $O(k)$-spanner of an input unweighted $n$-vertex graph with $O(n^{1+1/k})$ expected edges in $O(k)$ rounds of communication. In this paper we improve the result of \cite{MPVX15}, by showing a $k$-round distributed algorithm in the same model, that constructs a $(2k-1)$-spanner with $O(n^{1+1/k}/\epsilon)$ edges, with probability $1- \epsilon$, for any $\epsilon>0$. Moreover, when $k = \omega(\log n)$, our algorithm produces (still in $k$ rounds) {\em ultra-sparse} spanners, i.e., spanners of size $n(1+ o(1))$, with probability $1- o(1)$. To our knowledge, this is the first distributed algorithm in the CONGEST or in the PRAM models that constructs spanners or skeletons (i.e., connected spanning subgraphs) that sparse. Our algorithm can also be implemented in linear time in the standard centralized model, and for large $k$, it provides spanners that are sparser than any other spanner given by a known (near-)linear time algorithm.
We also devise improved bounds (and algorithms realizing these bounds) for $(1+\epsilon,\beta)$-spanners and emulators. In particular, we show that for any unweighted $n$-vertex graph and any $\epsilon > 0$, there exists a $(1+ \epsilon, ({{\log\log n} \over \epsilon})^{\log\log n})$-emulator with $O(n)$ edges. All previous constructions of $(1+\epsilon,\beta)$-spanners and emulators employ a superlinear number of edges, for all choices of parameters.
Finally, we provide some applications of our results to approximate shortest paths' computation in unweighted graphs. | cs.DS | cs |
Efficient Algorithms for Constructing Very Sparse Spanners and
Emulators∗
Michael Elkin†1 and Ofer Neiman‡1
1Department of Computer Science, Ben-Gurion University of the Negev, Beer-Sheva, Israel.
Email: {elkinm,neimano}@cs.bgu.ac.il
Abstract
Miller et al.
[MPVX15] devised a distributed1 algorithm in the CONGEST model, that given
a parameter k = 1, 2, . . ., constructs an O(k)-spanner of an input unweighted n-vertex graph with
O(n1+1/k) expected edges in O(k) rounds of communication. In this paper we improve the result of
[MPVX15], by showing a k-round distributed algorithm in the same model, that constructs a (2k − 1)-
spanner with O(n1+1/k/ǫ) edges, with probability 1 − ǫ, for any ǫ > 0. Moreover, when k = ω(log n),
our algorithm produces (still in k rounds) ultra-sparse spanners, i.e., spanners of size n(1 + o(1)), with
probability 1 − o(1). To our knowledge, this is the first distributed algorithm in the CONGEST or in
the PRAM models that constructs spanners or skeletons (i.e., connected spanning subgraphs) that sparse.
Our algorithm can also be implemented in linear time in the standard centralized model, and for large k, it
provides spanners that are sparser than any other spanner given by a known (near-)linear time algorithm.
We also devise improved bounds (and algorithms realizing these bounds) for (1 + ǫ, β)-spanners and
emulators. In particular, we show that for any unweighted n-vertex graph and any ǫ > 0, there exists a
(1 + ǫ, ( log log n
)log log n)-emulator with O(n) edges. All previous constructions of (1 + ǫ, β)-spanners
and emulators employ a superlinear number of edges, for all choices of parameters.
ǫ
Finally, we provide some applications of our results to approximate shortest paths' computation in
unweighted graphs.
1 Introduction
1.1 Setting, Definitions
We consider unweighted undirected n-vertex graphs G = (V, E). For a parameter α ≥ 1, a subgraph
H = (V, E′), E′ ⊆ E, is called an α-spanner of G, if for every pair u, v ∈ V of vertices, we have
dH (u, v) ≤ α · dG(u, v). Here dG(u, v) (respectively, dH (u, v)) stands for the distance between u and v
in G (resp., in H). The parameter α is called the stretch of the spanner H. More generally, if for a pair of
parameters α ≥ 1, β ≥ 0, for every pair u, v ∈ V of vertices, it holds that dH(u, v) ≤ α · dG(u, v) + β,
then the subgraph H is said to be an (α, β)-spanner of G. Particularly important is the case α = 1 + ǫ, for
∗A preliminary version [EN17] of this paper appeared in SODA'17.
†This research was supported by the ISF grant No. (724/15).
‡Supported in part by ISF grant No. (523/12) and by BSF grant No. 2015813.
1They actually showed a PRAM algorithm. The distributed algorithm with these properties is implicit in [MPVX15].
1
some small ǫ > 0. Such spanners are called near-additive. If H = (V, E′′, ω), where ω : E′′ → R+, is
not a subgraph of G, but nevertheless satisfies that for every pair u, v ∈ V of original vertices, dG(u, v) ≤
dH (u, v) ≤ (1 + ǫ)dG(u, v) + β, then H is called a near-additive β-emulator of G, or a (1 + ǫ, β)-emulator
of G.
Graph spanners have been introduced in [Awe85, PS89, PU89a], and have been intensively studied ever
since [ADD+93, ABCP93, Coh99, ACIM99, DHZ00, BS03, Elk04, Elk07a, EZ06, TZ06, Woo06, Elk07b,
Pet09, DGPV08, Pet10, BW15, MPVX15, AB16]. They were found useful for computing approximately
shortest paths [ABCP93, Coh99, Elk04, EZ06], routing [PU89b], distance oracles and labeling schemes
[Pel99, TZ05, EP15], synchronization [Awe85], and in other applications.
The simplest and most basic algorithm for computing a multiplicative α-spanner, for a parameter α ≥ 1,
is the greedy algorithm [ADD+93]. The algorithm starts with an empty spanner, and examines the edges of
the input graph G = (V, E) one after another. It tests if there is a path in H of length at most α between
the endpoints u and v of e. If it is not the case, the edge is inserted into the spanner. Otherwise the edge is
dropped.
It is obvious that the algorithm produces an α-spanner H. Moreover, the spanner H has no cycles of
length α + 1 or less, i.e., the girth of H, denoted g(H), satisfies g(H) ≥ α + 2. Denote m = m(n, g) the
maximum number of edges that a girth-g n-vertex graph may contain. It follows that H ≤ m(n, α + 2).
The function m(n, g) is known to be at most n1+ 2
g−2 , when g ≤ 2 log2 n, and for larger g (i.e., for m ≤ 2n),
it is given by m(n, g) ≤ n(1 + (1 + o(1)) ln(p+1)
), where p = m − n, [AHL02, BR10]. These bounds are
called "Moore's bounds for irregular graphs", or shortly, (generalized) Moore's bounds.
Any construction of multiplicative α-spanners for n-vertex graphs with at most m′(n, α + 2) edges
implies an upper bound m(n, α + 2) ≤ m′(n, α + 2) for the function m(n, g). (As running the construction
on the extremal girth-(α + 2) n-vertex graph can eliminate no edge.) Hence the greedy algorithm produces
(See also [FS16].)
multiplicative spanners with optimal tradeoff between stretch and number of edges.
However, the greedy algorithm is problematic from algorithmic perspective. In the centralized model of
computation, the best-known implementation of it [RZ04] requires O(α · n2+1/α) time. Moreover, the
greedy algorithm is inherently sequential, and as such, it is generally hard2 to implement it in distributed
and parallel models of computation.
g
In the distributed model [Pel00] we have processors residing in vertices of the graph. The processors
communicate with their graph neighbors in synchronous rounds. In each round, messages of bounded length
can be sent. (This is the assumption of the CONGEST model. In the LOCAL model, messages' size is
arbitrary.) The running time of an algorithm is this model is the number of rounds that it runs. By "parallel"
model we mean here PRAM EREW model [Rei93], and we are interested in algorithms with small running
time (aka depth) and work complexities. (The latter measures the overall number of operations performed
by all processors.)
Dubhashi et al. [DMP+03] devised a distributed implementation of the greedy algorithm in the LOCAL
model of distributed computation. Their algorithm runs in O(α · log2 n) rounds, i.e., suboptimal by a factor
of log2 n. Moreover, it collects graph topology to depth O(α), and conducts heavy local computations.
To our knowledge, there is no distributed-CONGEST or PRAM implementation of the greedy algorithm
known. There is also no known efficient3 centralized, distributed-CONGEST, or PRAM algorithm that
constructs ultra-sparse spanners, i.e., spanners with n + o(n) edges.
2In the sequel we discuss a distributed setting, specifically, the LOCAL model, in which a relatively efficient implementation of
the greedy is known.
3By "efficient" centralized algorithm in this paper we mean an algorithm with running time close to O(E). By efficient
distributed or parallel algorithm we mean an algorithm that runs in polylogarithmic, or nearly-polylogarithmic, time.
2
In the distributed and parallel settings it is often enough to compute a sparse skeleton H of the input
graph G, where a skeleton is a connected subgraph that spans all the vertices of G, i.e., the stretch require-
ment is dropped. Dubhashi et al. [DMP+03] devised a distributed-LOCAL algorithm that computes ultra-
sparse skeletons of size m(n, α) ≤ n + O(n · log n
α ) in O(α) rounds. Like their algorithm for constructing
spanners, this algorithm also collects topologies to depth O(α), and involves heavy local computations. To
our knowledge, no efficient distributed-CONGEST or PRAM algorithm for computing ultra-sparse skele-
tons is currently known. In this paper we devise the first such algorithms.
1.2 Prior Work and Our Results
In the centralized model of computation the best-known efficient algorithm for constructing multiplicative
spanners (for unweighted graphs) is due to Halperin and Zwick [HZ96]. Their deterministic algorithm, for
an integer parameter k ≥ 1, computes a (2k − 1)-spanner with n1+1/k + n edges in O(E) time. (Their
result improved previous pioneering work by [PS89, Coh99].) Note that their bound on the number of edges
is always at least 2n, i.e., in the range k = Ω(log n) it is very far from Moore's bound.
Our centralized randomized algorithm computes (with probability close to 1), a (2k − 1)-spanner with
n· (1 + O( log n
k )) edges in O(E) time, whenever k = Ω(log n). Note that when k = ω(log n), the number
of edges is n(1 + o(1)), i.e., in this range the algorithm computes an ultra-sparse spanner in O(E) time.
Moreover, whenever k ≤ n1−ǫ, for any constant ǫ > 0, up to a constant factor in the lower-order term, our
bound matches Moore's bound. In fact, our algorithm and its analysis can be viewed as an alternative proof
of (a slightly weaker version of) the generalized Moore's bound. Note that it is not the case for the greedy
algorithm and its implementations [ADD+93, RZ04, DMP+03]: the analysis of these algorithms relies on
Moore's bounds, but these algorithms cannot be used to derive them.
Another variant of our algorithm, which works for any k ≥ 2, computes with high probability a (2k−1)-
spanner with n1+1/k(1 + O( log k
ln ln n the
number of edges in our spanner is n1+1/k + o(n), improving the result of [HZ96] (albeit with a somewhat
worse running time for 2 ln n
k )) edges, in O(kE) time.4 In particular, for the range k ≥ 2 ln n
ln ln n ≤ k ≤ log n). Note that for any k ≥ 2 we have O(n1+1/k) edges.
Yet another variant of our algorithm computes a (2k − 1)-spanner with O(n1+1/k) edges, in expected
In the distributed-CONGEST and PRAM models, efficient algorithms for computing linear-size span-
O(E) time.
ners were given in [Pet09, MPVX15]. Specifically, [MPVX15] devised an O(k)-round distributed-CONGEST
randomized algorithm for constructing O(k)-spanner (with high probability) with expected O(n1+1/k)
In the PRAM model their algorithm has depth O(k log∗ n) and work O(E). There are also
edges.
k-round distributed-CONGEST randomized algorithms for constructing (2k − 1)-spanner with expected
O(k · n1+1/k) edges [BS07, Elk07a]. It is known that at least k rounds are required for this task, under
Erdos' girth conjecture [Elk07a, DGPV08].
Our randomized algorithm uses k rounds in the distributed-CONGEST model, and with probability at
least 1− ǫ it constructs a (2k − 1)-spanner with O(n1+1/k/ǫ) edges (for any desired, possibly sub-constant,
ǫ > 0). In the PRAM model the depth and work complexities of our algorithm are the same as in [MPVX15].
Furthermore, when k ≥ log n we can bound the number of edges by n · (1 + O( log n
ǫ·k )), again matching
Moore's bound up to a constant factor in the lower-order term.
This result improves the previous state-of-the-art in the entire range of parameters. In particular, it is
also the first efficient algorithm in the distributed-CONGEST or PRAM models that constructs an ultra-
sparse skeleton. Specifically, in O(log n) time it computes an O(log n)-spanner with n(1 + o(1)) edges,
4As usual, O(f (n)) stands for O(f (n)polylogf (n)).
3
with probability 1 − o(1).
We also use our algorithm for unweighted graphs to devise an improved algorithm for weighted graphs
as well. Specifically, our algorithm computes (2k − 1)(1 + ǫ)-spanner with O(n1+1/k · (log k)/ǫ) edges,
within expected O(E) time. See Theorem 2, and the discussion that follows it, for further details.
1.3 Near-Additive Spanners and Emulators
It was shown in [EP04] that for any ǫ > 0 and κ = 1, 2, . . ., and any (unweighted) n-vertex graph G =
(V, E), there exists a (1 + ǫ, β)-spanner with O(β · n1+1/κ) edges, where β = β(κ, ǫ) ≤ O( log κ
ǫ )log κ. Ad-
ditional algorithms for constructing such spanners were later given in [Elk04, EZ06, TZ06, DGPV08, Pet09,
Pet10]. Abboud and Bodwin [AB16] showed that multiplicative error of 1 + ǫ in [EP04]'s theorem cannot
be eliminated, while still keeping a constant (i.e., independent of n) additive error β, and more recently
.
In the regime of constant κ, the bound of [EP04] remains the state-of-the-art. Pettie [Pet09] showed that
one can construct a (1 + ǫ, β)-spanner with O(n log log(ǫ−1 log log n)) edges and β = O( log log n
)log log n.
This result of [Pet09] is not efficient in the sense considered in this paper, i.e., no distributed or parallel
implementations of it are known, and also no efficient (that is, roughly O(E)-time) centralized algorithm
computing it is known. Also, this result does not extend ([Pet16]) to a general tradeoff between β and the
number of edges.
[ABP17] showed that any such (1 + ǫ, β)-spanner of size O(n1+1/κ−δ), δ > 0, has β = Ω(cid:16) 1
ǫ·log κ(cid:17)log κ−1
ǫ
Improving upon previous results by [Elk05, EZ06], Pettie [Pet10] also devised an efficient distributed-
CONGEST algorithm, that for a parameter ρ > 0, constructs in O(nρ) rounds a (1 + ǫ, β)-spanner with
being the golden ratio. 5
Independently and simultaneously to our work, [ABP17] showed that there exist (1 + ǫ, β)-spanners with
< 3/4. This
O(n1+1/κ(ǫ−1 log κ)φ) edges and β = O(cid:16) log κ+1/ρ
O((ǫ−1 log κ)h · log κ · n1+1/κ) edges and β = O(cid:16) log κ
spanner has improved dependence on ǫ in the number of edges (at the cost of worse dependence on κ).
(cid:17)logφ κ+1/ρ
ǫ (cid:17)log κ−2
, where h = (3/4)κ−1−log κ
, for φ = 1+√5
In this paper we improve all of the tradeoffs [EP04, Pet10] in the entire range of parameters. Specifically,
log(1/ǫ)+log log log n , our distributed-CONGEST algorithm constructs
for any ǫ > 0, ρ > 0 and κ = 1, 2, . . . ,
in O(nρ) rounds a (1 + ǫ, β)-spanner with O(n1+1/κ) edges and
log n
κ
2
ǫ
β ≤ O(cid:18) log κ + 1/ρ
ǫ
(cid:19)log κ+1/ρ
.
Our algorithm also admits efficient implementations in the streaming and standard models of computation,
see Section 3. Our spanners are sparser and have polynomially smaller β than the previous best efficient con-
structions. They are even sparser than the state-of-the-art existential ones (with essentially the same β), with
the following exceptions: whenever ǫ < 1/ log3 log n our result and that of [ABP17] are incomparable,6 and
the spanner from [Pet09] that has O(n log(4) n) edges, while ours never gets sparser than O(n log log n)/ǫ.
In the complementary range, ǫ > 1/ log3 log n, our result is strictly stronger than that of [ABP17].
Moreover, a variant of our algorithm efficiently constructs very sparse (1+ǫ, β)-emulators. In particular,
we can obtain a linear-size (1 + ǫ, ( log log n
)log log n)-emulator. (We stress that the number of edges does not
depend even on ǫ.) All previous constructions of (1 + ǫ, β)-spanners or emulators employ a superlinear
number of edges, for all choices of parameters.
ǫ
5 In the range of κ = o( log n
log log n ), the result of [Pet10] is incomparable with [EP04], as spanners of [EP04] provide smaller β,
while spanners of [Pet10] are slightly sparser.
6The i-iterated logarithm is defined by log(i+1) n = log(log(i) n), for all i ≥ 0, and log(0) n = n.
4
We use our new algorithms for constructing near-additive spanners and emulators to improve approxi-
mate shortest paths' algorithms, in the centralized and streaming models of computation. One notable result
in this context is a streaming algorithm that for any constant ǫ > 0 and any subset S ⊆ V with S = nΩ(1),
computes (1+ǫ)-approximate shortest paths for S×V within O(S) passes over the stream, using O(n1+ǫ)
space. See Section 4 for more details, and additional applications of our spanners.
1.4 Technical Overview
Linial and Saks [LS93] were the first to employ exponential random variables to build network decomposi-
tions, i.e., partitions of graphs into clusters of small diameter, which possess some useful properties. This
technique was found useful for constructing padded partitions, hierarchically-separated trees, low-stretch
spanning trees [Bar96, Bar98, Bar04, EEST05, ABN11, AN12] and spanners [Coh99, BS03, Elk07b]. In
[LS93] every vertex v tosses a random variable rv from an exponential distribution, and broadcasts to all
vertices within distance rv from v. Every vertex v joins the cluster of a vertex u with largest identity number,
whose broadcast u heard.
Blelloch et al. [BGK+14] introduced a variant of this technique in which, roughly speaking, every vertex
v starts to broadcast at time −rv, and broadcasts indefinitely. A vertex x joins the cluster centered at a vertex
v, whose broadcast reaches x first. They called the resulting partition "exponential start time clustering",
and it was demonstrated in [BGK+14, MPX13, EN16a] that this approach leads to very efficient distributed
and parallel algorithms for constructing padded partitions and network decompositions.
Miller et al. [MPVX15] used this approach to devise an efficient parallel and distributed-CONGEST
O(k)-time algorithm for constructing O(k)-spanner with O(n1+1/k) edges. Specifically, they build the
exponential time clustering, add the spanning trees of the clusters into the spanner, and then every vertex x
adds into the spanner one edge (x, y) connecting x to every adjacent cluster Cy, y ∈ Cy.
The main property of the partition exploited by [MPVX15] in the analysis of their algorithm is that any
unit-radius ball in the input graph G intersects just O(n2/k) clusters, in expectation. Note also that their
algorithm is doomed to use at least n1+1/k + (n − 1) edges, because it starts with inserting the spanning
trees of all clusters (amounting to up to n− 1 edges), and then inserts the O(n1+2/k) edges crossing between
different clusters into the spanner. To get O(n1+1/k) edges, one rescales k′ = 2k.
In our algorithm we do not explicitly construct the exponential start time clustering. Rather we run the
procedure that builds it, but every vertex x connects not just to the neighbor y through which x received its
first broadcast message at time, say, ty, but also to all neighbors z whose broadcast x received witin time
interval [ty, ty + 1]. We show that, in expectation, x connects to n1/k neighbors altogether, and not just to
that many adjacent clusters. As a result we obtain both a sparser spanner, a smaller stretch, and a smaller
running time. The stretch and running time are smaller roughly by a factor of 2 than in [MPVX15], because
we do not need to consider unit balls, that have diameter 2. Rather we tackle individual edges (of length 1).
In the context of weighted graphs, [MPVX15] showed how their efficient algorithm for constructing
sparse (4k − 2)-spanners for unweighted graphs can be converted into an efficient algorithm that constructs
(16k − 8)-spanners with O(n1+1/k log k) edges for weighted graphs. By using their scheme naively on top
of our algorithm for unweighted graphs, one gets an efficient algorithm for computing (4k − 2)(1 + ǫ)-
spanners with O(n1+1/k · (log k)/ǫ) edges. Roughly speaking, the overhead of 2 in the stretch is because
in the analysis of [MPVX15], every vertex contributes expected O(n1/k) edges to the spanner on each of
roughly O(n1/k) phases of the algorithm in which it participates. By employing a more delicate probabilistic
argument, we argue that in fact, the expected total contribution of every vertex in all phases altogether is
O(n1/k), rather than O(n2/k). This enables us to eliminate another factor of 2 from the stretch bound. See
Section 2.3 for details.
5
Our constructions of (1 + ǫ, β)-spanners and emulators follow the [EP04] superclustering and intercon-
nection approach. One starts with a base partition P0. In [EP04] this was the partition of [Awe85, PS89,
AP92], obtained via region-growing technique. Then every cluster C ∈ P0 that has "many" unclustered
clusters of P0 "nearby", creates a supercluster around it. The "many" and the "nearby" are determined
by degree threshold deg 0 and distance threshold δ0, respectively. Once the superclustering phase is over,
the remaining unclustered clusters enter an interconnection phase, i.e., every pair of participating nearby
clusters gets interconnected by a shortest path in the spanner. This completes one iteration of the process.
The resulted superclustering P1 is the input for the next iteration of this process, which runs with different,
carefully chosen thresholds deg 1 and δ1. Such iterations continue until only very few clusters survive. The
latter are interconnected without further superclustering.
One bottleneck in devising efficient distributed algorithm based on this approach is the base parti-
tion. Known algorithms for constructing a region-growing partition of [Awe85] require almost linear dis-
tributed time [DMZ06]. We demonstrate that one can bypass it completely, and start from the base partition
P0 = {{v} v ∈ V }. This requires some modfication of the algorithm, and a more nuanced analysis. In
addition, we show that the superclustering and interconnection steps themselves can be implemented effi-
ciently. This part of the algorithm is based on our recent work on hopsets [EN16b], where we showed that
[EP04] approach is extremely useful in that context as well, and that it can be made efficient.
1.5 Related Work
Efficient algorithms for constructing (1 + ǫ, β)-spanners were also devised in [Elk05, EZ06, TZ06, Pet09,
Pet10]. These algorithms are based, however, on different approaches than that of the current paper. The
latter is based on [EP04]. Specifically, the approach of [Elk05, EZ06] is based on [Coh99, Coh00] construc-
tion of pairwise covers and hopsets, i.e., the algorithm works top-down. It recurses in small clusters, and
eliminates large ones. The approach of [TZ06, Pet09, Pet10] is based on [TZ05] collection of trees, used
originally for distance oracles.
Streaming algorithms for constructing multiplicative spanners were given in [FKM+05, Elk07b, Bas08],
and near-additive spanners in [Elk05, EZ06]. Spanners and emulators with sublinear error were given
in [TZ06, Pet09]. Spanners with purely additive error and lower bounds concerning them were given in
[ACIM99, EP04, BCE05, BKMP10, Che13, Woo06, BW15, AB16].
1.6 Organization
In Section 2 we present our algorithm for constructing multiplicative spanners and its analysis. In Section
2.3 we use this algorithm to provide improved spanners for weighted graphs as well. Our near-additive
spanners and emulators are presented in Section 3.
2 Sparse Multiplicative Spanners and Skeletons
Let G = (V, E) be a graph on n vertices, and let k ≥ 1 be an integer. Let c > 3 be a parameter governing
the success probability, and set β = ln(cn)/k. Recall the exponential distribution with parameter β, denoted
EXP(β), which has density
f (x) =(cid:26) β · e−βx
0
x ≥ 0
otherwise.
6
Construction. Each vertex u ∈ V samples a value ru from EXP(β), and broadcasts it to all vertices
within distance k. Each vertex x that received a message originated at u, stores mu(x) = ru − dG(x, u),
and also a neighbor pu(x) that lies on a shortest path from x to u (this neighbor sent x the message from u,
breaking ties arbitrarily if there is more than one). Let m(x) = maxu∈V {mu(x)}, then for every x ∈ V we
add to the spanner H the set of edges
The following lemma is implicit in [MPVX15]. We provide a proof for completeness.
C(x) = {(x, pu(x)) : mu(x) ≥ m(x) − 1} .
Lemma 1 ([MPVX15]). Let d1 ≤ . . . ≤ dn be arbitrary values and let δ1, . . . , δn be independent random
variables sampled from EXP(β). Define the random variables M = maxi{δi− di} and I = {i : δi− di ≥
M − 1}. Then for any 1 ≤ t ≤ n,
Pr[I ≥ t] = (1 − e−β)t−1 .
Proof. Denote by X (t) the random variable which is the t-th largest among {δi − di}. Then for any value
a ∈ R, if we condition on X (t) = a, then the event I ≥ t is exactly the event that all the remaining t − 1
values X (1), . . . , X (t−1) are at least a and at most a + 1. Using the memoryless property of the exponential
distribution and the independence of the {δi}, we have that
Pr[I ≥ t X (t) = a] = (1 − e−β)t−1 .
Since this bound does not depend on the value of a, applying the law of total probability we conclude that
Pr[I ≥ t] = (1 − e−β)t−1 .
Using this lemma, we can bound the expected size of the spanner.
Lemma 2. The expected size of H is at most (cn)1/k · n.
Proof. Fix any x ∈ V , and we analyze E[C(x)]. Note that the event C(x) ≥ t happens when there are at
least t shifted random variables ru − dG(u, x) that are within 1 of the maximum. By Lemma 1 this happens
with probability at most (1 − e−β)t−1 (we remark that if x did not hear at least t messages, then trivially
Pr[C(x) ≥ t] = 0). We conclude that
Xt=1
(1 − e−β)t = eβ = (cn)1/k ,
Pr[C(x) ≥ t] ≤
∞
Xt=0
n
E[C(x)] =
and the lemma follows by linearity of expectation.
We now argue about the stretch of the spanner.
Claim 3. With probability at least 1 − 1/c, it holds that ru < k for all u ∈ V .
Proof. For any u ∈ V , Pr[ru ≥ k] = e−βk = 1/(cn). By the union bound, Pr[∃u, ru ≥ k] ≤ 1/c.
Assume for now that the event of Claim 3 holds, i.e., that ru < k for all u ∈ V .
Corollary 4. For any x ∈ V , if u ∈ V is the vertex maximizing mu(x), then dG(u, x) < k.
7
Proof. First note that m(x) ≥ mx(x) ≥ 0, and using Claim 3 we have ru < k. So 0 ≤ m(x) = mu(x) =
ru − dG(u, x) < k − dG(u, x).
Claim 5. For any u, x ∈ V , if x adds an edge to pu(x), then there is a shortest path P between u and x
that is fully contained in the spanner H.
Proof. We prove by induction on dG(u, x). In the base case dG(x, u) = 1, then pu(x) = u, so (x, u) is
in the spanner. Assume that every vertex y ∈ V with dG(u, y) = t − 1 which added an edge to pu(y) has
a shortest path to u in H, and we prove for x that has dG(u, x) = t. We know that x added an edge to
y = pu(x), which lies on a shortest path to u, and thus satisfies dG(u, y) = t − 1. It remains to show that
this y added an edge to pu(y). First we claim that
m(y) ≤ m(x) + 1 .
(1)
Seeking contradiction, assume that (1) does not hold, and let v ∈ V be the vertex maximizing mv(y). By
Corollary 4 we have dG(v, y) < k, and thus dG(v, x) ≤ k. Hence x will hear the message of v. This means
that mv(x) ≥ mv(y) − 1 = m(y) − 1 > m(x), which is a contradiction. This establishes (1). Now, since x
added an edge to y = pu(x), by construction
mu(x) ≥ m(x) − 1 .
(2)
We conclude that
(2)
(1)
mu(y) = mu(x) + 1
≥ m(x) − 1 + 1
≥ m(y) − 1 ,
so y indeed adds an edge to pu(y), and by the induction hypothesis we are done.
Lemma 6. The spanner H has stretch at most 2k − 1.
Proof. Since H is a subgraph of G, it suffices to prove for any (x, y) ∈ E, that dH (x, y) ≤ 2k − 1. Let
u be the vertex maximizing m(x) = mu(x), and w.l.o.g assume m(x) ≥ m(y). By Corollary 4 we have
dG(u, x) ≤ k − 1, so dG(u, y) ≤ k, and y heard the message of u (which was sent to distance k). This
implies that mu(y) ≥ mu(x)− 1 = m(x)− 1 ≥ m(y)− 1, so y adds the edge (y, pu(y)) to the spanner. By
applying Claim 5 on x and y, we see that both have shortest paths to u that are fully contained in H. Since
dG(x, u) ≤ k − 1 and dG(y, u) ≤ k, these two paths provide stretch 2k − 1 between x, y.
2.1 Main Theorem
We now state our main theorem, from which we will derive several interesting corollaries in various settings.
Theorem 1. For any unweighted graph G = (V, E) on n vertices, any integer k ≥ 1, c > 3 and δ > 0,
there is a randomized algorithm that with probability at least (1− 1/c)· δ/(1 + δ) computes a spanner with
stretch 2k − 1 and number of edges at most
(cn)1+1/k
(1 + δ) ·
c − 1 − δ(n − 1) .
Proof. Let Z be the event that {∀u ∈ V, ru < k}. By Claim 3 we have Pr[Z] ≥ 1 − 1/c. Note that
conditioning on Z, by Lemma 6 the algorithm produces a spanner H = (V, E′) with stretch 2k − 1. In
particular, it must have at least n− 1 edges. Let X be the random variable E′− (n− 1), which conditioned
on Z takes only nonnegative values. By Lemma 2 we have E[X] = (cn)1/k · n − (n − 1). We now argue
8
that conditioning on Z will not affect this expectation by much. Indeed, by the law of total probability, for
any t, Pr[X = t] ≥ Pr[X = t Z] · Pr[Z]. Thus
Pr[Z] ≤
c − 1 ·h(cn)1/k · n − (n − 1)i .
E[X Z] ≤
E[X]
(3)
c
By Markov inequality,
Pr [X ≥ (1 + δ)E[X Z] Z] ≤
1
1 + δ
.
We conclude that
Pr [(X < (1 + δ)E[X Z]) ∧ Z] = Pr [X < (1 + δ)E[X Z] Z] · Pr[Z] ≥(cid:18)1 −
1
c(cid:19) ·
δ
1 + δ
.
If this indeed happens, then
E′ = X + n − 1
c
(3)
≤ (1 + δ) ·
= (1 + δ) ·
c − 1 ·h(cn)1/k · n − (n − 1)i + n − 1
(cn)1+1/k − (n − 1)
− δ(n − 1) .
c − 1
2.1.1
Implementation Details
Distributed Model.
It is straightforward to implement the algorithm in the LOCAL model of computation,
it will take k rounds to execute it – in each round, every vertex sends to its neighbors all the messages it
received so far. We claim that the algorithm can be implemented even when bandwidth is limited, i.e., in the
CONGEST model. This will require a small variation: in each round, every vertex v ∈ V will send to all its
neighbors the message (ru, dG(u, v)) for the vertex u that currently maximizes mu(v) = ru − dG(u, v). We
note that omitting all the other messages will not affect the algorithm, since if one such message would cause
some neighbor of v to add an edge to v, then the message about u will suffice, as the latter has the largest
mu(v) value. (Also recall that all vertices start their broadcast simultaneously, and do so for k rounds, so
any omitted message could not have been sent to further distance than the message from u, which implies
dropping it will have no effects on farther vertices as well.)
PRAM Model.
In the parallel model of computation, we can use a variant of the construction that appeared
in [MPX13, MPVX15]. Roughly speaking, vertex u will start its broadcast at time k − ⌈ru⌉, and every
vertex x will send only the first message that arrives to it (which realizes m(x)). As argued in [MPVX15],
the algorithm can be implemented in O(k log∗ n) depth and O(E) work.
Standard Centralized Model. Note that in the standard centralized model of computation, the running
time is at most the work of the PRAM algorithm, which is O(m). By taking constant c and δ, and repeating
the algorithm until the first success (we can easily check the number of edges of the spanner and that all
ru < k), we get a spanner with stretch 2k − 1 and O(n1+1/k) edges in expected time O(E).
9
2.2 Implications of Theorem 1
2.2.1 Standard Centralized Model and PRAM
The currently sparsest spanners which can be constructed in linear time are those of Halperin and Zwick
[HZ96]. They provide for any k ≥ 1, a deterministic algorithm running in O(m) time, that produces a
spanner with n1+1/k + n edges. We can improve their result for a wide range of k, albeit with a randomized
algorithm. First we show a near-linear time algorithm (which can be also executed in parallel), that provides
a spanner sparser than Halperin and Zwick in the range k ≥ 2 ln n/ ln ln n.7
Corollary 7. For any unweighted graph G = (V, E) on n vertices and m edges, and any integer k ≥ 2,
there is a randomized algorithm, that with high probability8 computes a spanner for G with stretch 2k − 1
k (cid:17) edges. The algorithm has O(k2 ln n ln∗ n) depth and the running time (or work)
and n1+1/k ·(cid:16)1 + O(ln k)
is O(kE).
k+1 ≥ 1
Proof. Apply Theorem 1 with parameters c = k and δ = 1/k. So with probability at least k−1
k
we obtain a spanner whose number of edges is at most
3k
1
·
(1 + 1/k) ·
(kn)1+1/k
k − 1 ≤ n1+1/k ·(cid:18)1 +
O(ln k)
k (cid:19) .
(4)
Run the algorithm C · k ln n times for some constant C. We noted in Section 2.1.1 that each run takes
O(k ln∗ n) depth and O(E) work, so the time bounds are as promised. Now, with probability at least
1 − (1 − 1/(3k))C·k ln n ≥ 1 − n−C/3, we achieved a spanner with number of edges as in (4) in one of the
executions.
Remark 1. Whenever k ≥ 2 ln n/ ln ln n, we have n1/k ≤ √ln n, so the number of edges in Corollary 7 is
n1+1/k + o(n), and the running time is O(kE).
2.2.2 Distributed Model
In a distributed setting we have the following result.
Corollary 8. For any unweighted graph G = (V, E) on n vertices, any k ≥ 1 and 0 < ǫ < 1, there is a
randomized distributed algorithm that with probability at least 1− ǫ computes a spanner with stretch 2k− 1
and O(n/ǫ)1+1/k edges, within k rounds.
Proof. Apply Theorem 1 with c = 3/ǫ and δ = 2/ǫ, so the success probability is at least
(1 − ǫ/3) · (1 − ǫ/(ǫ + 2)) > 1 − ǫ .
With these parameters, by Theorem 1, the number of edges in spanner will be bounded by O(n/ǫ)1+1/k.
7In fact, the factor 2 can be replaced by any 1 + ǫ for constant ǫ > 0.
8By high probability we mean probability at least 1 − n−C, for any desired constant C.
10
2.2.3 Ultra-Sparse Spanners and Skeletons
We now show that in the regime k ≥ ln n, our algorithm (that succeeds with probability close to 1) provides
a spanner whose number of edges is very close to n (as a function of k and the success probability). This
will hold in all computational models we considered. We note that for the centralized and PRAM models,
Corollary 7 gives high probability with roughly the same sparsity, albeit with larger depth and work.
Corollary 9. For any unweighted graph G = (V, E) on n vertices, and any integer k ≥ ln n and parameter
2/k < ǫ < 1, there is a randomized algorithm, that with probability at least 1− ǫ computes a spanner for G
with stretch 2k − 1 and n ·(cid:16)1 + O(ln n)
ǫ·k (cid:17) edges. The number of rounds in distributed model is k, in PRAM
it is O(k log∗ n) depth and O(E) work, and in the centralized model it is O(E) time.
Proof. Apply Theorem 1 with parameters c = k and δ = 2/ǫ, so with probability at least k−1
2/ǫ+1 ≥ 1−ǫ
we obtain a (2k − 1)-spanner. In the regime k ≥ ln n we have (kn)1/k ≤ e(2 ln n)/k ≤ 1 + O(ln n)/k, so
the number of edges is at most
k · 2/ǫ
(1 + 2/ǫ) ·
(kn)1+1/k
k − 1 − 2(n − 1)/ǫ ≤ n ·(cid:18)1 +
O(ln n)
ǫ · k (cid:19) .
(5)
Remark 2. The spanner of Corollary 9 can be used as a skeleton. E.g., one can take ǫ = o(1) and
k = O(log n), to obtain with probability 1 − o(1), a skeleton with n(1 + o(1)) edges, which is computed in
O(log n) rounds.
2.3 Weighted Graphs
Miller et al. [MPVX15] used their efficient algorithm for constructing O(k)-spanners with O(n1+1/k) edges
for unweighted graphs, to provide an efficient algorithm for constructing O(k)-spanners with O(n1+1/k log k)
edges for weighted graphs. In this section we argue that their scheme can be used to convert our algorithm
for constructing (2k − 1)-spanners with O(n1+1/k) edges for unweighted graphs (Theorem 1; see also Sec-
tion 2.1.1) into an efficient algorithm for constructing (2k − 1)(1 + ǫ)-spanners with O(n1+1/k · log k
ǫ ) edges
for weighted graphs.
The scheme of [MPVX15] works in the following way.
It partitions all edges of G = (V, E) into
⌈log1+ǫ ωmax⌉ = λ categories Et = {e ∈ E ω(e) ∈ [(1 + ǫ)t−1, (1 + ǫ)t)}, t = 1, 2, . . . , λ. (We assume
that the minimum weight is 1, and the maximum weight is ωmax. The last category Eλ should also contain
edges of weight exactly ωmax.) Now one defines ℓ = ⌈log1+ǫ kc⌉, for a sufficiently large constant c, graphs
Gj = (V, Ej ), j = 0, 1, . . . , ℓ − 1, Ej =S{Et t ≡ j( mod ℓ)}.
Observe that the edge weights in (each) Gj are well-separated, i.e., Ej is a disjoint union of at most
q = ⌈λ/ℓ⌉ edge sets E(1), . . . , E(q), such that the edge weights within each set are within a factor of 1 + ǫ
from one another. Moreover, if edge weights in E(i), i = 1, 2, . . . , q, are in the range [w(i), (1 + ǫ)w(i)),
then we have w(i) = w(1)(kc)i−1, for every i = 1, 2, . . . , q.
For each graph Gj, the scheme of [MPVX15] constructs an O(k)-spanner with O(n1+1/k) edges. It
then takes a union of ℓ = O( log k
ǫ ) such spanners as the ultimate O(k)-spanner of the original graphs. (In
fact, [MPVX15] used specifically ǫ = 1.) We will next outline the way in which [MPVX15] construct O(k)-
spanner with O(n1+1/k) edges for each Gj, and show how to modify it to provide a (2k− 1)(1 + ǫ)-spanner.
11
The scheme starts with running a routine of [MPVX15] that constructs an O(k)-spanner H (1) with
O(n1+1/k) edges for the unweighted graph (V (1), E(1)), V (1) = V , and constructing the exponential start
time partition P (1) for it. It then contracts each of the clusters of P (1) (which have unweighted radii at most
k − 1) into single vertices of V (2), and runs the unweighted spanner routine on (V (2), E(2)). As a result, it
constructs an O(k)-spanner H (2) and a partition P (2) of V (2), contracts all clusters of P (2) to get V (3), etc.
The final spanner returned by the scheme is H =Sq
The scheme guarantees stretch (1 + ǫ)(1 + O(k−(c−1)))O(k), because the blackbox routine for un-
weighted graphs provides stretch O(k) for each category of weights, but the weights are uniform only up to
a factor of 1 + ǫ. Also, the factor of 1 + O(k−(c−1)) appears, because one contracts clusters of unweighted
diameter O(k) of lower scales, on which all edge weights are a factor of roughly k−c smaller than the edge
weights on the current scale.
i=1 H (i).
In the analysis of H, [MPVX15] show that every vertex u is active (i.e., non-isolated vertex which
is not yet contracted into a larger super-vertex) for expected O(n1/k) phases, and when it is active, it con-
tributes expected O(n1/k) edges to the spanner of the current phase. Hence the overall size of the spanner
is O(n1+2/k), and by rescaling k′ = 2k, they ultimately get their result. (See the proof of Theorem 3.3 in
[MPVX15] for full details of this proof. We have sketched it for the sake of completeness.)
While the stretch analysis of [MPVX15] is sufficiently precise for our purposes, this is not the case with
the size analysis. Indeed, even when one plugs in stretch 2k − 1 of our unweighted spanner routine instead
of stretch O(k) of their routine, still one obtains a (2k − 1)(1 + ǫ)2(1 + k−(c−1))-spanner with O(n1+2/k)
edges, i.e., a (4k − 2)(1 + O(ǫ))-spanner with O(n1+1/k) edges (for each Gj).
In what follows we refine their size analysis, and show that, in fact, every vertex u contributes expected
O(n1/k) edges in all phases of the algorithm altogether (for a single graph Gj with well-separated edge
weights). Denote by r(i)
u the radius that u tosses from EXP(β) in the ith phase, i = 1, 2, . . . , q, assuming
that it is active on that phase. We say that a vertex v (which is active on phase i) is a candidate vertex
of phase i if its broadcast message reaches u no later than within one time unit after the time −r(i)
u , i.e.,
−r(i)
v + d(v, u), where d(v, u) is the unweighted distance between v and u in the graph on
which the unweighted spanner routine is invoked on phase i.
u + 1 ≥ −r(i)
Let X (i) denote the random variable counting the number of such candidate vertices on phase i, for
i = 1, 2, . . . , q. (Recall that these are the vertices which might cause u to add an edge to the spanner.) Let
j denote the random variable which is the phase in which u was contracted (and j = q if there is no such
phase). That is, j indicates the level in which the broadcast of some candidate vertex v has −r(i)
v +d(v, u) <
−r(i)
u . Denote also by X (i), i = 1, 2, . . . , q, the total number of candidates u sees between the beginning
of phase i, and up until phase j, where a candidate vertex v reaches u before time −r(j)
u . On that phase
j, X (i) counts the number of candidates with index not larger than that of the candidate vertex v (assume
every vertex has an arbitrary distinct index in {1, . . . n}). Note that for i > j, by definition, X (i) = 0. Also,
in particular, X = X (1) is at least the total contribution of edges u adds to the spanner, except for up to
expected O(n1/k) edges that it might contribute on phase j (as shown in Lemma 2).
We next argue that for any t = 0, 1, 2, . . .,
Pr( X > t) ≤ (1 − e−β)t .
(6)
This implies that
E( X) =
∞
Xt=0
Pr( X > t) ≤
∞
Xt=0
(1 − e−β)t = eβ = O(n1/k) .
12
First, note that
Pr( X > t) =
∞
Xt(1)=0
Pr(X (1) = t(1)) · Pr( X > t X (1) = t(1)) .
For t(1) > t, we have
Pr( X > t X (1) = t(1)) = (1 − e−β)t .
v + d(v, u) ≥ −r(1)
To justify this equation, note that the left-hand side is exactly the probability that on the first phase, all the
t first candidate vertices v have −r(1)
u , conditioned on them being candidates, i.e., on
−rv + d(v, u) ≤ −r(1)
u + 1. Since these are independent shifted exponential random variables, the equation
follows from the memoryless property of the exponential distribution.
For t(1) ≤ t, we have
Pr( X > t X (1) = t(1)) = (1 − e−β)t(1)
= (1 − e−β)t(1)
· Pr( X (2) > t − t(1) X (1) = t(1), X (1) ≥ t(1))
· Pr( X (2) > t − t(1) X (1) ≥ X (1)) .
(Again, (1 − e−β)t(1) is the probability that no candidate vertex of the first phase reached u before time
u , and so u was not contracted away at this phase.)
We conduct an induction on the phase, where the induction base is the last phase i = q. On the last
−r(1)
phase, for any h,
Pr( X (q) > h X (1) ≥ X (1), X (2) ≥ X (2), . . . , X (q−1) ≥ X (q−1)) ≤ (1 − e−β)h ,
because it is just the probability that none of the first h candidates (that have −r(q)
reaches u before time −r(q)
this probability is 0.)
u + 1)
u . (If there are fewer candidates or u was contracted in a previous phase, then
v + d(v, u) ≤ −r(1)
Hence by the inductive hypothesis,
Pr( X (2) > t − t(1) X (1) ≥ X (1)) ≤ (1 − e−β)t−t(1)
,
and so
for any t(1) ≤ t as well. Hence
Pr( X (1) > t X (1) = t(1)) ≤ (1 − e−β)t ,
Pr( X > t) =
≤
∞
Xt(1)=0
Xt(1)=0
∞
Pr(X (1) = t(1)) · Pr( X > t X (1) = t(1))
Pr(X (1) = t(1)) · (1 − e−β)t = (1 − e−β)t ,
as required.
Hence the expected contribution of every vertex is O(n1/k), and the overall spanner size for each Gj
is, in expectation, O(n1+1/k). The running time of the algorithm is expected to be O(E), following the
analysis of [MPVX15]. Moreover, as in [MPVX15], our algorithm can be implemented in PRAM model,
in O(log n · log∗ n · log Λ) depth, and O(E) work, where Λ is the aspect ratio of the input graph. We
summarize the result below.
13
Theorem 2. Given a weighted n-vertex graph G, and a pair of parameters k ≥ 1, 0 < ǫ < 1, our algorithm
computes a (2k − 1)(1 + ǫ)-spanner of G with O(n1+1/k · (log k)/ǫ) edges, in expected O(E) centralized
time, or in O(log n · log∗ n · log Λ) depth and O(E) work.
The result of Theorem 2 can be used in conjunction with the scheme of [ES16] to devise an algorithm
that computes (2k − 1)(1 + ǫ)-spanners of size O(n1+1/k(log k/ǫ)1/ǫ), with lightness (i.e., weight of the
spanner divided by the weight of the MST of the input graph) O(k · n1/k(1/ǫ)2+1/k), in expected time
O(E + min{n log n,Eα(n)}), where α(·) is an inverse-Ackermann function. This improves a result
of [ES16] that provides spanners with the same stretch and lightness, but with more edges (specifically,
O((k + (1/ǫ)2+1/k)n1+1/k), and using O(k·E + min{n log n,Eα(n)}) time. Recently, consequently to
our work, Alstrup et al. [ADF+17] further improved these bounds. We thus omit the details of our argument
that provides the aforementioned bounds.
3 An Efficient Centralized Construction of Nearly-Additive Spanners and
Emulators
3.1 A Basic Variant of the Algorithm
In this section we present an algorithm for constructing (1 + ǫ, β)-spanners, which can be efficiently imple-
mented in various settings. We start with the centralized setting. In this setting we present two variants of
our construction. The first variant presented in this section is somewhat simpler, while the second variant
presented in the next section provides better bounds.
Let G = (V, E) be an unweighted graph on n vertices, and let k ≥ 1, ǫ > 0 and 0 < ρ < 1/2 be
parameters. Unlike the algorithm of [EP04], our algorithm does not employ sparse partitions of [AP92].
The algorithm initializes the spanner H as an empty set, and proceeds in phases.
It starts with setting
P0 = {{v} v ∈ V } to be the partition of V into singleton clusters. The partition P0 is the input of phase 0
of our algorithm. More generally, Pi is the input of phase i, for every index i in a certain appropriate range,
which we will specify in the sequel.
Throughout the algorithm, all clusters C that we will construct will be centered at designated cen-
In particular, each singleton cluster C = {v} ∈ P0 is centered at v. We define Rad (C) =
ters rC.
max{dG(C)(rC, v) v ∈ C}, and Rad ( Pi) = maxC∈ Pi{Rad (C)}.
All phases of our algorithm except for the last one consist of two steps. Specifically, these are the
superclustering and the interconnection steps. The last phase contains only the interconnection step, and
the superclustering step is skipped. We also partition the phases into two stages. The first stage consists of
phases 0, 1, . . . , i0 = ⌊log(κρ)⌋, and the second stage consists of all the other phases i0 + 1, . . . , i1 where
κρ m− 2, except for the last phase ℓ = i1 + 1. The last phase will be referred to as the concluding
i1 = i0 +l κ+1
Each phase i accepts as input two parameters, the distance threshold parameter δi, and the degree param-
eter deg i. The difference between stage 1 and 2 is that in stage 1 the degree parameter grows exponentially,
while in stage 2 it is fixed. The distance threshold parameter grows in the same steady rate (increases by a
factor of 1/ǫ) all through the algorithm.
phase.
Next we describe the first stage of the algorithm. We start with describing its superclustering step. We
set deg i = n2i/κ, for all i = 0, 1, . . . , i0. Let R0 = 0, and δi = (1/ǫ)i + 4 · Ri, where Ri is determined by
the following recursion: Ri+1 = δi + Ri = (1/ǫ)i + 5· Ri. We will show that the inequality Rad ( Pi) ≤ Ri
will hold for all i.
14
On phase i, each cluster C ∈ Pi is sampled i.a.r. with probability 1/deg i. Let Si denote the set of
sampled clusters. We now conduct a BFS exploration to depth δi in G rooted at the set Si = SC∈Si{rC}.
As a result, a forest Fi is constructed, rooted at vertices of Si. For a cluster center r′ = rC′ of a cluster
C′ ∈ Pi\Si such that r′ is spanned by Fi, let rC be the root of the forest tree of Fi to which r′ belongs. (The
vertex rC is by itself a cluster center of a cluster C ∈ Si.) The cluster C′ becomes now superclustered in a
cluster C centered around the cluster C. (We also say that C′ is associated with C. We will view association
as a transitive relation, i.e., if C′ is associated with C and C′′ is associated with C′, we will think of C′′ as
associated with C as well.)
The cluster center rC of C becomes the new cluster center of C, i.e., r C = rC. The vertex set of the
new supercluster C is the union of the vertex set of C with the vertex sets of all clusters C′ which are
superclustered into C. The edge set T C of the new cluster C contains the BFS spanning trees of all these
clusters, and, in addition, it contains shortest paths from the forest Fi between rC and each rC′ as above. Si
is the set of superclusters created by this process. We set Pi+1 = Si. All edges that belong to the edgeset of
one of these superclusters are now added to the spanner H.
For each supercluster C, we write Rad ( C) = Rad (T C, r C ). Observe that Rad ( P0) ≤ R0 = 0, and
Rad ( S0) = max{Rad ( C) C ∈ S0} ≤ δ0 + R0 = R1 = 1. More generally we have
Rad ( Si) = max{Rad ( C) C ∈ Si} ≤ δi + Rad ( Pi) ≤ δi + Ri ≤ (1/ǫ)i + 5Ri = Ri+1.
(7)
Denote by Ui the set of clusters of Pi which were not superclustered into clusters of Si. In the intercon-
nection step for i ≥ 1, every cluster center rC of a cluster C ∈ Ui initiates a BFS exploration to depth 1
2 δi,
i.e., half the depth of the exploration which took place in the superclustering step. For each cluster center
rC′ for C′ ∈ Pi which is discovered by the exploration initiated in rC, the shortest path between rC and rC′
is inserted into the spanner H. The first phase i = 0 is slightly different: the exploration depth is set to be
1, and we add an edge from {v} ∈ U0 to all neighbors that are in U0. This completes the description of the
interconnection step.
Lemma 10. For any vertex v ∈ V , the expected number of explorations that visit v at the interconnection
step of phase i is at most deg i.
Proof. For i ≥ 1, assume that there are l clusters of Pi whose centers are within distance δi/2 from v. If
at least one of them is sampled to Si, then no exploration will visit v (since in the superclustering phase the
sampled center will explore to distance δi, and thus will supercluster all these centers). The probability that
none of them is sampled is (1− 1/deg i)l, in which case we get that l explorations visit v, so the expectation
is l · (1 − 1/deg i)l ≤ deg i (which holds for any l).
For i = 0, we note that we add an edge touching v iff none of its neighbors were sampled at phase 0
(as otherwise it would be clustered and thus not in U0). The expected number of edges added is once again
l · (1 − 1/deg i)l ≤ deg i (here l is the number of neighbors).
We also note the following lemma for future use, its proof follows from a simple Chernoff bound.
Lemma 11. For any constant c > 1, with probability at least 1 − 1/nc−1, for every vertex v ∈ V , at least
one among the deg i · c · ln n closest cluster centers rC′ with C′ ∈ Pi to v is sampled, i.e., satisfies C′ ∈ Si.
Observe that no vertex v ∈ V is explored by more than c · ln n · deg i explorations, with probability at
least 1 − n−(c−1). Indeed, otherwise when i ≥ 1 there would be more than c · ln n · deg i cluster centers
rC of unsampled clusters C ∈ Pi at pairwise distance at most δi. Applying Lemma 11 to any of them we
15
conclude that that the particular cluster C was superclustered by a nearby sampled cluster, i.e., C 6∈ Ui,
contradiction. In the case i = 0, we would have that v has at least c · ln n · deg i unsampled neighbors,
which occurs with probability at most n−c. Hence, by union-bound, every vertex v is explored by at most
c · ln n · deg i explorations, with probability at least 1 − n−(c−1).
Lemma 10 suggests that the interconnection step of phase i can be carried out in expected O(E · deg i)
time. Clearly, the superclustering step can be carried out in just O(E) time, and thus the running time of
the interconnection step dominates the running time of phase i. In order to control the running time, we
terminate stage 1 and move on to stage 2 when i0 = ⌊log(κρ)⌋, so that deg i0 ≤ nρ.
Observe also that the superclustering step inserts into the spanner at most O(n) edges (because we
insert a subset of edges of Fi, and Fi is a forest), and by Lemma 10 the interconnection step inserts in
expectation at most O( Pi · deg i · ((1/ǫ)i + Ri)) = O( Pi · deg i · (1/ǫ)i) edges. (We will soon show that
Ri = O((1/ǫ)i−1).) A more detailed argument providing an upper bound on the number of edges inserted
by the interconnection step will be given below.
Lemma 12. For all i, 1 ≤ i ≤ ℓ, for every pair of clusters C ∈ Ui, C′ ∈ Pi at distance at most 1
2 (1/ǫ)i
from one another, a shortest path between the cluster centers of C and C′ was inserted into the spanner H.
Moreover, for any pair {v} ∈ U0, {v′} ∈ P0, such that e = (v, v′) ∈ E, the edge e belongs to H.
Proof. We start with proving the first assertion of the lemma. For some index i, 1 ≤ i ≤ ℓ, and a pair
C ∈ Ui, C′ ∈ Pi of clusters, let rC , rC′ be the respective cluster centers. Then we have
dG(rC, rC′) ≤ Rad (C) + dG(C, C′) + Rad (C′)
≤ dG(C, C′) + 2 · Ri
(1/ǫ)i + 2 · Ri =
≤
1
2
1
2
δi,
and so a shortest path between rC and rC′ was inserted into the spanner H.
The second assertion of the lemma is guaranteed by the interconnection step of phase 0.
Next we analyze the radii of clusters' collections Pi, for i = 1, 2, . . ..
Lemma 13. For i = 0.1, . . . , ℓ, the value of Ri is given by
Ri =
i−1
Xj=0
(1/ǫ)j · 5i−1−j.
Proof. The proof is by induction of the index i. The basis (i = 0) is immediate as R0 = 0. For the induction
hypothesis, note that
Ri+1 = δi + Ri = (1/ǫ)i + 5 · Ri
(1/ǫ)j · 5i−1−j
i−1
= (1/ǫ)i + 5 ·
Xj=0
Xj=0
(1/ǫ)j · 5i−j ,
=
i
as required.
16
Observe that Lemma 13 implies that for ǫ < 1/10, we have Ri = 5i−1 · 1/(5ǫ)i−1
1−5ǫ · (1/ǫ)i−1 ≤
2 · (1/ǫ)i−1. Recall that Pi = Si−1. Hence Rad ( Pi) = Rad ( Si−1). By inequality (7), we have Rad ( Pi) ≤
(1/ǫ)i−1 + 5 · Ri−1 = Ri, for all i = 0, 1, . . . , ℓ.
1/(5ǫ)−1 ≤ 1
We analyze the number of clusters in collections Pi in the following lemma.
Lemma 14. For i = 0, 1, . . . , i0,
κ
Pi ≤ 2 · n1− 2i−1
κ )}.
with probability at least 1 − exp{−Ω(n1− 2i−1
Proof. The probability that a vertex v ∈ V will be a center of a cluster in Pi isQi−1
Thus the expected size of Pi is n1−(2i−1)/κ, and by Chernoff bound,
Pr[ Pi ≥ 2E[ Pi]] ≤ exp{−Ω(E[ Pi])} = exp{−Ω(n1− 2i−1
κ )} .
,
(8)
j=0 1/deg j = n−(2i−1)/κ.
Since for ρ < 1/2 and i ≤ i0 = ⌊log(κρ)⌋, we have n1− 2i+1−1
whp for all 0 ≤ i ≤ i0, Pi+1 = Si = O(n1− 2i+1−1
The total expected running time of the first stage is at most
κ ≥ n1−2ρ = ω(log n), we conclude that
). Hence in particular, Pi0+1 = O(n1−ρ+1/κ), whp.
κ
O(E)
i0
Xi=1
deg i = O(E · n
2i0
κ ) = O(E · nρ) .
Since each superclustering step inserts at most O(n) edges into the spanner, the overall number of edges
inserted by the i0 superclustering steps of stage 1 is O(n · i0) = O(n · log(κρ)). The expected number of
edges added to the spanner by the interconnection step of phase i is at most
O( Pi · deg i · (1/ǫ)i) = O(n1− 2i−1
κ
· (1/ǫ)log(κρ) · n
2i
κ ) = O(n1+1/κ · (1/ǫ)log(κρ)) .
Next we describe stage 2 of the algorithm, i.e., phases i = i0+1, i0+2, . . . , i1, where i1 = i0+⌈ κ+1
κρ ⌉−2.
All these phases are executed with the same fixed degree parameter deg i = nρ. On the other hand, the
distance threshold keeps growing in the same steady rate as in stage 1, i.e., it is given by δi = (1/ǫ)i + 4Ri.
The sets Pi0+1, Pi0+2, . . . , Pi1 on which phases i0 + 1, i0 + 2, . . . , i1, respectively, operate are defined by
Pi0+i = Si0+i−1, for 1 ≤ i ≤ i1 − i0.
Lemma 13 keeps holding for these additional i1 − i0 phases, i.e.,
Rad ( Pi) ≤ Ri ≤ 2 · (1/ǫ)i−1 .
(We assume all through that ǫ < 1/10.)
another, their centers are interconnected in the spanner by a shortest path between them.
Also, for every pair of clusters C ∈ Ui and C′ ∈ Pi which are at distance at most 1
In addition, for every i ∈ [i0, i1], the expected size of Pi+1 is
2 (1/ǫ)i from one
E[ Pi+1] = n ·
i
Yj=0
1/deg j ≤ n1+1/κ−(i+1−i0)ρ .
17
By Chernoff bound, for every such i, with probability at least 1 − exp{−Ω(nρ)}, we have
Assuming that nρ = ω(1), we conclude that whp
Pi+1 = Si ≤ 2 · n1+1/κ−ρ−(i−i0)ρ .
Pi1+1 = Si1 ≤ O(n1+1/κ−(i1+1−i0)ρ) = O(n1+1/κ−(⌈ κ+1
κρ ⌉−1)ρ) = O(nρ) .
(9)
(For the assumption above to hold we will need to assume that ρ ≥ log log n
this assumption is valid in our setting.)
2 log n , say. We will show soon that
O(E · deg i · (1/ρ)) = O(Enρ/ρ).
κρ ⌉ − 2 ≤ 1/ρ additional phases is expected to be at most
The time required to perform these ⌈ κ+1
The final collection of clusters Pi1+1 is created by setting Pi1+1 = Si1.
We will next bound the expected number of edges inserted into the spanner during stage 2 of the al-
gorithm. Each of the forests Fi, i ∈ [i0 + 1, i1], created during the superclustering steps contributes at
most n − 1 edges. The interconnection step of phase i + i0 contributes in expectation at most O( Pi+i0 ·
deg i+i0 · (1/ǫ)i) ≤ O(n1+1/κ−iρ · (1/ǫ)log(κρ)+i) edges. Assuming that 1/ǫ < nρ/2, this becomes a
geometric progression, so the overall expected number of edges inserted into the spanner on stage 2 is
O(n1+1/κ · (1/ǫ)log(κρ)). (We will show the validity of this assumption in the end of this section.)
Finally, we describe the concluding phase of the algorithm, i.e., phase ℓ = i1 + 1. In this phase we skip
the superclustering step (as the number of clusters is already sufficiently small), and proceed directly to the
interconnection step.
1
j=0
On this step each of the cluster centers rC for C ∈ Pℓ conducts a BFS exploration in G to depth
2 (1/ǫ)ℓ + 2Rℓ. (Essentially, we define Uℓ = Pℓ, and perform the usual interconnection step of the
2 δℓ = 1
algorithm.) By (9), the number of edges inserted by this step into the spanner is whp only O( Pℓ2·(1/ǫ)ℓ) =
O( Pi1+12·(1/ǫ)i1+1) = O(n2ρ·(1/ǫ)log(ρκ)+1/ρ). Recall that we assume that ρ < 1/2. Hence this number
of edges is sublinear in n.
Hence the overall expected number of edges in the spanner is H = O(n1+1/κ · (1/ǫ)log(κρ)). Observe
also that the running time of the last phase is O(E · nρ). Hence the overall expected running time of the
algorithm is O(Enρ/ρ). It remains to analyze the stretch of the resulting spanner H.
Let U = Sℓ
Uj. Observe that every singleton cluster {v} ∈ P0 is associated with exactly one cluster
of U, i.e., U is a partition of V . Note that Rad ( U0) = 0, Rad ( U1) ≤ 1 = R1, and for every j ∈ [ℓ], we have
Rad ( Uj) ≤ Rad ( Pj) ≤ Rj ≤ 2 · (1/ǫ)j−1. Denote c = 2. Recall also (see Lemma 12) that for j ≥ 1, for
every pair of clusters C, C′ ∈ Uj at distance at most 1
2 (1/ǫ)j from one another, a shortest path between the
cluster centers of this pair of clusters in G was added to the spanner H. Moreover, neighboring clusters of
U0 are also interconnected by a spanner edge.
Lemma 15. Consider a pair of indices 0 ≤ j < i ≤ ℓ, and a pair of neighboring (in G) clusters C′ ∈ Uj,
C ∈ Ui, and a vertex w′ ∈ C′ and the center r of C. Then the spanner H contains a path of length at most
3Rad ( Uj) + 1 + Rad ( Ui) between w′ and r.
Proof. Let (z′, z) ∈ E ∩ (C′ × C) be an edge connecting this pair of clusters. There exists a subcluster
C′′ ⊆ C, C′′ ∈ Pj such that z ∈ C′′. Hence the interconnection step of phase j inserted a shortest path
π(r′, r′′) in G between the cluster centers r′ of C′ and r′′ of C′′ into the spanner H. Note that the distance
between r′, r′′ is at most 1 + 2Rad ( Pj) ≤ 1 + 4(1/ǫ)j−1 < 1/2 · (1/ǫ)j, since we assume ǫ < 1/10.
Hence a path between w′ and r in H can be built by concatenating a path π(w′, r′) between w′ and r′ in
18
the spanning tree T (C′) of C′ with the path π(r′, r′′) in H, and with the path π(r′′, r) in the spanning tree
T (C) of C. (Note that both r′′ and r belong to C.) Its length is at most
π(w′, r′) + π(r′, r′′) + π(r′′, r) ≤ Rad (C′) + (Rad (C′) + 1 + Rad (C′′)) + Rad (C)
≤ 3Rad ( Uj) + 1 + Rad ( Ui) .
Now we are ready to analyze the stretch of our spanner.
Lemma 16. Suppose ǫ ≤ 1/10. Consider a pair of vertices u, v ∈ V . Fix a shortest path π(u, v) between
them in G, and suppose that for some index i ∈ [0, ℓ], all vertices of π(u, v) are clustered in the set U (i)
defined by U (i) =Si
Uj. Then
j=0
i
dH(u, v) ≤ (1 + 16c · ǫ · i)dG(u, v) + 4
Rj · 2i−j .
Xj=1
For the induction step, consider first a pair of vertices x, y such that π(x, y) ≤ 1
Proof. The proof is by induction on i. For the induction basis i = 0, observe that all vertices of π(u, v) are
clustered in U0, and thus all edges of π(u, v) are inserted into the spanner on phase 0. Hence dH (u, v) =
dG(u, v).
2 (1/ǫ)i, and V (π(x, y)) ⊆
U (i). Let z1 and z2 be the leftmost and the rightmost Ui-clustered vertices in π(x, y), if exist. (The case
when both these vertices exist is the one where the largest stretch is incurred; cf. [EP04].) Let C1, C2 ∈ Ui
be their respective clusters, i.e., z1 ∈ C1, z2 ∈ C2. Let w1 (respectively, w2) be the neighbor of z1 (resp.,
z2) on the subpath π(x, z1) (resp., π(z2, y)) of π(x, y), and denote by C′1 and C′2 the respective clusters of
w1 and w2. Observe that C′1, C′2 ∈ U (i−1).
Denote r1 and r2 the cluster centers of C1 and C2, respectively. The spanner H contains a path of length
at most dG(r1, r2) between these cluster centers. Also, by Lemma 15, since C′1 and C1 are neighboring
clusters, the spanner H contains a path of length at most 3Rj + 1 + Ri ≤ 2Ri + 1 between w1 and r1, and
a path of at most this length between r2 and w2. (For ǫ < 1/10, 3Rj ≤ Ri, for all j < i.) Observe also
that the subpaths π(x, w1) and π(w2, y) of π(x, y) have all their vertices clustered in U (i−1), and thus the
induction hypothesis is applicable to these subpaths.
Hence
dH (x, y) ≤ dH (x, w1) + dH (w1, r1) + dH (r1, r2) + dH(r2, w2) + dH(w2, y)
≤ (1 + 16c · ǫ(i − 1))dG(x, w1) + 4
Rj · 2i−1−j + 2Ri + 1 + (dG(C1, C2) + 2Ri)
i−1
Xj=1
+2Ri + 1 + (1 + 16c · ǫ(i − 1)) · dG(w2, y) + 4
i−1
Xj=1
Rj · 2i−1−j
= (1 + 16c · ǫ(i − 1)) · (dG(x, w1) + dG(w2, y)) + dG(C1, C2) + 4Ri + 2 + 8
Note also that
i−1
Xj=1
Rj · 2i−1−j.
dG(x, y) = dG(x, w1) + 1 + dG(z1, z2) + 1 + dG(w2, y) ≥ dG(x, w1) + dG(C1, C2) + 2 + dG(w2, y).
19
Hence
dH(x, y) ≤ (1 + 16c · ǫ(i − 1))dG(x, y) + 4Ri + 8
Rj · 2i−1−j
= (1 + 16c · ǫ(i − 1))dG(x, y) + 4
i
Xj=1
i−1
Xj=1
Rj · 2i−j.
Now consider a pair of vertices u, v such that all vertices of π(u, v) are clustered in U (i), without any
restriction on π(u, v). We partition π(u, v) into segments π(x, y) of length exactly ⌊ 1
2 (1/ǫ)i⌋, except
maybe one segment of possibly smaller length. Inequality (10) applies to all these segments. Hence
dH (u, v) ≤ (1 + 16c · ǫ(i − 1))dG(u, v) + 4
Rj · 2i−j
≤ 1 + 16c · ǫ(i − 1) +
i
dG(u, v)
2 (1/ǫ)i − 1⌋ + 4
1
Rj · 2i−j⌊
Xj=1
4Pi
2 (1/ǫ)i − 1 ! dG(u, v) + 4
j=1 Rj · 2i−j
1
i
Xj=1
Rj · 2i−j.
i
Xj=1
It remains to argue that 8 Pi
j=1 Rj 2i−j
(1/ǫ)i−2 ≤ 16c · ǫ. Recall that for every j, we have Rj ≤ c · (1/ǫ)j−1. Since
1/ǫ ≥ 10, the left-hand-side is at most
10c · ǫi
i
Xj=1
(1/ǫ)j−1 · 2i−j = 10c · ǫ
i
Xj=1
(2ǫ)i−j = 10c · ǫ
i−1
Xh=0
(2ǫ)h ≤ 16c · ǫ .
Observe that (as ǫ ≤ 1/10), we have
Rj · 2i−j ≤ c
i
i
Xj=1
i
2ǫ(cid:19)j−1
i−1
Xj=1
(1/ǫ)j−1 · 2i−j = c · 2i−1 ·
2ǫ(cid:19)j
Xj=0(cid:18) 1
= c · 2i−1
2 · (1/ǫ)i − 2i−1
Xj=1(cid:18) 1
= c · 2i−1 (1/2ǫ)i − 1
(1/2ǫ) − 1
= O (cid:18) 1
ǫ(cid:19)i−1! .
= c · ǫ ·
2 − ǫ
1
1
Note also that the condition of the last lemma holds with i = ℓ for every pair u, v ∈ V of vertices. Hence
Corollary 17. For every pair u, v ∈ V ,
dH (u, v) ≤ (1 + 16c · ℓ · ǫ)dG(u, v) + O((1/ǫ)ℓ−1) .
Recall that the spanner H contains, whp, H = O(n1+1/κ·log n·(1/ǫ)log(κρ)+n·log n·(1/ǫ)log(κρ)+1/ρ)
edges, and the expected running time required to construct it is O(E· nρ/ρ). Recall also that ℓ = i1 + 1 ≤
(cid:17)log κ+1/ρ(cid:19). The
log(κρ) + 1/ρ + 1. Set now ǫ′ = 16c · ℓ · ǫ. We obtain stretch (cid:18)1 + ǫ′, O(cid:16) log κ+1/ρ
condition ǫ < 1/10 translates now to ǫ′ ≤ 1.6c(log(κρ) + 1/ρ). We will replace it by a simpler stronger
condition ǫ ≤ 1.
ǫ′
20
Corollary 18. For any parameters 0 < ǫ ≤ 1, κ ≥ 2, and ρ > 0, and any n-vertex unweighted graph
(cid:17)log κ
G = (V, E), our algorithm computes a (1+ǫ, β)-spanner with expected number of edges O(cid:16) log κ+1/ρ
·
n1+1/κ, in expected time O(E · nρ/ρ), where
ǫ
β =(cid:18) O(log κ + 1/ρ)
ǫ
(cid:19)log κ+1/ρ
.
A particularly useful setting of parameters is ρ = 1/ log κ. Then we get a spanner with expected
1
ǫ (cid:17)log κ
· n1+1/κ edges, in time O(E · n
O(cid:16) log κ
We remark that it makes no sense to set ρ < 1/κ, as the resulting parameters will be strictly worse than
when ρ = 1/κ. Also our assumptions that ρ > log log n/(2 log n) and 16c · ℓ/ǫ < nρ/2 are justified, as
otherwise we get β ≥ n, so a trivial spanner will do.
3.2 An Improved Variant of the Algorithm
log κ · log κ), and β = O(cid:16) log κ
ǫ (cid:17)2 log κ
.
In this section we show that the leading coefficient O((log κ + 1/ρ)/ǫ)log κ of n1+1/κ in the size of the
spanner can be almost completely eliminated at essentially no price. We also devise here yet sparser con-
structions of emulators.
ǫ (cid:17)i
Recall also that the interconnection step of the ith phase contributes Ni · deg i ·(cid:16) c′ℓ
For i = 0, 1, . . . , ℓ, denote by Ni = Pi the expected number of clusters which take part in phase i.
edges in expectation,
where ℓ is the total number of steps, and c′ is a universal constant. Note that the contribution of the intercon-
nection step dominates the contribution of the superclustering step in the current variant of the algorithm,
and it will still be the case after the modification that we will now introduce. Hence we will now focus on
decreasing the number of edges contributed by the interconnection steps.
We keep the structure of the algorithm intact, and have the values of distance thresholds δi unchanged.
The only change is in the degree sequence deg 0, deg 1, . . . of degree parameters used in phases 0, 1, . . .,
respectively. Next, we describe our new setting of these parameters for stage 1 of the algorithm (i.e., phases
i, 1 ≤ i ≤ i0). In the case that κ ≥ 16 let a = log log κ, otherwise, when κ < 16, let a = 2. Define
i0 = min{⌊log(aκρ)⌋,⌊κρ⌋}, and for i = 0, 1, . . . , i0 let deg i = n(2i−1)/(aκ)+1/κ. We now have that for
i ≤ i0 + 1,
Ni = n
1/deg j = n1− 2i−1−i
aκ − i
κ ,
i−1
Yj=0
and in particular, when i0 = ⌊log(aκρ)⌋ we have Ni0+1 ≤ n1− aκρ−1−(i0+1)
i0 ≥ 1). Whenever i0 = ⌊κρ⌋ we also have Ni0+1 ≤ n1− i0+1
κ ≤ n1−ρ. Additionally, we always have
2i−1
aκ + 1
κ ≤ n1−ρ (since a ≥ 2 and
− i0+1
aκ
κ = n1+ i
aκ− i−1
κ .
We restrict ourselves to the case that
Ni · deg i = n1− 2i−1−i
aκ − i
c′ℓ
ǫ ≤ n
κ · n
which holds whenever κ ≤ c0·log n
inserted at phase i ≤ i0 is at most
1
2κ /2 ,
(10)
log(ℓ/ǫ) , for a sufficiently small constant c0. Now the expected number of edges
ǫ (cid:19)i
Ni · deg i ·(cid:18) c′ℓ
≤ n1+ i
2κ− i−1
2 !i
κ · n1/(2κ)
21
= n1+ 1
κ /2i .
(11)
Thus the total expected number of edges inserted in the first stage is O(n1+1/κ). The second stage proceeds
by setting deg i0+1 = nρ/2, and in all subsequent phases i0 + i, with i = 2, 3, . . . , i1− i0, we have deg i = nρ
as before. The "price" for reducing the degree in the first phase of stage two is that the number of phases i1
may increase by an additive 1. It follows that Ni0+1 · deg i0+1 ≤ n1−ρ/2. For i ≥ 2, at phase i0 + i we have
Ni0+i ≤ n1−3ρ/2−(i−2)ρ. We set i1 = ⌊1/ρ⌋, so that Ni1+1 ≤ nρ/2, and whp we have that Ni1+1 ≤ 2nρ/2.
We calculate
Ni0+i · deg i0+i ≤ n1−ρ/2−(i−2)ρ .
Note that i0/(2κ) ≤ ρ/2, which holds since i0 ≤ ⌊κρ⌋. Hence the condition (10) implies that
i0
(c′ℓ/ǫ)i0 ≤ n
2κ ≤ nρ/2. The total expected number of edges inserted at phase i0 + 1 is at most
Ni0+1 · deg i0+1 ·(cid:18) c′ℓ
ǫ (cid:19)i0+1
≤ n1−ρ/2 · nρ/2 ·
c′ℓ
ǫ ≤ n1+1/κ .
The expected contribution of phase i0 + i for i ≥ 2 is at most
Ni0+i · deg i0+i ·(cid:18)c′ℓ
ǫ (cid:19)i0+i
≤ n1−ρ/2−(i−2)ρ · nρ/2 · ni/(2κ)/2i ≤ n1+1/κ/2i ,
(12)
where the last inequality uses that ρ ≥ 1/κ (which we may assume w.l.o.g). This implies that the expected
number of edges in all these ⌊1/ρ⌋ phases is O(n1+1/κ).
.9
The upper bound on κ under which this analysis was carried out is c0·log n
Ω(log n)
log(ℓ/ǫ) ≥
log(1/ǫ)+log(1/ρ)+log(3) n
log(1/ǫ)+log(1/ρ)+log(3) n
We summarize this discussion with the following theorem.
c·log n
Theorem 3. For any unweighted graph G = (V, E) with n vertices , 0 < ǫ < 1/10, 2 ≤ κ ≤
for a constant c, and 1/κ ≤ ρ < 1/2, our algorithm computes a (1 + ǫ, β)-spanner with β = O( 1
ǫ (log κ +
1/ρ))log κ+1/ρ+max{1,log(3) κ} and expected number of edges O(n1+1/κ). The expected running time is
O(E · nρ/ρ).
.)
Note that the sparsest this spanner can be is O(n log log n), and at this level of sparsity its β =
O(log log n + 1/ρ)log log n+1/ρ. (To get this bound we set ǫ > 0 to be an arbitrary small constant, and
κ = c0 log n
log(3) n
This is sparser than the state-of-the-art efficiently-computable sparsest (1+ ǫ, β)-spanner due to [Pet10],
which has O(n(log log n)φ) edges, where φ = 1+√5
is the golden ratio. Moreover, this spanner has a
smaller β than the one of [Pet10] in its sparsest level. Denoting the latter as βP et, it holds that βP et ≈
O(log κ)1.44 log κ+1/ρ, i.e., for every setting of the time parameter ρ, the exponent of our β is smaller than
that of βP et.
2
3.2.1 Sparse Emulator
Finally, we note that if one allows an emulator instead of spanner, then we can decrease the size all the
way to O(n) when κ = log n. To achieve this, we insert single "virtual" edges instead of every path (of
length (c′ℓ/ǫ)i) between every pair of cluster centers that we choose to interconnect on phase i, for every
i. Analogously, in the superclustering step we also form a supercluster around a center rC of a cluster C
by adding virtual edges (rC, rC′) for each cluster C′ associated with C. The weight of each such edge is
9We denote log(k) n as the iterated logarithm function, e.g. log(3) n = log log log n.
22
defined by ω(rC, rC′) = dG(rC , rC′). The condition (10) was required to obtain converging sequences at
(11) and (12), but without the (c′ℓ/ǫ)i terms, the number of edges already forms a converging sequence at
each stage.
2i
Moreover, one can also use for emulators a shorter degree sequence than the one we used for spanners,
and as a result to save the additive term of log(3) κ in the exponent of β. Specifically, one can set deg i =
j=0 1/deg j = n1− 2i−1
κ /22i−1, for each i = 0, 1, . . . , i0 = ⌊log(κρ)⌋. As a result we get Ni = n ·Qi−1
n
·
22i−1−i, and thus the expected number of edges inserted at phase i ≤ i0 is at most
κ
Ni · deg i = n1+1/κ/2i .
As before, when the first stage concludes, we run one phase with deg i0+1 = nρ/2, and all subsequent phases
with deg i = nρ. To bound the expected number of edges added at phase i0 + 1 we need to note that
22i0 +1
≤ 22κρ ≤ nρ/2 as long as κ ≤ (log n)/4. (The latter can be assumed without affecting any of the
parameters by more than a constant factor). It follows that Ni0+1 · deg i0+1 = n1− 2i0+1−1
· 22i0 +1−1−(i0+1) ·
nρ/2 ≤ n1+1/κ. In the remaining phases Ni0+i ≤ n1+1/κ−(i−1)ρ for i ≥ 2, and the contribution of these
phases is a converging sequence. We conclude the discussion with the following theorem.
Theorem 4. For any unweighted graph G = (V, E) with n vertices, and for any parameters 0 < ǫ ≤
1, 2 ≤ κ ≤ (log n)/4, 1/κ ≤ ρ < 1/2, our algorithm computes a (1 + ǫ, β)-emulator with β =
O(cid:16) log κ+1/ρ
and expected number of edges O(n1+1/κ). The expected running time is O(E ·
(cid:17)log κ+1/ρ
nρ/ρ).
ǫ
κ
In particular, the algorithm produces a linear-size (1+ǫ, β)-emulator with β = O(cid:16) log log n+1/ρ
within this running time.
ǫ
(cid:17)log log n+1/ρ
3.3 Distributed and Streaming Implementations
In this section we provide efficient distributed and streaming algorithms for constructing sparse (1 + ǫ, β)-
spanners. The distributed algorithm works in the CONGEST model.
To implement phase 0, each vertex selects itself into S0 with probability n−1/κ, i.a.r.. In distributed
model vertices of S0 send messages to their neighbors. Each vertex u that receives at least one message,
picks an origin v ∈ S0 of one of these messages, and joins the cluster centered at v. It also sends negative
acknowledgements to all its other neighbors from S0. All unclustered vertices z insert all edges incident on
them into the spanner.
It is also straightforward to implement this in O(1) passes in the streaming model.
Each consecutive phase is now also implemented in a straightforward manner, i.e., BFS explorations
to depth δi in the superclustering steps are implemented via broadcasts and convergecasts in the distributed
model, and by δi passes in the streaming model. In the interconnection steps, however, we need to implement
many BFS explorations which may explore the same vertices. However, by Lemma 11 every vertex is
explored on phase i by up to O(degi · log n) explorations whp, it follows that in distributed setting this step
requires, whp, O(deg i · log n · δi) time. Also note thatPi δi = O(β).
In the streaming model we have two possible tradeoffs. The first uses expected O(n · deg i) space to
maintain for each vertex v the BFS parents and distance estimates, and requires just δi passes. To see that
such space suffices, recall that the expected number of explorations which visit any vertex v is at most deg i,
by Lemma 10. Whp, the space is O(n · deg i · log n) = O(n1+ρ · log n).
23
The second option in the streaming algorithm is to divide the interconnection step of phase i to c ·
deg i · log n subphases, for a sufficiently large constant c. On each subphase each exploration source, which
was not sampled on previous subphases, samples itself i.a.r. with probability 1/deg i. Then the sampled
exploration sources conduct BFS explorations to depth δi/2. For every vertex v, the expected number of
explorations that traverse it on each subphase is O(log n). Moreover, by Chernoff's inequality, whp, no
vertex v is ever traversed by more than O(log n) explorations. (Here we take a union-bound on all vertices,
all phases, and all subphases. The bad events are that some vertex is traversed by more than twice its
expectation explorations on some subphase.) Hence each subphase requires O(δi) passes, and whp, the
space requirement is O(n log n), plus the size of the spanner. After c · deg i · log n subphases, whp, each
exploration source is sampled on at least one of the subphases, and so the algorithm performs all the required
explorations.
Finally, the stretch analysis of distributed and streaming variants of our algorithm remains the same as
in the centralized case.
Hence we obtain the following distributed and streaming analogues of Theorem 3.
c·log n
Theorem 5. For any unweighted graph G = (V, E) with n vertices , 0 < ǫ < 1/10, 2 ≤ κ ≤
for a constant c, and 1/κ ≤ ρ < 1/2, our distributed algorithm (CONGEST model) computes a (1 + ǫ, β)-
ǫ (log κ + 1/ρ))log κ+1/ρ+max{1,log(3) κ} and expected number of edges O(n1+1/κ).
spanner with β = O( 1
The required number of rounds is whp O(nρ/ρ · β · log n).
expected space and O(nρ/ρ · log n · β) passes, or using O(n1+ρ · log n) space, whp, and O(β) passes.
Our streaming algorithm computes a spanner with the above properties, in either: O(n log n + n1+1/κ)
log(1/ǫ)+log(1/ρ)+log(3) n
The streaming algorithm described above can also be modified to provide a (1 + ǫ, β)-emulator as in
Theorem 4, within the same pass and space complexities.
4 Applications
In this section we describe applications of our improved constructions of spanners and emulators to com-
puting approximate shortest paths for a set S × V of vertex pairs, for a subset S ⊆ V of designated sources.
4.1 Centralized Setting
We start with the centralized setting. Here our input graph G = (V, E) is unweighted, and we construct a
(1 + ǫ, β)-emulator H of G with O(n1+1/κ) edges, in expected time O(E · nρ/ρ), where
β = O(cid:18) log κ + 1/ρ
ǫ
(cid:19)log κ+1/ρ
.
(13)
Observe that all edge weights in H are integers in the range [1, β]. We round all edge weights up to the
closest power of 1 + ǫ. Let H′ be the resulting emulator. Note that for any pair u, v of vertices, we have
dG(u, v) ≤ dH(u, v) ≤ dH ′(u, v) ≤ (1 + ǫ)dH(u, v)
≤ (1 + ǫ)2dG(u, v) + (1 + ǫ)β .
For a sufficiently small ǫ, (1 + ǫ)2 ≤ 3ǫ, and we rescale ǫ′ = 3ǫ. As a result the constant factor hidden by the
O-notation in the basis of β's exponent grows, but other than that H′ has all the properties of the emulator
24
H. Also, it employs only
t = O(cid:18) log β
ǫ (cid:19) = O(cid:18) (log 1/ǫ + log(log κ + 1/ρ)) · (log κ + 1/ρ)
ǫ
(cid:19)
different edge weights. Hence a single-source shortest path computation in H can be performed in O(H +
n log t) = O(n1+1/κ + n(log(1/ǫ) + log(log κ + 1/ρ))) time [OMSW10]. (See also [KMP11], Section 5.)
Hence computing S × V (1 + ǫ, β)-approximate shortest distances requires O(E · nρ/ρ + S(n1+1/κ +
n(log(1/ǫ) + log(log κ + 1/ρ)))) time.
Theorem 6. For any n, and for any parameters 0 < ǫ ≤ 1, 2 ≤ κ ≤ (log n)/4, 1/κ ≤ ρ ≤ 1/2, and any
n-vertex unweighted graph G = (V, E) with a set S ⊆ V , our algorithm computes (1 + ǫ, β)-approximate
S × V shortest distances in the centralized model in expected O(E · nρ/ρ + S(n1+1/κ + n(log(1/ǫ) +
log(log κ + 1/ρ)))) time, where β is given by (13).
If one is interested in actual paths rather than just in distances, then one can use our (1 + ǫ, β)-spanner
with
β = O(cid:18) log κ + 1/ρ
ǫ
(cid:19)log κ+1/ρ+max{1,log(3) κ}
,
(14)
O(log n)
log(1/ǫ)+log(1/ρ)+log(3) n
and O(n1+1/κ) edges, but restricting κ ≤
. After computing the spanner H with
these properties, we conduct BFS explorations on H originated at each vertex of S. The overall running
time becomes O(E · nρ/ρ + S · n1+1/κ).
Corollary 19. For any n, and for any parameters 0 < ǫ ≤ 1, κ ≤
, 1/κ ≤ ρ ≤ 1/2,
and any n-vertex unweighted graph G = (V, E), our algorithm computes (1 + ǫ, β)-approximate S × V
shortest paths in the centralized model in expected O(E · nρ/ρ + S · n1+1/κ) time, with β given by (14).
log κ . Then the running time of our algorithms from Theorem 6
and Corollary 19 become respectively O(E · n1/ log κ · log κ + S(n1+1/κ + n(log 1/ǫ + log log κ))) and
O(E · n1/ log κ · log κ + S · n1+1/κ) (we note that the former has smaller β given by (13), while the latter
has slightly larger β given by (14)).
The algorithm of Corollary 19 always outperforms the algorithm which can be derived by using the
spanner of [Pet10] within the same scheme. Specifically, the running time of that algorithm is at least
O(Enρ) + O(S · (n1+1/κ + n( log κ
ǫ )φ)), and the additive error βP et there is given by
A useful setting of parameters is ρ = 1
O(log n)
log(1/ǫ)+log(1/ρ)+log(3) n
βP et = O(cid:18) log κ + 1/ρ
ǫ
(cid:19)logφ κ+1/ρ
,
2
where φ = 1+√5
is the golden ratio. So βP et is typically polynomially larger than the additive error β in
our algorithm, e.g., for ρ = 1/ log κ we have βP et ≈ β1.22. (Setting ρ to be smaller than 1/ log κ makes less
sense, because then the additive errors β and βP et deteriorate. At any rate, as ρ tends to 0, the two estimates
approach each other.) Also, in the bound of [Pet10] there is a term of S · n · ( log κ
ǫ )φ)), which does not
occur in our construction.
25
4.2 Streaming Setting
In this section we show how efficient constructions of spanners and emulators for an unweighted graph G
in the streaming setting can be used for efficient computation of approximate shortest paths.
First, one can use O(nρ/ρ · log n · β) passes and expected space O(n1+1/κ + n · log n) to construct an
(1 + ǫ, β)-emulator H with β = (O(log κ + 1/ρ)/ǫ)log κ+1/ρ. One can now compute V × V (1 + ǫ, β)-
approximate shortest distances in G by computing exact V ×V shortest distances in H, using the same space
and without additional passes. (Observe that we do not store the output, as its size is larger than the size of
H.) In particular, one can use here space of O(n · log n), and have β = (O(log log n + 1/ρ)/ǫ)log log n+1/ρ.
It is also possible to set here ρ =
log log n ) passes, β = O( log log n
log log n , and obtain 2O( log n
)2 log log n.
Another option is to use space O(n1+ρ·log n), whp, and O(β) passes for constructing the same emulator
H. Again given the emulator we can compute V × V approximate shortest distances in G by computing
shortest distances in H offline.
Corollary 20. For any n and any parameters 0 < ǫ ≤ 1, 2 ≤ κ ≤ (log n)/4, 1/κ ≤ ρ ≤ 1/2, and any
n-vertex unweighted graph G, our streaming algorithm computes (1 + ǫ, β)-approximate shortest distances
for V × V with β given by (13). It uses in expectation either O(nρ/ρ · log n · β) passes and expected space
O(n1+1/κ + n · log n) or O(β) passes and space O(n1+ρ · log n), whp.
1
ǫ
If the actual paths rather than just distances are needed, then we compute a (1 + ǫ, β)-spanner H with
β given by (14) and with expected O(n1+1/κ) edges (with the restriction on κ as above). Then we compute
V × V (1 + ǫ, β)-approximate shortest paths in H offline, using space O(H).
Corollary 21. For any n, κ, ǫ, ρ and G as in Corollary 19, a variant of our streaming algorithm computes
(1 + ǫ, β)-approximate shortest paths for V × V with β given by (14). It uses either O(nρ/ρ · log n · β)
passes and expected space O(n1+1/κ + n · log n) or O(β) passes and space O(n1+ρ · log n), whp.
If one is interested only in S × V paths or distances, then it is possible to eliminate the additive term of
β by using O(S· β/ǫ) additional passes. These passes are used to compute exactly distances between pairs
(s, v) ∈ S × V , with dG(s, v) ≤ β/ǫ. The overall number of passes becomes O(S· β/ǫ + nρ/ρ· log n· β),
and space O(n1+1/κ + n · log n). Whenever S ≥ n1/κ log n, we can set ρ = log S
log n , and obtain
the following corollary.
Corollary 22. For any n, ǫ, ρ, κ and G as in Corollary 20, and any set S ⊆ V of size at least S ≥
n1/κ log n, a variant of our streaming algorithm computes (1 + ǫ)-approximate shortest distances for S × V
in expected
log n − log log n
ǫ · O(cid:18) 1
S
ǫ (cid:18)log κ +
log n
log S−log log n(cid:19)(cid:19)log κ+
log n
log S−log log n
(15)
passes, and space O(n1+1/κ + n · log n). To compute actual paths we have similar complexities10, but one
needs to restrict κ as in Corollary 19.
Observe that for a constant ǫ > 0 and S = nΩ(1), we can take a constant κ, so the number of passes
is O(S). One can also get for S = 2Ω(log n/ log log n), a streaming algorithm for computing (1 + ǫ)-
approximate shortest paths for S × V by setting κ = c log n/ log log log n, and obtaining O(S/ǫ) ·
O( log log n
)2 log log n passes and space O(n log n).
ǫ
10Though there is an additional additive term of log(3) κ in the exponent in (15).
26
References
[AB16]
Amir Abboud and Greg Bodwin. The 4/3 additive spanner exponent is tight. In Proceedings of
the 48th Annual ACM SIGACT Symposium on Theory of Computing, STOC 2016, Cambridge,
MA, USA, June 18-21, 2016, pages 351–361, 2016.
[ABCP93] Baruch Awerbuch, Bonnie Berger, Lenore Cowen, and David Peleg. Near-linear cost sequential
and distribured constructions of sparse neighborhood covers. In 34th Annual Symposium on
Foundations of Computer Science, Palo Alto, California, USA, 3-5 November 1993, pages
638–647, 1993.
[ABN11]
Ittai Abraham, Yair Bartal, and Ofer Neiman. Advances in metric embedding theory. Advances
in Mathematics, 228(6):3026 – 3126, 2011.
[ABP17]
Amir Abboud, Greg Bodwin, and Seth Pettie. A hierarchy of lower bounds for sublinear
additive spanners.
In Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on
Discrete Algorithms, pages 568–576, 2017.
[ACIM99] Donald Aingworth, Chandra Chekuri, Piotr Indyk, and Rajeev Motwani. Fast estimation of
diameter and shortest paths (without matrix multiplication). SIAM J. Comput., 28(4):1167–
1181, 1999.
[ADD+93]
I. Althofer, G. Das, D. Dobkin, D. Joseph, and J. Soares. On sparse spanners of weighted
graphs. Discrete Comput. Geom., 9:81–100, 1993.
[ADF+17] Stephen Alstrup, Soren Dahlgaard, Arnold Filtser, Morten Stockel, and Christian Wulff-Nilsen.
Personal communication, 2017.
[AHL02]
Noga Alon, Shlomo Hoory, and Nathan Linial. The moore bound for irregular graphs. Graphs
and Combinatorics, 18(1):53–57, 2002.
[AN12]
[AP92]
Ittai Abraham and Ofer Neiman. Using petal-decompositions to build a low stretch spanning
tree. In STOC, pages 395–406, 2012.
B. Awerbuch and D. Peleg. Routing with polynomial communication-space tradeoff. SIAM J.
Discrete Mathematics, 5:151–162, 1992.
[Awe85]
B. Awerbuch. Complexity of network synchronization. J. ACM, 4:804–823, 1985.
[Bar96]
[Bar98]
[Bar04]
Yair Bartal. Probabilistic approximations of metric spaces and its algorithmic applications. In
FOCS, pages 184–193, 1996.
Yair Bartal. On approximating arbitrary metrices by tree metrics. In Proceedings of the Thirti-
eth Annual ACM Symposium on Theory of Computing, STOC '98, pages 161–168, New York,
NY, USA, 1998. ACM.
Yair Bartal. Graph decomposition lemmas and their role in metric embedding methods. In
Algorithms - ESA 2004, 12th Annual European Symposium, Bergen, Norway, September 14-
17, 2004, Proceedings, pages 89–97, 2004.
27
[Bas08]
S. Baswana. Streaming algorithm for graph spanners - single pass and constant processing
time per edge. Inf. Process. Lett., 106(3):110–114, 2008.
[BCE05]
B´ela Bollob´as, Don Coppersmith, and Michael Elkin. Sparse distance preservers and additive
spanners. SIAM J. Discrete Math., 19(4):1029–1055, 2005.
[BGK+14] Guy E. Blelloch, Anupam Gupta, Ioannis Koutis, Gary L. Miller, Richard Peng, and Kanat
Tangwongsan. Nearly-linear work parallel SDD solvers, low-diameter decomposition, and
low-stretch subgraphs. Theor. Comp. Sys., 55(3):521–554, October 2014.
[BKMP10] Surender Baswana, Telikepalli Kavitha, Kurt Mehlhorn, and Seth Pettie. Additive spanners
and (alpha, beta)-spanners. ACM Transactions on Algorithms, 7(1):5, 2010.
[BR10]
[BS03]
[BS07]
[BW15]
[Che13]
[Coh99]
[Coh00]
Ajesh Babu and Jaikumar Radhakrishnan. An entropy based proof of the moore bound for
irregular graphs. CoRR, abs/1011.1058, 2010.
S. Baswana and S. Sen. A simple linear time algorithm for computing a (2k − 1)-spanner
of O(n1+1/k) size in weighted graphs. In Proceedings of the 30th International Colloquium
on Automata, Languages and Programming, volume 2719 of LNCS, pages 384–396. Springer,
2003.
Surender Baswana and Sandeep Sen. A simple and linear time randomized algorithm for
computing sparse spanners in weighted graphs. Random Struct. Algorithms, 30(4):532–563,
2007.
Gregory Bodwin and Virginia Vassilevska Williams. Very sparse additive spanners and emula-
tors. In Proceedings of the 2015 Conference on Innovations in Theoretical Computer Science,
ITCS 2015, Rehovot, Israel, January 11-13, 2015, pages 377–382, 2015.
Shiri Chechik. New additive spanners.
In Proceedings of the Twenty-Fourth Annual ACM-
SIAM Symposium on Discrete Algorithms, SODA '13, pages 498–512, Philadelphia, PA, USA,
2013. Society for Industrial and Applied Mathematics.
E. Cohen. Fast algorithms for t-spanners and stretch-t paths. SIAM J. Comput., 28:210–236,
1999.
Edith Cohen. Polylog-time and near-linear work approximation scheme for undirected shortest
paths. J. ACM, 47(1):132–166, 2000.
[DGPV08] Bilel Derbel, Cyril Gavoille, David Peleg, and Laurent Viennot. On the locality of distributed
sparse spanner construction. In Proceedings of the Twenty-Seventh Annual ACM Symposium
on Principles of Distributed Computing, PODC 2008, Toronto, Canada, August 18-21, 2008,
pages 273–282, 2008.
[DHZ00]
D. Dor, S. Halperin, and U. Zwick. All-pairs almost shortest paths. SIAM J. Comput., 29:1740–
1759, 2000.
[DMP+03] D. Dubhashi, A. Mei, A. Panconesi, J. Radhakrishnan, and A. Srinivisan. Fast distributed
algorithm for (weakly) connected dominating sets and linear-size skeletons, 2003.
28
[DMZ06]
Bilel Derbel, Mohamed Mosbah, and Akka Zemmari. Fast distributed graph partition and
application.
In 20th International Parallel and Distributed Processing Symposium (IPDPS
2006), Proceedings, 25-29 April 2006, Rhodes Island, Greece, 2006.
[EEST05] Michael Elkin, Yuval Emek, Daniel A. Spielman, and Shang-Hua Teng. Lower-stretch span-
ning trees. In STOC, pages 494–503, 2005.
[Elk04]
[Elk05]
[Elk07a]
M. Elkin. An unconditional lower bound on the time-approximation tradeoff of the minimum
spanning tree problem. In Proc. of the 36th ACM Symp. on Theory of Comput. (STOC 2004),
pages 331–340, 2004.
Michael Elkin. Computing almost shortest paths. ACM Trans. Algorithms, 1(2):283–323,
2005.
Michael Elkin. A near-optimal distributed fully dynamic algorithm for maintaining sparse
spanners. In Proceedings of the Twenty-Sixth Annual ACM Symposium on Principles of Dis-
tributed Computing, PODC 2007, Portland, Oregon, USA, August 12-15, 2007, pages 185–194,
2007.
[Elk07b] Michael Elkin. Streaming and fully dynamic centralized algorithms for constructing and main-
taining sparse spanners. In Automata, Languages and Programming, 34th International Col-
loquium, ICALP 2007, Wroclaw, Poland, July 9-13, 2007, Proceedings, pages 716–727, 2007.
[EN16a]
[EN16b]
[EN17]
[EP04]
[EP15]
[ES16]
[EZ06]
Michael Elkin and Ofer Neiman. Distributed strong diameter network decomposition.
In
Proceedings of the 2016 ACM Symposium on Principles of Distributed Computing, PODC
2016, Chicago, IL, USA, July 25-28, 2016, pages 211–216, 2016.
Michael Elkin and Ofer Neiman. Hopsets with constant hopbound, and applications to approx-
imate shortest paths. In IEEE 57th Annual Symposium on Foundations of Computer Science,
FOCS 2016, 9-11 October 2016, Hyatt Regency, New Brunswick, New Jersey, USA, pages
128–137, 2016.
Michael Elkin and Ofer Neiman. Efficient algorithms for constructing very sparse spanners and
emulators. In Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete
Algorithms, pages 652–669, 2017.
Michael Elkin and David Peleg. (1+epsilon, beta)-spanner constructions for general graphs.
SIAM J. Comput., 33(3):608–631, 2004.
Michael Elkin and Seth Pettie. A linear-size logarithmic stretch path-reporting distance oracle
for general graphs.
In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015, pages 805–821,
2015.
Michael Elkin and Shay Solomon. Fast constructions of lightweight spanners for general
graphs. ACM Trans. Algorithms, 12(3):29:1–29:21, 2016.
Michael Elkin and Jian Zhang. Efficient algorithms for constructing (1+epsilon, beta)-spanners
in the distributed and streaming models. Distributed Computing, 18(5):375–385, 2006.
29
[FKM+05]
J. Feigenbaum, S. Kannan, A. McGregor, S. Suri, and J. Zhang. Graph distances in the stream-
ing model: The value of space.
In Proc. of the ACM-SIAM Symp. on Discrete Algorithms,
pages 745–754, 2005.
[FS16]
Arnold Filtser and Shay Solomon. The greedy spanner is existentially optimal. In Proceedings
of the 2016 ACM Symposium on Principles of Distributed Computing, PODC 2016, Chicago,
IL, USA, July 25-28, 2016, pages 9–17, 2016.
[HZ96]
S. Halperin and U. Zwick. Linear time deterministic algorithm for computing spanners for
unweighted graphs, 1996. manuscript.
[KMP11]
Ioannis Koutis, Gary L. Miller, and Richard Peng. A nearly-m log n time solver for sdd
linear systems. In Proceedings of the 2011 IEEE 52Nd Annual Symposium on Foundations of
Computer Science, FOCS '11, pages 590–598, Washington, DC, USA, 2011. IEEE Computer
Society.
[LS93]
N. Linial and M. Saks. Decomposing graphs into regions of small diameter. Combinatorica,
13:441–454, 1993.
[MPVX15] Gary L. Miller, Richard Peng, Adrian Vladu, and Shen Chen Xu.
Improved parallel algo-
rithms for spanners and hopsets. In Proceedings of the 27th ACM Symposium on Parallelism in
Algorithms and Architectures, SPAA '15, pages 192–201, New York, NY, USA, 2015. ACM.
[MPX13]
Gary L. Miller, Richard Peng, and Shen Chen Xu. Parallel graph decompositions using random
shifts. In 25th ACM Symposium on Parallelism in Algorithms and Architectures, SPAA '13,
Montreal, QC, Canada - July 23 - 25, 2013, pages 196–203, 2013.
[OMSW10] James B. Orlin, Kamesh Madduri, K. Subramani, and M. Williamson. A faster algorithm
for the single source shortest path problem with few distinct positive lengths. J. of Discrete
Algorithms, 8:189–198, June 2010.
[Pel99]
David Peleg. Proximity-preserving labeling schemes and their applications.
In Graph-
Theoretic Concepts in Computer Science, 25th International Workshop, WG '99, Ascona,
Switzerland, June 17-19, 1999, Proceedings, pages 30–41, 1999.
[Pel00]
David Peleg. Distributed Computing: A Locality-sensitive Approach. Society for Industrial
and Applied Mathematics, Philadelphia, PA, USA, 2000.
[Pet09]
Seth Pettie. Low distortion spanners. ACM Transactions on Algorithms, 6(1), 2009.
[Pet10]
Seth Pettie. Distributed algorithms for ultrasparse spanners and linear size skeletons. Dis-
tributed Computing, 22(3):147–166, 2010.
[Pet16]
Seth Pettie. Personal communication, 2016.
[PS89]
D. Peleg and A. Schaffer. Graph spanners. J. Graph Theory, 13:99–116, 1989.
[PU89a]
D. Peleg and J. D. Ullman. An optimal synchronizer for the hypercube. SIAM J. on Comput.,
18:740–747, 1989.
30
[PU89b]
D. Peleg and E. Upfal. A tradeoff between size and efficiency for routing tables. J. of the ACM,
36:510–530, 1989.
[Rei93]
[RZ04]
John H. Reif. Synthesis of Parallel Algorithms. Morgan Kaufmann Publishers Inc., San Fran-
cisco, CA, USA, 1st edition, 1993.
Liam Roditty and Uri Zwick. On dynamic shortest paths problems. In Algorithms - ESA 2004,
12th Annual European Symposium, Bergen, Norway, September 14-17, 2004, Proceedings,
pages 580–591, 2004.
[TZ05]
Mikkel Thorup and Uri Zwick. Approximate distance oracles. J. ACM, 52(1):1–24, 2005.
[TZ06]
[Woo06]
M. Thorup and U. Zwick. Spanners and emulators with sublinear distance errors. In Proc. of
Symp. on Discr. Algorithms, pages 802–809, 2006.
David P. Woodruff. Lower bounds for additive spanners, emulators, and more. In Proceedings
of the 47th Annual IEEE Symposium on Foundations of Computer Science, FOCS '06, pages
389–398, Washington, DC, USA, 2006. IEEE Computer Society.
31
|
1312.0497 | 1 | 1312 | 2013-12-02T16:17:22 | Property-Testing in Sparse Directed Graphs: 3-Star-Freeness and Connectivity | [
"cs.DS"
] | We study property testing in directed graphs in the bounded degree model, where we assume that an algorithm may only query the outgoing edges of a vertex, a model proposed by Bender and Ron in 2002. As our first main result, we we present a property testing algorithm for strong connectivity in this model, having a query complexity of $\mathcal{O}(n^{1-\epsilon/(3+\alpha)})$ for arbitrary $\alpha>0$; it is based on a reduction to estimating the vertex indegree distribution. For subgraph-freeness we give a property testing algorithm with a query complexity of $\mathcal{O}(n^{1-1/k})$, where $k$ is the number of connected componentes in the queried subgraph which have no incoming edge. We furthermore take a look at the problem of testing whether a weakly connected graph contains vertices with a degree of least $3$, which can be viewed as testing for freeness of all orientations of $3$-stars; as our second main result, we show that this property can be tested with a query complexity of $\mathcal{O}(\sqrt{n})$ instead of, what would be expected, $\Omega(n^{2/3})$. | cs.DS | cs |
Property-Testing in Sparse Directed Graphs:
3-Star-Freeness and Connectivity
Frank Hellweg and Christian Sohler(cid:63)
Department of Computer Science, Technische Universität Dortmund
{frank.hellweg,christian.sohler}@tu-dortmund.de
Abstract. We study property testing in directed graphs in the bounded
degree model, where we assume that an algorithm may only query the
outgoing edges of a vertex, a model proposed by Bender and Ron [5].
As our first main result, we we present a property testing algorithm
for strong connectivity in this model, having a query complexity of
O(n1−/(3+α)) for arbitrary α > 0; it is based on a reduction to esti-
mating the vertex indegree distribution. For subgraph-freeness we give a
property testing algorithm with a query complexity of O(n1−1/k), where
k is the number of connected componentes in the queried subgraph which
have no incoming edge. We furthermore take a look at the problem of
testing whether a weakly connected graph contains vertices with a degree
of least 3, which can be viewed as testing for freeness of all orientations
of 3-stars; as our second main result, we show that this property can
√
be tested with a query complexity of O(
n) instead of, what would be
expected, Ω(n2/3).
1
Introduction
Property testing is a technique for solving decision problems that sacrifices some
accuracy for the benefit of a sublinear time complexity. The sacrifice of accuracy
is twofold: On the one hand, we allow property testing algorithms to accept a
small margin of inputs that do not have the queried property Π but are similar
to some inputs that have Π. More formally, for a proximity parameter < 1, we
say that an input is -far from having the property Π, if one must modify an
-fraction of the input's description in order to construct an input that has Π.
We only require a property testing algorithm for Π to give a reliable answer for
inputs that either have the property Π or are -far from it.
The second relaxation in accuracy is due to the randomized nature of prop-
erty testing algorithms: All those algorithms are Monte Carlo algorithms, which
means that they are allowed to have a small constant error probability.
The most important measure for the performance of a property testing algo-
rithm is its query complexity, which is the worst-case number of accesses to the
input that it needs for inputs of a given size. We aim for algorithms that have a
query complexity of o(n) or even O(1).
(cid:63) Research partly supported by DFG grant SO 514/3-2 and ERC Starting Grant
307696
In this paper we are particularly interested in property testing for sparse
directed graphs. Such graphs are assumed to be stored in adjacency list repre-
sentation and have both an in- and an outdegree of at most some constant d;
we require the adjacency lists to only contain the outgoing edges of a vertex, a
model which has been introduced in [5]. This is a quite natural model for directed
graphs: For example, the webgraph or, typically, graphs of social networks are
sparse graphs which have directed links; in particular, the incoming edges of a
vertex of these graphs might not be visible, for example in case of the incoming
links of a website during a web crawl. To gain this knowledge, basically the whole
graph has to be explored, and since these graphs are typically very large, this
may be inappropriate. Property testing algorithms for this graph model can be
useful to gain information about the structure of such graphs while exploring
only a small portion of it.
Property testing has been introduced by Rubinfeld and Sudan [16], while
Goldreich, Goldwasser, and Ron [10] have initiated the study of graph properties.
In this paper the authors introduced property testing in the dense graph model,
where graphs are assumed to be stored as an adjacency matrix. Furthermore,
Goldreich and Ron have introduced property testing in the sparse graph model
[11]. Since then a large variety of graph properties has been studied, including
[6,8,12,14] in the sparse graph model and [2] in the dense graph model. These
papers aim for identifying classes of testable properties: For the sparse graph
model, the above series of papers shows that every hyperfinite graph property
is testable, as well as every property in hyperfinite graphs; in the dense graph
model, a graph property is testable if and only if it can be reduced to a problem
of testing for satisfaction of one of a finite number of Szemerédi-Partitions.
Property testing in directed graphs can also be subdivided into property
testing in the dense graph and the sparse graph models. In the dense graph
model, Alon and Shapira have studied the property of subgraph-freeness [3].
Bender and Ron have studied the property of acyclicity in both the sparse graphs
and the dense graph model and the property of strong connectivity for sparse
graphs [5]. In the sparse graph model, they show that if a property testing
algorithm is only allowed to query the outgoing edges of a vertex, there are no
such algorithms with a query complexity of o(n1/3) for acyclicity and o(n1/2)
for strong connectivity, where n is the number of vertices of the input graph.
The assumption that only the outgoing edges of a vertex may be queried makes
testing strong connectivity much harder: As Bender and Ron show, there is a
one-sided error property testing algorithm with a query complexity of O(1/)
for strong connectivity without this constraint. Finally, Yoshida and Ito give
a constant-time property testing algorithm for k-edge connectivity of directed
graphs [17], which also relies on the visibility of incoming edges.
Our Results. In this paper we further study property testing in sparse directed
graphs where only the outgoing edges of a vertex may be queried. The first
property we study is subgraph-freeness, i.e., to test whether a graph H does not
occur as a subgraph of a graph G. Let k be the number of connected components
of H that have no incoming edge from another part of H: Then our algorithm
has a query complexity of O(n1−1/k).
A problem connected to subgraph-freeness is testing whether a weakly con-
nected graph is free of all orientations of 3-stars. Birthday-paradox type argu-
ments would imply a query complexity of Ω(n2/3) for this problem, but we can
give an algorithm with one of O(n1/2), which is the first main result of the pa-
per. This algorithm makes use of two facts: The first is that the above mentioned
class of forbidden subgraphs induces some strong properties for graphs that are
free of them; the second is that, when sampling edges, the probability of hitting
a vertex twice as the target vertex of two different edges is disproportionally high
if it has many incoming edges. This allows the algorithm to compute a ratio of
two estimators, which will be considerably larger if the input graph has many
vertices with a degree of at least 3.
The second main result of this paper is a property testing algorithm for strong
connectivity that achieves a query complexity of O(n1−/(3+α)) for arbitrary
α > 0. The algorithm is based on a reduction of the strong connectivity problem
to a problem of estimating the vertex indegrees of a graph: We show that it
is possible to define a locally computable partitioning of the input graph, such
that small connected components that have no incoming edges become their
own partitions; one can then construct a metagraph in which every partition
of the original graph becomes a vertex. If the input graph is far from strongly
connected, then the metagraph will contain many vertices with an indegree of
0, which can indeed be tested by statistics of the vertex indegrees. After first
publishing this result at ESA 2012 [13] we discovered that there already existed
a proof sketch for an algorithm similar to ours, which Oded Goldreich published
in the appendix of a survey article about graph property testing [9]. At the end
of the corresponding section we include a detailed discussion about similarities
and differences between Goldreich's and our algorithm.
2 Preliminaries
The graph model studied in this paper is the sparse graph model. If not explicitly
stated else, all graphs in this paper are directed graphs whose vertices have an
outdegree which is bounded by a constant d, as well as the indegree; this follows
the notion in [5]. The graphs are assumed to be stored as adjacency lists. We at
first define the notion of -farness:
Definition 1. Let G, H be directed graphs as above, both having n vertices. We
say that G is -far from H, if one has to change more than dn entries of the
adjacency lists of G to obtain a graph that is isomorphic to H.
Let Π be a graph property. We say that G is -far from Π, if it is -far from
any graph in Π.
Note that graphs as defined above have at most dn edges. This implies that
changing dn entries of adjacency lists means changing an -fraction of the
graph's description. We can now define the way property testing algorithms
get access to an input graph:
Definition 2. Let G = (V, E) be a directed graph with each vertex having an
outdegree of at most d ∈ N. We define fG : V ×N → V ∪{+} to be a function that
for querying f (v, i) returns the i-th neighbour of vertex v ∈ V in the adjacency
list representation of G, or +, if v has less than i neighbours.
Property testing algorithms get access to fG to gain knowledge about the input
graph. A call to fG takes O(1) time.
Definition 3. Let A be an algorithm that has parameters fG, and n. We define
the query complexity of A as the worst case number of calls to fG it performs
for any graph G with n vertices. A is a property testing algorithm for a graph
property Π, if:
1. The query complexity of A is sublinear in n.
2. A accepts every graph G ∈ Π with a probability of at least 2
3.
3. A rejects every graph G that is -far from Π with a probability of at least 2
3.
If A accepts every G ∈ Π with probability 1, we say it has 1-sided error, else we
say it has 2-sided error.
Finally, we define some graph properties that we will need throughout the rest
of this paper. Let G = (V, E) and H = (V (cid:48), E(cid:48)) be directed graphs.
We call H a subgraph of G, if there exists an injective mapping g : V (cid:48) → V
such that (g(u), g(v)) ∈ E for all (u, v) ∈ E(cid:48); we say that G is H-free, if H is
not a subgraph of G.
We call G (strongly) connected, if for all pairs of vertices u, v ∈ V there is a
directed path from u to v in G (we also say that v can be reached from u). We
call G weakly connected, if for all u, v ∈ V there is an undirected path between
u and v. U ⊆ V is a connected component of G, if the subgraph of G induced
by U is strongly connected and there is no set of vertices W ⊆ V − U such that
the subgraph of G induced by U ∪ W is strongly connected; i.e., U is maximal.
We have to distinguish between several types of connected components of
a graph that are witnesses to it being not strongly connected: Source and sink
components:
Definition 4 (source and sink components and dead ends). Let G =
(V, E) be a directed graph. A strongly connected compontent U ⊆ V is called
a source component, if there is no edge from V − U to U; U is called sink
component, if there is no vertex from U to V − U. We will call either of those
components dead ends.
We next define a special type of undirected graphs, k-stars. Such graphs
consist of a central vertex that is connected to k other vertices:
Definition 5 (k-star). An undirected graph H = (V, E) is called k-star, if the
following holds:
-- H has k + 1 vertices;
-- there is a vertex v ∈ V , such that for every vertex u ∈ V − {v} there is an
edge {u, v};
-- G does not contain any other edges.
In a directed setting, we will consider orientations of k-stars. If a k-star
orientation occurs as a subgraph of a graph G, then we call the central vertex
of this occurence k-star vertex. We call a vertex incoming k-star vertex if it has
at least k incoming edges and we call it outgoing k-star vertex if it has at least
k outgoing edges.
To simplify the analysis of our algorithms, we use a sampling technique that
deviates from the usual sampling of vertices (respectively, edges) with replace-
ment. Instead, we sample each vertex (edge) of the input graph with a certain
probability p. If a fixed number of vertex (edge) samples is exceeded, the algo-
rithm aborts by returning an arbitrary answer; the probability of this event will
be small. Note that in our analyses the case that the sample limit is exceeded
seperately is considered seperately from the rest of the particular analysis; after
that we use a union bound to bound the total error. Thus, in the rest of the
analyses we can assume that each vertex (edge) is independently sampled with
probability p.
3 Testing 3-Star-Freeness
We start by developing a simple property testing algorithm with one sided error
for a very basic graph property, subgraph freeeness.
Definition 6. Let G and H be directed graphs. We call G (induced) H-free if
H does not appear as an (induced) subgraph of G.
This algorithm has one-sided error and a query complexity of Θ(n1−1/k), if
the forbidden subgraph H has k source components. We will later use this al-
gorithm as a subroutine for a more complex algorithm that tests for freeness of
a certain class of subgraphs: The class of all orientations of 3-Stars. By running
the simple property tester for subgraph freeness for every possible 3-star orien-
tation, one would achieve a query complexity of O(n2/3). This is because there
is a 3-star orientation where the central vertex has 3 incoming edges and thus
this graph has 3 source components.
We give a more refined algorithm that requires the input graph to be weakly
connected and that uses statistical measures to distinguish between graphs that
have many occurences of such 3-star orientations with only incoming edges and
graphs that are 3-star-free; these statistical measures include 2-way collision-
statistics on the target vertices of samples edges, which can be done with O(n1/2)
queries. For all other types of 3-star orientations occuring in the input graph,
the simple subgraph freeness algorithm is called, each of these calls also having
a query complexity of O(n1/2). Hence, our algorithm has a query complexity
of O(n1/2) (considering as a constant) and thus breaks the trivial barrier of
TestSubgraphFreeness(G, H, )
Sample every vertex of G with probability p =(cid:0) 6m
if more than 4 ·(cid:0) 6m
(cid:1)1/k n1−1/k are sampledthen return true
(cid:1)1/k;
n
Foreach sampled vertex v do
Start a breadth first search at v having a maximum depth of m
If BFS explores an occurence of H in G or BFS completes
an occurence that was partly explored before then return false
return true
Algorithm 1: TestSubgraphFreeness
O(n2/3) for this problem. To our knowledge, this is the first algorithm that does
so for a nontrivial property of directed graphs.
At the end of the chapter we show that this algorithm is asymptotically
optimal for testing 3-star freeness: We show that any property testing algorithm
for this problem has a query complexity of Ω(n1/2), and that even Ω(n2/3)
queries are required if the input graph is allowed to be disconnected.
3.1 A Property Testing Algorithmus for H-Freeness
The algorithm TestSubgraphFreeness is given to directed bounded-degree
graphs G und H, where G has n vertices and H has m vertices. The algorithm is
also given a proximity parameter . Let k be the number of source components
of H.
In order to find an occurence of H in G, TestSubgraphFreiheit samples
every vertex of G with a probability of p = (cid:0) 6m
(cid:1)1/k, such that the expected
number of sampled vertices is np = O(−1m1/kn1−1/k). We will see that this is
sufficient for with high probability sampling a vertex in every source component
of at leastone occurence of H, if G is -far from H-free. The maximum depth of
m of the breadth first searches started in the sampled vertices ensures that this
suffices to completely explore that occurence of H.
We note that TestSubgraphFreeness can only reject the input, if it ac-
tually explores a occurence of H; hence, if G is H-free, it can never be rejected,
and thus TestSubgraphFreeness has a one-sided error. It remains to bound
the probability for inadvertantly accepting the input, if G is -far from H-free.
The following lemma bounds the number of vertex-disjoint occurences of H in
G from below in this case:
n
Lemma 1. Let G = (V, E) and H = (VH , EH ) be directed graphs with both
the vertex indegrees and outdegrees bounded by D ∈ N; let < 1 be a proximity
parameter and assume that G is -far from H-free. Let also V =: n and VH =:
m. Then G contains at least n
2m vertex-disjoint copies of H.
Proof. Let H consist of l weak connected components H1, . . . , Hl. We can assume
that each Hi contains at least two vertices, and, because it is weakly connected,
every vertex of Hi is incident to an edge. We assume that G contains less than
2m vertex-disjoint occurences of H lead this to a contradiction.
n
Let M be any inclusionwise maximal set of vertex-disjoint occurences of H
in G. Then there is at least one connected componentn Hi of H, for which all
the remaining occurences are not vertex-disjoint with the occurences of H in M.
Thus, at least one edge of each of these occurences of Hi is incident to a vertex
that belongs to some occurence of H in M.
Now consider the graph G(cid:48) that results from removing any edges that are
incident to vertices of occurences in M: G(cid:48) does not contain an occurence of Hi
any more, and thus does not contain a copy of H either. Thus, G(cid:48) is H-free. On
2m, each occurence of H has
the other hand, M has a cardinality of less than n
m vertices, and the maximum number of edges incident to any of these vertices
is 2D. Thus, the number of edges that are deleted from G to get G(cid:48) is less than
2Dm· n
2m = Dn. This is a contradiction to the assumption that G is -far from
H-free.
It remains to show that the size of the vertex sample in TestSubgraph-
Freeness is sufficiently large for, with high probability, exploring at least one
occurence of H, if G is -far from H-free. We at first assume that H is weakly
connected.
Theorem 1 Let G = (V, E) and H = (VH , EH ) be directed graphs with both the
vertex indegrees and outdegrees bounded by D ∈ N and let H be weakly connected;
let < 1 be a proximity parameter and let k be the number of source components
of H. Let also V =: n and VH =: m.
Then TestSubgraphFreeness(G, H, ) returns true, if G is H-free and
returns false with probability at least 2/3, is G is -far from H-free. TestSub-
graphFreeness has a query complexity of O(cid:16)
n1−1/k(cid:17)
.
Dm(cid:0) m
(cid:1)1/k
Proof. As discussed above, TestSubgraphFreeness can only reject if G con-
tains at least one occurence of H; thus, the algorithm always accepts if G is
H-free.
Now assume that G is -far from H-free and let M be a maximal set of
occurences of H in G; Lemma 1 ensures that there is a set M with a cardinality
of at least n/2m. Let Xi be the event that the i-th occurence in M (according
to an arbitrary but fixed ordering of the occurences in M) is explored completely
by TestSubgraphFreeness. Since the BFS traversals have a depth of m, this
happens if a vertex in each of the source components of this occurence is sampled;
since there are k of those, it holds Pr[Xi] ≥ pk for all i.
Since all the occurences in M are vertex-disjoint (and due to the sample
process), the events Xi are independent. Hence,
TestSubgraphFreenessAmplified(G, H, , p)
If not TestSubgraphFreeness(G, H, , p) then return false
(cid:109)
(cid:108) 1
p
For i → 1 to log3
return true
Algorithm 2: TestSubgraphFreenessAmplified
(cid:92)
1≤i≤M
=
¯Xi
Pr
=
(cid:89)
(1 − Pr[Xi]) ≤(cid:0)1 − pk(cid:1)n/2m
(cid:19)n/2m ≤ e−3 <
1
12
(cid:18)
1≤i≤M
1 − 6m
n
gives an upper bound for the probability that none of the occurences in M
is explored completely and thus an upper bound for the probability that none
of the occurences of H in G is identified.
It remains to bound the probability that the input is inadvertantly accepted
in the first line of TestSubgraphFreeness due to too many vertices being
sampled. The expected number of vertices sampled is np = (3m/)1/k · n1−1/k,
and, due to the Markov Inequality, the probability, that more than four times this
number of vertices is sampled is at most 1/4. Together with the above bound on
the probability of not exploring a single occurence of H, the union bound gives
a total probability of at most 1/4 + 1/12 = 1/3 for inatvertantly accepting the
input.
The number of vertices sampled by TestSubgraphFreeness is at most
n1−1/k of those a BFS traversal
of depth m is started. Since the vertex outdegree of G is bounded by D, the
number of edges queried in each of the BFS traversals is bounded by Dm, and
(cid:1)1/k
4 ·(cid:0) 3m
n1−1/k + 1, for at most 4 ·(cid:0) 3m
thus the total number of queries needed is O(cid:16)
Dm(cid:0) m
n1−1/k(cid:17)
(cid:1)1/k
(cid:1)1/k
.
Now we assume that H is disconnected, i.e. consists of l > 1 weakly connected
components H1, . . . , Hl. It suffices to identify occurences of each of the Hi inde-
pendently, all of those pairwise vertex-disjoint. In a maximal set of vertex-disjoint
occurences of H, as guaranteed by lemma 1, the occurences of the individual Hi
can be combined arbitrarily to get an occurence of H.
Thus, TestSubgraphFreeness can be run individually for each of the Hi.
For this purpose, we define the algorithm TestSubgraphFreenessAmplified,
which uses probability amplification to guarantee a better success probability p
instead of 2/3.
Lemma 2. Let G = (V, E) and H = (VH , EH ) be directed graphs with both
the vertex indegrees and outdegrees bounded by D ∈ N; let < 1 be a prox-
imity parameter and p ≤ 1
3. Then TestSubgraphFreenessAmplified re-
turns true if G is H-free and false with probability at least 1 − p if G is -far
from H-free. TestSubgraphFreenessAmplified has a query complexity of
O(cid:16)
p · Dm(cid:0) m
(cid:1)1/k
n1−1/k(cid:17)
.
log 1
Proof. In case G is H-free, the correctness follows directly from the correctness of
TestSubgraphFreeness. Now assume that G is -far from H-free. Theorem
1 guarantees that every call of TestSubgraphFreeness returns false with
probability at least 2/3. Thus, the probability that none of the calls returns false
(and thus TestSubgraphFreenessAmplified inadvertantly returns true) is
at most (1/3)log3(cid:100)1/p(cid:101) ≤ 3− log3 1/p = p.
times in TestSubgraphFreenessAmplified, the query complexity of the lat-
Since TestSubgraphFreeness is called at most log3
ter algorithm is O(cid:16)
p · Dm(cid:0) m
(cid:1)1/k
n1−1/k(cid:17)
.
log 1
(cid:109)
(cid:108) 1
p
= O(cid:16)
log 1
p
(cid:17)
For testing H-freeness for disconnected H, we run TestSubgraphFree-
nessAmplified individually for each of the l connected components of H while
3l. Assume that G is -far from H-free: Then the
setting the parameter p to 1
probability that at least one of the l calls to TestSubgraphFreenessAmpli-
fied fails to identify a complete occurence of the corresponding component is at
most l · p = 1/3 due to the union bound. Hence we get the following corollary:
Corollary 1. H = (VH , EH ) be a directed graph with m vertices and with both
the vertex indegrees and outdegrees bounded by D ∈ N; let H consist of l > 1
connected components. Let < 1 be a proximity parameter and let V = n
and VH = m. Let kmax the maximum number of source components of one
of the connected components of H and let kmin the minimum number of source
components of one of them.
Then there is a property testing algorithm for H-freeness, which in Graphs
with n vertices and both vertex indegree and outdegree bounded by D has a query
l log l · Dm(cid:0) m
(cid:1)1/kmin n1−1/kmax
complexity of O(cid:16)
(cid:17)
.
3.2 A Property Testing Algorithm for 3-Star-Freeness
We will now consider the problem of testing for a certain class of forbidden sub-
graphs, the class of all orientations of 3-stars. The algorithm that we introduce
distinguishes between directed graphs that contain many occurences of 3-stars,
regardless of how the edges of these occurences are oriented. We make two major
additional assumptions: The first is that the input graph is weakly connected.
The second assumption is that the input graph does not contain any double
edges, i.e. pairs of edges in both directions between two vertices; thus, the undi-
rected degree of a vertex is equal to its number of adjacent vertices. The latter
assumption is only made for simplicity and can be dropped using a simple re-
duction, as we will see later. The former assumption is, however, necessary, since
without this assumption testing 3-star-freeness requires Ω(n2/3) queries.
We will additionally assume that the undirected degree of each vertex of the
input graph is at most D ∈ N≥3. At last, we can assume n = ω( D4
3 ): Else an
algorithm could read all the input graph's edges with at most Dn = O( D3
3/2 )
queries and thus solve the problem deterministically.
In the following we will simply use the term k-star for any k-star orienta-
tion in directed graph contexts. We have to distinguish several types of k-stars,
depending on their number of incoming and outgoing edges:
Definition 7. Let G be a directed k-star and let 2 ≤ m ≤ k. We call G incoming
m-star, if for at least m of the edges of G the central vertex of G is the target
vertex. We call G outgoing m-star, if for at least m of the edges of G the central
vertex of G is the source vertex. If G is an incoming respectively outgoing m-star,
we call the central vertex of G outgoing respectively incoming k-star vertex.
As noted above, a simple algorithm for testing 3-star freeness works as follows:
Run the algorithm TestSubgraphFreenessAmplified from the previous sec-
tion for each of the 4 (neglecting isomorphism) possible orientations of 3-stars
and for a proximity parameter of /4; if one of the calls to TestSubgraphFree-
nessAmplified rejects, return false; else return true. This algorithm has a query
complexity of Θ(n2/3), since an incoming 3-star without outgoing edges has 3
source components. The algorithm that is introduced in this chapter will instead
have a query complexity of O(n1/2), but will in exchange have a two-sided error.
Since at this point the input graphs are restricted to weakly connected di-
rected graphs without double edges, we can formulate the problem differently:
We want to distinguish directed graphs that are orientations of a single circle
or a line from those that are -far from doing so. This is not trivial, since due
to the restriction that only outugoing edges of a vertex are possible to query,
the portion of the graph explored by a breadth first search may be very small,
particularly for vertices that have only incoming edges.
The algorithm makes use of a certain property of directed graphs that contain
many occurences of incoming 3-stars without outgoing edges: In such graphs, the
number of edges that are part of incoming 2-stars is significantly larger than the
number of edges that are part of outgoing two-stars (note that every incoming
3-star vertex is also an incoming 2-star vertex). Moreover, the probability of
sampling two incloming edges of a vertex grows quadratically in the total number
of its incoming edges. Thus, in a graph that contains many incoming 3-stars, the
number of incoming 2-stars found by collisions statistics on the target vertices of
edges will ne disproportionately high in comparision to the number of incoming
2-stars in this graph with high probability; on the other hand, in a graph is
3-star-free, these values will be roughly the same with high probability, as we
will see. Such collision statistics only require O(n1/2) edge samples.
For testing for any 3-stars that have incoming edges, we can rely on calling
TestSubgraphFreeness as sketched above. Since we now only have to call this
algorithm for orientations with at most 2 incoming edges, the query complexity
for all these calls is O(n1/2). Thus, we get a total query complexity of O(n1/2).
For formalizing these ideas, we state the algorithm Test3StarFreeness.
Test3StarFreeness(n, G, )
if EstimateEdgeCount(n, G,
Foreach 3-star orientation H that has at least one outgoing edge
16 then return false
16 )> n + n
if not TestSubgraphFreenessAmplified(G, H,
6 )
192D , 1
times the number of outgoing 2-star vertices in {v1, . . . , v3.1}
then return false
s3.1
12
vertices v1, . . . , v3.1 of G u.i.d.
sample s3.1 = 48
k ← n
k ← k + n
3/2√
sample each edge of G with probability p = 128D
n
if more than s3.2 = 2048·D
c ← 1
if c < 3
2 n then return true
if r := c
> 1 +
k
else return true
24 then return false
√
n
3/2
p2 times the number of collisions on target vertices of sampled edges
are sampled then return false
Algorithm 3: Test3StarFreeness
We will at first introduce a couple of Lemmas that state structural properties
of weakly connected directed graphs. All these Lemmas are connected to the
notion of the balance of a directed graph:
Definition 8. Let G = (V, E) be a directed graph. The balance B(G) of G is
the difference between the number of outgoing 2-star vertices and the number of
incoming 2-star vertices in G, i.e.:
B(G) = #{v ∈ V degout(v) ≥ 2} − #{v ∈ V degin(v) ≥ 2}
We will show several upper bounds on the balance of directed graphs, de-
pending on their structure. We will start with graphs that are orientations of an
undirected circle graph.
Lemma 3. Let G = (V, E) be an orientation of an undirected circle graph.
Then, B(G) = 0.
Proof. Since each vertex in G has an undirected degree of two, incoming 2-star
vertices do not have outgoing edges and outgoing 2-star vertices do not have
incoming edges. Each vertex that is neither of them has exactly one incoming
edge and one outgoint edge. Furthermore, the total number of incoming edges
in G equals the total number of outgoing edges, and hence, by the above con-
siderations, for each vertex that has two outgoing edges, there must be a vertex
that has two incoming edges and vice versa. Thus the number of incoming 2-star
vertices in G equals the number of outgoint 2-star vertices, i.e. B(G) = 0.
The following proofs make use of the techniques of contracting and re-expanding
and of deleting and re-inserting edges of graphs. Observe that inserting an edge
into a graph can only change its balance by one: Only the target vertex of the
(a)
(b)
e
Fig. 1: Orientations of undirected graphs: (a) Circle graph; (b) Tree; a feasible
edge e for choosing in the inductive step of the proof of Lemma 5 is drawn
dashed. In both cases incoming 2-star vertices are painted in red and outgoing
2-star vertices are painted in blue.
edge can become a new incoming 2-star, and only its source vertex can become
a new outgoing 2-star. Of course the same holds true in reverse direction for
deleting an edge.
We continue by showing that orientations of line-shaped graphs have a bal-
ance of at most 1, and, if the balance of such a graph is not 0, both edges at the
ends of the graph have the same orientation. A line-shaped graph is a tree that
has only two leafs.
Lemma 4. Let G = (V, E) be an orientation of a line-shaped graph. Then,
B(G) ≤ 1. If B(G) (cid:54)= 0, one of the following statements holds:
-- The number of outgoing 2-star vertices in G exceeds the number of incoming
2-star vertices by 1 and both leafs of G have incoming edges.
-- The number of incoming 2-star vertices in G exceeds the number of outgoing
2-star vertices by 1 and both leafs of G have outgoing edges.
Proof. At first we prove B(G) ≤ 1. Consider an arbitratily oriented edge e to
be inserted between the two leafs of G. The resulting graph G(cid:48) is an orientation
of a circle and, according to lemma 3, has a balance of 0. Since by re-deleting e
from G(cid:48) the balance can change by at most 1, we have B(G) ≤ 1.
Now assume that B(G) = 1. Then the insertion of e must have created a new
2-star vertex of the type that appears less in G and cannot have created one of
the other type, since G(cid:48) has a balance of 0. Thus, if G has more incoming 2-star
vertices than outgoing 2-star vertices, both leafs have an outgoing edge, so that
the insertion of e creates a new outgoing 2-star vertex at one of them and does
not create a new 2-star vertex at the other. Equvialently, if G has more outgoing
2-star vertices than incoming 2-star vertices, both leafs have an incoming edge.
We will now show that the maximum balance of a tree depends on its number
of leafs.
Lemma 5. Let G = (V, E) be an orientation of a tree with k leafs. Then B(G) ≤
k − 1.
Proof. We prove the lemma by induction over the number k of leafs of the tree.
For the base case let G be an orientation of a tree with k = 2 leafs. Then, by
Lemma 4, we have B(G) ≤ 1 = k − 1.
For the induction step we assume that G is an orientation of a tree with
k > 2 leafs and that every orientations of a tree with at most m < k leafs has a
balance of at most m − 1.
Since G has at least three leaf vertices, there exists at least one 3-star vertex
v in G. If we delete an edge e incident to v, G decomposes into two subgraphs
G1 and G2 which are not connected to each other. We can choose v and e in
such a way that G2 is a single vertex or an orientation of a line-shaped graph
by starting at an arbitrary leaf of G and traversing along the edges of G until
the first 3-star vertex is reached. Let v be a vertex found in such a way and let
e be the last edge visited when traversing from the corresponding leaf to v (see
figure 1).
G1 has k − 1 leafs and, by the induction hypothesis, is guaranteed to have a
balance of at most k − 2; G2 has a balance of 0, if it consists of a single vertex,
and, by the induction hypothesis, a balance of at most 1 elsewise.
We now re-insert e: We have B(G) ≤ B(G1) + B(G2) + 1. If B(G2) = 0 this
immediately yields B(G) ≤ k−1; if the majority of 2-stars in G1 is incoming and
the majority of 2-stars in G2 is outgoing (or vice versa), B(G) ≤ k − 1 follows
immediately, too.
Now assume that neither of it is the case, i.e. both G1 and G2 the same type
of 2-stars has the majority and the balance G2 is 1; without loss of generality,
assume that the majority of 2-stars in both graphs is incoming. Let u be the
vertex of G2 that e is connected to. Then, by Lemma 4, u has an outgoing edge.
Thus, if e = (u, v), a new outgoing 2-star at u is created, and at most one new
incoming 2-star at v is created. If e = (v, u), no incoming 2-star at u is created,
since the other edge incident to u is incoming, and, since e is an outgoing edge at
v, no additional incoming 2-star at v is created. Thus, the number of incoming
2-star created by re-inserting e is bounded by the number of outgoing 2-stars
created, and thus we have B(G) ≤ B(G1) + B(G2) ≤ k − 2 + 1 = k − 1.
Finally we will derive an upper bound on the balance of arbitrary weakly
connected graphs. We will need the following observation for this:
Observation 1 Let G = (V, E) be an orientation of a tree and let C3 ⊆ V be
deg(v)
the set of 3-star vertices of G. Then G has exactly 2 − 2C3 +(cid:80)
leafs.
v∈C3
Proof. We show the observation by induction over the number of 3-star vertices
in G. For the base case observe that if C3 = 0, G is an orientation of a line-
shaped graph and thus has 2 leafs; hence the observation holds. For C3 = 1,
assume that v is the only 3-star vertex in G and v has an undirected degree of
k. Then, G consists of orientations of k lines, which are connected in v; G thus
has k leafs and we have 2 − 2 · C3 +(cid:80)
v∈C3
= k; hence the observation holds if
G has at most one 3-star vertex.
For the induction step we assume that G is a graph with C3 = i ≥ 2 3-
star vertices and that the observation holds for all graphs G(cid:48) that have j < i
3-star vertices. Since G is a tree and contains at least 2 3-star vertices, there
exists an edge e which is incident to a 3-star vertex v and lies on a direct path
from v to another 3-star vertex u, but all the other edges incident to v lie on
paths to leafs of G. We delete all the vertices and edges from G that form paths
from v to leafs, but we sustain v and e. By the choice of v, v is a leaf in the
3 = C3\{v}. Thus,
resulting graphs G(cid:48), and the set of 3-star vertices in G(cid:48) is C(cid:48)
G(cid:48) has less than less than i 3-stars and by the induction hypothesis has at most
2 − 2C(cid:48)
deg(v) leafs; furthermore, the degree of every vertex v ∈ C(cid:48)
in G(cid:48) equals its degree in G.
3 +(cid:80)
v∈C(cid:48)
Now let G be the subgraph of G induced by the deleted vertices and v: Since
by construction G can have no 3-star vertices other than v, G contains at most
a single 3-star. Thus, by the base case, the number of leafs of G equals the
undirected degree of v in G; let k be this number. Now we join G(cid:48) and G: v now
has a degree of k + 1, and in the resulting graph G, v is not a leaf any more
compared to G(cid:48), but G contains k additional leafs from G; thus, the number of
leafs in G is larger than that of G(cid:48) by deg(v) − 2 and hence equals
3
3
(cid:88)
deg(v)
3 ∪ {v} +
(cid:88)
v∈C3
v∈C(cid:48)
3∪{v}
deg(v).
= 2 − 2C3 +
deg(v) − 2 + 2 − 2C(cid:48)
3 +
deg(v) = 2 − 2C(cid:48)
(cid:88)
v∈C(cid:48)
3
Now we can proceed by proving an upper bound on the balance of an arbi-
trary weakly connected graph:
Lemma 6. Let G = (V, E) be a weakly connected graph that has (1 + δ)V
edges, δ > 0. Also let C3 ⊆ V be the set of 3-star vertices of G. Then G has a
balance of at most 2 + δV − 2C3 +(cid:80)
by its number of leafs, which, by observation 1, is 2 − 2C3 +(cid:80)
hence it holds B(G(cid:48)) ≤ 1 − 2C3 +(cid:80)
2 + δV − 2C3 +(cid:80)
Proof. Delete δV + 1 edges from G in such a way that G remains weakly
connected; Since the resulting graph G(cid:48) is weakly connected and has V − 1
edges, G(cid:48) is an orientation of a tree. By Lemma 5 the balance of G(cid:48) is bounded
deg(v);
By re-inserting all the deleted edges into G(cid:48), we get G. Since every re-inserted
edge can only alter the balance by 1, we have B(G) ≤ B(G1) + δV + 1 ≤
deg(v).
deg(v).
deg(v).
v∈C3
v∈C3
v∈C3
v∈C3
The following simple Lemma states that every weakly connected graph whose
number of edges is larger than its number of vertices, contains at least one 3-star
vertices. Hence, a graph can be rejected if it contains too many edges.
EstimateEdgeCount(n, G, )
m ← 0
Sample s4 = 2D
for i ← 1 to s4 do
return m
pairs (v1, k1), . . . , (vs4 , ks4 ) of a vertex
and a number of an adjacency list slot, i.e. (vi, ki) ∈ V × {1 . . . D}
if vi has a ki-th edge then m ← m + Dn
s4
Algorithm 4: EstimateEdgeCount
Lemma 7. Let G = (V, E) be a weakly connected directed graph with n vertices
and m > n edges. Then, G contains at least 1 3-star vertex.
Proof. Since every edge of G has two incident vertices, the sum of all undirected
vertex degrees in G is 2m. Hence, the average undirected vertex degree in G is
2m/n > 2; thus, there exists at least one vertex that has an undirected degree
of more than 2 and therefore is a 3-star vertex (note that, by the assumption
from the beginning of the section, G does not contain any double-edges).
We will use the fact stated by Lemma 7 to reject every input graph, whose
edge count exceeds a certain number. To estimate the edge count of a given
directed graph, we use the algorithm EstimateEdgeCount. The algorithm
samples O(D/) pairs (v, i) ∈ V (G) × {1, . . . , D} of a vertex and the number of
a slot of its adjacency list uniformly and independently distributed at random; it
then checks for all those pairs whether the corresponding edge exists and extrap-
olates the fraction of edges found. Note that a similar technique has already been
used in a property testing algorithm for acyclicity in bounded-degree graphs by
Goldreich and Ron [11].
Lemma 8. Let G = (V, E) be a directed graph with n vertices and m edges and
let the undirected degree of every vertex of G be bounded by D ∈ N; let < 1
be a proximity parameter. Then EstimateEdgeCount returns a value m that,
with probability at least 1 − 2e−4, satisfies
m − n ≤ m ≤ m + n.
The running time of EstimateEdgeCount is O(D/).
neighbor, i = 1, . . . , s4; let X =(cid:80)
Proof. Let Xi be an indicator random variable for the event that vi has a ki-th
Xi. Each of the m edges of G is stored
in exactly one adjacency list slot as an outgoing edge and there are Dn adjacency
list slots in total. Since (vi, ki) is chosen u.i.d. at random among all pairs of a
vertex and an adjacency list slot, it holds
1≤i≤s4
E[Xi] = Pr[Xi = 1] =
m
Dn
and, by the linearity of expectation,
E[ m] = E
= E
(cid:21)
(cid:20) Dn
· (cid:88)
· X
s4
1≤i≤s4
Dn
s4
m
Dn
= m.
· (cid:88)
· (cid:88)
1≤i≤s4
Xi
1≤i≤s4
=
Dn
s4
E[Xi] =
Dn
s4
It remains to show that the probability that m differs from m by more than
n is at most 2e−4. For this purpose, we use an additive Chernoff Bound:
Pr [ m − E[ m] > n] = Pr
(cid:104)X − E[X] >
(cid:105) ≤ 2e−2s2
s4
D
4/D2
= 2e−4.
The running time of EstimateEdgeCount directly follows from the number
of samples taken.
Analogously to the above proof one can show that the value k that is com-
puted in algorithm Teste3SternFreiheit is a good estimate of the number of
outgoing 2-stars in the input graph:
Lemma 9. Let G = (V, E) be a directed graph with n vertices and a bounded
undirected vertex degree of D ∈ N. Let be a proximity parameter and assume
that G contains k ≤ n outgoing 2-star vertices. Then, for the value k computed
by Test3StarFreeness(n, G, ) it holds
k ≤ k ≤ k +
6
n
with a probability of at least 1 − 2e−16.
Proof. We consider the vertex sample drawn in the fifth line of the algorithm:
Let Xi be an indicator random variable for the event that the vertex vi is an
outgoing 2-star vertex. Both the probability for this event and the expected value
X. By the linearity
of Xi are k/n. Let X =(cid:80)
Xi; then, k =
1≤i≤s3.1
of expectation we get
E
(cid:105)
(cid:104)k
(cid:104)(cid:12)(cid:12)(cid:12)k − E[k]
Pr
k
n
12 n + n
s3.1
(cid:88)
= k +
(cid:105) ≤ 2e−2s2
1≤i≤s3.1
s3.1
12
=
12
n +
n
s3.1
E[X] =
12
n +
n
s3.1
12
n.
To bound the probability of k deviating from its expectation by more than
we again use an additive Chernoff Bound:
12 n
(cid:12)(cid:12)(cid:12) > n/12
(cid:105)
(cid:104)X − E[X] >
= Pr
3.1/144 = 2e−16.
We can now start with proving the correctness of Test3StarFreeness. We
divide the proof into four parts: The first part handles the case that the input
graph G (having n vertices) is 3-star free, the remaining three parts the case that
G ist -far from 3-star free. In this case we consider the following three subcases:
-- G has more than n +
-- G has at most n +
8 n edges;
have at least one outgoing edge;
8 n edges and contains at least
16D n 3-star vertices that
-- G has at most n +
16D n 3-star vertices that
have at least one outgoing edge (and thus there are many 3-star vertices that
have only incoming edges).
8 n edges and contains less than
For every possible input one of the above cases holds. We start our analysis
with the three subcases where we assume G to be -far from 3-star free. In case
8 n edges, Test3StarFreeness will reject in the first line
G has more than n +
with probability at least 2e−4; this follows directly from Lemma 8:
Lemma 10. Let G = (V, E) be a weakly connected directed graph with n vertices
and a bounded undirected vertex degree of D ∈ N. Let < 1 be a proximity param-
eter and assume that G has m > n+
returns false with a probability of at least 1 − 2e−4 false.
The next case we consider is that G has at most n +
8 n edges and contains
at least
16D n 3-star vertices that have an outgoing edge. Each of these 3-star
vertices is the central vertex of an occurence of a 3-star orientation in G, that
has at least one incoming edge; there are (not considering isomorphism) 3 such
orientations, and hence one of them has at least
8 n edges. Then, Test3StarFreeness(n, G, )
48D n occurences in G.
96D n ≥
Since deleting an edge from G can at most remove two of these occurences,
at least
192D n + 1 edge modifications in G are necessary to remove
all the occurences of this orientation1, and thus G is
192D -far from freeness
of this 3-star orientation. Hence and by Lemma 2, the corresponding call to
TestSubgraphFreenessAmplified in the third line of Test3StarFreeness
returns false with a probability of at least 5/6 false. These considerations yield
the following Lemma:
Lemma 11. Let G = (V, E) be a directed graph with n vertices and a bounded
undirected vertex degree of D ∈ N. Let < 1 be a proximity parameter and
assume that G has m ≤ n +
16D n 3-star vertices
that have at least one outgoing edge each. Then, Test3StarFreeness(n, G, )
returns false with a probabilty of at least 5/6.
8 n edges and contains at least
For the case of the input graph G being -far from 3-star free, it remains to
show that Test3StarFreeness rejects with high probability if G contains less
than
16D n 3-stars that have outgoing edges.
Lemma 12. Let G = (V, E) be a weakly connected directed graph with n vertices
and a bounded undirected vertex degree of D ∈ N. Let < 1 be a proximity
parameter and assume that G is -far from 3-star free, has m ≤ n +
8 n edges
and contains less than
16D n 3-star vertices that have at least one outgoing edge
each. Then, Test3StarFreeness(n, G, ) returns false with a probabilty of at
least 5/6.
1 n ≥ 192D
can be assumed since we have assumed n = ω
(cid:16) D4
above.
(cid:17)
2
e1
v
e2
Fig. 2: Two-way collision at an incoming 3-star vertex v; the sampled edges e1
and e2 are highlighted and the vertex colouring follows that of figure 1.
24 holds in the
Proof. We will show that under the given assumptions r > 1 +
first but las line of Test3StarFreeness with a probability of at least 1 − e−6;
hence the algorithm rejects the input with high probability. The basic idea of
this proof is the following: By Lemma 6, the maximum Balance of G has a linear
dependence on the number of its edges that are incident to 3-star vertices; the
probability of sampling at least two of the incoming edges of a 3-star vertex,
however, grows quadratically in the number of incoming edges that this vertex
has. Since there are (as we will show) at least n 3-star vertices in G and almost
all of them do not have outgoing edges, there are nearly n 3-star vertices that
have at least 3 incoming edges each. Hence the probability of sampling a collision
at one of them is relatively large: Counting the number of collisions and grossing
up overestimates the number of outgoing 2-star vertices (to which the outgoing
3-star vertices belong) by more than the outgoing 2-star vertices can outnumber
the incoming 2-star vertices due to the balance of G. Thus, the expected value
of c will be much larger than that of k and the expected value of c will be much
larger than 1 hence.
We will need some additional definitions to concretize this idea. At fist, let
C2 be the set of incoming 2-stars vertices that have exactly 2 incoming edges
3 ⊆ C3 the set of
and let C3 be the set of incoming 3-stars in G. We denote by C(cid:48)
incoming 3-stars that do not have an outgoing edge and by C3 ⊇ C3 the set of
all 3-star vertices in G. For the cardinalities of these sets we define c2 := C2,
3 and c3 := C3. Then it holds C := C2 + C3 for the set of
c3 := C3, c(cid:48)
all incoming 2-star vertices C and c := c2 + c3 for its cardinality. Analogously
to the last equation, we also say c = c2 + c3, where c2 denotes the contribution
of collisions on vertices in C2 and c3 that of collisions on vertices in C3.
3 := C(cid:48)
G contains at least n 3-star vertices: If the number of 3-star vertices was
smaller, there would be less than Dn edges adjacent to them, and deleting
all of these edges would remove every 3-star vertex from G; this would be a
contradiction to the assumption that G is -far from 3-star free.
Since by the assumption of the lemma less than
have incoming edges and D ≥ 3, there are at least c(cid:48)
3 > (cid:0)1 − 1
16D n of these 3-star vertices
48 n
(cid:1) n ≥ 47
16D
3-star vertices without outgoing edges. Since C(cid:48)
We define
c3 ≥ c(cid:48)
3 >
l :=
(deg(v) − 2),
(cid:19)
3 is a subset of C3, it also holds
n ≥ 47
48
n.
(1)
1 − 1
16D
(cid:18)
(cid:88)
v∈C3
and in the following we will derive an upper bound for k and a lower bound for
c; both bounds will relate on l, which will help us bounding r = c
from below.
k
Furthermore, it holds
l ≥ c3, ,
(2)
since every vertex in C3 has an undirected degree of at least 3.
We start by proving an upper bound for k. Since G has at most n +
by the assumption of the Lemma, the balance of G is at most
8 n edges
B(G) ≤
8
n + 2 +
=
<
8
8
n + 2 +
n + l +
(cid:88)
(cid:88)
v∈ C3
v∈C3
16
(deg(v) − 2)
(deg(v) − 2) +
(cid:88)
v∈ C3\C3
(deg(v) − 2)
n < l +
4
n,
sumption of the Lemma, and thus there are at most D(cid:0)
by Lemma 6 and because C3\C3 only contains 3-star vertices that have at least
16D n − 1 such 3-star vertices by the as-
one outgoing edge; there are at most
16 n − 2
Since c is the number of incoming 2-star vertices and the balance of G is at
4 n.
4 n, the number of outgoing 2-star vertices in G is ar most c + l +
most l +
By Lemma 9 we get that, with probability at least 1 − 2e−16, it holds
16D n − 1(cid:1) <
edges that are incident to one of them.
k ≤ c + l +
1
2
the second inequality follows from the inequality 5
12 n < 5
we get from (1).
n < c2 + c3 + l +
c3 ≤ c2 +
n +
4
6
12 · 48
l;
5
2
47 c3 < 1
2 c3, that
We will now bound c from above. Let Xv be an indicator random variable
for the event that for a vertex v at least 2 incoming edges are contained in the
edge sample of Test3StarFreeness, i.e., we have a (2-way) collision at v. Let
X (2) =(cid:80)
v∈C2
Xv und X (3) =(cid:80)
v∈C3
Xv. It holds
(cid:18)degin(v)
(cid:19)
2
(cid:18)degin(v)
(cid:19)
2
≥ αp2
E[Xv] = Pr[Xv = 1] = p2(1 − p)degin(v)−2
48 due to the minimum number of vertices we have assumed for G and by
n ≥ 47
for α := 47
the Bernoulli Inequality: (1− p)degin(v)−2 ≥ (1− p)D ≥ 1− pD = 1− D2
√
for n ≥ 2304D4
3/2
48
.
3
Hence, for the contribution of collisions on incoming 2-star vertices to c,
(cid:88)
v∈C3
=
1
p2
X (3)(cid:105)
(cid:104)
(cid:88)
1
p2 E
α
2
v∈C(cid:48)
3
deg(v)2 − deg(v) ≥ 3α
E[c3] =
=
E[Xv] ≥ α
(cid:18)degin(v)
(cid:19)
(cid:88)
(cid:18)deg(v)
(cid:19)
v∈C(cid:48)
3
2
≥ α
(cid:88)
(cid:88)
v∈C(cid:48)
3
2
v∈C3
deg(v) − 2
due to the linearity of expectation and since for every 3-star vertex it holds
deg(v)2 − deg(v) ≥ 6· (deg(v)− 2). We want to relate the above sum to l, but in
this sum only the vertices in C(cid:48)
3 are considered. However, by the assumption of
16D n vertices in C3\C(cid:48)
3 -- all of these vertices are
the Lemma there are at most
3-star vertices that have at least one outgoing edge. The sum over the degrees
47 l since it holds
of these vertices is at most 1
16 n ≤ 3
1
l ≥ c3 ≥ 47
48
n
(3)
by the inequalities (2) and (1). Hence we can conclude
(cid:88)
v∈C(cid:48)
3
deg(v) − 2 >
deg(v) − 2 − (cid:88)
(cid:88)
v∈C3
v∈C3\C(cid:48)
3
deg(v) ≥ l − 3
47
l =
44
47
l.
We can use this in the above bound for E[c3] and get E[c3] > 3αl = 11
4 ·
4 l ≥ 11
48 n since α = 47
48 and due to inequality (3). Now we can apply a multiplicative
47
Chernoff Bound, which yields an upper bound for the probability that c3 is
(cid:19)
smaller than this value by a factor of more than (1 − /44):
Pr [c3 < (1 − /44)E[c3]] = Pr
X (3) < (1 − /44)p2E [c3]
(cid:105) ≤ exp
(cid:19)
≤ exp
(cid:18)
(cid:18)
− 2p2
2 · 442 E[c3]
− 8 · 47D2
(cid:19)
33
< e−100.
≤ exp
− 12822D2
2 · 4423n
· 11
4
· 47
48
n
(cid:104)
(cid:18)
Analogously we can conclude Pr [c3 < (1 − /24)E[c3]] < e−600. In both cases we
use the fact D ≥ 3.
For estimating c it remains to calculate the expected number of collisions on
vertices in C2. Since these vertices have exactly 2 incoming edges, it holds
E[c2] =
E[Xv] =
p2 = c2.
v∈C2
Now we distinguish between two cases, c2 ≥ 1
v∈C2
16 l. In the former
case, c2 is relatively large, and we can use a Chernoff Bound to show that c2 is
a good estimate for c2; then we can argue that the vertices in C3 form at least
roughly an -fraction of the vertices in C and thus c = c2 + c3 is significantly
larger than c.
16 l and c2 < 1
(cid:88)
1
p2
(cid:88)
1
p2
of the vertices in C2 to c.
In the latter case, i.e., if c2 is small, we will simply neglect the contribution
We start with the first case, i.e., we have c2 ≥ 1
16·48 n by inequality
16 l ≥ 47
(3), and by applying a multiplicative Chernoff Bound we conclude
Pr [c2 < (1 − /24)E[c2]] = Pr
X (2) < (1 − /24)E
≤ exp
− 12822D2
2 · 2423n
· 47n
16 · 48
(cid:104)
(cid:18)
(cid:18)
(cid:104)
X (2)(cid:105)(cid:105) ≤ exp
(cid:19)
< e−7.
(cid:19)
− 2p2
2 · 242 c2
By adding the lower bounds for c3 and c2, we gain a lower bound for c, i.e.,
≥
(cid:1)
r =
=
(cid:16)
(cid:16)
c2 + c3
(cid:17)(cid:18)
with high probability it holds
(cid:1)(cid:0)c2 + 11
(cid:0)1 −
(cid:18) 5
(cid:19)(cid:19)
(cid:1) n, since G contains at least(cid:0)1 − 1
Furthermore we have c2 ≤ n−(cid:0)1 − 1
3-star vertices that do not have an incoming edge, and c3 ≥(cid:0)1 − 1
c
k
1 −
24
24
c2 + 5
2 c3
c2 + 5
1
4 c3
c2 + 5
1 −
24
(cid:17)(cid:18)
(cid:19)
1
10
4 c3
2 c3
2 c3
2 c3
1 +
1 +
16D
=
=
k
.
(cid:1) n
(cid:1) n by in-
16D
equality (1) and since every such vertex is contained in C3. We are allowed to
insert this lower bound for c3 into the fraction
c2+5c3/2 since the fraction is
smaller than 1 and by inserting both nominator and denominator are decreased
by the same value. Hence we can conclude
5c3/2
16D
(cid:32)
1 +
1
10
(cid:17)(cid:32)
(cid:17)(cid:32)
r ≥(cid:16)
(cid:16)
=
1 −
24
1 −
24
n −(cid:0)1 − 1
(cid:33)
5
2
16D
1
4
.
1 +
1
1− 1
16D
+ 3
2
(cid:1) n
(cid:0)1 − 1
(cid:1) n + 5
(cid:0)1 − 1
16D
2
16D
(cid:1) n
(cid:33)(cid:33)
48 and < 1 -- from which in particular follows > 2 -- , we
Since 1 − 1
finally get
r ≥(cid:16)
16D ≥ 47
(cid:17)(cid:18)
1 −
24
(47 · 24 − 474) − 472
1
4
47 + 3
2
1 +
48
= 1 +
24 · 474
(cid:19)
≥(cid:16)
(cid:17)(cid:18)
1 −
24
≥ 1 +
607
24 · 474
1 +
192
47 + 6
1
24
> 1 +
=
.
(cid:19)
(cid:16)
(cid:17)(cid:18)
1 −
24
1 +
47
474
(cid:19)
The probability that one of the bounds for c2 and c3 that we have assumed here
does not hold hold is, by the Union Bound, at most e−600 + e−7 < e−6.
We will now consider the case that there are only very few incoming 2-stars
16 l. Since, by inequality (2), it holds
that are not incoming 3-stars, i.e., c2 < 1
c3 ≤ l, we can conclude
(cid:0)1 −
(cid:1) · 11
4 l
44
c2 + 5
2 c3
(cid:16)
=
1 −
44
(cid:17) ·
(cid:16)
(cid:17) · 44
41
1 −
44
11
4 l
16 l + 5
2 l
1
=
≥
r =
=
≥ c3
c
k
k
44 −
41
≥ 41 + 2
41
> 1 +
1
24
,
with a probability of at least e−100 < e−6.
We have by now proven that r > 1 + 1
24 holds with a probability of 1− e−6 if
the assumptions of the lemma are fulfilled and the algorithm Test3StarFreeness
gets to the first but last line; hence, the algorithm rejects the input with at least
the above probability in this case. It remains to show that Test3StarFreeness
does not return true in the second but last line -- this would happen in case
c ≤ 3
4 l holds, and by in-
equality (3) we can conclude
2 n. We have, however, assumed that c3 ≥ (cid:0)1 −
(cid:1) 11
24
c ≥ c3 ≥(cid:16)
(cid:17) 11
1 −
24
l ≥ 47
48
· 11
4
n ≥ 2n
4
from this. Thus Test3StarFreeness does not return true in the second but
last line if r is estimated correctly.
at least 1 − e−6 > 5
6.
Hence the probability that Test3StarFreeness correctly returns false is
It remains to show that Test3StarFreeness works correctly if the input
graph is 3-star free.
Lemma 13. Let G = (V, E) be a weakly connected directed 3-star free graph
with n vertices and a bounded undirected vertex degree of D ∈ N. Let < 1 be a
proximity parameter. Then, Test3StarFreeness(n, G, ) returns true with a
probabilty of at least 3/4.
Proof. There are three possibilities for Test3StarFreeness inadvertantly re-
jecting the input: By considerably overestimating the edge count in the first line,
by drawing an edge sample of a size of more than 2048D
and by computing a
3/2
24. Since TestSubgraphFreenessAmplified has an one-sided
value r > 1 +
error and G is 3-star free, the subgraph freeness tests in the third line will always
return true.
Since G is weakly connected and does not contain any 3-star vertex, G con-
tains at most n edges (note that we still assume that there are no double edges
in G). By Lemma 8 the probability that EstimateEdgeCount(n, G, /16) re-
16 n is at most 2e−4; hence Test3StarFreeness
turns a value larger than n +
will not return false in the first line with a probability of at least 1 − 2e−4.
√
n
The expected number of edges that are drawn in the eigth line of Test3StarFreeness
√
n
; thus, by Markov's Inequality, the number of edges drawn will
is pn = 128D
3/2
16. Hence
not be larger than 2048D
3/2
Test3StarFreeness will not return false in the eigth line with at least this
probability.
= 16pn with a probability of at least 15
√
n
It remains to show that the estimator e does not exceed 1 +
24 with high
probability. Let k be the number of outgoing 2-star vertices in G and let c be
the number of incoming 2-star vertices. By Lemma 9 it holds k ≥ k with a
probability of at least 1 − 2e−16, and, since G is an orientation of a circle or
a line-shaped graph, the balance of G is at most 1 by lemma 3 respectively by
Lemma 5. Thus, we have k ≥ c − 1 with a probability of at least 1 − 2e−16.
Now let C2 be the set of incoming 2-star vertices; let, for v ∈ C2, Xv be an
indicator random variable for the event that both incoming edges of v are in
Xv. Es gilt E[Xv] =
the edge sample of Test3StarFreeness. Let X :=(cid:80)
v∈C2
Pr[Xv = 1] = p2. For the expected value of c we conclude
E[c] =
1
p2 X =
1
p2
E[Xv] = C2 = c.
(cid:88)
v∈C2
Now at first assume c ≥
(cid:104)
c ≥(cid:16)
Pr
1 +
32
(cid:17)
(cid:105)
4 n. Then, by a multiplicative Chernoff Bound
− 2p2c
3 · 1024
(cid:105) ≤ exp
32
E[X]
(cid:17)
= Pr
1 +
c
(cid:18)
(cid:19)
(cid:104)
X ≥(cid:16)
(cid:18)
(cid:19)
= exp
− 12823D2n
3 · 4 · 10243n
= e−2D2/3 ≤ e−6.
(cid:17)
(cid:0)1 +
24
(cid:1) −
≤(cid:16)
(cid:17)
(cid:0)1 +
32
(cid:1) c
(cid:16)
Thus, by the Union Bound, with at least a probability of 1 − 2e−16 − e−6 > 5
6:
25
24 −
c − 1
96·4 n
,
96 c
r ≤
=
+
+
1 +
1 +
24
c − 1
c − 1
24
and the last fraction is smaller or equal zero if n ≥ 400
2 ; however, this holds
because of the minimum number of vertices of G we have assumed above. Hence
it holds r ≤ 1 +
24 with a probability of at least 5
6.
Finally consider the second case for c, c <
4. By Markov's Inequality we have
6, and hence Test3StarFreeness returns true
Thus, the probability that Test3StarFreeness(n, G, ) inadvertantly re-
in the second but last line with a probability of at least 5
6.
turns false is at most 2e−4 + 1
2 n(cid:3) = Pr [c > 6E[c]] ≤ 1
Pr(cid:2)c > 3
4 by the Union-Bound.
16 + 1
6 < 1
The correctness of Teste3SternFreiheit directly follows from Lemmas 11,
12 and 13. The size of the edges sample drawn dominates the query complexity
of Teste3SternFreiheit, which is, hence, O(cid:16) √
(cid:17)
.
n
3/2
Theorem 2 Let G = (V, E) be a weakly connected directed graph with n ver-
tices and a bounded undirected vertex degree of D ∈ N. Let < 1 be a prox-
imity parameter and assume that G does not have any double edges. Then,
Test3StarFreeness(n, G, ) returns true with a probabilty of at least 3/4 if G
is 3-star free, and returns false with a probability of at least 5
6 if G is -far from
.
n
3/2
3-star free. The query complexity of the algorithm is O(cid:16) √
(cid:17)
Finally we will reconsider the restriction that the input graphs are not allowed
to have double edges. Assume that a graph G has double edges: Then let G(cid:48) be
the graph that results from deleting from every double edge in G the edge whose
starting vertex has the larger vertex number. Since in this way we delete exactly
one edge from every double edge, G(cid:48) does not have any double edges; on the
other hand, the number of adjacent vertices did not change for any vertex in G(cid:48).
Hence, G(cid:48) contains exactly the same number of 3-star vertices as G and thus is
3-star free if G is 3-star free and is -far from 3-star free if G is -far from 3-star
free. This reduction can be computed locally: For any sampled edge (u, v) it has
to be checked whether there is an edge (v, u); if this is the case and u has a
larger vertex number than v, (u, v) is considered non-existent and the sampled
edge is dismissed and, instead, a new edge is drawn uniformly by random.
We can now run Test3StarFreeness on G(cid:48) while construction G(cid:48) from G
locally. The expected number of edges we have to draw until the sample does
not get dismissed is at most 2, and by Markov's Inequality the number of edges
sampled in G is with high probability larger than the number of those returned
to Test3StarFreeness only by a constant number. If not so, the algorithm
can abort computation by returning true, which will result in a small increase
of the error probability in case the input graph is -far from 3-star free:
Corollary 2. Let G = (V, E) be a weakly connected directed graph with n ver-
tices and a bounded undirected vertex degree of D ∈ N. Let < 1 be a proximity
parameter. Then, Test3StarFreeness(n, G, ) returns true with a probabilty
of at least 3/4 if G is 3-star free, and returns false with a probability of at least 3
4
.
if G is -far from 3-star free. The query complexity of the algorithm is O(cid:16) √
(cid:17)
n
3/2
4 Lower Bounds for Testing 3-Star Freeness
√
In this section we give two lower bounds on the query complexity of property-
testing algorithms for 3-star freeness: A lower bound of Ω(
n) queries for testing
3-star freeness in weakly connected graphs, and a lower bound of Ω(n2/3) queries
if weak connectivity of the input graphs is not required. This means that the
algorithm given in the last section is asymptotically optimal in the number n
of vertices of the input graph; moreover, it means that testing 3-star freeness is
easier if the input graphs are guaranteed to be weakly connected. This brings
up the question whether there are more problems for which this holds true.
The first lower bound, that of Ω(
n) for testing in weakly connected graphs,
follows from graph classes given by Bender and Ron for their lower bound for
testing strong connectiviy [5]: Their class of strongly connected graphs consists
of orientations of circles, where all edges have the same orientation. Such graphs
are 3-star free. Their class of graphs that are -far from strongly connected
consists of orientations of circles where all edges have the same orientation, but
additionally there are more than Dn outer vertices that have an edge towards a
circle vertex. Each circle vertex at most one incoming edge from an outer vertex,
and since a circle vertex that has such an edge possesses three neighbours, it is a
√
3-star vertex. By construction there are more than Dn such vertices, and since
converting such a graph into a 3-star free graph requires deletion of one incedent
edge for each of these vertices, those graphs are -far from 3-star free. Since
Bender and Ron show that the two classes of graphs cannot be distinguished
√
with o(
n) queries, we can conclude the following corollary:
Corollary 3. In the adjacency list model for directed graphs where algorithms
√
cannot query the incoming edges of a vertex, every property testing algorithm
n), where n is the number
for 3-star freeness has a query complexity of Ω(
of vertices of the input graph. This even holds if the possible input graphs are
restricted to be weakly connected.
In the remainder of this section we will derive a lower bound of Ω(n2/3)
for the query complexity of every property testing algorithm for 3-star freeness
that does not require the input graph to be weakly connected. Note that for
the above bound of Ω(n1/2) both graph classes considererd consist of weakly
connected graphs. We will make use of a technique invented by Raskhodnikova
et al. [15] and construct two classes of problem instances -- one consisting of 3-
star free graphs, the other of -far ones -- such that the distributions of incoming
vertex degrees in graphs of these two classes have 2 proportional moments. By
the results of [15] that means that every algorithm that measures a significant
difference between these distributions uses Ω(n2/3).
The difficulty here is to show that measuring vertex in-degrees is the only
thing an algorithm can do to compare graphs from these classes. Since this is
hard to argue even for graphs that only consist of isolated k-stars, we will define
a helper problem that is very similar to the distinct elements problem examined
in [15]. We will then reduce testing this helper problem to testing 3-star freeness.
The helper problem is defined as follows: We are given a sequence A of m
integers Ai ∈ N, i = 1, . . . m; each value occurs in at most 3 elements of A. Let l
be the number of distinct values that occur in A: We assume that all these values
are from {1, . . . , l}. If there is no value a ∈ {1, . . . , l} that occurs in 3 elements
of A, we call A 3-value free. Note that this problem is completely characterized
by the sequence A.
For property testing we call a sequence A -far from 3-value free, if more
than m elements of A have to be changed to establish 3-value freeness; for that
matter it is allowed to assign values that are not yet assigned to one of the
elements of A, i.e., numbers that are larger than l. A property testing algorithm
knows the number m of elements and may query the value of the i-th element
of the sequence for i = 1, . . . , m.
We call an algorithm poisson-s algorithm if it determines randomly by the
poisson distribution how many random samples it draws. The analysis in [15]
requires poisson-s algorithms that only get access to the histograms of their
random samples; i.e., the algorithms get the information how many values in
the sample occur once, twice, thrice, etc., but not the numbers of the elements
or their values itself:
Definition 9. Let S be a multiset. The histogram H of S is a function that
assigns to each integer i ∈ N the number of elements of S that occur exactly i
times in S; i.e.,
H(i) := {s ∈ Ss is contained in S exactly i times} .
We make use of two Lemmas from [15] in order to show that for a lower
bound for testing 3-value freeness it suffices to consider poisson-s algorithms
that only get access to the histogram of their samples. The following Lemma is
a direct corollary of Lemma 3.12 and of Lemma 5.3 a) and c) from [15]:
Lemma 14. Let A be an arbitrary sampling-based algorithm that draws t ele-
ments of a sequence A; assume that A accepts with a probability of at least 3
4 if
A is 3-value free and rejects with a probability of at least 3
4 if A is -far from
3-value free.
Then there is a poisson-s algorithm A(cid:48) that has an expected number of s =
O(t) queries and only gets access to the histogram of these queries and that
4 − o(1) if A is 3-value free and rejects with
accepts with a probability of at least 3
a probability of at least 3
4 − o(1) if A is -far from 3-value free.
Note that the property of 3-value freeness of a sequence A is closed under
permutation of the sequence elements, as is the property of -farness of 3-value
freeness; thus these properties fulfill the premises of Lemma 3.1 and 5.3 c) in
[15]. The probability guarantees of A(cid:48) follow from those that Lemma 5.3 a) in
[15], since the statistical distance between the distributions of the results of A
and A(cid:48) is at most 4
Analogously to [15] define a frequency variable XA as a random variable
for the number of occurences in A that a value drawn randomly uniformly dis-
tributed from {1, . . . , l} has. The following Lemma is a also a direct corollary
from [15], namely from Corollary 5.7:
s = o(1)3.
Lemma 15. Seien A und B Probleminstanzen von 3-Wert-Freiheit und seien
XA und XB die zugehörigen Frequenzvariablen. Falls XA und XB k − 1 pro-
portionale Momente haben, dann gilt für jeden Poisson-s-Algorithmus A mit
s = o(n1−1/k), der nur die Histogramme seiner Samples sieht
Pr[A(A) = true] − Pr[A(B) = true] = o(1).
Two random variables X1 and X2 are said to have k−1 proportional moments,
E[X2] = E[X 2
We can now use the preceeding Lemmas in order to prove that any property-
2 ] = . . . = E[X k−1
E[X k−1
1 ]
E[X 2
]
]
if E[X1]
.
1
2
testing algorithm for 3-value freeness needs at least Ω(n2/3) samples.
Lemma 16. Any property-testing algorithm for 3-value freeness needs at least
Ω(n2/3) samples, where n is the length of the input sequence.
2 Cited by Raskhodnikova et al. from [4].
3 Due to corollary 3 we can assume s = Ω(
√
n), and hence it holds 4
s = o(1).
Proof. At first note that, due to Lemma 14, every property testing algorithm for
3-value freeness can be replaced by a poisson-s algorithm with an asymptotically
equal query complexity, which only gets access to the histogram of its queries --
for sufficiently large sequence length n the additional error probability is negleg-
ible. Hence it suffices to show that every poisson-s algorithm for 3-value freeness
needs to take at least an expected number of s = Ω(n2/3) queries. Furthermore,
due to Lemma 15, every such poisson-s-algorithm needs at least Ω(n2/3) queries
in expectation to distinguish two sequences A and B whose frequency variables
XA and XB have to proportional moments. We will now give two such classes,
CA being a class of 3-value free sequences and CB a class of sequences that are
-far from 3-value free. From the above considerations we can then conclude that
testing 3-value freeness requires Ω(n2/3) queries.
Let CA be the class of all sequences A of length n, such that n is a multiple
2 n distinct values, each of them two times. Every sequence
Let CB be the class of all sequences B of length n, such that n is a multiple
of 32 and the values that occur in B are distributed as follows: There are 17
32 n
distinct values in total, of whom 1
32 n values 2 times
32, all of the sequences in CB are -far from 3-value
and 3
free.
We will now show that the frequency variables of two sequences of equal
length from CA and CB have two proportional moments. Let, for an arbitrary
fixed n, A ∈ CA and B ∈ CB with length n each and let XA and XB be the corre-
sponding frequency variables. It holds E[XA] = 2 and E[X 2
A] = 4. Furthermore,
we have
of 32 and A contains 1
in A is 3-value free.
32 n values occur 3 times, 13
32 n values once. For < 1
E[XB] =
· 3 +
1
17
13
17
· 2 +
· 1 =
3
17
32
17
and
E[XB] =
1
17
· 9 +
· 4 +
13
17
3
17
· 1 =
64
17
.
We can conclude
E[X 2
A]
E[X 2
B]
and hence XA and XB have two proportional moments.
E[XA]
E[XB]
=
=
=
,
2 · 17
32
4 · 17
64
Wir werden nun durch Reduktion zeigen, dass jeder Property-Testing-Algorithmus
für 3-Stern-Freiheit Ω(n2/3) Anfragen benötigt, wenn dies auch eine untere
Schranke für das Testen von 3-Wert-Freiheit ist.
Theorem 3 In the adjacency list model for directed graphs where algorithms
cannot query the incoming edges of a vertex, every property testing algorithm
for 3-star freeness has a query complexity of Ω(n2/3), where n is the number of
vertices of the input graph.
Proof. For a proof by contradiction we assume that there is a property testing
algorithm A for 3-star freeness that has a query complexity of o(n2/3) for input
graphs with n vertices. We will construct an algorithm A(cid:48) that takes as input a
2 n and a proximity parameter (cid:48), and we will prove
sequence A of length n(cid:48) = 1
that A(cid:48) is a property testing algorithm for 3-value freeness that has a query
complexity of o(n2/3).
A(cid:48) works by calling A and returning the return value of A; the proximity
parameter of A is set to 2(cid:48) and A is given oracle access to a graph G = (V, E)
with n = 2n(cid:48) vertices that is dynamically constructed for each query of A. G
consists of isolated stars or single vertices and is defined as follows:
-- V consists of n(cid:48) central vertices v1, . . . , vn(cid:48) and n(cid:48) outer vertices u1, . . . , un(cid:48).
-- For i = 1, . . . , n(cid:48), there is an edge from ui to vAi; i.e., the outer vertices for
whose numbers the corresponding elements of A have the same value form
a star together with a common central vertex.
-- G does not contain any further vertices and edges; thus, there may be isolated
vertices in G.
1
Hence, the vertices ui represent the elements of the sequence A in G, and by
having a common central vertex in G it is represented that the corresponding
elements of A have the same value.
For a query of A, G can be constructed locally4: When an unknown vertex is
queries, it is determined by throwing a coin whether this vertex will be a central
n(cid:48)−c
2n(cid:48)−c−o, if so
vertex or an outer vertex; the probability for a central vertex is
far c central vertices and o outer vertices are known to A. If the queried vertex
turns out to be an outer vertex, A(cid:48) queries a sequence element that is randomly
uniformly distributed among those that have not already been queried; hence the
probability for each of the remaining elements is
n(cid:48)−o if, again, o is the number
of previously queried outer vertices. Basically this is equivalent to drawing from
A without replacement.
Now let i be the number of the sequence element that has been selected in
this way: A(cid:48) inserts an edge (u, vAi ) into the graph in construction. Since A can
only query the outgoing edges from vertices and all the edges of G are directed
from outer to central vertices, every edge that A queries can be constructed in
this way.
This construction procedure needs at most 1 query to A for every query of
A; hence the query complexity of A(cid:48) is at most that of A, which is o(n2/3) =
o((n(cid:48))2/3). Furthermore, each value that occurs i times in A corresponds to an
i-star in the graph G constructed by A(cid:48). Hence, G is 3-star free if and only if A is
3-value free, and, by the choices of and (cid:48), G is -far from 3-star free if and only
if A is (cid:48)-far from 3-value free. Thus, A(cid:48) is a property testing algorithm for 3-
value freeness that has a query complexity of o((n(cid:48))2/3), which is a contradiction
to Lemma 16. Hence there does not exist a property testing algorithm A for
3-star freeness that has a query complexity of o(n2/3).
4 The construction method given here does not necessarily create the above vertex
permutation, but the resulting graph is isomorphic to G. For the sake of simplicity,
we will refer to it as G.
We finally discuss lower bounds for the query complexity of testing k-star
freeness for k > 3. The property of k-value freeness is defined analogously to
3-value freeness, and the Lemmas 14 and 15 hold analogously for k-star freeness;
the reduction from k-value freeness to k-star freeness can be done analogously
to the special case k = 3. The difficulty is to give classes CA and CB like in the
proof of Lemma 16 that have k − 1 proportional moments. This can be easily
done for fixed k > 3, but it is an open problem to give closed formulae for the
frequency distributions of the different i-stars in graphs of these classes. We close
this chapter with the below conjecture:
Conjecture 1. In the adjacency list model for directed graphs where algorithms
cannot query the incoming edges of a vertex, every property testing algorithm
for k-star freeness has a query complexity of Ω(n1−1/k), where n is the number
of vertices of the input graph.
5 Testing Strong Connectivity
Strong connectivity, i.e., the question whether all pairs of vertices are connected
by paths in both directions, is a very basic property of directed graphs. In this
chapter we will give a property testing algorithm for strong connectivity that
Property-Testing for this graph property was at first considered by Bender and
Ron [5], who, amongst others, give a lower bound of Ω(n1/2) for any property
testing algorithm for this problem under the assumption that the algorithm may
only query the incoming edges of a vertex. After first publishing this result at
ESA 2012 [13] we discovered that there already existed a proof sketch for an
algorithm similar to ours, which Oded Goldreich published in the appendix of
a survey article about graph property testing [9]. At the end of this section
we will include a detailed discussion about similarities and differences between
Goldreich's and our algorithm.
The above-mentioned lower bound construction of Bender and Ron -- more
precisely the class of -far graphs that Bender and Ron give -- can be used for
deriving a lower bound of Ω(n) for any property testing algorithm with one-
sided error in the same graph model: Let G be a graph with n vertices that
consists of a directed circle of n − Dn − 1 vertices, whose edges all have the
same direction, and Dn + 1 outer vertices that have exactly one edge to one
of the circle vertices; each circle vertex has at most one incoming edge from an
outer vertex (see figure 3). G is -far from being strongly connected, since for
archieving strong connectivity, every outer vertex needs an additional incoming
edge. On the other hand, every subgraph H of G induced by at most n− Dn− 2
vertices can be completed to a strongly connected graph G(cid:48) with n vertices as
follows: There are Dn+2 vertices and their adjacent edges to add; for every outer
vertex of G, add an edge from one of the missing vertices; in particular, since at
least one of the circle vertices of G is missing in H, this missing circle vertex can
be connected to one of the outer vertices. After this, add the remaining edges
of G to G(cid:48), ignoring those that would create an outgoing 2-star at its source
G
H
G(cid:48)
Fig. 3: Graph G, example for a subgraph H explored by an algorithm and strongly
connected graph G(cid:48) constructed from H
vertex (see again figure 3). The resulting graph is strongly connected. Hence,
no algorithm that only queries n − Dn − 2 vertices of G and only its outgoing
edges can rule out the possibility that G is strongly connected; hence, every
such property testing algorithm for strong connectivity that has one-sided error
would have to accept G.
Corollary 4. Every property testing algorithm for strong connectivity of di-
rected graphs in the adjacency list model with only outgoing edges being visible
for the algorithm has a query complexity of Ω(n).
In the remainder of this section we will present a property testing algorithm
for strong connectivity in directed graphs that has a two-sided error and a query
complexity of O(n1−/(3+α)); α > 0 is an arbitrarily small constant. We assume
that both the in- and the outdegree of every vertex of the input graph is bounded
by a constant D ∈ N; hence, the undirected degree of a vertex can at most be
2D. The total number of edges is still at most Dn, such that a graph is -far
from another graph, if more than Dn adjacency list entries of the first graph
have to be modified to obtain the second one.
In the following we can assume D ≥ 2, since for D = 1 strong connectivity
can be easily tested as follows: If D = 1, every vertex has at most one incoming
edge and one outgoing edge; every weakly connected component in such a graph
is either an orientation of a circle or a line-shaped graph, and all the edges in
each of these orientations have the same direction. The only strongly connected
graph is a circle orientation with all n vertices where all edges have the same
direction. Every -far graph consists of Ω(n) weakly connected components,
which means that there are also Ω(n) weakly connected components that have
vertices each -- a vertex sample of size Θ(1/) contains one of them
at most 1
with high probability. By the above considerations, starting from an arbitrary
vertex of component, every single of these components can either be completely
queries (in case it is a circle) or a sink vertex can be found
explored with 1
TestSinkFreeness(n, G, β)
Sample s5 = 4
Foreach sampled vertex v do
βD vertices of G u.i.d. at random
Start a breadth first traversal at v which stops after exploring 2
If the breadth first traverdal completely explores a sink component
βD vertices
then return false
return true
Algorithm 5: TestSinkFreeness
within 1
query complexity of Θ( 1
queries. Hence, for D = 1, strong connectivity can be tested with a
2 ) with one-sided error.
Corollary 5. For D = 1, strong connectivity of a directed graph with both vertex
in- and outdegree bounded by D ∈ N can be tested with O( 1
2 ) queries.
For larger values of D, testing for strong connectivity turns out to be much
harder; one main obstacle is that one cannot rely on identifying whitnesses
against strong connectivity for -far graphs: Such whitnesses would either be
sink or source components of the input graph, and while it is easy to identify
a sink component if the input graph contains many of them, a source compo-
nent can never be identified without knowing the whole remaining graph, since
it cannot be ruled out that an explored area has incoming edges. On the other
hand, Bender and Ron have shown that the total number of dead ends in an
-far graph is large:
Lemma 17 ([5]). Let G be a directed graph with n vertices and both its vertex
in- and outdegree bounded by D ∈ N and let < 1 be a proximity parameter. If
G is -far from strongly connected, then G has more than 1
3 Dn dead ends.
Input graphs that contain many small sink components can be handled in the
same way as in the property testing algorithm of Bender and Ron for the model
where incomong edges are visible: It suffices to sample O(1/) vertices, explore
a small area of size O(1/) vertices around them using breadth first search and
reject the input graph, if for one of the sample vertices there are no outgoing
edges left during the breadth first search traversal. It can be shown that, if G has
many sink components, then G also has many sink components of size O(1/),
and hence the above approach will identify one of them with high probability.
Lemma 18. Let G be a directed graph with n vertices both its vertex in- and
outdegree bounded by D ∈ N and let β < 1 be a proximity parameter. If G does not
have any sink components, TestSinkFreeness(n, G, β) always returns true; if
G has at least βDn sink components, TestSinkFreeness(n, G, β) returns false
with a probability of at least 1− e−2. The query complexity and the running time
of the algorithm both are O(cid:16) 1
(cid:17)
.
β2D2
EstimateReachableVertices(n, G, D, )
for i ← D downto 1 do
pi
i
return m = Dn
i
1≤i≤D ni
Sample each edge of G with a probability of pi = ain−1/i,
If more than ti = 16Dain1−1/i edges are sampled then return n
ci ← number of vertices for which exactly i incoming edges have been sampled
ni ← ci
(cid:1)nj(1 − pi)j−i
(cid:0)j
−(cid:80)
16 +(cid:80)
i<j≤d
Algorithm 6: EstimateReachableVertices
Proof. If G does not have any sink components, the condition queried by the if
statement in the second but last line of TestSinkFreeness can never turn out
to be true; hence the algorithm alway accepts the input.
2 βDn of the sink components have
Now assume that G contains at least βDn sink components. Since G has n
vertices Knoten and each of them belongs to at most one sink component, at
least 1
βD or less vertices. If a vertex v of
such a sink component is sampled, then the breadth first search that starts at v
completely explores the sink component. Hence the algorithm correctly returns
false. The probability that none of the s5 sample vertices lie in one of the small
sink components is at most (1 − 1
2 βD)s5 ≤ e−2.
Running time and query complexity of the algorithm arise from the num-
ber s5 of sample vertices taken mulitplied by the maximum number of vertices
explored by the breadth first search.
2
As discussed above, source components that are not sink components at the
same time are much harder to identify as sink components. Indeed, corollary 4
implies that they cannot be directly identified (for example by a breadth first
search) in sublinear time. However, there may be graphs that are -far from
strong connectivity and that have very few sink components, and a property
testing algorithm for strong connectivity has to identify such graphs. Hence our
algorithm will identify the presence of many source components indirectly: We
will at first assume that the input graph G has many small source components
and that each of them contains of exactly one vertex; we will show that in this
case there is statistical evidence for G not being strongly connected. In the next
step we will give a reduction that is locally constructible -- that is, given G, the
reduced graph can be sampled from and explored with a constant number of
queries in G for each query to the reduced graph -- and that converts a graph
with many constant-size source components into a graph that has equally many
source components with a size of one each.
Hence we reduce the problem of testing a graph for strong connectivity to the
problem of approximating the number of vertices that have no incoming edge.
We solve the latter problem by computing collision statistics for i-way collisions
of common target vertices of a set of sampled edges, for i = 1, . . . , D; combining
these statistics allows us to measure the number of vertices that do not have an
incoming edge and thus cannot appear in any of the collision statistics, affecting
their outcomes significantly if there are sufficiently many such vertices. At this
moment, we only state the following lemma, the proof will be given at the end
of the section5:
Lemma 19. Let G be a directed graph with n vertices both its vertex in- and
outdegree bounded by D ∈ N and let < 1 be a proximity parameter. Let m
be the number of vertices of G that have at least one incoming edge. Then,
EstimateReachableVertices(n, G, ) returns an estimate m for which
m ≤ m ≤ m +
4. The algorithm needs at most O(cid:16) D8D+15 log D
3
holds with a probability of at least 3
queries to G.
Dn
8
n1−1/D(cid:17)
The basic proof strategy is to show by induction that, for appropriate choice
of the ai, the estimators ni approximate the number of vertices with exactly
i incoming edges with high probability (see algorithm 6). Then, by the Union
Bound, m satisfies the constraints given in Lemma 19 with high probability.
We will now introduce the reduction function C. For this purpose, we need
the notion of a compact component; the reduction will create a graph in which
every compact component in the original graph is contracted into a single vertex:
(a)
U1
U2
(b)
U
u2
u1
v
Fig. 4: (a) Compact components U1, U2; (b) for 3+α
component, since the highlighted path violates the third condition.
D = 5, U is no compact
Definition 10. Let G be a directed graph with n vertices both its vertex in- and
outdegree bounded by D ∈ N and let < 1 be a proximity parameter; let α > 0
be a fixed, but arbitrary constant. We call a set of vertices U ⊆ V compact
component if the following three conditions hold:
1. U ≤ 3+α
D ;
5 In [15], section 4, footnote 4, a similar approach for approximating a distribution with
several i-way collision statistics as in EstimateReachableVertices is mentioned,
but its correctness is not proved there.
2. The subgraph of G that is induced by U is strongly connected;
3. There are no vertices v ∈ V \U and u1, u2 ∈ U such that there are paths from
u1 tp v and from v to u2, each having a length of at most 3+α
D .
The following Lemma shows that the compact component of every vertex of
a graph is unique:
Lemma 20. Let G = (V, E), D, and α be defined as in definition 10. Then,
every vertex of G belongs to at most one compact component.
Proof. Let v ∈ V be an arbitrary vertex. Assume that v belongs to two distinct
compact components U and W . Since U (cid:54)= W , there is at least one vertex that
is in one of the components, but not in the other; let, without loss of generality,
u ∈ U\W be such a vertex. Since U is a compact component, U is strongly
D vertices. Hence, since v, u ∈ U, there are
connected and contains at most 3+α
paths from v to u and from u to v, each having a length of at most 3+α
D . This
is, however, a contradiction to the assumption that W is a compact component:
Since v ∈ W and u /∈ W , the existence of the above paths violates the third
condition for compact components. Hence there are not two distinct compact
components v belongs to.
There may be vertices of a graph that do not belong to any compact compo-
nent by the above definition. In the following we will assume that such vertices
form their own compact components. Together with the above Lemma this means
that every vertex of a graph can be mapped to a unique compact component:
Definition 11. Let G = (V, E), D, and α be defined as in definition 10. Then,
for a vertex v ∈ V , C(v) denotes the compact component that v belongs to, if
there exists such a component, or C(v) = {v} if v does not belong to a compact
component.
Let additionally C(G) be the graph that results from contracting every com-
pact component of G.
Next we show that every vertex of a compact component has the same com-
pact component.
Lemma 21. Let G = (V, E), D, and α be defined as in definition 10. Then,
C(u) = C(v) for all vertices v ∈ V and u ∈ C(v).
Proof. We assume C(v) > 1, since otherwise the Lemma holds trivially because
of C(v) = {v}. Furthermore we assume that there os a vertex u ∈ C(v) such
that C(u) (cid:54)= C(v); i.e., there exists a vertex that is only contained in either in
C(v) or in C(u).
First assume that there exists w ∈ C(u)\C(v): Then, analogously to the
proof of Lemma 20, there exist paths from u to w and from w to u of length
D each, and, since v ∈ C(v) but w /∈ C(v), this violates the third
at most 3+α
condition for compact components for C(v) -- a contradiction to the assumption
that C(v) is a compact component.
Now assume that there exists w ∈ C(v)\C(u) gibt. Analogously to the first
case this implies the existence of short paths from v to w and from w to v,
contradicting with the third condition for compact components for C(u).
By the above Lemma, the mapping C is well-defined. The fundamental prop-
erty of this mapping is that it maps every small source component of a graph to a
compact component; in C(G), each of these compact components gets contracted
to a single vertex, such that C(G) contains many vertices with an indegree of 0
if G contains many small source components.
Observation 2 Let G = (V, E), D, and α be defined as in definition 10 and
let U ⊆ V be a source component that contains at most 3+α
D vertices. Then, U
is a compact component.
Proof. By the assumption of the Lemma, U ≤ 3+α
D , and by definition of source
components, U is strongly connected. Additionally, there is no edge from V \U
to U, and thus there is not path from any vertex in V \U to a vertex in U. Hence
the three conditions for compact components are satisfied for U.
6+2α
O(cid:16) 1
D −1(cid:17)
For the remainder of this section we assume that α > 0 is chosen fixed but
arbitrary; we will assume an appropriate choice for α implicitly whenever we
refer to C(G) for a graph G. Analogously, is used implicitly there. It remains
to show that a compact component of a given vertex can be determined in a
constant number of queries and how sampling in a contracted graph C(G) can
be realized.
Determining the compact component of a given vertex v can be done with
· D
queries as follows: Start a breadth first search at v that explores
up to a maximal depth of 3+α
D in order to find every vertex that can possible be
in a compact component that includes v. Let U be the maximal set of vertices
that have been explored in this way such that v ∈ U and the subgraph of G
induced by U is strongly connected; if U > 3+α
D , it holds C(v) = {v}, elsewise
U is the candidate set for the compact component of v. We verify U by checking
for vertices that violate the third condition for compact components: For this
purpose we perform a breadth first search with a maximum depth of 6+2α
D ,
having U as the set of starting vertices. The depth boundary on this breadth
first search traversal makes for the above-mentioned query complexity.
Sampling a vertex in C(G) is realized by at first sampling a vertex of G and
then computing the connected component that the vertex belongs to. A vertex
sample is accepted with a probability of 1/i, if i is the number of vertices in the
corresponding component; in this case, the vertex of C(G) that represents the
sampled vertex of G is returned. Otherwise, the sample is thrown away. This
procedure ensures that all vertices have the same sample probability.
For this procedure, several tries may be necessary until a vertex or an edge
has been successfully sampled; however, the probability of a successful sample is
3+α, since this is the inverse of the maximum number of vertices that
at least D
can be represented by a vertex of C(G); hence, sampling k times yields k · D
3+α
EstimateVertexNumber(n, G, )
x ← 0
Sample s7 = 3
Foreach sampled vertex v do x ← x + 1C(v)
return n = n
2d2 vertices of G u.i.d. at random
t x
Algorithm 7: EstimateVertexNumber
p · 3+α
successful samples in expectation, and by Markov's Inequality, sampling k
yields k successful samples with a probability of at least (1 − p). This leads to
the following observation:
Observation 3 Let G = (V, E), D, and α be defined as in definition 10 and let
p ∈ (0, 1). Sampling k vertices of C(G) uniformly and independently distributed
and with an error probability of at most p can be done with O(cid:16) k
D −2(cid:17)
6+2α
2
2p · D
queries to G.
For using the algorithm TestSinkFreeness on a contracted graph C(G), an
estimate for the number of vertices of C(G) is needed. The algorithm Estimat-
eVertexNumber returns such an estimate with high probabilty; in contrast
to the approximation algorithm for the number of connected components of an
undirected graph by Chazelle et al. ([7]), here the idea is to sample vertices of G
and increase the estimate by the inverse of the size of the compact component
of the sample vertex.
Lemma 22. Let G be a directed graph with n vertices both its vertex in- and
outdegree bounded by D ∈ N and let < 1 be a proximity parameter; let n
be te number of vertices of C(G). Then, for the value n that is returned by
EstimateVertexNumber(n, G, ) it holds
with a probabilty of at least 1 − 2e−3. The algorithm has a query complexity of
n − Dn ≤ n ≤ n + Dn
O(cid:0) 1
3D3 · D(6+α)/(D)(cid:1).
X =(cid:80)
1≤j≤s7
Proof. Let, for i = 1, . . . , 3+α
D , ki be the number of compact components of
G that consist of exactly i vertices. Let Xj be a random variable for the con-
C(vj ). Let
tribution of the j-the sampe vertex to x, j = 1, . . . , s7, i.e., Xj = 1
(cid:88)
Xj. For every j it holds
· Pr[C(vj) = i] =
1
i
(cid:88)
1≤i≤ 3+α
D
E[Xj] =
1≤i≤ 3+α
D
· ki · i
n
1
i
=
1
n
· (cid:88)
1≤i≤ 3+α
D
ki =
n
n
,
since the sum in the first but last step is the number of compact components of
G and hence the number of vertices of C(G). Moreover,
E[n] =
· E[X] =
n
s7
n
s7
E[Xj] =
n
s7
n
n
= n,
· (cid:88)
1≤j≤s7
· (cid:88)
1≤j≤s7
TestStrongConnectivity(n, G, )
if D ≥ 3 + α then return true
if not TestSinkFreeness(n, G,
n ←EstimateVertexNumber(C(G),
m ←EstimateReachableVertices(n, C(G), 3+α
if m < n −
else return true
12(3+α) Dn then return false
24(3+α) )
α
α
α
6(3+α) ) then return false
, α2D
6(3+α)2 )
Algorithm 8: TestStrongConnectivity
and since all Xj are between 0 and 1, applying the Hoeffding Inequality yields
Pr[n−E[n] > dn] = Pr[X−E[X] > ds7] ≤ 2 exp
−
By the above considerations, determining the compact component of a vertex
of G takes at most O(cid:0) 1
2d2 vertices in the algorithm, the overall query complexity is O(cid:0) 1
D · D(6+α)/(D)(cid:1) queries to G; since this is done for s7 =
3D3 · D(6+α)/(D)(cid:1).
3
(cid:32)
(cid:80)
(cid:33)
2d2s2
7
(1 − 0)2
1≤i≤s7
= 2e−3.
Now we can state our property testing algorithm TestStrongConnectiv-
ity; it at first calls TestSinkFreeness in order to test whether there are many
sink components; if this is not thr case, the algorithm estimates the number of
vertices of C(G) by calling EstimateVertexNumber and uses this value to
measure the number of vertices that have at least one incoming edge by call-
ing EstimateReachableVertices. If the estimate returned by the latter call
is much smaller than the estimated number of vertices of C(G), then Test-
StrongConnectivity rejects; elsewise it accepts.
Theorem 4 Let G be a directed graph with n vertices both its vertex in- and
outdegree bounded by D ∈ N and let < 1 be a proximity parameter. Then,
TestStrongConnectivity(n, G, ) returns false with a probability of at least
3, if G is -far from strongly connected; it returns true with a probability of
2
at least 2
3, if G is strongly connected. The query complexity of the algorithm is
O(cid:16)
α−4(cid:0) 4
(cid:1)32/+23
D144/(α2)n1−/(3+α) log 1
.
(cid:17)
Proof. Let n be the number of vertices of C(G) and let m be the number of
those vertices that have at least one incoming edge. Let D be the maximum in-
or outdegree of a vertex in C(G): Because of the maximum size of a compact
component it holds D ≤ 3+α
At first assume that G is -far from strongly connected. We will show that
then the probability that TestStrongConnectivity inadvertantly returns
true is bounded by a constant. Initially note that this cannot happen in the
first line of the algorithm: For ≥ 3 + α there is no graph that is -far from
strongly connectes, since in this case it would be possible to remove all edges of
.
3 true.
In the second case at least
the graph and insert a circle over all vertices instead, not surpassing Dn edge
modifications; hence, for an -far graph the condition of the if-statement in line
1 will never hold true.
Since G is -far from strongly connected, G contains more than 1
3 Dn dead
6(3+α) Dn sink components or
ends by Lemma 17. Hence it contains at least
at least 1
6(3+α) Dn source components. In the first case
Lemma 18 guarantees that TesteSenkenFeiheit returns true with a proba-
bility of at most e−2 < 1
6(3+α) Dn = 6+α
3 Dn − α
α
6(3+α) Dn of the source components have a size of
D , since they are vertex-disjoint and together contain at most n ver-
6(3+α) Dn,
at most 3+α
tices. Thus the number of vertices without incoming edges is at least
too; i.e., it holds m ≤ n − α
α
α
6(3+α) Dn.
We can use this bound on the number of non-reachable vertices of C(G)
to gain an upper bound for m that holds with high probability: Lemma 19
6(3+α)2 · Dn holds with a probability of at least 3
guarantees that m ≤ m + 1
4;
moreover, by Lemma 22 we have n ≥ n −
24(3+α) Dn with a probability of at
least 1− 2e−3. These two facts combined with the bound of D ≤ D · 3+α
D = 3+α
for the in- and outdegree of a vertex of C(G) yield
8 · α2D
α
m ≤ m +
≤ n +
1
8
1
4
· α2D
6(3 + α)2 · Dn ≤ n −
·
α
6(3 + α)
Dn − 7
8
·
6(3 + α)
α
6(3 + α)
α
Dn +
1
8
Dn < n −
·
α
6(3 + α)
Dn
α
12(3 + α)
Dn.
Hence, by the Union Bound, TestStrongConnectivity(n, G, D, ) returns
false with a probability of at least 1 − 3
4 − 2e−3 > 2
3 in this case.
Finally we consider the case that G is strongly connected, i.e., there are no
dead ends in G. By Lemma 18, the call to TesteSenkenFreiheit will hence
return true true. Particularly, there are no source components in G, and thus
there are no vertices without incoming edges in C(G) -- every such vertex in
C(G) would correspond to a source component in G. Hence it holds m = n.
Moreover, Lemma 19 guarantees that m ≥ m holds with a probability of at least
4 and by Lemma 22 we have n ≤ n +
24(3+α) Dn with a probability of at least
1 − 2e−3. Altogether we can conclude
α
3
m ≥ m = n ≥ n −
α
24(3 + α)
Dn > n −
α
12(3 + α)
Dn,
and hence by the Union Bound TestStrongConnectivity(n, G, D, ) returns
3.
true in this case with a probability of at least 2
The query complexity of the algorihtm is dominated by the query complex-
, α2D
6(3+α)2 ): Since the
ity of the call EstimateReachableVertices(n, C(G), 3+α
degree bound of C(G) is 3+α
von C(G), for α < 1 this call queries C(G)
(cid:19)8(3+α)/+15
n1−/(3+α) log
1
(cid:33)
= O(cid:16)
(cid:19)
O
α−3
= O
6(3 + α)2
(cid:32)(cid:18) α2D
(cid:19)−3(cid:18) 3 + α
(cid:32)
(cid:18) 4
(cid:19)32/+21
(cid:18) 6(3 + α)2
(cid:32)
(cid:19)32/+23
α2D2
(cid:18) 4
O
α−4
times. Because each of these queries induces
· D6(6+α)(3+α)2/(α2D2)
O
queries to G, the overall query complexity is
D144/(α2)n1−/(3+α) log
(cid:33)
n1−/(3+α) log
3 + α
−2α−1D144/(α2)(cid:17)
(cid:33)
1
.
1
It remains to show 19, i.e., that the estimate m returned by algorithm Es-
timateReachableVertices satisfies m ≤ m ≤ m + 1
8 Dn with a probability
of at least 3
4, if m is the number of reachable vertices in the input graph C(G),
and that this algorithm has a query complexity of O(−3D8D+15n1−1/D log D).
Proof (Beweis (Lemma 19).). We will show by induction that the basic estima-
tors ni are a good approximation for the number ni of vertices that have exactly
i incoming edges; we then conclude that m is a good estimate for m. At first note
that, by Markov's Inequality, the probability that the algorithm aborts in the
third line (and possibly returns the wrong value) is at most
16D ; the probability
that this happens in at least one of the D iterations of the for-loop is at most 1
16
by the union bound hence. If the input graph G is a contracted graph, we can
set the error probability in Observation 3 to 1
16D ; the probability that in at least
one iteration of the for-loop the required number of samples is not attained is
at most 1
16 then because of the union bound. For the remainder of this proof we
will assume that neither of these events occurs.
Now let ai := aD/i für i = 1, . . . , D; we will determine the exact value of the
constant a later. For all these values for i let Vi be the set of vertices of G that
have exactly i incoming edges; we have ni = Vi hence.
We define the random variable Yi,j as the number of i-way collisions on
vertices in Vj in the iteration of the for-loop that has the loop counter i; an
i-way collision is the event that for a vertex v exactly i of its incoming edges are
contained in the edge sample taken in the second line of the algorithm. Moreover
let Xv,i be an indicator random variable for the event that in the iteration that
has the loop counter i the vertex v has an i-way collision. For every vertex v ∈ Vj
it holds
E[Xv,i] = Pr[Xv,i = 1] =
=
(ain−1/i)i(1 − pi)j−i =
i(1 − pi)j−i
pi
(cid:18)j
(cid:19)
aD(1 − pi)j−i,
1
n
i
(cid:18)j
(cid:19)
i
(cid:18)j
(cid:19)
i
since the events for the incoming edges of v being in the sample set are indepen-
dent from each other. Thus the number of incoming edges of v that are contained
in the sample set is binomially distributed. Hence we can conclude
(cid:88)
v∈Vj
(cid:18)j
(cid:19)
i
E[Yi,j] =
E[Xv,i] =
nj
n
aD(1 − pi)j−i.
aD
We can now bound the probability that Yi,j deviates from its expectation by too
much. The maximum deviation that we want to allow is δi :=
2i+4D2i−1 . Since
a Chernoff Bound can only be effectively used if the expected value is relatively
large, we distinguish two cases and use a Chernoff Bound, if nj is relatively large,
and Markov's Inequality, if nj is small.
At first consider the latter case and assume nj ≤ n(1−p)j−i
i)2i+7D2i+1 , that is, it
(j
holds E[Yi,j] ≤ aD(1−pi)2j−2i
and hence in particular E[Yi,j] < δi. The latter
fact implies that Yi,j can deviate from its expectation by more than δi only by
exceeding it. Hence we can conclude
Pr[Yi,j − E[Yi,j] > δi] = Pr[Yi,j − E[Yi,j] > δi] < Pr[Yi,j > δi] ≤ E[Yi,j]
≤ 1
2i+7D2i+1
δi
8D2 ,
applying Markov's Inequality.
Now consider the first of the two cases, i.e., nj > n(1−p)j−i
. Since nj ≤ n, we additionally have E[Yi,j] ≤(cid:0)j
i)2i+7D2i+1 ; this means that
(j
· δi. Since the events Xv,i = 1 are independent for
E[Yi,j] > aD(1−pi)2j−2i
pi)j−i ≤ 2i+4D2i−1(j
all vertices v, we can conclude
i)(1−p)j−i
2i+7D2i+1
i
(cid:1)aD(1 −
(cid:35)
E[Yi,j]
(cid:33)
Pr[Yi,j − E[Yi,j] > δi] ≤ Pr
Yi,j − E[Yi,j] >
(cid:34)
≤ 2 exp
(cid:1)(1 − p)j−i
(1 − p)2j−2i
E[Yi,j]
2i+4D2i−1(cid:0)j
3 · 22i+8D4i−2(cid:0)j
(cid:1)2
(cid:33)
3 · 23i+15D6i−1(cid:0)j
(cid:1)2
3aD
2
i
i
(cid:32)
(cid:32)
−
−
< 2 exp
by a multiplicatice Chernoff Bound; hence, for sufficiently large a = O(cid:16) D8+13/D
i
(cid:17)
log1/D D
3/D
this probability does not exceed 1
8D2 .
Basically we are only interested in i-way collisions that occur on vertices of
Vi, since this is the set whose size we want to estimate in the iteration of the
for-loop that has the loop counter i. However, for vertices with a degree if j > i,
i-way collisions can also happen, which can distort the estimate for ni. Hence
we have to estimate the number of i-way collisions Yi,j on vertices in Vj, j > i,
that will happen in the iteration of the for-loop that has a loop counter of i.
We can then subtract these estimates from the number of collisions measured in
this iteration and in this way gain an estimator for ni. For estimating the Yi,j
we can simply use the estimators nj from the previous iterations of the for-loop.
In what follows we will assume that Yi,j − E[Yi,j] ≤ δi holds for i = 1, . . . D
and j ≥ i; by the Union Bound, the probability for this to happen is at least
8, since the probability that a single Yi,j deviates from its expectation by more
7
8D2 ist. We will now show by induction over
than δi is, as shown above, at most
i that, under these assumptions, ni deviates from ni by at most
1
n
2i+3D2i−2 .
For the base case we chose the iteration of the for-loop that has i = D as the
loop counter. It holds
(cid:18)D
(cid:19)
D
E[nD] =
E[cD]
pD
D
=
n
aD
D
E[YD,D] =
n
aD · nD
n
aD(1 − pD)D−D = nD.
i
n
n
n
Moreover we have assumed that YD,D deviates from its expectation by at most
δD =
2D+3D2D−2
from its expectation E[nD] = nD.
2D+4D2D−1 . Hence nD deviates by at most δDn
2D+4D2D−1 <
aD =
aD
Now let i < D be an arbitrary value of the loop counter. As induction
2j+3D2j−2 .
hypothesis assume that, for all j > i, nj deviates from nj by at most
Since E[Yi,j] = nj
n
(cid:1)aD(1 − pi)j−i holds for all j > i, we can conclude
(cid:0)j
ci
(cid:18)j
(cid:19)
− (cid:88)
(cid:19)
(cid:18)j
− (cid:88)
E[ci] − (cid:88)
− (cid:88)
− (cid:88)
i<j≤d
n
aD E[Yi,j]
E[nj](1 − pi)j−i
nj(1 − pi)j−i =
=
nj(1 − pi)j−i
=
(cid:18)j
E[ni] =E
=
=
E[Yi,j]
E[Yi,i]
i<j≤d
i<j≤d
i<j≤d
(cid:19)
= ni.
E[ci]
pi
i
E[ci]
pi
i
E[ci]
pi
i
pi
i
i
i
i
i<j≤d
pi
i
1
pi
i
It remains to bound the maximum deviation of ni from its expected value.
There are two types of deviation that can occur: The first one is that, for j > i,
nj may deviate from E[nj]; due to the induction hypothesis, this deviation is at
most
ni, the contribution of nj to the overall deviation is at most
2j+3D2j−2 , and since nj gets multiplied by(cid:0)j
(cid:18)j
(cid:1)(1 − pi)j−i when computing
(cid:19)
n
n
n
i
·
2j+3D2j−2 ≤ Dj−i ·
i
2j+3D2j−2 =
n
2j+3Dj+i−2 ≤
n
2i+4D2i−1 .
αj :=
The second type of error results from the deviation of at most δi that, for j ≥ i,
may occur between Yi,j and E[Yi,j]; due to this deviation, ci may deviate from
E[ci]. The contribution to the overall deviation is at most
βj :=
δi
pi
i
=
n
aD ·
aD
2i+4D2i−1 =
n
2i+4D2i−1 .
Adding the contributions of both types of deviation, we can bound the overall
deviation by
αj+βj ≤ 2Dn
2i+4D2i−1 =
n
2i+3D2i−2 ,
ni−ni = ni−E[ni] ≤ (cid:88)
(cid:88)
(cid:88)
αj+
βj <
i<j≤D
i≤j≤D
i≤j≤D
which completes the induction step.
Altogether we therefor have
(cid:88)
ni − ni ≤ (cid:88)
1≤i≤D
1≤i≤D
2i+3D2i−2 ≤ (cid:88)
n
and hence m deviates from m+
on the tightness of the Yi,j hold, i.e.,
16 Dn by at most
n
24 =
16
Dn,
1≤i≤D
16 Dn, if the above assumptions
m ≤ m ≤ m +
8
Dn.
Union Bound this leads to a success probability of at least 1−(cid:0) 1
The success probability of the algorithm results from the probabilities for
successfully sampling and for all Yi,j being tight to their expected values; by the
4.
The query complexity of the algorithm results from the number of edges that
are sampled; since these are at most O(Dain1−1/i) = O(−3D8D+14n1−1/i log D)
in the iteration of the for-loop that has the loop counter i, the overall number
of samples is at most O(−3D8D+15n1−1/D log D).
(cid:1) = 3
16 + 1
16 + 1
8
At the end of the section we give a brief overview over the proof sketch
for a property testing algorithm for strong connectivity that Oded Goldreich
published earlier and independently from our result ([9], Appendix of the survey
article Introduction to Testing Graph Properties). Goldreich at first notes that
√
n) on
the graph classes that Bender and Ron use for their lower bound of Ω(
testing strong connectivity [5] also yield a lower bound of Ω(n) for testing strong
connectivity with a one-sided error; our proof at the beginning of this section
also makes use of this observation, considering only the class of -far graphs
Bender and Ron give.
For testing strong connectivity, Goldreich's approach is the same as ours,
except for his choice of a reduction function: In particular, he also proposes to
at first solve the problem for graphs in which all source components have size 1
by computing i-way collision statistics for the target vertices of sets of sampled
edges for i = 1, . . . , D.
considers directed circles with a length of at most s =(cid:6) 4
Goldreich's reduction function then has a similar effect than ours: Goldreich
(cid:7), and, for each vertex
v, the set Cv of all those vertices that lie on such a circle together with v. If
Cu = Cv for all vertices u ∈ Cv, then Cv is contracted. This approach guarantees
that all sufficiently small source components are contracted to a single vertex in
the resulting graph, and the reduction function can be computed locally.
For Goldreich's reduction, the maximal number of vertices in a contracted
component is at most s2 and hence the maximum indegree of a vertex in the
D
resulting graph is t = s2·D ≈ 16
of O(n1−1/t) ≈ O(1 − 2D
leaving open any technical details.
2D . This leads to a slightly worse query complexity
16 ). Goldreich only gives a brief sketch of his result,
References
1. N. Alon. Testing subgraphs in large graphs. Random Struct. Algorithms 21, 3-4,
pp. 359-370, 2002.
2. N. Alon, E. Fischer, I. Newman, and A. Shapira. A combinatorial characterization of
the testable graph properties: it's all about regularity. SIAM Journal on Computing,
39(1): 143-167, 2009.
3. N. Alon, A. Shapira. Testing subgraphs in directed graphs. In: Proc. of the 35th
ACM Symp. on the Theory of Computing (STOC), pp. 700-709, 2003.
4. Z. Bar-Yossef, R. Kumar, D. Sivakumar. Sampling algorithms: Lower bounds
and applications. In: Proc. of the 33rd ACM Symp. on the Theory of Computing
(STOC), pp. 266-275, 2001.
5. M.A. Bender, D. Ron. Testing properties of directed graphs: acyclicity and connec-
tivity. Random Structures & Algorithms, Volume 20 No. 2, 2002.
6. I. Benjamini, O. Schramm, A. Shapira. Every minor-closed property of sparse graphs
is testable. In: Proc. of the 40th ACM Symp. on the Theory of Computing (STOC),
pp. 393-402, 2008.
7. B. Chazelle, R. Rubinfeld, L. Trevisan. Approximating the Minimum Spanning Tree
Weight in Sublinear Time SIAM Journal on Computing, 34(6): 1370-1379, 2005
8. A. Czumaj, A. Shapira, and C. Sohler. Testing hereditary properties of non-
expanding bounded-degree graphs. SIAM Journal on Computing, 38(6): 2499-2510,
2009.
9. O. Goldreich. Property Testing: Current Research and Surveys. Springer 2010.
10. O. Goldreich, S. Goldwasser, D. Ron. Property Testing and its Connection to
Learning and Approximation. J. of the ACM, 45(4): 653-750, 1998.
11. O. Goldreich, D. Ron. Property Testing in Bounded Degree Graphs. In: Proc. of
the 29th ACM Symp. on the Theory of Computing (STOC), pp. 406-415, 1997.
12. A. Hassidim, J.A. Kelner, H.N. Nguyen, and K. Onak. Local graph partitions for
approximation and testing. In: Proc. of the 50th IEEE Symp. on Foundations of
Computer Science (FOCS), pp. 22-31, 2009.
13. F. Hellweg, C. Sohler. Property Testing in Sparse Directed Graphs: Strong Connec-
tivity and Subgraph Freeness. In: Proc. of the 20th European Symp. on Algorithms
(ESA), pp. 599-610, 2012.
14. I. Newman, C. Sohler. Every property of hyperfinite graphs is testable. In: Proc.
of the 43rd ACM Symp. on the Theory of Computing (STOC) pp. 675-684, 2011.
15. S. Raskhodnikova, D. Ron, A. Shpilka, A. Smith. Strong Lower Bounds for Approx-
imating the Distribution Support Size and the Distinct Elements Problem. SIAM
Journal on Computing 39(3), pp. 813-842, 2009.
16. R. Rubinfeld, M. Sudan. Robust Characterizations of Polynomials with Applica-
tions to Program Testing. SIAM Journal on Computing, 25(2): 252-271, 1996.
17. Y. Yoshida, H. Ito. Testing k-edge-connectivity of digraphs. Journal of System
Science and Complexity, 23(1) pp. 91-101, 2010.
|
1007.1726 | 5 | 1007 | 2016-07-26T17:26:40 | Vsep-New Heuristic and Exact Algorithms for Graph Automorphism Group Computation | [
"cs.DS",
"cs.DM",
"math.CO"
] | One exact and two heuristic algorithms for determining the generators, orbits and order of the graph automorphism group are presented. A basic tool of these algorithms is the well-known individualization and refinement procedure. A search tree is used in the algorithms - each node of the tree is a partition. All nonequivalent discreet partitions derivative of the selected vertices are stored in a coded form. A new strategy is used in the exact algorithm: if during its execution some of the searched or intermediate variables obtain a wrong value then the algorithm continues from a new start point losing some of the results determined so far. The algorithms has been tested on one of the known benchmark graphs and shows lower running times for some graph families. The heuristic versions of the algorithms are based on determining some number of discreet partitions derivative of each vertex in the selected cell of the initial partition and comparing them for an automorphism - their search trees are reduced. The heuristic algorithms are almost exact and are many times faster than the exact one. The experimental tests exhibit that the worst-cases running time of the exact algorithm is exponential but it is polynomial for the heuristic algorithms. Several cell selectors are used. Some of them are new. We also use a chooser of cell selector for choosing the optimal cell selector for the manipulated graph. The proposed heuristic algorithms use two main heuristic procedures that generate two different forests of search trees. | cs.DS | cs | Vsep-New Heuristic and Exact Algorithms for Computing Graph
Automorphism Group and Graph Isomorphism
Stoicho D. Stoichev
Department of Computer Systems, Technical University of Sofia
email: [email protected]
Abstract. Five new algorithms, named Vsep, are described. Four of them are for
determining the generators, orbits and order of an undirected graph automorphism group.
Vsep-e – exact, Vsep-orb and Vsep-sch – heuristic and Vsep-a automatically selects the
optimal version among Vsep-e, Vsep-orb and Vsep-sch. The fifth algorithm, Vsep-is, is for
finding an isomorphism between two graphs. Vsep-orb firstly finds heuristically the
generators and orbits and then uses the exact one on the orbital partition for determining
the order of the group. Vsep-sch differs from Vsep-orb in using the Schreier-Sims
algorithm for determining the order of the group. A basic tool of these algorithms is the
adjacency refinement procedure that gives finer output partition on a given input partition
of graph vertices. The refinement procedure is a simple iterative algorithm based on the
criterion of relative degree of a vertex toward a basic cell in the partition. A search tree is
used in the algorithms - each node of the tree is a partition. All nonequivalent discreet
partitions derivative of the selected vertices called a bouquet are stored in a coded form in a
hash table in order to reduce the necessary storage – this is a main difference of Vsep-e
with the known graph automorphism group algorithms. A new strategy is used in the exact
algorithm: if during its execution some of the searched or intermediate variables obtain a
wrong value then the algorithm continues from a new start point losing some of the results
determined so far. The new start point is such that the correct results can be obtained. The
proposed algorithms has been tested on the nauy&Traces benchmark graphs and compared
with Traces, and the results show that for some graph families Vsep-e outperforms Traces
and for some of the others Traces outperforms Vsep-e. The heuristic versions of Vsep are
based on determining some number of discreet partitions derivative of each vertex in the
selected cell of the initial partition and comparing them for an automorphism, i.e. their
search trees are reduced. The heuristic algorithms are almost exact and are many times
faster than the exact one. The heuristic algorithms are good choice for the user because of
their smaller running times. Several cell selectors are used in Vsep, some of them are
known and some are new. We also use a chooser of cell selector for choosing the optimal
cell selector for the manipulated graph. The experiments show that the running time of
Vsep algorithms does not depend on the vertex labeling.
Key words: graph, isomorphism, automorphism, group, stabilizer, exact algorithm,
heuristic algorithm, partition, numeration, generators, orbits and order of the graph
automorphism group.
- 1 -
1. Introduction
We assume some familiarity with the basics in the design and analysis of algorithms[1–
3], combinatorial algorithms[4,5], graph theory and group theory [6–10]. We consider
simple finite undirected graphs (without loops and multiple edges). The graph is denoted
by G(V,E), where : V={1,2,3, . . . ,n} is the set of vertices and E - the set of edges (v,
w), v, w V. The number E of the edges of G we denote by k, k=O(n2). Our algorithms
are applicable to any undirected graph including disconnected graphs but for them there is
more efficient algorithm we do not describe here. The set of all vertices adjacent to a
vertex x is denoted by Adj(x). We use a static list representation of a graph (by two static
arrays) because of its least required storage and fastest operation of finding all adjacent
vertices of a given vertex compared with the adjacency matrix and the dynamic list
representation.
An isomorphism [4–12] between two graphs G1(V1, E1) and G2(V2,E2) is called one-to-one
correspondence (mapping) yi = f(xi) between the vertices of the graphs (xiV1, yiV2,
i=1,2,...,n, n=V1=V2) such that two adjacent (nonadjacent) vertices from one of the
graphs correspond to two adjacent (nonadjacent) vertices from the another graph, i. e.
every edge (non-edge) (p, q ) from graph G1 corresponds to an edge (non-edge) (f(p),f(q))
from graph G2 and vice versa. So, the isomorphism preserves the adjacency relation of
vertices - this kind of bijection is commonly called "edge-preserving bijection".
Two graphs G1 and G2 are called isomorphic (G1G2) if there is at least one isomorphism
between them. Otherwise they are non-isomorphic, G1!G2.
An automorphism [4–11,13] of a graph is an isomorphism of the graph onto itself. Or, an
automorphism h of graph G(V,E) is called one-to-one correspondence yi=h(xi) between
the vertices of the graph (xi,yiV, i=1,2,...,n) that preserves the adjacency of the vertices, i.
e. there is unique corresponding edge (non-edge) (h(p),h(q))E to each edge (non-edge)
(p,q)E. A fixed point x of an automorphism h is called a vertex x for which x = h(x).
Trivial automorphism is an automorphism h0 if each its vertex is a fixed point, xi=h0(xi),
i=1,2,...,n and a non–trivial automorphism is an automorphism for which there is at least
one pair of vertices x,y such that y=h(x) x.
Two vertices xi and yi in the a graph G(V,E) are called similar (or symmetric) [6,7], xiyi,
when they are corresponding, yi=h(xi), in some automorphism h. Otherwise, if xi and yi are
not similar we use the notation xi! yi. An automorphism h can be presented by two
sequences Па and Пb of graph vertex labels:
a1, a2,..., ai,... an=Па
b1, b2,..., bi,... bn=Пb, where the corresponding vertices are ai and bi=h(ai),
i=1,2,...,n.
The automorphism h may be shortly written in the form Пb=h(Па). The automorphism can
be presented by n! pairs of rows - each row being derived from the other by transposing the
positions of the pairs of corresponding vertices. The corresponding pairs of vertices can be
set on to any place of the rows, but it is possible the place to depend on the sorting criterion
which does not depend on the vertex labeling. Each automorphism can be written uniquely
only with the permutation Пb if we assume that Пa = 1, 2, 3, … n. Even a simpler notation
- 2 -
called cycle notation [10] is often used. In a cycle (x1,x2, . . . ,xi, xi+1, . . . ,xp) xi maps to xi+1,
1 i p-1 and xp maps to x1. For example,
h=(1,8)(2,6,3,7)(4,5) =
1 2 3 4 5 6 7 8
8 6 7 5 4 3 2 1
Any automorphism can be written as the product of disjoint cycles and the product is
unique up to the order of the cycles [10]. The cycles of length 1 are omitted.
The operation function composition (or superposition) [4–11,13] of two automorphisms
. This
is a consecutive execution of and , i. e.
and
operation is usually called multiplication and is denoted by juxtaposition γ=..
The set of all automorphisms of a given graph G(V,E) form a graph automorphism group
(under the operation function composition of automorphisms) denoted Aut(G), shortly
A(G) or A ) [4–11,13]. The trivial automorphism is the identity of the group – we denote it
by I. The number of the automorphisms in A(G), Aut(G), is called an order of the
automorphism group. An order of an automorphism is the order of the cyclic group
generated by this automorphism. If the automorphism is written in cycle form then its order
is the least common multiple of the cycle lengths[10].
The subset gen(А)А, denoted gen(А) or <g1,g2,...,gd>, is called a generating set
[4,5,9,10,12] of the automorphism group if every automorphism hA can be written as a
finite product of elements
. The generating sets are not unique.
The subset of the graph vertices y similar to vertex x in any automorphism hAut(G),
y=h(x), is called an orbit of x, denoted Orb(x,A): Orb(x,A)={h(x)hAut(G)} and Orb(x,A)
is read as 'orbit of a vertex x under a group A'. By Orb(A) we denote the set of all orbits of
the group A.
A graph with only one orbit of Aut(G) is called transitive, and a graph is called rigid or
asymmetric if each of its vertices is in a separate orbit, i.e. Aut(G)=1. A stabilizer
А(х1,х2,...,хi) or point-wise stabilizer [5,10,13] is the subgroup of the automorphism group
A that contains only the automorphisms with fixed points х1,х2,...,хi.
There is a theorem called orbit-stabilizer theorem[5–8] for computing the order A of the
automorphism group of graph G(V,E). We denote it O-S theorem. The theorem claims that
A=A(х1)*Orb(х1,A)=A(х1)*d, where Orb(х1,A)={ х1, х2,..., хd} is the orbit of vertex x1
under the group A, and A(x1) is a stabilizer of a representative x1 of this orbit. In other
words, the order of the graph automorphism group is equal to the product of the order of a
stabilizer of one of its vertices and the length of the orbit of this vertex. Vsep algorithms
and other known graph isomorphism and automorphism algorithms use the orbit-stabilizer
theorem.
The stabilizer Ai= А(х1,...,хi) is called an ancestor stabilizer to the stabilizer Aj= А(х1,...,хj)
if j>i and Aj is called a successor (or descendant) stabilizer to Ai. Obviously, the orbits of
the successor stabilizer are subsets of the orbits of the ancestor stabilizer and the order of
the successor stabilizer is a divisor of the order of the ancestor stabilizer.
Given a set SV a set-wise stabilizer [10,12], denoted Aut(G,S), is a subgroup of Aut(G)
defined by:Aut(G,S)={Aut(G) x,(x)S}.
The graph isomorphism (GI) problem consists in deciding whether two given graphs are
isomorphic, i.e., whether there is an isomorphism between them. The graph isomorphism
- 3 -
...........XY...........YZ.............XZ)(,1Agenggii problem belongs to the class of NP problems [4–9,11,12,14], and has been conjectured
intractable, although probably not NP-complete.
It is neither known that this problem has polynomial time complexity nor to belong to the
class of NP–complete problems. Its unknown complexity status is a theoretical and
practical challenge. Two problems are called polynomial-time equivalent if there is a
polynomial time algorithm that transforms one of the problems to another [4,7,9,10]. The
problems [3,9,10] that are polynomial-time equivalent to graph isomorphism are called
graph isomorphism complete. The following problems are graph isomorphism complete
[15,16]: given two graphs G1 and G2: existence of isomorphism of the graphs; determine
the isomorphism of the graphs if it exists; determine the numbers of the isomorphisms
from G1 to G2; given a graph G: determine the generating set gen(Aut(G)), determine the
orbits Orb(Aut(G)), determine the order Orb(Aut(G)). The problems of determining the
orbits and order of the graph automorphism group A are also called automorphism
partitioning and automorphisms counting, respectively. The generators, orbits and order of
a graph G automorphism group A we denote by GOO(A) or GOO(Aut(G)) and by GO(A) -
the generators and orbits of the group. A class of graphs is called graph isomorphism
complete if the recognition of isomorphism for graphs from this class is a graph
isomorphism complete problem. The following classes are graph isomorphism complete
[16]: connected graphs, graphs of diameter 2 and radius 1, directed acyclic graphs, regular
graphs, bipartite graphs without non-trivial strongly regular subgraphs, bipartite Eulerian
graphs, bipartite regular graphs, line graphs, chordal graphs, regular self-complementary
graphs, etc. However, there are special cases of the graph isomorphism problem with
polynomial-time complexity: planar graphs (linear time), trees have a particularly simple
algorithm, graphs of bounded degree, interval graphs, permutation graphs and convex
graphs. At present it is not known a polynomial time algorithm for solving the graph
isomorphism complete problems in the worst-case [14,17] – all known algorithms have
exponential or moderately exponential worst-case time complexity. Graph isomorphism
problems are of great practical and theoretical interest [3–5,12,14,17]. Recently, Laszlo
Babai has claimed that the Graph Isomorphism problem can be solved in quasipolynomial
time [18]. There are several practical algorithms (their names are shown below in
parentheses ) for graph isomorphism and graph automorphism group, due to Brendan
McKay (nauty), Adolfo Piperno (Traces), William Kocay (Groups&Graphs), Schmidt and
Druffel, Jeffrey Ullman; L.P. Cordella, P. Foggia C. Sansone and M.Vento (VF2), Tommi
Junttila and Petteri Kaski (bliss), Hadi Katebi, Karem A. Sakallah, and Igor L. Markov
(saucy), Jose Luis Lopez-Presa, Luis Nunez Chiroque, and Antonio Fernandez Anta
(conauto), G. Tener and N. Deo (nishe), Nechepurenko [4], etc. There is a comparison in
[13,19] on running times of the following tools: nauty, Taces, bliss, conauto and saucy.
Their running time on random graphs is quite well but a major problem of these algorithms
is their exponential time performance in the worst- case. There are two main
generalizations of the graph isomorphism: subgraph problem (given two graphs determine
if one of them is a subgraph to another) and largest common subgraph problem (given two
graphs determine the common subgraph to both that has the maximum number of vertices
or edges). These two problems are NP-complete and have many applications and the
algorithms for them use the graph isomorphism and automorphism algorithms as basic
- 4 -
tools. Another important problem related to graph isomorphism is [5,12]: Compute a
complete invariant (certificate, signature) f for G, i.e. for all graphs G and H, G H
f(G) = f(H) (graph certificate identifies a graph uniquely up to isomorphism).
The graph isomorphism related problems (graph isomorphism itself, GOO(Aut(G))),
subgraph isomorphism, largest common subgraph, graph certificate and canonization of a
graph) arise in such fields as mathematics, chemistry, information retrieval, linguistics,
logistics, switching theory, bioinformatics, and network theory [4,20].
Our goal is to develop exact and heuristic algorithms for determining GOO(Aut(G)), i.e. to
solve the three problems by one algorithm with time complexity as lower as possible. In
addition, our requirements to the heuristic algorithms are to give results equal to the results
of the exact algorithm with the probability close to 1. There are a few heuristic algorithms
for the graph isomorphism problem [21–24] – there is no access to their program codes.
We propose three new heuristic algorithms (Vsep-orb, Vsep-hway, Vsep-sch) for
GOO(Aut(G)) with much lower polynomial time complexity than the exact one. The
experiments show that they are many times faster than the exact algorithm - even for
difficult graphs with large sizes they give correct results. One of the first step in the exact
algorithm is a call to a heuristic procedure for determining a representative of one of the
smallest orbits of Aut(G) as a starting selected vertex – this way we speed up the exact
algorithm and reduce the required storage.
2. Partitions and a refinement procedure
An ordered partition (or simply partition) П or П(G)
П = C1C2…Ci…Cp = C1C2...Ci...Cp
of the vertices of graph G(V, E) [5,12,15] is a sequence of disjoint non-empty subsets of V
whose union is V. П = x1,1, x1,2, …, x1,k1 … xi,1, xi,2, …, xi,ki … xp,1, xp,2, …, xp,kp is a
detailed presentation of П, where Ci= xi,1, xi,2, …, xi,ki , xi,j ϵ V, i=.1,…,p, j=1,…,ki. The
subsets C1, C2,… , Cp, are called cells (classes, blocks).
We denote the number of the cells in a partition П by П. Two cells are called adjacent if
there is at least one edge between their vertices, i.e. cells Ci,Cj П are adjacent if there is at
least one edge (x,y), x Ci, y Cj. A cell with cardinality one is called trivial (or
singleton). The vertex of such a cell is said to be fixed by П or it is called a fixed point of
П. A partition, of which each cell is trivial is called discrete or numeration having in
mind that in fact it is a permutation that can be viewed as a graph vertices renumbering –
vertex i corresponds to vertex x that is on the position i in the partition. By NC(x, П) we
denote the index of the cell C of П that contains vertex x, x C. The position (index) of a
vertex x in the partition (or in the cell) we denote by pos(x). The relative degree ρ(x,Ci) of
a vertex хСj toward a cell Сi is equal to the number of vertices of cell Сi adjacent to
vertex x. We denote by ν(x, П) a cell-degree vector defined as ν(x, П)=( ρ(x,Ci), i=1,…,p),
p=П – it is a vector whose components are the relative degrees of x to each cell in П. We
call a partition П stable (or equitable) if the cell-degree vectors ν(x, П)= ν(y, П) for each
two vertices x, y Ci, where Ci is any cell in П. We say that the partition П2 is finer than
П1, written П2 ≤ П1, if for every cell Ci П2 there exists a cell Cj П1 such that Сi Cj. In
order to get a finer partition П2=D1D2...Dq when given П1=С1С2...Ср , a refinement
- 5 -
procedure (RP, П2= RP(П1)) is used, that assign to each vertex х V a sorting criterion
according by which the vertices of each class СiП1 are sorted out in increasing or
decreasing order of their criterions [4,5,12]. Often a sorting criterion is the relative degree
ρ(x,W) of any vertex х П toward some cell WΠ.. The refinement procedure that uses
this criterion is called adjacency refinement procedure. We use only this version of the
procedure. Example of another criterion for sorting is the number of the subgraphs of a
given type (for example a triangle) that contain vertex x.
Two partitions Π1 and Π2 of the vertices of graph G(V,E) are called compatible [40] if: (1)
Π1 = Π2 = m; (2) if Π1 = W1W2…Wm and Π2 = U1U2…Um, then for all i ϵ [1:m], Wi =
Ui; (3) for all x,y ϵ V, NC(x,Π1) = NC(y,Π2) implies ν(x,Π1) = ν(y,Π2). Similar
definition is valid if Π1 and Π2 are for different graphs.
Two partitions П1=С1С2...Ср and П2=D1D2...Dq are called equivalent if there is an
automorphism αAut(G) such that NC(x,П1) = NC(y,П2) for each pair of vertices x, y=
α(x). In other words, the similar vertices are in cells with the same label of П1 and П2).
Obviously for the equivalent partitions П1 and П2 we have: p = q and Ci = Di, i[1:p].
We denote the equivalent partitions by П2 = α(П1) or П2 П1. Evidently the equivalence
relation is transitive. Two discrete partitions П1 = a1a2...ak...an, П2 = b1b2...bk...bn of graph
G(V,E) vertices are called equivalent if they form an automorphism h Aut(G), bk = h(ak),
k = 1,2, ... ,n. The testing if two discrete partitions П1 and П2 form an automorphism is a
basic operation in our algorithms.
Given sequence of equivalent discrete partitions П1, ... , Пm, we store one of them, for
example П1 and the orbits derived by the automorphisms Пi = i(П1), i=2, ... ,m.
A partition-wise stabilizer A(G, П) is defined by: A(G, П) = {Aut(G) x,(x) Ci,
where Ci is any cell of П} [12]. This is a subgroup of Aut(G) such that each automorphism
αAut(G) belongs to Aut(G, П) if to any vertex x of any cell of П corresponds a vertex
y=α(x) from the same cell. The orbits of A(G, П) are subsets of the cells of П as we'll see
below. If П= Пu then A(G, Пu) = Aut(G).Algorithms for determining the graph
automorphisms and isomorphisms use very often the refinement procedure - each cell of its
output partition contains at least one orbit of a graph automorphism group or its stabilizer.
It is still not known a refinement algorithm that gives output partition each cell of which
coincides with an orbit (orbit partition or automorphism partition) [20] on unit input
partition.
One of the most efficient RP is the RP with a base cell. The RP with a base cell sorts
(counting sort [1,2]) the vertices of any cell Cj according to their relative degree rdg(x,i)
toward a selected base cell Ci. The sorting continues until it reaches a partition П2 in which
there is no cell that can be divided into subcells toward any base cell – such partition is
called stable (equitable) [5,12] as we noted above: it holds the property П1=RP(П1).
Vertices in every cell of the stable partition have the same sorting criterion - in our case,
the same relative degree toward each cell. The base cell refinement algorithm has time
complexity О(к.log n), where k and n are respectively the number of edges and the number
of vertices of the graph [25,26]. The base cell RP refinement algorithm (Figure. 1) uses
the counting sort that does not use a comparison operation – it sorts integers (the relative
degree of the vertices) within the range 0 to some integer. In our algorithms for
- 6 -
GOO(Aut(G)) we use only the adjacency refinement procedure with a base cell (Figure. 1)
[25] and a W queue that contains the labels of all not selected cells as base cell. It differs
from the known refinement algorithms [5,12,20] in the way the base cell is selected. After
the current base cell ends sorting the adjacent cells as a new base cell is selected the first
new smallest subcell and if there is no such subcell the label of the new base cell is taken
from a W queue - the first cell label in W queue. There is a property that speeds up RP
procedure: the label of the new largest subcell of C cell (adjacent to B cell) is not included
in W queue if C cell label is not in W. There is a version of RP refinement algorithm with
base cell that always takes the new base cell from the queue.
The individualization-refinement operation (denoted IR), used in the known GA
algorithms, has two steps: individualization and refinement. Given an equitable partition π
and a vertex x at the individualization step a new partition π1 is obtained: the cell C(x) of π
with index i is divided into 2 subcells: {x}-the first subcell, with index i and the second
subcell {C(x)\{x}} with index i+1, other cells of π are not changed. At the refinement step
the partition π1 is refined with the refinement procedure RP obtaining a new equitable
partition π2=RP(π1) finer than π1. Given a partition π and a vertex x we denote by
IR(π,x,brcl) the resulting partition π2 from the application of IR operation on π and x, where
brcl is the number of cells of π2.
Input: graph G(V,E); П1 is the input partition on the graph vertices, W is a queue of
some cells of П1,BRCL-the number of cells of П1
Output: a better stable partition П2 ( П2 П1), BRCL-the number of cells of П2
S1: П2:= П1;
S2: Base cell В:= first cell in W; Delete the label B from W;
S3: repeat{each the loop execution is performed for different base cell B}
S4: Each cell С П2 adjacent to B is divided into subcells according to its relative
vertex degrees toward В;
S5: if there are new subcells from S4 then В:= the label of the subcell with
minimum length. If there are more cells with minimum length then the one with the
smallest label is chosen. Put into W the labels of the new subcells in П2, excluding
the label of one of the largest subcell if it is not in W; Delete the label B from W;
S6: if there are no new subcells from S4 then В:= first cell label in W;
S7: until there are no new subcells from S4 and W becomes empty
Figure. 1. RP refinement procedure with base cell B
Cells in our algorithms are not consecutively labeled by 1, 2, … . The label NC(Cj,П) of a
cell Cj in the partition П is determined by
, i. e.
the label of the cell Ci (respectively of each of its vertices ) is the first vertex index in the
cell, or it is greater by 1 than the number of the vertices in all cells preceding Ci in the
partition. NC(х,П) denotes the label of the cell that contains vertex x. This way of labeling
is time saving because changing the labels of a given cell does not cause change of the
labels of other cells.
- 7 -
111)),(ijjiCПCNC Two cells of a partition are called non-trivially joined (have non-trivial join, non-uniformly
joined) if the number of edges between them is greater than 0 and less than maximum
possible.
A channel of a cell C (new notion), denoted as Ch(C)), is the number of the edges
adjacent to any vertex in C cell. A channel of two cells C1,C2, denoted as Ch(C1,C2)), is the
number of all edges (x,y) between any vertex xC1 and any vertexC2. Example: Let
П=1,84,52,3,6,7=C1 C2 C3 be a partition on the vertices of the graph on Fig. 2. The
the cells are: Ch(C1,C1)=0, Ch(C1,C3)=4, Ch(C1)=4, Ch(C3,C3)=2,
channels of
Ch(C3,C2)=4, Ch(C3)=10, Ch(C2,C2)=1, ChC(C2)=5. Only the edges of cells with non-
trivial join to C are included in Ch(C). For example, the cells D5 and D6 of the partition
П=18546, 72, 3=D1 D2 D3 D4 D5 D6 of the vertices of the above graph have channels
Ch(D5)=1, Ch(D6)=1 since the edges incident to vertices from D5 and D6 and connected to
vertices from the trivial cells D1 D2 D3 D4 are not included – only the edge (6,7) is included
in Ch(D5) and (2,3) in Ch(D6).
A channel vector of a cell C П, ChVect(C), is a vector VC whose component VC(i) is
equal to Ch(Ci,C) channel where Ci is a cell П with non-trivial join to C. A channel
graph of a partition П, ChG (П), is an weighted graph with loops: each vertex of ChG
corresponds uniquely to a cell of Π and its weight is the channel of the cell; each edge of
ChG corresponds to a channel of the corresponding cells of Π and the weight of this edge
is equal to the weight of this channel. Similar notion is a quotient graph in [27]. A selected
non-trivial cell, SC(П), of a partition П is the cell Cj, Cj>1, that is selected by a defined
criterion (often it is called target cell, for example in [12]). The procedure that finds the
target cell is called cell selector.
Fig. 2. GIRA graph
The selection of the target cell has strong influence on the search tree size (see the next
section) and consequently, on the running time of the GA algorithms.
In Vsep algorithms, at given an equitable partition π we use the following 5 cell selectors
(their numbering is denoted by izb):
izb=1, MXNACS1: (a) it selects the nontrivial cell C of π with maximal number of
adjacent non-trivially joined cells to C, and (b) if there are several such cells it selects
the first one of the smallest size;
izb=2, MXNACS2: it differs from (1) only in (b) selecting the first cell with largest size.
izb=3, MXVECTCHVAL: it selects the first nontrivial cell C of π with the largest
vector of relative degrees of C to other cells – the vector is considered as a number with
leftist least significant digit.
- 8 -
izb=4, MXPRCHVAL: it selects the first nontrivial cell C of π with largest channel.
izb=5, ICLMXBRCL: it selects the first non-trivial cell C with maximal number of cells
of the derived partition π(x), x ϵ π, obtained after IR.
To select the optimal cell selector we use a cell selector chooser (CSLCh) – it chooses the
cell selector by three criteria: a) the smallest maximal level LMAX of the first path in the
search tree., b) the smallest product of the sizes of all target cells of the first path in the
search tree and c) the smallest product of the numbers of the cells of all partitions on the
first path in the search tree.
The partitions in the developed algorithms can be:
1. Unit partition (denoted by Пu=V=1, . . ., n) - all its vertices are in one cell С1 and the cell
number is NC(i)=1, i=1, . . .,n.
2. Еquitable partition - it is obtained as a result of the execution of the refinement procedure
(RP) on given input partition (see the definition given above);
3. Transformed partition ПТ is obtained from a stable partition П by separating from SC(П)
given vertex x in a singleton cell. The difference between ПТ and П is that SC(П) with cell
label i, is divided into two cells in ПТ: cell С'={x} with label i containing vertex x and cell
С''= SC(П)\{x} with label i+1 containing the other vertices of SC(П). The transformation
operation is denoted by TR: ПТ=TR(x, П) and is called an individualization in many
references.
The following theorem holds for the RP.
Theorem 1
Given a graph G(V,E), two different equivalent partitions П1 and П2 on V and a non-trivial
automorphism fAut(G). Then, NC(x,Па) = NC(y,Пb) for each pair x, y=f(x), where
Пa=RP(П1), Пb=RP(П2).
In other words, RP gives output partitions where the similar vertices remain in cells with
the same label.
This theorem is equivalent to the Theorem 7.1 in [5] – for equivalent partitions under an
isomorphism of two graphs and the statement that the RP is invariant under an
automorphism [28]. This is the reason for not giving here the proof of the theorem.
Corollary 1 of Theorem 1 Let П be a partition on V of a graph G(V,E) with an
automorphism fAut(G) and let for each pair of similar vertices x, y=f(x) the property
NC(x,П) = NC(y,П) holds. Then, NC(x,Пс) = NC(y,Пс), where Пс=RP(П).
Corollary 1 means that the similar vertices in any cell of a given input partition remain
in one cell of the resulting partition of RP.
Proof Let П=П1=П2
NC(x,Пс)=NC(y,Пс) is obtained by replacing Пс in NC(x,Па)=NC(y,Пb).
theorem 1. Then, we obtain Пс=Па=Пb=RP(П) and
in
Corollary 2 of Theorem 1 The refinement procedure RP is invariant under the vertex
orbits of a subgroup HА=Aut(G(V,E)): if Orb(x,H)C(x,П) for each vertex хV of a
given partition П, then Orb(x,H)C(x,Пc) for a partition Пc=RP(П). (In other words, if
- 9 -
each orbit of a given automorphism subgroup is a subset of a cell of the input partition П,
then the same is true for the output partition Пc=RP(П). Obviously, this corollary holds for
any stabilizer А(х1,х2,...,хi) of the automorphism group.)
Proof The condition Orb(x,H)C(x,П) means that NC(x,П)=NC(y,П), y=f(x) for each f
Н, i. e. the condition of Corollary 1 of Theorem 1 holds for each f Н and from it follows
NC(x,Пс)=NC(y,Пс), y=f(x) for each f Н, i.e. Orb(x,H)C(x,Пс).
Corollary 3 of Theorem 1 Orb(x,A)C(x,П1) holds for each vertex xV of the initial
partition П1=RP(Пu) of the graph G(V,E) where Пu is the unit partition.
Proof Let H=A and П=Пu
Orb(x,A)C(x,П1) follows from Corollary 2 where П1=Пс=RP(Пu).
for which obviously Orb(x,A)C(x, Пu). Then,
3. Auxiliary algorithms
3.1. A1 algorithm for determining one discrete partition
The output of A1 algorithm (Figure 3) is a series of better partitions the last of which is
discrete on given input partition П. Using A1 we introduce new notions and prove a
property used for speeding up our algorithms.
The following basic operations are performed in the A1 algorithm: {П1=RP(П), L=1};
{ determine SC(ПL) by some cell selector; select any vertex ХL in the SC(ПL); L:=L+1;
ПL:=IR(ПL-1, XL-1)}; L=2, . . . , LK, where ПLK is discrete and LK<n.
Given any input partition П the initial partition П1=RP(П) is determined first. In most cases
we'll consider that the starting input partition П is the unit partition, П= Пu.
A selection level is called the successive number L of the refinement procedure RP calls for
obtaining the partition ПL in A1 algorithm. ПL=П(х1,х2, ... , хL-1) denotes the partition that
is obtained by making L-1 selections х1,х2, ... , хL-1 starting with the partition П1. The
selected non-singleton cell in partition ПL is denoted by SC(ПL) or SC(х1,х2, ... , хL-1). Five
operations (rows 4, 5) are executed in the loop C: determining SC(ПL), a selection of
vertex xL in SC(ПL), a move to next level (a forward move, L+1), obtaining the
T) (refinement with RP) – the last two
transformed partition ПL
operations are IR . Obviously, ПL is a stable partition. The determination of SC(ПL) is
made by some cell selector. This execution continues until a discrete partition ПLK is
obtained – its level of selection LK is called terminal or final. The process of obtaining the
sequence П1, П2, . . . , ПLK is called series of forward moves (SFM).
The selection of a vertex also is made by some criterion but here we consider (for
definiteness) the selection of the first vertex in SC( ПL).
The five instructions (lines 4, 5) we denote by CSVSIR (cell selection, vertex selection,
individualization and refinement) – cell selection in ПL-1, vertex Xl-1 selection in SC(ПL-1),
L=L+1, individualization of XL-1 (ПL:=IR(ПL-1,XL-1)) and refinement of ПL
Each partition ПL, L=1, 2, ... , LK is called a partition-child of ПL-1 (partition-parent) and
a successor of each of the partitions ПL-1, ПL-2, ... , П1, П0 (partitions-ancestors of ПL).
T:=TR(xL-1,ПL-1), ПL:=RP(ПL
T.
- 10 -
The selected cells SC (L), L=1, 2, . . ., L-1 for a given partition ПL are called supporting SC
for ПL, and the selected vertices х1,х2, ... , хL-1 of ПL are called base points or a base of ПL
(or supporting points).
A search tree (ST) is an oriented tree whose root represents the initial partition П1=RP(П).
Each node of the search tree corresponds to an equitable partition ПL. Each arc coming out
of a node ПL corresponds to a selected vertex from the SC(ПL). Thus, the partition ПL+1 of a
particular node q on a level L+1 could be obtained from the partition ПL (node p on a level
L) by selecting a vertex х SC (ПL), x being depicted on the edge p-q in ST. Let r =
SC(ПL), then each node ПL=П(х1,х2, ... ,хL-1) with SC (ПL) = {a1a2 ... ar} has r nodes-
children of level L+1:
П(х1,х2, ... ,хL-1,a1), П(х1,х2, ... ,хL-1,a2), ... , П(х1,х2, ... ,хL-1,ar). Thus, each vertex in each
SC(ПL), for each level L is selected in ST. Only in П1 one vertex is selected – the first
vertex in SC(П1).
The just described search tree ST may be called a full search tree in contrast to the reduced
Input: a graph G(V,E), a partition П of V
Output: Series of better partitions П1<П2<…< ПL< ..< ПLK(discrete)
1: L:=1; ПL:=П; ПL:=RP(ПL); Output(ПL);
2: if ПL is discrete then message; return end if
3: repeat {loop C}
4: determine SC(ПL) by some cell selector; select a vertex ХL in the
SC(ПL)
5: L:=L+1; ПL:=IR(ПL-1, XL-1,brcl); Output(ПL);
6: until brcl=n
Figure 3. A1 algorithm
search tree RST, in which some vertices from SC(ПL) are not selected according to a
certain criterion. In the reduced search tree, for each orbit of а stabilizer AL-1=А(х1,х2, ...
,хL-1) only one representative of the orbit(AL-1)SC(ПL) is selected – this is the first
orbits vertex met during the traversal of the SC(ПL). This way of a selection of vertices
is used in Vsep algorithms. It is based on the theorems proved below.
The reduced search tree used in Vsep algorithms is not explicitly presented in the
algorithm – only the partitions of the path from the root to the current ПL are stored, i.e.,
the sequence П1, . . . , ПL.
3.2. A2 algorithm for generating the full search tree
A2 algorithm (Figure 4) generates all partitions of the search tree on a given input
partition П of graph G(V,E) and the selection of a vertex Х1 SC(L) of the initial
partition П1 = RP(П). The leaf nodes of ST are discrete partitions. A new partition ПL (a
new node of ST) is obtained after each execution of the instruction 10. In A2, each of
the vertices in every SC(ПL) is successively selected and after that the algorithm
continues with a forward move until a discrete partition is obtained. There are two loops
- 11 -
in A2: the loop C1 (lines 4-15) performs a forward move and the loop C2 (lines 9-13)
performs the backward move. In the forward move, as in A1 algorithm, we begin from
the partition ПL (on a level L) and selected vertex xSC(ПL), and we obtain
successively the partitions ПL+1, ПL+2, ... , ПLK (discrete). The backward move (L:=L-1)
is made when all vertices in the SC(L) are selected. We introduce a new notion – a level
of the last backward move, denoted by LP and meaning the last value assigned to L by
the instruction 10 (L:=L-1), i. e. LP is the level to which the last backward move is
made. We call LP a working level and SC(ПLP) - a working cell since the main
operations of the algorithms are with the vertices in SC(ПLP). Level LP is a significant
variable for Vsep-e algorithm. A sequence of forward moves (SFM) starts from the
level LP. A2 algorithm (as well as Vsep-e algorithm) generates the search tree in
preorder [1,2]– first visiting the root (a vertex SC(ПL) then visiting the subtree of
each vertex SC(ПL+1) from left to right.
Input: a partition П of vertices of G(V,E) graph
Output: all partitions –nodes of the full search tree
1: L:=1;ПL= П; ПL:=RP(П); Output ПL;
2: if ПL is discrete then message ' the graph is rigid'; return end if
3: determine SC(П1) by the cell selector; select a vertex X1 SC(ПL);
4: do { C1: loop of the forward move}
5: L:=L+1; ПL:=IR(ПL-1, XL-1,brcl); Output ПL;
6: if ПL is not discrete then
7: determine SC(ПL) by the cell selector; Х L:=first vertex in the SC(ПL)
8: else
9: repeat{ C2: loop of the backward move}
10: L:=L-1;
11: if L=1 then return end if;
12: Х L:=first unselected vertex in SC(ПL);
13: until Х L 0;
14: end if
15: end do
Figure 4. A2 algorithm
A bouquet В(хL) or В(х1,х2, ... , хL) of a selected vertex хLSC(ПL) is called the set of
all mutually non-equivalent discrete partitions derived from the partition ПL and the
selection of хL, i.e. obtained from П(х1,х2, ... , хL)=ПL+1. Only the first numeration
obtained by Vsep-e algorithm is stored from the set of equivalent discrete partitions and
is used as a representative in В(хL). The bouquet В(х1,х2, ... , хL) consists of the
bouquets of the orbits
representatives of vertices in the SC(ПL+1), i. e. it consists of the bouquets В(х1,х2, ... ,
хL,
are orbits representatives of the stabilizer А(х1, х2, ... , хL) in SC(L+1). Given the
), . . . , В(х1,х2, ... , хL,
), В(х1,х2, ... , хL,
), where
, . . . ,
,
- 12 -
11La21LapLa111La21LapLa1 selected vertices х1, х2, . . . , хL-1, хL the following relation holds for the bouquets: В(х1)
В(хL), i.e., the bouquet of a selected vertex хL is included
. . . В(хL-1)
В(х2)
Т,Qi
Т), Рi=RP(Рi
Т).
Т and Qi
Т) =NC(y, Qi
Т the property NC(x, Pi
in the bouquets of all preceding selected vertices.
Theorem 2 Given a graph G(V,E), a partition П on V, an automorphism fAut(G, П),
NC(x,П) = NC(y,П) for each pair x, y=f(x) (a property- equal cell labels for each pair
of similar vertices), two partitions PL=П(а1, а2, . . . , аL-1 ), QL= П(b1, b2, . . . , bL-1 )
obtained by 2 series of executions of CSVSIR on the input partition П and
NC(ai)=NC(bi) for bi = f(ai), i=1,2, . . . , L-1 (equal cell labels of corresponding base
vertices). Then, NC(x, PL) = NC(y, QL) for each pair x, y=f(x), x PL, y QL (the
property holds for each pair of similar vertices).
Proof (by the mathematical induction): (1). The base case: By assumption the property
NC(x, П) = NC(y, П) holds for the initial partition П. In particular, NC(a1, П)=NC(b1,
П).
(2). The inductive step: Let the property holds for i-1: NC(x, Pi-1) = NC(y, Qi-1) for
Т) holds since by
each pair x, y=f(x). For Pi
assumption NC(ai-1)=NC(bi-1) for ai-1,bi-1=f(ai-1) is true. For all other vertices of the cells
С(ai-1) and С(bi-1) the cell labels are equal to NC(ai-1)+1 – this does not violate the
property. For the vertices of other cells of Pi-1, Qi-1 there is no change of cell labels in
Т
Pi
i. Hence, according to Corollary 1 of Theorem 1 the property NC(x, Пi) =NC(y,
Pi) also holds for Пi=RP(Пi
Since both the base and the inductive step have been proved, it has now been proved
that NC(x, ПL) =NC(y, PL) holds for all vertices x and y=f(x), x PL, y QL.
Corollary of Theorem 2 Given a partition ПL=П(х1,х2, ... ,хL-1). Then, Orb(x, AL-1)C(x,
ПL) for each vertex хV of the graph G(V,E), L=1, . . . , LK. (In other words, the orbits
of AL-1 are subsets of the cells of ПL).
Proof Let bi=ai, i=1,2, . . . , L-1 in Theorem 2. Thus, these L-1vertices become fixed
points of automorphism f and PL=QL. From NC(x, ПL) = NC(y, ПL) for each pair x, y =
f(x) and for each fАL-1=А(а1, а2, . . . , аL-1) follows Orb(x,AL-1)C(x,ПL).
Applying the corollary for L=1 we have Orb(x, A0)= Orb(x, A(П)) C(x, П1). Applying
it to L=LK we obtain Orb(x,ALK-1)C(x,ПLK) and Orb(x,ALK-1)C(x,ПLK)=1 since
ПLK is discrete, i.e., Orb(x, ALK-1)=1 and ALK-1={I}. Thus, for the series of partitions
П1, . . . , ПLK only the stabilizer ALK-1 is known – the stabilizer of the first selected
vertex XLK-1 in SC(ПLK-1) .
Theorem 3 Given a graph G(V,E), n=V, an automorphism fAut(G) with fixed points
а1, а2, ... , аj-1 and a partition Пj=П(а1, а2, ... , аj-1), obtained after (j-1) successive
applications of CSVSIR operation on а1, а2, . . . , аj-1 with starting input partition П and
vertices
p, q SC (Пj), q=f(p).
Then, each numeration
- 13 -
П′=П(а1, а2, ..., аj-1, p, аj+1, аj+2, ..., аLК-1)=x1 x2 . . . xi . . . xn,
derivative of П(а1, а2, ... , аj-1,p) has a corresponding numeration
П″=П(а1, а2, ..., аj-1, q, bj+1, bj+2, ..., bLК-1)= y1 y2 . . . yi . . . yn,
derived from П(а1, а2, ... , аj-1,q), such that yi=f(xi), i=1, 2, ... , n and
NC(xi,П′)=NC(yi, П″).
(Note: There is an equivalent theorem of B. McKay - theorem 2.15 in [12]).
.
Proof Let we consider two executions (labeled I and II) of the operation CSVSIR on
starting input partition П of graph G(V, E). The first j-1 selections а1, а2, ... , аj-1 are
equal for both executions and the resulting partitions are equal to Пj. Under the
conditions of the theorem, there are vertices p, q=f(p) in the cell SC (Пj).
Let the jth selection be p in the first execution, and the jth selection be q in the second
execution, i.e. the obtained partitions are: for execution I – ПI=П(а1, а2, ..., аj-1, p), for
execution II – ПII=П( а1, а2, ..., аj-1, q), for which the conditions of Theorem 2 hold:
аi=f(аi) for i = 1, 2, ..., j-1 and q=f(p). Therefore, according to Theorem 2, the similar
vertices are in cells with the same label, and thus the vertex аj+1 in SC (ПI) will
correspond to the vertex bj+1= f(аj+1) in SC (ПII). So, the selection аj+1 is possible in
execution I and the selection bj+1 = f(аj+1) is also possible in execution II. The conditions
of Theorem 2 also hold for these selections and at the (j+2)th selection similar vertices
can be selected again, i. e. аj+2, in the execution I and bj+2 = f(аj+2) in the execution II.
This process continues until the last selections аLК-1, bLК-1=f(аLК-1) in both executions
have been done – after these selections the partitions will be discrete. This means that
each numeration
П(а1, а2, ..., аj-1, p, аj+1, аj+2, ..., аLК-1)= x1 x2 . . . xi . . . xn=П′
has a corresponding numeration
П(а1, а2, ..., аj-1, q, bj+1, bj+2, ... , bLК-1)= y1 y2 . . . yi . . . yn=П″ , such that
NC(xi,П′)=NC(yi, П″), yi=f(xi), i=1, 2, ... , n.
There are four obvious corollaries of Theorem 3:
Corollary 1 of Theorem 3 Under the conditions of Theorem 3 the discrete partitions of
type П″, successors of the partition П(а1, а2, ..., аj-1, q), do not find new automorphisms
(new similar vertices). Consequently, it is not necessary to determine them if we
pre.liminarily know the discrete partitions of type П′ - successors of the partition
П(а1, а2, ..., аj-1, p).
Corollary 2 of Theorem 3 The bouquets of two similar vertices in a given SC(ПL) are
of the same size.
This statement is obvious because to each numeration of the one bouquet uniquely
corresponds a numeration of the other bouquet.
Corollary 3 of Theorem 3 To determine whether two vertices X and Y in SC(ПL) are
similar we need to know the bouquet of one of the vertices, say B(L, X), and generate
one numeration n1 derived of a selection Y in SC(L) and compare n1 with the
numerations B(L, X). Even more, B(L, X) should not contain equivalent numerations
- 14 -
because of the transitivity of the equivalence: if n1 is equivalent to one of them it is
equivalent to the another.
Corollary 4 of Theorem 3 The bouquet B(L, X) contains all bouquets derived from
each representative of an orbit in SC(ПL+1).
.Important conclusions follow from Theorem 3 and its corollaries. There are three
possibilities to determine whether two vertices X and Y in SC(ПL) are similar under
A(x1, . . . ,xL-1): (a) The bouquet B(L, X) of the vertex X should be stored and for the
vertex Y we should generate only one numeration and compare it with the
numerationsB(L, X) - this version is used in Vsep-e algorithm; (b) One numeration
should be stored for vertex X and the whole bouquet for vertex Y should be generated.
This version is used in Nauty [12] and in the most of the known algorithms; (c) Two
bouquets В(L, X) and В(L, Y) are partially generated and their numerations are
compared for determining an automorphism (with a certain probability) that maps X to
Y. This probability might be near to 1 if we choose an appropriate selection of the
bouquets size. This is the basis for the heuristic algorithms described in Section 5.
NS-number
stored
numerations
m
1
of
NG-number of
generated
numerations
m+q-1
m+m(q-1)=q.m
of
number
-
NC
comparisons
of numerations
c.m(q-1)
c.m(q-1)
versio
n
a
b
Table 1
Let's compare versions (a) and (b) (Table 1). Let m=B(L, X) and let's consider that the
numerations of the bouquets are stored in a hash table with a maximum number c of
collisions of some hash function (characteristic of the numeration) we'll explain below.
Let's also consider the worst-case – a rigid regular graph for L=1 and SC(П1)=q – in this
case all vertices in SC(П1) are not similar each other. This is the worst-case since: (i) for
L=1 the bouquets have the larger size than the bouquets for L>1 and (ii) the bouquets for
rigid graphs are full – each vertex at each level is selected. The advantage of a version (b)
is a low storage – only one numeration is stored and the disadvantage of version (a) is the
large required storage – the whole bouquet of size m for the first vertex x1 SC(П1) is
stored. A version (a) is faster since the number of the generated numerations is smaller:
NG=m+q-1. In this case the bouquet of the first vertex x1 SC(ПL) is generated and stored
and for each of the other q-1 vertices only one numeration is generated - totally NG =m+q-
1. In case of version (b) for each vertex x SC(П1) all m numerations of bouquet В(1, x)
are generated, i.e. NG= m.q – we suppose that the size of each bouquet is m or m is the
largest size. Since m.q >> m+q-1 version (a) is many times faster than version (b). The
number (NC) of the comparisons of numerations is c.m(q-1) for both versions. In version
(a) one numeration for each vertex in SC(1) is compared with c.m numerations of B(X1). In
- 15 -
version (b) the only stored numeration n1 derived from the selection X1 is compared with
c.m numerations of each bouquet B(X), x SC(1), x≠x1.
Examples:
Graph A29_1 (rigid regular graph from [29]):n=29, m=14 (this size is for each vertex in
SC(П1), q=n=29; NG(a)=m+q-1=42, NG(b)=q*m=29*14=406 (in the brackets is the
numbering of the version). We see the big difference between the .numbers of
generated numerations of the two versions.
Graph G1275 (Rigid affine plane of order 25, received from R. Mathon in private
communication): n=1275, bipartite graph with k=625*26=650*25=16250; m=8322060;
q=625; NG(a)=m+q-1=8322060+625-1=8322684; NG( b)=q*m = 625*8322060
=5201287500. In this case the difference between NG (a) and NG (b) is impressive.
4. S-code of a partition and storing the bouquets
We propose new code, named S-code, of a partition of the graph vertices. The partition
code is a number depending on the labels, sizes of the partition cells and the number of the
edges between the cells. S-code is used for reducing the time of comparing the partitions in
the graph isomorphism and automorphism algorithms. The code of a given partition can be
computed directly from the partition and the graph representation or from the code of the
parent partition and the differences between the partition and its parent partition.
In our algorithms a large number of discrete partitions (numerations) of graph vertices
are generated and stored. The length of each partition is n (n is the number of the graph
vertices). One way of reducing storage requirements is the coding of partitions. To every
partition is assigned a code (a number, characteristic value). The codes of two partitions are
compared (instead of comparing the corresponding partitions) and if they are equal then the
partitions are compared to determine if they form an automorphism. In this case the
partitions have to be regenerated using the stored base B, p=B of the partition and
applying the IR operation p times. The cardinality p of the base is many times less than n.
Let's consider the storing of a partition L obtained from the start partition 0 by applying
the IR operation successively. There are 3 ways of storing the partition L: a) storing the
partition itself, i.e. n numbers are stored; b) storing the base B(L) and the code c(L) , i.e.
p+1 numbers are stored – one for the code and p for the base. This way the amount of the
stored information is reduced from n to p+1 numbers, where p << n. In this case a
regeneration of L is made when using of L is needed; c) storing the code c(L) and a
polynomial code of B, i.e. only 2 numbers are stored but a regeneration of both L and B is
needed. In our implementations the version (b) is used.
The requirements for the code are:
i) The codes of the equivalent partitions have to be equal;
ii) The splitting ability of the code has to be maximal. This means that the number of not
equivalent partitions with equal codes have to be minimal (minimum collisions);
iii) The computation of the code should have minimal number of operations (easy to
compute);
We have examined a few versions of coding and the code with the best satisfaction of the
requirements is the following:
- 16 -
, where:
(4.1)
- the adjacency refinement partition of the vertices of graph G(V,E), (x,y)- an edge of
the graph, C(x),C(y) – the cells of the vertices x,yV and L(C(x)), L(C(y)) – the labels of
the cells C(x),C(y). The label of a cell is the index of the first vertex in the cell representing
a partition as an array.
The following theorem proves that the requirement (i) holds for the code (4.1):
Theorem Given a graph G and 1 ≡ 2, then code(1) = code(2).
Proof Each edge (x,y)E(G) has unique image (α(x),α(y))E(G) for an automorphism α
defining the equivalency of 1,2. Moreover, C(x) = C(α(x)), C(y) = C(α(y)) – the similar
vertices are in namesake cells. Therefore,
L(C(x)) = L(C(α(x))), L(C(y)) = L(C(α(y))) and
L(C(x)).L(C(y)) { code(1) }= L(C(α(x))).L(C(α(y))) { code(2)}.
Consequently, code(1) = code(2), since the last equation holds for each
edge E(G).
Evidently, the time complexity of computing the code by (4.4.1) is T=k=O(n2)
multiplications (k-the number of the graph edges) since k= O(n2). The code of can be
computed directly by (4.4.1) or indirectly by the code of the parent partition of . The
maximal value of the code Max(Code()) is obtained for a discrete partition of a
complete graph on n vertices (since its number of vertices is largest) :
Max(code(π))
1
=
2
1
=
2
𝑛−1
((𝑛2 + 𝑛) ∑
𝑖=1
((𝑛2 + 𝑛) 𝑛(𝑛−1)
𝑖 −
−
2
𝑛−1
𝑖=1 − ∑
𝑖3
∑
𝑛2(𝑛−1)2
= ∑
𝑛−1
𝑖=1
𝑖
𝑛−1
𝑖=1
𝑖2)
=
(𝑛+𝑖+1)(𝑛−𝑖)
=
2
−
𝑛(𝑛−1)(2𝑛−1)
6
) =
𝑛(𝑛−1)(𝑛+1)(3𝑛+2)
24
(4.2)
4
The second multiplicand of the first expression in (4.4.2) is a sum of an arithmetic
progression.
Example: Let's consider the graph in Figure 5 and the series of partitions and their codes:
Figure 5. G8 graph
Figure 5. G8 graph
π0=π(0) = 1,2,3,4,5,6,7,8 = C1
1 C4
2 C4
1785,63,42) = C1
17,85,63,42 = C1
1 C2
2 C2
1 C3
2 C3
0 ; Code(π0)=12; π1= π(1) = RP (1 2,3,4,5,6,7,8) =
1 C5
2 C5
1 ; Code(π1)=208; π2=π(1,7) = RP (1785,63,42) =
2 C6
2 C7
3 C2
2 Code(π2)=218; π3=π(1,7,5) =
3 C3
3 C6
3 C5
3 C7
3 C4
3 C8
3 ; Code(π3)=234. We
RP(178563,42) =17856432) = C1
illustrate only how Code(π1) is computed (Table 2).
- 17 -
EyxyCLxCLcode),())(()).(()(iiinnin(...)12111 1,2
1,3
1,4
2,3
2,4
3,5
4,6
5,7
5,8
6,7
6,8
7,8
Code (π1)
edge
(x,y)
L(C(x))*
1*8
1*6
1*6
8*6
8*6
6*4
6*4
4*2
4*2
6*2
6*2
2*2
208
L(C(y))
=8
=6
=6
=48
=48
=24
=24
=8
=8
=12
=12
=4
5. The exact Vsep-e algorithm
Table 2
In this and the following section we describe the proposed algorithms some of which
were preliminary presented in arXiv [30].
,
5.1. Basics of the algorithm
We need the following theorem for the reasoning of Vsep-e algorithm.
Theorem 4 Given: A=Aut(G(V,E)), an orbit Q=Orb(x1,A) of vetex x1V and a generating
set gen(A(x1)) of a stabilizer A(x1). Then, there is a tower of acsending subgroups of A
A(x1)=A(1) A(2) . . . A(i) . . . A(m)=A, such that: (a) A(i)
is a proper subgroup of A(i+1)
1 i m-1; (b) gen(A(i))={ gen(A(i-1)),i}, 2 i m, where the autmorphism i is such that
xi=i(x1); (c) A(i) 2. A(i-1), Orb(x1, A(i)) 2.Orb(x1,A(i-1)), 2 i m; (d) m is the
minimal value of i such that Orb(x1, A(m))=Q.
{Note: Evidently, the consequence of (b) is gen(A)=gen(A(m)))={gen(A(x1)), 2, 3,…,m.
Theorem 4 can be considered as an extension or modification of Theorem 5 of C. M.
Hoffmann in [10], page 25}
Proof (inductive, it follows the proof of Theorem 5 in [10]). Let's construct a tower of
ascending subgroups of A: A(x1)=A(1) A(2) . . . A(i) . . . A(m)=A, such that A(i)
is a
proper subgroup of A(i+1)
, i=1,2, . . . ,m-1; m is finite since A is a finite group. For i=1 we
have A(1)= A(x1) and gen(A(1)) =gen(A(x1)), known. Assume inductively that A(i) is a proper
subgroup of A and let we have orbits Q= Orb(x1, A(i-1))≠Q= Orb(x1, A(i-1)), Q, QQ.
Then, we'll have an automorphism iA\ A(i-1) such that xi=i(x1) and Orb(x1, gen(A(i)))
Q Q. Thus, i is a generator of A and A(i) since it does not belong to A(i-1), i.e.
gen(A(i))=gen(A(i-1))i and therefore A(i)2. A(i-1) (since at least one new coset appear in
the partitioning of A(i) into cosets of A(i-1)) and Orb(x1,A(i)) 2.Orb(x1,A(i-1)) (from the
orbit-stabilizer theorem: Orb(x1,A(i))= A(i)/.A(x1) ≥Orb(x1,A(i-1))=A(i-1)/A(x1)).
The first value of i when Orb(x1,gen(A(i)))=Q and A-A(i)= is i=m and consequently
gen(A)=gen(A(m)))={gen(A(x1)), 2, 3, . . . , m}.
Theorem 4 gives us the idea how to find the generators of a group if we know an orbit Q
of the group and the generators of a stabilizer of a representative x1 of this orbit. This is
done by traversing the orbit step by step. At each step we find one new generator of a new
subgroup of A knowing the generators and the orbits of the previous subgrpoup of A.
Before the first step the previous subgroup is equal to A(x1) with its orbits and generators.
- 18 -
Visiting each vertex x of the orbit, xx1, we select x only if x is not similar to x1 under the
previous subgroup. Thus a new automorphism-generator that maps x to x1 and new orbits
for a new subgroup are found. The new subgroup is a proper supergroup of the previous
subgroup.. At each step the generators, orbits and order of the current group A (subgroup
of A) are defined by the position of the selected vertex x. This process stops when the orbit
of x1 under the new subgroup becomes equal to the given orbit Q. The described process is
presented in Table 3. The selected vertex xi, i=2, . . . ,m is the first vertex after xi-1 in Q that
is not similar to x1 under A(i-1). The generators of A(x1) are called proper generators of A
and the generators 2, 3, . . . , m - mutual generators of A. The following corollary is
obvious:
Selec-
ted
vertex
x
x1
Mutual
generators
Visited
vertices
Orb(x1,A)
1={x1}
gen(A)
A(x1)
A
{x1}
-
A
A(1)=
A(x1)
gen(A(x1))
(proper generators of
A)
x2
x3
x4
. . .
xi
. . .
xm
{x1:x2} x2=2(x1) gen(A(x1)) 2
{x1:x3} x3=3(x1) gen(A(x1)) 23
{x1:x4} x4=4(x1) gen(A(x1))
. . .
. . .
{x1:xi} xi=i(x1)
234
. . .
gen(A(x1)) 2 . . .
i
. . .
. . .
xi=m(x1) gen(A(x1)) 2. . .
. . .
{x1:xm}
=Q
m
2
4
8
. . .
2i-1
. . .
Q2m-1
2.A(x1) A(2)
4.A(x1) A(3)
8.A(x1) A(4)
. . .
2i-
1.A(x1)
. . .
2m-
1.A(x1)
. . .
A(i)
. . .
A(m)=A
Table 3
Corollary of Theorem 4 The number of the mutual generators of the graph automorphism
group A toward the stabilizer A(x1) is m=1+log2Orb(x1, A).
Based on Theorems 3 4 and O-S theorem we build a new algorithm called A3. Actually,
Theorem 4 tell us that there is a generator xi=i(x1) if xx1 is not similar to x1 under the
current group but according to Theorem 3 to determine this generator we should know the
bouquet B(x1) and one numeration derivative of x and to compare them. Knowing the
partition П1 and SC(П1) according to Theorem 4 to traverse the unknown Q orbit we
should traverse its superset, SC(П1) Q. Thus, we come to the idea of A3 algorithm
(Figure 6). We denote by FRPO(X) – all first representatives of the orbits of A positioned
in SC(П1) before the selected vertex X and by BFRPO(X) – the bouquets of FRPO(X).
Then, the bouquet B may be considered as a union of the bouquets BFRPO(X).
- 19 -
A3 algorithm determines GOO(A) of the graph automorphism group A=Aut(G,1) ) and
the bouquet of each representative of an orbit Orb(A)SC(1) given GOO(A(X1)) of a
stabilizer A(X1) of the first vertex X1SC(1) and the bouquet B(X1).
According to Theorem 3 and 4 we'll determine GOO(A) and the bouquet of each
representative of an orbit Orb(A)SC(1) visiting each vertex xSC(1) positioned after
X1. Before the traversal of SC(1) we have GOO(A)=GOO(A(X1)). Each visited vertex
xSC(1) is selected if it is not similar to any previous vertexSC(1) under the current
group A. Then, we determine if there is an automorphism , x=(y), where y is one of the
roots of numerations in B. This is made (according to Theorem3) by comparing the first
numeration LK derived from the selection x with the numerationsB. Two cases are
Input: a graph G; a partition 1=RP() for a given input partition ; SC(1); a vertex XF
SC(1) and its index in SC(1); GOO(XF): the generating set gen(A(XF)) of a stabilizer A(XF),
the orbits of A(XF) and A(XF) and the bouquet B(XF); the bouquets BFRPO(XF), i.e., the
bouquet of each first representative y of an orbit of A in SC(1) with index (y) < index (XF); each
representative y and XF has no similar vertices in a position in SC(1) before it.
Output: The generators, the orbits and the order (shortly GOO) of the graph automorphism group
A=Aut(G, 1); the bouquet B of each representative of an orbit of A in SC(1).
1. orbits of A:=orbits of A(XF); gen(A):=gen(A(XF)); (A):= A(XF)
2. X:=XF;
3. Do
4.
5.
6.
7.
select next X SC(1); if there is no selected vertex X then exit;
determine a numeration n1 by SFM(X);
compare the numeration n1 with the numerations of the BFRPO(X);
if n1 is equivalent to some numeration from BFRPO(X), i.e. there is new automorphism
mapping X to some vertex from FRPO(X) then
gen(A):= gen(A); recompute the orbits of A; A:=Orb(XF).A(XF)
else {n1 is not equivalent to any numeration from BFRPO(X), i.e. there is no new
automorphism mapping X to some vertex from FRPO(X)}
build the search tree ST(X) for determining GOO(A(X)) and the bouquet B(X)
end; {if}
8.
9.
10.
11.
12. end do
Figure 6. A3 algorithm
possible: (a) If there is an automorphism then it belongs to a new subgroup of A since it
unites the orbits of x and X1, i.e., gen(A): = gen(A) {}; (b) If LK is not equivalent to any
numeration B then x becomes a representative of a new orbit and a new search tree ST(x)
is built - it determines gen(A(x)) and B(x). If during the generation of ST(x) an
automorphism is discovered and if it unites the orbits of A then, it is also a generator for A
since A(x)A. Thus, after case (b) for the next selected vertex we should compare the
numeration LK derived from the selection of x with the numerations of all known
bouquets B.
To determine the orbit of the vertex X1 it is sufficient to do the above comparison only
with the bouquet B(X1) but we do this comparison with the numerations of all bouquets of
- 20 -
B since we store them and the earlier finding of a generator is better since we can use it
earlier. If the found automorphism , x=(y), yX1 unites orbits of the current A it is a
generator of A.
After each of both cases has been handled the traversal of SC(1) continue by selection of
a new vertex x. Thus, after the traversal of SC(1) all verticesOrb(X1, A) will be visited
and the bouquet of each first representative xiSC(1) of an orbit of A and GOO(A) will be
determined (according to Theorem 4).
During the execution of A3 algorithm (as in Theorem 4) a tower of ascending subgroups of
A are built implicitly: A(x1)=A(0) A(1) . . . A(i) . . . A(m)=A, such that A(i)
is a proper
subgroup of A(i+1), 0 i m-1.
A3 algorithm is applied differently for L partition of level L=1 and L>1. If L=1 then
GOO(X1) and B(X1) are determined only for the first vertex X1 SC(1) and for other
orbit representatives in SC(1) only one numeration is generated and compared with
B(X1). This means that instruction in line 10 of A3 is not necessary if L=1. If L>1 then
GOO(XL) and B(XL) are determined for all orbit representatives XL SC(1) since they
are needed for A3 application to L=L-1.
A3 algorithm can be applied for determining any GOO(A(XL-1)) and B(XL-1), L=2, . . . ,
LK-1 under the following requirements:
R1. For L known are: ПL, SC(ПL), a vertex XFLSC(ПL); its index in SC(ПL,
,
,
is not similar to any previous vertex in SC(ПL) under A(XL-1), XL-selected
vertex, bouquets BFRPO(XL); GOO(XL-1)= GOO(XL) at the start;
R2. XL is selected in interval index(XFL)+1 to the end of SC(ПL), XL is selected as a first
not similar vertex after the current XL under the current A(XL-1). At the start XL is the
vertex with index=index(XFL)+1. This means that the vertices of
in
SC(ПL) are after
requirement of Theorem 4;
R3. If XL is similar to some vertexFRPO(XL) then the discovered automorphism is a
generator of A(XL-1): gen(A):= gen(A). Otherwise, each new numeration in B(XL)
(instruction 10) belongs also to B(XL-1).
. Thus, traversing SC(ПL) we'll traverse Orb(XFL,A(XL-1))– this is the
5.2. Vsep-e exact algorithm: cases CS1, CS2, CS3 and CS4
Let we have the starting series SFM1: П1, П2,…, ПLK. We can apply A3 algorithm
directly only to the partition ПLK-1 because for the other partitions the required input
variables are not known. For the partition ПLK-1 we have B(xLK-1) ={ПLK} and
gen(A(xLK-1))=, i.e., A(xLK-1)=1 and discrete orbits of A(xLK-1). After the application of
A3 to ПLK-1 we have determined correctly B(xLK-2) and GOOGA(A(xLK-2)). Then, A3 can
be applied to ПLK-2, i.e., a backward move is done from LK-1 to LK-2. Thus, applying A3
to the series ПLK-1, ПLK-2,…, П2, П1 we can determine GOO(A). The lowest level to which
a backward move has been made we denote by LMIN, i.e., LMIN is the level for which we
determine GOO(A(XLMIN-1)). In A3 algorithm the process of the backward moves is not
included and the instruction 10 is not revealed. All this is taken into account in PART1
(Figure 9) and PART2 ( Figure 11) algorithms called from the Vsep-e algorithm
- 21 -
))((LXFAGOO)(LXFBLXF))(,(1LLXAXFOrbLXF (Figure 8) that determines GOO(A) of the partition-wise stabilizer A=Aut(G, П) given a
graph G(V,E) and the input partition П on V. S3.
Before calling PART1 and PART2 Vsep-e algorithm determines (step S2) the orbits of A
by an TREE4 heuristic algorithm (see section 7.1) and selects X1 SC(П1) - a
representative of one of the smallest orbit of the vertices in SC(П1). Experimental tests
show with very rare exceptions that if the staring vertex X1 is a representative of one of the
smallest orbit of A then the size of the bouquet B(X1) built by PART1 is the smallest and
the running time is minimal. PART1 algorithm (Figure 9) can be considered as an
application of A3 algorithm with added the backward moves and revealed instruction 10 –
all above requirements are implemented in it. PART1 algorithm determines GOO(A(x1)
and B(x1) given П1, SC(П1) and x1SC(П1). PART2 algorithm (the second part of Vsep-e
algorithm, line S5 in Figure 8) determines GOO(A) given GOO(A(x1) and B(x1) obtained
from PART1 algorithm.
PART2 algorithm can be considered also as an application of A3 algorithm to the partition
П1 with replacing the instruction 10 by determining one derivative numeration ПLK of each
selected vertex xSC(П1) if x is not similar to x1 under the current A. PART2 algorithm
may be also considered as a direct application of theorems 3 and 4. The satisfaction of the
requirements of these theorems guarantees .the correctness of PART2 At the start of
PART2 GOO(Aut(G):= GOO(Aut(G,x1)). Each vertex x in SC(П1) that is not similar to x1
under the current Aut(G) is selected (line B2) and a comparison (line B5) of the first
derived from the selection x (line B4) with the numerationsB(x1) is
numeration
made. If there is an automorphism between some numeration B(x1) and
then is a
generator for A since it unites the orbits of x and x1. In both cases (existence or
nonexistence of ) the traversing of SC(П1) continues until its end. When the traversal of
SC(П1) completes, the generators and the orbits of A are determined and the 'orbit-
stabilizer' theorem is applied for determining A=Orb(x1, A). A(x1) (line B3).
We'll describe PART1 algorithm considering an intermediate state of ST search tree
(Figure 7) being built by the algorithm during its execution. The series of partitions ПL,
L=2, . . . , LK-1 can be divided into three intervals: the first is from П1 to ПLMIN-1, the
second – from ПLMIN to ПLP and the third – from ПLP+1 to ПLK-1. The search tree is built in a
preorder: first visiting the root (a partition ПL) and then its subtrees (the partitions ПL+1
derived from each selected vertex) in a defined order.
Applying A3 algorithm for determining gen(A(xL)) we have:
,
is the generating set of the current stabilizer
,
is the
,
is the generating set of the current stabilizer
where
generating set of the stabilizer
A(xL) and
conditions hold for the intermediate state of PART1 algorithm execution (Figure 9) shown
as a search tree ST on Figure 7:
C1: On the current path of ST tree known are: L, ПL, SC(ПL), XL, L=1, 2, . . . ,LK-1, XL is
the current selected vertex in SC(ПL). The current numeration is ПLK-1=n1;
C2: LMIN, LP, LMINLP, LP2 are known;
C3: X1 is the first vertex SC(П1);
is th.e set of the mutual generators of
. The following
- 22 -
LKLK))((),())(())((111111LLLLLxAgenxxMGxAgenxAgen))((LxAgen)(LxA))((11LxAgen)(11LxA),(111LLxxMG))((1LxAgen)(1LxA C4: Each selected vertex XL is not similar to any previous vertex in SC(ПL) under A(XL-1),
L=2, . . . ,LP-1;
C5: Each selected vertex XL is the first vertex in the SC(ПL) for L=LP+1, . . . ,LK-1; The
partitions in this section of the path are a result of the forward move SFM(XLP);
C6: Known are: GOO(XLMIN-1), B(XLMIN-1), FRPO(XLMIN);
C7: Known are the bouquets BFRPO(XL), L=1, 2, . . . , LP;
C8: Known are the orbits and orders A(X) of vertices X FRPO(XL) and X= XL,
L=LMIN,…,LP, under A(XL-1).
C9: On the current path known are the computed orbits Orbc(Xl) and computed orders
AC(Xl) for L=LMIN+1,…,LK-1 under the current A(Xlmin-1). For L=LMIN we have
Orbc(Xl)= Orb(Xl) and AC(Xl)= A(Xl) under the current A(Xlmin-1) since all generators
found so far have the same fixed points with A(Xlmin-1).
The action that follows the above state is a comparison of n1 numeration with the
numerations of BFRPO(XLP) for discovering a new generator of A(Xlmin-1).
All above conditions can be considered as an invariant for correctness of the loop C1 of
PART1algorithm.
Let's now describe PART1 algorithm. It calls SFM1 (Figure 10) and COMP (Figure 13)
algorithm. At the start all of the searched variables are not known and for each partition
ПL, L=2, . . . , LK-1 we select the first vertex XLSC(ПL) and obtain the partition ПL+1, i.e.
the only action we do is a forward move (line I1) until a discrete ПLK is obtained. Thus, the
conditions R1 to R3 hold for only for ПLK-1.Let's now consider the above requirements R1
to R3 for determining GOO(A(XLP-1))=GOO(A(x1, , . . . , xLP-1)) and the bouquet B(XLP-
1)=B(x1, , . . . , xLP-1) given the partitions П, П1, . . . , ПLP. For ПLP=П(x1, , . . . , xLP-1) also
SC(ПLP) and the selected vertex XLP are known. Besides, the requirements hold for the
position of the current vertex XLP in SC(ПLP). By the loop C1 (lines I1- I12) in PART1
algorithm each selected cell SC(ПLP) is visited and the following four basic steps are
performed:
A1 {Selection}: The selection of a vertex XLP in SC(ПLP) (line I3) is made starting from the
position next to the current XLP. The vertex XLP should not be similar to any previous
vertex in SC(ПLP) under the current A(XLMIN-1). For each level L the position i(XL) of the
selected vertex XL is stored and when a backward move to this level is performed then the
selection of a new vertex starts from the next position, i.e. i(XL)+1. If there is no selected
vertex in SC(ПLP), i.e. the SC(ПLP) has been traversed then, a backward move follows (step
A4). If there is a selected vertex XLP in SC(ПLP) then, step A2 follows.
A2 {Series of forward moves}: A series of forward moves SFM1 is performed determining
the partitions ПL=ПLP+1, . . . , ПLK with discrete ПLK (line I5). In each of these partitions the
selected vertex XL is the first vertex in SC(ПL). This way the requirements hold for the
orbit Orb(XL, A(XL-1)).
A3 {Comparison}: A check if there is a new automorphism that not belong to the current
subgroup of A(xLP-1) and maps the vertex X to any vertex from FRPO(X) is made, i.e. if
belongs to the next subgroup of A(xLP-1). This check is made by comparing ПLK with
BFRPO(X) (line I6, COMP algorithm).
A4 {Backward move}: After the traversal of SC(ПLP) is completed then
GOO(A(xLP-1)) and B(xLP-1) are determined and a backward move
- 23 -
LP:=LP-1 is made. Stop follows if LP=1. Otherwise, a selection of a new
vertex in SC(ПLP) is made applying the step A1 to it.
If there is an automorphism then, it is a generator: gen(A(XLP-1)) := gen(A(XLP-1)){}
and the orbits and the order of A(XLP-1) are recomputed. It is a generator also for A(X1):
gen(A(X1)):=gen(A(X1)){} and the orbits and the order of A(X1) are recomputed. If
there is no mapping x to a vertexFRPO(X) then a move back to LK-1 follows. This
way the building of the tree ST(x) starts from LK-1 performing the step A1 to SC(ПLK-1).
ST(x) is necessary since it determines the bouquet B(x) that belongs to B(XLP-1). After
ST(x) has been built we continue with a selection of a new vertex in SC(LP) applying the
step A1 to SC(ПLP). The search tree is built in preorder traversal: first visiting the root (ПLP
partition) and then its subtrees (the partitions ПLP+1 derived from each selected vertex) in a
defined order.
For the selected vertex XLP (line I3 – the start of ST(XLP) building) by SFM1 (line I5) is
built the first (leftmost) tree ST(XLP+1), ST(XLP+2), . . . , ST(XLK-1) for each previous
subtree. Each of these subtrees is built in backward ord..er. When the subtree ST(XLP+1) has
been built then the building of the subtree for the next selected vertex XLP+1 starts (XLP+1
should hold the requirements). When there is no selected vertex XLP+1 then a backward
move LP+1 to LP is made – this means that the ST(XLP) is built. If
LP=1 the algorithm stops.
Figure 7. Search tree of VSEP-e
LMIN-1
LMIN
xLMIN-1
.
.
.
.
.
.
.
.
xLMIN
.
.
.
.
.
.
.
.
.
.
LP-1
.
.
.
.
LP
xLP
1
xLP
2
.
.
.
.
B(a1,1
LP)
B(a2,1
LP)
.
.
.
.
xLP-1
.
.
.
.
.
.
.
.
xL
i
P
.
.
.
.
LP+1
LP+2
.
.
.
.
.
.
.
.
LK-1
LK
n1
.
.
.
.
x2
x3
x1
L=1
L=2
L=3
5.2.1. Cases CS1 and CS3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
SC(1)
SC(2)
SC(3)
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
SC(LMIN-1)
SC(LMIN)
SC(LP-1)
SC(LP)
SC(LP+1)
SC(LP+2)
SC(LK-1)
- 24 -
(5.2.2)
(5.2.1)
Let us consider the cases when the numeration .ПLK (Figure 7) does not form an
automorphism with any numeration BFRPO(ПLP). Knowing that A(XLP-1) is correct and
supposing that each orbit Orb(XL, A(XL-1)) for L=LP, . . . , LK-1 is also correct, and
applying the Theorem 'O-S' we obtain
A(XLP-1)/Orb(XLP, A(XLP-1))/Orb(XLP+1, A(XLP))/ … /Orb(XLK-1,A(XLK-2)) =
АLK-1 = 1.
However, the orbits in (5.2.1) are unknown and consequently we cannot use it. Instead, the
sets W=COrb(XL, A(XLP-1)) of the vertices in SC(ПL) similar to XL under A(XLP-1) are
known, i.e.
W=COrb(XL, A(XLP-1))= Orb(XL, A(XLP-1)) SC(ПL).
We call the set W 'computed orbit' to distinguish it from the real orbit U=Orb(XL, A(XL)).
The following relation holds U=Orb(xL,A(xL-1))COrb(xL,A(xLP-1))=W, (5.2.3)
since A(xL-1)) A(xLP-1).
Considering (5.2.3) and applying the Theorem 'O-S' for the computed orbits we obtain
A(XLP-1)/COrb(XLP,A(XLP-1))/COrb(XLP+1,A(XLP-1))/ … / COrb(XLK-1,A(XLP-1)) =
АLK-1 1.
The computation on (5.2.4) is performed by SFM1 algorithm (Figure 10) called from the
instruction I5 of PART1. If the sign in (5.2.4) is = (i.e., the real and the computed orbit are
equal), then we call the orbit Orb(xL, A(xL-1)) separated. If the sign in (5.2.4) is < , then,
obviously, W=UR, R>1, i.e., W (and R) include vertices that belong to SC(ПL) but are
not similar to XL under A(XL-1). It can be proved that when W contains some vertex, then it
contains the whole orbit of this vertex under A(XL-1).
Thus, W can be considered as an union of orbits of .A(XL-1). We call this case non-
separation of orbits (denoted by NSO) and the orbit U is called non-separated (non-
partitioned). Since the representatives of the orbits belonging to R are not known we
cannot select them during the traversal of the SC(ПL). Thus, the search tree of such a
(5.2.4)
Input: Graph G(V,E) and a partition П on V
Output: Generators, orbits and order of the partition-wise stabilizer A=Aut(G, П), denoted GOO(A)
S1: Initialization: П1:=RP(П, NCL);{NCL-the number of the cells in П1}
S2: if NCL=n {П1is discrete} then message'Trivial group'); return end if;
S3: CSLCh {call cell selector chooser-finds the cell selector};
S4: Determine SC(П1) and х1:=first vertex in SC(П1);
S5: Use a filter that defines whether or not to call the following procedures that:
a) Give a better partition П1 obtained by S code sorting of the vertices in П1. Determine a new
SC(П1) and х1:=first vertex in SC(П1);
b) Determine: (i) the generators and orbits of A by TREE4 heuristic algorithm,
starting from discrete orbits; (ii) SC(П1); X1:=a representative of the smallest
orbit of the vertices in SC(П1); Put the vertex х1 on the first position in SC(П1);
S6: PART1: Determine GOO(A(х1)), B(х1) given П1, SC(П1), х1SC(П1)
S7: PART2: Determine GOO(A) given GOO(A(х1)), B(х1)
Figure 8. Vsep-e algorithm
- 25 -
vertex cannot be built and its bouquet will not be determined. This is an unallowable error
since these bouquets belong to B(XL-1), B(XL-2), . . . , B(X1) and they are needed (as we
know from A3 algorithm) for determining
GOO(A(XL-2)), GOO(A(XL-3)), . . . , GOO(A(X1)). If there is only one partition with non-
separated orbit, then the sign in (4.2.1.4) is <, since the length of the computed orbit is
greater or equal to the length of the real orbit – this is the condition to detect the presence
of NSO. We call this case CS3. If there is no NSO in any partition, i.e. each computed
orbit is equal to the real one, then the sign in (4.2.1.4) is = and the case is denoted by CS1.
If the case is CS1, then at the exit of COMP algorithm (line I6 in PART1) the invariant for
correctness holds for LP=LK-1: only LP is changed, LMIN remains the same. When the
case is CS3 we lose GOO(XLMIN-1) determined so far and the algorithm continues with a
new start point: LP=LK-1, LMIN=LP, XLMIN= XLK-1 is the first vertex in SC(LK-1) and
A(XLMIN-1)={I}, i.e., each vertex is put into a separate orbit. Obviously, the invariant for
correctness holds for CS3 case.
Input: П1, SC(П1), х1
Output: GOO(A(х1)), B(х1)
I1:SFM1:Determine ПL,A(XL-1)=1,SC(L),L=2, . . . ,LK given ,П1,SC(П1),X1;
LP:=LK-1; LMIN:=LP; gen(A(х1)):=; A(х1):=1; B(XLK-1):={ПLK};
I2: do {loop C1: Visit vertices in SC(ПLP) }
I3: Select XLP in SC(LP);
I4: if XLP0 then {forward move}
I5: SFM1: Determine ПL, A(XL-1), SC(L),L=LP+1, . . . ,LK given LP, ПLP ,
SC(ПLP), XLP;
I6: COMP: Compare ПLK with the numerationsBFRPO(XLP) for determining an
automorphism A(XLP-1). Determine GOO(A(XLMIN-1)), LP;
I7: else {backward move: SC(LP) has been traversed}
I8: LP:=LP-1;
I9: if LP<LMIN then LMIN:=LP; Aut(XLMIN):= Aut(XLMIN+1) end if;
I10: if LP=1 then return end if;
I11: end if
I12:end do;{ loop C1}
Figure 9. PART1 algorithm (instruction S6 of Vsep-e algorithm (Figure 8))
Input: L=LP, ПL, SC(ПL), XL; Output: ПL, A(XL-1), L=LP+1,… ,LK
1. do
2. L:=L+1; ПL:=RP(ПL-1,XL-1,BRCL);
3. if BRCL=n then return end if
2. 4. determine SC(L); XL:=first vertex in SC(L);
5. A(XL):= A(XL-1)/Orb(XL, A(XLMIN-1)SC(ПL)
5. end do
Figure 10. SFM1 algorithm (instruction I5 of PART1 algorithm (Figure 9)
- 26 -
The operations in the case CS3 may be considered as an error correction of the incorrect
orbits of some A(XL-1) determined by the moment since the algorithm interrupts its current
execution and starts from the new start point for finding the correct orbits of A(XL-1) and
the bouquet В(ХL-1) .
3.2.1. Cases CS2 and CS4
Let's consider the cases when there is an automorphism mapping XLP to some vertex
ULPFRPO(XLP), i.e. the numeration ПLK forms an automorphism with some
numerationB(ULP), XLP=α(ULP).
Input: L=1, П1, SC(П1), х1, GOO(A(х1)), B(х1)
Outut: GOO(A)
B1: do
B2: Selecte next vertex Х in SC(П1);
B3: if X=0 then A=Orb(х1, A.A(х1); return end if;
B4: SFM1A: Determine ПLK given L=1, П1, SC(П1), X SC(П1)
B5: COMP: Compare ПLK with the numerationsbouquet В(x1) for an
automorphism . If there is then gen(A):=gen(A) and recompute the
orbits of A;
B6: end do;
Figure 11. PART2 algorithm (instruction S5 of Vsep-e algorithm (Figure 8))
Input: L=1, П1, SC(П1), X SC(П1)
Output: ПLK
1: do
2: L=L+1; ПL=IR(ПL-1,XL-1);
3: if NCL=n then return end if;
4: else determine SC(L); XL=first vertex in SC(L)
5: end do
Figure 12. SFM1A algorithm (instruction B4 of PART2 algorithm ( Figure 11))
Then, there is a possibility of NSO for some orbits of the vertices of the target cells of the
current path for the levels LMIN+1 L LP: we denote by CS2 the case when there is no
NSO and by CS4 the case when there is at least one case of NSO in this interval. The
automorphism is a generator of A(XLMIN-1) since A(XLP-1) A(XLMIN-1): gen(A(XLMIN-1))
:= gen(A(XLMIN-1)) {}. Thus, the Orb(A(XLMIN)) and A(XLMIN) are changed (line 11)
and we denote by +α any variable with changed value. Before each orbit and each order
of the current stabilizers are correct: A(ZL) =A(XL-1)/Orb(ZL, A(XL-1)) for each
representative of orbit ZL SC(L), L=LMIN, . . . , LP.
The following actions are:
- 27 -
a) For the interval L=LMIN to LP (loop: lines 11 to 24) of the current ST path the
stabilizers of the selected vertices XL before and after are determined (line 13) ;
Input: ПL, SC(ПL), XL, A(XL-1), L=2,3, . . . , LK; LP, BFRPO(XLP), LMIN,
GOO(A(XLMIN-1))
Output: GOO(A(XLMIN-1)), LP
1: Compare ПLK with the numerationsBFRPO(XLP)
2: if ПLK is not equivalent to any numerationBFRPO(XLP)
3: then {CS1 or CS3}
4: if A(XLK-1)=1
5: then {CS1}
6: LP=LK-1; B(XLP)={ПLK}; if LP<LMIN then LMIN=LP
7: else {CS3}
8: LP=LK-1; B(XLP)={ПLK}; LMIN=LP; A(XLMIN-1)={I};XLMIN:=the the first vertex in
SC(LMIN); Aut(XLMIN):=1
9: end if
10: else{CS2 or CS4: there is an automorphism , XLP=(ULP), between ПLK and some
numerationBFRPO(XLP) derivative of the vertex ULPFRPO(XLP)}
11: Determine GOO(A+(XLMIN-1)): gen(A+(XLMIN-1))=A(XLMIN-1) {};
determine orbits of A+(XLMIN-1) and
A+(XLMIN-1)=A(XLMIN)Orb(XLMIN, A+(XLMIN-1))
12: for L=LMIN, LMIN+1, . . . , LP do
13: A(XL)= A(XL-1)/Orb(XL, A(XL-1)); A+(XL)= A+(XL-1)/COrb(XL,
A+(XLMIN-1));
14: if L>LMIN then
15: for each vertex ZLFRPO(XL) do{check if A(ZL) has changed after }
16: If A+ (ZL)= A(XL-1)/COrb(ZL, A(XLMIN-1))< A(ZL)= A(XL-1) / Orb(ZL,
A(XL-1))
17: then CS4=true
18: end if
19: if L=LP and index(ZL)= index(ZLp)
20: then IULP:= index(ULP); RSTBULP:=A(ULP)
21: end if
22: end for {loop for from line 15}
23: end if
24: end for {loop for from line 12}
25: if (not CS4) then return end if
26: if RSTBULP=1
27: then LMIN:=LP;L:=LP; gen(A(XLMIN-1)):={};orbits of A(XLMIN-1) :=cycles of ;
A(XLMIN-1):= Orb(ULP, A(XLMIN-1)) ;XFLMIN:=ULP;
A(XFLMIN-1) :=1; the execution continues by starting the selection with a vertex with
index(ULP)+1
28: else SFM(ULP);L:=LK-1;LP:-L;LMIN:=L;A(XLMIN):={I};the next selected vertex is
the first vertex in SC(LP)
29: end if
30:end if
Figure 13. . COMP algorithm (instruction I6 of PART1 algorithm (Figure 9))
- 28 -
b) For each representative ZLFRPO(XL), L>LMIN a check if A(ZL) has changed after
(line 16) is performed. This check is excluded for L=LMIN since all orbits and
c) stabilizers are correct – all so far found generators have the same fixed points with
A(XLMIN-1); If there is a change of A(ZL) for some L then CS4=true. Also, the
index(ZLP) and A(ULP) are stored;
d) If CS4=false (line 25), i.e. the case is CS2 then an exit from COMP follows, no
changes of the variables computed so far are made and the next selection starts from
the current selected vertex in SC (LP).
e) In CS4 GOO(XLMIN-1) found so far are lost and the algorithm continue from a new
start point as in CS3 case. The invariant for correctness holds for the new state of the
ST tree. The main requirement that hold is that the selected vertices XL of the current
path (ULP for L=P) are not similar to the previous vertices in SC(L) under A(XL-1).
If the case is CS4 (lines 26 to 30) then there are possible two subcases: i) A(ULP)=1
and ii) A(ULP)>1. If A(ULP)=1 then LMIN:=LP, L:=LP and a new GOO(XLMIN-!)
determined (line 17) are determined – the new start point is index(ULP) in SC(LP).
The case when A(ULP)>1 (line 28) means that there are generators we do not know –
that's why we do a forward move FM(index(ULP)) to come to a new start point:
L=LK-1;LP=L, LMIN=L, A(XLMIN):={I};the next selected vertex is the first vertex
in SC(LP). From registered CS4 cases in experiments on benchmark graphs there are
only cases with A(ULP)=1.
There is another way for CS2CS4 check (not shown in COMP procedure): the check
for difference of the new and old stabilizer is made only for the selected vertices XL
of the path. The experiments show that both ways work correctly.
The experiments on the benchmark graphs in [19] show that CS4 case occurs only for the
graphs B52 (Mathon doubling of b25-1 graph [18]), latin-16 and 24, and had-96.
5.3. Examples
Figure 14. G10 graph
Let us consider the search tree traversal in Figure 15 (in preorder) for the graph of
5.3.1. Simple example
Figure 14 . Starting from the partition П1= П()= RP(Пu)=2,94,5,6,71,3,8,10 we do a series
of forward moves SFM: П2 = П(4) = 2947568,101,3, П3=П(4,8) = 2947568101,3,
П4=П(4,8,1)=29475681013 - the first numeration n1. The selected cell SC(ПL) in this
SFM is the cell with the largest number of adjacency cells, the selected vertex is always the
first in SC(ПL) and the order of the stabilizers A(0), A(1), A(2),…, A(L), L=1,2,…,LK-1 is 1.
Then, a back move follows, L=L-1, L=3, vertex 3 is selected and П4 = П(4,8,1)=
2947581031 - n2, n2 =α1(n1), α1=(1,3). Then, again back move to L=3, where there is
no selected vertex in SC(3), new back move to L=2, A(4,8)=Orb(1).A(4,8,1)=2,1=2 and
- 29 -
selected vertex in SC(2) is 10, П3=П(4,10) = 2947561081,3, П43 = П(4,10,1) =
29475610813 - n3, n3 =α2(n1), α2=(8,10). At this point, there are no selected vertices
in SC(3) and SC(2), that's why back moves to L=3,2,1 are made and A(4) =
Orb(8).A(4,8)=2.2=4. Here PART1 ends and PART2 starts generating two SFM, for
vertices 5, 6 SC(1). The first SFM is: П(5)= 29564768,101,3,
П(5,8)=29564768101,3, П(5,8,1)= 295647681013 - n4, n4 =α3(n1),
α3=(4,5).(6,7). The second SFM is: П(6)= 9267541,38,10, П(6,1)=
926754138,10, П(6,1,8)= 92675413810 - n5, n5 =α4(n1),
α4=(1,8)(2,9)(3,10)(4,6).(5,7). Orbits of A are: (2,9)(1,3,8,10)(4,5,6,7) and
A=Orb(4).A(4)=4.4=16.
A
α4
α3
α2
α1
16
8
4
2
1
SC(1)
SNSC(
1)
L=1
П( )=
П1
)
4
(
A
4
2
1
4
5
6
7
SNSC(
SC(2)
2)
L=2
П(4)=
П2
)
8
,
4
(
A
2
1
8
10
L=2
П(5)=П2
8
10
L=2
П(6)=П2
1
3
SC(3)
L=3
П(4,8)=
П3
1
3
L=3
П(4,10)=
П3
)
1
,
8
,
4
(
A
1
)
1
,
0
1
,
4
(
A
1
1
3
L=3
П(5,8)=
П3
1
3
L=3
П(6,1)=
П3
8
10
n1 = 29475681013 =
П(4,8,1)
α1 = (1,3) = (n1 ~ n2);
orbits: (1,3)(2)(4)(5)(6)(7)(8)(9)(10)
n3 = 29475610813 = П(4,10,1)
α2 = (8,10) = (n1 ~ n3);
orbits: (1,3)(2)(4)(5)(6)(7)(8,10)(9)
n4 = 29546781013 =
П(5,8,1)
α3 = (4,5)(6,7) = (n1 ~ n4);
orbits: (1,3)(8,10)(4,5)(6,7)(2)(9)
n5 = 92675413810 =
П(6,1,8)
α4 = (2,9)(4,6)(5,7)(1,8) = (n1 ~ n5);
orbits: (1,3,8,10)(2,9)(4,5,6,7)
Figure 15. Search tree for the graph of Fig, 14
- 30 -
5.3.2. Example with all cases (CS1, CS2, CS3, CS4) (Figure 16 )
The example is for graph G=B52 (Mathon [29]), regular graph, n=52, degree=25,
Aut(G)=12, orbit lengths: 2*2+4*6+2*12; orbits:
(4,30)(24,50) (21,37, 12,38,11,47) (17,36,16,42,10,43) (8,44,39,18,13,34)
(26,35,28,2,9,52) (29,25,22,15,20,33,7,46,41,48,3,51) (40, 6,49,45,5,1,27,31,19,23,32,14).
We show in Figure 16 only the subtrees of the selections (4,11), (4,13), (4,14) and (4,24) of
the search tree. The first selected vertex in П1=1,2, . . . ,52 is the vertex 4 since it is from
one of the smallest orbits – the orbits are found by the heuristic algorithm. We start the
consideration from the selections (4,11,51) – this is the numeration n52: it is not equivalent
to any numeration from B(4,2), B(4,3), B(4,10). Before these selections there were
determined the bouquets of the representatives B(4,2), B(4,2)=9, B(4,3), B(4,3)=37,
B(4,10), B(4,10)=15, i.e., totally 51 nonequivalent numerations. These bouquets are
derivatives of FRPO of the set {2, 3, 7, 9, 10} – these vertices precede the selected vertex
11 in SC(П2), П2=П(4). There are no bouquets for the vertices 7 and 9 since they are
similar to previous vertices in SC(П2) under A(4): 7-3, 9-2. We have LP=LMIN=2 at the
selections (4,11,51). Before these selections there are found 4 generators of A(4), its order,
orbits and some stabilizers. At the selections (4,11,51) the case CS3 has been discovered.
That's why the selected vertex 51 in SC(П3), П3=П(4, 11) becomes a new starting start
point: all information about the stabilizer A(4) obtained so far is lost and
A(4)=A(4,11)=A(4,11,51)={I}, LP=LMIN=3, XFLMIN=51, A(XFLMIN)=1. The next
selected vertex in SC(П3) is 7. The numeration П(4, 11, 7) is equivalent to numeration
n52.
Thus a new generator α5 for A(4,11) and A(4) is found (this is case CS2), α5=
(1,6)(2,52)(3,41)(4)(5,45)(7,51)(8,39)(9)(10,43)(11)(12,21)(13,34)(14,23)(15,29)(16)(17,3
6)(18)(19,31)(20,22)(24)(25,33)(26,28)(27,32)(30)(35)(37)(38,47)(40,49)(42)(44)(46,48)(5
0). Then, we compute
A(XLMIN-1)=A(XFLMIN)Orb(XFLMIN, A(XLMIN)), i.e., A(4,11)=Orb(51,
A(4,11)).A(4,11,51)=2*1=2. The next selected vertex in SC(П3) is 9 - the partition
П4=П(4, 11,9) is not discrete: A(4, 11,9)= A(4, 11)/Orb(9, A(4, 11))=2/{9}=2/1=2.
Then, the next selected vertex in SC(П4) is 46 and П5=П(4, 11,9,46) is discrete
(numeration n53). The numeration n53 is not equivalent to the numeration n52 and A(4,
11,9,46)= A(4, 11,9)/Orb(46, A(4, 11,9))=2/{46,48}=2/2=1 – this is CS1. The vertex 48
is not selected in SC(П4) since it is similar to the vertex 46 under A(4, 11,9). Then, a
backward move to L=3 and a selection of the vertex 10 are made. The partition n54 is
discrete and not equivalent to any numeration in B(4,11), A(4,11,10)= A(4,11)/Orb(10,
A(4,11))=2/2=1 (the case is CS1). The vertex 43 in SC(П3) is not selected because it's
similar to the vertex 10 under A(4, 11). The next selected vertex in SC(П3) is 48- the
partition П4=П(4,11,48) is not discrete, so we do forward move to L=4 and choose vertex
27 in SC(П4). The numeration П5=П(4,11,48,27)=n55 is not equivalent to any numeration
in B(4,11): A(4,11,48,27)=A(4,11)/Orb(48,A(4, 11))/Orb(27, A(4,11,48)) =
2/{46,48}/{27,32} =2/2/2=0.5<1. This is case CS3. So, the vertex 27 in SC(П(4,11,48))
becomes a new starting start point: all information about the stabilizer A(4,11) is lost,
A(4)=A(4,11)=A(4,11,48)=A(4,11,48,27)={I}, LP=LMIN=4, XFLMIN=27, A(XFLMIN)=
- 31 -
Figure 16. The search tree ST for the graph B52
- 32 -
L=1L=2L=3L=45225...3114CS1α523791011134578951791043484218324624274648273274318244648322727323474610311445679101112146101112131415675...1254125657n1n9n36n51n52n53n54n55n56n57n58n59n60n61n62n63n64n655148n665α6α72622222212221122122CS4CS1CS1CS1CS1CS2CS1CS1CS1CS1CS1CS3CS1CS3CS2CS1CS1CS1.........n8n50n35622348181431512734891113CS1CS1CS1CS1n67n68n69n701111
Figure 16A. The search tree ST for the graph B52 (continued)
A(4,11,48,27)=1. We omit the description of the next selections in SC(П3)= SC(П(4,11)).
We only mention the occurrence of CS2: the generator 6=5 that leads to gen(A(4,11))=
{6}, A(4,11)=2 and the orbits of A(4,11) equal to the cycles of A(4,11). After the
SC(П(4,11)) has been traversed a backward move to L=2 follows: LP=LMIN=2,
B(4,11)={n52,n52,...,n64}, B(4)={n1,n2,...,n63}), XFLMIN=11, A(XFLMIN)= A(4,11)=2.
The next selected vertex in SC(П2) is 13: the partition П3=П(4,13) is not discrete, we do
forward move to L=3 and choose the vertex 3 in SC(П3). The numeration П4=П(4,13,3)
- 33 -
L=1L=2L=3L=45225314...CS12379101424457101451791043484218324624273843471010424567910111215...12412559n1n9n36n78n77n75n73n71n72n74n76n79n813214n83α9221CS1CS1CS3CS1.........n8n50n356...231918131189CS1CS1CS1CS1CS1CS1CS1CS1CS181314n80n823α812...3112...1132544CS2CS2n51...n63n64...n70...n67CS1n84n8511122262 (n64) is not equivalent to any numeration in B(4). The order of its stabilizer is
A(4,13,3)=A(4)/Orb(13, A(4))/Orb(3,A(4,13)) =2/{13,34}/{3}
=2/2/1=1 (CS1). After the selection 47 in SC(П3) and 31 in SC(П4) we obtain the
numeration n65=П5=П(4,13,47,31) that is not equivalent to n65 and
A(4,13,47,31)=A(4)/Orb(13,A(4))/Orb(47,A(4,13))/Orb(31,A(4,13,47))=2/{13,34}/
{47}/{31}=2/2/1/1=1 (CS1). After the selection 14 in SC(П4) we have the numeration
П(4,13,47,14), LP=4, LMIN=2. Numeration П(4,13,47,14) is equivalent to previous one
n65. Thus, a new generator α7 = (1,49)
(2,9)(3,46)(4)(5,40)(6,45)(7,41)(8,44)(10)(11,38)(12,37)(13)(14,31)(15,33)(16,17)(18,34)(
19,32)(20,29)(21)(22,25)(23,27)(24)(26)(28 35)(30)(36)(39)(42,43)(47)(48,51)( 50)(52)
for A(4) is found and the new orbits are
Orb(A(4))={1,5,6,40,45,49}{2,9,52}{3,7,41,46,48,51}{4}{8,39,44} {10,42,43}{11,38,47}
{12,21,37}{13,18,34}{14,19,23,27,31,32}{15,20,22,25,29,33}{16,17,36}{24}{26,28,35}
{30}{50} and the order of the stabilizer A(4) is A(4)=A(4,11).Orb(11,A(4))=2.3=6
(XFLMIN=13). Then, a check for CS2/CS4 follows (LMIN=2<LP=4). We check for NRO
for each vertex ZLFRPO(XL), L=LMIN+1,...,LP=3,4, i.e., if there are changes of the
orders A(4,13,3) and A(4,13,47,31). The order of A(4,13,3) before α7 is A-α7(4,13,3) =
A-α7(4)/Orb(13,A-α7(4))/Orb(3,A-α7(4,13))=2/2/1=1 and after α7 it is
A(4,13,3)=A(4)/COrb(13,A(4))/COrb(3,A(4))=6/13,18,34/3,46,51,48=6/3/4=0.5.
This difference (A-α7(4,13,3) =1 A(4,13,3) = 0.5) shows that the orbit
Orb(3,A(4,13))={3,46,51,48} under A(4,13) is incorrect, it is united orbit, i.e., the case is
CS4. (As we'll see later, the correct orbits are {3,46}{51,48}). Hence, the check for
A(4,13,47,31) is not necessary. As the case is CS4 we set LP=4 (not changed), LMIN=LP,
gen(A(XLMIN-1)=gen(A(4,13,47))={α7}; Orb(A(4,13,47))=cycles of α7 and A(4,13,47))=2
(the least multiple of the cycle lengths of α7). We also set XFLMIN=31 and we start the
selection of a new vertex from the current XLP=14 and since it is the last vertex in SC(П4)
we make a move back to the level L=3 selecting the vertex 46. We omit the following
actions of the algorithm. We only mention the last generators
α8=(1,5)(2)(3,51)(4)(6,40)(7,46)(8)(9,52)(10,42)(11,47)(12)(13,18)(14,32)(15,22)(16,36)(1
7)(19,23)(20,33)(21,37)(24)(25,29)(26,35)(27,31)(28)(30)(34)(38)(39,44)(41,48)(43)(45,4
9)(50) (14,27,19) (15,20,25)(16,17,36)(22,29,33)(23,31,32)(24)(26,35,28). The orbits due
to the generators α8 and α9 are:
Orb(A(4))={1,5,6,40,45,49}{2,9,52}{3,7,41,46,48,51}{4}{8,39,44}{10,42,43}{11,38,47}
{12,21,37}{13,18,34}{14,19,23,27,31,32}{15,20,22,25,29,33}{16,17,36}{24}{26,28,35}
{30}{50} and A(4)=Orb(XFLMIN,A(4)).A(4,XFLMIN)=6.1=6, where XFLMIN=14. The last
generator found by PART2 is
α10=(1,27)(2,28)(3,29)(4,30)(5,31)(6,32)(7,33)(8,34)(9,35)(10,36)(11,37)(12,38)(13,39)
(14,40)(15,41)(16,42)(17,43)(18,44)(19,45)(20,46)(21,47)(22,48)(23,49)(24,50)(25,51)
(26,52). The orbits of A due to the generators α8, α9 and α10 are given at the beginning of
this section and A=A(4)Orb(4,A)=6*2=12. Thus, the output is: A=12, Orb(A) and
generators α8, α9 and α10.
- 34 -
Figure 17. The search tree for A50 rigid graph
5.3.3. Example of a search tree for a rigid graph
The example is for graph A50, Mathon [29], n=50, regular bipartite graph,
k=105=15*7+35*3, Error! Reference source not found..). The bouquet B(1)=168,
SC(П2)=14, each selected cell SC(П3) has size SC(П2)=12. The search tree is full.
5.4. Correctness and analysis of the algorithm
We do not prove formally the correctness of the algorithm but it follows from the
description of the algorithm. If the bouquets of each representative of an orbit in SC(ПLP)
are correct and the whole orbit of each selected vertex in SC(ПLP) is traversed then
according to Theorem 4 GOO(A(XLP-1)) and B(A(XLP-1)) will be determined correctly. The
main problem is to guarantee the correctness of the bouquets - this is shown in the
description of the cases CS1, CS2, CS3 and CS4. The traversal of the whole orbit of each
selected vertex in SC(ПLP) is also guaranteed (Figure 9). The formal correctness prove of
the algorithms and their analysis will be presented in a new article..
6. Vsep-is algorithm for graph isomorphism(GI algorithm)
Vsep-is algorithm finds an isomorphism between two graphs G1(n1,k1), G2(n2,k2), if any.
It is based on Corollary 3 of theorem 3 and theorem 4. Firstly, it checks for some necessary
conditions for isomorphism: a) n1=n2?; b) k1=k2? c) equality of sorted degree sequences
of vertices of G1,G2;c) compatibility of π1=RP(G1,πu); π2=RP(G2,πu); d) equality of
indices of the target cells of π1and π2. If one of these conditions does not hold the graphs
are not isomorphic. Otherwise, the bouquet B(x1, G1) is determined by applying PART1 to
the first vertex x1SC(π1).
Then, by applying PART2 for each vertex ySC(π2,G2) is generated one numeration that
is compared with the numerations of the bouquet B(x1,G1) for an isomorphism, If there is
an isomorphism for the current numeration then a message 'G1G2', printing the
isomorphism and stop follows. If all numerations (instruction 6) derived from each vertex
- 35 -
38 37 36 LK=4 L=1 L=2 L=3 15 14 1 12 12 12 12 ySC(π2,G2) do not form an isomorphism with the bouquet B(x1,G1) then a message that
the graphs are not isomorphic follows.
Input: Graphs G1(n1,k1), G2(n2,k2),
Output: G1 and G2 are isomorphic or no; print the isomorphism if any
1: if (n1≠n2 or k1≠k2 or ordered sequences of vertex degrees of G1,G2 are not equal)
then output 'G1!G2';return end if
2: Define π1=RP(G1,πu); π2=RP(G2,πu); if π1,π2 are not compatible then message
'G1!G2'; return end if
3. Define SC(π1), SC(π2); if(index(SC(π1)) ≠ index(SC(π2)))then output 'G1!G2';
return end if
4. Apply PART1 to the first vertex x1SC(π1) to determine the bouquet B(x1,G1)
5. for each vertex ySC(π2,G2) do
determine one numeration num by applying SFM(y);
compare num with the bouquet B(x1,G1) for isomorphism;
if there is an isomorphism then output 'G1G2' and the isomorphism; return end if
end for
Figure 18. Vsep-is algorithm
6. output 'G1!G2'
7. The heuristic algorithms (Vsep-orb, Vsep-hway, Vsep-sch)
They are based on Theorem 3. For determining whether two vertices x and y are similar
two partial bouquets are built for both vertices and then, some automorphisms between the
numerations of these bouquets are determined. To determine certainly that x and y are
similar one of the bouquets should be full. Consequently, the probability to find at least one
automorphism mapping x to y is less than 1 if we use the algorithm with partial bouquets.
This makes the algorithms inexact. Instead, less time is needed for bouquets building and
less storage is needed for them because of their smaller sizes. We describe two heuristic
algorithms Vsep-orb, and Vsep-hway in this section.
7.1. TABLE1 and TREE4 procedures
These procedures are needed for the following description of the heuristic algorithms.
TABLE1 heuristic procedure generates in SC(П) numerations by the operations fork and
RRST. Each new numeration is compared with previously found numerations (stored in
HT) for an automorphism. If there is an automorphisn it is added to previously found set of
generators. If there is no an automorphism the numeration is put into HT. RRST tree is
defined by the parameters W (width) and D (depth): for each level L, 1< L ≤ 1+D, in
SC(ПL) the number of selected vertices is W. The experiments show that when the selected
vertices are evenly distributed in SC(ПL) of RRST the results are more correct compared
- 36 -
Input: G- a graph, П – a stable partition, W, D – a width and a depth of the reduced
regular search tree RRST;
Output: Generators and orbits of A=А(G, П) stabilizer, norb - a number of orbits;
Local variables: HT- a hash table of partial bouquets of generated numerations
T1: Initialization: Set Orb(A(G, П) to discrete orbits; gen( A):=; HT empty;
T2: Find SC(П) target cell;
T3: Select each vertex xSC(П) that is not similar to a previous vertex in SC(LB)
under current orbits. Generate a numeration n1 for x vertex by a series of forward
move and compare it for an automorphism with the numerations in HT .If there is
then: gen(A) :=gen(A) {} and determine new Orb(A(G, П) and norb. If there is no
then add n1 into HT. We call this operation a fork.
T4: A reduced regular search tree (RRST) is built for each representative of the
orbits in SC(LB) determined so far. The orbits are sorted in increasing order of their
lengths starting from the smallest one. Each representative of an orbit is a root of a
tree. The depth of RRST is D. Each node of the RRST corresponds to a selection level
L, 1 < L ≤ 1+D and a target cell SC(L). In each SC(ПL) a fixed number W of vertices
y are selected, then a forward move to next level L is made for each y. In the 1+D
level only a series of forward moves SFM is made for each selected vertex until a
discrete partition (numeration n1) is obtained. The number W may be regarded as out-
degree of the RRST node. The vertices in SC (ПL) are selected successively by step
st= SC(ПL) / W ≥ 1 starting from the beginning of SC ПL), i.e. they are evenly
distributed in SC(ПL). To determine an automorphism and new orbits each
numeration n1 is compared with the numerations in HT table. If there is then:
gen(A) :=gen(A) {}, new Orb(A(G, П) and norb are .determined. If there is no
then add n1 into HT. The procedure stops when RRSTs are generated for all
representatives of the orbits in SC(ПL).
Figure 19. TABLE1 procedure
Input: graph G, equitable partition П, parameters mbrsyvp,wth1, b1a, dpth1, br2a
Output: GO(A(П)), norb – number of orbits of A(П)
1: DO wth=wth1, wth1+br1a
2: DO dpth=dpth1, dpth1+br2a
3: TABLE1(П, WTH, DPTH, norb)
4: IF (norb=1 or SC(П1) contains one orbit or norb remains unchanged after
mbrsyvp number of successive calls to TABLE1) then
5: exit
6: end if
7: END DO
8:END DO
Figure 20. TREE4 procedure
- 37 -
with other ways of distribution. One call to TABLE1 procedure with fixed parameters W
and D usually is not sufficient to determine а correct GO(A). That's why several calls to
TABLE1 with different parameters (W, D) are needed – this is done by TREE4 procedure
(Figure 20). The parameters of TREE4 have the following meaning: mbrsyvp – maximal
number of successive calls to TABLE1 with no changed norb, wth1 – the start value of W,
wth1+br1a - the last value of W, dpth1 – the start value of D, dpth1+br2a - the last value of
D. The parameters have default values but there is a possibility the user to change them.
7.2. Vsep-orb algorithm
It has two main steps S1 and S2:
S1) Determine the orbits Orb(G, Π) by an heuristic TREE4 algorithm (Figure 20);
S2) Use PART1, PART2 algorithms to determine GOO(G, ΠORB), where ΠORB is a
partition which cells to the orbits found in S1 (orbit partition).
TREE4 does not guarantee the computation of the exact orbits of GOO(G, Π) and
consequently Vsep-orb is an heuristic algorithm - it does not guarantee the computation of
the exact GOO(G, Π). But if the parameters of Vsep-orb are selected in proper way the
probability of the exact computation of GOO(G, Π) is very close to 1. The experiments
confirm this.
7.3. Vsep-sch algorithm
Vsep-sch algorithm has two main steps:
S1) Determine the base points, generators and orbits of Aut(G, Π) by TABLE1 heuristic
algorithm (fig. 18) applied with a parameter LB = 1;
S2) Determine the order Aut(G, Π) using Schreier-Sims algorithm [5,31,32] on the base
points and generators found in S1. Obviously, at this step the strong generators (see the
definition below) of Aut(G,Π) are determined.
Recall [32] that a base for a group A is a sequence of points B=[1,2,…, m] such that the
stabilizer 𝐴𝛽1,𝛽2 ,…,𝛽𝑚 = 𝐼. Schreier-Sims algorithm determines a chain of stabilizers
A=A(1) ≥ A(2) ≥ . . . ≥ A(m) ≥ A(m+1) = I, where 𝐴(𝑗) = 𝐴𝛽1,𝛽2,…,𝛽𝑖−1.
Each stabilizer A(j) is represented by a series of all right cosets to A(j+1)
A(j) = A(j+1) \ A(j) = A(j+1) ∪ A(j+1) α2 ∪ A(j+1) α3 ∪ . . . ∪ A(j+1) αn
The collection of elements in A(j+1) \ A(j) , j=1,2, . . . ,m are called strong generators for A
toward B. The algorithm [33] is constructing a table T(n,m) of strong generators, where
T(i,j) is an element of the group A which fixes 1, 2 ,…,j-1 and maps j to i. If no such
element exists, then the entry is empty. The non-empty elements of the j-th column of T
form a set of coset representatives U(j) of A(j+1) in A(j).
The step S2 algorithm (fig. 21) uses a procedure called sift (fig. 22) that for a given
automorpism α compares (j) with y=α((j)), j=1,2,…,m: (a) if y= (j) then a pass to the
next base point (j+1) follows; (b ) if y ≠ (j) then a check if α belongs to the stabilizer
A(j) follows, i.e. if T(y,j) is empty or no. If T(y,j) is empty then T(y,j) = α, the number of
nonempty elements in column j of T is increased by 1, c(j)=c(j)+1, and the sift ends. If
- 38 -
T(y,j)=γ then a new automorphism α = γα-1 is determined and all above actions are applied
to α for the next base points (j+1). The procedure sift modifies the table T by inserting at
most one new coset representative.
Step S2 algorithm is with base points - that is the difference from FHL version of Schreier-
Sims method in [33]. When the first 3 instructions of S2 algorithm are finished, the table T
should have the property that α ε A iff α can be expressed as a1a2 . . . am , where ai is a
member of the j-th column of T. This is called canonical representation of α toward B. The
time complexity of step S2 is O(n6) [33,34]
Input: n, generators, B - base points, m==B
Local variables: T(n, m)-2 dimensional integer array, c – integer array of size m, c(j) is
the number of nonempty elements in column j of T; Output: Aut(G, Π)
1; Set each element of c to 1;Set each element T(i,j) empty;
2: sift(n, α,B,m,T,c,) {Pass each generator to the procedure sift};
3: sift(n,γ,B,m,T,c,) {Pass the product γ of each pair of representatives in T to the
procedure sift until no new element is inserted in T};
4: Determine the order Aut(G, Π)=∏ 𝑐(𝑗)
.
𝑚
𝑗=1
Figure 21. S2 step of Vsep-sch algorithm
Input: n,α, B, m, T, c; c-array of size m=B,c(j) is the number of nonempty elements
in column j of T;Output: T,c
1: for j:=1 to m do
2: y:= α(B(j))
3: if(y=B(j)) cycle
4: if (T(y,j) is empty) then
5: T(y,j) := α; c(j) := c(j)+1; return
6: else
7: γ := T(y,j); α := γα-1
8: end if
9: end for
Figure 22. sift procedure
8. Vsep-a automatic version of Vsep algorithm
Experiments show that Vsep-sch algorithm runs long for graphs with large Aut(G) (small
number of orbits), especially for transitive graphs. In these cases T table is large and full or
almost full and evidently this causes more computation. On the other hand, Vsep-orb runs
fast on such graphs. This is the reason for developing Vsep-a algorithm that chooses Vsep-
orb when the number of orbits is small (≤ 3) and Vsep-sch otherwise. There are also other
criteria for this selection. Vsep-a algorithm also chooses Vsep-e algorithm for many graphs
on the base of some criteria, for example for some non-regular graphs.
- 39 -
9. Program implementation of Vsep algorithms
All proposed algorithms in this paper are implemented in Fortran programs that can be
compiled by any Compaq Fortran compiler. The program version published on author's
web site section My Programs [35] is called VSEP_PUB4. It implements Vsep-e, Vsep-
orb, Vsep-sch and Vsep-a version of VSEP algorithms. The user may choose any version
and set different parameters (see user guide).
10. Experimental results
In this section we present: a) experiments on almost all benchmark graphs from [19] that
compare the performance of VSEP_PUB4 program with Traces (nauty2.6r5 [19]) (Table
4) –one of the most competitive known tool for the worst cases; b) The running times of
VSEP_PUB4 for whole families of some benchmark graphs are given in Table 5.
Some of the experimental results from VSEP _PUB4 (as they are received from the
program) are in RESVSE4P_PUB4OBSHT file given in [35]. The most difficult graphs for
our algorithms are the graphs with Aut(G)=1 or with small Aut(G). It is known that none
of the known algorithms outperform others for all graphs. For each algorithm there are
specific difficult and easy graph families. A graph family may be easy for one algorithm
and very difficult for another. The same is for Vsep and Traces. Even more, for Vsep
different cell selectors give different running times – none of the cell selectors outperform
others. The chooser of cell selectors does not always choose the optimal cell selector. For
each result we show the cell selector for which it is obtained. The experiments were carried
out on a laptop Dell, CPU: Intel(R) Core (TM) i5-3317U@ 1.7 GHz, Memory: 8 GB, OS:
64 bit Microsoft Windows 7 Profesional. For the experiments, we have used all the
benchmark graphs of nauty&Traces page [19] that includes a variety of graph families
with different characteristics. We show mostly the results for the graphs that are worst
cases for either of the compared tools and cell selectors for which they are obtained (only
for Vsep). It is evident (See Table 4) that Vsep outperforms Traces considerably for the
graphs tnn(39)_1014-1, chh_cc(7-7)_1078-1, f-lex-srg-10-1 and f-lex-srg-50-1 . On the
other side Traces outperforms Vsep-e considerably for the graphs of projective planes (pp-
16-14, 15, 22, pp-25-90, 116) , had-112, had-176 and latin-sw-112 but for the same
graphs Vsep-a outperforms Traces considerably. There are no essential differences
between Vsep and Traces on the other graphs on the Table 4! Traces is slow for the graphs
with large order of the automorphism group and large number of generators – my
experience show that this is may be due to the use of Schreier-Sims method. The main
disadvantage of Vsep-e is the storing of the whole bouquet of the first selected vertex-
millions of words for some graphs. For almost all benchmark graph families the heuristic
versions are many times faster than the exact one and at correctly chosen parameters give
correct results.
- 40 -
G graph, n*, minval*, maxval*,
Aut(G), norb*
tnn(39)_1014-1, 1014,
4,312, 6.314790834154e174,12
cmz-50, 1200,2,5,
1.267650600228e32,8
ag-49,4851,49,50,
2.710632960000e10,2
cfi-200,2000,
3,3, 2.535301200456e30, 800
chh_cc(7-7)_1078-1
1078,3,45,4.907372642035e71,8
had-128,512,129,129,
1.073807313661e19,1
had-112,448,113,113, 1677312,1
T in seconds
Traces
vsep-e*
izb*,
bouquet
size*.par*
5654.38
0.08
1.01, 1,1,nop
0.83, 4,1,nop
0.14, 1, 1,nop
vsep-a*
selected
version(orb for
vsep-orb,shc for
vsep-sch,e-for
vsep-e),izb*,par*
1.03,e,1,nop
0.80,e,4,nop
0.13,e,1,nop
0.11
0.14, 1,21,nop
0.14,e,1,nop
0.14
NA
1.17, 1,4,nop
1.06,4,4,nop
0.17, 1,1,nop
1.17,e,1,nop
1.07,e,1,nop
0.17,e,1,nop
0.02
0.11,3,def
0.05
had-176,704,177,177. 15257088,1
0.14
had-256,1024,257,257,
1.401962828716e24,1
latin-30.dre,900 ,87,87, 43200.1
latin-sw-30-1,900 ,87,87, 1,900
0.05
0.03
0.17
had-sw-112,448,113,113, 2,224
1.61
lattice-30.dre, 900 , 58,58,
1.407181592771e65,1
10cube,1024,10,10, 3.715891200e9,1 0.02
0.05
85.75,1,32287,de
f
429.75,1,21210,d
ef
0.61,1,1,def
0.71,1,1106,def
0.66,4,962,def
4.33, 1,812;
1,2,1,1,1
179.7,1,
519624;
1,2,1,1,1
0.09, 1 ,1,def
0.11, 1,1,def
0.11,e,1,def
0.42.e,1,def
1.88,e,1,def
0.62,e,1,def
1.33.e,1,def
3.16,e,1;
1,2,1,1,1
16.69,sch,1,def
0.61,sch,
1,2,1,1,1
0.12,orb,1,def
0.11,orb, 1,def
0.15,orb, 1,def
2.00,orb,1,def
paley-461.dre ,461,230,230,
1.06030e5,1
pg2-49, 4902,50,50,
1.328752276992e14,1
pp-16-14,546,17,17, 2304,14
0.02
0.14, 1,1,def
0.11
2.00, 1,21,def
0.66
55.24, 1, 420960; 0.44,sch,1;
- 41 -
G graph, n*, minval*, maxval*,
Aut(G), norb*
T in seconds
Traces
vsep-e*
izb*,
bouquet
size*.par*
pp-16-22,546,17,17, 9216,10
0.55
pp-25-1,1302,26,26, 609336000000,2 0.05
pp-25-90,1302,26,26,1000,40
42.62
pp-25-116,1302,26,26, 500,64
104.82
pp-27-10, 1514,28, 28, 122472,6
1.05
pp-49-1,4902,50,50,288120,10
236.87
3,3,2,1,1
8.50, 2, 52680;
3,3,2,1,1
15.4, 1,117264;
3,3,2,1,1
2.26,2,13336;
3,3,2,1,1
0.12,2,10;
3,3,2,1,1
242.6, 2,
522995,
3,3,2,1,1;
2244.03,1,
5016276,def
500.0, 2 ,
1058568,
3,3,2,1,1
40.0,,2, 73008
3 5 2 1 1
5631.0, 2,
1975680;
2,5,2,2,1
5600.0, 4,
1975680
2,5,2,2,1
5529.05, 2,
1975680
2,5,2,2,1
vsep-a*
selected
version(orb for
vsep-orb,shc for
vsep-sch,e-for
vsep-e),izb*,par*
3,3,2,1,1
0.26,sch,1;
3,3,2,1,1
0.12,orb,1,
3,3,2,1,1
9.7,sch,1;
3,3,2,1,1
6.00,sch,1,
3,3,2,1,1
4.59,sch,2;
3 5 2 1 1
117.4,1;
2,5,2,2,1;
90.1,2;
2,4,2,2,1
0.22, orb,1,def
mz-aug-50,1000,3,6,
101412048018258352119736256430
08,250
mz-aug2-50,1200,2,5,
507060240091291760598681282150
4,700
mz-50,1000,3,3,
0.00
0.22,1,1,def
0.00
0.54,1,1,def
0.53,orb, 1,def
0.00
0.22,1,1,def
0.20,orb, 1,def
- 42 -
G graph, n*, minval*, maxval*,
Aut(G), norb*
101412048018258352119736256430
08,250
f-lex-reg-50-5,1700,
8,30,16384,1016
f-lex-srg-10-1,790,16,75,16,,439
T in seconds
Traces
vsep-e*
izb*,
bouquet
size*.par*
vsep-a*
selected
version(orb for
vsep-orb,shc for
vsep-sch,e-for
vsep-e),izb*,par*
0.06
0.42,1,2,def
0.42,orb,1,def
NA
NA
NA
0.17,sch,2;
2 2 1 2 2
10.69,sch,2;
2 2 1 2 2
NA
f-lex-srg-50-5,3950,
16,75,32768,2360
Table 4. Experimental comparison of VSEP4P_PUB4 program with Traces. n - number
of vertices, norb-number of orbits, vsep-e - exact version, vsep-a – automatic version, izb -
the numbering of the chosen cell selector; minval, maxval - minimal and maximal degree
of a vertex of the graph, bouquet size – number of the stored numerations; PAR –
parameters mbrsyvp, wth1, dpth1, br1a, br2a; 4,5,2 ,1,1 are default parameters
(def=default),nop-no parameters since there is no call to tree4;NA-not attended
Graph family
Parameters: mbrsyvp, wth1, dpth1, br1a, br2a; 4,5,2
,1,1-default (def)
mathon
mathon
dobling
tnn
tnn
cmz
ag
chh
cfi
cfi
latin
lattice
hypercubes3
paley
ppsmall
pp16
pp16
pp16
pp25
pp27
pp27
Def
Def
Def
Def
Def
Def
Def
Def
Def
Def
Def
Def
Def
Def
3 3 2 4 1
3 3 2 2 1
3 3 2 1 1
3 3 2 1 1
Def
3 5 2 1 1
izb T[secs]-running
time
0.1248
0.312
3.97
3.31
2.37
0.71
2.19
38.94
36.77
3.79
0.99
0.25
1.90
0.56
7.28
7.22
7.19
746.11
60.66
57.60
1
1
1
4
1
1
1
1
4
1
1
1
1
1
1
1
1
1
2
2
- 43 -
Graph family
Parameters: mbrsyvp, wth1, dpth1, br1a, br2a; 4,5,2
,1,1-default (def)
pp27
pp27
pp27
pp27
pp27
pp27
pp27
pp27
pp27
pp27
pp27
pp49
pp49
pp49
had
had
had
f-lex=srg
f-lex=reg
pg
pg
pg
mz-aug
mz-aug2
total run time for all cases in the table
3 5 2 1 1
3 4 3 4 2
3 4 3 4 2
2 4 3 4 2
3 4 3 3 2
3 4 3 4 1
2 4 3 3 2
2 4 3 3 1
2 4 3 4 1
2 4 3 2 1
2 4 3 1 1
3 5 2 2 1
2 5 2 2 1
2 5 2 2 1
4 5 2 1 1
4 5 2 1 1
3 6 2 1 1
2 2 1 2 2
Def
Def
3 4 2 1 1
3 3 2 1 1
Def
Def
izb T[secs]-running
3
1
2
2
2
2
2
2
2
2
2
1
2
1
1
3
3
2
1
1
1
1
1
1
time
58.26
68.53
88.54
70.76
88.90
60.31
76.32
68.62
75.48
69.28
68.82
505.00
349.60
365.10
293.28
278.40
281.36
589.29
18.47
7.84
7.80
7.80
2.09
3.77
4478.15
Table 5. Experimental results from VSEP4P_PUB4 program for whole families of the
most of the benchmark graphs in [19], some of the results are in RESVSEP4OBSHT file
[35], izb is the numbering of the chosen cell selector
11. Concluding remarks and open problems
Five new algorithms, named Vsep, are described. Four of them are for determining the
generators, orbits and order of an undirected graph automorphism group:. Vsep-e – exact,
Vsep-orb and Vsep-sch – heuristic and Vsep-a automatically selects the optimal version
among Vsep-e, Vsep-orb and Vsep-sch. The fifth algorithm, Vsep-is, is for finding an
isomorphism between two graphs. A new approach is used in the exact algorithm: if during
its execution some of the searched or intermediate variables obtain a wrong value then the
algorithm continues from a new start point losing some of the results determined so far
(cases CS3, CS4). The new start point is such that the correct results are obtained.
A new code, named S-code, of a partition of the graph vertices is proposed. S-code is used
for reducing the time of comparing the partitions in the proposed algorithms.
The experiments show that the worst case time complexity of Vsep-e for an arbitrary graph
is exponential but for some classes it is polynomial. The main difference of the exact one
and the widely known tools is the storing of the whole bouquet of numerations of the first
- 44 -
selected vertex in the first level. Five cell selectors are used in the algorithms and a chooser
of optimal cell selector is presented. Some of the cell selectors are new, namely
mxvectchval and mxprchval.
Experimental comparison of the proposed algorithms with Traces algorithm is made - it
shows their worst and best cases. A disadvantage of Vsep-e algorithm is its higher
requirements for memory (for some worst cases several millions of numbers are stored.
The worst cases for the algorithm Vsep-e are the graphs with smaller order Aut(G),
especially the rigid graphs. Vsep-orb and Vsep-sch heuristic algorithms are extremely fast
(with some exceptions) compared with the exact one and are almost exact - for all tested
thousands of graphs they give correct results. Practically, their requirements for memory
are very small. Heuristic algorithms are an important approach for solving so hard problem
as graph isomorphism is.
The future work on developing Vsep algorithms will include: a) search for a new cell
selector that will reduce the size of the search tree; b) search for a new chooser of a cell
selector; c) conduct a comparison of Vsep algorithms with other known GOO algorithms;
d) develop parallel versions of Vsep algorithms. Open problems are: i) Develop a new
chooser of cell selector that given a graph can determine the best cell selector (that yields a
minimal running time of the algorithm); ii) Develop an algorithm that given a graph and
the first numeration can determine the parameters mbrsyvp, wth1, dpth1, br1a, br2a of the
forest of reduced regular trees such that when used the heuristic versions of VSEP
algorithms will give correct results for a minimal running time; iii) Develop an algorithm
that given a graph on the base of some criteria to choose the optimal algorithm from the
known published algorithms.
Acknowledgements
The author thanks to Brendan McKay, Bill Kocay, Rudi Mathon, Vladimir Tonchev,
Ulrich Dempwolff, Gordon Royle, Adolfo Piperno, Jose Luis López-Presa, Petteri Kaski
and Hadi Katebi for the extensive discussions on the graph isomorphism problem and
exchanging graphs with interesting properties and papers on this problem. Special thanks
to my former diploma student Apostol Garchev who contributed in performing and
repeating some of the experiments on the developed algorithms. The author would like to
thank all colleagues and students who contributed to this study.
References
[1] A. V Aho, J.E. Hopcroft, J.D. Ullman, The Design and Analysis of Computer
Algorithms, 1974.
[2] C.C.E. Leiserson, R.R.L. Rivest, C. Stein, T.H. Cormen, Introduction to Algorithms,
Third Edition, 2009. doi:10.2307/2583667.
[3] E. by J. van Leeuwen, Handbook of Theoretical Computer Science, Volume B
Formal Models and Semantics, MIT Press, 1994.
https://mitpress.mit.edu/books/handbook-theoretical-computer-science-0.
edited by M.I. Nechepurenko, Algorithms and Programs for Solving Problems on
- 45 -
[4]
Graphs and Networks, Nauka, Novosibirsk, 1990.
http://mt3.org/bloog/wodolaz/39.html.
[5] K. D. L., D. R.Stinson, Combinatorial Algorithms, CRS Press (1999)., CRS Press,
1999.
[6] F. Harary, Graph theory 1736–1936, Hist. Math. 4 (1977) 480–481.
doi:10.1016/0315-0860(77)90099-4.
[7] F. Harary, E.M. Palmer, Graphical Enumeration, 1973.
http://oai.dtic.mil/oai/oai?verb=getRecord&metadataPrefix=html&identifie
r=AD0765600.
[8] R.B. Ash, Abstract Algebra : The Basic Graduate Year, Nature. 169 (1952) 167–168.
[9]
doi:10.1038/169167b0.
J.L. Gross, J. Yellen, Handbook of graph theory, 2004. doi:10.1016/0304-
3878(86)90037-4.
[10] C.M. Hoffmann, Group-Theoretic Algorithms and Graph Isomorphism, Lecture No,
Springer-Verlag ., 1982. http://link.springer.com/book/10.1007/3-540-11493-9.
[11] P.J. Cameron, Automorphisms of graphs, J. Math. Sci. 72 (1994) 3341–3343.
doi:10.1007/BF01261692.
[12] B.D. McKay, Practical Graph Isomorphism, Congr. Numer. 30 (1981) 45–87.
http://users.cecs.anu.edu.au/~bdm/papers/pgi.pdf.
[13] B.D. McKay, A. Piperno, Practical graph isomorphism, II, J. Symb. Comput. 60
(2014) 94–112. doi:10.1016/j.jsc.2013.09.003.
[14] L. Babai, Moderately Exponential Bound for Graph Isomorphism, Proceeding FCT
'81 Proc. 1981 Int. FCT-Conference Fundam. Comput. Theory. (1981) 34–50.
http://dl.acm.org/citation.cfm?id=739270.
[15] R. Mathon, A note on the graph isomorphism counting problem, Inf. Process. Lett. 8
(1979) 131–136.
http://www.sciencedirect.com/science/article/pii/0020019079900048.
[16] K.S.B. and C.J. Colbourn, Problems Polynomially Equivalent to Graph Isomorphism,
1979. https://cs.uwaterloo.ca/research/tr/1977/CS-77-04.pdf.
[17] V.N. Zemljachenko, N.M. Korneenko, R.I. Tishkevich, Graph isomorphism problem,
(1982) 83–158. doi:1007/BF02104746.
[18] L. Babai, Graph Isomorphism in Quasipolynomial Time, arXiv. (2016) 89.
http://arxiv.org/abs/1512.03547.
[19] B.D. McKay, A. Piperno, Nauty&Traces website, (n.d.). http://pallini.di.uniroma1.it/.
[20] D.G. Corneil, C.C. Gotlieb, An Efficient Algorithm for Graph Isomorphism, J. ACM.
17 (1970) 51–64. doi:10.1145/321556.321562.
[21] A.V.P. Faizullin,R.T., Heuristic algorithm for solving the graph isomorphism prob
lem, 2002. http://arxiv.org/pdf/math/0205220.pdf.
[22] G.-H.A.W.-G.T. KING, A New Graph Invariant for Graph Isomorphism:Probability
Propagation Matrix, J. Inf. Sci. Eng. 15 (1999) 337–352.
http://www.iis.sinica.edu.tw/page/jise/1999/199905_01.pdf.
[23] H. Ogata, W. Fujibuchi, S. Goto, M. Kanehisa, A heuristic graph comparison
algorithm and its application to detect functionally related enzyme clusters., Nucleic
Acids Res. 28 (2000) 4021–4028. doi:10.1093/nar/28.20.4021.
- 46 -
[24] H.B. Mittal, A Fast Backtrack Algorithm for Graph Isomorphism, Inf. Process. Lett.
29 (1988) 105–110. doi:10.1016/0020-0190(88)90037-3.
[25] S.D. Stoichev, Fast Adjacency Refinement Algorithm, in: Intern. Conf. Digit. Signal
Process., Limassol, Cyprus, 1995: pp. 870–875.
[26] J. Hopcroft, An n log n algorithm for minimizing states in a finite automaton,
Reproduction. (1971) 189–196.
http://oai.dtic.mil/oai/oai?verb=getRecord&metadataPrefix=html&identifier=AD071
9398.
[27] A. Piperno, Search Space Contraction in Canonical Labeling of Graphs (Preliminary
Version), arXiv. (2008) 1–21. http://arxiv.org/abs/0804.4881.
[28] D.G. Corneil, C.C. Gotlieb, An Efficient Algorithm for Graph Isomorphism, J. ACM.
17 (1970) 51–64. doi:10.1145/321556.321562.
[29] R. Mathon, Sample graphs for isomorphism testing, in: Proc. 9th S.-E.conf. Comb.
Graph Theory Comput. Boca - Raton, 1978: pp. 499–517.
[30] S.D. Stoichev, Vsep-New Heuristic and Exact Algorithms for Graph Automorphism
Group Computation, (n.d.). arXiv:1007.1726v4 [cs.DS].
[31] M. Jaggi, Implementations of 3 Types of the Schreier-Sims Algorithm, (2005).
http://www.m8j.net/data/List/Files-118/Documentation.pdf.
[32] E.A. O'Brien., Toward effective algorithms for linear groups, pp. 163–190, in: Finite
Geom. Groups, Comput. , Gruyter, Berlin (2006), Berlin, 2006: pp. 163–190.
https://www.math.auckland.ac.nz/~obrien/research/survey.pd.
[33] G.. Butler, Chapter 14. Complexity of the Schreier-Sims method, in: Lect. Notes
Comput. Sci. Vol. 559 1991 Fundam. Algorithms Permut. Groups, Springer, 1991.
http://link.springer.com/chapter/10.1007/3-540-54955-2_33#page-1.
[34] M. Furst, J. Hopcroft, E. Luks, Polynomial-time algorithms for permutation groups,
in: Found. Comput. Sci. …, 1980: pp. 36–41.
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4567802.
[35] S.D. Stoichev, S.D. Stoichev's web page, (n.d.).
https://sites.google.com/site/stoichostoichev2/my-programs.
- 47 -
|
1510.01800 | 4 | 1510 | 2017-04-10T19:21:24 | Logarithmic regret bounds for Bandits with Knapsacks | [
"cs.DS"
] | Optimal regret bounds for Multi-Armed Bandit problems are now well documented. They can be classified into two categories based on the growth rate with respect to the time horizon $T$: (i) small, distribution-dependent, bounds of order of magnitude $\ln(T)$ and (ii) robust, distribution-free, bounds of order of magnitude $\sqrt{T}$. The Bandits with Knapsacks model, an extension to the framework allowing to model resource consumption, lacks this clear-cut distinction. While several algorithms have been shown to achieve asymptotically optimal distribution-free bounds on regret, there has been little progress toward the development of small distribution-dependent regret bounds. We partially bridge the gap by designing a general-purpose algorithm with distribution-dependent regret bounds that are logarithmic in the initial endowments of resources in several important cases that cover many practical applications, including dynamic pricing with limited supply, bid optimization in online advertisement auctions, and dynamic procurement. | cs.DS | cs |
Logarithmic Regret Bounds for Bandits with Knapsacks*
Arthur Flajolet
Operations Research Center
Massachusetts Institute of Technology
Cambridge, MA 02139
[email protected]
Patrick Jaillet
Department of Electrical Engineering and Computer Science
Laboratory for Information and Decision Systems
Operations Research Center
Massachusetts Institute of Technology
Cambridge, MA 02139
[email protected]
Optimal regret bounds for Multi-Armed Bandit problems are now well documented. They can be classified into two cat-
egories based on the growth rate with respect to the time horizon T : (i) small, distribution-dependent, bounds of order of
magnitude ln(T ) and (ii) robust, distribution-free, bounds of order of magnitude √T . The Bandits with Knapsacks model,
an extension to the framework allowing to model resource consumption, lacks this clear-cut distinction. While several
algorithms have been shown to achieve asymptotically optimal distribution-free bounds on regret, there has been little
progress toward the development of small distribution-dependent regret bounds. We partially bridge the gap by design-
ing a general-purpose algorithm with distribution-dependent regret bounds that are logarithmic in the initial endowments
of resources in several important cases that cover many practical applications, including dynamic pricing with limited
supply, bid optimization in online advertisement auctions, and dynamic procurement.
1. Introduction.
1.1. Motivation. Multi-Armed Bandit (MAB) is a benchmark model for repeated decision making in
stochastic environments with very limited feedback on the outcomes of alternatives. In these circumstances,
a decision maker must strive to find an overall optimal sequence of decisions while making as few subop-
timal ones as possible when exploring the decision space in order to generate as much revenue as possible,
a trade-off coined exploration-exploitation. The original problem, first formulated in its predominant ver-
sion in Robbins [21], has spurred a new line of research that aims at introducing additional constraints that
reflect more accurately the reality of the decision making process. Bandits with Knapsacks (BwK), a model
formulated in its most general form in Badanidiyuru et al. [8], fits into this framework and is characterized
by the consumption of a limited supply of resources (e.g. time, money, and natural resources) that comes
with every decision. This extension is motivated by a number of applications in electronic markets such as
dynamic pricing with limited supply, see Besbes and Zeevi [11] and Babaioff et al. [6], online advertising,
see Slivkins [22], online bid optimization for sponsored search auctions, see Tran-Thanh et al. [26], and
crowdsourcing, see Badanidiyuru et al. [7]. A unifying paradigm of online learning is to evaluate algo-
rithms based on their regret performance. In the BwK theory, this performance criterion is expressed as the
gap between the total payoff of an optimal oracle algorithm aware of how the rewards and the amounts of
resource consumption are generated and the total payoff of the algorithm. Many approaches have been pro-
posed to tackle the original MAB problem, where time is the only limited resource with a prescribed time
* Research funded in part by ONR, grants N00014-12-1-0033 and N00014-15-1-2083.
1
2
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
horizon T , and the optimal regret bounds are now well documented. They can be classified into two cate-
gories with qualitatively different asymptotic growth rates. Many algorithms, such as UCB1, see Auer et al.
[4], Thompson sampling, see Agrawal and Goyal [3], and ǫ-greedy, see Auer et al. [4], achieve distribution-
dependent, i.e. with constant factors that depend on the underlying unobserved distributions, asymptotic
bounds on regret of order Θ(ln(T )), which is shown to be optimal in Lai and Robbins [19]. While these
results prove very satisfying in many settings, the downside is that the bounds can get arbitrarily large if a
malicious opponent was to select the underlying distributions in an adversarial fashion. In contrast, algo-
rithms such as Exp3, designed in Auer et al. [5], achieve distribution-free bounds that can be computed in an
online fashion, at the price of a less attractive growth rate Θ(√T ). The BwK theory lacks this clear-cut dis-
tinction. While provably optimal distribution-free bounds have recently been established, see Agrawal and
Devanur [1] and Badanidiyuru et al. [8], there has been little progress toward the development of asymptot-
ically optimal distribution-dependent regret bounds. To bridge the gap, we introduce a template algorithm
with proven regret bounds which are asymptotically logarithmic in the initial supply of each resource, in
four important cases that cover a wide range of applications:
• Case 1, where there is a single limited resource other than time, which is not limited, and the amount
of resource consumed as a result of making a decision is stochastic. Applications in online advertising, see
Tran-Thanh et al. [23], fit in this framework;
• Case 2, where there are arbitrarily many resources and the amounts of resources consumed as a result of
making a decision are deterministic. Applications to network revenue management of perishable goods, see
Besbes and Zeevi [11], shelf optimization of perishable goods, see Graczov´a and Jacko [15], and wireless
sensor networks, see Tran-Thanh et al. [25], fit in this framework;
• Case 3, where there are two limited resources, one of which is assumed to be time while the consump-
tion of the other is stochastic, under a nondegeneracy condition. Typical applications include online bid
optimization in sponsored search auctions, see Tran-Thanh et al. [26], dynamic pricing with limited supply,
see Babaioff et al. [6], and dynamic procurement, see Badanidiyuru et al. [7];
• Case 4, where there are arbitrarily many resources, under a stronger nondegeneracy condition than for
Case 3. Typical applications include dynamic ad allocation, see Slivkins [22], dynamic pricing of multiple
products, see Badanidiyuru et al. [8], and network revenue management, see Besbes and Zeevi [11].
In terms of applicability and significance of the results, Case 3 is the most important case. Case 4 is the
most general one but the analysis is more involved and requires stronger assumptions which makes it less
attractive from a practical standpoint. The analysis is easier for Cases 1 and 2 but their modeling power is
more limited.
In fast-paced environments, such as in ad auctions, the stochastic assumptions at the core of the BwK
model are only valid for a short period of time but there are typically a large number of actions to be
performed per second (e.g. submit a bid for a new ad auction). In these situations, the initial endowments
of resources are thus typically large and logarithmic regret bounds can be significantly more attractive than
distribution-free ones.
1.2. Problem statement and contributions. At each time period t ∈ N, a decision needs to be made
among a predefined finite set of actions, represented by arms and labeled k = 1,··· , K. We denote by at
the arm pulled at time t. Pulling arm k at time t yields a random reward rk,t ∈ [0, 1] (after scaling) and
incurs the consumption of C ∈ N different resource types by random amounts ck,t(1),··· , ck,t(C) ∈ [0, 1]C
(after scaling). Note that time itself may or may not be a limited resource. At any time t and for any arm k,
the vector (rk,t, ck,t(1),··· , ck,t(C)) is jointly drawn from a fixed probability distribution νk independently
from the past. The rewards and the amounts of resource consumption can be arbitrarily correlated across
arms. We denote by (Ft)t∈N the natural filtration generated by the rewards and the amounts of resource
consumption revealed to the decision maker, i.e. ((rat,t, cat,t(1),··· , cat,t(C)))t∈N. The consumption of any
resource i ∈ {1,··· , C} is constrained by an initial budget B(i) ∈ R+. As a result, the decision maker can
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
3
keep pulling arms only so long as he does not run out of any of the C resources and the game ends at time
period τ ∗, defined as:
τ ∗ = min{t ∈ N ∃i ∈ {1,··· , C},
t
Xτ =1
caτ ,τ (i) > B(i)}.
(1)
Note that τ ∗ is a stopping time with respect to (Ft)t≥1. When it comes to choosing which arm to pull next,
the difficulty for the decision maker lies in the fact that none of the underlying distributions, i.e. (νk)k=1,··· ,K,
are initially known. Furthermore, the only feedback provided to the decision maker upon pulling arm at (but
prior to selecting at+1) is (rat,t, cat,t(1),··· , cat,t(C)), i.e. the decision maker does not observe the rewards
that would have been obtained and the amounts of resources that would have been consumed as a result
of pulling a different arm. The goal is to design a non-anticipating algorithm that, at any time t, selects at
based on the information acquired in the past so as to keep the pseudo regret defined as:
RB(1),··· ,B(C) = EROPT(B(1),··· , B(C))− E[
τ ∗−1
Xt=1
rat,t],
(2)
as small as possible, where EROPT(B(1),··· , B(C)) is the maximum expected sum of rewards that can be
obtained by a non-anticipating oracle algorithm that has knowledge of the underlying distributions. Here,
an algorithm is said to be non-anticipating if the decision to pull a given arm does not depend on the future
observations. We develop algorithms and establish distribution-dependent regret bounds, that hold for any
choice of the unobserved underlying distributions (νk)k=1,··· ,K, as well as distribution-independent regret
bounds. This entails studying the asymptotic behavior of RB(1),··· ,B(C) when all the budgets (B(i))i=1,··· ,C
go to infinity. In order to simplify the analysis, it is convenient to assume that the ratios (B(i)/B(C))i=1,···,C
are constants independent of any other relevant quantities and to denote B(C) by B.
b(i) is a positive quantity.
ASSUMPTION 1. For any resource i ∈ {1,··· , C}, we have B(i) = b(i) · B for some fixed constant
b(i) ∈ (0, 1]. Hence b = min
i=1,··· ,C
When time is a limited resource, we use the notation T in place of B. Assumption 1 is widely used in the
dynamic pricing literature where the inventory scales linearly with the time horizon, see Besbes and Zeevi
[11] and Johnson et al. [17]. Assumption 1 will only prove useful when deriving distribution-dependent
regret bounds and it can largely be relaxed, see Section A of the Appendix.
As the mean turns out to be an important statistics, we denote the mean reward and amounts of resource
consumption by µr
k(C) and their respective empirical estimates by ¯rk,t, ¯ck,t(1),··· , ¯ck,t(C).
These estimates depend on the number of times each arm has been pulled by the decision maker up to, but
not including, time t, which we write nk,t. We end with a general assumption, which we use throughout the
paper, meant to have the game end in finite time.
k, µc
k(1),··· , µc
ASSUMPTION 2. For any arm k ∈ {1,··· , K}, we have max
i=1,··· ,C
µc
k(i) > 0.
Note that Assumption 2 is automatically satisfied if time is a limited resource.
Contributions. We design an algorithm that runs in time polynomial in K for which we establish
O(K C · ln(B)/∆) (resp.pK C · B · ln(B)) distribution-dependent (resp. distribution-free) regret bounds,
where ∆ is a parameter that generalizes the optimality gap for the standard MAB problem. We estab-
lish these regret bounds in four cases of increasing difficulty making additional technical assumptions that
become stronger as we make progress towards tackling the general case. We choose to present these inter-
mediate cases since: (i) we get improved constant factors under weaker assumptions and (ii) they subsume
many practical applications. Note that our distribution-dependent regret bounds scale as a polynomial func-
tion of K of degree C, which may be unacceptable when the number of resources is large. We provide
evidence that suggests that a linear dependence on K can be achieved by tweaking the algorithm, at least in
4
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
some particular cases of interest. Finally, we point out that the constant factors hidden in the O notations are
not scale-free, in the sense that jointly scaling down the amounts of resources consumed at each round along
with their respective initial endowments worsens the bounds. As a consequence, initially scaling down the
amounts of resource consumption in order to guarantee that they lie in [0, 1] should be done with caution:
the scaling factors should be as small as possible.
1.3. Literature review. The Bandits with Knapsacks framework was first introduced in its full gener-
ality in Badanidiyuru et al. [8], but special cases had been studied before, see for example Tran-Thanh et al.
[23], Ding et al. [13], and Babaioff et al. [6]. Since the standard MAB problem fits in the BwK framework,
with time being the only scarce resource, the results listed in the introduction tend to suggest that regret
bounds with logarithmic growth with respect to the budgets may be possible for BwK problems but very few
such results are documented. When there are arbitrarily many resources and a time horizon, Badanidiyuru
et al. [8] and Agrawal and Devanur [1] obtain O(√K · T ) distribution-free bounds on regret that hold on
average as well as with high probability, where the O notation hides logarithmic factors. These results were
later extended to the contextual version of the problem in Badanidiyuru et al. [9] and Agrawal et al. [2].
Johnson et al. [17] extend Thompson sampling to tackle the general BwK problem and obtain distribution-
dependent bounds on regret of order O(√T ) (with an unspecified dependence on K) when time is a limited
resource, under a nondegeneracy condition. Trovo et al. [27] develop algorithms for BwK problems with a
continuum of arms and a single limited resource constrained by a budget B and obtain o(B) regret bounds.
Combes et al. [12] consider a closely related framework that allows to model any history-dependent con-
straint on the number of times any arm can be pulled and obtain O(K · ln(T )) regret bounds when time
is a limited resource. However, the benchmark oracle algorithm used in Combes et al. [12] to define the
regret is significantly weaker than the one considered here as it only has knowledge of the distributions of
the rewards, as opposed to the joint distributions of the rewards and the amounts of resource consumption.
Babaioff et al. [6] establish a Ω(√T ) distribution-dependent lower bound on regret for a dynamic pric-
ing problem which can be cast as a BwK problem with a time horizon, a resource whose consumption is
stochastic, and a continuum of arms. This lower bound does not apply here as we are considering finitely
many arms and it is well known that the minimax regret can be exponentially smaller when we move from
finitely many arms to uncountably many arms for the standard MAB problem, see Kleinberg and Leighton
[18]. Tran-Thanh et al. [24] tackles BwK problems with a single limited resource whose consumption is
deterministic and constrained by a global budget B and obtain O(K · ln(B)) regret bounds. This result was
later extended to the case of a stochastic resource in Xia et al. [30]. Wu et al. [29] study a contextual version
of the BwK problem when there are two limited resources, one of which is assumed to be time while the
consumption of the other is deterministic, and obtain O(K · ln(T )) regret bounds under a nondegeneracy
condition. Logarithmic regret bounds are also derived in Slivkins [22] for a dynamic ad allocation problem
that can be cast as a BwK problem.
Organization. The remainder of the paper is organized as follows. We present applications of the BwK
model in Section 2. We expose the algorithmic ideas underlying our approach in Section 3 and apply theses
ideas to Cases (1), (2), (3), and (4) in Sections 4, 5, 6, and 7 respectively. We choose to discuss each case
separately in a self-contained fashion so that readers can delve into the setting they are most interested in.
This comes at the price of some overlap in the analysis. We relax some of the assumptions made in the
course of proving the regret bounds and discuss extensions in Section A of the Appendix. To provide as
much intuition as possible, the ideas and key technical steps are all included in the main body, sometimes
through proof sketches, while the technical details are deferred to the Appendix.
Notations. For a set S, S denotes the cardinality of S while 1S is the indicator function of S. For
a vector x ∈ Rn and S a subset of {1,··· , n}, xS refers to the subvector (xi)i∈S. For a square matrix A,
det(A) is the determinant of A while adj(A) denotes its adjugate. For x ∈ R, x+ is the positive part of x.
We use standard asymptotic notations such as O(·), o(·), Ω(·), and Θ(·).
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
5
2. Applications. A number of applications of the BwK framework are documented in the literature. For
the purpose of being self-contained, we review a few popular ones that satisfy our technical assumptions.
2.1. Online advertising.
Bid optimization in repeated second-price auctions. Consider a bidder participating in sealed second-
price auctions who is willing to spend a budget B. This budget may be allocated only for a period of time (for
the next T auctions) or until it is completely exhausted. Rounds, indexed by t ∈ N, correspond to auctions
the bidder participates in. If the bid submitted by the bidder for auction t is larger than the highest bid
submitted by the competitors, denoted by mt, the bidder wins the auction, derives a private utility vt ∈ [0, 1]
(whose monetary value is typically difficult to assess), and is charged mt. Otherwise, mt is not revealed
to the bidder and vt cannot be assessed. We consider a stochastic setting where the environment and the
competitors are not fully adversarial: ((vt, mt))t∈N is assumed to be an i.i.d. stochastic process. The goal
for the bidder is to design a strategy to maximize the expected total utility derived given that the bidder has
selected a grid of bids to choose from (b1,··· , bK) (e.g. b1 = $0.10,··· , bK = $1). This is a BwK problem
with two resources: time and money. Pulling arm k at round t corresponds to bidding bk in auction t, costs
ck,t = mt · 1bk≥mt , and yields a reward rk,t = vt · 1bk≥mt . Weed et al. [28] design bidding strategies for a
variant of this problem where the bidder is not limited by a budget and rk,t = (vt − mt)· 1bk≥mt .
This model was first formalized in Tran-Thanh et al. [26] in the context of sponsored search auctions.
In sponsored search auctions, advertisers can bid on keywords to have ads (typically in the form of a link
followed by a text description) displayed alongside the search results of a web search engine. When a user
types a search query, a set of relevant ads are selected and an auction is run in order to determine which
ones will be displayed. The winning ads are allocated to ad slots based on the outcome of the auction and,
in the prevailing cost-per-click pricing scheme, their owners get charged only if the user clicks on their
ads. Because the auction is often a variant of a sealed second-price auction (e.g. a generalized second-
price auction), very limited feedback is provided to the advertiser if the auction is lost. In addition, both
the demand and the supply cannot be predicted ahead of time and are thus commonly modeled as random
variables, see Ghosh et al. [14]. For these reasons, bidding repeatedly on a keyword can be formulated
as a BwK problem. In particular, when the search engine has a single ad slot per query, this problem can
be modeled as above: B is the budget the advertiser is willing to spend on a predetermined keyword and
rounds correspond to ad auctions the advertiser has been selected to participate in. If the advertiser wins the
auction, his or her ad gets displayed and he or she derives a utility vt = 1At , where At is the event that the
ad gets clicked on. The goal is to maximize the expected total number of clicks given the budget constraint.
The advertiser may also be interested in optimizing the ad to be displayed, which will affect the probability
of a click. In this case, the modeling is similar but arms correspond to pairs of bid values and ads.
Dynamic ad allocation. This problem was first modeled in the BwK framework in Slivkins [22]. A
publisher, i.e. the owner of a collection of websites where ads can be displayed, has previously agreed with
K advertisers, indexed by k ∈ {1,··· , K}, on a predetermined cost-per-click pk. Additionally, advertiser
k is not willing to spend more than a prescribed budget, Bk, for a predetermined period of time (which
corresponds to the next T visits or rounds). Denote by Ak
t the event that the ad provided by advertiser k gets
clicked on at round t. We consider a stochastic setting where the visitors are not fully adversarial: (1Ak
)t∈N
is assumed to be an i.d.d. stochastic process for any advertiser k. The goal for the publisher is to maximize
the total expected revenues by choosing which ad to display at every round, i.e. every time somebody visits
one of the websites. This situation can be modeled as a BwK problem with K + 1 resources: time and
money for each of the K advertisers. Pulling arm k ∈ {1,··· , K} at round t corresponds to displaying the
ad owned by advertiser k, incurs the costs ck,t(i) = pk · 1Ak
t · 1i=k to advertiser i ∈ {1,··· , K}, and yields
a revenue rk,t = pk · 1Ak
.
t
t
2.2. Revenue management.
6
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
Dynamic pricing with limited supply. This BwK model was first proposed in Babaioff et al. [6]. An
agent has B identical items to sell to T potential customers that arrive sequentially. Customer t ∈ {1,··· , T}
is offered a take-it-or-leave-it price pt and purchases the item only if pt is no larger than his or her own
valuation vt, which is never disclosed. Customers are assumed to be non-strategic in the sense that their
valuations are assumed to be drawn i.i.d. from a distribution unknown to the agent. The goal for the agent is
to maximize the total expected revenues by offering prices among a predetermined list (p1,··· , pK). This
is a BwK problem with two resources: time and item inventory. Pulling arm k ∈ {1,··· , K} at round t
corresponds to offering the price pk, depletes the inventory of ck,t = 1pk≤vt unit, and generates a revenue
rk,t = pk · 1pk≤vt . Badanidiyuru et al. [8] propose an extension where multiple units of M different products
may be offered to a customer, which then buys as many as needed of each kind in order to maximize his or
her own utility function. In this case, the modeling is similar but arms correspond to vectors of dimension
2M specifying the number of items offered along with the price tag for each product and there are M + 1
resources: time and item inventory for each of the M products.
t
mk
t
t
t
Network revenue management.
Non-perishable goods. This is an extension of the dynamic pricing problem developed in Besbes and
Zeevi [11] which is particularly suited for applications in the online retailer industry, e.g. the online fashion
sample sales industry, see Johnson et al. [17]. Each product m = 1,··· , M is produced from a finite amount
of C different kinds of raw materials (which may be products themselves). Producing one unit of product
m ∈ {1,··· , M} consumes a deterministic amount of resource i ∈ {1,··· , C} denoted by cm(i). Customer
t ∈ {1,··· , T} is offered a product mt ∈ {1,··· , M} along with a take-it-or-leave-it price pmt
and purchases
is larger than pmt
it if his or her valuation vmt
. Products are manufactured online as customers order them.
t ,··· , vM
We assume that ((v1
t ))t∈N is an i.i.d. stochastic process with distribution unknown to the agent.
This is a BwK problem with C + 1 resources: time and the initial endowment of each resource. Given a
predetermined list of arms ((mk, pk))k=1,··· ,K, pulling arm k at round t corresponds to offering product mk
at the price pk, incurs the consumption of resource i by an amount ck,t(i) = cmk(i)· 1pk≤v
, and generates
a revenue rk,t = pk · 1pk≤v
Perishable goods. This is a variant of the last model developed for perishable goods, with appli-
cations in the food retail industry and the newspaper industry. At each time period t ∈ {1,··· , T},
a retailer chooses how many units λm
t ∈ N of product m ∈ {1,··· , M} to manufacture along with a
price offer for it pm
t . At time t, the demand for product m sold at the price p is a random quantity
denoted by dm
t (p). We assume that customers are non-strategic: for any vector of prices (p1,··· , pm),
((d1
t (p1),··· dM
t (pM )))t∈N is an i.i.d. stochastic process with distribution unknown to the agent. Products
perish at the end of each round irrespective of whether they have been purchased. Given a predetermined
list of arms ((λ1
k,··· , λM
k units of
product m at the price pm
k for any m ∈ {1,··· , M}, incurs the consumption of resource i by a determinis-
tic amount ck,t(i) =PM
m=1 λm
k · cm(i) (where cm(i) is defined in the previous paragraph), and generates a
revenue rk,t =PM
k · min(dm
Shelf optimization for perishable goods. This is a variant of the model introduced in Graczov´a and
Jacko [15]. Consider a retailer who has an unlimited supply of M different types of products. At each time
period t, the retailer has to decide how many units, λm
t , of each product, m ∈ {1,··· , M}, to allocate to
a promotion space given that at most N items fit in the limited promotion space. Moreover, the retailer
also has to decide on a price tag pm
for each product m. All units of product m ∈ {1,··· , M} perish
t
by time period Tm and the retailer is planning the allocation for the next T time periods. At round t,
the demand for product m is a random quantity denoted by dm
t (p). Customers are non-strategic: for any
vector of prices (p1,··· , pm), ((d1
t (pM )))t∈N is an i.i.d. stochastic process with distribution
unknown to the agent. This is a BwK problem with M + 1 resources: time horizon and time after which each
product perishes. Given a predetermined list of arms ((λ1
k ≤
K for any k ∈ {1,··· , K}, pulling arm k at round t corresponds to allocating λm
k units of product m to
k ))k=1,··· ,K, pulling arm k at round t corresponds to offering λm
k ))k=1,··· ,K satisfyingPM
t (p1),··· dM
t (pm
k ), λm
k ).
k, p1
k,··· , λM
k , pM
m=1 λm
k, p1
k , pM
.
mk
t
m=1 pm
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
7
the promotion space for every m ∈ {1,··· , M} with the respective price tags (p1
consumption of resource i by a deterministic amount ck,t(i) = 1, and generates a revenue rk,t =PM
min(dm
k,··· , pM
k ), incurs the
m=1 pm
k ·
k ).
t (pm
k ), λm
2.3. Dynamic procurement. This problem was first studied in Badanidiyuru et al. [7]. Consider a
buyer with a budget B facing T agents arriving sequentially, each interested in selling one good. Agent
t ∈ {1,··· , T} is offered a take-it-or-leave-it price, pt, and makes a sell only if the value he or she attributes
to the item, vt, is no larger than pt. We consider a stochastic setting where the sellers are not fully adver-
sarial: (vt)t∈N is an i.i.d. stochastic process with distribution unknown to the buyer. The goal for the buyer
is to maximize the total expected number of goods purchased by offering prices among a predetermined
list (p1,··· , pK). This is a BwK problem with two resources: time and money. Pulling arm k at round t
corresponds to offering the price pk, incurs a cost ck,t = pk · 1pk≥vt , and yields a reward rk,t = 1pk≥vt . It
is also possible to model situations where the agents are selling multiple types of products and/or multiple
units, in which case arms correspond to vectors specifying the number of units of each product required
along with their respective prices, see Badanidiyuru et al. [8].
Applications of this model to crowdsourcing platforms are described in Badanidiyuru et al. [7] and
Badanidiyuru et al. [8]. In this setting, agents correspond to workers that are willing to carry out microtasks
which are submitted by buyers (called "requesters") using a posted-price mechanism. Requesters are typi-
cally submitting large batches of jobs and can thus adjust the posted prices as they learn about the pool of
workers.
2.4. Wireless sensor networks. This is a variant of the model introduced in Tran-Thanh et al. [25].
Consider an agent collecting information using a network of wireless sensors powered by batteries. Acti-
vating sensor k ∈ {1,··· , K} consumes some amount of energy, ck, which is depleted from the sensor's
initial battery level, Bk, and triggers a measurement providing a random amount of information (measured
in bits), rk,t, which is transmitted back to the agent. Sensors cannot harvest energy and the goal for the agent
is to maximize the total expected amount of information collected over T actions. This is a BwK problem
with K + 1 resources: time and the energy stored in the battery of each sensor. Pulling arm k ∈ {1,··· , K}
corresponds to activating sensor k, incurs the consumption of resource i ∈ {1,··· , K} by a deterministic
amount ck,t = ck · 1k=i, and yields a random reward rk,t.
3. Algorithmic ideas.
3.1. Preliminaries. To handle the exploration-exploitation trade-off, an approach that has proved to be
particularly successful hinges on the optimism in the face of uncertainty paradigm. The idea is to consider
all plausible scenarios consistent with the information collected so far and to select the decision that yields
the most revenue among all the scenarios identified. Concentration inequalities are intrinsic to the paradigm
as they enable the development of systematic closed form confidence intervals on the quantities of interest,
which together define a set of plausible scenarios. We make repeated use of the following result.
LEMMA 1. Hoeffding's inequality Consider X1,··· , Xn n random variables with support in [0, 1].
If ∀t ≤ n E[Xt X1,··· , Xt−1] ≤ µ, then P[X1 +··· + Xn ≥ nµ + a] ≤ exp(− 2a2
If ∀t ≤ n E[Xt X1,··· , Xt−1] ≥ µ, then P[X1 +··· + Xn ≤ nµ− a] ≤ exp(− 2a2
n ) ∀a ≥ 0.
n ) ∀a ≥ 0.
Auer et al. [4] follow the optimism in the face of uncertainty paradigm to develop the Upper Confidence
Bound algorithm (UCB1). UCB1 is based on the following observations: (i) the optimal strategy always
consists in pulling the arm with the highest mean reward when time is the only limited resource, (ii) infor-
k ∈ [¯rk,t − ǫk,t, ¯rk,t + ǫk,t] at time t with probability at least 1 − 2/t3 for
mally, Lemma 1 shows that µr
ǫk,t =p2 ln(t)/nk,t, irrespective of the number of times arm k has been pulled. Based on these observa-
tions, UCB1 always selects the arm with highest UCB index, i.e. at ∈ arg maxk=1,··· ,K Ik,t, where the UCB
8
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
index of arm k at time t is defined as Ik,t = ¯rk,t + ǫk,t. The first term can be interpreted as an exploita-
tion term, the ultimate goal being to maximize revenue, while the second term is an exploration term, the
smaller nk,t, the bigger it is. This fruitful paradigm go well beyond this special case and many extensions of
UCB1 have been designed to tackle variants of the MAB problem, see for example Slivkins [22]. Agrawal
and Devanur [1] embrace the same ideas to tackle BwK problems. The situation is more complex in this
all-encompassing framework as the optimal oracle algorithm involves pulling several arms. In fact, find-
ing the optimal pulling strategy given the knowledge of the underlying distributions is already a challenge
in its own, see Papadimitriou and Tsitsiklis [20] for a study of the computational complexity of similar
problems. This raises the question of how to evaluate EROPT(B(1),··· , B(C)) in (2). To overcome this
issue, Badanidiyuru et al. [8] upper bound the total expected payoff of any non-anticipating algorithm by
the optimal value of a linear program, which is easier to compute.
LEMMA 2. Adapted from Badanidiyuru et al. [8]
The total expected payoff of any non-anticipating algorithm is no greater than B times the optimal value
of the linear program:
K
sup
(ξk)k=1,··· ,K
subject to
plus the constant term max
k=1,··· ,K
i=1,··· ,C
with µc
k(i)>0
µr
k
k (i) .
µc
µr
k · ξk
K
Xk=1
Xk=1
ξk ≥ 0,
k = 1,··· , K.
µc
k(i)· ξk ≤ b(i),
i = 1,··· , C
(3)
The optimization problem (3) can be interpreted as follows. For any arm k, B · ξk corresponds to the
expected number of times arm k is pulled by the optimal algorithm. Hence, assuming we introduce a dummy
arm 0 which is equivalent to skipping the current round, ξk can be interpreted as the probability of pulling
arm k at any round when there is a time horizon T . Observe that the constraints restrict the feasible set of
expected number of pulls by imposing that the amounts of resources consumed are no greater than their
respective budgets in expectations, as opposed to almost surely which would be a more stringent constraint.
This explains why the optimal value of (3) is larger than the maximum achievable payoff. In this paper,
we use standard linear programming notions such as the concept of a basis and a basic feasible solution.
We refer to Bertsimas and Tsitsiklis [10] for an introduction to linear programming. A pseudo-basis x is
described by two subsets Kx ⊂ {1,··· , K} and Cx ⊂ {1,··· , C} such that Kx = Cx. A pseudo-basis x is
a basis for (3) if the matrix Ax = (µc
k(i))(i,k)∈Cx×Kx is invertible. Furthermore, x is said to be a feasible basis
for (3) if the corresponding basic solution, denoted by (ξx
k = 0 for k 6∈ Kx
and Axξx
Kx = bCx (where bCx is the subvector (b(i))i∈Cx), is feasible for (3). When x is a feasible basis for
k )k=1,··· ,K and determined by ξx
k its objective function. From Lemma 2, we derive:
(3), we denote by objx =PK
k=1 µr
k · ξx
τ ∗
Xt=1
RB(1),··· ,B(C) ≤ B · objx∗ − E[
rat,t] + O(1),
(4)
where x∗ is an optimal feasible basis for (3). For mathematical convenience, we consider that the game
carries on even if one of the resources is already exhausted so that at is well defined for any t ∈ N. Of
course, the rewards obtained for t ≥ τ ∗ are not taken into account in the decision maker's payoff when
establishing regret bounds.
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
9
3.2. Solution methodology. Lemma 2 also provides insight into designing algorithms. The idea is to
incorporate confidence intervals on the mean rewards and the mean amounts of resource consumption into
the offline optimization problem (3) and to base the decision upon the resulting optimal solution. There
are several ways to carry out this task, each leading to a different algorithm. When there is a time horizon
T , Agrawal and Devanur [1] use high-probability lower (resp. upper) bounds on the mean amounts of
resource consumption (resp. rewards) in place of the unknown mean values in (3) and pull an arm at random
according to the resulting optimal distribution. Specifically, at any round t, the authors suggest to compute
an optimal solution (ξ∗k,t)k=1,··· ,K to the linear program:
K
sup
(ξk )k=1,··· ,K
subject to
(¯ck,t(i)− ǫk,t)· ξk ≤ (1− γ)· b(i),
i = 1,··· , C − 1
(5)
(¯rk,t + ǫk,t)· ξk
K
Xk=1
Xk=1
Xk=1
ξk ≥ 0,
K
ξk ≤ 1
k = 1,··· , K,
for a well-chosen γ ∈ (0, 1), and then to pull arm k with probability ξ∗k,t or skip the round with probability
1 −PK
k=1 ξ∗k,t. If we relate this approach to UCB1, the intuition is clear: the idea is to be optimistic about
both the rewards and the amounts of resource consumption. We argue that this approach cannot yield log-
arithmic regret bounds. First, because γ has to be of order 1/√T . Second, because, even if we were given
an optimal solution to (3), (ξx∗
k )k=1,··· ,K, before starting the game, consistently choosing which arm to pull
at random according to this distribution at every round would incur regret Ω(√T ), as we next show.
LEMMA 3. For all the cases treated in this paper, pulling arm k with probability ξx∗
yields a regret of order Ω(√T ) unless pulling any arm in {k ∈ {1,··· , K} ξx∗
deterministic amount of resource consumption for all resources in Cx∗ and for all rounds t ∈ N.
Proof. We use the shorthand notation K∗ = {k ∈ {1,··· , K} ξx∗
i ∈ {1,··· , C − 1}, we have:
Xt=1
T · objx∗ − E[
k at any round t
k > 0} incurs the same
k > 0}. Observe that, for any resource
τ ∗
≥ E[(
Xt=τ ∗
Xt=1
{cat,t(i)− b(i)})+]· objx∗ + O(1),
where the inequality is derived using cat,t(i) ≤ 1 for all rounds t and Pτ ∗−1
i ∈ Cx∗ , (cat,t(i))t∈N is an i.i.d. bounded stochastic process with mean b(i), we have:
= E[(
T
cat,t(i)− B(i))+]· objx∗ + O(1)
t=1 cat,t(i) ≤ B(i). Since, for
rat,t] = (T − E[τ ∗])· objx∗
Xt=1
cat,t(i) +
τ ∗−1
T
E[(
T
{cat,t(i)− b(i)})+] = Ω(√T ),
Xt=1
(6)
provided that cat,t(i) has positive variance, which is true if there exists at least one arm k ∈ K∗ such
that ck,t(i) has positive variance or if there exist two arms k, l ∈ K∗ such that ck,t(i) and cl,t(i) are not
almost surely equal to the same deterministic value. Strictly speaking, this is not enough to conclude that
RB(1),··· ,B(C−1),T = Ω(√T ) as T · objx∗ is only an upper bound on the maximum total expected payoff.
10
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
However, in Sections 4, 5, 6, and 7, we show that there exists an algorithm that satisfies T · objx∗ −
E[Pτ ∗
t=1 rat,t] = O(ln(T )) for all the cases considered in this paper. This result, together with (6) and
k at any round is Ω(√T ).
Lemma 2, implies that the regret incurred when pulling arm k with probability ξx∗
The fundamental shortcoming of this approach is that it systematically leads us to plan to consume the same
average amount of resource i per round b(i), for any resource i = 1,··· , C − 1, irrespective of whether we
have significantly over- or under-consumed in the past. Based on this observation, a natural idea is to solve
the linear program:
K
sup
(ξk)k=1,··· ,K
subject to
(¯ck,t(i)− ǫk,t)· ξk ≤ (1− γ)· bt(i),
i = 1,··· , C − 1
(7)
(¯rk,t + ǫk,t)· ξk
K
Xk=1
Xk=1
Xk=1
ξk ≥ 0,
K
ξk ≤ 1
k = 1,··· , K,
instead of (5), where bt(i) denotes the ratio of the remaining amount of resource i at time t to the remaining
time horizon, i.e. T − t + 1. Bounding the regret incurred by this adaptive algorithm is, however, difficult
from a theoretical standpoint. To address this issue, we propose the following family of algorithms, whose
behaviors are similar to the adaptive algorithm but lend themselves to an easier analysis.
Algorithm: UCB-Simplex
Take λ ≥ 1 and (ηi)i=1,··· ,C ≥ 0 (these quantities will need to be carefully chosen). The algorithm is
preceded by an initialization phase which consists in pulling each arm a given number of times, to be
specified. For each subsequent time period t, proceed as follows.
Step-Simplex: Find an optimal basis xt to the linear program:
K
(¯rk,t + λ· ǫk,t)· ξk
sup
(ξk )k=1,··· ,K
subject to
K
Xk=1
Xk=1
ξk ≥ 0,
k = 1,··· , K
(¯ck,t(i)− ηi · ǫk,t)· ξk ≤ b(i),
i = 1,··· , C
(8)
.
Kx,t = bKxt
, and the corresponding basic feasible
k,t)k=1,··· ,K determined by ξxt
k,t = 0 for k 6∈ Kxt and ¯Axt,tξxt
Adapting the notations, xt is described by two subsets Kxt ⊂ {1,··· , K} and Cxt ⊂ {1,··· , C} such that
Kxt = Cxt, the matrix ¯Axt,t = (¯ck,t(i)− ηi · ǫk,t)(i,k)∈Cxt×Kxt
solution (ξxt
Step-Load-Balance: Identify the arms involved in the optimal basis, i.e. Kxt . There are at most min(K, C)
such arms. Use a load balancing algorithm Axt, to be specified, to determine which of these arms to pull.
For all the cases considered in this paper, (8) is always bounded and Step-Simplex is well defined. The
Simplex algorithm is an obvious choice to carry out Step-Simplex, especially when ηi = 0 for any resource
i ∈ {1,··· , C}, because, in this case, we only have to update one column of the constraint matrix per
round which makes warm-starting properties attractive. However, note that this can also be done in time
polynomial in K and C, see Grotschel et al. [16]. If we compare (8) with (5), the idea remains to be overly
optimistic but, as we will see, more about the rewards than the amounts of resource consumption through the
exploration factor λ which will typically be larger than ηi, thus transferring most of the burden of exploration
from the constraints to the objective function. The details of Step-Load-Balance are purposefully left out
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
11
and will be specified for each of the cases treated in this paper. When there is a time horizon T , the general
idea is to determine, at any time period t and for each resource i = 1,··· , C, whether we have over- or
under-consumed in the past and to perturb the probability distribution (ξxt
k,t)k=1,··· ,K accordingly to get back
on track.
k=1 ξx
k=1 ξx
k,t)k=1,··· ,K and objx,t are plug-in estimates of (ξx
k,t · ¯rk,t, and the exploration term, Ex,t = λ ·PK
The algorithm we propose is intrinsically tied to the existence of basic feasible optimal solutions to (3)
and (8). We denote by B (resp. Bt) the subset of bases of (3) (resp. (8)) that are feasible for (3) (resp.
(8)). Step-Simplex can be interpreted as an extension of the index-based decision rule of UCB1. Indeed,
Step-Simplex assigns an index Ix,t to each basis x ∈ Bt and outputs xt ∈ arg maxx∈Bt Ix,t, where Ix,t =
objx,t + Ex,t with a clear separation (at least when ηi = 0 for any resource i) between the exploitation term,
objx,t =PK
k,t · ǫk,t. Observe that, for x ∈ Bt that
is also feasible for (3), (ξx
k )k=1,··· ,K and objx when ηi = 0
for any resource i. Also note that when λ = 1 and ηi = 0 for any resource i and when time is the only
limited resource, UCB-Simplex is identical to UCB1 as Step-Load-Balance is unambiguous in this special
case, each basis involving a single arm. For any x ∈ B, we define ∆x = objx∗ − objx ≥ 0 as the optimality
gap. A feasible basis x is said to be suboptimal if ∆x > 0. At any time t, nx,t denotes the number of times
basis x has been selected at Step-Simplex up to time t while nx
k,t denotes the number of times arm k has
been pulled up to time t when selecting x at Step-Simplex. For all the cases treated in this paper, we will
show that, under a nondegeneracy assumption, Step-Simplex guarantees that a suboptimal basis cannot be
selected more than O(ln(B)) times on average, a result reminiscent of the regret analysis of UCB1 carried
out in Auer et al. [4]. However, in stark contrast with the situation of a single limited resource, this is merely
a prerequisite to establish a O(ln(B)) bound on regret. Indeed, a low regret algorithm must also balance
the load between the arms as closely as possible to optimality. Hence, the choice of the load balancing
algorithms Ax is crucial to obtain logarithmic regret bounds.
4. A single limited resource.
In this section, we tackle the case of a single resource whose consump-
tion is limited by a global budget B, i.e. C = 1 and b(1) = 1. To simplify the notations, we omit the indices
k, ck,t, ¯ck,t, and η as opposed to µc
identifying the resources as there is only one, i.e. we write µc
k(1), ck,t(1),
¯ck,t(1), and η1. We also use the shorthand ǫ = mink=1,··· ,K µc
k. Recall that, under Assumption 2, ǫ is positive
and a priori unknown to the decision maker. In order to derive logarithmic bounds, we will also need to
assume that the decision maker knows an upper bound on the optimal value of (3).
ASSUMPTION 3. The decision maker knows κ ≥ max
k=1,··· ,K
µr
k
µc
k
ahead of round 1.
Assumption 3 is natural in repeated second-price auctions, as detailed in the last paragraph of this section.
Moreover, note that if ǫ happens to be known ahead of round 1 we can take κ = 1/ǫ.
Specification of the algorithm. We implement UCB-Simplex with λ = 1 + κ and η = 0. The initializa-
tion step consists in pulling each arm until the amount of resource consumed as a result of pulling that arm
is non-zero. The purpose of this step is to have ¯ck,t > 0 for all periods to come and for all arms. Step-Load-
Balance is unambiguous here as basic feasible solutions involve a single arm. Hence, we identify a basis x
such that Kx = {k} and Cx = {1} with the corresponding arm and write x = k to simplify the notations. In
particular, k∗ ∈ {1,··· , K} identifies an optimal arm in the sense defined in Section 3. For any arm k, the
exploration and exploitation terms defined in Section 3 specialize to:
objk,t =
¯rk,t
¯ck,t
and Ek,t = (1 + κ)·
ǫk,t
¯ck,t
,
while objk = µr
k/µc
k, so that:
k∗ ∈ arg max
k=1,··· ,K
µr
k
µc
k
, at ∈ arg max
k=1,··· ,K
¯rk,t + (1 + κ)· ǫk,t
¯ck,t
, and ∆k =
µr
k∗
k∗ −
µc
µr
k
µc
k
.
12
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
We point out that, for the particular setting considered in this section, UCB-Simplex is almost identical to
the fractional KUBE algorithm proposed in Tran-Thanh et al. [24] to tackle the case of a single resource
whose consumption is deterministic. It only differs by the presence of the scaling factor 1 + κ to favor
exploration over exploitation, which becomes unnecessary when the amounts of resource consumed are
deterministic, see Section A of the Appendix.
Regret analysis. We omit the initialization step in the theoretical analysis because the amount of
resource consumed is O(1) and the reward obtained is non-negative and not taken into account in the deci-
sion maker's total payoff. Moreover, the initialization step ends in finite time almost surely as a result of
Assumption 2. First observe that (4) specializes to:
RB ≤ B ·
τ ∗
µr
k∗
Xt=1
k∗ − E[
µc
rat,t] + O(1).
(9)
To bound the right-hand side, we start by estimating the expected time horizon.
t=1 cat,t ≤ B. Taking expectations on both sides yields
Sketch of proof.
LEMMA 4. For any non-anticipating algorithm, we have: E[τ ∗] ≤ B+1
B ≥ E[Pτ ∗
By definition of τ ∗, we havePτ ∗−1
t=1 µc
ǫ
.
at ]− 1 ≥ E[τ ∗]· ǫ− 1 by Assumption 2. Rearranging this last inequality yields the claim.
The next result is crucial. Used in combination with Lemma 4, it shows that any suboptimal arm is pulled
at most O(ln(B)) times in expectations, a well-known result for UCB1, see Auer et al. [4]. The proof
is along the same lines as for UCB1, namely we assume that arm k has already been pulled more than
Θ(ln(τ ∗)/(∆k)2) times and conclude that arm k cannot be pulled more than a few more times, with the
additional difficulty of having to deal with the random stopping time and the fact that the amount of resource
consumed at each step is stochastic.
LEMMA 5. For any suboptimal arm k, we have:
E[ln(τ ∗)]
(∆k)2 +
Sketch of proof. We use the shorthand notation βk = 25(λ/µc
E[nk,τ ∗] ≤ 26(
λ
µc
k
)2 ·
4π2
3ǫ2 .
k)2·(1/∆k)2. First observe that if we want
to bound E[nk,τ ∗], we may assume, without loss of generality, that arm k has been pulled at least βk · ln(t)
times at any time t up to an additive term of 2βk· E[ln(τ ∗)] in the final inequality. We then just have to bound
by a constant the probability that k is selected at any time t given that nk,t ≥ βk · ln(t). If k is selected at
time t, it must be that k is optimal for (8), which, in particular, implies that objk,t + Ek,t ≥ objk∗,t + Ek∗,t.
This can only happen if either: (i) objk,t ≥ objk + Ek,t, i.e. the objective value of k is overly optimistic, (ii)
objk∗,t ≤ objk∗ − Ek∗,t, i.e. the objective value of k∗ is overly pessimistic, or (iii) objk∗ < objk + 2Ek,t, i.e.
the optimality gap of arm k is small compared to its exploration factor. The probability of events (i) and (ii)
can be bounded by ∼ 1/t2 in the same fashion, irrespective of how many times these arms have been pulled
in the past. For example for event (i), this is because if ¯rk,t/¯ck,t = objk,t ≥ objk + Ek,t = µr
k + Ek,t, then
k − ǫk,t and both of these events have probability at most ∼ 1/t2 by
either (a) ¯rk,t ≥ µr
Lemma 1. Indeed, if (a) and (b) do not hold, we have:
k + ǫk,t or (b) ¯ck,t ≤ µc
k/µc
¯rk,t
¯ck,t −
µr
k
µc
k
=
<
k)µc
(¯rk,t − µr
ǫk,t
ǫk,t
¯ck,t ·
¯ck,t
+
k + (µc
¯ck,t · µc
µr
k
k ≤ (1 + κ)·
µc
k − ¯ck,t)µr
ǫk,t
¯ck,t
k
k
= Ek,t.
As for event (iii), observe that if objk∗ < objk + 2Ek,t and nk,t ≥ βk · ln(t) then we have ¯ck,t ≤ µc
k/2, which
happens with probability at most ∼ 1/t2 by Lemma 1 given that arm k has already been pulled at least
∼ ln(t)/(µk
c )2 times.
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
13
Building on the last two results, we derive a distribution-dependent regret bound which improves upon the
one derived in Xia et al. [30]: the decision maker is only assumed to know κ, as opposed to a lower bound
on ǫ, ahead of round 1. This is more natural in bidding applications as detailed in the last paragraph of this
section. This bound generalizes the one obtained by Auer et al. [4] when time is the only scarce resource.
THEOREM 1. We have:
RB ≤ 26λ2 · ( Xk∈{1,··· ,K} ∆k>0
1
µc
k · ∆k
)· ln(
B + 1
ǫ
) + O(1).
Sketch of proof. We build upon (9):
RB ≤ B ·
τ ∗
K
rat,t] + O(1)
µr
k∗
Xt=1
k∗ − E[
µc
µr
k∗
Xk=1
µr
k∗ −
k · E[nk,τ ∗] + O(1)
µc
k∗ · (B − Xk ∆k=0
µc
k · E[nk,τ ∗])− Xk ∆k>0
= B ·
µr
k∗
=
µc
µr
k · E[nk,τ ∗ ] + O(1).
t=1 cat,t. Taking expectations on both
By definition of τ ∗, the resource is exhausted at time τ ∗, i.e. B ≤Pτ ∗
sides yields B ≤PK
µc
k · ∆k · E[nk,τ ∗ ] + O(1).
k=1 µc
k · E[nk,τ ∗ ]. Plugging this last inequality back into the regret bound, we get:
RB ≤ Xk ∆k>0
Using the upper bound of Lemma 4, the concavity of the logarithmic function, and Lemma 5, we derive:
B + 1
4π2
1
µc
k · ∆k
ǫ
) +
k∗ /µc
)· ln(
k∗ ≤ κ and µc
µc
k · ∆k) + O(1)
3ǫ2 · ( Xk ∆k>0
k ≤ 1 for any arm k.
RB ≤ 26λ2 · ( Xk ∆k>0
which yields the claim since ∆k ≤ µr
Observe that the set of optimal arms, namely arg maxk=1,··· ,K µr
k, does not depend on B and that ∆k
is a constant independent of B for any suboptimal arm. We conclude that RB = O(K · ln(B)/∆) with
∆ = mink∈{1,··· ,K} ∆k>0 ∆k. Interestingly, the algorithm we propose does not rely on B to achieve this
regret bound, much like what happens for UCB1 with the time horizon, see Auer et al. [4]. This result is
optimal up to constant factors as the standard MAB problem is a special case of the framework considered
in this section, see Lai and Robbins [19] for a proof of a lower bound in this context. It is possible to improve
the constant factors when the consumption of the resource is deterministic as we can take λ = 1 in this
scenario and the resulting regret bound is scale-free, see Section A of the Appendix. Building on Theorem
1, we can also derive a near-optimal distribution-free regret bound in the same fashion as for UCB1.
k/µc
THEOREM 2. We have:
RB ≤ 8λ·rK ·
B + 1
ǫ
B + 1
ǫ
· ln(
) + O(1).
Proof
To get the distribution-free bound, we start from the penultimate inequality derived in the proof
sketch of Theorem 1 and apply Lemma 5 only if ∆k is big enough, noting that:
K
Xk=1
E[nk,τ ∗] = E[τ ∗] ≤ (B + 1)/ǫ.
14
Specifically, we have:
RB ≤
sup
(n1,··· ,nK )≥0
k=1 nk≤ B+1
PK
ǫ
≤
sup
(n1,··· ,nK )≥0
k=1 nk≤ B+1
PK
ǫ
min(µc
min(µc
{ Xk ∆k>0
{ Xk ∆k>0
{ Xk ∆k>0r26λ2 · nk · ln(
B + 1
ǫ
· ln(
) + O(1),
ǫ
sup
(n1,··· ,nK )≥0
k=1 nk≤ B+1
PK
ǫ
B + 1
≤
≤ 8λ·rK ·
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
ln( B+1
ǫ )
µc
k · ∆k
ln( B+1
ǫ )
µc
k · ∆k
+
4π2
3ǫ2 · µc
k · ∆k) } + O(1)
) } + K ·
4π2κ
3ǫ2 + O(1)
k · ∆k · nk, 26λ2 ·
k · ∆k · nk, 26λ2 ·
B + 1
ǫ
) } + O(1)
where the second inequality is obtained with ∆k ≤ µr
by maximizing on (µc
inequality.
k ≤ 1, the third inequality is derived
k· ∆k) ≥ 0 for all arms k, and the last inequality is obtained with the Cauchy−Schwarz
We conclude that RB = O(pK · B · ln(B)), where the hidden constant factors are independent of the
underlying distributions (νk)k=1,··· ,K.
k∗ ≤ κ and µc
k∗/µc
Applications. Assumption 3 is natural for bidding in repeated second-price auctions when the auction-
eer sets a reserve price R (this is common practice in sponsored search auctions). Indeed, then we have:
E[ck,t] = E[mt · 1bk≥mt]
≥ R· E[1bk≥mt]
≥ R· E[vt · 1bk≥mt] = R· E[rk,t],
for any arm k ∈ {1,··· , K} and Assumption 3 is satisfied with κ = 1/R.
5. Arbitrarily many limited resources whose consumptions are deterministic.
In this section, we
study the case of multiple limited resources when the amounts of resources consumed as a result of pulling
an arm are deterministic and globally constrained by prescribed budgets (B(i))i=1,··· ,C , where C is the
number of resources. Note that time need not be a constraint. Because the amounts of resources consumed
are deterministic, we can substitute the notation µc
k(i) with ck(i) for any arm k ∈ {1,··· , K} and any
resource i ∈ {1,··· , C}. We point out that the stopping time need not be deterministic as the decision to
select an arm at any round is based on the past realizations of the rewards. We define ρ ≤ min(C, K) as the
rank of the matrix (ck(i))1≤k≤K,1≤i≤C.
Specification of the algorithm. We implement UCB-Simplex with an initialization step which consists
in pulling each arm ρ times. The motivation behind this step is mainly technical and is simply meant to
have:
nk,t ≥ ρ + Xx∈B k∈Kx
nx
k,t ∀t ∈ N,∀k ∈ {1,··· , K}.
(10)
Compared to Section 4, we choose to take λ = 1 and ηi = 0 for any i ∈ {1,··· , C}. As a result and since
the amounts of resource consumption are deterministic, the exploration (resp. exploitation) terms defined in
Section 3 specialize to objx,t =PK
k · ǫk,t). Compared to the case of a single
resource, we are required to specify the load balancing algorithms involved in Step-Load-Balance of UCB-
Simplex as a feasible basis selected at Step-Simplex may involve several arms. Although Step-Simplex will
also need to be specified in Sections 6 and 7, designing good load balancing algorithms is arguably easier
here as the optimal load balance is known for each basis from the start. Nonetheless, one challenge remains:
k · ¯rk,t (resp. Ex,t =PK
k=1 ξx
k=1 ξx
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
15
we can never identify the (possibly many) optimal bases of (3) with absolute certainty. As a result, any
basis selected at Step-Simplex should be treated as potentially optimal when balancing the load between the
arms involved in this basis, but this inevitably causes some interference issues as an arm may be involved
in several bases, and worst, possibly several optimal bases. Therefore, one point that will appear to be of
particular importance in the analysis is the use of load balancing algorithms that are decoupled from one
another, in the sense that they do not rely on what happened when selecting other bases. More specifically,
we use the following class of load balancing algorithms.
Algorithm: Load balancing algorithm Ax for a feasible basis x ∈ B
If basis x is selected at time t, pull any arm k ∈ Kx such that nx
k,t ≤ nx,t ·
ξx
k
PK
l=1 ξx
l
.
The load balancing algorithms Ax thus defined are decoupled because, for each basis, the number of times
an arm has been pulled when selecting any other basis is not taken into account. The following lemma
shows that Ax is always well defined and guarantees that the ratios (nx
l,t)k,l∈Kx remain close to the
optimal ones (ξx
k,t/nx
k /ξx
l )k,l∈Kx at all times.
LEMMA 6. For any basis x ∈ B, Ax is well defined and moreover, at any time t and for any arm k ∈ Kx,
we have:
nx,t ·
ξx
k
l=1 ξx
l
PK
− ρ ≤ nx
k,t ≤ nx,t ·
+ 1,
ξx
k
l=1 ξx
l
PK
while nx
k,t = 0 for any arm k /∈ Kx.
Proof. We need to show that there always exists an arm k ∈ Kx such that nx
k,t ≤ nx,t · ξx
Suppose there is none, we have:
l=1 ξx
l .
k /PK
nx,t = Xk∈Kx
nx
k,t > Xk∈Kx
nx,t ·
= nx,t,
ξx
k
l=1 ξx
l
PK
a contradiction. Moreover, we have, at any time t and for any arm k ∈ Kx, nx
Indeed, suppose otherwise and define t∗ ≤ t as the last time arm k was pulled, we have:
k,t ≤ nx,t · ξx
k /PK
l=1 ξx
l + 1.
k,t∗ = nx
nx
k,t − 1 > nx,t ·
ξx
k
l=1 ξx
l ≥ nx,t∗ ·
ξx
k
l=1 ξx
l
,
PK
k /PK
PK
k,t since nx,t =Pk∈Kx nx
l − ρ ≤ nx
which shows, by definition, that arm k could not have been pulled at time t∗. We also derive as a byproduct
that, at any time t and for any arm k ∈ Kx, nx,t · ξx
k,t and since
a basis involves at most ρ arms.
l=1 ξx
Observe that the load balancing algorithms Ax run in time O(1) but may require a memory storage capacity
exponential in C and polynomial in K, although always bounded by O(B) (because we do not need to keep
track of nx
k,t if x has never been selected). In practice, only a few bases will be selected at Step-Simplex,
so that a hash table is an appropriate data structure to store the sequences (nx
k,t)k∈Kx . In Section A of the
Appendix, we introduce another class of load balancing algorithms that is both time and memory efficient
while still guaranteeing logarithmic regret bounds (under an additional assumption) but no distribution-free
regret bounds.
Regret Analysis. We use the shorthand notation:
ck(i).
ǫ = min
k=1,··· ,K
i=1,··· ,C
with ck(i)>0
16
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
Note that ǫ < ∞ under Assumption 2. We discard the initialization step in the theoretical study because the
amounts of resources consumed are bounded by a constant and the total reward obtained is non-negative
and not taken into account in the decision maker's total payoff. We again start by estimating the expected
time horizon.
Proof.
LEMMA 7. For any non-anticipating algorithm, we have: E[τ ∗] ≤ PC
t=1 cat,t(i) ≤ B(i) for any resource i ∈ {1,··· , C}. Summing
up these inequalities and using Assumption 2 and the fact that (ck,t(i))t=1,··· ,T are deterministic, we get
By definition of τ ∗, we havePτ ∗−1
i=1 B(i). Taking expectations on both sides and using Assumption 1 yields the result.
i=1 b(i)·B
+ 1.
ǫ
(τ ∗ − 1)· ǫ ≤PC
We follow by bounding the number of times any suboptimal basis can be selected at Step-Simplex in the
same spirit as in Section 4.
LEMMA 8. For any suboptimal basis x ∈ B, we have:
E[nx,τ ∗] ≤ 16ρ· (
ξx
k )2 ·
E[ln(τ ∗)]
(∆x)2 + ρ·
π2
3
.
K
Xk=1
Sketch of proof. We use the shorthand notation βx = 8ρ· (PK
k /∆x)2. The proof is along the same
lines as for Lemma 5. First note that we may assume, without loss of generality, that x has been selected at
least βx· ln(t) times at any time t up to an additive term of 2βx· E[ln(τ ∗)] in the final inequality. We then just
have to bound by a constant the probability that x is selected at any time t given that nx,t ≥ βx · ln(t). If x is
selected at time t, x is an optimal basis to (8). Since the amounts of resources consumed are deterministic,
x∗ is feasible to (8) at time t, which implies that objx,t + Ex,t ≥ objx∗,t + Ex∗,t. This can only happen if
either: (i) objx,t ≥ objx + Ex,t, (ii) objx∗,t ≤ objx∗ − Ex∗,t, or (iii) objx∗ < objx + 2Ex,t. First note that
(iii) is impossible because, assuming this is the case, we would have:
k=1 ξx
∆x
2
ξx
ξx
nk,t
ρ + nx
k,t
k ·s 2 ln(t)
k ·s 2 ln(t)
k · Xk∈Kxpξx
k ·r 2
< Xk∈Kx
≤ Xk∈Kx
≤sXk∈Kx
≤ √ρ · Xk∈Kx
βx
ξx
ξx
k ·s 2 ln(t)
nx,t
=
∆x
2
,
where we use (10), Lemma 6 for each k ∈ Kx, the Cauchy−Schwarz inequality, and the fact that a basis
involves at most ρ arms. Along the same lines as for Lemma 5, the probability of events (i) and (ii) can be
bounded by ∼ ρ/t2 in the same fashion, irrespective of how many times x and x∗ have been selected in the
past. For example for event (i), this is because if objx,t ≥ objx + Ex,t, then there must exist k ∈ Kx such
k + ǫk,t, but any of these events have individual probability at most ∼ 1/t2 by Lemma 1. Indeed
that ¯rk,t ≥ µr
otherwise, if ¯rk,t < µr
k + ǫk,t for all k ∈ Kx, we would have:
objx,t − objx = Xk∈Kx
< Xk∈Kx
k
(¯rk,t − µr
ǫk,t · ξx
k)· ξx
k = Ex,t,
a contradiction
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
17
Lemma 8 used in combination with Lemma 7 shows that a suboptimal basis is selected at most O(ln(B))
times. To establish the regret bound, what remains to be done is to lower bound the expected total payoff
derived when selecting any of the optimal bases. This is more involved than in the case of a single limited
resource because the load balancing step comes into play at this stage.
THEOREM 3. We have:
RB(1),··· ,B(C) ≤ 16
ρ·PC
i=1 b(i)
ǫ
· ( Xx∈B ∆x>0
1
∆x
)· ln(PC
i=1 b(i)· B
ǫ
+ 1) + O(1).
Sketch of proof.
The proof proceeds along the same lines as for Theorem 1. We build upon (4):
RB(1),··· ,B(C) ≤ B ·
= B ·
K
K
Xk=1
Xk=1
τ ∗
k · ξx∗
Xt=1
µr
k − E[
k · ξx∗
Xk=1
k −Xx∈B
µr
K
rat,t] + O(1)
µr
k · E[nx
k,τ ∗ ] + O(1).
Using the properties of the load balancing algorithm established in Lemma 6, we derive:
K
K
E[nx,τ ∗]
k=1 ξx
K
= (
k · ξx∗
Xk=1
k −Xx∈B
µr
RB(1),··· ,B(C) ≤ B ·
{
PK
k · ξx∗
k )· (B − Xx∈B ∆x=0
Xk=1
µr
Xk=1
− Xx∈B ∆x>0
{(
PK
k · ξx
µr
k )·
K
E[nx,τ ∗]
k=1 ξx
µr
k · ξx
k · (
E[nx,τ ∗]
k=1 ξx
k
Xk=1
PK
k } + O(1).
)
k )} + O(1)
Now observe that, by definition, at least one resource is exhausted at time τ ∗. Hence, there exists i ∈
{1,··· , C} such that:
where we use Lemma 6 again and the fact that any basis x ∈ B satisfies all the constraints of (3). We
conclude that:
Taking expectations on both sides and plugging the result back into the regret bound yields:
ck(i)· ξx
k
k,τ ∗
B(i) ≤Xx∈B Xk∈Kx
ck(i)· nx
nx,τ ∗
≤ O(1) +Xx∈B
PK
k=1 ξx
≤ O(1) + b(i)·Xx∈B
nx,τ ∗
k=1 ξx
k
k · Xk∈Kx
PK
k ≥ B − Xx∈B ∆x>0
nx,τ ∗
k=1 ξx
,
Xx∈B ∆x=0
PK
∆x
k=1 ξx
RB(1),··· ,B(C) ≤ Xx∈B ∆x>0
PK
RB(1),··· ,B(C) ≤ 16ρ· ( Xx∈B ∆x>0PK
ρ· ( Xx∈B ∆x>0
PK
k· ξx∗
i=1 b(i)/ǫ,PK
k ≤PC
k=1 ξx
∆x
∆x
k=1 ξx
k=1 µr
π2
3
+
k
k
)· ln(PC
) + O(1)
+ O(1).
nx,τ ∗
k=1 ξx
k
PK
k · E[nx,τ ∗] + O(1).
i=1 b(i)· B
+ 1)
ǫ
Using Lemma 7, Lemma 8, and the concavity of the logarithmic function, we obtain:
which yields the claim since ∆x ≤PK
k=1 ξx
k ≥ b, andPK
k=1 ξx
k ≤PC
i=1 b(i)/ǫ.
18
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
We point out that, if time is a limited resource with a time horizon T , we can also derive the (possibly better)
regret bound:
RB(1),··· ,B(C) ≤ 16ρ· ( Xx∈B ∆x>0
Since the number of feasible bases to (3) is at most(cid:0)K+ρ
K (cid:1) ≤ 2K ρ, we get the distribution-dependent regret
bound O(ρ · K ρ · ln(B)/∆) where ∆ = minx∈B ∆x>0 ∆x. In Section A of the Appendix, we introduce
an alternative class of load balancing algorithms which yields a better dependence on K and C with a
regret bound of order O(ρ3 · K · ln(B)/∆2) provided that there is a unique optimal basis to (3). Along the
sames lines as in Section 4, the distribution-dependent bound of Theorem 3 almost immediately implies a
distribution-free one.
)· ln(T ) + O(1).
1
∆x
THEOREM 4. We have:
RB(1),··· ,B(C) ≤ 4sρ·B· (PC
i=1 b(i)· B
ǫ
+ 1)· ln(PC
i=1 b(i)· B
ǫ
+ 1) + O(1).
Sketch of proof.
The proof is along the same lines as for the case of a single limited resource, we start
from the penultimate inequality derived in the proof sketch of Theorem 3 and apply Lemma 8 only if ∆x is
big enough, taking into account the fact thatPx∈B
We conclude that RB(1),··· ,B(C) = O(pρ· K ρ · B · ln(B)), where the hidden constant factors are indepen-
dent of the underlying distributions (νk)k=1,··· ,K. If time is a limited resource, we can also derive the (pos-
sibly better) regret bound:
E[nx,τ ∗] ≤ E[τ ∗] ≤PC
i=1 b(i)· B/ǫ + 1.
RB(1),··· ,B(C) ≤ 4pρ·B· T · ln(T ) + O(1).
In any case, we stress that the dependence on K and C is not optimal since Badanidiyuru et al. [8] and
Agrawal and Devanur [1] obtain a O(√K · B) bound on regret, where the O notation hides factors logarith-
mic in B. Observe that the regret bounds derived in this section do not vanish with b. This can be remedied
by strengthening Assumption 2, additionally assuming that ck,t(i) > 0 for any arm k ∈ {1,··· , K} and
resource i ∈ {1,··· , C}. In this situation, we can refine the analysis and substitutePC
i=1 b(i) with b in the
regret bounds of Propositions 3 and 4 which become scale-free.
Applications. BwK problems where the amounts of resources consumed as a result of pulling an arm
are deterministic find applications in network revenue management of perishable goods, shelf optimization
of perishable goods, and wireless sensor networks, as detailed in Section 2.
6. A time horizon and another limited resource.
In this section, we investigate the case of two lim-
ited resources, one of which is assumed to be time, with a time horizon T , while the consumption of the
other is stochastic and constrained by a global budget B. To simplify the notations, we omit the indices
identifying the resources since the second limited resource is time and we write µc
k, ck,t, ¯ck,t, B, and T as
opposed to µc
k(1), ck,t(1), ¯ck,t(1), B(1), and B(2). Moreover, we refer to resource i = 1 as "the" resource.
Observe that, in the particular setting considered in this section, τ ∗ = min(τ (B), T + 1) with τ (B) =
min{t ∈ N Pt
τ =1 caτ ,τ > B}. Note that the budget constraint is not limiting if B ≥ T , in which case the
problem reduces to the standard MAB problem. Hence, without loss of generality under Assumption 1, we
assume that the budget scales linearly with time, i.e. B = b · T for a fixed constant b ∈ (0, 1), and we study
the asymptotic regime T → ∞. Motivated by technical considerations, we make two additional assumptions
for the particular setting considered in this section that are perfectly reasonable in many applications, such
as in repeated second-price auctions, dynamic procurement, and dynamic pricing, as detailed in the last
paragraph of this section.
ASSUMPTION 4. There exists σ > 0 such that µr
k ≤ σ · µc
k for any arm k ∈ {1,··· , K}.
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
19
ASSUMPTION 5. The decision maker knows κ > 0 such that:
µr
k − µr
l ≤ κ·µc
k − µc
l ∀(k, l) ∈ {1,··· , K}2,
ahead of round 1.
Note that σ, as opposed to κ, is not assumed to be known to the decision maker. Assumption 4 is relatively
weak and is always satisfied in practical applications. In particular, note that if µc
k > 0 for all arms k ∈
{1,··· , K}, we can take σ = 1/ mink=1,··· ,K µc
k.
Specification of the algorithm. We implement UCB-Simplex with λ = 1 + 2κ, η1 = 1, η2 = 0, and
an initialization step which consists in pulling each arm once. Because the amount of resource consumed
at each round is a random variable, a feasible basis for (8) may not be feasible for (3) and conversely. In
particular, x∗ may not be feasible for (8), thus effectively preventing it from being selected at Step-Simplex,
and an infeasible basis for (3) may be selected instead. This is in contrast to the situation studied in Section
5 and this motivates the choice η1 > 0 to guarantee that any feasible solution to (3) will be feasible to (8)
with high probability at any round t.
t
= {0, kt} (resp. {kt, K + kt}) and Cx′
Just like in Section 5, we need to specify Step-Load-Balance because a basis selected at Step-Simplex
may involve up to two arms. To simplify the presentation, we introduce a dummy arm k = 0 with reward
0 and resource consumption 0 (pulling this arm corresponds to skipping the round) and K dummy arms
k = K + 1,··· , 2K with reward identical to arm K − k but resource consumption 1 so that any basis
involving a single arm can be mapped to an "artificial" one involving two arms. Note, however, that we do
not introduce new variables ξk in (8) for these arms as they are only used for mathematical convenience
in Step-Load-Balance once a basis has been selected at Step-Simplex. Specifically, if a basis xt involving
a single arm determined by Kxt = {kt} and Cxt = {1} (resp. Cxt = {2}) is selected at Step-Simplex, we
map it to the basis x′t determined by Kx′
= {1, 2}. We then use a
load balancing algorithm specific to this basis, denoted by Axt, to determine which of the two arms in Kx′
t
to pull. Similarly as in Section 5, using load balancing algorithms that are decoupled from one another is
crucial because the decision maker can never identify the optimal bases with absolute certainty. This implies
that each basis should be treated as potentially optimal when balancing the load between the arms, but this
inevitably causes interference issues as an arm may be involved in several bases. Compared to Section 5, we
face an additional challenge when designing the load balancing algorithms: the optimal load balances are
initially unknown to the decision maker. It turns out that we can still approximately achieve the unknown
optimal load balances by enforcing that, at any round t, the total amount of resource consumed remains
close to the pacing target b· t with high probability, as precisely described below.
Algorithm: Load balancing algorithm Ax for any basis x
For any time period t, define bx,t as the total amount of resource consumed when selecting x in the past
t− 1 rounds. Suppose that x is selected at time t. Without loss of generality, write Kx = {k, l} with
¯ck,t − ǫk,t ≥ ¯cl,t − ǫl,t. Pull arm k if bx,t ≤ nx,t · b and pull arm l otherwise.
Observe that a basis x with Kx = {k, l} is feasible for (3) if either µc
we are in the first situation, the exploration and exploitation terms defined in Section 3 specialize to:
k. Assuming
k > b > µc
l > b > µc
l or µc
t
objx,t = ξx
l,t · ¯rl,t + ξx
k,t · ¯rk,t and Ex,t = λ· (ξx
l,t · ǫl,t + ξx
k,t · ǫk,t)
with:
ξx
l,t =
(¯ck,t − ǫk,t)− b
(¯ck,t − ǫk,t)− (¯cl,t − ǫl,t)
and ξx
k,t =
b− (¯cl,t − ǫl,t)
(¯ck,t − ǫk,t)− (¯cl,t − ǫl,t)
,
provided that ¯ck,t − ǫk,t > b > ¯cl,t − ǫl,t. Moreover, their offline counterparts are given by:
objx = ξx
l · µr
l + ξx
k · µr
k, ξx
l =
µc
k − b
µc
k − µc
l
, and ξx
k =
l
b− µc
µc
k − µc
l
.
20
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
Regret Analysis. We start by pointing out that, in degenerate scenarios, using the linear relaxation (3)
as an upper bound on EROPT(B, T ) already dooms us to Ω(√T ) regret bounds. Precisely, if there exists a
unique optimal basis x∗ to (3) that happens to be degenerate, i.e. Kx∗ = {k∗} (pre-mapping) with µc
k∗ = b,
then, in most cases, T · objx∗ ≥ EROPT(B, T ) + Ω(√T ) as shown below.
LEMMA 9.
Sketch of proof.
√T
this is also nothing more than an i.i.d. strategy which implies, along the same lines as in Lemma 3, that
If there exists k∗ ∈ {1,··· , K} such that: (i) the i.i.d. process (ck∗,t)t∈N has positive vari-
k∗ = b, and (iii) (ξk)k=1,··· ,K determined by ξk∗ = 1 and ξk = 0 for k 6= k∗ is the unique optimal
ance, (ii) µc
solution to (3), then there exists a subsequence of ( T·objx∗−EROPT(B,T )
)T∈N that does not converge to 0.
For any time horizon T ∈ N and any arm k ∈ {1,··· , K}, we denote by nopt
k,T the
expected number of times arm k is pulled by the optimal non-anticipating algorithm when the time horizon
is T and the budget is B = b· T . We expect that consistently pulling arm k∗ is near-optimal. Unfortunately,
E[τ ∗] = T − Ω(√T ) so that the total expected payoff is E[τ ∗]· µr
k∗ = T · objx∗ − Ω(√T ). To formalize these
k∗,T = o(√T ) (Case B) and we show
ideas, we study two cases: T − nopt
that EROPT(B, T ) = T · objx∗ − Ω(√T ) in both cases. In Case A, this is because the optimal value of (3)
remains an upper bound on the maximum total expected payoff if we add the constraint ξk∗ ≤ nopt
k∗,T /T to
the linear program (3) by definition of nopt
k∗,T . Since the constraint ξk∗ ≤ 1 is binding for (3), the optimal
value of this new linear program can be shown to be smaller than objx∗ − Ω((T − nopt
k∗,T )/T ) (by strong
duality and strict complementary slackness). In Case B, up to an additive term of order o(√T ) in the final
bound, the optimal non-anticipating algorithm is equivalent to consistently pulling arm k∗, which is an i.i.d.
strategy so the study is very similar to that of Lemma 3.
k∗,T = Ω(√T ) (Case A) and T − nopt
Dealing with these degenerate scenarios thus calls for a completely different approach than the one taken on
in the BwK literature and we choose instead to rule them out in such a way that there can be no degenerate
optimal basis to (3).
ASSUMPTION 6. We have µc
k − b > 0 for any arm k ∈ {1,··· , K}.
We use the shorthand notation ǫ = mink=1,··· ,K µc
k − b. Assumption 6 is equivalent to assuming that any
basis for (3) is non-degenerate. This assumption can be relaxed to some extent at the price of more tech-
nicalities. However, in light of Lemma 9, the minimal assumption is that there is no degenerate optimal
basis to (3). As a final remark, we stress that Assumption 6 is only necessary to carry out the analysis but
Step-Simplex can be implemented in any case as ǫ is not assumed to be known to the decision maker.
We are now ready to establish regret bounds. Without loss of generality, we can assume that any pseudo-
basis for (3) involves two arms, one of which may be a dummy arm introduced in the specification of the
algorithm detailed above. As stressed at the beginning of this section, UCB-Simplex may sometimes select
an infeasible basis or even a pseudo-basis x with det(Ax) = 0 (i.e. such that µc
l assuming Kx = {k, l}).
Interestingly the load balancing algorithm plays a crucial role to guarantee that this does not happen very
often.
k = µc
LEMMA 10. For any basis x /∈ B, we have:
E[nx,T ] ≤
26
ǫ3 · ln(T ) +
10π2
3ǫ2 .
The same inequality holds if x is a pseudo-basis but not a basis for (3).
k, µc
Proof. We use the shorthand notation βx = 25/ǫ3. Without loss of generality, we can assume that
Kx = {k, l} with µc
l > b (the situation is symmetric if the reverse inequality holds). Along the same lines
as in Lemma 5, we only have to bound by a constant the probability that x is selected at any round t given
that x has already been selected at least βx · ln(t) times. If x is selected at round t and nx,t ≥ βx · ln(t),
then bx,t must be larger than nx,t · b by at least a margin of ∼ 1/ǫ2 · ln(t) with high probability given that
l > b. Moreover, at least one arm, say k, has been pulled at least ∼ 1/ǫ3 · ln(t) times and, as a result,
k, µc
µc
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
21
¯ck,τ − ǫk,τ ≥ b with high probability for the last s ∼ 1/ǫ2 · ln(t) rounds τ = τ1,··· , τs where x was selected.
This implies that arm l must have been pulled at least ∼ 1/ǫ2 · ln(t) times already by definition of the load
balancing algorithm but then we have ¯cl,t − ǫl,t ≥ b with high probability and x cannot be feasible for (8)
at time t with high probability.
What remains to be done is to: (i) show that suboptimal bases are selected at most O(ln(T )) times and (ii)
lower bound the expected total payoff derived when selecting any of the optimal bases. The major diffi-
culty lies in the fact that the amounts of resource consumed, the rewards obtained, and the stopping time
are correlated in a non-trivial way through the budget constraint and the decisions made in the past. This
makes it difficult to study the expected total payoff derived when selecting optimal bases independently
from the amounts of resource consumed and the rewards obtained when selecting suboptimal ones. How-
ever, a key point is that, by design, the pulling decision made at Step-Load-Balance is based solely on the
past history associated with the basis selected at Step-Simplex because the load balancing algorithms are
decoupled. For this reason, the analysis proceeds in two steps irrespective of the number of optimal bases.
In a first step, we show that, for any basis x for (3), the amount of resource consumed per round when
selecting x remains close to the pacing target b with high probability. This enables us to show that the ratios
(E[nx
l,T ])k,l∈Kx are close to the optimal ones (ξx
l )k,l∈Kx , as precisely stated below.
k,T ]/E[nx
k /ξx
LEMMA 11. For any basis x ∈ B and time period t, we have:
P[bx,t − nx,t · b ≥ u + (
4
ǫ
)2 · ln(t)] ≤
4
ǫ2 · exp(−ǫ2 · u) +
8
ǫ2 · t2
∀u ≥ 1,
which, in particular, implies that:
E[nx
E[nx
k,T ] ≥ ξx
l,T ] ≥ ξx
k · E[nx,T ]− 13/ǫ5 − 16/ǫ3 · ln(T )
l · E[nx,T ]− 13/ǫ5 − 16/ǫ3 · ln(T ).
(11)
k > b > µc
Sketch of proof. Without loss of generality, we can assume that Kx = {k, l} with µc
l . Observe
that, if the decision maker knew that µc
l ahead of round 1, he would always pull the "correct"
arm in order not to deviate from the pacing target nx,t · b and bx,t − nx,t · b would remain small with high
probability given Assumption 6. However, because this information is not available ahead of round 1, the
decision maker is led to pull the incorrect arm when arm k and l are swapped, in the sense that ¯ck,t − ǫk,t ≤
¯cl,t− ǫl,t. Fortunately, at any time t, there could have been at most 1/ǫ2· ln(t) swaps with probability at least
∼ 1 − 1/t2 given Assumption 6. To derive (11), we use: E[bx,T − nx,T · b] ≤R T
P[bx,T − nx,T · b ≥ u]du
and E[bx,T ] = µk
c · E[nx
c · E[nx
k > b > µc
k,T ] + µl
l,T ].
0
The next step is to show, just like in Section 5, that any suboptimal feasible basis is selected at most
O(ln(T )) times on average. Interestingly, the choice of the load balancing algorithm plays a minor role in
the proof. Any load balancing algorithm that pulls each arm involved in a basis at least a constant fraction
of the time this basis is selected does enforce this property.
LEMMA 12. For any suboptimal basis x ∈ B, we have:
ln(T )
(∆x)2 +
E[nx,T ] ≤ 29 λ2
ǫ3 ·
10π2
ǫ2
.
Sketch of proof. We use the shorthand notation βx = 28/ǫ3 · (λ/∆x)2. Without loss of generality, we
can assume that Kx∗ = {k∗, l∗} with µc
l . Along the same lines
as in Lemma 5, we only have to bound by a constant the probability that x is selected at any round t given
that x has already been selected at least βx·ln(t) times. If x is selected at time t, x is optimal for (8). Observe
that (ξx∗
l∗ , which happens
with probability at least ∼ 1− 1/t2. As a result, objx,t + Ex,t ≥ objx∗ when additionally ¯rk∗,t + ǫk∗,t ≥ µr
k∗
k )k=1,··· ,K is a feasible solution to (8) when ¯ck∗,t − ǫk∗,t ≤ µc
l∗ and Kx = {k, l} with µc
k∗ and ¯cl∗,t − ǫl∗,t ≤ µc
k∗ > b > µc
k > b > µc
22
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
and ¯rl∗,t + ǫl∗,t ≥ µr
l∗ , which also happens with probability at least ∼ 1− 1/t2. If objx,t + Ex,t ≥ objx∗ then
we have either (i) objx,t ≥ objx + Ex,t or (ii) objx∗ < objx + 2Ex,t. Observe that (ii) can only happen with
probability at most ∼ 1/t2 given that nx,t ≥ βx· ln(t) because (ii) implies that either nl,t ≤ 8(λ/∆x)2· ln(t)
or nk,t ≤ 8(λ/∆x)2 · ln(t) but the load balancing algorithm guarantees that each arm is pulled a fraction
of the time x is selected (using Lemma 11). As for (i), if objx,t ≥ objx + Ex,t, then, using Assumption 4,
either ¯rk,t ≥ µr
l + ǫl,t] but all of these
events have individual probability at most ∼ 1/t2 by Lemma 1.
In a last step, we show, using Lemma 11, that, at the cost of an additive logarithmic term in the regret bound,
we may assume that the game lasts exactly T rounds. This enables us to combine Lemmas 10, 11, and 12
to establish a distribution-dependent regret bound.
l + ǫl,t, or ¯cl,t /∈ [µc
k + ǫk,t, ¯ck,t /∈ [µc
k + ǫk,t], ¯rl,t ≥ µr
k − ǫk,t, µc
l − ǫl,t, µc
THEOREM 5. We have:
RB,T ≤ 29 λ2
ǫ3 · ( Xx∈B ∆x>0
1
∆x
)· ln(T ) + O(
K 2 · σ
ǫ3
· ln(T )),
where the O notation hides universal constant factors.
Sketch of proof. We build upon (4):
RB,T ≤ T ·
≤ T ·
K
K
Xk=1
Xk=1
τ ∗
k · ξx∗
µr
k · ξx∗
µr
Xt=1
k − E[
Xt=1
k − E[
T
rat,t] + O(1)
rat,t] + σ · E[(
T
Xt=1
cat,t − B)+] + O(1),
where we use Assumption 4 for the second inequality. Moreover:
E[(
T
Xt=1
cat,t − B)+] ≤Xx∈B
E[bx,T − nx,T · b] +Xx /∈B
+ Xx pseudo-basis for (3)
with det(Ax)=0
E[nx,T ] = O(
E[nx,T ]
K 2
ǫ3 ln(T )),
using Lemmas 10 and 11. Plugging this last inequality back into the regret bound yields:
RB,T ≤ T ·
T
(
K
K
K
K
K
K
≤ T ·
k · ξx∗
Xk=1
Xt=1
µr
k − E[
k · ξx∗
Xk=1
Xk=1
k −Xx∈B
µr
k · ξx∗
Xk=1
k −Xx∈B
Xk=1
µr
≤ T ·
k · ξx∗
k · (T − Xx∈B ∆x=0
Xk=1
µr
k · ξx∗
Xk=1
k · ( Xx∈B ∆x>0
µr
Xk=1
− Xx∈B ∆x>0
µr
k · ξx
=
=
K
K
(
k )· E[nx,T ] + O(
K 2 · σ
ǫ3
ln(T ))
E[nx,T ])
rat,t] + O(
K 2 · σ
ǫ3
ln(T ))
µr
k · E[nx
k,T ] + O(
K 2 · σ
ǫ3
ln(T ))
k · ξx
µr
k )· E[nx,T ] + O(
E[nx,T ])− Xx∈B ∆x>0
ln(T ))
K 2 · σ
ǫ3
K
(
Xk=1
k · ξx
µr
E[nx,T ] + Xx pseudo-basis for (3)
K 2 · σ
ǫ3
with det(Ax)=0
ln(T ))
E[nx,T ] +Xx /∈B
k )· E[nx,T ] + O(
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
23
≤ Xx∈B ∆x>0
≤ 29 λ2
ǫ3 · ( Xx∈B ∆x>0
∆x · E[nx,T ] + O(
K 2 · σ
ǫ3
ln(T ))
1
∆x
)· ln(T ) + O(
K 2 · σ
ǫ3
ln(T )),
the fourth inequality, and Lemma 12 for the last inequality.
where we use Lemma 11 for the third inequality, Lemma 10 along withPK
Since there are at most 2K 2 feasible bases, we get the regret bound O(K 2 · (1/∆ + σ/ǫ3) · ln(T )), where
∆ = minx∈B ∆x>0 ∆x. Along the sames lines as in Sections 4 and 5, pushing the analysis further almost
immediately yields a distribution-free regret bound.
k · ξx∗
k ≤PK
k ≤ 1 for
k=1 ξx∗
k=1 µr
THEOREM 6. We have:
RB,T ≤ 25 λ
ǫ3/2 ·pB· T · ln(T ) + O(
where the O notation hides universal constant factors.
K 2 · σ
ǫ3
ln(T )),
Sketch of proof.
The proof is along the same lines as for Theorems 2 and 4, we start from the penul-
timate inequality derived in the proof sketch of Theorem 5 and apply Lemma 12 only if ∆x is big enough,
E[nx,T ] ≤ T .
taking into account the fact thatPx∈B
We conclude that RB,T = O(pK 2 · T · ln(T )), where the hidden factors are independent of the underlying
distributions (νk)k=1,··· ,K. Just like in Section 5, we stress that the dependence on K is not optimal since
Badanidiyuru et al. [8] and Agrawal and Devanur [1] obtain a O(√K · T ) bound on regret, where the O
notation hides factors logarithmic in T . Observe that the regret bounds derived in Theorems 5 and 6 do
not vanish with b, which is not the expected behavior. This is a shortcoming of the analysis that can easily
be remedied when mink=1,··· ,K µc
k > 0 provided that instead of pulling the dummy arm 0 we always pull
the other arm involved in the basis (i.e. we never skip rounds). Note that not skipping rounds can only
improve the regret bounds derived in Theorems 5 and 6: arm 0 was introduced only in order to harmonize
the notations for mathematical convenience.
THEOREM 7. Relax Assumption 6 and redefine ǫ = mink=1,··· ,K µc
k. Suppose that b ≤ ǫ/2 and that we
never skip rounds, then we have:
RB,T ≤ 212 λ2
1
∆x
)· ln(
B + 1
ǫ
) + O(
K 2 · κ
ǫ3
· ln(
B + 1
ǫ
))
ǫ3 · ( Xx∈B ∆x>0
ǫ3/2 ·rK ·
and
· ln(
where the O notations hide universal constant factors.
ǫ
RB,T ≤ 26 λ
B + 1
B + 1
ǫ
) + O(
K 2 · κ
ǫ3
· ln(
B + 1
ǫ
)),
Applications. Similarly, as in the case of a single resource, Assumptions 4 and 5 are natural when
bidding in repeated second-price auctions if the auctioneer sets a reserve price R (which is common practice
in sponsored search auctions). Indeed, we have:
E[ck,t]− E[cl,t] = E[mt · 1bk≥mt>bl]
≥ R · E[1bk≥mt>bl]
≥ R · E[vt · 1bk≥mt>bl] = R·E[rk,t]− E[rl,t],
24
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
for any pair of arms (k, l) ∈ {1,··· , K} with bk ≥ bl. Hence, Assumption 4 (resp. 5) is satisfied with
σ = 1/R (resp. κ = 1/R).
In dynamic procurement, Assumptions 4 and 5 are satisfied provided that the agents are not willing to
sell their goods for less than a known price P . Indeed, in this case, pulling any arm k associated with a price
pk ≤ P is always suboptimal and we have:
E[ck,t]− E[cl,t] = pk · P[pk ≥ vt]− pl · P[pl ≥ vt]
≥ pk · P[pk ≥ vt > pl]
≥ P · P[pk ≥ vt > pl] = P ·E[rk,t]− E[rl,t],
for any pair of arms (k, l) ∈ {1,··· , K} with pk ≥ pl ≥ P . Hence, Assumption 4 (resp. 5) is satisfied with
σ = 1/P (resp. κ = 1/P ).
In dynamic pricing, Assumptions 4 and 5 are satisfied if the distribution of valuations has a positive
probability density function f (·). Indeed, in this case, we have:
E[rk,t]− E[rl,t] = pl · P[pl ≤ vt]− pk · P[pk ≤ vt]
= pl · P[pl ≤ vt < pk] + (pl − pk)· P[pk ≤ vt]
≤ max
r=1,··· ,K
≤ ( max
r=1,··· ,K
= ( max
pr · P[pl ≤ vt < pk] +pk − pl
)· P[pl ≤ vt < pk]
pr +
)·E[rk,t]− E[rl,t],
pr +
r=1,··· ,K
1
1
inf f (·)
inf f (·)
for any pair of arms (k, l) ∈ {1,··· , K} with k ≥ l. Hence, Assumption 4 (resp. 5) is satisfied with σ =
maxk=1,··· ,K pk + 1/ inf f (·) (resp. κ = maxk=1,··· ,K pk + 1/ inf f (·)).
7. Arbitrarily many limited resources.
In this section, we tackle the general case of arbitrarily many
limited resources. Additionally, we assume that one of them is time, with index i = C, but this assumption
is almost without loss of generality, as detailed at the end of this section. To simplify the presentation, we
consider the regime K ≥ C, which is the most common in applications. This implies that Kx = Cx ≤ C
for any pseudo-basis x. We also use the shorthand notation ¯At = (¯ck,t(i))(i,k)∈{1,··· ,C}×{1,··· ,K} at any round
t. For similar reasons as in Section 6, we are led to make two additional assumptions which are discussed
in the last paragraph of this section.
ASSUMPTION 7. There exists σ > 0 such that rk,t ≤ σ · min
i=1,··· ,C
ck,t(i) for any arm k ∈ {1,··· , K} and
for any round t ∈ N.
Note that Assumption 7 is stronger than Assumption 4 given that the amounts of resources consumed at
each round have to dominate the rewards almost surely, as opposed to on average. Assumption 7 is not
necessarily satisfied in all applications but it simplifies the analysis and can be relaxed at the price of an
additive term of order O(ln2(T )) in the final regret bounds, see the last paragraph of this section.
ASSUMPTION 8. There exists ǫ > 0, known to the decision maker ahead of round 1, such that every
basis x for (3) is ǫ-non-degenerate for (3) and satisfy det(Ax) ≥ ǫ.
Without loss of generality, we assume that ǫ ≤ 1. Observe that Assumption 8 generalizes Assumption 6 but
is more restrictive because ǫ is assumed to be known to the decision maker initially. Just like in Section
6, this assumption can be relaxed to a large extent. For instance, if ǫ is initially unknown, taking ǫ as a
vanishing function of T yields the same asymptotic regret bounds. However, note that Lemma 9 carries over
to this more general setting and, as a result, the minimal assumption we need to get logarithmic rates is that
any optimal basis for (3) is non-degenerate.
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
25
Specification of the algorithm. We implement UCB-Simplex with λ = 1 + 2(C + 1)!2/ǫ, ηi = 0 for
any i ∈ {1,··· , C}, and an initialization step which consists in pulling each arm 28(C + 2)!4/ǫ6 · ln(T )
times in order to get i.i.d. samples. Hence, Step-Simplex is run for the first time after round tini = K ·
28(C + 2)!4/ǫ6 · ln(T ). Compared to Section 6, the initialization step plays as a substitute for the choice ηi >
0 which was meant to incentivize exploration. This significantly simplifies the analysis but the downside is
that ǫ has to be known initially. Similarly, as in Section 6, we introduce a dummy arm which corresponds to
skipping the round (i.e. pulling this arm yields a reward 0 and does not consume any resource) so that any
basis can be mapped to one for which the time constraint is always binding, i.e. without loss of generality
we assume that C ∈ Cx for any pseudo-basis x. Following the ideas developed in Section 6, we design
load balancing algorithms for any basis x that pull arms in order to guarantee that, at any round t, the total
amount of resource i consumed remains close to the target t · b(i) with high probability for any resource
i ∈ Cx. This is more involved that in Section 6 since we need to enforce this property for multiple resources
but, as we show in this section, this can be done by perturbing the probability distribution solution to (8)
taking into account whether we have over- or under-consumed in the past for each binding resource i ∈ Cxt.
Algorithm: Load balancing algorithm Ax for any basis x
For any time period t > tini and i ∈ Cx −{C}, define bx,t(i) as the total amount of resource i consumed
when selecting basis x in the past t− 1 rounds. Suppose that basis x is selected at time t and define the
vector ex
i,t = 1) if bx,t(i) ≥ nx,t · b(i) (resp. bx,t(i) < nx,t · b(i)) for any
i ∈ Cx −{C}. Since x is selected at round t, ¯Ax,t is invertible and we can define, for any δ ≥ 0,
k,t(δ) = ( ¯A−1
px
k,t(δ) = 0 otherwise, which together define the probability
distribution px
x,t(bCx + δ · ex
t (δ) = (px
t by ex
C,t = 0 and ex
i,t = −1 (resp. ex
t ))k for k ∈ Kx and px
k,t(δ))k∈{1,··· ,K}. Define
δ∗x,t =
δ
( ¯Atpx
t (δ))i≤b(i),i /∈Cx
max
δ≥0
px
t (δ)≥0
t = px
and px
Step-Simplex. Pull an arm at random according to the distribution px
t .
t (δ∗x,t). Note that δ∗x,t is well defined as x must be feasible for (8) if it is selected at
Observe that the load balancing algorithms generalize the ones designed in Section 6 (up to the change
ηi = 0). Indeed, when there is a single limited resource other than time, the probability distribution px
t is
a Dirac supported at the arm with smallest (resp. largest) empirical cost when bx,t ≥ nx,t · b (resp. bx,t <
nx,t· b). Similarly, as in Section 5, the load balancing algorithms Ax may require a memory storage capacity
exponential in C and polynomial in K, but, in practice, we expect that only a few bases will be selected at
Step-Simplex, so that a hash table is an appropriate data structure to store the sequences (bx,t(i))i∈Cx. Note,
however, that the load balancing algorithms are computationally efficient because px
t can be computed in
O(C 2) running time if ¯A−1
x,t is available once we have computed an optimal basic feasible solution to (8),
which is the case if we use the revised simplex algorithm.
Regret analysis. The regret analysis follows the same recipe as in Section 6 but the proofs are more
technical and are thus deferred to Section F of the Appendix. First, we show that the initialization step
guarantees that infeasible bases or pseudo-bases x with det(Ax) = 0 cannot be selected more than O(ln(T ))
times on average at Step-Simplex.
LEMMA 13. For any basis x /∈ B, we have:
E[nx,T ] ≤ 29 (C + 3)!4
ǫ6
.
The same inequality holds if x is a pseudo-basis but not a basis for (3).
26
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
The next step is to show that the load balancing algorithms guarantee that, for any basis x, the amount
of resource i ∈ Cx (resp. i /∈ Cx) consumed per round when selecting x remains close to (resp. below) the
pacing target b(i) with high probability. This enables us to show that the ratios (E[nx
l,T ])k,l∈Kx are
close to the optimal ones (ξx
k,T ]/E[nx
k /ξx
l )k,l∈Kx .
LEMMA 14. For any feasible basis x and time period t, we have:
P[bx,t(i)− nx,t · b(i) ≥ u] ≤ 25 (C + 1)!2
for any resource i ∈ Cx while
ǫ4
· exp(−u· (
ǫ2
4· (C + 1)!
)2) + 29 (C + 3)!4
ǫ6
1
T
·
∀u ≥ 1,
(12)
P[bx,t(i)− nx,t · b(i) ≥ 28 (C + 3)!3
ǫ6
for any resource i /∈ Cx. In particular, this implies that:
· ln(T )] ≤ 210 (C + 4)!4
ǫ6 · T
,
E[nx
k,T ] ≥ ξx
k · E[nx,T ]− 210 (C + 3)!4
ǫ9
,
(13)
(14)
for any arm k ∈ Kx.
Next, we show that a suboptimal basis cannot be selected more than O(ln(T )) times on average at Step-
Simplex. Just like in Section 6, the exact definition of the load balancing algorithms has little impact on the
result: we only need to know that, for any feasible basis x, each arm k ∈ Kx is pulled at least a fraction of
the time x is selected with high probability.
LEMMA 15. For any suboptimal basis x ∈ B, we have:
E[nx,T ] ≤ 210 (C + 3)!3 · λ2
ǫ6
·
ln(T )
(∆x)2 + 211 (C + 4)!4
ǫ6
.
We are now ready to derive both distribution-dependent and distribution-independent regret bounds.
THEOREM 8. We have:
RB(1),··· ,B(C−1),T ≤ 210 (C + 3)!3 · λ2
ǫ6
· ( Xx∈B ∆x>0
1
∆x
where the O notation hides universal constant factors.
)· ln(T ) + O(
σ ·B· (C + 3)!4
ǫ6
· ln(T )),
THEOREM 9. We have:
RB(1),··· ,B(C−1),T ≤ 25 (C + 3)!2 · λ
ǫ3
where the O notation hides universal constant factors.
·pB· T · ln(T ) + O(
σ ·B· (C + 3)!4
ǫ6
· ln(T )),
Since the number of feasible bases is at most 2K C, we get the distribution-dependent regret bound
RB(1),··· ,B(C−1),T = O(K C · (C + 3)!4/ǫ6 · (λ2/∆ + σ) · ln(T )) where ∆ = minx∈B ∆x>0 ∆x and the
distribution-independent bound RB(1),··· ,B(C−1),T = O((C + 3)!2 · λ/ǫ3 ·pK C · T · ln(T )). We stress that
the dependence on K and C is not optimal since Agrawal and Devanur [1] obtain a O(√K · T ) distribution-
independent bound on regret, where the O notation hides factors logarithmic in T . Just like in Section 6,
we can also derive regret bounds that vanish with b under the assumption that pulling any arm incurs some
positive amount of resource consumption in expectations for all resources, but this requires a minor tweak
of the algorithm.
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
THEOREM 10.
Suppose that:
ǫ ≤ min
i=1,··· ,C−1
k=1,··· ,K
µc
k(i)
27
and that b ≤ ǫ. If the decision maker artificially constrains himself or herself to a time horizon T = b· T /ǫ ≤
T , then the regret bounds derived in Theorems 8 and 9 hold with T substituted with T .
Similarly, if the decision maker is not constrained by a time horizon, artificially constraining himself or
herself to a time horizon T = mini=1,··· ,C B(i)/ǫ yields the regret bounds derived in Theorems 8 and 9 with
T substituted with T .
Applications.
In dynamic pricing and online advertising applications, Assumption 7 is usually not satis-
fied as pulling an arm typically incurs the consumption of only a few resources. We can relax this assumption
but this comes at the price of an additive term of order O(ln2(T )) in the final regret bounds.
THEOREM 11.
If Assumption 7 is not satisfied, the regret bounds derived in Theorems 8 and 9 hold
with σ = 0 up to an additive term of order O( (C+4)!4·B2
b·ǫ6
· ln2(T )).
As for Assumption 8, the existence of degenerate optimal bases to (3) is determined by a complex interplay
between the mean rewards and the mean amounts of resource consumption. However, we point out that
the set of parameters (µr
k(i))k=1,··· ,K,i=1,··· ,C that satisfy these conditions has Lebesgue
measure 0, hence such an event is unlikely to occur in practice. Additionally, while ǫ is typically not known
in applications, taking ǫ as a vanishing function of T yields the same asymptotic regret bounds.
k)k=1,··· ,K and (µc
8. Concluding remark.
In this paper, we develop an algorithm with a O(K C · ln(B)/∆) distribution-
dependent bound on regret, where ∆ is a parameter that generalizes the optimality gap for the standard
MAB problem. It is however unclear whether the dependence on K is optimal. Extensions discussed in
Section A of the Appendix suggest that it may be possible to achieve a linear dependence on K but this
calls for the development of more efficient load balancing algorithms.
Appendix A: Extensions.
A.1. Improving the multiplicative factors in the regret bounds.
A.1.1. A single limited resource whose consumption is deterministic.
If the amounts of resource
consumed are deterministic, we can substitute the notation µc
k for ck. Moreover, we can take λ = 1 and,
going through the analysis of Lemma 5, we can slightly refine the regret bound. Specifically, we have
3 , for any arm k such that ∆k > 0. Moreover, τ ∗ ≤ B/ǫ + 1 in this setting
E[nk,τ ∗] ≤ 16
(ck)2 · E[ln(τ ∗)]
(∆k)2 + π2
since:
B ≥
τ ∗−1
Xt=1
cat,t ≥ (τ ∗ − 1)· ǫ,
by definition of τ ∗. As a result, the regret bound derived in Theorem 1 turns into:
RB ≤ 16( Xk ∆k>0
1
ck · ∆k
)· ln(
B
ǫ
+ 1) + O(1),
which is identical (up to universal constant factors) to the bound obtained by Tran-Thanh et al. [24]. Note
that this bound is scale-free.
28
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
A.1.2. Arbitrarily many limited resources whose consumptions are deterministic. We propose
another load balancing algorithm that couples bases together. This is key to get a better dependence on K
because, otherwise, we have to study each basis independently from the other ones.
Algorithm: Load balancing algorithm Ax for a feasible basis x ∈ B.
If x is selected at time t, pull any arm at ∈ arg max
k∈Kx
ξx
k
nk,t
.
Observe that this load balancing algorithm is computationally efficient with a O(K) runtime (once we have
computed an optimal basic feasible solution to (8)) and requires O(K) memory space. The shortcoming of
this approach is that, if there are multiple optimal bases to (3), the optimal load balance for each optimal
basis will not be preserved since we take into account the number of times we have pulled each arm when
selecting any basis (for which we strive to enforce different ratios). Hence, the following assumption will
be required for the analysis.
ASSUMPTION 9. There is a unique optimal basis to (3).
Regret Analysis. All the proofs are deferred to Section G. We start by bounding, for each arm k, the
number of times this arm can be pulled when selecting any of the suboptimal bases. This is in stark contrast
with the analysis carried out in Section 5 where we bound the number of times each suboptimal basis has
been selected.
LEMMA 16. For any arm k ∈ {1,··· , K}, we have:
ρ· (PC
E[ Xx∈B k∈Kx, x6=x∗
where ∆k = minx∈B k∈Kx, x6=x∗ ∆x.
nx
k,τ ∗] ≤ 16
i=1 b(i))2
ǫ2
E[ln(τ ∗)]
(∆k)2 + K ·
π2
3
,
·
l,t)k,l∈Kx remain close to the optimal
l )k,l∈Kx at all times for the optimal basis x = x∗. This will not allow us to derive distribution-free
k,t/nx
k /ξx
In contrast to Section 5, we can only guarantee that the ratios (nx
ones (ξx
regret bounds for this particular class of load balancing algorithms.
LEMMA 17. At any time t and for any arm k ∈ Kx∗ , we have:
l − ρ· ( Xx∈B,x6=x∗
nk,t ≥ nx∗,t ·
ξx∗
l=1 ξx∗
k
and
PK
nk,t ≤ nx∗,t ·
PK
k
ξx∗
l=1 ξx∗
l
nx,t + 1)
+ Xx∈B,x6=x∗
nx,t + 1.
(15)
(16)
+ 1) + O(1),
Bringing everything together, we are now ready to establish regret bounds.
THEOREM 12. We have:
RB(1),··· ,B(C) ≤ 32
ρ3 · (PC
i=1 b(i))3
ǫ3 · b
· (
K
Xk=1
where the O notation hides universal constant factors.
1
(∆k)2 )· ln(PC
i=1 b(i)· B
ǫ
We derive a distribution-dependent regret bound of order O(ρ3· K · ln(B)
no non-trivial distribution-free regret bound.
∆2 ) where ∆ = minx∈B ∆x>0 ∆x but
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
29
A.1.3. Arbitrarily many limited resources. A straightforward extension of the load balancing algo-
rithm developed in the case of deterministic resource consumption in Section A.1.2 guarantees that the total
number of times any suboptimal basis is pulled is of order O(K · ln(T )). However, in contrast to Section
A.1.2, this is not enough to get logarithmic regret bounds as ξx
k,t fluctuates around the optimal load balance
k,t with a magnitude of order at least ∼ 1/√t, and, as a result, the ratios (E[nx
ξx
l,T ])k,l∈Kx might be
very different from the optimal ones (ξx
k,T ]/E[nx
k /ξx
l )k,l∈Kx.
Algorithm: Load balancing algorithm Ax for a feasible basis x ∈ B.
If x is selected at time t, pull any arm at ∈ arg max
k∈Kx
ξx
k,t
nk,t
.
LEMMA 18. For any arm k ∈ {1,··· , K}, we have:
E[ Xx∈B k∈Kx, ∆x>0
where ∆k = minx∈B k∈Kx, ∆x>0 ∆x.
nx
k,T ] ≤ 16C · λ2 ·
ln(T )
(∆k)2 + 212 K · (C + 3)!2
ǫ6
,
A.2. Relaxing Assumption 1. The regret bounds obtained in Sections 5, 6, and 7 can be extended
when the ratios converge as opposed to being fixed, as precisely stated below, but this requires slightly more
work.
ASSUMPTION 10. For any resource i ∈ {1,··· , C}, the ratio B(i)/B(C) converges to a finite value
b(i) ∈ (0, 1]. Moreover, b = min
i=1,··· ,C
To state the results, we need to redefine some notations and to work with the linear program:
b(i) is a positive quantity.
K
sup
(ξk)k=1,··· ,K
subject to
µr
k · ξk
K
Xk=1
Xk=1
ξk ≥ 0,
µc
k(i)· ξk ≤
,
B(i)
B(C)
k = 1,··· , K.
i = 1,··· , C
(17)
We redefine B as the set of bases that are feasible to (17) and, for x ∈ B, ∆x is redefined as the optimality
gap of x with respect to (17). We also redefine O = {x ∈ B ∆x = 0} as the set of optimal bases to (17).
Moreover, we define B∞ (resp. O∞) as the set of feasible (resp. optimal) bases to (3) and, for x ∈ B∞, ∆∞x
is the optimality gap of x with respect to (3). Our algorithm remains the same provided that we substitute
b(i) with B(i)/B(C) for any resource i ∈ {1,··· , C}. Specifically, Step-Simplex consists in solving:
K
sup
(ξk)k=1,··· ,K
subject to
K
Xk=1
Xk=1
ξk ≥ 0,
(¯ck,t(i)− ηi · ǫk,t)· ξk ≤
B(i)
B(C)
,
i = 1,··· , C
(18)
(¯rk,t + λ· ǫk,t)· ξk
k = 1,··· , K
and Step-Load-Balance is identical up to the substitution of b(i) with B(i)/B(C).
30
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
Regret Analysis. As it turns out, the logarithmic regret bounds established in Theorems 3, 5, and 8 do
not always extend when Assumption 1 is relaxed even though these bounds appear to be very similar to the
one derived in Theorem 1 when there is a single limited resource. The fundamental difference is that the
set of optimal bases may not converge while it is always invariant in the case of a single limited resource.
Typically, the ratios (B(i)/B(C))i=1,···,C may oscillate around (b(i))i=1,··· ,C in such a way that there exist
two optimal bases for (3) while there is a unique optimal basis for this same optimization problem whenever
the right-hand side of the inequality constraints is slightly perturbed around this limit. This alternately
causes one of these two bases to be slightly suboptimal, a situation difficult to identify and to cope with
for the decision maker. Nevertheless, this difficulty does not arise in several situations of interest which
generalize Assumption 1, as precisely stated below. The proofs are deferred to Section G.
Arbitrarily many limited resources whose consumptions are deterministic.
THEOREM 13.
Suppose that Assumption 10 holds. If there exists a unique optimal basis to (3) or if
B(i)/B(C)− b(i) = O(ln(B(C))/B(C)) for all resources i ∈ {1,··· , C − 1} then, we have:
RB(1),··· ,B(C) = O(
ρ·PC
i=1 b(i)
ǫ· b
· ( Xx∈B∞ ∆∞
x >0
1
∆∞x
)· ln(PC
i=1 b(i)· B(C)
ǫ
+ 1) +O∞·
ln(B(C))
ǫ· b
),
where the O notation hides universal constant factors.
A time horizon and another limited resource.
THEOREM 14.
(0, 1]. If there exists a unique optimal basis to (3) or if B/T − b = O(ln(T )/T ), then, we have:
Suppose that Assumptions 4, 5, and 6 hold and that the ratio B/T converges to b ∈
RB,T = O(
λ2
ǫ3 · ( Xx∈B∞ ∆∞
x >0
1
∆∞x
)· ln(T ) +
K 2 · σ
ǫ3
· ln(T )),
where the O notation hides universal constant factors.
Arbitrarily many limited resources with a time horizon.
THEOREM 15.
Suppose that Assumptions 7, 8, and 10 hold. If there exists a unique optimal basis to (3)
or if B(i)/T − b(i) = O(ln(T )/T ) for all resources i ∈ {1,··· , C − 1}, then, we have:
RB(1),··· ,B(C−1),T = O(
(C + 3)!3 · λ2
ǫ6
· ( Xx∈B∞ ∆∞
x >0
1
∆∞x
)· ln(T ) +
σ ·B∞· (C + 3)!4
ǫ6
· ln(T )),
where the O notation hides universal constant factors.
Appendix B: Proofs for Section 3.
B.1. Proof of Lemma 2. The proof can be found in Badanidiyuru et al. [8]. For the sake of complete-
ness, we reproduce it here. The optimization problem (3) is a linear program whose dual reads:
C
inf
(ζi)i=1,··· ,C
subject to
b(i)· ζi
C
Xi=1
Xi=1
ζi ≥ 0,
µc
k(i)· ζi ≥ µr
k,
k = 1,··· , K
(19)
i = 1,··· , C.
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
31
Observe that (3) is feasible therefore (3) and (19) have the same optimal value. Note that (3) is bounded
under Assumption 2 as ξk ∈ [0, b(i)/µc
k(i)] for any feasible point and any resource i ∈ {1,··· , C} such that
k(i) > 0. Hence, (19) has an optimal basic feasible solution (ζ∗1 ,··· , ζ∗C). Consider any non-anticipating
µc
algorithm. Let Zt be the sum of the total payoff accumulated in rounds 1 to t plus the "cost" of the remaining
i=1 ζ∗i · (B(i)−Pt
τ =1 raτ ,τ +PC
resources, i.e. Zt =Pt
τ =1 caτ ,τ (i)). Observe that (Zt)t is a supermartingale
i=1 ζ∗i · µc
with respect to the filtration (Ft)t as E[Zt Ft−1] = E[µr
at(i) Ft−1] + Zt−1 ≤ Zt−1 since
(ζ∗1 ,··· , ζ∗C) is feasible for (19). Moreover, note that (Zt)t has bounded increments since Zt − Zt−1 =
i=1 ζ∗i · cat,t(i) ≤ 1 +PC
rat,t −PC
i=1 ζ∗i < ∞. We also have E[τ ∗] < ∞ as:
at −PC
∞
E[τ ∗] =
t
∞
∞
P[
t−1
≤
Xt=1
P[τ ∗ ≥ t]
Xτ =1
Xt=1
Xt=1
≤ 1 +
≤ (PC
< ∞,
i=1 B(i)
P[
ǫ
C
caτ ,τ (i) ≤ B(i), i = 1,··· , C]
Xτ =1
caτ ,τ (i) ≤ t· ǫ− (t· ǫ−
Xt≥
exp(−
Xi=1
+ 2) +
i=1
ǫ
PC
∞
B(i)
C
B(i))]
Xi=1
2(t· ǫ−PC
i=1 B(i))2
t
)
where the third inequality results from an application of Lemma 1 and
µc
k(i).
ǫ = min
k=1,··· ,K
i=1,··· ,C
with µc
k (i)>0
By Doob's optional stopping theorem, E[Zτ ∗] ≤ E[Z0] =PC
C
i=1 ζ∗i · B(i). Observe that:
Xi=1
ζ∗i · caτ ∗ ,τ ∗(i) + Zτ ∗−1]
τ ∗−1
Xt=1
Using Assumption 2 and since (ζ∗i )i=1,··· ,C is a basic feasible solution, for every i ∈ {1,··· , C} such that
ζ∗i > 0 there must exist k ∈ {1,··· , K} such that ζ∗i ≤ µr
E[Zτ ∗] = E[raτ ∗ ,τ ∗ −
Xi=1
k(i) > 0. We get:
k(i) with µc
≥ E[−
rat,t].
ζ∗i +
k/µc
C
E[Zτ ∗] ≥ E[
τ ∗−1
Xt=1
rat,t]− max
k=1,··· ,K
i=1,··· ,C
with µc
k(i)>0
µr
k
k(i)
µc
and finally:
E[
τ ∗−1
Xt=1
rat,t] ≤
C
Xi=1
Xi=1
= B ·
C
ζ∗i · B(i) + max
k=1,··· ,K
i=1,··· ,C
with µc
k (i)>0
µr
k
k(i)
µc
ζ∗i · b(i) + max
k=1,··· ,K
i=1,··· ,C
with µc
k(i)>0
µr
k
k(i)
µc
.
By strong duality,PC
i=1 ζ∗i · b(i) is also the optimal value of (3).
32
Appendix C: Proofs for Section 4.
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
C.1. Proof of Lemma 4. By definition of τ ∗, we havePτ ∗−1
sides yields:
t=1 cat,t ≤ B. Taking expectations on both
τ ∗−1
τ ∗
=
∞
cat,t]
Xt=1
B ≥ E[
Xt=1
cat,t]− 1
≥ E[
∞
Xt=1
E[Iτ ∗≥t · cat,t]− 1
Xt=1
Xt=1
Xt=1
≥
= E[τ ∗]· ǫ− 1,
E[Iτ ∗≥t · ǫ]− 1
E[Iτ ∗≥t · µc
at ]− 1
∞
∞
=
=
E[Iτ ∗≥t · E[cat,t Ft−1]]− 1
where we use the fact that ck,t ≤ 1 for all arms k to derive the second inequality, the fact that τ ∗ is a stopping
time for the second equality, the fact that at is deterministically determined by the past, i.e. at ∈ Ft−1, for
the third equality and Assumption 2 for the third inequality. We conclude that E[τ ∗] ≤ B+1
.
ǫ
C.2. Proof of Lemma 5. We break down the analysis in a series of facts. Consider any arm k such that
∆k > 0. We use the shorthand notation βk = 25( λ
µc
k
)2 · ( 1
∆k
)2.
Fact 1
τ ∗
Xt=1
E[nk,τ ∗] ≤ 2βk · E[ln(τ ∗)] + E[
Iat=k · Ink,t≥βk ln(t)].
(20)
Proof. Define the random variable Tk = βk · ln(τ ∗). We have:
E[nk,τ ∗] = E[nk,τ ∗ · Ink,τ ∗ <Tk] + E[nk,τ ∗ · Ink,τ ∗≥Tk ]
≤ βk · E[ln(τ ∗)] + E[nk,τ ∗ · Ink,τ ∗≥Tk ].
Define T ∗k as the first time t such that nk,t ≥ Tk and T ∗k = ∞ if no such t exists. We have:
τ ∗
Xt=1
Iat=k · Ink,τ ∗≥Tk ]
E[nk,τ ∗ · Ink,τ ∗≥Tk] = E[
Xt=1
k −1 · Ink,τ ∗≥Tk] + E[
≤ E[nk,T ∗
Iat=k · Ink,τ ∗≥Tk ] + E[
T ∗
k −1
= E[
τ ∗
k
τ ∗
Iat=k · Ink,τ ∗≥Tk]
Xt=T ∗
Xt=T ∗
Iat=k · Ink,t≥Tk ]
k
τ ∗
Xt=1
≤ βk · E[ln(τ ∗)] + E[
Iat=k · Ink,t≥βk ln(t)],
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
33
since, by definition of T ∗k , nk,T ∗
k −1 ≤ Tk if T ∗k if finite, which is always true if nk,τ ∗ ≥ Tk (the sequence
(nk,t)t is non-decreasing and τ ∗ is finite almost surely as a byproduct of Lemma 4). Conversely, nk,t ≥ Tk ≥
βk ln(t) for t ∈ {T ∗k ,··· , τ ∗}. Wrapping up, we obtain:
Xt=1
E[nk,τ ∗] ≤ 2βk · E[ln(τ ∗)] + E[
Iat=k · Ink,t≥βk ln(t)].
τ ∗
Fact 1 enables us to assume that arm k has been pulled at least βk ln(t) times out of the last t time periods.
The remainder of this proof is dedicated to show that the second term of the right-hand side of (20) can be
bounded by a constant. Let us first rewrite this term:
τ ∗
E[
Xt=1
τ ∗
τ ∗
Xt=1
Iat=k · Ink,t≥βk ln(t)] ≤ E[
Xt=1
≤ E[
Xt=1
Xt=1
+ E[
+ E[
τ ∗
τ ∗
Iobjk,t+Ek,t≥objk∗,t+Ek∗,t · Ink,t≥βk ln(t)]
Iobjk,t≥objk+Ek,t]
Iobjk∗,t≤objk∗−Ek∗,t]
Iobjk∗ <objk+2Ek,t · Ink,t≥βk ln(t)].
(21)
(22)
(23)
To derive this last inequality, simply observe that if objk,t < objk + Ek,t and objk∗,t > objk∗ − Ek∗,t while
objk,t + Ek,t ≥ objk∗,t + Ek∗,t, it must be that objk∗ < objk + 2Ek,t. Let us study (21), (22), and (23)
separately.
Fact 2
τ ∗
Iobjk∗ <objk+2Ek,t · Ink,t≥βk ln(t)] ≤
2π2
3ǫ2 .
E[
Xt=1
Proof. Observe that when both nk,t ≥ βk ln(t) and objk∗ < objk + 2Ek,t, we have:
∆k
2
< Ek,t
λ
¯ck,t ·r 2
µc
k
2¯ck,t ·
βk
∆k
2
,
≤
≤
by definition of βk. This implies that ¯ck,t ≤ µc
k/2. Thus:
τ ∗
τ ∗
E[
Xt=1
Xt=1
Iobjk∗ <objk+2Ek,t · Ink,t≥βk ln(t)] ≤ E[
I¯ck,t<µc
k /2 · Ink,t≥βk ln(t)]
We upper bound this last term using the concentration inequalities of Lemma 1 observing that:
E[
∞
Xt=1
I¯ck,t<µc
k /2 · Ink,t≥βk ln(t)] =
≤
∞
∞
Xt=1
Xt=1
µc
k
2
; nk,t ≥ βk ln(t)]
P[¯ck,t < µc
k −
µc
k
2
; nk,t = s].
P[¯ck,t <
t
Xs=βk ln(t)
34
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
Denote by t1,··· , ts the first s random times at which arm k is pulled (these random variables are finite
almost surely). We have:
P[¯ck,t < µc
k −
µc
k
2
s
Xl=1
; nk,t = s] ≤ P[
ck,tl < s· µc
k − s·
µc
k
2
].
Observe that, for any l ≤ s:
∞
Xτ =1
E[ck,tl ck,t1,··· , ck,tl−1] = E[
Xτ =1
= E[
∞
= µc
k,
Itl=τ · E[ck,τ Fτ−1] ck,t1,··· , ck,tl−1]
Itl=τ · µc
k ck,t1,··· , ck,tl−1]
since the algorithm is not randomized ({tl = τ} ∈ Fτ−1) and using the tower property. Hence, we can apply
Lemma 1 to get:
P[¯ck,t <
µc
k
2
∞
Xt=1
; nk,t ≥ βk ln(t)] ≤
k)2
(µc
2
)
· βk ln(t))
∞
∞
k)2
2
∞
1
∞
≤
≤
exp(−s·
k)2
2
Xs=βk ln(t)
exp(− (µc
1− exp(− (µc
Xt=1
Xt=1
Xt=1
1− exp(− (µc
2π2
3(µc
k)2
2π2
3ǫ2 ,
)2 ≥ 25( 1+ 1
≤
≤
k)2
2
(µc
)
)
1
t2
where we use the fact that βk ≥ 25( 1+κ
Assumption 3), the fact that exp(−x) ≤ 1 − x
µc
k
the last inequality.
µr
k∗
k∗
)2 · ( µc
k)2 for the third inequality (using
2 for x ∈ [0, 1] for the fourth inequality, and Assumption 2 for
)2 ≥ 4
κ
µc
k
Let us now elaborate on (21).
Fact 3
τ ∗
Proof. Note that if ¯rk,t/¯ck,t = objk,t ≥ objk + Ek,t = µr
µc
k − ǫk,t, otherwise we would have:
¯rk,t
¯ck,t −
k)µc
=
µr
k
µc
k
E[
Xt=1
Iobjk,t≥objk+Ek,t] ≤
k/µc
π2
3
.
k + Ek,t, then either ¯rk,t ≥ µr
k + ǫk,t or ¯ck,t ≤
k − ¯ck,t)µr
k
k + (µc
¯ck,t · µc
µr
k
µc
k
k
<
(¯rk,t − µr
ǫk,t
ǫk,t
¯ck,t ·
¯ck,t
ǫk,t
¯ck,t
+
≤ (1 + κ)·
= Ek,t,
a contradiction. Therefore:
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
35
k + ǫk,t] + P[¯ck,t ≤ µc
k − ǫk,t]
τ ∗
E[
Xt=1
Iobjk,t≥objk+Ek,t] ≤
≤
+
=
+
≤
=
∞
∞
∞
∞
Xt=1
Xt=1
Xt=1
Xt=1
Xt=1
Xt=1
∞
∞
π2
3
,
t
t
t
s
s
k +r 2 ln(t)
k −r 2 ln(t)
rk,tl ≥ s· µr
P[¯rk,t ≥ µr
P[¯rk,t ≥ µr
Xs=1
Xs=1
Xs=1
Xs=1
Xs=1
P[¯ck,t ≤ µc
Xl=1
Xl=1
2 exp(−4 ln(t))
ck,tl ≤ s· µc
P[
P[
s
s
t
t
; nk,t = s]
; nk,t = s]
k +ps· 2 ln(t) ; nk,t = s]
k −ps· 2 ln(t) ; nk,t = s]
where the random variables (tl)l are defined similarly as in the proof of Fact 2 and the last inequality results
from an application of Lemma 1.
What remains to be done is to bound (22).
Fact 4
τ ∗
Iobjk∗,t≤objk∗−Ek∗,t] ≤
π2
3
.
E[
Xt=1
Proof. We proceed along the same lines as in the proof of Fact 3. As a matter of fact, the situation is
perfectly symmetric because, in the course of proving Fact 3, we do not rely on the fact that we have pulled
k∗ − Ek∗,t,
arm k more than βk ln(t) times at any time t. If ¯rk∗,t/¯ck∗,t = objk∗,t ≤ objk∗ − Ek∗,t = µr
then we have either ¯rk∗,t ≤ µr
k∗ + ǫk∗,t, otherwise we would have:
k∗ /µc
k∗ − ǫk∗,t or ¯ck∗,t ≥ µc
¯rk∗,t
¯ck∗,t −
µr
k∗
µc
k∗
=
k∗ )µc
k∗ + (µc
¯ck∗,t · µc
k∗
µr
k∗
µc
k∗
ǫk∗,t
¯ck∗,t ·
ǫk∗,t
¯ck∗,t
(¯rk∗,t − µr
ǫk∗,t
> −
¯ck∗,t −
≥ −(1 + κ)·
= −Ek∗,t,
k∗ − ¯ck∗,t)µr
k∗
a contradiction. Therefore:
τ ∗
E[
Xt=1
∞
k∗ +ǫk,t]
t
Xt=1
Iobjk∗,t≤objk∗−Ek∗,t] ≤ E[
I¯rk∗,t≤µr
∞
Xs=1
Xt=1
Xs=1
Xt=1
≤
∞
+
t
k∗−ǫk,t + I¯ck∗,t≥µc
k∗ −r 2 ln(t)
k∗ +r 2 ln(t)
s
s
P[¯rk∗,t ≤ µr
P[¯ck∗,t ≥ µc
; nk∗,t = s]
; nk∗,t = s]
36
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
2
t4
t
Xs=1
≤
=
∞
Xt=1
π2
3
,
where the third inequality is obtained using Lemma 1 in the same fashion as in Fact 3.
We conclude:
E[nk,τ ∗] ≤ 2βk · E[ln(τ ∗)] +
4π2
3ǫ2 .
C.3. Proof of Theorem 1. First observe that:
τ ∗
E[
Xt=1
rat,t] =
=
=
=
=
∞
∞
∞
Xt=1
Xt=1
Xt=1
Xk=1
Xk=1
K
K
E[Iτ ∗≥t · E[rat,t Ft−1]]
K
E[Iτ ∗≥t · µr
at]
Xk=1
µr
k · E[Iτ ∗≥t · Iat=k]
∞
Xt=1
µr
k · E[
µr
k · E[nk,τ ∗],
Iτ ∗≥t · Iat=k]
since τ ∗ is a stopping time. Plugging this equality into (9) yields:
µr
k · E[nk,τ ∗] + O(1)
RB ≤ B ·
µr
k∗
µc
=
K
Xk=1
µr
k∗
k∗ −
µc
k∗ · (B − Xk ∆k=0
µc
k · E[nk,τ ∗])− Xk ∆k>0
Along the same lines as for the rewards, we can show that E[Pτ ∗
of τ ∗, we have B ≤Pτ ∗
K
µr
k · E[nk,τ ∗ ] + O(1).
t=1 cat,t almost surely. Taking expectations on both sides yields:
t=1 cat,t] =PK
k=1 µc
k · E[nk,τ ∗]. By definition
Xk=1
B ≤
= Xk ∆k=0
µc
k · E[nk,τ ∗]
k · E[nk,τ ∗] + Xk ∆k>0
µc
µc
k · E[nk,τ ∗].
Plugging this inequality back into the regret bound, we get:
RB ≤ Xk ∆k>0
= Xk ∆k>0
(
k − µr
µr
k∗
k∗ · µc
µc
µc
k · ∆k · E[nk,τ ∗ ] + O(1).
k)· E[nk,τ ∗] + O(1)
(24)
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
37
Using the upper bound of Lemma 4, the concavity of the logarithmic function, and Lemma 5, we derive:
RB ≤ 26λ2 · ( Xk ∆k>0
≤ 26λ2 · ( Xk ∆k>0
k∗ ≤ κ and µc
since ∆k ≤ µr
k∗/µc
k ≤ 1 for any arm k.
1
µc
k · ∆k
1
µc
k · ∆k
)· ln(
)· ln(
B + 1
ǫ
B + 1
ǫ
) +
4π2
3ǫ2 · ( Xk ∆k>0
) + K ·
4π2κ
3ǫ2 + O(1),
µc
k · ∆k) + O(1)
Appendix D: Proofs for Section 5.
D.1. Proof of Lemma 8. Consider any suboptimal basis x ∈ B. The proof is along the same lines as
for Lemma 5 and follows the exact same steps. We use the shorthand notation βx = 8ρ· ( PK
k=1 ξx
k
∆x
)2.
Fact 5
Xt=1
E[nx,τ ∗] ≤ 2βx · E[ln(τ ∗)] + E[
Ixt=x · Inx,t≥βx ln(t)].
τ ∗
(25)
We omit the proof as it is analogous to the proof of Fact 1. As in Lemma 5, we break down the second term
in three terms and bound each of them by a constant:
τ ∗
E[
Xt=1
τ ∗
τ ∗
Xt=1
Ixt=x · Inx,t≥βx ln(t)] = E[
Xt=1
≤ E[
Xt=1
Xt=1
+ E[
+ E[
τ ∗
τ ∗
Iobjx,t+Ex,t≥objx∗,t+Ex∗,t · Inx,t≥βx ln(t)]
Iobjx,t≥objx+Ex,t]
Iobjx∗,t≤objx∗−Ex∗,t]
Iobjx∗ <objx+2Ex,t · Inx,t≥βx ln(t)].
(26)
(27)
(28)
Fact 6
τ ∗
Iobjx∗ <objx+2Ex,t · Inx,t≥βx ln(t)] = 0.
E[
Xt=1
Proof. If objx∗ < objx + 2Ex,t, we get:
∆x
2
ξx
< Xk∈Kx
≤ Xk∈Kx
≤sXk∈Kx
nk,t
k ·s 2 ln(t)
k ·s 2 ln(t)
k · Xk∈Kxpξx
ρ + nx
k,t
ξx
ξx
k ·s 2 ln(t)
nx,t
,
38
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
where we use (10) and Lemma 6 for each k ∈ Kx such that ξx
k 6= 0 (otherwise, if ξx
k = 0, the inequality is
trivial). This implies:
nx,t < 8ρ· (PK
k=1 ξx
∆x
= βx · ln(t),
k
)2 · ln(t)
using the Cauchy−Schwarz inequality and the fact that a basis involves at most ρ arms.
Fact 7
τ ∗
Iobjx,t≥objx+Ex,t] ≤ ρ·
π2
6
.
E[
Xt=1
Proof. If objx,t ≥ objx + Ex,t, there must exist k ∈ Kx such that ¯rk,t ≥ µr
k + ǫk,t, otherwise:
objx,t − objx = Xk∈Kx
< Xk∈Kx
= Ex,t,
k)· ξx
k
(¯rk,t − µr
ǫk,t · ξx
k
where the inequality is strict because there must exist l ∈ Kx such that ξx
l > 0 as a result of Assumption 2
(at least one resource constraint is binding for a feasible basis to (3) aside from the basis x associated with
Kx = ∅). We obtain:
τ ∗
E[
Xt=1
Iobjx,t≥objx+Ex,t] ≤ Xk∈Kx
≤ ρ·
π2
6
∞
Xt=1
,
P[¯rk,t ≥ µr
k + ǫk,t]
where the last inequality is derived along the same lines as in the proof of Fact 3.
Fact 8
Proof. Similar to Fact 7.
τ ∗
E[
Xt=1
Iobjx∗,t≤objx∗−Ex∗,t] ≤ ρ·
π2
6
.
D.2. Proof of Theorem 3. The proof proceeds along the same lines as for Theorem 1. We build upon
(4):
K
τ ∗
RB(1),··· ,B(C) ≤ B ·
= B ·
= B ·
K
Xk=1
Xk=1
Xk=1
K
k · ξx∗
µr
k · ξx∗
µr
k −
Xt=1
k − E[
∞
Xt=1
E[Iτ ∗≥t ·
Xt=1
E[Iτ ∗≥t ·
∞
k · ξx∗
µr
k −
rat,t] + O(1)
K
K
Xk=1Xx∈B
Xk=1Xx∈B
rk,t · Ixt=x,at=k] + O(1)
Ixt=x,at=k · E[rk,t Ft−1]] + O(1)
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
39
= B ·
= B ·
K
K
Xk=1
Xk=1
k · ξx∗
k −Xx∈B
µr
k · ξx∗
k −Xx∈B
µr
K
K
Xk=1
Xk=1
τ ∗
Xt=1
µr
k · E[
k · E[nx
µr
Ixt=x,at=k] + O(1)
k,τ ∗] + O(1),
where we use the fact that xt and at are determined by the events of the first t − 1 rounds and that τ ∗ is a
stopping time. Using the properties of the load balancing algorithm established in Lemma 6, we derive:
K
ξx
k
l=1 ξx
K
K
K
= (
RB(1),··· ,B(C) ≤ B ·
{µr
k ·
E[nx,τ ∗]
k=1 ξx
l · E[nx,τ ∗]− ρ} + O(1)
k )− (ρ)2} + O(1)
µr
k · ξx
k · ξx∗
Xk=1
k −Xx∈B Xk∈Kx
µr
k · ξx∗
Xk=1
k −Xx∈B
µr
{
= B ·
PK
k · ξx∗
Xk=1
k )· (B − Xx∈B ∆x=0
µr
Xk=1
− Xx∈B ∆x>0
{(
PK
Now observe that, by definition, at least one resource is exhausted at time τ ∗. Hence, there exists i ∈
{1,··· , C} such that the following holds almost surely:
ck(i)· nx
PK
Xk=1
k · (
E[nx,τ ∗]
PK
k=1 ξx
k } + O(1).
E[nx,τ ∗]
k=1 ξx
µr
k · ξx
k )·
K
)
k
k,τ ∗
ξx
k
l=1 ξx
B(i) ≤Xx∈B Xk∈Kx
≤Xx∈B Xk∈Kx
[ck(i)· (
PK
k · Xk∈Kx
= B· ρ +Xx∈B
PK
≤ B· ρ + b(i)·Xx∈B
PK
nx,τ ∗
k=1 ξx
nx,τ ∗
k=1 ξx
k
,
l · nx,τ ∗ + 1)]
ck(i)· ξx
k
where we use Lemma 6 again and the fact that any basis x ∈ B satisfies all the constraints of (3). We
conclude that the inequality:
Xx∈B ∆x=0
nx,τ ∗
k=1 ξx
k
PK
≥ B − Xx∈B ∆x>0
nx,τ ∗
k=1 ξx
k
− B· ρ
b
PK
holds almost surely. Taking expectations on both sides and plugging the result back into the regret bound
yields:
K
k=1 µr
(PK
RB(1),··· ,B(C) ≤ Xx∈B ∆x>0
k )· B· ρ
k · ξx∗
Xk=1
µr
≤ Xx∈B ∆x>0
PK
∆x
k=1 ξx
+ (
b
k=1 µr
k · ξx
k )
k · ξx∗
k −PK
PK
k=1 ξx
+ O(1)
k
k · E[nx,τ ∗] + O(1),
· E[nx,τ ∗]
(29)
(30)
40
where we use the fact that:
K
Xk=1
k · ξx∗
µr
k ≤
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
(31)
i=1 ck(i)
k
· ξx∗
ck(i)· ξx∗
k
K
C
=
ǫ
K
Xk=1 PC
1
Xk=1
Xi=1
ǫ ·
≤ PC
i=1 b(i)
ǫ
,
using Assumption 2 and the fact that x∗ is a feasible basis to (3). Using Lemma 7, Lemma 8, and the
concavity of the logarithmic function, we obtain:
k
ǫ
+
+ 1)
π2
3
i=1 b(i)· B
)· ln(PC
k=1 ξx
∆x
∆x
k=1 ξx
k
RB(1),··· ,B(C) ≤ 16ρ· ( Xx∈B ∆x>0PK
ρ· ( Xx∈B ∆x>0
PK
ρ·PC
· ( Xx∈B ∆x>0
To derive this last inequality, we use: (i) ∆x ≤PK
i=1 b(i)/ǫ (see (31)), (ii) the fact that,
for any basis x ∈ B, at least one of the first C inequalities is binding in (3), which implies that there exists
i ∈ {1,··· , C} such that:
)· ln(PC
k ≤PC
i=1 b(i)· B
i=1 b(i)
ǫ
+ 1) + O(1).
k · ξx∗
) + O(1)
≤ 16
k=1 µr
1
∆x
ǫ
ck(i)· ξx
k
K
Xk=1
K
Xk=1
ξx
k ≥
= b(i)
≥ b,
and (iii) the inequality:
K
Xk=1
i=1 ck(i)
· ξx
k
ck(i)· ξx
k
ξx
k ≤
K
C
=
ǫ
K
Xk=1 PC
1
Xk=1
Xi=1
ǫ ·
≤ PC
i=1 b(i)
ǫ
,
for any basis x ∈ B.
As a side note, observe a possibly better regret bound is given by:
RB(1),··· ,B(C) ≤ 16ρ· ( Xx∈B ∆x>0
)· ln(T ) + O(1),
if time is a limited resource since, in this case, τ ∗ ≤ T and the constraintPK
1
∆x
D.3. Proof of Theorem 4. Along the same lines as for the case of a single limited resource, we start
from inequality (30) derived in the proof of Theorem 3 and apply Lemma 8 only if ∆x is big enough, taking
into account the fact that:
k=1 ξx
k ≤ 1 is part of (3).
E[nx,τ ∗] ≤ E[τ ∗] ≤PC
ǫ
i=1 b(i)· B
+ 1.
Xx∈B
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
Specifically, we have:
41
RB(1),··· ,B(C)
≤
sup
Px∈B nx≤
(nx)x∈B≥0
ǫ
PC
i=1 b(i)·B
k=1 ξx
∆x
k
16ρ·PK
sup
Px∈B nx≤
(nx)x∈B≥0
ǫ
PC
i=1 b(i)·B
k=1 ξx
∆x
k
16ρ·PK
sup
≤
≤
+1
ǫ
{ Xx∈B ∆x>0
· ln(PC
{ Xx∈B ∆x>0
· ln(PC
{ Xx∈B ∆x>0
ǫ
+1
(nx)x∈B≥0
ǫ
PC
i=1 b(i)·B
+1
i=1 b(i)· B
Px∈B nx≤
ǫ
≤ 4sρ· ln(PC
≤ 4sρ·B· (PC
i=1 b(i)· B
where we use ∆x ≤PC
i=1 b(i)/ǫ and PK
Theorem 3), we maximize over ∆x/PK
Cauchy-Schwartz for the last inequality.
ǫ
Appendix E: Proofs for Section 6.
min(
∆x
k=1 ξx
k · nx,
PK
i=1 b(i)· B
+ 1) +
min(
∆x
k=1 ξx
PK
+ 1)) } +
i=1 b(i)· B
) } + O(1)
∆x
k=1 ξx
k
π2
3
ρ·
PK
k · nx,
π2
3
B· ρ
ǫ
+ O(1)
s16ρ· ln(PC
i=1 b(i)· B
ǫ
+ 1)· nx } + O(1)
+ 1)·
sup
(nx)x∈B≥0
ǫ
PC
i=1 b(i)·B
Px∈B nx≤
{ Xx∈B ∆x>0
+1
√nx } + O(1)
ǫ
+ 1) + O(1),
i=1 b(i)· B
+ 1)· ln(PC
k ≥ b for the second inequality (see the end of the proof of
k ≥ 0 for each x ∈ B to derive the third inequality, and we use
k=1 ξx
k=1 ξx
E.1. Proof of Lemma 9. For any T ∈ N and any arm k ∈ {1,··· , K}, we denote by nopt
k,T the expected
number of times that arm k is pulled by the optimal non-anticipating algorithm (which is characterized by a
high-dimensional dynamic program) when the time horizon is T and the budget is B = b · T . We prove the
k∗,T = Ω(√T ) (Case A) or T − nopt
k∗,T = o(√T ) (Case B) then
claim in two steps. First, we show that if T − nopt
k∗,T 6= Ω(√T )
EROPT(B, T ) = T · objx∗ − Ω(√T ). This is enough to establish the result because if T − nopt
k∗,T )/√T that converges to 0 and we can conclude with Case
then we can extract a subsequence of (T − nopt
B.
Case A: T − nopt
k∗,T = Ω(√T ). Consider the linear program:
K
sup
(ξk)k=1,··· ,K
subject to
µr
k · ξk
µc
k · ξk ≤ b
ξk ≤ 1
K
Xk=1
Xk=1
Xk=1
ξk∗ ≤ Γ
ξk ≥ 0,
K
k = 1,··· , K
(32)
42
parametrized by Γ and its dual:
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
inf
b· ζ1 + ζ2 + Γ· ζ3
(ζ1,ζ2,ζ3)
subject to µc
k · ζ1 + ζ2 ≥ µr
k,
k∗ · ζ1 + ζ2 + ζ3 ≥ µr
µc
ζ1, ζ2, ζ3 ≥ 0.
k 6= k∗
k∗
(33)
Since the vector (ξk)k=1,··· ,K determined by ξk∗ = 1 and ξk = 0 for k 6= k∗ is the only optimal solution
to (32) when Γ = 1 (by assumption), we can find a strictly complementary optimal solution to the dual
(33) ζ∗1 , ζ∗2 , ζ∗3 > 0. Moreover, by definition of nopt
k∗,T , we can show, along the same lines as in the proof of
Lemma 2, that EROPT(B, T ) is no larger than T times the optimal value of (32) when Γ = nopt
k∗,T /T (up to
a constant additive term of order O(1)). By weak duality, and since (ζ∗1 , ζ∗2 , ζ∗3 ) is feasible for (33) for any
Γ, this implies:
EROPT(B, T ) ≤ T · (b· ζ∗1 + ζ∗2 +
≤ T · (b· ζ∗1 + ζ∗2 + ζ∗3 −
≤ T · objx∗ − Ω(√T ),
nopt
k∗,T
T · ζ∗3 ) + O(1)
T − nopt
k∗,T
T
· ζ∗3 ) + O(1)
where we use the fact that b · ζ∗1 + ζ∗2 + ζ∗3 is the optimal value of (32) when Γ = 1 by strong duality (both
(32) and (33) are feasible) and ζ∗3 > 0.
k∗,T = o(√T ). First observe that since the vector (ξk)k=1,··· ,K determined by ξk∗ = 1 and
ξk = 0 for k 6= k∗ is the only optimal solution to (3), it must be that µr
k∗ > 0 (since 0 is a feasible solution
to (3) with objective value 0). For any t ∈ N, denote by at the arm pulled by the optimal non-anticipating
algorithm at time t and define τ ∗T as the corresponding stopping time when the time horizon is T . We have:
Case B: T − nopt
EROPT(B, T ) = E[
τ ∗
T −1
K
k,T
=
rat,t]
Xt=1
Xk=1
k · nopt
µr
≤ Xk6=k∗
nopt
k,T + µr
≤ (T − nopt
k∗,T ) + µr
= T · µr
k∗ − µr
= T · objx∗ − µr
≤ T · objx∗ − µr
≤ T · objx∗ − µr
= T · objx∗ − µr
k∗,T
k∗ · (E[τ ∗T ]− 1)
k∗ · nopt
k∗ · (T − E[τ ∗T ] + 1) + o(√T )
k∗ · E[
1] + o(√T )
T
T
T
τ ∗
T −1
ck∗,t +
Xt=1
Xt=τ ∗
Xt=τ ∗
cat,t − B)+]
k∗ · E[(
k∗ ·(cid:16)E[(
Xt=1
{ck∗,t − b})+]− Xk6=k∗
{ck∗,t − b})+] + o(√T ).
Xt=1
k∗ · E[(
T
T
T
nopt
k,T(cid:17) + o(√T )
The first inequality is obtained using the fact that the rewards are bounded by 1. The second inequality
k,T = E[τ ∗T ] − 1 ≤ T . The third inequality is obtained along the
k=1 nopt
is obtained using the fact that PK
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
43
same lines as in the proof of Lemma 3, usingPτ ∗
T −1
t=1 cat,t ≤ B by definition of τ ∗T . We use the inequality
(y + z)+ ≥ y+ − z (true for any (y, z) ∈ R2) and the fact the amount of resource consumed at any step
is no larger than 1 for the fourth inequality. Since (ck∗,t − b)t∈N is an i.i.d. zero-mean bounded stochastic
t=1{ck,t − b})+] converges to a positive value and we conclude:
process with positive variance,
1√T · E[(PT
EROPT(B, T ) ≤ T · objx∗ − Ω(√T ),
since µr
k∗ > 0.
E.2. Proof of Lemma 10. Consider x either an infeasible basis to (3) or a pseudo-basis for (3). Without
loss of generality, we can assume that x involves two arms (one of which may be a dummy arm introduced
in the specification of the algorithm given in Section 6) and that Kx = {k, l} with µc
l > b (the situation
is symmetric if the reverse inequality holds). Defining βx = 32/ǫ3, we have:
k, µc
T
Ixt=x · Inx,t≥βx ln(t) · Ibx,t≥nx,t·(b+ǫ/2)]
t
T
T
+
Ixt=x · Inx,t≥βx ln(t)]
Xt=1
E[nx,T ] ≤ 2βx · ln(T ) + E[
Xt=1
≤ 2βx · ln(T ) + E[
Xs=βx ln(T )
Xt=1
P[bx,t < s· (b + ǫ)− s· ǫ/2, nx,t = s]
Xt=1
≤ 2βx · ln(T ) + E[
Xt=1
≤ 2βx · ln(T ) + E[
T
T
Ixt=x · Inx,t≥βx ln(t) · Ibx,t≥nx,t·(b+ǫ/2)] +
Xt=1
Ixt=x · Inx,t≥βx ln(t) · Ibx,t≥nx,t·b+ǫβx/4·ln(t)] +
Xs=βx ln(T )
2π2
3ǫ2 .
T
∞
exp(−s
ǫ2
2
)
The first inequality is derived along the same lines as in Fact 1. The third inequality is obtained by observing
that, as a result of Assumption 6, the average amount of resource consumed any time basis x is selected
at Step-Simplex is at least b + ǫ no matter which of arm k or l is pulled. Finally, we use the same bounds
as in Fact 2 for the last two inequalities. Observe that if x is selected at time t, either ¯ck,t − ǫk,t ≤ b or
¯cl,t − ǫl,t ≤ b, otherwise x would have been infeasible for (8). Moreover, if nx,t ≥ βx ln(T ), then we have
either nx
k,t ≥ βx/2 ln(T ) or nx
l,t ≥ βx/2 ln(T ) since there are only two arms in Kx. By symmetry, we study
the first situation and look at:
T
E[
k,t≥βx/2·ln(t) · Ibx,t≥nx,t·b+ǫβx/4·ln(t)]
k,t≥βx/2·ln(t) · Ibx,t≥nx,t·b+ǫβx/4·ln(t) · I¯ck,τq−ǫk,τq≥b, q=qt−ǫβx/4 ln(t),··· ,qt ]
t
T
T
+
Xt=1
Ixt=x · Inx
Xt=1
Ixt=x · Inx
≤ E[
Xs=βx/4·ln(t)
Xτ =1
Xt=1
Xt=1
Ixt=x · Inx
≤ E[
Xs=βx/4·ln(t)
Xτ =1
Xt=1
∞
+
T
T
t
t
P[¯ck,τ < b +
ǫ
2
, nk,τ = s]
exp(−s·
ǫ2
2
)
k,t≥βx/2·ln(t) · Ibx,t≥nx,t·b+ǫβx/4·ln(t) · I¯ck,τq−ǫk,τq≥b, q=qt−ǫβx/4 ln(t),··· ,qt ]
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
Ixt=x · Inx
k,t≥βx/2·ln(t) · Ibx,t≥nx,t·b+ǫβx/4·ln(t) · I¯ck,τq−ǫk,τq≥b, q=qt−ǫβx/4 ln(t),··· ,qt ]
44
T
Xt=1
≤ E[
2π2
3ǫ2 ,
+
k,t − ǫβx/4 ln(t) ≥ βx/4 ln(t) for q = qt − ǫβx/4 ln(t),··· , qt and nk,τq ≥ nx
where (τq)q∈N denote the random times at which basis x is selected and, for a time t at which basis x
is selected, qt denotes the index q ∈ N such that τq = t. The first inequality is a consequence of nx
k,τq =
k,t − (qt − q) ≥ nx
nx
k,τq , which
implies ǫk,τq ≤ ǫ/2. We use the same bounds as in Fact 2 for the last two inequalities. Now observe that, for
any q ∈ {qt − ǫβx/4 ln(t),··· , qt}, we have ¯cl,τq − ǫl,τq ≤ b since ¯ck,τq − ǫk,τq ≥ b and since x is feasible
basis to (8) at time τq (by definition). This implies that, for any q ∈ {qt − ǫβx/4 ln(t),··· , qt}, arm l was
pulled at time τq by definition of the load balancing algorithm since the amount of resource consumed at any
round cannot be larger than 1 and bx,τq ≥ bx,t− (qt− q) ≥ nx,t· b + ǫβx/4· ln(t)− (qt− q) ≥ nx,t· b ≥ nx,τqb.
Hence, we get:
T
E[
k,t≥βx/2·ln(t) · Ibx,t≥nx,t·b+ǫβx/4·ln(t)]
l,t≥ǫβx/4·ln(t) · I¯cl,t−ǫl,t≤b] +
2π2
3ǫ2
T
≤
Inx
Xt=1
Ixt=x · Inx
Xt=1
≤ E[
Xt=1
P[¯cl,t ≤ b +
Xs=ǫβx/4·ln(t)
Xt=1
Xs=ǫβx/4·ln(t)
Xt=1
4π2
3ǫ2 .
≤
≤
≤
T
T
T
t
∞
ǫ
2
, nx
l,t ≥ ǫβx/4· ln(t)] +
P[¯cl,t ≤ b +
ǫ2
2
exp(−s·
ǫ
2
, nl,t = s] +
) +
2π2
3ǫ2
2π2
3ǫ2
2π2
3ǫ2
Bringing everything together, we derive:
E[nx,T ] ≤
26
ǫ3 · ln(T ) +
10π2
3ǫ2 .
E.3. Proof of Lemma 11. Without loss of generality, we can assume that x involves two arms (one of
which may be a dummy arm introduced in the specification of the algorithm given in Section 6) and that
Kx = {k, l} with µc
l . We say that a "swap" occurred at time τ if basis x was selected at time τ and
¯ck,τ − ǫk,τ ≤ b ≤ ¯cl,τ − ǫl,τ . We define nswap
as the total number of swaps that have occurred before time t,
x,t =Pt−1
i.e. nswap
τ =1 Ixτ =x · I¯ck,τ −ǫk,τ≤b≤¯cl,τ−ǫl,τ . Consider u ≥ 1 and define γx = (4/ǫ)2. First note that:
k > b > µc
x,t
t
P[nswap
x,t ≥ γx ln(t)] ≤
≤
+
t
Xq=γx ln(t)
Xq=γx ln(t)
Xq=γx ln(t)
t
P[¯ck,τq − ǫk,τq ≤ b ≤ ¯cl,τq − ǫl,τq ]
P[¯ck,τq ≤ b +
ǫ
2
γx
2
, nk,τq ≥
γx
2
ln(t)]
ln(t)]
P[b−
ǫ
2 ≤ ¯cl,τq, nl,τq ≥
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
45
exp(−s·
ǫ2
2
)
∞
Xs=γx/2·ln(t)
t
Xq=1
≤ 2
ǫ2 · t2 ,
≤
8
where (τq)q∈N are defined as the times at which basis x is selected. The first inequality is derived observing
that if nswap
x,t ≥ γx ln(t) then it must be that basis x was selected for the qth time, for some q ≥ γx ln(t), and
that we had ¯ck,τq − ǫk,τq ≤ b ≤ ¯cl,τq − ǫl,τq . To obtain the second inequality, we observe that, at any time τ ,
at least one of arm k and l must have been pulled nx,τ /2 times and that ǫk,τ ≤ ǫ/2 when nk,τ ≥ γx/2 ln(t)
(a similar inequality holds for arm l). The last two inequalities are derived in the same fashion as in Lemma
10. This yields:
P[bx,t − nx,t · b ≥ u + γx ln(t)]
≤ P[bx,t − nx,t · b ≥ u + γx ln(t) ; nswap
≤ P[bx,t − nx,t · b ≥ u + γx ln(t) ; nswap
x,t ≤ γx ln(t)] + P[nswap
x,t ≤ γx ln(t)] +
ǫ2 · t2 .
8
x,t ≥ γx ln(t)]
Note that, by definition of the load balancing algorithm, we are led to pull arm k (resp. arm l) at time τq if the
budget spent so far when selecting basis x, denoted by bx,τq , is below (resp. above) the "target" of nx,τq · b
assuming there is no "swap" at time τq (i.e. ¯ck,τq − ǫk,τq ≥ ¯cl,τq − ǫl,τq ). Hence, if bx,t− nx,t· b ≥ u + γx ln(t)
and nswap
x,t ≤ γx ln(t), we must have been pulling arm l for at least s ≥ ⌊u⌋ rounds t1 ≤ ··· ≤ ts ≤ t − 1
where basis x was selected since the last time, denoted by t0, where basis x was selected and the budget was
below the target, i.e. bx,t0 ≤ nx,t0 · b (because the amounts of resource consumed at each round are almost
surely bounded by 1). Moreover, we have:
t−1
Xτ =t0+1
Ixτ =x·(ck,τ · I¯ck,τ −ǫk,τ ≥¯cl,τ−ǫl,τ + cl,τ · I¯ck,τ −ǫk,τ <¯cl,τ−ǫl,τ )
= bx,t − bx,t0+1
≥ (nx,t − nx,t0)· b + u− 1 + γx ln(t)
≥ s· b + u− 1 + γx ln(t).
This implies:
s
Xq=1
τ =t0+1 Ixτ =x · I¯ck,τ −ǫk,τ <b · ck,τ ≤ nswap
x,t ≤ γx ln(t). Hence, if u ≥ 1:
x,t ≤ γx ln(t)]
P[bx,t − nx,t · b ≥ u + γx ln(t) ; nswap
cl,tq ≥ s· b + u− 1
s
t
sincePt−1
cl,tq ≥ s· b + u− 1]
cl,tq ≥ s· µc
l + s· (b− µc
l )]
cl,tq ≥ s· µc
l + s· ǫ]
≤
=
≤
≤
≤
≤
t
t
t
s
s
P[
P[
P[
Xq=1
Xq=1
Xq=1
exp(−2ǫ2 · s)
Xs=⌊u⌋
Xs=⌊u⌋
Xs=⌊u⌋
Xs=⌊u⌋
exp(−2ǫ2 ·⌊u⌋)
1− exp(−2ǫ2)
2
ǫ2 · exp(−ǫ2 · u),
46
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
where we use Lemma 1 for the third inequality and the fact that exp(−2v) ≤ 1 − v/2 for v ∈ [0, 1] for the
last inequality. With a similar argument, we conclude:
P[bx,t − nx,t· b ≥ u + γx ln(t) ; nswap
x,t ≤ γx ln(t)] ≤
4
ǫ2 · exp(−ǫ2 · u).
This last result enables us to show that:
0
E[bx,T ]− E[nx,T ]· b
≤ E[bx,T − nx,T · b]
=Z T
≤Z T
≤Z T
ǫ2 ·Z T
≤
13
ǫ4 + (
=
4
ǫ
0
4
0
0
P[bx,T − nx,T · b ≥ u]du
P[bx,T − nx,T · b ≥ u ; nswap
P[bx,T − nx,T · b ≥ u + 1 + γx ln(T ) ; nswap
exp(−ǫ2 · u)du + 1 + γx ln(T ) +
)2 ln(T ).
8
ǫ2
x,T ≤ γx ln(T )]du + T · P[nswap
x,T ≥ γx ln(T )]
x,T ≤ γx ln(T )]du + 1 + γx ln(T ) +
8
ǫ2
We get:
which, in combination with E[nx
E[nx
k,T ]· µc
k + E[nx
l,T ]· µc
k,T ] + E[nx
l = E[bx,T ] ≥ E[nx,T ]· b−
l,T ] = E[nx,T ], shows that:
13
ǫ4 − (
4
ǫ
)2 ln(T ),
E[nx
l
b− µc
k,T ] ≥ (
µc
k − µc
≥ ξx
k · E[nx,T ]−
13
)· E[nx,T ]−
ǫ4 · (µc
k − µc
13
16
ǫ5 −
ǫ3 ln(T ).
l
l ) −
42
ǫ2 · (µc
k − µc
l )
ln(T )
Symmetrically, we get:
E[nx
l,T ] ≥ ξx
l · E[nx,T ]−
13
ǫ5 −
16
ǫ3 ln(T ).
28
∆x
E.4. Proof of Lemma 12. Consider any suboptimal basis x ∈ B. We use the shorthand notation βx =
ǫ3 ·( λ
)2. Without loss of generality, we can assume that both x and x∗ involve two arms (one of which may
be a dummy arm introduced in the specification of the algorithm given in Section 6) and that Kx∗ = {k∗, l∗}
l∗ and Kx = {k, l} with µc
with µc
l . The proof is along the same lines as for Lemmas 5
and 8. We break down the analysis in a series of facts where we emphasize the main differences. We start
off with an inequality analogous to Fact 1.
k∗ > b > µc
k > b > µc
T
Ixt=x · Inx,t≥βx ln(t)]
Ixt=x · Inx,t≥βx ln(t) · I¯ck∗,t−ǫk∗,t≤µc
k∗ · I¯cl∗ ,t−ǫl∗,t≤µc
l∗ ]
T
Xt=1
E[nx,T ] ≤ 2βx · ln(T ) + E[
Xt=1
≤ 2βx · ln(T ) + E[
Xt=1
Xt=1
≤ 2βx · ln(T ) + E[
P[¯cl∗,t > µc
+
T
T
l∗ + ǫl∗,t] + P[¯ck∗,t > µc
k∗ + ǫk∗,t]
Ixt=x · Inx,t≥βx ln(t) · I¯ck∗,t−ǫk∗,t≤µc
k∗ · I¯cl∗ ,t−ǫl∗,t≤µc
l∗ ] +
π2
3
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
T
Ixt=x · Iobjx,t+Ex,t≥Pk∈{k∗,l∗}(¯rk,t+λǫk,t)·ξx∗
k · Inx,t≥βx ln(t)] +
47
π2
3
T
T
Xt=1
≤ 2βx · ln(T ) + E[
Xt=1
≤ 2βx · ln(T ) + E[
Xt=1
l∗ − ǫl∗,t] + P[¯rk∗ ,t < µr
Xt=1
≤ 2βx · ln(T ) + E[
P[¯rl∗,t < µr
+
T
Ixt=x · Iobjx,t+Ex,t≥objx∗ · Inx,t≥βx ln(t)]
k∗ − ǫk∗,t] +
π2
3
Ixt=x · Iobjx,t+Ex,t≥objx∗ · Inx,t≥βx ln(t)] +
2π2
3
.
The first inequality is derived in the same fashion as in Fact 1 substituting k with x. The third and last
inequalities are obtained using Lemma 1 in the same fashion as in Fact 3. The fourth inequality is obtained
by observing that (i) if xt = x then xt is optimal for (8) and (ii) (ξ∗k)k=1,··· ,K is feasible for (8) if ¯cl∗,t− ǫl∗,t ≤
l∗ and ¯ck∗,t − ǫk∗,t ≤ µc
µc
k . The
second term in the last upper bound can be broken down in two terms similarly as in Lemmas 5 and 8:
k∗ . The fifth inequality results from λ ≥ 1 and objx∗ =Pk∈{k∗,l∗}
k · ξx∗
µr
T
E[
T
Xt=1
Ixt=x · Iobjx,t+Ex,t≥objx∗ · Inx,t≥βx ln(t)]
Xt=1
≤ E[
Xt=1
Ixt∈Bt · Iobjx,t≥objx+Ex,t · Inx,t≥βx ln(t)]
Ixt∈Bt · Iobjx∗≤objx+2Ex,t · Inx,t≥βx ln(t)].
+ E[
T
(34)
(35)
We carefully study each term separately.
Fact 9
E[
T
Xt=1
Ix∈Bt · Iobjx∗≤objx+2Ex,t · Inx,t≥βx ln(t)] ≤
6π2
ǫ2 .
Proof. Using the shorthand notations αx = 8( λ
∆x
)2 and γx = ( 4
ǫ )2, we have:
T
E[
T
I∆x≤2λ·max(ǫk,t,ǫl,t) · Inx,t≥βx ln(t)]
Xt=1
Ix∈Bt · Iobjx∗≤objx+2Ex,t · Inx,t≥βx ln(t)]
Xt=1
≤ E[
Xt=1
Xt=1
P[nl,t ≤ αx ln(t) ; nx,t ≥ βx ln(t)] +
Imin(nk,t,nl,t)≤αx ln(t) · Inx,t≥βx ln(t)]
= E[
≤
T
T
T
Xt=1
k,t · ǫk,t + ξx
P[nk,t ≤ αx ln(t) ; nx,t ≥ βx ln(t)].
The first inequality is derived using Ex,t = λ · (ξx
l,t ≤ 1 (this is imposed as
a constraint in (8)). Observe now that αx/βx is a constant factor independent of ∆x. Thus, we just have
to show that if x has been selected at least βx ln(t) times, then both k and l have been pulled at least a
l,t · ǫl,t) and ξx
k,t + ξx
48
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
constant fraction of the time with high probability. This is the only time the load balancing algorithm comes
into play in the proof of Lemma 12. We study the first term and we conclude the study by symmetry. We have:
P[nl,t ≤ αx ln(t) ; nx,t ≥ βx ln(t)]
t
Ixτ =x · Iaτ =k · ck,τ ≥ (b + ǫ/2)· nx
k,t]
Xτ =1
t
t
t
+
P[
Xτ =1
Ixτ =x · Iaτ =k · ck,τ ≤ (b + ǫ/2)· s ; nx
≤ P[nl,t ≤ αx ln(t) ; nx,t ≥ βx ln(t) ;
Xs=4/ǫ2 ln(t)
≤ P[nl,t ≤ αx ln(t) ; nx,t ≥ βx ln(t) ; bx,t − nx,t · b ≥ ǫ/2· nx
Xs=4/ǫ2 ln(t)
+
≤ P[bx,t − nx,t · b ≥ 2γx ln(t)] + (
≤
ǫ2 · t2 .
exp(−s·
2
ǫ· t
ǫ2
2
16
)2
)
k,t = s]
k,t − nx
l,t]
The first inequality is obtained observing that if nl,t ≤ αx ln(t) and nx,t ≥ βx ln(t), we have:
derive the second inequality, we use Lemma 1 for the second term and remark that:
l,t ≥ (
k,t = nx,t − nx
nx
because λ ≥ 1 and ∆x ≤ objx∗ =PK
bx,t − nx,t · b ≥ (
Xτ =1
since b ≤ 1. The third inequality is derived using:
k=1 µr
k · ξx∗
t
)2 · ln(t) ≥
λ
∆x
k=1 ξx∗
28
ǫ3 − 8)· (
k ≤PK
Ixτ =x · Iaτ =k · ck,τ − nx
k ≤ 1 since x∗ is a feasible basis to (3). To
4
ǫ2 · ln(t)
k,t · b)− nx
l,t
ǫ/2· nx
k,t − nx
l,t ≥ ǫ/2· nx,t − 2· nx
l,t ≥ (
27
ǫ2 − 16)· ln(t) ≥ 2γx ln(t)
and the last inequality is obtained with Lemma 11.
Fact 10
T
E[
Xt=1
Proof. First observe that:
T
Ixt∈Bt · Iobjx,t≥objx+Ex,t · Inx,t≥βx ln(t)] ≤
3π2
ǫ2 .
E[
T
T
Ixt∈Bt · Iobjx,t≥objx+Ex,t · Ink,t≥βx/2·ln(t)]
Xt=1
Ixt∈Bt · Iobjx,t≥objx+Ex,t · Inx,t≥βx ln(t)]
Xt=1
≤ E[
Xt=1
Xt=1
≤ E[
Ixt∈Bt · Iobjx,t≥objx+Ex,t · I¯ck,t−ǫk,t>b]
Ixt∈Bt · Iobjx,t≥objx+Ex,t · Inl,t≥βx/2·ln(t)]
+ E[
T
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
T
49
t
t
T
T
T
T
+
+
+ E[
P[¯ck,t ≤ b + ǫ/2, nk,t = s] + P[¯cl,t ≥ b + ǫ/2, nl,t = s]
P[¯ck,t − ǫk,t ≤ b, nk,t = s] + P[¯cl,t − ǫl,t ≥ b, nl,t = s]
Iobjx,t≥objx+Ex,t · I¯ck,t−ǫk,t≥b≥¯cl,t−ǫl,t · I¯ck,t−ǫk,t>¯cl,t−ǫl,t]
Xt=1
Ixt∈Bt · Iobjx,t≥objx+Ex,t · I¯cl,t−ǫl,t<b]
Xs=βx/2·ln(t)
Xt=1
Xt=1
≤ 2· E[
Xs=βx/2·ln(t)
Xt=1
Xt=1
Iobjx,t≥objx+Ex,t · I¯ck,t−ǫk,t≥b≥¯cl,t−ǫl,t · I¯ck,t−ǫk,t>¯cl,t−ǫl,t]
≤ 2· E[
Xs=βx/2·ln(t)
Xt=1
Xt=1
Iobjx,t≥objx+Ex,t · I¯ck,t−ǫk,t≥b≥¯cl,t−ǫl,t · I¯ck,t−ǫk,t>¯cl,t−ǫl,t] +
≤ 2· E[
k· ξx∗
k ≤PK
2 for nk,t, nl,t ≥ βx
exp(−s
+ 2·
ǫ2
2
)
T
T
t
4π2
3ǫ2 .
k=1 µr
k=1 ξx∗
The third inequality is obtained by observing that ǫk,t, ǫl,t ≤ ǫ
2 ln(t) (because λ ≥ 1 and
∆x ≤ objx∗ =PK
k ≤ 1) and that, if xt ∈ Bt and (for example) ¯cl,t− ǫl,t < b, it must be
that ¯ck,t− ǫk,t ≥ b. The last two inequalities are obtained in the same fashion as in Lemma 10 observing that
βx/2 ≥ 4/ǫ2. At this point, the key observation is that if objx,t ≥ objx + Ex,t, ¯ck,t − ǫk,t ≥ b ≥ ¯cl,t − ǫl,t, and
¯ck,t − ǫk,t > ¯cl,t − ǫl,t, at least one of the following six events occurs: {¯rk,t ≥ µr
l + ǫl,t},
{¯ck,t ≤ µc
l + ǫl,t}. Otherwise, using the shorthand
notations ck = ¯ck,t − ǫk,t and cl = ¯cl,t − ǫl,t, we have:
l − ǫl,t} or {¯cl,t ≥ µc
k − ǫk,t}, {¯ck,t ≥ µc
k + ǫk,t}, {¯rl,t ≥ µr
k + ǫk,t}, {¯cl,t ≤ µc
objx,t − objx
= [
< [
=
=
l
l +
l + ǫl,t) +
k − µr
(µr
l · µr
k]
l · µr
l +
b− µc
µc
k − µc
µc
k − b
k − µc
µc
µc
k − b
b− cl
ck − b
l · µr
ck − cl · ¯rk,t]− [
ck − cl · ¯rl,t +
µc
k − µc
ck − b
b− cl
ck − cl · (µr
ck − cl · (µr
k + ǫk,t)]− [
b− µc
b− cl
1
λ · Ex,t + (µr
l )· [
ck − cl −
µc
k − µc
k − µr
1
l )
l ) · [(µc
k − b)(µc
l − cl) + (b− µc
λ · Ex,t +
k − µc
(ck − cl)· (µc
1
ck − cl ·(µc
k − b)(µc
l − cl) + (b− µc
l )(µc
k − ck)
λ · Ex,t +
1
λ · Ex,t + 2
ck − cl · [(ck − b)· ǫl,t + (b− cl)· ǫk,t]
1
2κ
λ · Ex,t
λ · Ex,t +
κ
κ
]
l
l
≤
≤
=
= Ex,t,
l
b− µc
k − µc
µc
l · µr
k]
l )(µc
k − ck)]
a contradiction. The first inequality is strict because either ck > b or cl < b. The second inequality is derived
using Assumption 5. The third inequality is derived from the observation that the expression (µc
l −
cl) + (b − µc
l cancels out) so that
k − b)(µc
(µc
l ) and the maximum of this expression over the
polyhedron [¯ck,t − ǫk,t, ¯ck,t + ǫk,t]× [¯cl,t − ǫl,t, ¯cl,t + ǫl,t] is attained at an extreme point. We obtain:
k, µc
k − ck) is convex in (µc
l ) (since the cross term µc
k, µc
k − ck) is a linear function of (µc
l )(µc
l − cl) + (b − µc
k − b)(µc
k · µc
l )(µc
T
E[
Xt=1
Ix∈BtIobjx,t≥objx+Ex,t · Inx,t≥βx ln(t)]
50
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
∞
+
∞
≤
Xt=1
Xt=1
Xt=1
≤ π2 +
3π2
≤
ǫ2 ,
using the same argument as in Fact 3.
∞
+
P[¯rk,t ≥ µr
k + ǫk,t] + P[¯rl,t ≥ µr
l + ǫl,t]
P[¯cl,t ≥ µc
l + ǫl,t] + P[¯ck,t ≥ µc
k + ǫk,t]
k − ǫk,t] + P[¯cl,t ≤ µc
l − ǫl,t] +
4π2
3ǫ2
P[¯ck,t ≤ µc
4π2
3ǫ2
E.5. Proof of Theorem 5. We build upon (4):
K
τ ∗
RB,T ≤ T ·
= T ·
≤ T ·
≤ T ·
K
K
Xk=1
Xk=1
Xk=1
Xk=1
K
k · ξx∗
µr
k · ξx∗
µr
k · ξx∗
µr
k · ξx∗
µr
T
Xt=1
k − E[
Xt=1
k − E[
Xt=1
k − E[
Xt=1
k − E[
T
T
rat,t] + O(1)
T
T
rat,t] + E[
Xt=τ ∗+1
Xt=τ ∗+1
rat,t] + σ · E[
Xt=1
rat,t] + σ · E[(
T
rat,t] + O(1)
cat,t] + O(1)
cat,t − B)+] + O(1).
The second inequality is a consequence of Assumption 4:
Iτ ∗≥t · E[cat,t Ft−1]]
T
E[
Xt=τ ∗+1
cat,t] = E[
T
τ ∗
Xt=1
cat,t]− E[
cat,t]
∞
Xt=1
E[cat,t Ft−1]]− E[
Iτ ∗≥t · µc
at ]
Xt=1
µc
at]− E[
∞
= E[
= E[
= E[
T
T
Xt=1
Xt=1
Xt=1
Xt=τ ∗+1
1
Xt=τ ∗+1
σ · E[
T
T
µc
at]
≥
µr
at] =
1
σ · E[
T
Xt=τ ∗+1
rat,t],
since τ ∗ is a stopping time. To derive the third inequality, observe that if τ ∗ = T + 1, we have:
T
T
Xt=τ ∗+1
cat,t = 0 ≤ (
cat,t − B)+,
while if τ ∗ < T + 1 we have run out of resources before round T , i.e.Pτ ∗
cat,t − B
cat,t ≤
cat,t +
Xt=1
τ ∗
T
T
Xt=τ ∗+1
Xt=τ ∗+1
Xt=1
t=1 cat,t ≥ B, which implies:
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
51
Now observe that:
E[(
T
Xt=1
≤ (
T
Xt=1
cat,t − B)+.
cat,t − B)+] = E[( Xx pseudo-basis for (3)
{bx,T − nx,T · b})+]
≤Xx∈B
= O(
E[bx,T − nx,T · b] +Xx /∈B
K 2
ǫ3 ln(T )),
E[nx,T ] + Xx pseudo-basis for (3)
with det(Ax)=0
E[nx,T ]
where we use the fact that ck,t ≤ 1 at any time t and for all arms k for the first inequality and Lemma 10
along with the proof of Lemma 11 for the last equality. Plugging this last inequality back into the regret
bound yields:
RB,T ≤ T ·
T
K
K
K
(
K
K
K
K
=
≤ T ·
k · ξx∗
Xt=1
Xk=1
µr
k − E[
k · ξx∗
Xk=1
Xk=1
k −Xx∈B
µr
k · ξx∗
Xk=1
Xk=1
k −Xx∈B
µr
≤ T ·
k · ξx∗
Xk=1
k · (T − Xx∈B ∆x=0
µr
k · ξx∗
Xk=1
k · ( Xx∈B ∆x>0
µr
Xk=1
− Xx∈B ∆x>0
µr
k · ξx
≤ Xx∈B ∆x>0
∆x · E[nx,T ] + O(
≤ 29 λ2
ǫ3 · ( Xx∈B ∆x>0
1
∆x
=
K
(
rat,t] + O(
K 2 · σ
ǫ3
ln(T ))
k · E[nx
µr
k,T ] + O(
K 2 · σ
ǫ3
ln(T ))
k · ξx
µr
k )· E[nx,T ] + O(
E[nx,T ])− Xx∈B ∆x>0
ln(T ))
K 2 · σ
ǫ3
K
(
Xk=1
k · ξx
µr
E[nx,T ] + Xx pseudo-basis for (3)
K 2 · σ
ǫ3
with det(Ax)=0
ln(T ))
E[nx,T ] +Xx /∈B
k )· E[nx,T ] + O(
K 2 · σ
ǫ3
ln(T ))
)· ln(T ) + O(
K 2 · σ
ǫ3
ln(T )),
k )· E[nx,T ] + O(
E[nx,T ])
K 2 · σ
ǫ3
ln(T ))
(36)
where we use Lemma 11 for the third inequality, Lemma 10 along withPK
the fourth inequality, and Lemma 12 for the last inequality.
k=1 µr
k · ξx∗
k ≤PK
k=1 ξx∗
k ≤ 1 for
E.6. Proof of Theorem 6. Along the same lines as for the case of a single limited resource, we start
from inequality (36) derived in the proof of Theorem 5 and apply Lemma 12 only if ∆x is big enough,
taking into account the fact that:
Specifically, we have:
RB,T ≤ sup
(nx)x∈B≥0
Px∈B nx≤T
{ Xx∈B ∆x>0
E[nx,T ] ≤ T.
Xx∈B
min(∆x · nx, 29 λ2
ǫ3 ·
ln(T )
∆x
+
10π2
3ǫ2 · ∆x) } + O(
K 2 · σ
ǫ3
ln(T ))
min(∆x · nx, 29 λ2
ǫ3 ·
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
ln(T )
∆x
) } + O(
K 2 · σ
ǫ3
ln(T ))
52
≤ sup
(nx)x∈B≥0
Px∈B nx≤T
{ Xx∈B ∆x>0
{Xx∈Br29 λ2
ǫ3 · ln(T )· nx } + O(
K 2 · σ
ǫ3
ln(T ))
√nx } + O(
K 2 · σ
ǫ3
ln(T ))
(nx)x∈B≥0
Px∈B nx≤T
≤ sup
≤ 25 λ
≤ 25 λ
sup
(nx)x∈B≥0
Px∈B nx≤T
{Xx∈B
ǫ3/2 ·pln(T )·
K 2 · σ
ǫ3/2 ·pB· T · ln(T ) + O(
ǫ3
k · ξx∗
k ≤PK
k=1 µr
ln(T )),
where we use the fact that ∆x ≤PK
over each ∆x ≥ 0 to derive the third inequality, and we use Cauchy-Schwartz for the last inequality.
E.7. Proof of Theorem 7. When b ≤ ǫ/2, the analysis almost falls back to the case of a single limited
k ≤ 1 for the second inequality, we maximize
resource. Indeed, we have τ ∗ = τ (B) with high probability given that:
k=1 ξx∗
P[τ (B) > T + t] = P[
T +t
T +t
Xτ =1
1
T + t ·
caτ ,τ ≤ B]
Xt=1
≤ exp(−(T + t)·
≤ P[
ǫ2
2
),
cat,t ≤ ǫ− (ǫ− b)]
for any t ∈ N using Lemma 1. Now observe that, since b ≤ ǫ/2, the feasible bases for (3) are exactly the
bases x such that Kx = {k} and Cx = {1} for some k ∈ {1,··· , K}, which we denote by (xk)k=1,··· ,K.
This shows that EROPT(B, T ) = B· maxk=1,··· ,K µr
k. Moreover note that Assumption 6 is automatically
satisfied when b ≤ ǫ/2. Hence, with a minor modification of the proof of Lemma 10, we get:
k/µc
E[nx,τ (B)] ≤
212
ǫ3 · E[ln(τ (B))] +
40π2
3ǫ2 ,
for any pseudo-basis x involving two arms or any basis x such that Kx = {k} and Cx = {2} for some arm
k ∈ {1,··· , K}. Similarly, a minor modification of Lemma 12 yields:
E[nxk,τ (B)] ≤ 212 λ2
ǫ3 ·
E[ln(τ (B))]
∆2
xk
+
40π2
ǫ2
,
for any k ∈ {1,··· , K}. What is left is to refine the analysis of Theorem 5 as follows:
RB,T ≤ B · max
k=1,··· ,K
τ ∗
τ (B)
rat,t] + O(1)
µr
Xt=1
k
k − E[
µc
µr
Xt=1
k
k − E[
µc
µr
Xk=1
k · E[nxk
k
µr
k −
µc
k · (B − Xk ∆xk =0
k · E[nxk
µc
K
≤ B · max
k=1,··· ,K
≤ B · max
k=1,··· ,K
µr
k
≤ max
µc
k=1,··· ,K
k,τ (B)] + O(1)
rat,t] + T · P[τ (B) > T ] + E[(τ (B)− T )+] + O(1)
k,τ (B)])− Xk ∆xk >0
k · E[nxk
µr
k,τ (B)] + O(1).
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
53
K
By definition of τ (B), we have B ≤Pτ (B)
µc
k · E[nk,τ (B)]
k · E[nxk
µc
Xk=1
B ≤
= Xk ∆xk =0
t=1 cat,t. Taking expectations on both sides yields:
k,τ (B)] + Xk ∆xk >0
k · E[nxk
µc
k,τ (B)] +Xx /∈B
E[nx,τ (B)].
Plugging this inequality back into the regret bound, we get:
RB,T = Xk ∆xk >0
≤ Xk ∆xk >0
≤ 212 λ2
k · ∆xk · E[nxk
µc
k,τ (B)] + max
k · ∆xk · E[nxk
µc
k,τ (B)] +
µr
k
µc
k=1,··· ,K
212K 2 · κ
ǫ3
E[nx,τ (B)] + O(1)
k ·Xx /∈B
· E[ln(τ (B))] + O(1)
ǫ3 · ( Xk ∆xk >0
1
∆xk
)· ln(
B + 1
ǫ
) +
212K 2 · κ
ǫ3
· ln(
B + 1
ǫ
) + O(1),
where we use Assumption 5 for the second inequality and Lemma 4 for the third inequality. A distribution-
ǫ )) can be derived from the penul-
independent regret bound of order O(qK · B+1
timate inequality along the same lines as in Theorem 6.
ǫ ) + K2·κ
· ln( B+1
· ln( B+1
ǫ3
ǫ
Appendix F: Proofs for Section 7.
F.1. Preliminary work for the proofs of Section 7.
Concentration inequality. We will use the following inequality repeatedly. For a given round τ ≥ tini
and a basis x:
P[∃(k, i) ∈ Kx ×{1,··· , C},¯ck,τ (i)− µc
k(i) >
T
16· (C + 2)!2 ]
ǫ3
P[¯ck,τ (i)− µc
k(i) >
ǫ3
16· (C + 2)!2 , nk,τ = s]
i∈{1,··· ,C}
∞
≤
Xs=tini /K Xk∈Kx
Xs=tini /K
≤ 2· C 2 ·
≤ 29 (C + 3)!4
ǫ6 · T2
,
exp(−s·
ǫ6
27 · (C + 2)!4 )
(37)
using Lemma 1, the inequality exp(−x) ≤ 1−x/2 for x ∈ [0, 1], and the fact that we pull each arm tini /K ≥
28 (C+2)!4
· ln(T ) times during the initialization phase.
ǫ6
Useful matrix inequalities. For any basis x, assume that {¯ck,t(i) − µc
k(i) ≤
ǫ3
16·(C+2)!2}, for any arm
k ∈ Kx and resource i ∈ Cx. We have:
det( ¯Ax,t)− det(Ax) = Xσ∈S(Kx,Cx)
¯ck,t(σ(k))− Yk∈Kx
µc
k(σ(k))]
¯ck,t(σ(k))· [¯cl,t(σ(l))− µc
¯cl,t(σ(l))− µc
l (σ(l))
l (σ(l))]·Yk>l
µc
k(σ(k))
[ Yk∈Kx
= Xσ∈S(Kx,Cx)Xl∈KxYk<l
≤ Xσ∈S(Kx,Cx)Xl∈Kx
≤
16· (C + 2)!
ǫ
≤
,
2
ǫ3
54
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
since the amounts of resources consumed at any round are no larger than 1. This yields:
det( ¯Ax,t) ≥ det(Ax)− det( ¯Ax,t)− det(Ax) ≥
using Assumption 8. Now consider any vector c such that kck∞ ≤ 1, we have:
ǫ
2
,
(38)
cT ¯A−1
x,tbKx − cTA−1
x bKx
1
det(Ax)· det( ¯Ax,t)· det(Ax)· cT adj( ¯Ax,t)bKx − det( ¯Ax,t)· cT adj(Ax)bKx
=
det( ¯Ax,t)·cT(adj( ¯Ax,t)− adj(Ax))bKx
det(Ax)· det( ¯Ax,t)· det( ¯Ax,t)− det(Ax)·cT adj(Ax)bKx
≤
+
1
1
ǫ
+
8· (C + 2)! ·kck2 ·kadj(Ax)bKxk2
ǫ2
8
ǫ
.
4
≤
≤
The second inequality is obtained using Assumption 8 and (38) by proceeding along the same lines as above
to bound the difference between two determinants for each component of adj( ¯Ax,t) − adj(Ax). The last
inequality is obtained using kck2 ≤ √C and the fact that each component of Ax is smaller than 1. If we take
c = ek for k ∈ Kx, this yields:
(39)
(40)
(41)
x,tbKx − cTA−1
x bKx
x,tbKx − cTA−1
x,tbKx +cT ¯A−1
det( ¯Ax,t) ·(cid:13)(cid:13)adj( ¯Ax,t)bKx(cid:13)(cid:13)2 +
16· (C + 2)!2 ·
2
ǫ · C! +
x bKx
ǫ
4
ǫ
4
ǫ3
If we take c = (µc
and
Xk∈Kx
k ≤
ǫ
4
.
ξx
k,t − ξx
µc
k(i)· ξx
Xk∈Kx
k,t − Xk∈Kx
ǫ
k ≤
4
k(i))k∈Kx, for any i ∈ {1,··· , C}, we get:
k,t − Xk∈Kx
µc
k(i)· ξx
k = ¯cT ¯A−1
≤ (¯c− c)T ¯A−1
≤ k¯c− ck2 ·
√C ·
≤
≤
¯ck,t(i)· ξx
µc
k(i)· ξx
ǫ
2
1
,
where ¯c = (¯ck,t(i))k∈Kx.
F.2. Proof of Lemma 13. First, consider a basis x /∈ B. Since x is ǫ-non-degenerate by Assumption
8, there must exist k ∈ Kx such that ξx
k ≥ b(i) + ǫ. Let
us assume that we are in the first situation (the proof is symmetric in the other scenario). Using (39) in the
preliminary work of Section F.1, we have:
k ≤ −ǫ or i ∈ {1,··· , C} such that PK
k(i) · ξx
k=1 µc
ξx
k,t = ξx
k + (ξx
k,t − ξx
k ) ≤ −
ǫ
2
,
(42)
if {¯ck,t(i)− µc
k(i) ≤
ǫ3
16·(C+2)!2} for all arms k ∈ Kx and resources i ∈ {1,··· , C}. Hence:
T
E[nx,T ] = E[
Xt=tini
Ixt=x]
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
T
55
k,t ≥ 0]
T
≤
P[ξx
Xt=tini
Xt=tini Xk∈Kx
≤
≤ 29 (C + 3)!4
i∈{1,··· ,C}
,
ǫ6
P[¯ck,t(i)− µc
k(i) >
ǫ3
16· (C + 2)!2 ]
where the third inequality is derived with (37).
Second, consider a pseudo-basis x for (3) that is not a basis. Since det(Ax) = 0, either every component
of Ax is 0, in which case det( ¯Ax,t) = 0 at every round t and x can never be selected, or there exists a
basis x for (3) with Kx ⊂ Kx and Cx ⊂ Cx along with coefficients (αkl)k∈Kx−Kx,l∈Kx such that µc
k(i) =
Pl∈Kx
l (i) for any resource i ∈ Cx. Assuming we are in the second scenario and since x is ǫ-non-
degenerate by Assumption 8, we havePK
k ≤ b(j)− ǫ for any j /∈ Cx. Take i ∈ Cx −Cx. Suppose
that x is feasible for (8) at round t and assume by contradiction that {¯ck,t(j)− µc
16·(C+2)!2} for all
arms k ∈ Kx and resources j ∈ {1,··· , C}. Using the notations ξx
αkl · ξx
l,t, we have,
for any resource j ∈ Cx:
k(j) ≤
k,t +Pl∈Kx−Kx
k(j)· ξ x
αkl · µc
k,t = ξx
k=1 µc
ǫ3
b(j) = Xl∈Kx
l,t
¯cl,t(j)· ξx
= α(j) +Xl∈Kx
µc
l (j)· ξx
l (j)· ξx
= α(j) +Xl∈Kx
µc
l,t,
l,t
16·(C+2)!2 since ξk ∈ [0, 1] ∀k ∈ {1,··· , K} for any feasible solution to (8). We get
ǫ3
where α(j) ≤
Ax
ξx
Kx,t = bCx − αCx while Axξ x
Kx
= bCx . We derive:
K
µc
k(i)· ξ x
k +
ǫ3
16· (C + 2)!2
K
Xk=1
K
K
k,t −
Xk=1
k(i)· ξ x
µc
k
Xk=1
µc
k(i)· ξx
k,t −
k(i)· ( ξx
k,t − ξ x
µc
k ) +
(i)TA−1
x αCx
1
¯ck,t(i)· ξx
Xk=1
≤
≤ Xk∈Kx
+µc
Kx
+√C ·
=
ǫ
4
det(Ax) ·kadj(Ax)αCxk2
ǫ3
(C + 1)!
ǫ
·
16· (C + 2)!2
ǫ
4
ǫ
4
ǫ
4
ǫ
2
≤
≤
≤
+
.
Thus we obtain:
K
Xk=1
¯ck,t(i)· ξx
k,t ≤
K
Xk=1
µc
k(i)· ξ x
k +
ǫ
2 ≤ b(i)−
ǫ
2
< b(i),
56
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
a contradiction since this inequality must be binding by definition if x is selected at round t. We finally
conclude:
T
E[nx,T ] = E[
T
Ixt=x]
Xt=tini
Xt=tini Xk∈Kx
≤
≤ 29 (C + 3)!4
ǫ6
j∈{1,··· ,C}
.
P[¯ck,t(j)− µc
k(j) >
ǫ3
16· (C + 2)!2 ]
F.3. Proof of Lemma 14.
Proof of (12). Consider a resource i ∈ Cx and u ≥ 1. We study P[bx,t(i)− nx,t · b(i) ≥ u] but the same
technique can be used to bound P[bx,t(i) − nx,t · b(i) ≤ −u]. If bx,t(i) − nx,t · b(i) ≥ u, it must be that
ex
i,τ = −1 for at least s ≥ ⌊u⌋ rounds τ = t1 ≤ ··· ≤ ts ≤ t − 1 where x was selected at Step-Simplex since
the last time, denoted by t0 < t1, where x was selected at Step-Simplex and the budget was below the target,
i.e. bx,t0(i) ≤ nx,t0 · b(i) (because the amounts of resources consumed at each round are bounded by 1).
Moreover, we have:
s
Xq=1
Hence:
P[bx,t(i)− nx,t · b(i) ≥ u]
s
t
catq ,tq (i) =
t−1
Ixτ =x · caτ ,τ (i)
Xτ =t0+1
= bx,t(i)− bx,t0+1(i)
≥ (nx,t − nx,t0)· b(i) + u− 1
≥ s· b(i) + u− 1.
catq ,tq (i) ≥ s· b(i) + u− 1 ; ex
i,tq = −1 ∀q ∈ {1,··· , s}]
K
Xk=1
k(j) >
µc
k(i)· px
k,tq ≤ b(i)−
ǫ3
16· (C + 2)!2 ]
ǫ2
4· (C + 1)! ∀q ∈ {1,··· , s}]
catq ,tq (i) ≥ s· b(i) ;
P[¯ck,τ (j)− µc
t
s
T
+
P[
P[
≤
≤
Xq=1
Xs=⌊u⌋
Xq=1
Xs=⌊u⌋
Xτ =tini Xk∈Kx
Xs=⌊u⌋
≤ 16
(C + 1)!2
≤
∞
ǫ4
j∈{1,··· ,C}
exp(−2s· (
ǫ2
4· (C + 1)!
)2) + 29 (C + 3)!4
ǫ6
1
T
·
· exp(−u· (
ǫ2
4· (C + 1)!
)2) + 29 (C + 3)!4
ǫ6
1
T
·
.
The last inequality is obtained using exp(−x) ≤ 1− x/2 for x ∈ [0, 1]. The third inequality is derived using
Lemma 1 for the first term and (37) for the second term. The second inequality is obtained by observing
that if x was selected at time τ and ¯ck,τ (j) − µc
16·(C+2)!2 for any arm k ∈ Kx and resource j ∈
{1,··· , C}, then we must have δ∗x,τ ≥ ǫ2
4·(C+1)! and
any arm k ∈ Kx:
4·(C+1)! . Indeed, using (38) and (39), we have, for δ ≤ ǫ2
k(j) ≤
ǫ3
cT ¯A−1
x,τ (bCx + δ · ex
τ ) = ξx
k,τ + δ · cT ¯A−1
x,τ ex
τ
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
57
x,τ ex
τ
k,τ− δ ·cT ¯A−1
k − ξx
≥ ξx
k −ξx
ǫ
¯A−1
2 − δ ·(cid:13)(cid:13)
τ(cid:13)(cid:13)2
x,τ ex
≥
1
det( ¯Ax,τ ) ·(cid:13)(cid:13)adj( ¯Ax,τ )ex
τ(cid:13)(cid:13)2
≥
√C · C!
≥
≥ 0,
ǫ
2 − δ ·
ǫ
2δ
2 −
ǫ ·
where c = ek and since x is ǫ-non-degenerate by Assumption 8. Similarly, using (38) and (41), we have, for
δ ≤ ǫ2
x,τ (bCx + δ · ex
4·(C+1)! and any resource j /∈ Cx:
cT ¯A−1
τ ) = Xk∈Kx
≤ Xk∈Kx
≤ b(j)−
≤ b(j)−
≤ b(j),
τ
x,τ ex
¯ck,τ (j)· ξx
µc
k(j)· ξx
ǫ
+ δ ·
2
2δ
ǫ
ǫ · (C + 1)!
+
2
k,τ + δ · cT ¯A−1
k,τ + Xk∈Kx
√C ·(cid:13)(cid:13)
¯ck,τ (j)· ξx
τ(cid:13)(cid:13)2
x,τ ex
¯A−1
k,τ − Xk∈Kx
µc
k(j)· ξx
k,τ + δ ·cT ¯A−1
x,τ ex
τ
where c = (¯ck,τ (j))k∈Kx and since x is ǫ-non-degenerate by Assumption 8.
Proof of (14). First observe that, using (12), we have:
i∈Cx E[bx,T (i)]− E[nx,T ]· b(i)
max
i∈Cx bx,T (i)− nx,T · b(i)]
P[max
i∈Cx bx,T (i)− nx,T · b(i) ≥ u]du
P[bx,T (i)− nx,T · b(i) ≥ u]du
ǫ2
0
(C + 1)!2
exp(−u· (
·Z T
+ C + C · 29 (C + 3)!4
4· (C + 1)!
ǫ6
0
≤ E[max
=Z T
=Xi∈CxZ T
0
≤ 32C ·
= 29C ·
≤ 210C ·
(C + 1)!4
(C + 3)!4
.
ǫ4
ǫ8
ǫ8
)2)du + C + C · 29 (C + 3)!4
ǫ6
Now observe that, for any resource i ∈ Cx, we have E[bx,T (i)] =Pk∈Kx µc
E[nx
E[nx,T ] )k∈Kx , we get:
vector p = (
k,T ]
k(i)· E[nx
k,T ]. Hence, defining the
E[nx,T ]·kp− ξxk2 = E[nx,T ]·(cid:13)(cid:13)A−1
1
=
x Ax(p− ξx)(cid:13)(cid:13)2
√C · 210C ·
x (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2 ·kE[nx,T ]· Ax(p− ξx)k2
≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)A−1
det(Ax) ·adj(Ax)2 ·k(E[bx,T (i)])i∈Cx − (E[nx,T ]· b(i))i∈Cxk2
1
ǫ · (C + 1)!·
≤
≤ 210 (C + 3)!5
(C + 3)!4
ǫ9
ǫ8
,
58
using Assumption 8. Finally we obtain:
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
E[nx,T ]· ξx
k − E[nx
k,T ] ≤ E[nx,T ]·kp− ξxk2
≤ 210 (C + 3)!5
ǫ9
,
ǫ4
contradiction, suppose that:
· ln(T ) for all resources j ∈ Cx,
8·(C+2)! for all resources j ∈ {1,··· , C} and for all arms k ∈ Kx such that nx
for any arm k ∈ Kx.
Proof of (13). Consider a resource i /∈ Cx and assume that bx,t(i) − nx,t · b(i) ≥ 28 (C+3)!3
• bx,t(j)− nx,t · b(j) ≤ 16 (C+1)!2
• µc
26 (C+2)!2
basis x and pulling arm k, i.e. ck(j) = 1
nx
Observe that if bx,t(i)− nx,t · b(i) ≥ 28 (C+3)!3
ln(T ) times at Step-Simplex since tini, i.e. nx,t ≥ 28 (C+3)!3
and K2
for any j ∈ Cx:
k(j) − ck(j) ≤ ǫ2
k,t ≥
· ln(T ), where ck(j) denotes the empirical average amount of resource j consumed when selecting
· ln(T ), it must be that x has been selected at least 28 (C+3)!3
·
· ln(T ). We can partition Kx into two sets K1
x
k,t < 26 (C+2)!2
x. We get,
Ixτ =x · Iaτ =k · ck,τ (j).
· ln(T ) for all k ∈ K2
· ln(T ) for all k ∈ K1
k,t ·Pt−1
k,t ≥ 26 (C+2)!2
· ln(T ). By
ǫ6
x and nx
x such that nx
τ =tini
ǫ4
ǫ4
ǫ4
ǫ6
ǫ6
ǫ6
(C + 1)!2
16
ǫ4
· ln(T ) ≥ bx,t(j)− nx,t · b(j)
≥ nx,t · Xk∈K1x
≥ nx,t · Xk∈K1x
nx
k,t
ck(j)· pk − b(j)− Xk∈K2
ck(j)· pk − b(j)− 26C ·
x
(C + 2)!2
ǫ4
· ln(T ),
where pk =
nx
k,t
nx,t
1 and pk = 0 otherwise. Hence:
for k ∈ Kx
Xk∈Kx
µc
k(j)· pk − b(j) ≤ max
k∈K1x µc
k(j)− ck(j) + Xk∈K1x
ck(j)· pk − b(j)
ǫ2
(16 (C+1)!2
ǫ4 + 26C · (C+2)!2
ǫ4
)· ln(T )
nx,t
≤
≤
+
,
ǫ2
8· (C + 2)!
4· (C + 2)!
where we use the fact that PK
· ln(T ) for the last one. We get:
28 (C+3)!3
ǫ6
k=1 pk ≤ 1 and pk ≥ 0 for any arm k for the first inequality and nx,t ≥
1
x Ax(p− ξx)(cid:13)(cid:13)2
kp− ξxk2 =(cid:13)(cid:13)A−1
≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)A−1
x (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2 ·kAx(p− ξx)k2
det(Ax) ·adj(Ax)2 ·
√C ·
1
ǫ · (C + 1)!·
4·√C
≤
≤
≤
ǫ
,
ǫ2
4· (C + 2)!
√C ·
ǫ2
4· (C + 2)!
using Assumption 8. Hence:
28 (C + 3)!3
ǫ6
· ln(T ) ≤ bx,t(i)− nx,t · b(i)
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
59
nx
k,t
ck(i)· pk − b(i)) + Xk∈K2
ck(i)· pk − b(i)) + 26C ·
≤ nx,t · (Xk∈K1x
≤ nx,t · (Xk∈K1x
k(i))k∈Kx, this implies:
x
(C + 2)!2
ǫ4
· ln(T ).
Using the shorthand notation c = (µc
ck(i)· pk − b(i)
0 ≤ Xk∈K1x
≤ Xk∈K1x
≤ cT(p− ξx)−
√C ·kp− ξxk2 −
≤
< 0,
µc
k(i)· pk +
ǫ
2
ǫ
2
ǫ2
8· (C + 2)! − (Xk∈Kx
µc
k(i)· ξx
k + ǫ)
k=1 pk ≤ 1 and pk ≥ 0 for any arm k and Assumption 8 for
· ln(T )]
(C + 1)!2
ǫ4
ǫ6
the second inequality. We conclude that:
a contradiction. Note that we use the fact thatPK
P[bx,t(i)− nx,t · b(i) ≥ 28 (C + 3)!3
≤Xj∈Cx
P[bx,t(j)− nx,t · b(j) ≥ 16
+ Xk∈Kx
≤ 25C ·
+ Xk∈Kx
≤ 2C · T · 29 (C + 3)!4
≤ 210 (C + 4)!4
ǫ6 · T
(C + 1)!2
ǫ4 · T
Xs=26·(C+2)!2/ǫ4·ln(T )
ǫ6 · T2 +28C 2 ·
k(j)− ck(j) ≥
+ C · T · 29 (C + 3)!4
ǫ6 · T2
P[µc
j∈{1,··· ,C}
j∈{1,··· ,C}
P[µc
ǫ2
,
(C + 2)!2
ǫ4 · T 2
· ln(T )]
k,t ≥ 26 (C + 2)!2
ǫ4
; nx
· ln(T )]
8· (C + 2)!
k(j)− ck(j) ≥
ǫ2
8· (C + 2)!
; nx
k,t = s]
where we use (12) for the second inequality and Lemma 1 for the third inequality.
F.4. Proof of Lemma 15. Consider any suboptimal basis x ∈ B. The proof is along the same lines
as for Lemmas 5, 8, and 12. We break down the analysis in a series of facts where we emphasize the
main differences. We start off with an inequality similar to Fact 1. We use the shorthand notation βx =
210 (C+3)!3
)2.
ǫ6
· ( λ
∆x
Fact 11
E[nx,T ] ≤2βx · ln(T ) + 29 (C + 3)!4
ǫ6
T
+ E[
Xt=tini
Ixt=x · Inx,t≥βx ln(t) · Ix∗∈Bt].
(43)
60
Proof. Similarly as in Fact 1, we have:
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
T
E[nx,T ] ≤2βx · ln(T ) + E[
Xt=tini
Ixt=x · Inx,t≥βx ln(t)].
This yields:
T
T
E[nx,T ] ≤ 2βx · ln(T ) + E[
Xt=tini
Ixt=x · Inx,t≥βx ln(t) · Ix∗∈Bt] + E[
Xt=tini
Ix∗ /∈Bt].
Using (39), (41), and Assumption 8, we have:
ξx∗
k,t = ξx∗
k − (ξx∗
k − ξx∗
k,t) ≥
ǫ
2 ≥ 0,
¯ck,t(i)· ξx∗
for any k ∈ Kx∗ and
Xk∈Kx∗
k,t = Xk∈Kx∗
≤ b(i)− ǫ +
ǫ
≤ b(i)−
2
≤ b(i),
for any resource i /∈ Cx∗ if {¯cl,t(j)− µc
l (j) ≤
ǫ
2
Hence:
k(i)· ξx∗
µc
k + ( Xk∈Kx∗
¯ck,t(i)· ξx∗
k,t − Xk∈Kx∗
k(i)· ξx∗
µc
k )
ǫ3
16·(C+2)!2} for any arm l ∈ Kx∗ and resource j ∈ {1,··· , C}.
E[nx,T ] ≤ 2βx · ln(T ) + E[
Ixt=x · Inx,t≥βx ln(t) · Ix∗∈Bt]
T
Xt=tini
P[¯cl,t(j)− µc
T
+
Xt=tini Xl∈Kx∗
j∈{1,··· ,C}
≤ 2βx · ln(T ) + E[
T
Xt=tini
l (j) >
ǫ3
16· (C + 2)!2 ]
Ixt=x · Inx,t≥βx ln(t) · Ix∗∈Bt] + 29 (C + 3)!4
ǫ6
,
where we bound the third term appearing in the right-hand side using (37).
The remainder of this proof is dedicated to show that the last term in (43) can be bounded by a constant.
This term can be broken down in three terms similarly as in Lemmas 5 and 8.
T
Iobjx,t+Ex,t≥objx∗,t+Ex∗,t · Inx,t≥βx ln(t) · Ix∈Bt,x∗∈Bt]
E[
Xt=tini
≤ E[
≤ E[
+ E[
+ E[
T
T
Iobjx,t≥objx+Ex,t · Ix∈Bt]
Ixt=x · Inx,t≥βx ln(t) · Ix∗∈Bt]
Xt=tini
Xt=tini
Xt=tini
Xt=tini
T
T
Iobjx∗,t≤objx∗−Ex∗,t · Ix∗∈Bt]
Iobjx∗ <objx+2Ex,t · Ix∈Bt · Inx,t≥βx ln(t)].
(44)
(45)
(46)
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
Fact 12
61
E[
T
Xt=tini
Iobjx∗ <objx+2Ex,t · Ix∈Bt · Inx,t≥βx ln(t)] ≤ 211 (C + 4)!4
ǫ6
.
Proof. Using the shorthand notation αx = 8( λ
∆x
)2, we have:
T
E[
Xt=tini
Iobjx∗ <objx+2Ex,t · Ix∈Bt · Inx,t≥βx ln(t)]
T
T
≤ E[
Xt=tini
Xt=tini
≤ E[
Xt=tini Xk∈Kx
≤
T
I∆x<2λ·maxk∈Kx ǫk,t · Inx,t≥βx ln(t)]
Imink∈Kx nk,t≤αx ln(t) · Inx,t≥βx·ln(t)]
P[nk,t ≤ αx ln(t) ; nx,t ≥ βx ln(t)],
ǫ4
l=1 ξx
l,t ≤ 1 and ξx
· ln(t) for any resource i ∈ Cx and that µc
since PK
l,t ≥ 0 for any arm l when x is feasible for (8) at time t. Consider k ∈ Kx and
assume that nk,t ≤ αx · ln(t) and nx,t ≥ βx · ln(t). Suppose, by contradiction, that bx,t(i) − nx,t · b(i) ≤
32 (C+1)!2
8·(C+2)! for any resource i ∈ {1,··· , C}
for all arms l ∈ Kx such that nx
· ln(t), where cl(i) is the empirical average amount of resource
l,t ·Pt−1
i consumed when selecting basis x and pulling arm l, i.e. cl(i) = 1
Ixτ =x · Iaτ =l · cl,τ (i). We
nx
can partition Kx − {k} into two sets K1
l,t <
27 (C+2)!2
x and nx
x. Similarly as in the proof of Lemma 14, we have, for any resource i ∈ Cx:
l (i) − cl(i) ≤ ǫ2
· ln(t) for all l ∈ K2
· ln(t) for all l ∈ K1
l,t ≥ 27 (C+2)!2
l,t ≥ 27 (C+2)!2
x such that nx
x and K2
τ =tini
ǫ4
ǫ4
ǫ4
32
(C + 1)!2
ǫ4
· ln(t) ≥ bx,t(i)− nx,t · b(i)
≥ nx,t ·Xl∈K1x
≥ nx,t ·Xl∈K1x
cl(i)· pl − b(i)− nk,t −Xl∈K2
cl(i)· pl − b(i)− αx · ln(t)− C ·
nx
l,t
x
27 · (C + 2)!2
ǫ4
· ln(t),
where pl =
nx
l,t
nx,t
1 and pl = 0 otherwise. Hence:
for l ∈ Kx
Xl∈Kx
µc
l (i)· pl − b(i) ≤ max
l∈K1x µc
l (i)− cl(i) +Xl∈K1x
cl(i)· pl − b(i)
ǫ2
(αx + C · 27·(C+2)!2
)· ln(t)
+
ǫ4
nx,t
≤
≤
ǫ2
8· (C + 2)!
4· (C + 2)!
To derive the first inequality, we use the fact that PK
inequality, we use nx,t ≥ 210 (C+3)!3
)2 · ln(t) along with λ ≥ 1 and ∆x ≤ objx∗ =PK
k=1 ξx∗
PK
k ≤ 1 because of the time constraint imposed in (3). We get:
l=1 pl ≤ 1 and pl ≥ 0 for any arm l. For the last
k ≤
k · ξx∗
k=1 µr
· ( λ
∆x
ǫ6
.
ξx
k ≤ kp− ξxk2
62
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
1
=(cid:13)(cid:13)A−1
x Ax(p− ξx)(cid:13)(cid:13)2
≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)A−1
x (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2 ·kAx(p− ξx)k2
det(Ax) ·adj(Ax)2 ·
√C ·
1
ǫ · (C + 1)!·
ǫ
,
2
≤
≤
≤
ǫ2
4· (C + 2)!
√C ·
ǫ2
4· (C + 2)!
a contradiction since x is ǫ-non-degenerate by Assumption 8. We conclude that:
T
E[
Xt=tini
Iobjx∗ <objx+2Ex,t · Ix∈Bt · Inx,t≥βx·ln(t)]
T
≤ C ·
+ C ·
T
Xt=tiniXi∈Cx
Xt=tini Xl∈Kx
i∈Cx
P[bx,t(i)− nx,t · b(i) ≥ 32
ǫ2
P[µc
l (i)− cl(i) ≥
8· (C + 2)!
π2
6
·
(C + 1)!2
ǫ4
≤ 32C 2 ·
+Xl∈Kx
≤ 4C · 29 (C + 3)!4
≤ 211 (C + 4)!4
Xs=27·(C+2)!2/ǫ4·ln(T )
+27C 2 ·
j∈Cx
ǫ6
ǫ6
,
+ C · 29 (C + 3)!4
P[µc
l (j)− cl(j) ≥
ǫ6
(C + 2)!2
ǫ4
π2
6
·
(C + 1)!2
ǫ4
· ln(t)]
; nx
l,t ≥ 27 (C + 2)!2
ǫ4
· ln(t)]
ǫ2
8· (C + 2)!
; nx
l,T = s]
where we use (12) for the second inequality and Lemma 1 for the third inequality.
Fact 13
Proof. First observe that:
T
E[
Xt=tini
Iobjx,t≥objx+Ex,t · Ix∈Bt] ≤ 210 (C + 3)!2
ǫ6
.
T
T
E[
Xt=tini
Iobjx,t≥objx+Ex,t · Ix∈Bt] ≤ E[
Iobjx,t≥objx+Ex,t · Ix∈Bt · I det( ¯Ax,t)≥ǫ/2]
T
T
+ E[
Xt=tini
Xt=tini
Xt=tini
≤ E[
Xt=tini Xk∈Kx
Xt=tini
≤ E[
+ 29 (C + 3)!4
+
T
T
ǫ6
,
I det( ¯Ax,t)<ǫ/2]
Iobjx,t≥objx+Ex,t · Ix∈Bt · I det( ¯Ax,t)≥ǫ/2]
P[¯ck,t(i)− µc
k(i) >
ǫ3
16· (C + 2)!2 ]
i∈Cx
Iobjx,t≥objx+Ex,t · Ix∈Bt · I det( ¯Ax,t)≥ǫ/2]
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
63
where we use the preliminary work of Section F.1 and in particular (38). The key observation now is that if
k + ǫk,t},
objx,t ≥ objx + Ex,t, x ∈ Bt, and det( ¯Ax,t) ≥ ǫ/2, at least one of the following events {¯rk,t ≥ µr
for k ∈ Kx, or {¯ck,t(i)− µc
k(i) ≥ ǫk,t}, for k ∈ Kx and i ∈ {1,··· , C}, occurs. Otherwise we have:
objx,t − objx = ( ¯A−1
< ( ¯A−1
x,tbCx)T¯rKx,t − (A−1
x,tbCx)T(µr
x bCx)Tµr
Kx
Kx + ǫKx,t)− (A−1
x bCx)Tµr
Kx
x,t − A−1
x )bCx)Tµr
Kx,
1
λ · Ex,t + (( ¯A−1
=
where the first inequality is a consequence of the fact that x is feasible for (8), i.e. ¯A−1
x,tbCx ≥ 0 with at least
one non-zero coordinate since ¯ck,t(i) ≥ ǫ for all arms k and resource i by Assumption 8. Writing µc
k(i) =
¯ck,t(i) + ui,k · ǫk,t, with ui,k ∈ [−1, 1] for all (i, k) ∈ Cx ×Kx, and defining the matrix U = (ui,k)(i,k)∈Cx×Kx,
we get:
objx,t − objx
<
=
≤
=
Ex,t
λ
Ex,t
λ
Ex,t
λ
Ex,t
λ
Ex,t
λ
Ex,t
λ
Ex,t
λ
Ex,t
λ
≤
≤
≤
≤
= Ex,t.
+( ¯A−1
+( ¯A−1
+
x,t − ( ¯Ax,t + U diag(ǫKx,t))−1)bCx)Tµr
Kx
x,tU (I + diag(ǫKx,t) ¯A−1
x,tU )−1 diag(ǫKx,t) ¯A−1
x,tbCx)Tµr
Kx
x,tU )−1 diag(ǫKx,t) ¯A−1
x,tU (I + diag(ǫKx,t) ¯A−1
x,tU ) diag(ǫKx,t) ¯A−1
x,tbCx)Tµr
1
ǫ · det( ¯Ax,t + U diag(ǫKx,t))·( ¯A−1
1
ǫ ·(adj( ¯Ax,t)U adj(I + diag(ǫKx,t) ¯A−1
1
ǫ ·(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)adj( ¯Ax,t)U adj(I + diag(ǫKx,t) ¯A−1
x,tU )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2 ·(cid:13)(cid:13)diag(ǫKx,t) ¯A−1
Kx(cid:13)(cid:13)2
x,tbCx(cid:13)(cid:13)2 ·(cid:13)(cid:13)µr
√C
ǫ ·(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)adj( ¯Ax,t)U(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2 ·(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)adj(I + diag(ǫKx,t) ¯A−1
ǫ · (C + 1)!·(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)adj(I + diag(ǫKx,t) ¯A−1
x,tU )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2 ·
x,tU )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2 ·
Ex,t
λ ·
Ex,t
λ
2
ǫ · (C + 1)!2 ·
Ex,t
λ
Kx
1
1
+
+
+
+
+
x,tbCx)Tµr
Kx
We use the Woodbury matrix identity to derive the first equality and the matrix determinant lemma
for the second equality. The second inequality is derived from Assumption 8 since ǫ ≤ det(Ax) =
det( ¯Ax,t + U diag(ǫKx,t)) by definition of U . The fourth inequality is derived from the observation that
diag(ǫKx,t) ¯A−1
k,t)k∈Kx . The fifth inequality is obtained by observing that the com-
ponents of ¯Ax,t and U are all smaller than 1 in absolute value. The sixth inequality is obtained by observing
that the elements of ¯A−1
2·C! for all arms
k as a result of the initialization phase. We get:
ǫ · (C − 1)! since det( ¯Ax,t) ≥ ǫ
x,tbCx is the vector (ǫk,t · ξx
2 and that ǫk,t ≤ ǫ
x,t are smaller than 2
T
E[
Xt=tini
Iobjx,t≥objx+Ex,t · Ix∈Bt] ≤
P[ck,t(i)− µc
k(i) ≥ ǫk,t]
T
T
+
i∈Cx
Xt=tini Xk∈Kx
Xt=tini Xk∈Kx
+ 29 (C + 3)!4
≤ 2· 29 (C + 3)!4
ǫ6
ǫ6
.
P[rk,t ≥ µr
k + ǫk,t]
64
Fact 14
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
T
E[
Xt=tini
Iobjx∗,t≤objx∗−Ex,t · Ix∗∈Bt] ≤ 210 (C + 3)!2
ǫ6
.
We omit the proof since it is almost identical to the proof of Fact 13.
F.5. Proof of Theorem 8. Along the same lines as for Theorem 5, we build upon (4):
RB(1),··· ,B(C−1),T ≤ T ·
= T ·
≤ T ·
≤ T ·
K
K
K
Xk=1
Xk=1
Xk=1
Xk=1
K
k · ξx∗
µr
k · ξx∗
µr
k · ξx∗
µr
k · ξx∗
µr
τ ∗
T
Xt=1
k − E[
Xt=1
k − E[
Xt=1
k − E[
Xt=1
k − E[
T
T
rat,t] + O(1)
T
T
rat,t] + E[
rat,t] + O(1)
Xt=τ ∗+1
Xt=τ ∗+1
rat,t] + σ · E[ min
i=1,··· ,C
T
C
Xt=1
cat,t(i)− B(i))+] + O(1).
cat,t(i)] + O(1)
rat,t] + σ ·
Xi=1
E[(
The second inequality is a direct consequence of Assumption 7. To derive the last inequality, observe that
if τ ∗ = T + 1, we have:
T
Xt=τ ∗+1
cat,t(j) = 0 ≤
C
Xi=1
E[(
T
Xt=1
cat,t(i)− B(i))+],
for any j ∈ {1,··· , C} while if τ ∗ < T + 1 we have run out of resources before the end of the game, i.e.
there exists j ∈ {1,··· , C} such thatPτ ∗
t=1 cat,t(j) ≥ B(j), which implies that:
T
T
T
min
i=1,··· ,C
cat,t(i) ≤
Xt=τ ∗+1
Xt=τ ∗+1
Xt=τ ∗+1
Xt=1
Xi=1
Now observe that, for any resource i ∈ {1,··· , C}:
= (
≤
≤
(
C
T
cat,t(j)
τ ∗
cat,t(j) +
Xt=1
cat,t(j)− B(j))+
Xt=1
T
cat,t(i)− B(i))+.
cat,t(j)− B(j)
E[(
T
Xt=1
cat,t(i)− B)+] ≤ E[(
T
Xt=tini
cat,t(i)− b(i))+] + K · 28 (C + 2)!4
{bx,T (i)− nx,T · b(i)})+] + O(K ·
ǫ6
· ln(T )
E[(bx,T (i)− nx,T · b(i))+] +Xx /∈B
E[nx,T ]
E[nx,T ] + O(K ·
ǫ6
· ln(T ))
(C + 2)!4
= E[( Xx basis for (3)
≤Xx∈B
+ Xx pseudo-basis for (3)
with det(Ax)=0
(C + 2)!4
ǫ6
· ln(T ))
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
65
(C + 3)!4
ǫ6
· ln(T ))
≤Xx∈BZ T
≤Xx∈B
+ 28B·
= O(B· (C + 3)!4
(C + 3)!4
ǫ6
ǫ6
0
P[bx,T (i)− nx,T · b(i) ≥ u]du + O(K ·
T · P[bx,T (i)− nx,T · b(i) ≥ 28 (C + 3)!4
ǫ6
· ln(T )]
(C + 3)!4
ǫ6
· ln(T ))
· ln(T ) + O(K ·
· ln(T )),
where we use the fact that the amounts of resources consumed at any time period are no larger than 1 for
the first and second inequalities, Lemma 13 for the third inequality and inequalities (12) and (13) from
Lemma 14 along with the fact that there are at least K feasible basis for (3) (corresponding to single-armed
strategies) for the last equality. Plugging this back into the regret bound yields:
RB(1),··· ,B(C−1),T ≤ T ·
T
K
K
K
(
K
K
(
K
K
=
= T ·
k · ξx∗
µr
Xt=tini
Xk=1
k − E[
k · ξx∗
Xk=1
Xk=1
k −Xx∈B
µr
k · ξx∗
Xk=1
Xk=1
k −Xx∈B
µr
≤ T ·
k · ξx∗
Xk=1
k · (T − Xx∈B ∆x=0
µr
Xk=1
− Xx∈B ∆x>0
k · ξx∗
Xk=1
k · (tini + Xx∈B ∆x>0
µr
Xk=1
− Xx∈B ∆x>0
≤ Xx∈B ∆x>0
∆x · E[nx,T ] + O(
≤ 210 (C + 3)!3 · λ2
· ( Xx∈B ∆x>0
µr
k · ξx
µr
k · ξx
=
ǫ6
K
K
(
rat,t] + O(
σ ·B· (C + 3)!4
· ln(T ))
ǫ6
σ ·B· (C + 3)!4
· ln(T ))
ǫ6
σ ·B· (C + 3)!4
ǫ6
µr
k · E[nx
k,T ] + O(
µr
k · ξx
k )· E[nx,T ] + O(
· ln(T ))
(47)
E[nx,T ])
k )· E[nx,T ] + O(
σ ·B· (C + 3)!4
ǫ6
· ln(T ))
E[nx,T ] +Xx /∈B
E[nx,T ] + Xx pseudo-basis for (3)
with det(Ax)=0
E[nx,T ])
k )· E[nx,T ] + O(
σ ·B· (C + 3)!4
· ln(T ))
ǫ6
σ ·B· (C + 3)!4
ǫ6
1
∆x
)· ln(T ) + O(
· ln(T ))
σ ·B· (C + 3)!4
ǫ6
(48)
· ln(T )),
where we use (14) from Lemma 14 for the second inequality, Lemma 13 for the third inequality. To derive
the last inequality, we use Lemma 15 and the fact that ∆x ≤ objx∗ =PK
k=1 µr
k · ξx∗
k ≤PK
F.6. Proof of Theorem 9. Along the same lines as for the proof of Theorem 6, we start from inequality
(48) derived in the proof of Theorem 8 and apply Lemma 15 only if ∆x is big enough, taking into account
the fact that:
k=1 ξx∗
k ≤ 1.
E[nx,T ] ≤ T.
Xx∈B
Specifically, we have:
RB(1),··· ,B(C−1),T
66
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
ln(T )
∆x
·
+ 211 (C + 4)!2
ǫ6
· ∆x) })
ln(T )
∆x
·
) } + O(
σ ·B· (C + 3)!4
ǫ6
· ln(T ))
min(∆x · nx, 210 (C + 3)!3 · λ2
ǫ6
· ln(T ))
min(∆x · nx, 210 (C + 3)!3 · λ2
ǫ6
= sup
(nx)x∈B≥0
Px∈B nx≤T
+ O(
ǫ6
{ Xx∈B ∆x>0
σ ·B· (C + 3)!4
{ Xx∈B ∆x>0
{Xx∈Br210 (C + 3)!3 · λ2
ǫ6
= sup
(nx)x∈B≥0
Px∈B nx≤T
(nx)x∈B≥0
Px∈B nx≤T
≤ sup
≤ 25 (C + 3)!2 · λ
≤ 25 (C + 3)!2 · λ
ǫ3
ǫ3
· ln(T )· nx } + O(
√nx } + O(
{Xx∈B
·pln(T )·
σ ·B· (C + 3)!4
·pσ ·B· T · ln(T ) + O(
(nx)x∈B≥0
Px∈B nx≤T
sup
ǫ6
σ ·B· (C + 3)!4
σ ·B· (C + 3)!4
ǫ6
ǫ6
· ln(T ))
· ln(T ))
· ln(T )),
where we use the fact that ∆x ≤ 1 (see the proof of Theorem 8) for the second equality, we maximize over
each ∆x ≥ 0 to derive the first inequality, and we use Cauchy-Schwartz for the last inequality.
F.7. Proof of Theorem 10. Define b(i) = B(i)/ T for any i ∈ {1,··· , C}. If the decision maker stops
pulling arms at round T at the latest, all the results derived in Section 7 hold as long as we substitute T with
T and we get:
min(τ ∗, T )
T · opt− E[
Xt=1
rat,t] ≤ X,
where X denotes the right-hand side of the regret bound derived in either Theorem 8 or Theorem 9 and
opt denotes the optimal value of (3) when b(i) is substituted with b(i) for any i ∈ {1,··· , C}. The key
observation is that T · opt = T · opt, where opt denotes the optimal value of (3), because the time constraint
is redundant in (3) even when b(i) is substituted with b(i) for any i ∈ {1,··· , C}. This is enough to show
the claim as we get:
τ ∗
Xt=1
X ≥ T · opt− E[
≥ RB(1),··· ,B(C−1),T ,
rat,t]
where we use Lemma 2 for the last inequality.
F.8. Proof of Theorem 11. The only difference with the proofs of Theorems 8 and 9 lies in how we
t=τ ∗+1 rat,t]. We have:
T
E[
bound E[PT
Xt=τ ∗+1
Xt=0
Xi=1
≤
=
C
T
rat,t] ≤ E[(T − τ ∗)+]
T
P[τ ∗ ≤ T − t]
Xt=0
Xτ =1
T−t
P[
caτ ,τ (i) ≥ B(i)]
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
67
(bx,T−t(i)− nx,T−t · b(i)) ≥ t· b(i)]
C
T
T−t
nx,T +Xx∈B
nx,T ≥ t·
C
T
C
T
C
T
=
=
=
P[
with det(Ax)=0
Xτ =1
P[tini +Xx /∈B
P[tini +Xx /∈B
Xt=0
Xi=1
(caτ ,τ (i)− b(i)) ≥ t· b(i)]
nx,T + Xx pseudo-basis for (3)
Xt=0
Xi=1
nx,T + Xx pseudo-basis for (3)
Xt=0
Xi=1
Xt=0Xx∈B
Xi=1
P[bx,T−t(i)− nx,T−t · b(i) ≥ t·
tini +Px /∈B
Xt=1
Xi=1
≤ 2
Xt=1Xx∈B
Xi=1
+ 2B·
(C + 4)!4 ·B2
b· ǫ6
b(i)
2·B
E[nx,T ] +Px pseudo-basis for (3)
E[bx,T−t(i)− nx,T−t · b(i)]
· ln2(T )),
t· b(i)
t· b(i)
with det(Ax)=0
with det(Ax)=0
= O(
+
C
T
C
T
+ O(1)
b(i)
]
2
]
E[nx,T ]
where we use Lemma 13 and we bound E[bx,T−t(i) − nx,T−t · b(i)] in the same fashion as in the proof of
Theorem 8 using Lemma 14.
Appendix G: Proofs for Section A.
G.1. Proof of Lemma 16. The proof follows the same steps as for Lemma 8. We use the shorthand
k,t. Along the same lines as in Fact 1,
)2 and n6=x∗
i=1 b(i))2
ǫ2
notations βk = 8 ρ·(PC
we have:
· ( 1
∆k
k,t =Px∈B k∈Kx,x6=x∗ nx
τ ∗
k,t ≥βk ln(t)],
and we can focus on bounding the second term, which can be broken down as follows:
Xt=1
E[nk,τ ∗] ≤ 2βk · E[ln(τ ∗)] + E[
Ixt6=x∗ · Iat=k · In6=x∗
τ ∗
E[
Xt=1
Ixt6=x∗ · Iat=k · In6=x∗
k,t ≥βk ln(t)]
τ ∗
τ ∗
τ ∗
= E[
Xt=1
Xt=1
≤ E[
Xt=1
Xt=1
+ E[
+ E[
τ ∗
Iobjxt,t+Ext,t≥objx∗,t+Ex∗,t · Ixt6=x∗ · Iat=k · In6=x∗
k,t ≥βk ln(t)]
Iobjxt,t≥objxt
+Ext,t]
Iobjx∗,t≤objx∗−Ex∗,t]
Iobjx∗ <objxt
+2Ext,t · Ixt6=x∗ · Iat=k · In6=x∗
k,t ≥βk ln(t)].
We study each term separately, just like in Lemma 8.
Fact 15
τ ∗
E[
Xt=1
Iobjxt,t≥objxt
+Ext,t] ≤ K ·
π2
6
.
68
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
Proof. If objxt,t ≥ objxt + Ext,t, there must exist l ∈ Kxt such that ¯rl,t ≥ µr
l + ǫl,t, otherwise:
objxt,t − objxt = Xl∈Kxt
< Xl∈Kxt
= Ext,t,
l )· ξxt
l
(¯rl,t − µr
ǫl,t · ξxt
l
where the inequality is strict because there must exist l ∈ Kxt such that ξxt
l > 0 (at least one resource
constraint is binding for a feasible basis to (3) aside from the basis x associated with Kx = ∅). We obtain:
τ ∗
Iobjxt,t≥objxt
E[
Xt=1
τ ∗
K
K
∞
I¯rl,t≥µr
Xl=1
Xt=1
+Ext,t] ≤ E[
Xl=1
Xt=1
P[¯rl,t ≥ µr
≤ K ·
π2
6
≤
,
l +ǫl,t]
l + ǫl,t]
where the last inequality is derived along the same lines as in the proof of Fact 3.
Similarly, we can show that:
We move on to study the last term.
Fact 16
τ ∗
τ ∗
E[
Xt=1
Iobjx∗,t≤objx∗−Ex∗,t] ≤ K ·
π2
6
.
E[
Xt=1
Iobjx∗ <objxt
+2Ext,t · Ixt6=x∗ · Iat=k · In6=x∗
k,t ≥βk ln(t)] = 0.
Proof. If objx∗ < objxt + 2Ext,t, xt 6= x∗, and at = k, we have:
∆xt
2
∆k
2 ≤
< Xl∈Kxt
≤ Xl∈Kxt
ξxt
l
·s 2 ln(t)
nl,t
s 2ξxt
· ξxt
k ln(t)
nk,t
l
,
where we use the fact that, by definition of the load balancing algorithm and since at = k, ξxt
arm k would not have been selected) and:
k 6= 0 (otherwise
nl,t ≥
ξxt
l
ξxt
k
nk,t,
(49)
for all arms l ∈ Kxt . We get:
nk,t <
≤
≤
l )2 · ln(t)
· ln(t)
8
8
k · (Xl∈Kxtpξxt
(∆k)2 · ξxt
k · ρ· Xl∈Kxt
(∆k)2 · ξxt
(∆k)2 · ρ· (Xl∈Kxt
ξxt
l
8
ξxt
l )2 · ln(t),
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
69
using the Cauchy−Schwarz inequality and the fact that a basis involves at most ρ arms. Now observe that:
Xl∈Kxt
i=1 cl(i)
ξxt
l ≤ Xl∈Kxt PC
≤PC
i=1 b(i)
ǫ
ǫ
· ξxt
l
as xt is a feasible basis to (8) and using Assumption 2. We obtain:
n6=x∗
k,t ≤ nk,t
ρ· (PC
< 8·
ǫ2 · (∆k)2
= βk · ln(t).
i=1 b(i))2
· ln(t)
G.2. Proof of Lemma 17. We first show (16) by induction on t. The base case is straightforward.
Suppose that the inequality holds at time t− 1. There are three cases:
while the right-hand side can only increase, hence the inequality still holds at time t,
• arm k is not pulled at time t− 1, in which case the left-hand side of the inequality remains unchanged
• arm k is pulled at time t − 1 after selecting xt−1 6= x∗, in which case both sides of the inequality
• arm k is pulled at time t− 1 after selecting xt−1 = x∗. First observe that there must exist l ∈ Kx∗ such
increase by one and the inequality still holds at time t,
that nl,t−1 ≤ (t− 1)·
ξx∗
l
r=1 ξx∗
r
PK
. Suppose otherwise, we have:
t− 1 =
nl,t
nl,t
K
Xl=1
≥ Xl∈Kx∗
> Xl∈Kx∗
= t− 1,
(t− 1)·
l
ξx∗
r=1 ξx∗
r
PK
a contradiction. Suppose now by contradiction that inequality (16) no longer holds at time t, we have:
k
ξx∗
l=1 ξx∗
nk,t−1 = nk,t − 1
> nx∗,t ·
PK
≥ (nx∗,t + Xx∈B,x6=x∗
= (t− 1)·
ξx∗
l=1 ξx∗
k
l
l
PK
nx,t
k
ξx∗
l=1 ξx∗
l
+ Xx∈B,x6=x∗
nx,t)·
PK
,
which implies, using the preliminary remark above, that
definition of the load balancing algorithm.
ξx∗
k
nk,t−1
ξx∗
l
nl,t−1
< max
l∈Kx∗
, a contradiction given the
70
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
We conclude that inequality (16) holds for all times t and arms k ∈ Kx∗. We also derive inequality (15) as a
byproduct, since, at any time t and for any arm k ∈ Kx∗:
nl,t
nk,t ≥ nx∗,t − Xl∈Kx∗ ,l6=k
PK
l − ρ· ( Xx∈B,x6=x∗
≥ nx∗,t · (1−Pl∈Kx∗ ,l6=k ξx∗
= nx∗,t ·
ξx∗
l=1 ξx∗
l=1 ξx∗
k
l
l
PK
nx,t + 1)
)− ρ· ( Xx∈B,x6=x∗
nx,t + 1),
as a basis involves at most ρ arms.
G.3. Proof of Theorem 12. The proof proceeds along the same lines as for Theorem 3. We build upon
(4):
K
K
RB(1),··· ,B(C) ≤ B ·
µr
k · E[nk,τ ∗ ] + O(1)
K
≤ B ·
Xk=1
Xk=1
≤ (B −
PK
k · ξx∗
Xk=1
µr
k −
k · ξx∗
k − Xk∈Kx∗
µr
E[nx∗,τ ∗]
Xk=1
)·
k=1 ξx∗
K
k
µr
k · E[nk,τ ∗ ] + O(1)
k · ξx∗
k + ρ2 · Xx∈B,x6=x∗
µr
E[nx,τ ∗] + O(1),
where we use (15) to derive the third inequality. Now observe that, by definition, at least one resource is
exhausted at time τ ∗. Hence, there exists i ∈ {1,··· , C} such that the following holds almost surely:
K
B(i) ≤
ck(i)· nk,τ ∗
nk,τ ∗ + Xk∈Kx∗
nx,τ ∗ + Xk∈Kx∗
Xk=1
≤ Xk /∈Kx∗
≤ Xx∈B,x6=x∗
≤ ρ· ( Xx∈B,x6=x∗
≤ ρ· ( Xx∈B,x6=x∗
ck(i)· nk,τ ∗
ck(i)· nk,τ ∗
ck(i)·
,
k
ξx∗
l=1 ξx∗
l
PK
nx∗,τ ∗
k=1 ξx∗
k
nx,τ ∗ + 2) + nx∗,τ ∗ · Xk∈Kx∗
nx,τ ∗ + 2) + b(i)·
PK
b · ( Xx∈B,x6=x∗
ρ
k ≥ B −
nx,τ ∗ + 2),
nx∗,τ ∗
k=1 ξx∗
PK
where we use (16) and the fact that x∗ is a feasible basis to (3). Rearranging yields:
almost surely. Plugging this last inequality back into the regret bound, we get:
RB(1),··· ,B(C) ≤ ρ· Xx∈B,x6=x∗
≤ ρ· Xx∈B,x6=x∗
+ ρ) + O(1)
k
E[nx,t]· (PK
E[nx,t]· (PK
k=1 µr
b
k · ξx∗
i=1 ck(i)· ξx∗
k=1PC
ǫ· b
k
+ ρ) + O(1)
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
71
K
= (
≤ (
E[nx,t] + O(1)
ρ·PC
i=1 b(i)
ǫ· b
ρ·PC
i=1 b(i)
ǫ· b
ρ3 · (PC
i=1 b(i))3
≤ 32
ǫ3 · b
ρ3 · (PC
i=1 b(i))3
ǫ3 · b
+ (ρ)2)· Xx∈B,x6=x∗
Xk=1
E[ Xx∈B k∈Kx,x6=x∗
+ (ρ)2)·
Xk=1
(∆k)2 )· E[ln(τ ∗)] + O(1)
(∆k)2 )· ln(PC
i=1 b(i)· B
Xk=1
≤ 32
· (
· (
nx
1
1
ǫ
K
K
k,τ ∗] + O(1)
+ 1) + O(1),
where we use the fact that x∗ is a feasible basis to (3) for the third inequality, Lemma 16 for the fourth
inequality, the concavity of the logarithmic function along with Lemma 7 for the last inequality.
G.4. Proof of Lemma 18. We use the shorthand notations βk = 8C · ( λ
n /∈Ok,t =Px∈B k∈Kx, x /∈O
k,t. Similarly as in Fact 11, we have:
nx
∆k
E[n /∈Ok,T ] ≤ 2βk · ln(T ) + 29 (C + 3)!4
Ixt /∈O · Iat=k · In /∈O
+ E[
ǫ6
T
Xt=tini
k,t ≥βk ln(t) · Ix∗∈Bt],
)2 and, for any round t,
and what remains to be done is to bound the second term, which we can break down as follows:
T
Iobjxt,t+Ext,t≥objx∗,t+Ex∗,t · Ixt /∈O · Iat=k · In /∈O
k,t ≥βk ln(t) · Ix∗∈Bt]
k,t ≥βk ln(t) · Ix∗∈Bt]
E[
Xt=tini
≤ E[
≤ E[
+ E[
+ E[
T
T
Iobjxt,t≥objxt
Ixt /∈O · Iat=k · In /∈O
Xt=tini
Xt=tini
Xt=tini
Xt=tini
Iobjx∗ <objxt
T
T
+Ext,t · Ixt∈Bt]
Iobjx∗,t≤objx∗−Ex∗,t · Ix∗∈Bt]
+2Ext,t · Ixt /∈O · Iat=k · In /∈O
k,t ≥βk ln(t)].
The study of the second term is the same as in the proof of Lemma 15. We can also bound the first term in
the same fashion as in the proof of Lemma 15 since there is no reference to the load balancing algorithm in
the proof of Fact 13. The major difference with the proof of Lemma 15 lies in the study of the last term.
Fact 17
T
Iobjxt,t≥objxt
+Ext,t · Ix∈Bt] ≤ 210 K · (C + 3)!2
ǫ6
.
E[
Xt=tini
Proof. The only difference with the proof of Fact 13 is that the number of arms that belong to Kx for x
ranging in {x ∈ B k ∈ Kx, x /∈ O} can be as big as K, as opposed to C when we are considering one
basis at a time. This increases the bound by a multiplicative factor K.
72
Fact 18
T
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
E[
Xt=tini
Iobjx∗ <objxt
+2Ext,t · Ixt /∈O · Iat=k · In /∈O
k,t ≥βk ln(t)] = 0.
Proof. Assume that objx∗ < objxt + 2Ext,t, xt /∈ O, and at = k. We have:
∆xt
2
∆k
2 ≤
< λ· Xl∈Kxt
≤ λ· Xl∈Kxt
ξxt
l,t ·s 2 ln(t)
nl,t
s 2ξxt
l,t · ξxt
nk,t
k,t ln(t)
,
where we use the fact that, by definition of the load balancing algorithm and since at = k, ξxt
arm k would not have been selected) and:
k,t 6= 0 (otherwise
nl,t ≥
ξxt
l,t
k,t · nk,t,
ξxt
(50)
for any arm l ∈ Kxt . We get:
n /∈Ok,t ≤ nk,t
λ
< 8(
∆k
λ
∆k
λ
∆k
≤ 8(
≤ 8(
≤ 8C · (
l,t)2 · ln(t)
ξxt
l,t · ln(t)
ξxt
l,t)2 · ln(t)
)2 · ξxt
k,t · (Xl∈Kxtqξxt
k,t · C · Xl∈Kxt
)2 · ξxt
)2 · C · (Xl∈Kxt
)2 · ln(t),
λ
∆k
using the Cauchy−Schwarz inequality, the fact that a basis involves at most C arms, and the fact that
xt is feasible for (8) whose linear constraints include PK
l=1 ξl ≤ 1 and ξl ≥ 0,∀l ∈ {1,··· , K}. We get
n /∈Ok,t < βk ln(t) by defintion of βk.
G.5. Proof of Theorem 13. Substituting b(i) with B(i)/B(C) for every resource i ∈ {1,··· , C}, the
regret bound obtained in Theorem 3 turns into:
RB(1),··· ,B(C) ≤ 16
ρ
ǫ ·PC
i=1 B(i)
B(C)
· ( Xx∈B ∆x>0
1
∆x
i=1 B(i)
)· ln(PC
ǫ
+ 1) + O(1).
(51)
and O∁
are defined by strict
Observe that B∁
∞ ⊂ O∁ and thus
(B(1)/B(C),··· , B(C − 1)/B(C)). Hence, for B(C) large enough, B∁
B ⊂ B∞ and O ⊂ O∞. We now move on to prove each claim separately.
First claim. Suppose that there exists a unique optimal basis to (3), which we denote by x∗. Then, we must
have O = {x∗} = O∞ for B(C) large enough. Indeed, using the set inclusion relations shown above, we
have O ⊂ O∞ = {x∗} and O can never be empty as there exists at least one optimal basis to (3) (this linear
program is feasible and bounded). We get O∁ ∩B ⊂ O∁
∞ ∩B∞ for B(C) large enough. Note moreover that,
∞ ⊂ B∁ and O∁
that are linear
in the vector
inequalities
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
73
for any x ∈ B, ∆x converges to ∆∞x (because both the objective value of a feasible basis and the optimal
value of a linear program are Lipschitz in the right-hand side of the inequality constraint), which implies
that ∆x > ∆∞
2 > 0 when x ∈ B ∩O∁ for B(C) large enough. We conclude with (51) that:
x
RB(1),··· ,B(C) ≤ 32
ρ
ǫ ·PC
i=1 B(i)
B(C)
· ( Xx∈B∞ ∆∞
x >0
1
∆x
i=1 B(i)
)· ln(PC
ǫ
+ 1) + O(1),
for B(C) large enough. This yields the result since B(i)/B(C) → b(i) > 0 for any resource i = 1,··· , C −
1.
Second claim. Suppose that B(i)
B(C) ) for any resource i ∈ {1,··· , C − 1}. Starting from
(30) derived in the proof of Theorem 3 and applying Lemma 8 only if ∆x is big enough, we have:
B(C) − b(i) = O( ln(B(C))
min(
PK
min(∆x ·
∆x
k=1 ξx
k
· E[nx,τ ∗], 16ρ·PK
mini=1,··· ,C B(i) ·PC
B(C)
ǫ
k
k=1 ξx
∆x
i=1 B(i)
ǫ
· ln(PC
ρ·PC
, 16
i=1 B(i)
+ 1) +
π2
3
i=1 B(i)/B(C)
ǫ
) + O(1)
∆x
k=1 ξx
k
i=1 B(i)
ρ·
PK
∆x · ln(PC
1
ǫ
+ 1))
·
RB(1),··· ,B(C)
≤ Xx∈B ∆x>0
≤ Xx∈B ∆x>0
ρ·PC
≤ 16
+ ( Xx∈B∩O∁∩O∞
+ O(1)
where we use:
and
i=1 B(i)/B(C)
· ( Xx∈B∩O∁∩O∁
∞
1
∆x
)· ln(PC
ǫ
i=1 B(i)
+ 1)
B(C)
mini=1,··· ,C B(i) ·PC
ǫ
i=1 B(i)
+ O(1), '
ǫ
∆x)·
K
Xk=1
ξx
k ∈ [ min
i=1,··· ,C
B(i)/B(C),PC
i=1 B(i)/B(C)
]
ǫ
,
∆x ≤ PC
i=1 B(i)/B(C)
ǫ
as shown in the proof of Theorem 3 (substituting b with mini=1,··· ,C B(i)/B(C)). For x ∈ B∩O∁∩O∁
∞
have x ∈ B∞ and ∆x > ∆∞
2 > 0 for B(C) large enough, as shown for the first claim. For x ∈ B ∩O∁ ∩O∞,
we have ∆x = O(ln(B(C))/B(C)) as both the objective value of a feasible basis and the optimal value
of a linear program are Lipschitz in the right-hand side of the inequality constraints. We conclude that, for
B(C) large enough:
, we
x
RB(1),··· ,B(C)
≤ 32
1
ǫ ·
i=1 B(i)/B(C)
ρ·PC
· ( Xx∈B∞∩O∁
mini=1,··· ,C B(i) · Xx∈B∩O∁∩O∞
B(C)
+
ǫ
∞
1
∆∞x
)· ln(PC
ǫ
i=1 B(i)
+ 1)
O(ln(B(C))) + O(1).
This yields the result since B ∩ O∁ ∩ O∞ ≤ O∞ and B(i)/B(C) → b(i) > 0 for any resource i =
1,··· , C − 1.
74
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
G.6. Proof of Theorem 14. The proof is along the same lines as for Theorem 13. Specifically, in a first
step, we observe that all the proofs of Section 6 remain valid (up to universal constant factors) for T large
enough as long as we substitute b with B/T . Indeed, for T large enough, we have B
T > ǫ
for all arms k ∈ {1,··· , K} under Assumption 6. In a second step, just like in the proof of Theorem 13,
we show that we can substitutePx∈B ∆x>0
in the regret bound up to universal
constant factors.
withPx∈B∞ ∆∞
T ≤ 2 and µc
k − B
1
∆x
1
x >0
∆∞
x
2
G.7. Proof of Theorem 15. The proof is along the same lines as for Theorem 13. Specifically, in a
first step, we observe that all the proofs of Section 7 remain valid (up to universal constant factors) for T
large enough as long as we substitute b with mini=1,··· ,C−1 B(i)/T . Indeed, for T large enough, we have
mini=1,··· ,C−1 B(i)/T ≤ 2 and, under Assumption 8, any basis to (17) has determinannt larger than ǫ/2 in
absolute value and is ǫ/2−non-degenerate by continuity of linear functions. In a second step, just like in the
proof of Theorem 13, we show that we can substitutePx∈B ∆x>0
in the regret
bound up to universal constant factors.
withPx∈B∞ ∆∞
1
∆x
x >0
∆∞
x
1
References
[1] Agrawal S, Devanur NR (2014) Bandits with concave rewards and convex knapsacks. Proc. 15th ACM Conf.
Economics and Computation, 989 -- 1006.
[2] Agrawal S, Devanur NR, Li L (2016) An efficient algorithm for contextual bandits with knapsacks, and an
extension to concave objectives. Proc. 29th Annual Conf. Learning Theory, 4 -- 18.
[3] Agrawal S, Goyal N (2012) Analysis of thompson sampling for the multi-armed bandit problem. Proc. 25th
Annual Conf. Learning Theory, volume 23.
[4] Auer P, Cesa-Bianchi N, Fischer P (2002) Finite-time analysis of the multiarmed bandit problem. Machine learn-
ing 47(2-3):235 -- 256.
[5] Auer P, Cesa-Bianchi N, Freund Y, Schapire RE (2002) The nonstochastic multiarmed bandit problem. SIAM J.
Comput. 32(1):48 -- 77.
[6] Babaioff M, Dughmi S, Kleinberg R, Slivkins A (2012) Dynamic pricing with limited supply. Proc. 13th ACM
Conf. Electronic Commerce, 74 -- 91.
[7] Badanidiyuru A, Kleinberg R, Singer Y (2012) Learning on a budget: posted price mechanisms for online pro-
curement. Proc. 13th ACM Conf. Electronic Commerce, 128 -- 145.
[8] Badanidiyuru A, Kleinberg R, Slivkins A (2013) Bandits with knapsacks. Proc. 54th IEEE Annual Symp. Foun-
dations of Comput. Sci., 207 -- 216.
[9] Badanidiyuru A, Langford J, Slivkins A (2014) Resourceful contextual bandits. Proc. 27th Annual Conf. Learn-
ing Theory, volume 35, 1109 -- 1134.
[10] Bertsimas D, Tsitsiklis JN (1997) Introduction to linear optimization, volume 6 (Athena Scientific).
[11] Besbes O, Zeevi A (2012) Blind network revenue management. Oper. Res. 60(6):1537 -- 1550.
[12] Combes R, Jian C, Srikant R (2015) Bandits with budgets: Regret lower bounds and optimal algorithms. Proc.
2015 ACM SIGMETRICS Int. Conf. Measurement and Modeling Computer Systems, 245 -- 257.
[13] Ding W, Qin T, Zhang X, Liu T (2013) Multi-armed bandit with budget constraint and variable costs. Proc. 27th
AAAI Conf. Artificial Intelligence, 232 -- 238.
[14] Ghosh A, Rubinstein BIP, Vassilvitskii S, Zinkevich M (2009) Adaptive bidding for display advertising. Proc.
18th Int. Conf. World Wide Web, 251 -- 260.
[15] Graczov´a D, Jacko P (2014) Generalized restless bandits and the knapsack problem for perishable inventories.
Oper. Res. 62(3):696 -- 711.
[16] Grotschel M, Lov´asz L, Schrijver A (2012) Geometric algorithms and combinatorial optimization, volume 2
(Springer).
[17] Johnson K, Simchi-Levi D, Wang H (2015) Online network revenue management using thompson sampling.
Working Paper .
Flajolet and Jaillet: Logarithmic Regret Bounds for BwK
75
[18] Kleinberg R, Leighton T (2003) The value of knowing a demand curve: Bounds on regret for online posted-price
auctions. Proc. 44th IEEE Annual Symp. Foundations of Comput. Sci., 594 -- 605.
[19] Lai TL, Robbins H (1985) Asymptotically efficient adaptive allocation rules. Adv. Applied Math. 6(1):4 -- 22.
[20] Papadimitriou CH, Tsitsiklis JN (1999) The complexity of optimal queuing network control. Math. Oper. Res.
24(2):293 -- 305.
[21] Robbins H (1952) Some aspects of the sequential design of experiments. Bulleting American Math. Society
58(5):527 -- 535.
[22] Slivkins A (2013) Dynamic ad allocation: Bandits with budgets. arXiv preprint arXiv:1306.0155 .
[23] Tran-Thanh L, Chapman A, Rogers A, Jennings NR (2010) ǫ-first policies for budget-limited multi-armed ban-
dits. Proc. 24th AAAI Conf. Artificial Intelligence, 1211 -- 1216.
[24] Tran-Thanh L, Chapman A, Rogers A, Jennings NR (2012) Knapsack based optimal policies for budget-limited
multi-armed bandits. Proc. 26th AAAI Conf. Artificial Intelligence, 1134 -- 1140.
[25] Tran-Thanh L, Rogers A, Jennings NR (2012) Long-term information collection with energy harvesting wireless
sensors: a multi-armed bandit based approach. Autonomous Agents and Multi-Agent Systems 25(2):352 -- 394.
[26] Tran-Thanh L, Stavrogiannis C, Naroditskiy V, Robu V, Jennings NR, Key P (2014) Efficient regret bounds for
online bid optimisation in budget-limited sponsored search auctions. Proc. 30th Conf. Uncertainty in Artificial
Intelligence, 809 -- 818.
[27] Trovo F, Paladino S, Restelli M, Gatti N (2016) Budgeted multi -- armed bandit in continuous action space. Work-
ing Paper .
[28] Weed J, Perchet V, Rigollet P (2016) Online learning in repeated auctions. Proc. 29th Annual Conf. Learning
Theory, volume 49, 1562 -- 1583.
[29] Wu H, Srikant S, Liu X, Jiang C (2015) Algorithms with logarithmic or sublinear regret for constrained contextual
bandits. Adv. Neural Inform. Processing Systems, 433 -- 441.
[30] Xia Y, Ding W, Zhang X, Yu N, Qin T (2015) Budgeted bandit problems with continuous random costs. Proc.
7th Asian Conf. Machine Learning, 317 -- 332.
|
1902.02526 | 2 | 1902 | 2019-02-14T09:18:46 | Going Far From Degeneracy | [
"cs.DS",
"cs.DM"
] | An undirected graph G is d-degenerate if every subgraph of G has a vertex of degree at most d. By the classical theorem of Erd\H{o}s and Gallai from 1959, every graph of degeneracy d>1 contains a cycle of length at least d+1. The proof of Erd\H{o}s and Gallai is constructive and can be turned into a polynomial time algorithm constructing a cycle of length at least d+1. But can we decide in polynomial time whether a graph contains a cycle of length at least d+2? An easy reduction from Hamiltonian Cycle provides a negative answer to this question: deciding whether a graph has a cycle of length at least d+2 is NP-complete. Surprisingly, the complexity of the problem changes drastically when the input graph is 2-connected. In this case we prove that deciding whether G contains a cycle of length at least d+k can be done in time 2^{O(k)}|V(G)|^{O(1)}. In other words, deciding whether a 2-connected n-vertex G contains a cycle of length at least d+log n can be done in polynomial time.
Similar algorithmic results hold for long paths in graphs. We observe that deciding whether a graph has a path of length at least d+1 is NP-complete. However, we prove that if graph G is connected, then deciding whether G contains a path of length at least d+k can be done in time 2^{O(k)}n^{O(1)}. We complement these results by showing that the choice of degeneracy as the `above guarantee parameterization' is optimal in the following sense: For any \epsilon>0 it is NP-complete to decide whether a connected (2-connected) graph of degeneracy d has a path (cycle) of length at least (1+\epsilon)d. | cs.DS | cs | Fedor V. Fomin†
Going Far From Degeneracy∗
Daniel Lokshtanov‡
Petr A. Golovach†
Saket Saurabh§
Meirav Zehavi¶
Fahad Panolan†
9
1
0
2
b
e
F
4
1
]
S
D
.
s
c
[
2
v
6
2
5
2
0
.
2
0
9
1
:
v
i
X
r
a
Abstract
An undirected graph G is d-degenerate if every subgraph of G has a vertex of degree at
most d. By the classical theorem of Erdos and Gallai from 1959, every graph of degeneracy
d > 1 contains a cycle of length at least d + 1. The proof of Erdos and Gallai is constructive
and can be turned into a polynomial time algorithm constructing a cycle of length at least
d + 1. But can we decide in polynomial time whether a graph contains a cycle of length at
least d + 2? An easy reduction from Hamiltonian Cycle provides a negative answer to
this question: Deciding whether a graph has a cycle of length at least d + 2 is NP-complete.
Surprisingly, the complexity of the problem changes drastically when the input graph is 2-
connected. In this case we prove that deciding whether G contains a cycle of length at least
d + k can be done in time 2O(k)V (G)O(1). In other words, deciding whether a 2-connected
n-vertex G contains a cycle of length at least d + log n can be done in polynomial time.
Similar algorithmic results hold for long paths in graphs. We observe that deciding
whether a graph has a path of length at least d + 1 is NP-complete. However, we prove
that if graph G is connected, then deciding whether G contains a path of length at least
d + k can be done in time 2O(k)nO(1). We complement these results by showing that the
choice of degeneracy as the "above guarantee parameterization" is optimal in the following
sense: For any ε > 0 it is NP-complete to decide whether a connected (2-connected) graph
of degeneracy d has a path (cycle) of length at least (1 + ε)d.
1
Introduction
The classical theorem of Erdos and Gallai
Theorem 1 (Erdos and Gallai [12]). Every graph with n vertices and more than (n − 1)(cid:96)/2
edges ((cid:96) ≥ 2) contains a cycle of length at least (cid:96) + 1.
[12] says that
Recall that a graph G is d-degenerate if every subgraph H of G has a vertex of degree at
most d, that is, the minimum degree δ(H) ≤ d. Respectively, the degeneracy of graph G, is
dg(G) = max{δ(H) H is a subgraph of G}. Since a graph of degeneracy d has a subgraph H
with at least d · V (H)/2 edges, by Theorem 1, it contains a cycle of length at least d + 1. Let
us note that the degeneracy of a graph can be computed in polynomial time, see e.g. [28], and
thus by Theorem 1, deciding whether a graph has a cycle of length at least d + 1 can be done in
polynomial time. In this paper we revisit this classical result from the algorithmic perspective.
We define the following problem.
∗The research leading to these results has received funding from the Research Council of Norway via the
projects "CLASSIS" and "MULTIVAL".
†Department of Informatics, University of Bergen, Norway.
‡Department of Computer Science, University of California Santa Barbara, USA.
§Institute of Mathematical Sciences, HBNI, Chennai, India.
¶Ben-Gurion University, Israel.
1
Longest Cycle Above Degeneracy
Input:
Task:
A graph G and a positive integer k.
Decide whether G contains a cycle of length at least dg(G) + k.
Let us first sketch why Longest Cycle Above Degeneracy is NP-complete for k = 2
even for connected graphs. We can reduce Hamiltonian Cycle to Longest Cycle Above
Degeneracy with k = 2 as follows. For a connected non-complete graph G on n vertices, we
construct connected graph H from G and a complete graph Kn−1 on n−1 vertices as follows. We
identify one vertex of G with one vertex of Kn−1. Thus the obtained graph H has V (G) + n− 2
vertices and is connected; its degeneracy is n − 2. Then H has a cycle with dg(H) + 2 = n
vertices if and only if G has a Hamiltonian cycle.
Interestingly, when the input graph is 2-connected, the problem becomes fixed-parameter
tractable being parameterized by k. Let us remind that a connected graph G is (vertex) 2-
connected if for every v ∈ V (G), G − v is connected. Our first main result is the following
theorem.
Theorem 2. On 2-connected graphs Longest Cycle Above Degeneracy is solvable in
time 2O(k) · nO(1).
Similar results can be obtained for paths. Of course, if a graph contains a cycle of length
d + 1, it also contains a simple path on d + 1 vertices. Thus for every graph G of degeneracy
d, deciding whether G contains a path on dg(G) + 1 vertices can be done in polynomial time.
Again, it is a easy to show that it is NP-complete to decide whether G contains a path with d+2
vertices by reduction from Hamiltonian Path. The reduction is very similar to the one we
sketched for Longest Cycle Above Degeneracy. The only difference that this time graph
H consists of a disjoint union of G and Kn−1. The degeneracy of H is d = n − 2, and H has
a path with d + 2 = n vertices if and only if G contains a Hamiltonian path. Note that graph
H used in the reduction is not connected. However, when the input graph G is connected, the
complexity of the problem change drastically. We define
Longest Path Above Degeneracy
Input:
Task:
A graph G and a positive integer k.
Decide whether G contains a path with at least dg(G) + k vertices.
The second main contribution of our paper is the following theorem.
Theorem 3. On connected graphs Longest Path Above Degeneracy is solvable in time
2O(k) · nO(1).
Let us remark that Theorem 2 does not imply Theorem 3, because Theorem 2 holds only
for 2-connected graphs.
We also show that the parameterization lower bound dg(G) that is used in Theorems 3 and
2 is tight in some sense. We prove that for any 0 < ε < 1, it is NP-complete to decide whether
a connected graph G contains a path with at least (1 + ε)dg(G) vertices and it is NP-complete
to decide whether a 2-connected graph G contains a cycle with at least (1 + ε)dg(G) vertices.
Related work. Hamiltonian Path and Hamiltonian Cycle problems are among the oldest
and most fundamental problems in Graph Theory. In parameterized complexity the following
generalizations of these problems, Longest Path and Longest Cycle, we heavily studied.
The Longest Path problem is to decide, for given an n-vertex (di)graph G and an integer k,
whether G contains a path of length at least k. Similarly, the Longest Cycle problem is to
2
decide whether G contains a cycle of length at least k. There is a plethora of results about
parameterized complexity (we refer to the book of Cygan at al. [10] for the introduction to the
field) of Longest Path and Longest Cycle (see, e.g., [5, 6, 8, 7, 13, 15, 22, 23, 24, 33]) since
the early work of Monien [29]. The fastest known randomized algorithm for Longest Path on
undirected graph is due to Bjorklund et al. [5] and runs in time 1.657k · nO(1). On the other
hand very recently, Tsur gave the fastest known deterministic algorithm for the problem running
in time 2.554k · nO(1) [32]. Respectively for Longest Cycle, the current fastest randomized
algorithm runs in time 4knO(1) was given by Zehavi in [34] and the best deterministic algorithm
constructed by Fomin et al. in [14] runs in time 4.884knO(1).
Our theorems about Longest Path Above Degeneracy and Longest Cycle Above
Degeneracy fits into an interesting trend in parameterized complexity called "above guaran-
tee" parameterization. The general idea of this paradigm is that the natural parameterization
of, say, a maximization problem by the solution size is not satisfactory if there is a lower bound
for the solution size that is sufficiently large. For example, there always exists a satisfying
assignment that satisfies half of the clauses or there is always a max-cut containing at least
half the edges. Thus nontrivial solutions occur only for the values of the parameter that are
above the lower bound. This indicates that for such cases, it is more natural to parameterize
the problem by the difference of the solution size and the bound. The first paper about above
guarantee parameterization was due to Mahajan and Raman [26] who applied this approach
to the Max Sat and Max Cut problem. This approach was successfully applied to various
problems, see e.g. [1, 9, 17, 18, 19, 20, 25, 27].
For Longest Path, the only successful above guarantee parameterization known prior to
our work was parameterization above shortest path. More precisely, let s, t be vertices of an
undirected graph G. Clearly, the length of any (s, t)-path in G is lower bounded by the shortest
distance, d(s, t), between these vertices. Based on this observation, Bez´akov´a et al.
in [4]
introduced the Longest Detour problem that asks, given a graph G, two vertices s, t, and a
positive integer k, whether G has an (s, t)-path with at least d(s, t) + k vertices. They proved
that for undirected graphs, this problem can be solved in time 2O(k)nO(1). On the other hand,
the parameterized complexity of Longest Detour on directed graphs is still open. For the
variant of the problem where the question is whether G has an (s, t)-path with exactly d(s, t)+k
vertices, a randomized algorithm with running time O∗(2.746k) and a deterministic algorithm
with running time O∗(6.745k) were obtained [4]. These algorithms work for both undirected and
directed graphs. Parameterization above degeneracy is "orthogonal" to the parameterization
above the shortest distance. There are classes of graphs, like planar graphs, that have constant
degeneracy and arbitrarily large diameter. On the other hand, there are classes of graphs, like
complete graphs, of constant diameter and unbounded degeneracy.
1.1 Our approach
Our algorithmic results are based on classical theorems of Dirac [11], and Erdos and Gallai [12]
on the existence of "long cycle" and "long paths" and can be seen as non-trivial algorithmic
extensions of these classical theorems. Let δ(G) be the minimum vertex degree of graph G.
Theorem 4 (Dirac [11]). Every n-vertex 2-connected graph G with minimum vertex degree
δ(G) ≥ 2, contains a cycle with at least min{2δ(G), n} vertices.
Theorem 5 (Erdos and Gallai [12]). Every connected n-vertex graph G contains a path with at
least min{2δ(G) + 1, n} vertices.
Theorem 4 is used to prove Theorem 2 and Theorem 5 is used to prove Theorem 3.
We give a high-level overview of the ideas used to prove Theorem 2. The ideas behind the
proof of Theorem 3 are similar. Let G be a 2-connected graph of degeneracy d. If d = O(k),
3
we can solve Longest Cycle Above Degeneracy in time 2O(k) · nO(1) by making of use one
of the algorithms for Longest Cycle. Assume from now that d ≥ c · k for some constant c,
which will be specified in the proof. Then we find a d-core H of G (a connected subgraph of
G with the minimum vertex degree at least d). This can be done in linear time by one of the
known algorithms, see e.g. [28]. If the size of H is sufficiently large, say V (H) ≥ d + k, we use
Theorem 4 to conclude that H contains a cycle with at least V (H) ≥ d + k vertices.
The most interesting case occurs when V (H) < d + k. Suppose that G has a cycle of length
at least d + k. It is possible to prove that then there is also a cycle of length at leat d + k that it
hits the core H. We do not know how many times and in which vertices of H this cycle enters
and leaves H, but we can guess these terminal points. The interesting property of the core H is
that, loosely speaking, for any "small" set of terminal points, inside H the cycle can be rerouted
in a such way that it will contain all vertices of H.
A bit more formally, we prove the following structural result. We define a system of segments
in G with respect to V (H), which is a family of internally vertex-disjoint paths {P1, . . . , Pr}
in G (see Figure 1). Moreover, for every 1 ≤ i ≤ r, every path Pi has at least 3 vertices, its
endpoints are in V (H) and all internal vertices of Pi are in V (G) \ V (H). Also the union of all
the segments is a forest with every connected component being a path.
P1
Pr
H
Figure 1: Reducing Longest Cycle Above Degeneracy to finding a system of segments
P1, . . . , Pr; complementing the segments into a cycle is shown by dashed lines.
We prove that G contains a cycle of length at least k + d if and only if
• either there is a path of length at least k + d − V (H) with endpoints in V (H) and all
internal vertices outside H, or
• there is a system of segments with respect to V (H) such that the total number of vertices
outside H used by the paths of the system, is within the interval [k + d − V (H), 2 · (k +
d − V (H))].
The proof of this structural result is built on Lemma 1, which describes the possibility of routing
in graphs of large minimal degree. The crucial property is that we can complement any system
of segments of bounded size by segments inside the core H to obtain a cycle that contains all
the vertices of H as is shown in Figure 1.
Since V (H) > d, the problem of finding a cycle of length at least k + d in G boils down to
one of the following tasks. Either find a path of length c(cid:48) · k with all internal vertices outside
H, or find a system of segments with respect to V (H) such that the total number of vertices
used by the paths of the system is c(cid:48)(cid:48) · k, here c(cid:48) and c(cid:48)(cid:48) are the constants to be specified in the
proof. In the first case, we can use one of the known algorithms to find in time 2O(k) · nO(1)
such a long path. In the second case, we can use color-coding to solve the problem.
Organization of the paper. In Section 2 we give basic definitions and state some known
fundamental results. Sections 3 -- 4 contain the proof of Theorems 3 and 2.
In Section 3 we
state structural results that we need for the proofs and in Section 4 we complete the proofs. In
Section 5, we give the complexity lower bounds for our algorithmic results. We conclude the
paper in Section 6 by stating some open problems.
4
2 Preliminaries
vertices that are adjacent to v in G. For a set U ⊆ V (G), NG(U ) = ((cid:83)
We consider only finite undirected graphs. For a graph G, we use V (G) and E(G) to denote its
vertex set and edge set, respectively. Throughout the paper we use n = V (G) and m = E(G).
For a graph G and a subset U ⊆ V (G) of vertices, we write G[U ] to denote the subgraph of G
induced by U . We write G−U to denote the graph G[V (G)\U ]; for a single-element set U = {u},
we write G− u. For a vertex v, we denote by NG(v) the (open) neighborhood of v, i.e., the set of
v∈S NG(v))\S. The degree
of a vertex v is dG(v) = NG(v). The minimum degree of G is δ(G) = min{dG(v) v ∈ V (G)}.
A d-core of G is an inclusion maximal induced connected subgraph H with δ(H) ≥ d. Every
graph of degeneracy at least d contains a d-core and that can be found in linear time (see [28]). A
vertex u of a connected graph G with at least two vertices is a cut vertex if G−u is disconnected.
A connected graph G is 2-connected if it has no cut vertices. An inclusion maximal induced
2-connected subgraph of G is called a biconnected component or block. Let B be the set of
blocks of a connected graph G and let C be the set of cut vertices. Consider the bipartite graph
Block(G) with the vertex set B ∪ C, where (B, C) is the bipartition, such that B ∈ B and c ∈ C
are adjacent if and only if c ∈ V (B). The block graph of a connected graph is always a tree
(see [21]).
A path in a graph is a self-avoiding walk. Thus no vertex appears in a path more than once.
A cycle is a closed self-avoiding walk . For a path P with end-vertices s and t, we say that
the vertices of V (P ) \ {s, t} are internal. We say that G is a linear forest if each component of
G is a path. The contraction of an edge xy is the operation that removes the vertices x and
y together with the incident edges and replaces them by a vertex uxy that is adjacent to the
vertices of NG({x, y}) of the original graph. If H is obtained from G by contracting some edges,
then H is a contraction of G.
We summarize below some known algorithmic results which will be used as subroutines by
our algorithm.
Proposition 1. Longest Path and Longest Cycle are solvable in time 2O(k) · nO(1).
We also need the result about the variant of Longest Path with fixed end-vertices. In the
(s, t)-Longest Path, we are given two vertices s and t of a graph G and a positive integer k.
The task is to decide, whether G has an (s, t)-path with at least k vertices. Using the results of
Bez´akov´a et al. [3], we immediately obtain the following.
Proposition 2. (s, t)-Longest Path is solvable in time 2O(k) · nO(1).
3 Segments and rerouting
In this section we define systems of segments and prove structural results about them. These
combinatorial results are crucial for our algorithms for Longest Path Above Degeneracy
and Longest Cycle Above Degeneracy.
The following rerouting lemma is crucial for our algorithms.
Lemma 1. Let G be an n-vertex graph and k be a positive integer such that δ(G) ≥ max{5k −
3, n − k}. Let {s1, t1}, . . . ,{sr, tr}, r ≤ k, be a collection of pairs of vertices of G such that (i)
si /∈ {sj, tj} for all i (cid:54)= j, i, j ∈ {1, . . . , r}, and (ii) there is at least one index i ∈ {1, . . . , r}
such that si (cid:54)= ti. Then there is a family of pairwise vertex-disjoint paths P = {P1, . . . , Pr} in G
i=1 V (Pi) = V (G), that is, the paths cover all vertices
such that each Pi is an (si, ti)-path and(cid:83)r
of G.
5
that each Pi has at most 3 vertices. Let T = (cid:83)r
i=1 V (Pi) = V (G), then the lemma holds. Assume that (cid:83)r
Proof. We prove the lemma in two steps. First we show that there exists a family P(cid:48) of pairwise
vertex-disjoint paths connecting all pairs {si, ti}. Then we show that if the paths of P(cid:48) do not
cover all vertices of G, it is possible to enlarge a path such that the new family of paths covers
more vertices.
We start by constructing a family of vertex-disjoint paths P(cid:48) = {P1, . . . , Pr} in G such
that each Pi ∈ P(cid:48) is an (si, ti)-path. We prove that we can construct paths in such a way
i=1{si, ti} and S = V (G) \ T . Notice that
S ≥ n− 2k ≥ δ(G) + 1− 2k ≥ 3k− 2. We consecutively construct paths of P(cid:48) for i ∈ {1, . . . , r}.
If si = ti, then we have a trivial (si, ti)-path. If si and ti are adjacent, then edge siti forms
an (si, ti)-path with 2 vertices. Assume that si (cid:54)= ti and siti /∈ E(G). The already constructed
paths contain at most r− 1 ≤ k− 1 vertices of S in total. Hence, there is a set S(cid:48) ⊆ S of at least
2k−1 of vertices that are not contained in any of already constructed paths. Since δ(G) ≥ n−k,
each vertex of G has at most k − 1 non-neighbors in G. By the pigeonhole principle, there is
v ∈ S(cid:48) such that siv, tiv ∈ E(G). Then we can construct the path Pi = sivti.
We proved that there is a family P(cid:48) = {P1, . . . , Pr} of vertex-disjoint (si, ti)-paths in G.
Among all such families, let us select a family P = {P1, . . . , Pr} covering the maximum number
i=1 V (Pi) <
i=1 V (Pi) ≤ 3k − 1. Since si (cid:54)= ti for some i, there is an edge uv in one of
the paths. Since n ≥ δ(G) + 1 ≥ 5k − 2, there are at least 2k − 1 vertices uncovered by paths of
P. Since δ(G) ≥ n − k, each vertex of G has at most k − 1 non-neighbors in G. Thus there is
i=1 V (Pi)) adjacent to both u and v. But then we can extend the path containing
uv by replacing uv by the path uwv. The paths of the new family cover more vertices than the
paths of P, which contradicts the choice of P.
i=1 V (Pi) ≥ 3k. Because the paths of P are vertex-disjoint, the union of edges
of paths from P contains a k-matching. That is, there are k edges u1v1, . . . , ukvk of G such
that for every i ∈ {1, . . . , k}, vertices ui, vi are consecutive in some path from P and ui (cid:54)= uj,
i=1 V (Pi)). We again use the
observation that w has at most k − 1 non-neighbors in G and, therefore, there is j ∈ {1, . . . , k}
such that ujw, vjw ∈ E(G). Then we extend the path containing ujvj by replacing edge ujvj
by the path ujwvj, contradicting the choice of P. We conclude that the paths of P cover all
vertices of G.
of vertices of V (G). If(cid:83)r
V (G). Suppose (cid:83)r
w ∈ V (G)\ ((cid:83)r
Suppose (cid:83)r
ui (cid:54)= vj for all non-equal i, j ∈ {1, . . . , k}. Let w ∈ V (G) \ ((cid:83)r
Let G be a graph and let T ⊂ V (G) be a set of terminals. We need the following definitions.
Definition 1 (Terminal segments). We say that a path P in G is a one-terminal T -segment if
it has at least two vertices, exactly one end-vertex of P is in T and other vertices are not in T .
Respectively, P is a two-terminal T -segment if it has at least three vertices, both end-vertices of
P are in T and internal vertices of P are not in T .
For every cycle C hitting H, removing the vertices of H from C turns it into a set of
two-terminal T -segments for T = V (H). So here is the definition.
Definition 2 (System of T -segments). We say that a set {P1, . . . , Pr} of paths in G is a system
of T -segments if it satisfies the following conditions.
(i) For each i ∈ {1, . . . , r}, Pi is a two-terminal T -segment,
(ii) P1, . . . , Pr are internally vertex-disjoint, and
(iii) the union of P1, . . . , Pr is a linear forest.
Let us remark that we do not require that the end-vertices of the paths {P1, . . . , Pr} cover
all vertices of T . System of segments will be used for solving Longest Cycle Above Degen-
eracy.
6
For Longest Path Above Degeneracy we need to modify the definition of a system of
T -segments to include the possibility that path can start or end in H.
Definition 3 (Extended system of T -segments). We say that a set {P1, . . . , Pr} of paths in G
is an extended system of T -segments if the following holds.
(i) At least one and at most two paths are one-terminal T -segments and the other are two-
terminal T -segments.
(ii) P1, . . . , Pr are internally vertex-disjoint and the end-vertices of each one-terminal segment
that is in V (G) \ T is pairwise distinct with the other vertices of the paths.
(iii) The union of P1, . . . , Pr is a linear forest and if {P1, . . . , Pr} contains two one-terminal
segments, then the vertices of these segments are in distinct components of the forest.
The following lemma will be extremely useful for the algorithm solving Longest Path
Above Degeneracy. Informally, it shows that if a connected graph G is of large degeneracy
but has a small core H, then deciding whether G has a path of length k + d can be reduced to
checking whether G has an extended system of T -segments with terminal set T = V (H) such
that the total number of vertices used by the system is O(k).
Lemma 2. Let d, k ∈ N. Let G be a connected graph with a d-core H such that d ≥ 5k − 3 and
d > V (H) − k. Then G has a path on d + k vertices if and only if G has an extended system
of T -segments {P1, . . . , Pr} with terminal set T = V (H) such that the total number of vertices
contained in the paths of the system in V (G) \ V (H) is p = d + k − V (H).
Proof. We put T = V (H). Suppose first that G has an extended system {P1, . . . , Pr} of T -
segments and that the total number of vertices of the paths in the system outside T is p =
d + k − T. Let si and ti be the end-vertices of Pi for ∈ {1, . . . , r} and assume without loss
of generality that for 1 ≤ i < j ≤ r, the vertices of Pi and Pj are pairwise distinct with the
possible exception ti = sj when i = j − 1. We also assume without loss of generality that P1 is
a one-terminal segment and t1 ∈ T and if {P1, . . . , Pr} has two one-terminal segments, then the
second such segment is Pr and sr ∈ T .
Suppose that {P1, . . . , Pr} contains one one-terminal segment P1. Let sr+1 be an arbitrary
i=1 V (Pi)) ≤
2p − 1 ≤ 2k − 1 and T ≥ d + 1 ≥ 5k − 3. Consider the collection of pairs of vertices
{t1, s2},{t2, s3}, . . . ,{tr, sr+1}. Notice that vertices from distinct pairs are distinct and tr (cid:54)=
sr+1. By Lemma 1, there are vertex-disjoint paths P (cid:48)
1, . . . , P (cid:48)
i is
a (ti, si+1)-path for i ∈ {1, . . . , r}. By concatenating P1, P (cid:48)
r we obtain a path in
G with T + p = d + k vertices.
Assume now that {P1, . . . , Pr} contains two one-terminal segments P1 and Pr. Consider the
collection of pairs of vertices {t1, s2}, . . . ,{tr−1, sr}. Notice that vertices from distinct pairs are
distinct and there is i ∈ {2, . . . , r} such that ti−1 (cid:54)= si by the condition (iii) of the definition
of an extended system of segments. By Lemma 1, there are vertex-disjoint paths P (cid:48)
r−1
i is a (ti, si+1)-path for i ∈ {1, . . . , r − 1}. By concatenating
in H that cover T such that P (cid:48)
P1, P (cid:48)
i=1 V (Pi)). Notice that such a vertex exists, because T ∩ ((cid:83)r
r in H that cover T such that P (cid:48)
1, P2, . . . , Pr, P (cid:48)
vertex of T \ ((cid:83)r
1, . . . , P (cid:48)
1, . . . , P (cid:48)
To show the implication in the opposite direction, let us assume that G has and (x, y)-path
r−1, Pr we obtain a path in G with T + p = d + k vertices.
P with d + k vertices. We distinguish several cases.
Case 1: V (P ) ∩ T = ∅. Consider a shortest path P (cid:48) with one end-vertex s ∈ V (P ) and the
second end-vertex t ∈ T . Notice that such a path exists, because G is connected. Denote by
Px and Py the (s, x) and (s, y)-subpaths of P respectively. Because d ≥ 5k − 3, V (Px) ≥ k or
V (Py) ≥ k. Assume that V (Px) ≥ k. Then the concatenation of P (cid:48) and Px is a path with at
7
least k + 1 vertices and it contains a subpath P (cid:48)(cid:48) with the end-vertex t with p + 1 vertices. We
have that {P (cid:48)} is an extended system of T -segments and P (cid:48)(cid:48) has p vertices outside T .
Case 2: V (P ) ∩ T (cid:54)= ∅ and E(P ) ∩ E(H) = ∅. Let S = V (P ) ∩ T . Note that k > p,
because V (H) > d. Since H is an induced subgraph of G and E(P ) ∩ E(H) = ∅, V (P ) \ S ≥
(d + k)/2 − 1 ≥ 3k − 5/2 > 3p − 5/2 ≥ 2p − 2. Then for every t ∈ S, either the (t, x)-subpath
Px of P contains at least p vertices outside T or the (t, y)-subpath Py of P contains at least p
vertices outside T . Assume without loss of generality that Px contains at least p vertices outside
T . Consider the minimal subpath P (cid:48) of Px ending at t such that V (P (cid:48)) \ T = p. Then the
start vertex s of P (cid:48) is not in T . Let {t1, . . . , tr} = V (P (cid:48)) ∩ T and assume that t1, . . . , tr are
ordered in the same order as they occur in P (cid:48) starting from s. In particular, tr = t. Let t0 = s.
Consider the paths P1, . . . , Pr where Pi is the (ti−1, ti)-subpath of P (cid:48) for i ∈ {1, . . . , r}. Since
k ≥ p, r ≤ k. We obtain that {P1, . . . , Pr} is an extended system of T -segments with p vertices
outside T .
Case 3: E(P )∩ E(H) (cid:54)= ∅. Then there are distinct s, t ∈ T ∩ V (P ) such that the (s, t)-subpath
of P lies in H. Since P has at least p vertices outside T , there are s(cid:48), t(cid:48) ∈ V (P ) \ T such that
the (s(cid:48), t(cid:48))-subpath P (cid:48) of P is a subpath with exactly p vertices outside T with s, t ∈ V (P (cid:48)).
Let P1, . . . , Pr be the family of inclusion maximal subpaths of P (cid:48) containing the vertices of
V (P (cid:48)) \ T such that the internal vertices of each Pi are outside T . The set {P1, . . . , Pr} is a
required extended system of T -segments.
The next lemma will be used for solving Longest Cycle Above Degeneracy.
Lemma 3. Let d, k ∈ N. Let G be a 2-connected graph with a d-core H such that d ≥ 5k − 3
and d > V (H) − k. Then G has a cycle with at least d + k vertices if and only if one of the
following holds (where p = d + k − V (H)).
(i) There are distinct s, t ∈ V (H) and an (s, t)-path P in G with all internal vertices outside
V (H) such that P has at least p internal vertices.
(ii) G has a system of T -segments {P1, . . . , Pr} with terminal set T = V (H) and the total
number of vertices of the paths outside V (H) is at least p and at most 2p − 2.
Proof. We put T = V (H). First, we show that if (i) or (ii) holds, then G has a cycle with at
least d + k vertices. Suppose that there are distinct s, t ∈ T and an (s, t)-path P in G with all
internal vertices outside T such that P has at least p internal vertices. By Lemma 1, H has a
Hamiltonian (s, t)-path P (cid:48). By taking the union of P and P (cid:48) we obtain a cycle with at least
T + p = d + k vertices.
Now assume that G has a system of T -segments {P1, . . . , Pr} and the total number of vertices
of the paths outside T is at least p. Let si and ti be the end-vertices of Pi for i ∈ {1, . . . , r} and
assume without loss of generality that for 1 ≤ i < j ≤ r, the vertices of Pi and Pj are pairwise
distinct with the possible exception ti = sj when i = j − 1. Consider the collection of pairs of
vertices {t1, s2}, . . . ,{tr−1, sr},{tr, s1}. Notice that vertices from distinct pairs are distinct and
tr (cid:54)= s1. By Lemma 1, there are vertex-disjoint paths P (cid:48)
r in H that cover T such that
i is a (ti, si+1)-path for i ∈ {1, . . . , r − 1} and P (cid:48)
P (cid:48)
r is a (tr, s1)-path. By taking the union of
P1, . . . , Pr and P (cid:48)
r we obtain a cycle in G with at least T + p = d + k vertices.
1, . . . , P (cid:48)
1, . . . , P (cid:48)
To show the implication in the other direction, assume that G has a cycle C with at least
d + k vertices.
Case 1: V (C) ∩ T = ∅. Since G is a 2-connected graph, there are pairwise distinct vertices
s, t ∈ T and x, y ∈ V (C) and vertex-disjoint (s, x) and (y, t)-paths P1 and P2 such that the
internal vertices of the paths are outside T ∪ V (C). The cycle C contains an (x, y)-path P with
at least (d + k)/2 + 1 ≥ p vertices. The concatenation of P1, P and P2 is an (s, t)-path in G
with at least p internal verices and the internal vertices are outside T . Hence, (i) holds.
8
Case 2: V (C) ∩ T = 1. Let V (C) ∩ T = {s} for some vertex s. Since G is 2-connected, there
is a shortest (x, t)-path P in G − s such that x ∈ V (C) and t ∈ T . The cycle C contains an
(s, x)-path P (cid:48) with at least (d + k)/2 + 1 ≥ p vertices. The concatenation of P (cid:48) and P is an
(s, t)-path in G with at least p internal vertices and the internal vertices of the path are outside
T . Therefore, (i) is fulfilled.
Case 3: V (C) ∩ T ≥ 2. Since V (C) ≥ d and T < d, we have that V (C) \ T (cid:54)= ∅. Then
we can find pairs of distinct vertices {s1, t1} . . . ,{s(cid:96), t(cid:96)} of T ∩ V (C) and segments P1, . . . , P(cid:96)
of C such that (a) Pi is an (si, ti)-path for i ∈ {1, . . . , (cid:96)} with at least one internal vertex and
the internal vertices of Pi are outside T , (b) for 1 ≤ i < j ≤ (cid:96), the vertitces of Pi and Pj
(cid:83)(cid:96)
are distinct with the possible exception ti = sj if i = j − 1 and, possibly, t(cid:96) = s1, and (c)
i=1 V (Pi)\ T = V (C)\ T . If there is i ∈ {1, . . . , (cid:96)} such that Pi has at least p internal vertices,
We select an inclusion minimal set of indices I ⊆ {1, . . . , (cid:96)} such that (cid:83)
then (i) is fulfilled.
Now assume that each Pi has at most p − 1 internal vertices; notice that p ≥ 2 in this case.
Notice that because each path has at most p− 1 internal vertices, (cid:83)
i∈I V (Pi) \ T ≥ p.
i∈I V (Pi)\ T ≤ 2p− 2. Let
I = {i1, . . . , ir} and i1 < . . . < ir. By the choice of Pi1, . . . , Pir , the union of Pi1, . . . , Pir is either
the cycle C or a linear forest. Suppose that the union of the paths is C. Then I = {1, . . . , (cid:96)},
(cid:96) ≤ p and V (P )∩ T = (cid:96). We obtain that C has at most (2p− 2) + p ≤ 3p− 2 ≤ 3k − 2 < d + k
vertices (the last inequality follows from the fact that d ≥ 5k − 3); a contradiction. Hence, the
union of the paths is a linear forest. Therefore, {Pi1, . . . , Pir} is a system of T -segments with
terminal set T = V (H) and the total number of vertices of the paths outside T is at least p and
at most 2p − 2, that is, (ii) is fulfilled.
We have established the fact that existence of long (path) cycle is equivalent to the existence
of (extended) system of T -segments for some terminal set T with at most p ≤ k vertices
from outside T . Towards designing algorithms for Longest Path Above Degeneracy and
Longest Cycle Above Degeneracy, we define two auxiliary problems which can be solved
using well known color-coding technique.
Segments with Terminal Set
A graph G, T ⊂ V (G) and a positive integers p and r.
Decide whether G has a system of segments {P1, . . . , Pr} w.r.t. T such
that the total number of internal vertices of the paths is p.
Input:
Task:
Input:
Task:
Extended Segments with Terminal Set
A graph G, T ⊂ V (G) and a positive integers p and r.
Decide whether G has an extended system of segments {P1, . . . , Pr}
w.r.t. T such that the total number of vertices of the paths outside T
is p.
Lemma 4. Segments with Terminal Set and Extended Segments with Terminal
Set are solvable in time 2O(p) · nO(1).
Proof. We start with the algorithm for Segments with Terminal Set. Then we show how
to modify it for Extended Segments with Terminal Set. Our algorithm uses the color
coding technique introduced by Alon, Yuster and Zwick in [2]. As it is usual for algorithms of
this type, we first describe a randomized Monte-Carlo algorithm and then explain how it could
be derandomized.
Let (G, T, p, r) be an instance of Segments with Terminal Set.
9
Notice that if paths P1, . . . , Pr are a solution for the instance, that is, {P1, . . . , Pr} is a system
of T -segments and the total number of internal vertices of the paths is p, then ∪r
i=1 V (Pi) ≤
p + 2r. If r > p, then because each path in a solution should have at least one internal vertex,
(G, T, p, r) is a no-instance. Therefore, we can assume without loss of generality that r ≤ p. Let
q = p + 2r ≤ 3p. We color the vertices of G with q colors uniformly at random. Let P1, . . . , Pr
be paths and G and let si, ti be the end-vertices of Pi for i ∈ {1, . . . , r}. We say that the paths
P1, . . . , Pr together with the ordered pairs (si, ti) of their end-vertices form a colorful solution
if the following is fulfilled:
(i) {P1, . . . , Pr} is a system of T -segments,
(ii) ∪r
(iii) if 1 ≤ i < j ≤ r, u ∈ V (Pi) and v ∈ V (Pj), then the vertices u and v have distinct colors
i=1 V (Pi) \ T = p,
unless i = j − 1, u = ti and v = sj (in this case the colors can be distinct or same).
It is straightforward to see that any colorful solution is a solution of the original problem. From
qq > e−q all
the other side, if (G, T, p, r) has a solution P1, . . . , Pr, then with probability at least q!
distinct vertices of the paths of a solution are colored by distinct colors and for such a coloring,
P1, . . . , Pr is a colorful solution. Since q ≤ 3p, we have that the probability is lower bounded
by e−3p. This shows that if there is no colorful solution, then the probability that (G, T, p, r)
is a yes-instance is at most 1 − e−3p. It immediately implies that if after trying e3p random
colorings there is no colorful solution for any of them, then the probability that (G, T, p, r) is a
yes-instance is at most (1 − e−3p)e3p < e−1 < 1.
solution. Denote by c : V (G) → {1, . . . , q} the considered random coloring.
We construct a dynamic programming algorithm that decides whether there is a colorful
In the first step of the algorithm, for each non-empty X ⊆ {1, . . . , q} and distinct i, j ∈ X,
we compute the Boolean function α(X, i, j) such that α(X, i, j) = true if and only if there
are s, t ∈ T and an (s, t)-path P such that P is a two-terminal T -segment, V (P ) = X,
c(s) = i, c(t) = j and the vertices of P are colored by pairwise distinct colors from X. We
define α(X, i, j) = f alse if X < 3. For other cases, we use dynamic programming.
We use a dynamic-programming algorithm to compute α(X, i, j). For each v ∈ V (G)\T and
each non-empty Y ⊆ X \ {i}, we compute the Boolean function β(Y, i, v) such that β(Y, i, v) =
true if and only if there is s ∈ T and an (s, v)-path P (cid:48) such that V (P (cid:48)) \ {s} ⊆ V (G) \ T ,
c(s) = i, V (P )\{s} = Y and the vertices of V (P )\{s} are colored by pairwise distinct colors
from Y .
We compute β(Y, i, v) recursively starting with one-element sets. For every Y = {h}, where
h (cid:54)= i, and every v ∈ V (G)\T , we set β(Y, i, v) = true if c(v) = h and v is adjacent to a vertex of
T colored i, and we set β(Y, i, v) = f alse otherwise. For Y ⊆ {1, . . . , q}\{i} of size at least two,
we set β(Y, v, i) = true if c(v) ∈ Y and there is w ∈ NG(v) \ T with β(i, Y \ {c(v)}, w) = true,
and β(Y, i, v) = f alse otherwise.
We set α(X, i, j) = true if and only if there are t ∈ T and v ∈ NG(t) \ T such that c(t) = j
and β(X \ {i, j}, i, v) = true.
The correctness of computing β and α is proved by standard arguments in a straightforward
way. Notice that we can compute the tables of values of β and α in time 2q · nO(1). First,
we compute the values of β(Y, i, v) for all v ∈ V (G) \ T , i ∈ {1, . . . , q} and non-empty Y ⊆
{1, . . . , q} \ {i}. Then we use the already computed values of β to compute the table of values
of α.
Next, we use the table of values of α to check whether a colorful solution exists. We
introduce the Boolean function γ0(i, X, (cid:96), j) such that for each i ∈ {1, . . . , r}, X ⊆ {1, . . . , q},
integer (cid:96) ≤ p and j ∈ X, γ0(i, X, (cid:96), j) = true if and only if there are paths P1, . . . , Pi and
10
ordered pairs (s1, t1), . . . , (si, ti) of distinct vertices of T such that each Ph is an (sh, th)-path
and the following is fulfilled:
(i) {P1, . . . , Pi} is a system of T -segments,
(ii) ∪i
(iii) if 1 ≤ f < g ≤ i, u ∈ V (Pf ) and v ∈ V (Pg), then the vertices u and v have distinct colors
h=1 V (Ph) \ T = (cid:96),
unless f = g − 1, u = tf and v = sg when the colors could be same,
(iv) c(ti) = j.
Notice, that if (cid:96) < i, then γ0(i, X, (cid:96), j) = f alse. Our aim is to compute γ0(r, X, p, j) for
X ⊆ {1, . . . , q} and j ∈ {1, . . . , q}. Then we observe that a colorful solution exists if and only
if there is X ⊆ {1, . . . , q} and j ∈ {1, . . . , q} such that γ0(r, X, p, j) = true.
If i = 1 and (cid:96) ≥ 1, then
γ0(1, X, (cid:96), j) =(cid:0) (cid:95)
α(X, h, j)(cid:1) ∧(cid:0)X = (cid:96) + 2(cid:1).
h∈X\{j}
For (cid:96) ≥ i > 1, we use the following recurrence:
(1)
(2)
γ0(i, X, (cid:96), j) =(cid:0)
∨(cid:0)
(cid:95)
(cid:95)
(α(Y, h, j) ∧ γ0(i − 1, (X \ Y ) ∪ {h}, (cid:96) − Y + 2, h))(cid:1)
(α(Y, h, j) ∧ γ0(i − 1, X \ Y, (cid:96) − Y + 2, h(cid:48)))(cid:1).
j∈Y ⊂X,h∈Y \{j}
j∈Y ⊂X,h∈Y \{j},h(cid:48)∈X\Y
The correctness of (1) and (2) is proved by the standard arguments. Since the size of the
table of values of α is 2q · nO(1) and the table can be constructed in time 2q · nO(1), we obtain
that the values of γ0(r, X, p, j) for X ⊆ {1, . . . , q} and j ∈ {1, . . . , q} can be computed in time
3q · nO(1). Therefore, the existence of a colorful solution can be checked in time 3q · nO(1).
This leads us to a Monte-Carlo algorithm for Segments with Terminal Set. We try at
most e3p random colorings. For each coloring, we check the existence of a colorful solution. If
such a solution exists, we report that we have a yes-instance of the problem. If after trying e3p
random colorings we do not find a colorful solution for any of them, we return the answer no. As
we already observed, the probability that this negative answer is false is at most (1− e−3p)e3p <
e−1 < 1, that is, the probability is upper bounded by the constant e−1 < 1 that does not depend
on the problem size and the parameter. The running time of the algorithm is (3e)3p · nO(1).
The algorithm can be derandomized, as it was explained in [2] (we also refer to [10] for the
detailed introduction to the technique), by the replacement of random colorings by a family of
perfect hash functions. The currently best explicit construction of such families was done by
Naor, Schulman and Srinivasan in [30]. The family of perfect hash function in our case has size
e3ppO(log p) log n and can be constructed in time e3ppO(log p)n log n [30]. It immediately gives the
deterministic algorithm for Segments with Terminal Set running in time (3e)3ppO(log p) ·
nO(1).
Now we explain how the algorithm for Segments with Terminal Set can be modified
for Extended Segments with Terminal Set.
Let (G, T, p, r) be an instance of Extended Segments with Terminal Set.
If paths P1, . . . , Pr are a solution for the instance, that is, {P1, . . . , Pr} is an extended system
i=1 V (Pi) ≤
of T -segments and the total number of vertices of the paths outside T is p, then ∪r
p + 2r − 1.
If r > p, then because each path in a solution should have at least one vertex
outside T , (G, T, p, r) is a no-instance. Therefore, we can assume without loss of generality that
11
r ≤ p. The total number of distinct vertices of the paths q ∈ {p + r, . . . , p + 2r − 1} and q ≤ 3p.
We guess the value of q and color the vertices of G with q colors uniformly at random. Let
P1, . . . , Pr be paths and G and let si, ti be the end-vertices of Pi for i ∈ {1, . . . , r}. We say that
the paths P1, . . . , Pr together with the ordered pairs (si, ti) of their end-vertices form a colorful
solution if the following is fulfilled:
(i) {P1, . . . , Pr} is an extended system of T -segments,
(ii) if {P1, . . . , Pr} has one one-terminal segment, then this is P1 and t1 ∈ T , and if {P1, . . . , Pr}
has two one-terminal segments, then these are P1, Pr and t1, sr ∈ T ,
i=1 V (Pi) \ T = p,
(iii) ∪r
(iv) if 1 ≤ i < j ≤ r, u ∈ V (Pi) and v ∈ V (Pj), then the vertices u and v have distinct colors
unless i = j − 1, u = ti and v = sj (in this case the colors could be distinct or same), and
if {P1, . . . , Pr} contains two one-terminal segments, then there is i ∈ {2, . . . , r} such that
ti−1 and si have distinct colors.
In the same way as before, any colorful solution is a solution of the original problem and if after
trying e3p random colorings there is no colorful solution for any of them, then the probability
that (G, T, p, r) is a yes-instance is at most (1 − e−3p)e3p < e−1 < 1.
solution. Denote by c : V (G) → {1, . . . , q} the considered random coloring.
We construct a dynamic programming algorithm that decides whether there is a colorful
First, we construct the tables of values of the Boolean functions α and β defined above
exactly in the same way as in the algorithm for Segments with Terminal Set. Now we
consider the following two possibilities.
We check the existence of a colorful solution such that {P1, . . . , Pr} has one one-terminal
segment P1. We introduce the Boolean function γ1(i, X, (cid:96), j) for each i ∈ {1, . . . , r}, X ⊆
{1, . . . , q}, integer (cid:96) ≤ p and j ∈ X such that γ1(i, X, (cid:96), j) = true if and only if there are paths
P1, . . . , Pi and ordered pairs (s1, t1), . . . , (si, ti) of distinct vertices of T such that each Ph is
(sh, th)-path for h ∈ {1, . . . , i} and the following is fulfilled:
(i) {P1, . . . , Pi} is an extended system of T -segments with one one-terminal segment P1 and
t1 ∈ T ,
h=1 V (Ph) \ T = (cid:96),
(ii) ∪i
(iii) if 1 ≤ f < g ≤ i, u ∈ V (Pf ) and v ∈ V (Pg), then the vertices u and v have distinct colors
unless f = g − 1, u = tf and v = sg when the colors could be same,
(iv) c(ti) = j.
As with γ0, γ1(i, X, (cid:96), j) = f alse if (cid:96) < i. A colorful solution exists if and only if there is
X ⊆ {1, . . . , q} and j ∈ {1, . . . , q} such that γ1(r, X, p, j) = true.
If i = 1 and (cid:96) ≥ 1, then
γ1(1, X, (cid:96), j) =(cid:0) (cid:95)
β(X \ {j}, j, v)(cid:1) ∧(cid:0)X = (cid:96) + 1(cid:1).
(3)
(4)
v∈V (G)\T
For (cid:96) ≥ i > 1, we use the same recurrence as (2):
γ1(i, X, (cid:96), j) =(cid:0)
∨(cid:0)
(cid:95)
(cid:95)
(α(Y, h, j) ∧ γ1(i − 1, (X \ Y ) ∪ {h}, (cid:96) − Y + 2, h))(cid:1)
(α(Y, h, j) ∧ γ1(i − 1, X \ Y, (cid:96) − Y + 2, h(cid:48)))(cid:1).
j∈Y ⊂X,h∈Y \{j}
j∈Y ⊂X,h∈Y \{j},h(cid:48)∈X\Y
12
j=1V (Pj) and the vertices of ∪r
Again, it is standard to prove correctness of (3) and (4) and the existence of a colorful solution
can be checked in time 3q · nO(1).
Now we check the existence of a colorful solution such that {P1, . . . , Pr} has two one-terminal
segments P1 and Pr. It is possible to write down a variant of the dynamic programming algo-
rithm tailored for this case, but it is more simple to reduce this case to the already considered.
Recall that we are interested in a colorful solution with the property that there is i ∈ {2, . . . , r}
such that the vertices of ∪i−1
j=iV (Pj) are colored by distinct colors.
We obtain that a colorful solution that we are looking for can be seen as disjoint union of two
partial colorful solutions {P1, . . . , Pi−1} and {Pi, . . . , Pr} such that each of them has one one-
terminal segment. To find them, we use the function γ1 constructed above. We guess the value
of i ∈ {2, . . . , r}. Recall that we are looking for a solution that uses all colors from {1, . . . , q}.
We construct the tables of values of γ1(i− 1, X, (cid:96), j) and γ1(r− i + 1, X(cid:48), (cid:96)(cid:48), j(cid:48)). It remains to ob-
serve that a colorful solution exists if and only if there X ⊆ {1, . . . , q}, j ∈ X, j(cid:48) ∈ {1, . . . , q}\ X
and (cid:96) ∈ {1, . . . , p − 1} such that γ1(i − 1, X, (cid:96), j) ∧ γ1(r − i + 1,{1, . . . , q} \ X, p − (cid:96), j(cid:48)) = true.
This implies that the existence of a colorful solution with two one-terminal segments can be
checked in time 3q · nO(1).
As with Segments with Terminal Set, we obtain the Monte-Carlo algorithm running in
time (2e)3p · nO(1) and then we can derandomize it to obtain the deterministic algorithm with
running time (3e)3ppO(log p) · nO(1).
4 Putting all together: Final proofs
Proof of Theorem 3. Let G be a connected graph of degeneracy at least d and let k be
a positive integer. If d ≤ 5k − 4, then we check the existence of a path with d + k ≤ 6k − 4
vertices using Proposition 1 in time 2O(k) · nO(1). Assume from now that d ≥ 5k − 3. Then
we find a d-core H of G. This can be done in linear time using the results of Matula and
If V (H) ≥ d + k, then by Theorem 5, H, and hence G, contains a path with
Beck [28].
min{2d + 1,V (H)} ≥ d + k vertices. Assume that V (H) < d + k. By Lemma 2, G has a path
with d + k vertices if and only if G has paths P1, . . . , Pr such that {P1, . . . , Pr} is an extended
system of T -segments for T = V (H) and the total number of vertices of the paths outside T is
p = d + k − T. Since the number of vertices in every graph is more than its minimum degree,
we have that T > d, and thus p < k. For each r ∈ {1, . . . , p}, we verify if such a system exists
in time 2O(k) · nO(1) by making use of Lemma 4. Thus the total running time of the algorithm
is 2O(k) · nO(1).
Proof of Theorem 2 Let G be a 2-connected graph of degeneracy at least d and let
k ∈ N. If d ≤ 5k − 4, then we check the existence of a cycle with at least d + k ≤ 6k − 4 vertices
using Proposition 1 in time 2O(k) · nO(1). Assume from now on that d ≥ 5k − 3. Then we find
a d-core H of G in linear time using the results of Matula and Beck [28].
We claim that if V (H) ≥ d + k, then H contains a cycle with at least d + k vertices. If H is
2-connected, then this follows from Theorem 4. Assume that H is not a 2-connected graph. By
the definition of a d-core, H is connected. Observe that V (H) ≥ d + 1 ≥ 5k − 2 ≥ 3. Hence,
H has at least two blocks and at least one cut vertex. Consider the block graph Block(H)
of H. Recall that the vertices of Block(H) are the blocks and the cut vertices of H and a
cut vertex c is adjacent to a block B if and only if c ∈ V (B). Recall also that Block(H) is
a tree. We select an arbitrary block R of H and declare it to be the root of Block(H). Let
S = V (G) \ V (H). Observe that S (cid:54)= ∅, because G is 2-connected and H is not. Let F1, . . . , F(cid:96)
be the components of G[S]. We contract the edges of each component and denote the obtained
vertices by u1, . . . , u(cid:96). Denote by G(cid:48) the obtained graph. It is straightforward to verify that G(cid:48)
13
has no cut vertices, that is, G(cid:48) is 2-connected. For each i ∈ {1, . . . , (cid:96)}, consider ui. This vertex
has at least 2 neighbors in V (H). We select a vertex vi ∈ NG(cid:48)(ui) that is not a cut vertex
of H or, if all the neighbors of ui are cut vertices, we select vi be a cut vertex at maximum
distance from R in Block(H). Then we contract uivi. Observe that by the choice of each vi,
the graph G(cid:48)(cid:48) obtained from G(cid:48) by contracting u1v1, . . . , u(cid:96)v(cid:96) is 2-connected. We have that G(cid:48)(cid:48)
is a 2-connected graph of minimum degree at least d with at least d + k vertices. By Theorem 4,
G(cid:48)(cid:48) has a cycle with at least min{2d,V (G(cid:48)(cid:48))} ≥ d + k vertices. Because G(cid:48)(cid:48) is a contraction of
G, we conclude that G contains a cycle with at least d + k vertices as well.
From now we can assume that V (H) < d + k. By Lemma 3, G has a cycle with d + k
vertices if and only if one of the following holds for p = d + k − T where T = V (H).
(i) There are distinct s, t ∈ T and an (s, t)-path P in G with all internal vertices outside T
such that P has at least p internal vertices.
(ii) G has a system of T -segments {P1, . . . , Pr} and the total number of vertices of the paths
outside T is at least p and at most 2p − 2.
Notice that p ≤ k (because d − T ≤ 0). We verify whether (i) holds using Proposition 2.
To do it, we consider all possible choices of distinct s, t. Then we construct the auxiliary graph
Gst from G by the deletion of the vertices of T \ {s, t} and the edges of E(H). Then we check
whether Gst has an (s, t)-path of length at least p+1 in time 2O(k)·nO(1) applying Proposition 2.
Assume that (i) is not fulfilled. Then it remains to check (ii). For every r ∈ {1, . . . , p}, we
verify the existence of a system of T -segments {P1, . . . , Pr} in time 2O(k) · nO(1) using Lemma 4.
We return the answer yes if we get the answer yes for at least one instance of Segments with
Terminal Set and we return no otherwise.
5 Hardness for Longest Path and Cycle above Degeneracy
In this section we complement Theorems 3 and 2 by some hardness observations.
Proposition 3. 1 Longest Path Above Degeneracy is NP-complete even if k = 2 and
Longest Cycle Above Degeneracy is NP-complete even for connected graphs and k = 2.
Proof. To show that Longest Path Above Degeneracy is NP-complete for k = 2, consider
the graph G(cid:48) that is a disjoint union of a non-complete graph G with n vertices and a copy
the complete (n − 1)-vertex graph Kn−1. Because G is not a complete graph, dg(G(cid:48)) ≤ n − 2.
Therefore, dg(G(cid:48)) = n − 2, because dg(Kn−1) = n − 2. Observe that G(cid:48) has a path with
dg(G(cid:48)) + 2 = n vertices if and only if G is Hamiltonian. Since Hamiltonian Path is a well-
known NP-complete problem (see [16]), the claim follows.
Similarly, for Longest Cycle Above Degeneracy, consider G(cid:48) that is a union of a
connected non-complete graph G with n vertices and Kn−1 with one common vertex. We
have that G(cid:48) has a cycle with dg(G(cid:48)) + 2 = n vertices if and only if G has a Hamiltonian cycle.
Using the fact that Hamiltonian Cycle is NP-complete [16], we obtain that Longest Cycle
Above Degeneracy is NP-complete for connected graphs and k = 2.
Recall that a graph G has a path with at least dg(G) + 1 vertices and if dg(G) ≥ 2, then G
has a cycle with at least dg(G) + 1 vertices. Moreover, such a path or cycle can be constructed
in polynomial (linear) time. Hence, Proposition 3 gives tight complexity bounds. Neverthe-
less, the construction used to show hardness for Longest Path Above Degeneracy uses a
disconnected graph, and the graph constructed to show hardness for Longest Cycle Above
1Proposition 3 and its proof was pointed to us by Nikolay Karpov.
14
Degeneracy has a cut vertex. Hence, it is natural to consider Longest Path Above De-
generacy for connected graphs and Longest Cycle Above Degeneracy for 2-connected
graphs. We show in Theorems 3 and 2 that these problems are FPT when parameterized by k in
these cases. Here, we observe that the lower bound dg(G) that is used for the parameterization
is tight in the following sense.
Proposition 4. For any 0 < ε < 1, it is NP-complete to decide whether a connected graph G
contains a path with at least (1 + ε)dg(G) vertices and it is NP-complete to decide whether a
2-connected graph G contains a cycle with at least (1 + ε)dg(G) vertices.
Proof. Let 0 < ε < 1.
First, we consider the problem about a path with (1 + ε)dg(G) vertices. We reduce Hamil-
tonian Path that is well-known to be NP-complete (see [16]). Let G be a graph with n ≥ 2
vertices. We construct the graph G(cid:48) as follows.
• Construct a copy of G.
• Let p = 2(cid:100) n
• For each v ∈ V (G), construct an edge vu1.
• Let q = (cid:100)(1 + ε)(p − 1) − (n + p)(cid:101). Construct vertices w1, . . . , wq and edges u1w1, wqu2
ε(cid:101) and construct p pairwise adjacent vertices u1, . . . , up.
and wi−1wi for i ∈ {2, . . . , q}.
ε(cid:101) − n − 1 − ε(cid:101) ≥ (cid:100)n − 1 − ε(cid:101) ≥ 1 as
Notice that q = (cid:100)(1 + ε)(p − 1) − (n + p)(cid:101) = (cid:100)2ε(cid:100) n
n ≥ 2. Observe also that G is connected. We claim that G has a Hamiltonian path if and
only if G(cid:48) has a path with at least (1 + ε)dg(G(cid:48)) vertices. Notice that dg(G(cid:48)) = p − 1 and
V (G(cid:48)) = n + p + q = (cid:100)(1 + ε)dg(G(cid:48))(cid:101). Therefore, we have to show that G has a Hamiltonian
path if and only if G(cid:48) has a Hamiltonian path. Suppose that G has a Hamiltonian path P with
an end-vertex v. Consider the path Q = vu1w1 . . . wqu2u3 . . . up. Clearly, the concatenation of
P and Q is a Hamiltonian path in G(cid:48). Suppose that G(cid:48) has a Hamiltonian path P . Since u1 is
a cut vertex of G(cid:48), we obtain that P has a subpath that is a Hamiltonian path in G.
Consider now the problem about a cycle with at least (1 + ε)dg(G) vertices. We again
reduce Hamiltonian Path and the reduction is almost the same. Let G be a graph with n ≥ 2
vertices. We construct the graph G(cid:48) as follows.
• Construct a copy of G.
• Let p = 2(cid:100) n
• For each v ∈ V (G), construct edges vu1 and vu2.
• Let q = (cid:100)(1 + ε)(p − 1) − (n + p)(cid:101). Construct vertices w1, . . . , wq and edges u2w1, wqu3
ε(cid:101) and construct p pairwise adjacent vertices u1, . . . , up.
and wi−1wi for i ∈ {2, . . . , q}.
As before, we have that q ≥ 1. Notice additionally that p ≥ 3, i.e., the vertex u3 exists. It is
straightforward to see that G(cid:48) is 2-connected. We claim that G has a Hamiltonian path if and
only if G(cid:48) has a cycle with at least (1 + ε)dg(G(cid:48)) vertices. We have that dg(G(cid:48)) = p − 1 and
V (G(cid:48)) = (cid:100)(1 + ε)dg(G(cid:48))(cid:101). Hence, we have to show that G has a Hamiltonian path if and only
if G(cid:48) has a Hamiltonian cycle. Suppose that G has a Hamiltonian path P with end-vertices
x and y. Consider the path Q = xu2w1 . . . wqu3u4 . . . upy. Clearly, P and Q together form a
Hamiltonian cycle in G(cid:48). Suppose that G(cid:48) has a Hamiltonian cycle C. Since {u1, u2} is a cut
set of G(cid:48), we obtain that C contains a path that is a Hamiltonian path of G.
15
6 Conclusion
We considered the lower bound dg(G) + 1 for the number of vertices in a longest path or cycle
in a graph G. It would be interesting to consider the lower bounds given in Theorems 4 and 5.
More precisely, what can be said about the parameterized complexity of the variants of Long
Path (Cycle) where given a (2-connected) graph G and k ∈ N, the task is to check whether G
has a path (cycle) with at least 2δ(G)+k vertices? Are these problems FPT when parameterized
by k? It can be observed that the bound 2δ(G) is "tight". That is, for any 0 < ε < 1, it is
NP-complete to decide whether a connected (2-connected) G has a path (cycle) with at least
(2 + ε)δ(G) vertices. See also [31] for related hardness results.
Acknowledgement We thank Nikolay Karpov for communicating to us the question of find-
ing a path above the degeneracy bound and Proposition 3.
References
[1] N. Alon, G. Gutin, E. J. Kim, S. Szeider, and A. Yeo, Solving MAX-r-SAT above a
tight lower bound, in Proceedings of the 21st Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), SIAM, 2010, pp. 511 -- 517. 3
[2] N. Alon, R. Yuster, and U. Zwick, Color-coding, J. ACM, 42 (1995), pp. 844 -- 856. 9,
11
[3] I. Bez´akov´a, R. Curticapean, H. Dell, and F. V. Fomin, Finding detours is fixed-
parameter tractable, CoRR, abs/1607.07737 (2016). 5
[4]
, Finding detours is fixed-parameter tractable, in 44th International Colloquium on
Automata, Languages, and Programming, ICALP 2017, vol. 80 of LIPIcs, Schloss Dagstuhl
- Leibniz-Zentrum fuer Informatik, 2017, pp. 54:1 -- 54:14. 3
[5] A. Bjorklund, T. Husfeldt, P. Kaski, and M. Koivisto, Narrow sieves for param-
eterized paths and packings, CoRR, abs/1007.1161 (2010). 3
[6] H. L. Bodlaender, On linear time minor tests with depth-first search, J. Algorithms, 14
(1993), pp. 1 -- 23. 3
[7] J. Chen, J. Kneis, S. Lu, D. Molle, S. Richter, P. Rossmanith, S.-H. Sze, and
F. Zhang, Randomized divide-and-conquer: improved path, matching, and packing algo-
rithms, SIAM J. Comput., 38 (2009), pp. 2526 -- 2547. 3
[8] J. Chen, S. Lu, S.-H. Sze, and F. Zhang, Improved algorithms for path, matching, and
packing problems, in Proceedings of the18th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), SIAM, 2007, pp. 298 -- 307. 3
[9] R. Crowston, M. Jones, G. Muciaccia, G. Philip, A. Rai, and S. Saurabh,
Polynomial kernels for lambda-extendible properties parameterized above the Poljak-Turzik
bound, in IARCS Annual Conference on Foundations of Software Technology and Theoreti-
cal Computer Science (FSTTCS), vol. 24 of Leibniz International Proceedings in Informat-
ics (LIPIcs), Dagstuhl, Germany, 2013, Schloss Dagstuhl -- Leibniz-Zentrum fuer Informatik,
pp. 43 -- 54. 3
[10] M. Cygan, F. V. Fomin, (cid:32)L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk,
M. Pilipczuk, and S. Saurabh, Parameterized Algorithms, Springer, 2015. 3, 11
16
[11] G. A. Dirac, Some theorems on abstract graphs, Proc. London Math. Soc. (3), 2 (1952),
pp. 69 -- 81. 3
[12] P. Erdos and T. Gallai, On maximal paths and circuits of graphs, Acta Math. Acad.
Sci. Hungar, 10 (1959), pp. 337 -- 356 (unbound insert). 1, 3
[13] F. V. Fomin, D. Lokshtanov, F. Panolan, and S. Saurabh, Efficient computation
of representative families with applications in parameterized and exact algorithms, J. ACM,
63 (2016), pp. 29:1 -- 29:60. 3
[14] F. V. Fomin, D. Lokshtanov, F. Panolan, S. Saurabh, and M. Zehavi, Long
directed ( s, t)-path: FPT algorithm, Inf. Process. Lett., 140 (2018), pp. 8 -- 12. 3
[15] H. N. Gabow and S. Nie, Finding a long directed cycle, ACM Transactions on Algo-
rithms, 4 (2008). 3
[16] M. R. Garey and D. S. Johnson, Computers and Intractability: A Guide to the Theory
of NP-Completeness, W. H. Freeman, 1979. 14, 15
[17] S. Garg and G. Philip, Raising the bar for vertex cover: Fixed-parameter tractabil-
ity above a higher guarantee, in Proceedings of the Twenty-Seventh Annual ACM-SIAM
Symposium on Discrete Algorithms (SODA), SIAM, 2016, pp. 1152 -- 1166. 3
[18] G. Gutin, E. J. Kim, M. Lampis, and V. Mitsou, Vertex cover problem parameterized
above and below tight bounds, Theory of Computing Systems, 48 (2011), pp. 402 -- 410. 3
[19] G. Gutin, L. van Iersel, M. Mnich, and A. Yeo, Every ternary permutation constraint
satisfaction problem parameterized above average has a kernel with a quadratic number of
variables, J. Computer and System Sciences, 78 (2012), pp. 151 -- 163. 3
[20] G. Z. Gutin and V. Patel, Parameterized traveling salesman problem: Beating the
average, SIAM J. Discrete Math., 30 (2016), pp. 220 -- 238. 3
[21] F. Harary, A characterization of block-graphs, Canad. Math. Bull., 6 (1963), pp. 1 -- 6. 5
[22] F. Huffner, S. Wernicke, and T. Zichner, Algorithm engineering for color-coding
with applications to signaling pathway detection, Algorithmica, 52 (2008), pp. 114 -- 132. 3
[23] J. Kneis, D. Molle, S. Richter, and P. Rossmanith, Divide-and-color, in Proceedings
of the 34th International Workshop Graph-Theoretic Concepts in Computer Science (WG),
vol. 4271 of Lecture Notes in Computer Science, Springer, 2008, pp. 58 -- 67. 3
[24] I. Koutis, Faster algebraic algorithms for path and packing problems, in Proceedings of
the 35th International Colloquium on Automata, Languages and Programming (ICALP),
vol. 5125 of Lecture Notes in Comput. Sci., Springer, 2008, pp. 575 -- 586. 3
[25] D. Lokshtanov, N. S. Narayanaswamy, V. Raman, M. S. Ramanujan, and
S. Saurabh, Faster parameterized algorithms using linear programming, ACM Trans. Al-
gorithms, 11 (2014), pp. 15:1 -- 15:31. 3
[26] M. Mahajan and V. Raman, Parameterizing above guaranteed values: Maxsat and max-
cut, J. Algorithms, 31 (1999), pp. 335 -- 354. 3
[27] M. Mahajan, V. Raman, and S. Sikdar, Parameterizing above or below guaranteed
values, J. Computer and System Sciences, 75 (2009), pp. 137 -- 153. 3
17
[28] D. W. Matula and L. L. Beck, Smallest-last ordering and clustering and graph coloring
algorithms, J. ACM, 30 (1983), pp. 417 -- 427. 1, 4, 5, 13
[29] B. Monien, How to find long paths efficiently, in Analysis and design of algorithms for com-
binatorial problems (Udine, 1982), vol. 109 of North-Holland Math. Stud., North-Holland,
Amsterdam, 1985, pp. 239 -- 254. 3
[30] M. Naor, L. J. Schulman, and A. Srinivasan, Splitters and near-optimal derandom-
ization, in Proceedings of the 36th Annual Symposium on Foundations of Computer Science
(FOCS 1995), IEEE, 1995, pp. 182 -- 191. 11
[31] I. Schiermeyer, Problems remaining np-complette for sparse or dense graphs, Discus-
siones Mathematicae Graph Theory, 15 (1995), pp. 33 -- 41. 16
[32] D. Tsur, Faster deterministic parameterized algorithm for k-path, CoRR, abs/1808.04185
(2018). 3
[33] R. Williams, Finding paths of length k in O∗(2k) time, Inf. Process. Lett., 109 (2009),
pp. 315 -- 318. 3
[34] M. Zehavi, A randomized algorithm for long directed cycle, Inf. Process. Lett., 116 (2016),
pp. 419 -- 422. 3
18
|
1802.10488 | 1 | 1802 | 2018-02-28T15:49:11 | An Approximate Pareto Set for Minimizing the Maximum Lateness and Makespan on Parallel Machines | [
"cs.DS"
] | We consider the two-parallel machines scheduling problem, with the aim of minimizing the maximum lateness and the makespan. Formally, the problem is defined as follows. We have to schedule a set J of n jobs on two identical machines. Each job i in J has a processing time p_i and a delivery time q_i. Each machine can only perform one job at a given time. The machines are available at time t=0 and each of them can process at most one job at a given time. The problem is to find a sequence of jobs, with the objective of minimizing the maximum lateness L_max and the makespan C_max. With no loss of generality, we consider that all data are integers and that jobs are indexed in non-increasing order of their delivery times: q_1 >= q_2 >= ... >= q_n. This paper proposes an exact algorithm (based on a dynamic programming) to generate the complete Pareto Frontier in a pseudo-polynomial time. Then, we present an FPTAS (Fully Polynomial Time Approximation Scheme) to generate an approximate Pareto Frontier, based on the conversion of the dynamic programming. The proposed FPTAS is strongly polynomial. Some numerical experiments are provided in order to compare the two proposed approaches. | cs.DS | cs | An Approximate Pareto Set for Minimizing the
Maximum Lateness and Makespan on Parallel
Machines
Gais Alhadi1, Imed Kacem2, Pierre Laroche3, and Izzeldin M. Osman4
1
8
1
0
2
b
e
F
8
2
]
S
D
.
s
c
[
1
v
8
8
4
0
1
.
2
0
8
1
:
v
i
X
r
a
Abstract-We consider the two-parallel machines scheduling
problem, with the aim of minimizing the maximum lateness and
the makespan. Formally, the problem is defined as follows. We
have to schedule a set J of n jobs on two identical machines.
Each job i ∈ J has a processing time pi and a delivery time
qi. Each machine can only perform one job at a given time.
The machines are available at time t = 0 and each of them
can process at most one job at a given time. The problem is
to find a sequence of jobs, with the objective of minimizing the
maximum lateness Lmax and the makespan Cmax. With no loss
of generality, we consider that all data are integers and that
jobs are indexed in non-increasing order of their delivery times:
q1 ≥ q2 ≥ . . . ≥ qn. This paper proposes an exact algorithm
(based on a dynamic programming) to generate the complete
Pareto Frontier in a pseudo-polynomial time. Then, we present
an FPTAS (Fully Polynomial Time Approximation Scheme) to
generate an approximate Pareto Frontier, based on the conversion
of the dynamic programming. The proposed FPTAS is strongly
polynomial. Some numerical experiments are provided in order
to compare the two proposed approaches.
I. INTRODUCTION
We consider the two-parallel machines scheduling prob-
lem, with the aim of minimizing the maximum lateness and
makespan. Formally, the problem is defined as follows. We
have to schedule a set J of n jobs on two identical machines.
Each job i ∈ J has a processing time pi and a delivery time
qi. The machines are available at time t=0 and each of them
can process at most one job at a time. The problem is to
find a sequence of jobs, with the objective of minimizing the
maximum lateness Lmax and the makespan Cmax. With no
loss of generality, we consider that all data are integers and
that jobs are indexed in non-increasing order of their delivery
times q1 ≥ q2 ≥ . . . ≥ qn.
For self-consistency, we recall some necessary definitions
related to the approximation area. An algorithm A is called
a ρ−approximation algorithm for a given problem, if for any
instance I of that problem the algorithm A yields, within a
polynomial time, a feasible solution with an objective value
A(I) − OP T (I) ≤ .OP T (I), where
A(I) such that:
is
1Gais Alhadi
a member
Sciences, University
Computer
[email protected]
Pierre
Kacem2
Imed
and
of Faculty
of Gezira, Wad-Madani,
of Mathematical
and
Sudan
LCOMS
the
Metz,
[email protected]
Laboratory,
France
of
F-57045
[email protected],
Laroche3
of
are
Lorraine,
University
members
Izzeldin M. Osman4 is from Sudan University of Science and Technology,
Khartoum, Sudan [email protected]
it
OP T (I) is the optimal value of I and ρ is the performance
guarantee or the worst-case ratio of the approximation al-
gorithm A. It can be a real number greater or equal to 1
for the minimization problems ρ = 1 + (that
leads
to inequality A(I) ≤ (1 + )OP T (I)), or it can be real
number from the interval [0, 1] for the maximization problems
ρ = 1− (that it leads to inequality A(I) ≥ (1− )OP T (I)).
The Pareto-optimal solutions are the solutions that are not
dominated by other solutions. Thus, we can consider that
the solution is Pareto-optimal if there does not exist another
solution that is simultaneously the best for all the objectives.
Noteworthy, Pareto-optimal solutions represent a range of
reasonable optimal solutions for all possible functions based
on the different objectives. A schedule is called Pareto-optimal
if it is not possible to decrease the value of one objective
without increasing the value of the other.
It is noteworthy that during the last decade the multi-objective
scheduling problems have attracted numerous researchers from
all the world and have been widely studied in the litera-
ture. For the scheduling problems on serial-batch machine,
Geng et al.[17] studied scheduling problems with or without
precedence relations, where the objective is to minimize
makespan and maximum cost. They have provided highly
efficient polynomial-time algorithms to generate all Pareto
optimal points. An approximate Pareto set of minimal size
that approximates within an accuracy for multi-objective
optimization problems have been studied by Bazgan et al.[2].
They proposed a 3-approximation algorithm for two objec-
tives and also proposed a study of the greedy algorithm
performance for a three-objective case when the points are
given explicitly in the input. They showed that the three-
objective case is NP-hard. Chen and Zou [5] proposed a
runtime analysis of a (µ + 1) multi-objective evolutionary
algorithm for three multi-objective optimization problems with
unknown attributes. They showed that when the size of the
population is less than the total number of Pareto-vector, the
(µ + 1) multi-objective evolutionary algorithm cannot obtain
the expected polynomial runtime for the exact discrete multi-
objective optimization problems. Thus, we must determine the
size of the population equal to the total number of leading
ones, trailing zeros. Furthermore, the expected polynomial
runtime for the exponential discrete multi-objective optimiza-
tion problem can be obtained by the ratio of n/2 to µ − 1
over an appropriate period of time. They also showed that the
(µ + 1) multi-objective evolutionary algorithm can be solved
efficiently in polynomial runtime by obtaining an − adaptive
to produce all
Pareto front. Florios and Mavrotas [6] used AUGMECON2,
a multi-objective mathematical programming method (which
is suitable for general multi-objective integer programming
problems),
the Pareto-optimal solutions for
multi-objective traveling salesman and set covering problems.
They showed that the performance of the algorithm is slightly
better than it already exists. Moreover, they showed that their
results can be helpful for other multi-objective mathematical
programming methods or even multi-objective meta-heuristics.
In [10], Sabouni and Jolai proposed an optimal method for
the problem of scheduling jobs on a single batch processing
machine to minimize the makespan and the maximum lateness.
They showed that the proposed method is optimal when the
set with maximum lateness objective has the same process-
ing times. They also proposed an optimal method for the
group that has the maximum lateness objective and the same
processing times. Geng et al.[3] considered the scheduling
problem on an unbounded p-batch machine with family jobs
to find all Pareto-optimal points for minimizing makespan and
maximum lateness. They presented a dynamic programming
algorithm to solve the studied problem. He et al.[4] showed
that the Pareto optimization scheduling problem on a single
bounded serial-batching machine to minimize makespan and
maximum lateness is solvable in O(n6). They also presented
an O(n3)- time algorithm to find all Pareto optimal solutions
where the processing times and deadlines are agreeable. For
the bi-criteria scheduling problem, He et al.[8] showed that the
problem of minimizing maximum cost and makespan is solv-
able in O(n5) time. The authors presented a polynomial-time
algorithm in order to find all Pareto optimal solutions. Also,
He et al.[9] showed that the bi-criteria batching problem of
minimizing maximum cost and makespan is solvable in O(n3)
time. The bi-criteria scheduling problem on a parallel-batching
machine to minimize maximum lateness and makespan have
been considered in [11]. The authors presented a polynomial-
time algorithm in order to find all Pareto optimal solutions.
Allahverdi and Aldowaisan [13] studied the no-wait flow-shop
scheduling problem with bi-criteria of makespan or maximum
lateness. They also proposed a dominance relation and a
branch-and-bound algorithm and showed that these algorithms
are quite efficient.
The remainder of this paper is organized as follows. In
Section 2, we describe the proposed dynamic programming
(DP) algorithm. Section 3, provides the description and the
analysis of the FPTAS. In Section 4, we present a practical
example for DP and FPTAS. Finally, Section 5 concludes the
paper.
II. DYNAMIC PROGRAMMING ALGORITHM
The following dynamic programming algorithm A, can
be applied to solve exactly this problem. This algorithm A
generates iteratively some sets of states. At every iteration i,
a set χi composed of states is generated (0 ≤ i ≤ n). Each
state [k, Lmax, Cmax] in χi can be associated to a feasible
partial schedule for the first i jobs. Let variable k ∈ {0, 1}
denote the most loaded machine, Lmax denote the maximum
lateness and Cmax denote the maximum completion time
– add
j=1 pj − Cmax)
if (Cmax ≥(cid:80)i
[k,max{Lmax,(cid:80)i
– add [1 − k,max{Lmax,(cid:80)i
qi},(cid:80)i
qi},Cmax] to χi
else
j=1 pj − Cmax] to χi
j=1 pj − Cmax +
j=1 pj − Cmax +
2
of the corresponding schedule. The dynamic programming
algorithm can be described as follows.
ALGORITHM A
1) Set χ1 = {[1, p1 + q1, p1]}.
2) For i ∈ {2, 3, ..., n},
a) χi = φ.
b) For every state [k, Lmax, Cmax] in χi−1 :
• (schedule job i on machine k)
add [k,max{Lmax,Cmax + pi + qi},Cmax + pi]
to χi
• (schedule job i on machine 1 − k)
c) For every k, for every Cmax: keep only one state
with the smallest possible Lmax.
d) Remove χi−1.
dominated states.
3) Return the Pareto front of χn, by only keeping non-
Remark: To destroy the symmetry, we start by χ1 = {[1, p1 +
q1, p1]} (i.e., we perform job 1 on the first machine).
III. APPROXIMATE PARETO FRONTIER
The main idea of the Approximate Pareto Frontier is to
remove a special part of the states generated by the dynamic
programming algorithm A. Therefore, the modified algorithm
A(cid:48) described in Lemma III.1 produces an approximation
solution instead of the optimal solution.
Given an arbitrary > 0, we define the following parameters:
and
δ1 =
P/2
n
,
δ2 =
(P + qmax)/3
n
.
where qmax is the maximum delivery time and P is the total
sum of processing times.
Let L∗
max be the optimal solutions for our
two objectives. Let LM AX and CM AX be the upper
bounds for the two considered criteria (scheduling all the
jobs on the same machine), such that,
max and C∗
0 ≤ LM AX = P + qmax ≤ 3L∗
max
0 ≤ CM AX = P ≤ 2C∗
max
We divide the intervals [0, CM AX] and [0, LM AX] into
equal sub-intervals respectively of lengths δ1 and δ2. Then,
an FPTAS is defined by following the same procedure as in
the dynamic programming, except the fact that it will keep
only one representative state for every couple of the defined
subintervals produced from [0, CM AX] and [0, LM AX].
Thus, our FPTAS will generate approximate sets χ#
i of states
• 2nd
(cid:48)
[k
,max{L
max,(cid:80)i
Case:
(cid:48)
[k, Lmax, Cmax]
max + qi},C
(cid:48)
(cid:48)
max]
j=1 pj − C
∈
(cid:48)
[k
(cid:48)
(cid:48)
,L
Since
max,C
there
max]
max] ∈ χ#
(cid:48)#, L
(cid:48)#
(cid:48)#
i−1, such that:
[k
max, C
≤
max + (i − 1)max{δ1, δ2}
(cid:48)#
(cid:48)
L
L
max − (i − 1)δ1 ≤ C
max ≤ C
max
(cid:48)
(cid:48)#
C
max + (i − 1)δ1.
χi−1,
(cid:48)
3
=
exists
and
• Sub-case 2.1:(cid:80)i
Consequently, two sub-cases can occur:
max ≤ C
(cid:48)#
(cid:48)#
max
j=1 pj − C
(cid:48)#
max,(cid:80)i
(cid:48)#, max{L
j=1 pj−C
(cid:48)#
Here, the state [k
max]
at iteration i. However, it may be re-
is created by algorithm A
moved when reducing the state subset. Let [α, λ, µ] be the state
j=1 pj −
in χ#
i
max + qi}, C
(cid:48)#
C
(cid:48)#
max]. Hence, we have:
that is in the same box as [k
max,(cid:80)i
(cid:48)#, max{L
max +qi},C
(cid:48)#
(cid:48)#
(cid:48)
λ ≤ max{L
(cid:48)#
max,
i(cid:88)
j=1
pj − C
max + qi} + δ2
(cid:48)#
i(cid:88)
≤ max{L
(cid:48)
−(C
(cid:48)
max + (i − 1) max{δ1, δ2},
max − (i − 1)δ1) + qi} + δ2
max + qi}
i(cid:88)
pj − C
(cid:48)
≤ max{L
(cid:48)
max,
pj
j=1
j=1
+(i − 1) max{δ1, δ2} + δ2
≤ Lmax + (i − 1) max{δ1, δ2} + δ2
< Lmax + i. max{δ1, δ2}
Moreover,
max + δ1 ≤ C
(cid:48)#
(cid:48)
max + (i − 1)δ1 + δ1 = Cmax + iδ1.
µ ≤ C
And,
(4)
(5)
(6)
instead of χi. The following lemma shows the closeness of
the result generated by the FPTAS compared to the dynamic
programming.
Lemma III.1. For every state [k, Lmax, Cmax] ∈ χi there
max] ∈ χ#
exists at least one approximate state [m, L#
such that:
max, C #
i
max ≤ Lmax + i. max{δ1, δ2},
L#
and
Cmax − i.δ1 ≤ C #
max ≤ Cmax + i.δ1.
Proof. By induction on i.
First, for i = 0 we have χ#
i = χ1. Therefore, the statement is
trivial. Now, assume that the lemma holds true up to level i−1.
Consider an arbitrary state [k, Lmax, Cmax] ∈ χi. Algorithm
A introduces this state into χi when job i is added to some
feasible state for the first i − 1 jobs. Let [k
(cid:48)
(cid:48)
max] be
, L
max, C
the above feasible state. Three cases can be distinguished:
max + pi + qi},
j=1 pj − C
, max{L
, max{L
max,(cid:80)i
1) [k, Lmax, Cmax] = [k
(cid:48)
max,C
(cid:48)
(cid:48)
(cid:48)
(cid:48)
(cid:48)
(cid:48)
max + pi]
C
qi}, C
(cid:48)
max]
2) [k, Lmax, Cmax] = [k
3) [k, Lmax, Cmax] = [1 − k
(cid:48)
max]
max + qi},(cid:80)i
j=1 pj − C
C
(cid:48)
(cid:48)
max,(cid:80)i
(cid:48)
max +
j=1 pj −
(cid:48)
, max{L
(cid:48)
, max{L
We will prove the statement for level i in the three cases.
(cid:48)
max,C
(cid:48)
max +
(cid:48)
[k
∈
exists
χi−1,
• 1st Case:
[k, Lmax, Cmax] = [k
pi + qi}, C
(cid:48)
max + pi]
(cid:48)
(cid:48)
there
max,C
,L
Since
max]
max] ∈ χ#
(cid:48)#
(cid:48)#, L
(cid:48)#
i−1, such that:
max, C
[k
max + (i − 1) max{δ1, δ2}
≤
(cid:48)
(cid:48)#
L
L
max + (i − 1)δ1.
max ≤ C
max − (i − 1)δ1 ≤ C
max
(cid:48)#
(cid:48)
C
(cid:48)#, max{L
(cid:48)#
(cid:48)#
Consequently,
max + pi +
max, C
qi}, C
(cid:48)
is created by algorithm A
at
iteration i.
However, it may be removed when reducing the state subset.
Let [α, λ, µ] be the state in χ#
that is in the same box as the
i
(cid:48)#
max + pi]. Hence,
sate [k
we have:
max + pi + qi}, C
(cid:48)#
(cid:48)#
max + pi]
(cid:48)#, max{L
the state [k
(cid:48)#
max, C
and
(cid:48)
λ ≤ max{L
≤ max{L
(cid:48)#
max, C
(cid:48)
max, C
max + pi + qi} + δ2
(cid:48)#
max + pi + qi}
(cid:48)
+(i − 1). max{δ1, δ2} + δ2
≤ Lmax + i. max{δ1, δ2}
In addition,
µ ≤ C
≤ C
(cid:48)#
max + pi + δ1
max + (i − 1)δ1 + pi + δ1 = Cmax + iδ1.
(cid:48)
and,
µ ≥ C
max + pi − δ1 ≥ C
(cid:48)#
≥ Cmax − iδ1.
(cid:48)
max − (i − 1)δ1 + pi − δ1
(2)
(3)
Consequently, [α, λ, µ] is an approximate state verifying the
two conditions.
(1)
µ ≥ C
max − δ1 ≥ C
(cid:48)#
(cid:48)
max − (i − 1)δ1 − δ1 = Cmax − iδ1.
Consequently, [α, λ, µ] is an approximate state verifying the
two conditions.
• Sub-case 2.2:(cid:80)i
qi},(cid:80)i
Here,
j=1 pj − C
(cid:48)#
max > C
(cid:48)#, max{L
j=1 pj − C
the state [1 − k
j=1 pj − C
(cid:48)#
max] is created by algorithm A
(cid:48)#
max +
at iteration i.
However, it may be removed when reducing the state subset.
Let [α, λ, µ] be the state in χ#
that is in the same box as
[1−k
(cid:48)#, max{L
j=1 pj−C
i
(cid:48)#
(cid:48)#
max].
Hence, we have:
max,(cid:80)i
j=1 pj−C
(cid:48)#
(cid:48)
(cid:48)#
(cid:48)#
max
max,(cid:80)i
max+qi},(cid:80)i
λ ≤ max{L
(cid:48)#
max,
i(cid:88)
j=1
pj − C
max + qi} + δ2
(cid:48)#
i(cid:88)
≤ max{L
(cid:48)
−(C
(cid:48)
max + (i − 1) max{δ1, δ2},
max + (i − 1)δ1) + qi} + δ2
max + qi}
i(cid:88)
pj − C
(cid:48)
≤ max{L
(cid:48)
max,
j=1
j=1
+(i − 1) max{δ1, δ2} + δ2
≤ Lmax + (i − 1) max{δ1, δ2} + δ2
< Lmax + i. max{δ1, δ2}
Moreover,
i(cid:88)
j=1
µ ≤
pj − C
(cid:48)#
max + δ1
pj
(7)
(8)
(cid:48)
max − (i − 1)δ1, then the following relation
max ≥ C
(cid:48)#
i(cid:88)
Since C
holds
µ ≤
(since(cid:80)i
And,
j=1 pj − C
i(cid:88)
µ ≥
(cid:48)
max ≤ C
(cid:48)
max).
pj − C
max − δ1 ≥ C
(cid:48)#
max − δ1
(cid:48)#
j=1
≥ Cmax − (i − 1)δ1 − δ1 = Cmax − iδ1.
(10)
max,(cid:80)i
(cid:48)
Thus, [α, λ, µ] verifies the necessary conditions.
max + qi},(cid:80)i
(cid:48)
(cid:48)
,
[1 − k
(cid:48)
max]
exists
and
(cid:48)
[k
• 3rd Case:
[k, Lmax, Cmax]
=
max{L
j=1 pj − C
j=1 pj − C
(cid:48)
(cid:48)
Since
,L
max,C
there
max]
max] ∈ χ#
(cid:48)#, L
(cid:48)#
(cid:48)#
i−1, such that:
[k
max, C
≤
max + (i − 1) max{δ1, δ2}
(cid:48)#
(cid:48)
L
L
max − (i − 1)δ1 ≤ C
max ≤ C
max
(cid:48)#
(cid:48)
C
Consequently, two sub-cases can occur:
max + (i − 1)δ1.
χi−1,
∈
(cid:48)
• Sub-case 3.1:(cid:80)i
qi},(cid:80)i
j=1 pj − C
max ≥ C
(cid:48)#
(cid:48)#,max{L
max,(cid:80)i
(cid:48)#
max
j=1 pj − C
Here,
the state [1 − k
j=1 pj − C
(cid:48)#
max +
at iteration i.
However, it may be removed when reducing the state subset.
Let [α, λ, µ] be the state in χ#
that is in the same box as
i
(cid:48)#
max] is created by algorithm A
(cid:48)#
(cid:48)
pj − C
(cid:48)
max + (i − 1)δ1 + δ1 ≤ Cmax + iδ1
µ ≥
pj − C
max − δ1
(cid:48)#
j=1
(9)
(11)
(12)
(13)
max,(cid:80)i
(cid:48)#
[1−k
(cid:48)#,max{L
Hence, we have:
j=1 pj−C
i(cid:88)
λ ≤ max{L
(cid:48)#
max,
pj − C
4
max +qi},(cid:80)i
(cid:48)#
j=1 pj−C
(cid:48)#
max].
max + qi} + δ2
(cid:48)#
i(cid:88)
j=1
i(cid:88)
j=1
≤ max{L
(cid:48)
max + (i − 1)max{δ1, δ2},
max − (i − 1)δ1) + qi} + δ2
(cid:48)
−(C
pj
j=1
≤ max{L
(cid:48)
max,
pj
−C
(cid:48)
max + qi} + (i − 1) max{δ1, δ2} + δ2
≤ Lmax + (i − 1) max{δ1, δ2} + δ2
≤ Lmax + i max{δ1, δ2}
and
µ ≤
pj − C
(cid:48)#
max + δ1
(cid:48)
max − (i − 1)δ1) + δ1
pj − (C
≤ Cmax + iδ1.
j=1
In the other hand, we have
i(cid:88)
≤ i(cid:88)
j=1
i(cid:88)
i(cid:88)
j=1
(cid:48)
max + (i − 1)δ1) − δ1
≥
pj − (C
≥ Cmax − iδ1.
j=1
Thus, [α, λ, µ] fulfills the conditions.
• Sub-case 3.2:(cid:80)i
j=1 pj − C
(cid:48)#
max,(cid:80)i
(cid:48)#
max < C
(cid:48)#
max
j=1 pj−C
(cid:48)#, max{L
(cid:48)#
Here, the state [k
max]
at iteration i. However, it may be re-
is created by algorithm A
moved when reducing the state subset. Let [α, λ, µ] be the state
j=1 pj −
in χ#
i
max + qi},C
(cid:48)#
C
(cid:48)#
max]. Hence, we have:
that is in the same box as [k
max,(cid:80)i
(cid:48)#, max{L
max +qi},C
(cid:48)#
(cid:48)#
(cid:48)
i(cid:88)
λ ≤ max{L
(cid:48)#
max,
pj − C
max + qi}
(cid:48)#
j=1
+δ2
≤ max{L
−(C
≤ max{L
(cid:48)
(cid:48)
max + (i − 1) max{δ1, δ2},
max − (i − 1)δ1) + qi} + δ2
max + qi}
i(cid:88)
pj − C
(cid:48)
max,
(cid:48)
j=1
+(i − 1) max{δ1, δ2} + δ2
≤ Lmax + (i − 1) max{δ1, δ2} + δ2
≤ Lmax + i max{δ1, δ2}
i(cid:88)
j=1
pj
(14)
In the other hand, we have
µ ≥ C
pj − C
max − δ1
(cid:48)#
(cid:48)#
max − δ1 ≥ i(cid:88)
i(cid:88)
j=1
pj − C
(15)
(16)
5
max. We can see that
max/C∗
C A(cid:48)
the FPTAS algorithms are
finding solutions closer to the optimal ones when the number
of jobs is increasing. Cmax values are closer to the optimal
than Lmax values, which is not a surprise, as Lmax depends
on Cmax. Worth to mention, our FPTAS with = 0.3 gives
better results than with = 0.9, which is consistent with the
theory.
We have also studied the influence of processing and
delivery times, in Table I and II. These tables present average
results for our benchmark, considering the 5 sets of instances.
Results are presented for = 0.3; the analysis is the same
with = 0.9. Table I shows that instances composed of jobs
with various processing times lead to optimal solutions with
a bigger Pareto front, as seen in column 2. At the opposite,
columns 3 to 5 show that our FPTAS algorithm is not really
influenced by this parameter. Table II shows that delivery
times ranges have more influence on the results : the size of
the Pareto front of non-dominated solutions grows faster. Our
FPTAS algorithm has the same behavior, and we can see that
the results are more close to the optimal with smaller values
of qi.
and
µ ≤ C
max + δ1 ≤ C
(cid:48)#
(cid:48)
max + (i − 1)δ1 + δ1
≤ Cmax + iδ1.
≥
(cid:48)
max − iδ1 = Cmax − iδ1.
j=1
Therefore, [α, λ, µ] fulfills the conditions.
In conclusion,
level
the third case, and this completes our inductive proof.
the statement holds also for
i
in
Based on the lemma, we deduce easily that for every non-
dominated state [k, Lmax, Cmax] ∈ χn, it must remain a close
state [m, L#
max] ∈ χ#
n such that:
max, C #
max ≤ Lmax + n. max{δ1, δ2} ≤ (1 + ).Lmax,
L#
and
max ≤ Cmax + n.δ1 ≤ (1 + ).Cmax.
C #
Table I: Quality of FPTAS as a function of processing time
ranges (for = 0.3)
Moreover, it is clear that the FPTAS runs polynomially in n
and 1/. The overall complexity of our FPTAS is O(n3/2).
IV. RESULTS
The following results have been obtained after testing the
performance of the proposed algorithms. The code has been
done in Java and the experiments were performed on an
Intel(R) Core(TM)-i7 with 8GB RAM. We randomly generate
five sets of instances, with different numbers of jobs and
various processing and delivery times:
• number of jobs: from 5 to 25, 26 to 50, 51 to 75, 76 to
100 and 100 to 200
• processing times : from 1 to 20, 1 to 100 and 1 to 1000
• delivery times : from 1 to 20, 1 to 100 and 1 to 1000
That gave us 135 instances in each set of instances. Finally,
the FPTAS has been tested with two values of : 0.3 and 0.9.
To ensure the consistency of running times, each test has been
run three times.
Figure 1 presents a comparison of FPTAS and Dynamic
Programming. The left part of this figure shows the average
size of the Pareto Front (i.e. the number of solutions) found
by the Dynamic Programming algorithm and our FPTAS with
the two values we used. The sizes are given for our five sets
of instances, from small instances (5-25 instances) to bigger
ones (100-200 jobs). We can see that the number of solutions
decreases as the number of jobs increases. With a lot of jobs,
it is more likely to obtain very similar solutions, a lot of them
being dominated by others. At the opposite, the number of
jobs has no real influence on the Pareto front sizes found by
our FPTAS algorithm, whatever the value of .
quality of the two objectives of our study: LA(cid:48)
On the right part of the same figure are given the average
max and
max/L∗
pi range
1-20
1-100
1-500
DP size of
Pareto front
2.26
4.07
5.65
FPTAS size of
Pareto front
1.23
1.84
1.73
max
CA(cid:48)
max/C∗
1.0006
1.0008
1.0005
max
LA(cid:48)
max/L∗
1.003
1.007
1.004
Table II: Quality of FPTAS as a function of delivery time
ranges (for = 0.3)
qi range
1-20
1-100
1-500
DP size of
Pareto front
2.02
3.03
6.93
FPTAS size of
Pareto front
1.25
1.57
1.99
max
CA(cid:48)
max/C∗
1.0007
1.0007
1.0005
max
LA(cid:48)
max/L∗
1.001
1.002
1.008
Computing times are given in Tables III, IV and V. They
compare our Dynamic Programming algorithm and our FP-
TAS, considering two values : 0.3 and 0.9. All values
are in milliseconds. Table III shows that all algorithms are
slower when the number of states is growing. Table IV
shows an interesting result: while the Dynamic Programming
algorithm becomes slower when the processing times ranges
are growing, the FPTAS has an opposite behavior. The FPTAS
with = 0.3 is even slower than the exact algorithm for the
smallest range of processing times. Table V shows that the
delivery times ranges have a smaller influence on the Dynamic
Programming algorithm: computing times are growing, but
slower than the delivery times ranges. The FPTAS computing
times are also growing in function of the delivery times ranges.
Note that tables IV and V are based on mean values from all
the set of instances.
6
Size of Pareto Front
Quality of Cmax and Lmax
Figure 1: Quality of our FPTAS algorithm with = 0.3 and = 0.9
Table III: Average computing times vs size of instances (ms)
#jobs
5-25
26-50
51-75
76-100
100-200
DP
67
1278
7917
24937
164332
FPTAS
= 0.3
0.9
5.4
22
58
281
= 0.9,
0.3
0.9
3.4
7.8
32
times, small delivery times and big make the FPTAS faster
and vice versa).
In our future works, the study of the multiple-machine schedul-
ing problems seems to be a challenging perspective in the
extension of our work.
REFERENCES
Table IV: Average computing times (ms) vs processing time
ranges
piranges
1-20
1-100
1-500
DP
91
2668
116362
FPTAS
= 0.3
144
51
25
= 0.9,
15
7
5
V. CONCLUSIONS AND PERSPECTIVES
The two-parallel machines scheduling problem has been
considered to minimize the maximum lateness and the
makespan. We have proposed an exact algorithm (based on
dynamic algorithm) to generate the complete Pareto Frontier
in a pseudo-polynomial time. Then, we present an FPTAS
(Fully Polynomial Time Approximation Scheme) to generate
an approximate Pareto Frontier, based on the conversion of
the exact dynamic programming. For the proposed algorithms,
we randomly generated several instances with different ranges,
and, for each job Ji, its processing time pi and delivery time qi
are sets to be integer numbers.The results of the experiments
showed that the proposed algorithms for the considered prob-
lem are very efficient. It is clear that optimizing the maximum
lateness (Lmax) implies to minimize implicitly the makespan
(Cmax). Moreover, the values of and processing and delivery
times play an important role in the results (i.e., big processing
Table V: Average computing times (ms) vs delivery time
ranges
qiranges
1-20
1-100
1-500
DP
31447
40482
47190
FPTAS
= 0.3
26
46
149
= 0.9,
5
7
15
[1] G. Sapienza, G. Brestovac, R. Grgurina, and T. Seceleanu. "On applying
multiple criteria decision analysis in embedded systems design." Design
automation for embedded systems,20(3), Vol 20(3), pp 211–238 (2016).
[2] C. Bazgan, F. Jamain, and D. Vanderpooten. "Approximate Pareto sets
of minimal size for multi-objective optimization problems." .Operations
Research Letters, Vol 43(1), pp 1–6 (2015).
[3] Z. Geng and J. Yuan. "Pareto optimization scheduling of family jobs
on a p-batch machine to minimize makespan and maximum lateness."
Theoretical Computer Science, Vol 570, pp 22–29 (2015).
[4] C. He, H. Lin, and Y. Lin. "Bounded serial-batching scheduling for
minimizing maximum lateness and makespan." Discrete Optimization,
Vol 16, pp 70–75 (2015).
[5] Y. Chen and X. Zou. "Runtime analysis of a multi-objective evolutionary
algorithm for obtaining finite approximations of Pareto fronts." Informa-
tion Sciences, Vol 262, pp 62–77 (2014).
[6] K. Florios and G. Mavrotas. "Generation of the exact Pareto set in
Multi-Objective Traveling Salesman and Set Covering Problems." Applied
Mathematics and Computation, Vol 237, pp 1–19 (2014).
[7] Q. Feng, J. Yuan, H. Liu, and C. He. "A note on two-agent scheduling on
an unbounded parallel-batching machine with makespan and maximum
lateness objectives." Applied Mathematical Modelling, Vol 37(10–11), pp
7071–7076 (2013).
[8] C. He, H. Lin, Y. Lin, and J. Tian. "Bicriteria scheduling on a series-
batching machine to minimize maximum cost and makespan." Central
European Journal of Operations Research, pp 1–10 (2013).
[9] C. He, X. M. Wang, Y. X. Lin, and Y. D. Mu. "An Improved Algorithm for
a Bicriteria Batching Scheduling Problem." RAIRO-Operations Research,
Vol 47(1), pp1–8 (2013).
[10] M. T. Y. Sabouni and F. Jolai. "Optimal methods for batch processing
problem with makespan and maximum lateness objectives." Applied
Mathematical Modelling, Vol 34(2), pp 314–324 (2010).
[11] C. He, Y. Lin, and J. Yuan. "Bicriteria scheduling on a batching machine
to minimize maximum lateness and makespan." Theoretical Computer
Science, Vol 381(1-3), pp 234–240 (2007).
[12] D. Sarkar and J. M. Modak. "Pareto-optimal solutions for multi-objective
optimization of fed-batch bioreactors using nondominated sorting genetic
algorithm." Chemical Engineering Science, Vol 60(2), pp 481–492 (2005).
[13] A. Allahverdi and T. Aldowaisan. "No-wait flowshops with bicriteria
of makespan and maximum lateness." European Journal of Operational
Research,Vol 152(1), pp 132–147 (2004).
[14] C. M. Sil and E. C. Biscaia. "Genetic algorithm development for multi-
objective optimization of batch free-radical polymerization reactors."
Computers & chemical engineering, Vol 27(8), pp 1329–1344 (2003).
[15] P. Kumar. "A Framework for Multi-objective Optimization and Multi-
criteria Decision Making for Design of Electrical Drives." Universiteit
Delft. (2008).
[16] S. Chakhar and J. Martel. "Multi-Criteria Evaluation Functions Inside
Geographical Information Systems Towards a Spatial Decision Support
System."(2006).
[17] Z. Geng and J. Yuan. "Scheduling with or without precedence relations
on a serial-batch machine to minimize makespan and maximum cost."
submitted, 2017
7
|
1705.05105 | 3 | 1705 | 2017-05-24T19:52:38 | Assembling sequences of DNA using an on-line algorithm based on DeBruijn graphs | [
"cs.DS"
] | The problem of assembling DNA fragments starting from imperfect strings given by a sequencer, classified as NP hard when trying to get perfect answers, has a huge importance in several fields, because of its relation with the possibility of detecting similarities between animals, dangerous pests in crops, and so on. Some of the algorithms and data structures that have been created to solve this problem are Needleman Wunsch algorithm, DeBruijn graphs and greedy algorithms working on overlaps graphs; these try to work out the problem from different approaches that give place to certain advantages and disadvantages to be discussed.
In this article we first expose a summary of the research done on already created solutions for the DNA assembly problem, to present later an on-line solution to the same matter, which, despite not considering mutations, would have the capacity of using only the necessary amount of readings to assemble an user specified amount of genes. | cs.DS | cs | Assembling sequences of DNA using an on-line algorithm based
on DeBruijn graphs
Andres Felipe Zapata Palacio
Juan Manuel Ciro Restrepo
Universidad Eafit
Colombia
[email protected]
Universidad Eafit
Colombia
[email protected]
Mauricio Toro
Universidad Eafit
Colombia
[email protected]
ABSTRACT
The problem of assembling DNA fragments starting from
imperfect strings given by a sequencer, classified as
NP-hard when trying to get perfect answers [4], has a huge
importance in several fields, because of its relation with the
possibility of detecting similarities between animals,
and so on. Some of the
dangerous pests in crops,
algorithms and data structures that have been created to
solve this problem are Needleman–Wunsch algorithm, De
bruijn graphs and greedy algorithms working on overlaps
graphs; these try to work out the problem from different
approaches that give place to certain advantages and
disadvantages to be discussed.
In this article we first expose a summary of the research
done on already created solutions for the DNA assembly
problem, to present later an on-line solution to the same
matter, which, despite not considering mutations, would
have the capacity of using only the necessary amount of
readings to assemble an user specified amount of genes.
Author Keywords
DNA assembly, DeBruijn graphs, eulerian walks, on-line
algorithms, complexity.
ACM Classification Keywords
Applied computing→ Life and medical
sciences→Computational biology→Recognition of genes
and regulatory elements;
INTRODUCTION
As many other sciences, the biology progress has been
markedly accelerated by the use of computational and
the
statistical analysis in many fields; one of
assembly of DNA sequences. This specific problem has
undergone a lot of variation in its constraints, from the hand
work at the very beginning to the length of the reads of real
time devices for DNA sequencing.
These devices, however, do not sequence the DNA in a
predictable way, because the reads length is not enough to
cover the whole DNA and it is not currently possible to start
them,
one exactly after another. Algorithms for genome assembly
are then playing a crucial role in order to get entire genes.
PROBLEM
Currently, the DNA sequencers turn DNA in strings within
the alphabet {A, G, T, C}, that represents its Nitrogenous
bases. However, because a sequencer can not read a whole
DNA directly, the DNA of a single creature is given in
many fragments, which must be matched and conjugated
on-line, it is, while the sequencer is still working.
Because of that, the solution has to work with the data
introduced in real-time by the sequencer, and produce
results before the reading process is finished. (It would take
too much time and memory to wait for the reading to be
completed)
The objective of this work is then finding one solution to
process the reads of a sequencer with such characteristics,
finding the DNA strand of the examined living being and
extracting the genes found in the partially assembled DNA
sequence.
The program will finish when the sequencer stops reading,
when the number of genes required are found or when the
limit time of computation is spent.
RELATED WORK
De Bruijn graph and Eulerian walks [7]
One algorithm created to solve the genome assembly
problem is based on traversing a Bruijn graph following an
Eulerian path (a trail in a graph which visits every edge
exactly once[9]).
A Bruijn graph is a directed graph that represents each
k-mer (Substring of length k) of the genome with nodes
containing its prefix and suffix (k-1)-mers and then
establishing an edge A → B for every A and B where B has
a prefix that is also a suffix of A, that is, there is an overlap
between A and B.
After building such a graph the genome can be obtained by
following the aforementioned Eulerian path of the graph.
ALGORITHM
This algorithm uses one DeBruijn graph as in the Figure 1,
to store the k-mers of each fragment read:
The program reads each DNA fragment given and updates
the DeBruijn graph:, getting the right and left K-1mers for
each kmer, inserting them into the graph (if they are not
already added) and then setting a directed edge from the
left to the right one. Also, it is necessary to discard the
possibility that the right K1-mer (the one to which the edge
"arrives") is initial, this aspect is going to be clarified later.
Once the graph is updated, the next step is to traverse it. It
is done by following the next K-1mers of each initial
K-1mer (each K-1mer to which no other K-1mer points),
until a K-1mer without next is found. This process gave us
a set of partially assembled strands.
The last step is processing every strand and finding the
genes in it. At first the program goes over the strand and
saves the indexes of the start and stop codons, after that it
"greedily" matches the first start codon found with the first
stop codon (a gen has been found) and the start codons in
between are removed, that because there can be many start
codons before a stop codon and the "correct" one is the one
that produces the largest gen.
If one gen is found, the program checks if it is already
found(the found genes are stored in a HashSet so that it is
easy to check) if that is not the case, the gen is printed in the
screen and stored in that HashSet.
DATA STRUCTURES
graph
that
The
representation had to suffice are basically three: Each K- 1
mer needs to be added at most once, the graph must be able
to be efficiently traversed, and, as a consequence of the
former, the initial nodes should be able to be found easily.
At a first glance, a HashMap<K-1mer, nextK-1mer> seems
to be a good option: It is O(1) for inserting, and also allows
to get the "next" of each K-1mer in constant time, but when
it
there is no
straightforward way to do that.
Keeping in mind the HashMap advantages we implemented
a bijection between each K-1mer and an index using a
bidirectional HashMap, so that getting the index of a
K-1mer and the K-1mer of an index are both performed in
O(1) constant time. That indexing makes it possible to
the graph as an one-dimensional "adjacency
represent
every edge A → B is represented as
array" where
array[indexof A] = indexof B. The array is one-dimensional
since A and B are guaranteed to be unique , which causes
that edge to be unique, and the only outgoing edge from A.
to finding the initial nodes
data-structure-level
needings
comes
the
Figure 1: graphic representation of a DeBruijn graph.
Global alignment of two genetic sequences [3]
Needleman-Wunsch and Christian Wunsch implemented in
1970 an algorithm that aligns protein and nucleic acid
sequences. This computation allows you to compare two
sequences and determine how different or similar they are
the one of the another.
the alphabet(set of
This algorithm needs some inputs:
symbols that make up the sequences), the two strings that
contains the two sequences, a function that determines a
mark of similarity between each pair of symbols that make
up the alphabet, and the mark of no matching a symbol.
Greedy shortest common superstring [4]
Another approach to the genome assembly problem,
simplified to finding the shortest common superstring, are
greedy algorithms, which are quick at the cost of not
guaranteeing that the found superstring is the shortest one.
More precisely, the greedy approximations can compute in
O(n log n)
time a superstring that in the worst case is,
however, "only β times ( where 2 ≤ β ≤ 4) longer than the
shortest common superstring"[6].
The strategy behind the greedy approach is basically
building and "greedily" reducing an overlaps graph. This
graph is a directed weighted graph, where A → B shows
that A overlaps B, and the weight of the edge indicates the
number of overlapped characters. The reduction, on the
other hand, merges (eliminating redundancy) the nodes
attached to the edge with the greatest weight at each
iteration, randomly choosing if there is more than one edge
labeled with that weight and concatenating the remaining
nodes after all edges have been merged.
Comparing biological sequence information [5]
BLAST is an informatic program that computes the
statistical level of similarity between a given nucleotide
sequence and all the sequences stored in its database, using
an heuristic algorithm.
Although the heuristic does not ensures the answer is
correct,
in most of the cases, BLAST works not only
successfully, but also efficiently.
To find the initial nodes the the strategy is pretty simple:
every K-1mer, at the beginning, is a "possible initial", and
each time an edge A → B is added, B stops being an
initial. We save that information in a 1D boolean array,
indexed the same way, and which values are all True at the
beginning. In the end those K-1mers whose values in their
corresponding indices are still True are the initial nodes:
Figure 2: Strategy used to determine the initial nodes.
the same edge or
This implementation assures us that
K-1mer will not be stored twice. Shown below, there is a
summary of the complexity of the operations gotten by
using this implementation.
V: number of K-1mers stored in the graph
Operation
Complexity
Get the index of one K-1mer
Get the K-1mer referred by one index
Get the next K-1mer of one K-1mer
Check whether a K-1mer is an initial
node
Get all the initial nodes of the graph
Traverse the graph(starting at every initial
node)
O(1)
O(1)
O(1)
O(1)
O(V)
O(V)
Table 1: Complexity of operations on the proposed
implementation of the graph.
01)
COMPLEXITY
Since the length of the K-mer is 201, the graph is going to
store K1-mers of 200. When the graph is being updated, for
each K-mer in the read of length L, two K-1mers are taken,
. After that, theY new K-1 mers are
that
inserted and their edges added to the graph, both operations
in constant time. Updating the graph is then
Once the graph is updated it has to be traversed in order to
find the genes. First, we get all the initials O(V), and then
(L)O
( − 2
* 2
L
.
is
)
(V )
= O
(2V )
the specific lengths of all reads. It
traverse the graph from every initial. Since the group of
K-1mers pointed by each initial are disjoint every K-1mer is
visited only once and then the complexity of traversing
the graph is
O + V = O
(V
together, the complexity of processing N
Putting it all
lectures is developed next:
For the sake of simplicity we use the average read length
is fair
instead of
considering that they were going to be used altogether in a
summatory and, because of that, the final value does not
change.
Let A be the average read length. Each time the graph is
traversed every already-added node is visited, that would be
A for the first read, plus A + A for the second one and so
on. In the nth lecture, the amount of traveled nodes would
be:
A + 2 + 3 + . + N
A
A
A
..
Factoring A we get:
( + 2 + . + N * A
1
..
)
Which, using Gauss summation, can be reduced to:
N(N+1) * A =
2
AN² + AN
2
2
+
)
(AN²
=
AN² + AN
= O
AN) O(AN²)
Now, applying Big O product, constant and sum rules we
get:
O
(
That would be the complexity of traversing the graph N
times. The complexity of updating the graph is way more
simple: it walks each read completely, which is O(AN).
The overall complexity, adding the cost of traversing and
updating the graph is O(AN + AN²), and, using the product
we conclude that the complexity of
rule once again,
processing N reads is O(AN²) where A is the average
read length.
Operation
Complexity
Update the graph with a read of length L
O(L)
Traverse the graph when it has V
K-1mers added
O(V)
Update and traverse the graph N times,
with N reads of average length A
Table 2: Complexity of the operations used by the algorithm.
O(AN²)
IMPLEMENTATION
The program execution requires three arguments:
the
filename where the DNA fragments are stored, the number
of genes to be found and the computing timeout.
Figure 3: Help menu when the arguments are incorrects
A fourth parameter is optional, represented by the flag -c;
this argument allows the user to write the start and stop
codons, separated by commas.
Figure 4: Starting the program with the -c argument
If the flag -c is not written, the default start codon is ATG
and the default stop codons are TGA, TAA and TAG.
During the program execution, the found genes are printed
in the screen one by one.
Figure 5: Final results of the execution
At the end of the execution, the program prints the number
of found fragments and the computing time expressed in
milliseconds.
The code can be found at:
https://svn.riouxsvn.com/edya-dnassembly
RESULTS WITH TWO SMALL DNAS
The implementation was tested at assembling mitochondrial
genes from two species: Acipenser transmontanus (White
sturgeon) and Acanthisitta chloris (rifleman). The memory
and time results are shown below:
Species: Acipenser transmontanus
Number of requested
genes
15
30
45
60
Execution time(ms)
38
70
132
284
Memory(MB)
62,8
Table 3: Time and memory results from the tests with
mitochondrial Acipenser transmontanus DNA
23,2
29,4
35,3
Species: Acanthisitta chloris
Number of requested genes
Execution time(ms)
15
53
30
45
60
126
212
249
Memory(MB)
25,9 42,8 60,8
62,6
Table 4: Time and memory results from the tests with
mitochondrial Acanthisitta chloris DNA
Figure 6: Results of time for both tests.
Figure 7: Memory used by both tests.
CONCLUSIONS
The problem of assembling DNA sequences from a group
of fragments using an on-line algorithm is a wide problem
for which every assumption, made or not, can influence in
high degree the difficulty or ease on the solution. This
project's difficulty was increased by the fact that the
algorithm had to be on-line, but also reduced because
mutations were not considered.
REFERENCES
1. Anon. 2017. Computational genomics
https://en.wikipedia.org/wiki/Computational_genomics
Retrieved March 12, 2017.
2. Anon. 2017. Algoritmo Needleman-Wunsch
3. Anon. 2017. Neeedlman-Wunsch Algorithm
http://algoritmoneedleman-wunsch-bio-alevaca.blogspot.c
om.co/ Retrieved March 12, 2017.
https://es.wikipedia.org/wiki/Algoritmo_Needleman-Wun
sch Retrieved March 12, 2017.
https://en.wikipedia.org/wiki/BLAST Retrieved March
12, 2017.
4. Anon. 2017. BLAST
5. Ben Langmead. Assembly & shortest common
superstring. Retrieved March 11, 2017 from
http://www.cs.jhu.edu/~langmea/resources/lecture_notes/
assembly_scs.pdf
6. Alan Frieze and Wojciech Szpankowski. 1997. Greedy
algorithms for the shortest common superstring that are
asymptotically optimal. (April 1997). Retrieved March
11, 2017 from
https://www.math.cmu.edu/~af1p/Texfiles/super.pdf
7. Lisandra Aguilar Bultet and Laurent Falquet. 2015.
Sequencing and de novo assembly of bacterial genomes:
an approach to study new pathogens. Revista de Salud
Animal, 37(2), 125-132 (August 2015). Retrieved March
12, 2017 from
http://scielo.sld.cu/scielo.php?script=sci_arttext&pid=S02
53-570X2015000200008
8. Anon. 2017. K-mer. https://en.wikipedia.org/wiki/K-mer
Retrieved March 12, 2017.
9. Anon. 2017. Eulerian path.
https://en.wikipedia.org/wiki/Eulerian_path Retrieved
March 12, 2017.
10.Ben Langmead. De Bruijn Graph assembly. Retrieved
April 14, 2017 from
http://www.cs.jhu.edu/~langmea/resources/lecture_notes/
assembly_dbg.pdf
from DNAs under, or
The results of time and memory were totally satisfactory
and it is certain that the software performs well assembling
slightly above 20000
genes
nucleobases long. We also conclude that it is impossible to
assembly genes with 100% of correctness by using on-line
algorithms, specially because there is no way to make
predictions about start or stop codons that are yet to come in
further reads.
Thanks to this project, we got started in the world of
scientific computing, using our programming knowledge
for solving a computational genomic problem: assembling a
DNA sequence and finding the genes using an on-line
algorithm.
This approach to the world of science shows us the large
number of problems where the use of computing can make
a difference. Science and computing are nowadays meant to
walk together in order to find answers and solutions for the
benefit of mankind.
ACKNOWLEDGEMENT
Special thanks to Ben Langmead from the John Hopkins
School of engineering, whose courses and online-sources
gave us the theoretical framework for this project.
Also to Santiago Passos Patiño and Juan David Arcila from
EAFIT university,
the space to discuss different
approaches to solve the proposed problem.
FUTURE WORK
The future works could be focused on reducing the
computation time, because we are still solving some
important
subproblems more
improvement would be modifying the algorithm in order to
consider mutations and the statistical models behind this.
than once.
Another
for
|
1705.09177 | 2 | 1705 | 2018-06-06T16:22:13 | On the (parameterized) complexity of recognizing well-covered (r,l)-graphs | [
"cs.DS",
"cs.CC"
] | An $(r, \ell)$-partition of a graph $G$ is a partition of its vertex set into $r$ independent sets and $\ell$ cliques. A graph is $(r, \ell)$ if it admits an $(r, \ell)$-partition. A graph is well-covered if every maximal independent set is also maximum. A graph is $(r,\ell)$-well-covered if it is both $(r,\ell)$ and well-covered. In this paper we consider two different decision problems. In the $(r,\ell)$-Well-Covered Graph problem ($(r,\ell)$WCG for short), we are given a graph $G$, and the question is whether $G$ is an $(r,\ell)$-well-covered graph. In the Well-Covered $(r,\ell)$-Graph problem (WC$(r,\ell)$G for short), we are given an $(r,\ell)$-graph $G$ together with an $(r,\ell)$-partition of $V(G)$ into $r$ independent sets and $\ell$ cliques, and the question is whether $G$ is well-covered. We classify most of these problems into P, coNP-complete, NP-complete, NP-hard, or coNP-hard. Only the cases WC$(r,0)$G for $r\geq 3$ remain open. In addition, we consider the parameterized complexity of these problems for several choices of parameters, such as the size $\alpha$ of a maximum independent set of the input graph, its neighborhood diversity, its clique-width, or the number $\ell$ of cliques in an $(r, \ell)$-partition. In particular, we show that the parameterized problem of deciding whether a general graph is well-covered parameterized by $\alpha$ can be reduced to the WC$(0,\ell)$G problem parameterized by $\ell$. In addition, we prove that both problems are coW[2]-hard but can be solved in XP-time. | cs.DS | cs | On the (parameterized) complexity of recognizing
well-covered (r, (cid:96))-graphs∗
Konrad K. Dabrowski‡
Sancrey Rodrigues Alves†
Sulamita Klein¶
Ignasi Sau(cid:107) ∗∗
Uéverton S. Souza††
Luerbio Faria§
8
1
0
2
n
u
J
6
]
S
D
.
s
c
[
2
v
7
7
1
9
0
.
5
0
7
1
:
v
i
X
r
a
Abstract
An (r, (cid:96))-partition of a graph G is a partition of its vertex set into r independent sets
and (cid:96) cliques. A graph is (r, (cid:96)) if it admits an (r, (cid:96))-partition. A graph is well-covered
if every maximal independent set is also maximum. A graph is (r, (cid:96))-well-covered if
it is both (r, (cid:96)) and well-covered. In this paper we consider two different decision
problems. In the (r, (cid:96))-Well-Covered Graph problem ((r, (cid:96))wc-g for short), we
are given a graph G, and the question is whether G is an (r, (cid:96))-well-covered graph. In
the Well-Covered (r, (cid:96))-Graph problem (wc-(r, (cid:96))g for short), we are given an
(r, (cid:96))-graph G together with an (r, (cid:96))-partition, and the question is whether G is well-
covered. This generates two infinite families of problems, for any fixed non-negative
integers r and (cid:96), which we classify as being P, coNP-complete, NP-complete, NP-
hard, or coNP-hard. Only the cases wc-(r, 0)g for r ≥ 3 remain open. In addition,
we consider the parameterized complexity of these problems for several choices of
parameters, such as the size α of a maximum independent set of the input graph,
its neighborhood diversity, its clique-width, or the number (cid:96) of cliques in an (r, (cid:96))-
partition. In particular, we show that the parameterized problem of determining
whether every maximal independent set of an input graph G has cardinality equal
to k can be reduced to the wc-(0, (cid:96))g problem parameterized by (cid:96). In addition, we
prove that both problems are coW[2]-hard but can be solved in XP-time.
Keywords: well-covered graph; (r, (cid:96))-graph; coNP-completeness; FPT-algorithm; pa-
rameterized complexity; coW[2]-hardness.
∗This work was supported by FAPERJ, CNPq, CAPES Brazilian Research Agencies, EPSRC
(EP/K025090/1), the Leverhulme Trust (RPG-2016-258), and the French ANR projects DEMOGRAPH
(ANR-16-CE40-0028) and ESIGMA (ANR-17-CE40-0028).
†FAETEC, Fundação de Apoio à Escola Técnica do Estado do Rio de Janeiro, Brazil.
[email protected].
‡Department
of Computer
[email protected].
Science, Durham University, Durham, United Kingdom.
§UERJ, DICC, Universidade do Estado do Rio de Janeiro, Brazil. [email protected].
¶UFRJ, COPPE-Sistemas, Universidade Federal do Rio de Janeiro, Brazil. [email protected].
(cid:107)CNRS, LIRMM, Université de Montpellier, Montpellier, France. [email protected].
∗∗Departamento de Matemática, Universidade Federal do Ceará, Fortaleza, Brazil.
††UFF, IC, Universidade Federal Fluminense, Niterói, Brazil. [email protected].
1
1
Introduction
One of the most important combinatorial problems is Maximum Independent Set
(MIS), where the objective is to find a maximum sized subset S ⊆ V of pairwise non-
adjacent vertices in a graph G = (V, E). Maximum independent sets appear naturally
in a wide range of situations, and MIS also finds a number of "real world" relevant
applications.
Unfortunately, the decision version of MIS is an NP-complete problem [23], and thus
it cannot be solved in polynomial time unless P = NP. In spite of the fact that finding a
maximum independent set is a computationally hard problem, a maximal independent
set of a graph can easily be found in linear time. Indeed, a naive greedy algorithm for
finding maximal independent sets consists simply of selecting an arbitrary vertex v to
add to a set S, and updating the current graph by removing the closed neighborhood
N [v] of v. This algorithm always outputs a maximal independent set in linear time, but
clearly not all choices lead to a maximum independent set.
Well-covered graphs were first introduced by Plummer [29] in 1970. Plummer defined
that "a graph is said to be well-covered if every minimal point cover is also a minimum
cover". This is equivalent to demanding that all maximal independent set have the same
cardinality. Therefore, well-covered graphs can be equivalently defined as the class of
graphs for which the naive greedy algorithm discussed above always outputs a maximum
independent set.
The problem of recognizing a well-covered graph, which we denote by Well-
Covered Graph, was proved to be coNP-complete by Chvátal and Slater [3] and
independently by Sankaranarayana and Stewart [33]. On the other hand, the Well-
Covered Graph problem is in P when the input is known to be a perfect graph of
bounded clique size [12] or a claw-free graph [26, 35].
Let r, (cid:96) ≥ 0 be two fixed integers. An (r, (cid:96))-partition of a graph G = (V, E) is
a partition of V into r independent sets S1, . . . , Sr and (cid:96) cliques K1, . . . , K(cid:96). For
convenience, we allow these sets to be empty. A graph is (r, (cid:96)) if it admits an (r, (cid:96))-
partition. Note that the notion of (r, (cid:96))-graphs is a generalization of that of r-colorable
graphs.
A P versus NP-complete dichotomy for recognizing (r, (cid:96))-graphs was proved by Brand-
städt [1]: the problem is in P if max{r, (cid:96)} ≤ 2, and NP-complete otherwise. The class
of (r, (cid:96))-graphs and its subclasses have been extensively studied in the literature. For
instance, list partitions of (r, (cid:96))-graphs were studied by Feder et al. [17].
In another
paper, Feder et al. [18] proved that recognizing graphs that are both chordal and (r, (cid:96))
is in P.
A graph is (r, (cid:96))-well-covered if it is both (r, (cid:96)) and well-covered. In this paper we
2
analyze the complexity of the (r, (cid:96))-Well-Covered Graph problem, which consists
of deciding whether a graph is (r, (cid:96))-well-covered.
In particular, we give a complete
classification of the complexity of this problem.
Additionally, we analyze the complexity of the Well-Covered-(r, (cid:96))-Graph prob-
lem, which consists of deciding, given an (r, (cid:96))-graph G = (V, E) together with an (r, (cid:96))
partition, whether G is well-covered or not. We classify the complexity of this problem
for every pair (r, (cid:96)), except for the cases when (cid:96) = 0 and r ≥ 3, which we leave open.
We note that similar restrictions have been considered in the literature. For instance,
Kolay et al. [24] recently considered the problem of removing a small number of vertices
from a perfect graph so that it additionally becomes (r, (cid:96)).
To the best of our knowledge, this is the first time in the literature that a deci-
sion problem obtained by "intersecting" two recognition NP-complete and coNP-complete
properties has been studied. From our results, the (r, (cid:96))wc-g problem has a very pecu-
liar property, namely that some cases of the problem are in NP, but other cases are in
coNP. And if P (cid:54)= NP, there are some cases where the decision problem is neither in NP
nor in coNP.
In addition, according to the state of the art for the Well-Covered Graph prob-
lem, to the best of our knowledge this is the first work that associates the hardness of
Well-Covered Graph with the number of independent sets and the number of cliques
of an (r, (cid:96))-partition of the input graph. This shows an important structural property
for classifying the complexity of subclasses of well-covered graphs.
As a by-product of this paper, an infinite class of decision problems was classified as
being both NP-hard and coNP-hard. Hence, unless P = NP these decision problems are
neither in NP nor in coNP.
More formally, in this paper we focus on the following two decision problems.
(r, (cid:96))-Well-Covered Graph ((r, (cid:96))wc-g )
Input: A graph G.
Question:
Is G (r, (cid:96))-well-covered?
Well-Covered (r, (cid:96))-Graph (wc-(r, (cid:96))g )
Input: An (r, (cid:96))-graph G, together with a partition of V (G) into
Question:
r independent sets and (cid:96) cliques.
Is G well-covered?
We establish an almost complete characterization of the complexity of the (r, (cid:96))wc-g
and wc-(r, (cid:96))g problems. Our results are shown in the following tables, where r (resp. (cid:96))
3
corresponds to the rows (resp. columns) of the tables, and where coNPc stands for
coNP-complete, NPh stands for NP-hard, NPc stands for NP-complete, and (co)NPh
stands for both NP-hard and coNP-hard. The symbol '?' denotes that the complexity
of the corresponding problem is open.
(r, (cid:96))wc-g
0
1
2
≥ 3
wc-(r, (cid:96))g
0
1
2
≥ 3
0
−
P
P
NPh
0
−
P
P
?
1
P
P
2
P
P
≥ 3
NPc
NPc
coNPc
(co)NPh
coNPc
(co)NPh
(co)NPh
(co)NPh
1
P
P
2
P
P
≥ 3
P
P
coNPc
coNPc
coNPc
coNPc
coNPc
coNPc
We note the following simple facts, which we will use to fill the above tables:
Fact 1. If (r, (cid:96))wc-g is in P, then wc-(r, (cid:96))g is in P.
Fact 2. If wc-(r, (cid:96))g is coNP-hard, then (r, (cid:96))wc-g is coNP-hard.
Note that wc-(r, (cid:96))g is in coNP, since a certificate for a NO-instance consists just
of two maximal independent sets of different size. On the other hand, for (r, (cid:96))wc-g we
have the following facts, which are easy to verify:
Fact 3. For any pair of integers (r, (cid:96)) such that the problem of recognizing an (r, (cid:96))-graph
is in P, the (r, (cid:96))wc-g problem is in coNP.
Fact 4. For any pair of integers (r, (cid:96)) such that the wc-(r, (cid:96))g problem is in P, the
(r, (cid:96))wc-g problem is in NP.
In this paper we prove that (r, (cid:96))wc-g with (r, (cid:96)) ∈ {(0, 1), (1, 0), (0, 2), (1, 1), (2, 0),
(1, 2)} can be solved in polynomial time, which by Fact 1 yields that wc-(r, (cid:96))g with
(r, (cid:96)) ∈ {(0, 1), (1, 0), (0, 2), (1, 1), (2, 0), (1, 2)} can also be solved in polynomial time.
On the other hand, we prove that wc-(2, 1)g is coNP-complete, which by Fact 2 and
Fact 3 yields that (2, 1)wc-g is also coNP-complete. Furthermore, we also prove that
wc-(0, (cid:96))g and wc-(1, (cid:96))g are in P, and that (r, (cid:96))wc-g with (r, (cid:96)) ∈ {(0, 3), (3, 0), (1, 3)}
are NP-hard. Finally, we state and prove a "monotonicity" result, namely Theorem 1,
4
stating how to extend the NP-hardness or coNP-hardness of wc-(r, (cid:96))g (resp. (r, (cid:96))wc-
g) to wc-(r + 1, (cid:96))g (resp. (r + 1, (cid:96))wc-g), and wc-(r, (cid:96) + 1)g (resp. (r, (cid:96) + 1)wc-g).
Together, these results correspond to those shown in the above tables.
In addition, we consider the parameterized complexity of these problems for several
choices of the parameters, such as the size α of a maximum independent set of the input
graph, its neighborhood diversity, its clique-width or the number (cid:96) of cliques in an (r, (cid:96))-
partition. We obtain several positive and negative results. In particular, we show that
the parameterized problem of determining whether every maximal independent set of
an input graph G has cardinality equal to k can be reduced to the wc-(0, (cid:96))g problem
parameterized by (cid:96). In addition, we prove that both problems are coW[2]-hard, but can
be solved in XP-time.
The rest of this paper is organized as follows. We start in Section 2 with some
basic preliminaries about graphs, parameterized complexity, and width parameters. In
Section 3 we prove our results concerning the classical complexity of both problems, and
in Section 4 we focus on their parameterized complexity. We conclude the paper with
Section 5.
2 Preliminaries
Graphs. We use standard graph-theoretic notation, and we refer the reader to [13] for
any undefined notation. A graph G = (V, E) consists of a finite non-empty set V of
vertices and a set E of unordered pairs (edges) of distinct elements of V . If uv ∈ E(G),
then u, v are said to be adjacent, and u is said to be a neighbor of v. A clique (resp.
independent set) is a set of pairwise adjacent (resp. non-adjacent) vertices. A vertex
cover is a set of vertices containing at least one endpoint of every edge in the graph. The
open neighborhood N (v) or neighborhood, for short, of a vertex v ∈ V is the set of vertices
adjacent to v. The closed neighborhood of a vertex v is defined as N [v] = N (v) ∪ {v}. A
v∈S N [v] = V . Given S ⊆ V and
dominating set is a set of vertices S ⊆ V such that(cid:83)
v ∈ V , the neighborhood NS(v) of v in S is the set NS(v) = N (v) ∩ S.
Throughout the paper, we let n denote the number of vertices in the input graph for
the problem under consideration.
Parameterized complexity. We refer the reader to [10,14,19,27] for basic background
on parameterized complexity, and we recall here only some basic definitions. A param-
eterized problem is a language L ⊆ Σ∗ × N. For an instance I = (x, k) ∈ Σ∗ × N, k
is called the parameter. A parameterized problem is fixed-parameter tractable (FPT) if
there exists an algorithm A, a computable function f, and a constant c such that given
an instance I = (x, k), A (called an FPT-algorithm) correctly decides whether I ∈ L in
time bounded by f (k)Ic.
5
Within parameterized problems, the class W[1] may be seen as the parameterized
equivalent to the class NP of classical optimization problems. Without entering into
details (see [10, 14, 19, 27] for the formal definitions), a parameterized problem being
W[1]-hard can be seen as a strong evidence that this problem is not FPT. The canonical
example of a W[1]-hard problem is Independent Set parameterized by the size of the
solution1.
The class W[2] of parameterized problems is a class that contains W[1], and so the
problems that are W[2]-hard are even more unlikely to be FPT than those that are W[1]-
hard (again, see [10, 14, 19, 27] for the formal definitions). The canonical example of a
W[2]-hard problem is Dominating Set parameterized by the size of the solution2.
For i ∈ [1, 2], to transfer W[i]-hardness from one problem to another, one uses an
fpt-reduction, which given an input I = (x, k) of the source problem, computes in time
f (k)Ic, for some computable function f and a constant c, an equivalent instance I(cid:48) =
(x(cid:48), k(cid:48)) of the target problem, such that k(cid:48) is bounded by a function depending only on k.
Hence, an equivalent definition of W[1]-hard (resp. W[2]-hard) problem is any prob-
lem that admits an fpt-reduction from Independent Set (resp. Dominating Set)
parameterized by the size of the solution.
Even if a parameterized problem is W[1]-hard or W[2]-hard, it may still be solvable
in polynomial time for fixed values of the parameter; such problems are said to belong
to the complexity class XP. Formally, a parameterized problem whose instances consist
of a pair (x, k) is in XP if it can be solved by an algorithm with running time f (k)xg(k),
where f, g are computable functions depending only on the parameter and x represents
the input size. For example, Independent Set and Dominating Set parameterized
by the solution size are easily seen to belong to XP.
Width parameters. A tree-decomposition of a graph G = (V, E) is a pair (T,X ),
where T = (I, F ) is a tree, and X = {Bi}, i ∈ I is a family of subsets of V (G), called
bags and indexed by the nodes of T , such that
1. each vertex v ∈ V appears in at least one bag, i.e.,(cid:83)
i∈I Bi = V ;
2. for each edge e = {x, y} ∈ E, there is an i ∈ I such that x, y ∈ Bi; and
3. for each v ∈ V the set of nodes indexed by {i i ∈ I, v ∈ Bi} forms a subtree of T .
The width of a tree-decomposition is defined as maxi∈I{Bi− 1}. The treewidth of G,
denoted by tw(G), is the minimum width of a tree-decomposition of G.
1Given a graph G and a parameter k, the problem is to decide whether there exists an independent
2Given a graph G and a parameter k, the problem is to decide whether there exists a dominating set
set S ⊆ V (G) such that S ≥ k.
S ⊆ V (G) such that S ≤ k.
6
The clique-width of a graph G, denoted by cw(G), is defined as the minimum number
of labels needed to construct G, using the following four operations:
1. Create a single vertex v with an integer label (cid:96) (denoted by (cid:96)(v));
2. Take the disjoint union (i.e., co-join) of two graphs (denoted by ⊕);
3. Join by an edge every vertex labeled i to every vertex labeled j for i (cid:54)= j (denoted
by η(i, j));
4. Relabel all vertices with label i by label j (denoted by ρ(i, j)).
An algebraic term that represents such a construction of G and uses at most k labels is
said to be a k-expression of G (i.e., the clique-width of G is the minimum k for which G
has a k-expression).
Graph classes with bounded clique-width include cographs [2], distance-hereditary
graphs [20], graphs of bounded treewidth [9], graphs of bounded branchwidth [32], and
graphs of bounded rank-width [22].
3 Classical complexity of the problems
We start with a monotonicity theorem that will be very helpful to fill the tables presented
in Section 1. The remainder of this section is divided into four subsections according to
whether (r, (cid:96))wc-g and wc-(r, (cid:96))g are polynomial or "hard" problems.
Theorem 1. Let r, (cid:96) ≥ 0 be two fixed integers. Then it holds that:
(i) if wc-(r, (cid:96))g is coNP-complete then wc-(r + 1, (cid:96))g and wc-(r, (cid:96) + 1)g are coNP-
complete;
(ii) if (r, (cid:96))wc-g is NP-hard (resp. coNP-hard) then (r, (cid:96) + 1)wc-g is NP-hard (resp.
coNP-hard);
(iii) supposing that r ≥ 1, if (r, (cid:96))wc-g is NP-hard (resp. coNP-hard) then (r + 1,
(cid:96))wc-g is NP-hard (resp. coNP-hard).
Proof. (i) This follows immediately from the fact that every (r, (cid:96))-graph is also an
(r + 1, (cid:96))-graph and an (r, (cid:96) + 1)-graph.
(ii) Let G be an instance of (r, (cid:96))wc-g. Let H be an (r, (cid:96) + 1)wc-g instance
defined as the disjoint union of G and a clique Z with V (Z) = {z1, . . . , zr+1}. Clearly G
is well-covered if and only if H is well-covered. If G is an (r, (cid:96))-well-covered graph then H
is an (r, (cid:96) + 1)-well-covered graph. Suppose H is an (r, (cid:96) + 1)-well-covered graph, with
7
a partition into r independent sets S1, . . . , Sr and (cid:96) + 1 cliques K1, . . . , K(cid:96)+1. Each
independent set Si can contain at most one vertex of the clique Z. Therefore, there
must be a vertex zi in some clique Kj. Assume without loss of generality that there
is a vertex of Z in K(cid:96)+1. Then K(cid:96)+1 cannot contain any vertex outside of V (Z), so
we may assume that K(cid:96)+1 contains all vertices of Z. Now S1, . . . , Sr, K1, . . . , K(cid:96) is an
(r, (cid:96))-partition of G, so G is an (r, (cid:96))-well-covered graph. Hence, H is a YES-instance of
(r, (cid:96) + 1)wc-g if and only if G is a YES-instance of (r, (cid:96))wc-g.
(iii) Let G be an instance of (r, (cid:96))wc-g. Let G(cid:48) be an (r + 1, (cid:96))wc-g instance
obtained from G by adding (cid:96) + 1 isolated vertices. (This guarantees that every maximal
independent set in G(cid:48) contains at least (cid:96) + 1 vertices.) Since r ≥ 1, it follows that G(cid:48) is
an (r, (cid:96))-graph if and only if G is. Clearly G(cid:48) is well-covered if and only if G is.
Next, find an arbitrary maximal independent set in G(cid:48) and let p be the number of
vertices in this set. Note that p ≥ (cid:96)+1. Let H be the join of G(cid:48) and a set of p independent
vertices Z = {z1, . . . , zp}, i.e., NH (zi) = V (G(cid:48)) for all i. Every maximal independent set
of H is either Z or a maximal independent set of G(cid:48) and every maximal independent set
of G(cid:48) is a maximal independent set of H. Therefore, H is well-covered if and only if G(cid:48)
is well-covered. Clearly, if G(cid:48) is an (r, (cid:96))-graph then H is an (r + 1, (cid:96))-graph. Suppose H
is an (r + 1, (cid:96))-graph, with a partition into r + 1 independent sets S1, . . . , Sr+1 and (cid:96)
cliques K1, . . . , K(cid:96). Each clique set Ki can contain at most one vertex of Z. Therefore
there must be a vertex zi in some independent set Sj. Suppose that there is a vertex of Z
in Sr+1. Then Sr+1 cannot contain any vertex outside of Z. Without loss of generality,
we may assume that Sr+1 contains all vertices of Z. Now S1, . . . , Sr, K1, . . . , K(cid:96) is an
(r, (cid:96))-partition of G, so G is an (r, (cid:96))-graph. Thus H is a YES-instance of (r + 1, (cid:96))wc-g
if and only if G is a YES-instance of (r, (cid:96))wc-g.
3.1 Polynomial cases for wc-(r, (cid:96))g
Theorem 2. wc-(0, (cid:96))g and wc-(1, (cid:96))g are in P for every integer (cid:96) ≥ 0.
Proof. It is enough to prove that wc-(1, (cid:96))g is in P. Let V = (S, K1, K2, K3, . . . , K(cid:96))
be a (1, (cid:96))-partition for G. Then each maximal independent set I of G admits a partition
I = (IK, S \ NS(IK)), where IK is an independent set of K1 ∪ K2 ∪ K3 ∪ ···∪ K(cid:96).
Observe that there are at most O(n(cid:96)) choices for an independent set IK of K1∪ K2∪
K3∪···∪ K(cid:96), which can be listed in time O(n(cid:96)), since (cid:96) is constant and (K1, K2, K3, . . . ,
K(cid:96)) is given. For each of them, we consider the independent set I = IK ∪ (S \ NS(IK)).
If I is not maximal (which may happen if a vertex in (K1 ∪ K2 ∪ K3 ∪ ···∪ K(cid:96)) \ IK
has no neighbors in I), we discard this choice of IK. Hence, we have a polynomial
number O(n(cid:96)) of maximal independent sets to check in order to decide whether G is a
well-covered graph.
8
3.2 Polynomial cases for (r, (cid:96))wc-g
Fact 5. The graph induced by a clique or by an independent set is well-covered.
The following corollary is a simple application of Fact 5.
Corollary 1. G is a (0, 1)-well-covered graph if and only if G is a (0, 1)-graph. Similarly,
G is a (1, 0)-well-covered graph if and only if G is a (1, 0)-graph.
The following is an easy observation.
Lemma 1. (0, 2)wc-g can be solved in polynomial time.
Proof. By definition, a graph G = (V, E) is a (0, 2)-graph if and only if its vertex set
can be partitioned into two cliques, and this can be tested in polynomial time. It follows
that every (0, 2)-graph has maximum independent sets of size at most 2. Let G be
a (0, 2)-graph with (0, 2)-partition (K1, K2). If V is a clique, then G is a (0, 1)-well-
covered graph, and hence a (0, 2)-well-covered graph. If V is not a clique, then G is a
(0, 2)-well-covered graph if and only if G has no universal vertex.
In the next three lemmas we give a characterization of (1, 1)-well-covered graphs in
terms of their graph degree sequence. Note that (1, 1)-graphs are better known in the
literature as split graphs.
Lemma 2. Let G = (V, E) be a (1, 1)-well-covered graph with (1, 1)-partition V = (S, K),
where S is a independent set and K is a clique. If x ∈ K, then NS(x) ≤ 1.
Proof. Suppose that G is a (1, 1)-well-covered graph with (1, 1)-partition V = (S, K),
where S is a independent set and K is a clique. Let I be a maximal independent set of G
such that x ∈ I ∩ K. Suppose for contradiction that NS(x) ≥ 2, and let y, z ∈ NS(x).
Since y, z ∈ S, NG(y), NG(z) ⊆ K. Since K is a clique, vertex x is the only vertex
of I in K. Hence, we have that NG(y) ∩ (I \ {x}) = NG(z) ∩ (I \ {x}) = ∅. Therefore
I(cid:48) = (I \ {x}) ∪ {y, z} is an independent set of G such that I(cid:48) = I + 1. Thus, I
is a maximal independent set that is not maximum, so G is not well-covered. Thus,
NS(x) ≤ 1.
Lemma 3. A graph G is a (1, 1)-well-covered graph if and only if it admits a (1, 1)-
partition V = (S, K) such that either for every x ∈ K, NS(x) = 0, or for every x ∈ K,
NS(x) = 1.
Proof. Let G be a (1, 1)-well-covered graph. By Lemma 2 we have that, given a vertex
x ∈ K, either NS(x) = 0 or NS(x) = 1. Suppose for contradiction that there are
two vertices x, y ∈ K such that NS(x) = 0 and NS(y) = 1. Let z be the vertex of S
9
adjacent to y. Let I be a maximal independent set containing vertex y. Note that the
vertex x is non-adjacent to every vertex of I \ {y} since there is at most one vertex of I
in K. The same applies to the vertex z. Hence, a larger independent set I(cid:48), with size
I(cid:48) = I+1, can be obtained from I by replacing vertex y with the non-adjacent vertices
x, z, i.e., I is a maximal independent set of G that is not maximum, a contradiction.
Thus, either for every x ∈ K, NS(x) = 0, or for every x ∈ K, NS(x) = 1.
Conversely, suppose that there is a (1, 1)-partition V = (S, K) of G such that either
for every x ∈ K, NS(x) = 0, or for every x ∈ K, NS(x) = 1. If K = ∅, then G is (1, 0)
and then G is well-covered. Hence we assume K (cid:54)= ∅. If for every x ∈ K, NS(x) = 0,
then every maximal independent set consists of all the vertices of S and exactly one
vertex v ∈ K. If for every x ∈ K, NS(x) = 1, then every maximal independent set is
either I = S, or I = {x} ∪ (S \ NS(x)) for some x ∈ K. Since NS(x) = 1 we have
I = 1 + S − 1 = S, and hence G is a (1, 1)-well-covered graph.
Corollary 2. (1, 1)wc-g can be solved in polynomial time.
Proof. Since we can check in polynomial time whether G is a (1, 1) graph [1], and one
can enumerate all (1, 1)-partitions of a split graph in polynomial time, we can solve the
(1, 1)wg-g problem in polynomial time.
The next lemma shows that (1, 1)-well-covered graphs can be recognized from their
degree sequences.
sequence of V is either (k, k, k, . . . , k, i1, i2, . . . , is, 0, 0, 0, . . . , 0) with (cid:80)s
Lemma 4. G is a (1, 1)-well-covered graph if and only if there is a positive integer k such
that G is a graph with a (1, 1)-partition V = (S, K) where K = k, such that the degree
j=1(ij) = k,
or (k − 1, k − 1, k − 1, . . . , k − 1, 0, 0, 0, . . . , 0), where the subsequences k, . . . , k (resp.
k − 1, . . . , k − 1) have length k.
Proof. Let G be a (1, 1)-well-covered graph. Then G admits a (1, 1)-partition V =
(S, K) where k := K, k ≥ 0. If k = 0, then the degree sequence is (0, 0, 0, . . . , 0). If
k ≥ 1, then by Lemma 3 either for every x ∈ K, NS(x) = 0, or for every x ∈ K,
NS(x) = 1.
If for every x ∈ K, NS(x) = 0, then the degree sequence of G is
(k − 1, k − 1, k − 1, . . . , k − 1, 0, 0, 0, . . . , 0). If for every x ∈ K, NS(x) = 1, then the
degree sequence of G is (k, k, k, . . . , k, i1, i2, . . . , is, 0, 0, 0, . . . , 0), with(cid:80)s
0, 0, 0, . . . , 0), or (k − 1, k − 1, k − 1, . . . , k − 1, 0, 0, 0, . . . , 0), such that(cid:80)s
Suppose that there is a positive integer k such that G is a graph with (1, 1)-partition
V = (S, K) where K = k, with degree sequence either (k, k, k, . . . , k, i1, i2, . . . , is,
j=1(ij) = k. If
the degree sequence of G is (k, k, k, . . . , k, i1, i2, . . . , is, 0, 0, 0, . . . , 0), then the vertices
of K are adjacent to k − 1 vertices of K and exactly one of S, since the vertices with
j=1(ij) = k.
10
degree i1, i2, . . . , is, have degree at most k and the vertices with degree 0 are isolated. If
the degree sequence of G is (k − 1, k − 1, k − 1, . . . , k − 1, 0, 0, 0, . . . , 0), then the vertices
of K are adjacent to k − 1 vertices of K and none of S and the vertices with degree 0
are isolated. By Lemma 3 we have that G is a well-covered graph.
Ravindra [31] gave the following characterization of (2, 0)-well-covered graphs.
Proposition 3 (Ravindra [31]). Let G be a connected graph. G is a (2, 0)-well-covered
graph if and only if G contains a perfect matching F such that for every edge e = uv
in F , G[N (u) ∪ N (v)] is a complete bipartite graph.
We now prove that Proposition 3 leads to a polynomial-time algorithm.
Lemma 5. (2, 0)wc-g can be solved in polynomial time.
Proof. Assume that G is connected and consider the weighted graph (G, ω) with ω :
E(G) → {0, 1} satisfying ω(uv) = 1, if G[N (u) ∪ N (v)] is a complete bipartite graph,
and 0 otherwise. By Proposition 3, G is well-covered if and only if (G, ω) has a weighted
perfect matching with weight at least n/2, and this can be decided in polynomial
time [15].
Lemma 6. (1, 2)wc-g can be solved in polynomial time.
Proof. We can find a (1, 2)-partition of a graph G (if such a partition exists) in polynomial
time [1]. After that, we use the algorithm for wc-(1, (cid:96))g given by Theorem 2.
Below we summarize the cases for which we have shown that wc-(r, (cid:96))g or (r, (cid:96))wc-g
can be solved in polynomial time.
Theorem 4. (r, (cid:96))wc-g with (r, (cid:96)) ∈ {(0, 1), (0, 2), (1, 0), (1, 1), (1, 2), (2, 0)} and wc-
(r, (cid:96))g with r ∈ {0, 1} or (r, (cid:96)) = (2, 0) can be solved in polynomial time.
Proof. The first part follows from Corollary 1, Lemma 1, Corollary 2, Lemma 6, and
Lemma 5, respectively. The second part follows from Theorem 2, and Lemma 5 together
with Fact 1.
3.3 coNP-complete cases for wc-(r, (cid:96))g
We note that the well-Covered Graph instance G constructed in the reduction
of Chvátal and Slater [3] is a (2, 1)-graph, directly implying that wc-(2, 1)g is coNP-
complete.
Indeed, Chvátal and Slater [3] take a 3-sat instance I = (U, C) = ({u1, u2, u3,
. . . , un}, {c1, c2, c3, . . . , cm}), and construct a Well-Covered Graph instance G =
11
1:
and
Slater's
Chvátal
[3] Well-Covered Graph instance
Figure
G = (V, E) obtained from the satisfiable 3-sat instance I = (U, C) =
({u1, u2, u3},{(u1, u2, u3), (u1, u2, u3), (u1, u2, u3)}), where {c1, c2, . . . , cm} is a clique
of G. Observe that I is satisfiable if and only if G is not well-covered, since there is a
maximal independent set with size n + 1 (e.g. {c1, u1, u2, u3}) and there is a maximal
independent set of size n (e.g. {u1, u2, u3}). Note also that G is a (2, 1)-graph with
(2, 1)-partition V = ({u1, u2, . . . , un},{u1, u2, . . . , un},{c1, c2, . . . , cm} ).
(V, E) = ( {u1, u2, u3, . . . , un, u1, u2, u3, . . . , un, c1, c2, c3, . . . , cm}, {xcj : x occurs in
cj} ∪ {uiui : 1 ≤ i ≤ n} ∪ {cicj : 1 ≤ i < j ≤ m} ). Note that {cicj : 1 ≤ i < j ≤ m}
is a clique, and that {u1, u2, u3, . . . , un}, and {u1, u2, u3, . . . , un} are independent sets.
Hence, G is a (2, 1)-graph. An illustration of this construction can be found in Figure 1.
This discussion can be summarized as follows.
Proposition 5 (Chvátal and Slater [3]). wc-(2, 1)g is coNP-complete.
As (2, 1)-graphs can be recognized in polynomial time [1], we obtain the following.
Corollary 3. (2, 1)wc-g is coNP-complete.
3.4 NP-hard cases for (r, (cid:96))wc-g
Now we prove that (0, 3)wc-g is NP-complete. For this purpose, we slightly modify an
NP-completeness proof of Stockmeyer [34].
Stockmeyer's [34] NP-completeness proof of 3-coloring considers a 3-sat in-
stance I = (U, C) = ( {u1, u2, u3, . . . , un},{c1, c2, c3, . . . , cm} ), and constructs
a 3-coloring instance G = (V, E) = ({u1, u2, u3, . . . , un, u1, u2, u3, . . . , un} ∪
{v1[j], v2[j], v3[j], v4[j], v5[j], v6[j]
:
i ∈ {1, 2, 3, . . . , n}} ∪ {v1[j]v2[j], v2[j]v4[j], v4[j]v1[j], v4[j]v5[j], v5[j]v6[j], v6[j]v3[j],
: j ∈ {1, 2, 3, . . . , m}} ∪ {v1[j]x, v2[j]y, v3[j]z : cj = (x, y, z)} ∪ {t1ui,
v3[j]v5[j]
t1 ¯ui : i ∈ {1, 2, 3, . . . , n}} ∪ {t2v6[j] : j ∈ {1, 2, 3, . . . , m}} ); see Figure 2(a).
Lemma 7. (0, 3)wc-g is NP-complete.
{1, 2, 3, . . . , m}} ∪ {t1, t2},{uiui
∈
j
:
Proof. As by Theorem 2 the Well-Covered Graph problem can be solved in poly-
nomial time on (0, 3)-graphs, by Fact 4 (0, 3)wc-g is in NP.
12
u1u1u2u2u3u3c1c2c3 Figure 2: (a) Stockmeyer's [34] 3-coloring instance G obtained from the 3-sat instance
I = (U, C) = ({u1, u2, u3},{(u3, u2, u1), (u1, u2, u3)}). (b) The graph G(cid:48) obtained from G
by adding a vertex xuv with NG(cid:48)(xuv) = {u, v} for every edge uv of G not belonging to
a triangle.
Let I = (U, C) be a 3-sat instance. We produce, in polynomial time in the size of I,
a (0, 3)wc-g instance H, such that I is satisfiable if and only if H is (0, 3)-well-covered.
Let G = (V, E) be the graph of [34] obtained from I, and let G(cid:48) be the graph obtained
from G by adding to V a vertex xuv for every edge uv of G not belonging to a triangle,
and by adding to E edges uxuv and vxuv; see Figure 2(b). Finally, we define H = G(cid:48) as
the complement of G(cid:48). Note that, by [34], I is satisfiable if and only if G is 3-colorable.
Since xuv is adjacent to only two different colors of G, clearly G is 3-colorable if and only
if G(cid:48) is 3-colorable. Hence, I is satisfiable if and only if H is a (0, 3)-graph. We prove
next that I is satisfiable if and only if H is a (0, 3)-well-covered graph.
Suppose that I is satisfiable. Then, since H is a (0, 3)-graph, every maximal indepen-
dent set of H has size 3, 2, or 1. If there is a maximal independent set I in H with size
1 or 2, then I is a maximal clique of G(cid:48) of size 1 or 2. This contradicts the construction
of G(cid:48), since every maximal clique of G(cid:48) is a triangle. Therefore, G is well-covered.
Suppose that H is (0, 3)-well-covered. Then G(cid:48) is 3-colorable, so G is also 3-colorable.
Thus, by [34], I is satisfiable.
We next prove that (3, 0)wc-g is NP-hard. For this, we again use the proof of
Stockmeyer [34], together with the following theorem.
Proposition 6 (Topp and Volkmann [36]). Let G = (V, E) be an n-vertex graph,
V = {v1, v2, v3, . . . , vn}, and let H be obtained from G such that V (H) = V ∪
13
v3[1]v2[1]v5[1]v4[1]v6[1]v1[1]v1[2]v4[2]v5[2]v2[2]v3[2]v6[2]v3[1]v2[1]v5[1]v4[1]v6[1]v1[1]v1[2]v4[2]v5[2]v2[2]v3[2]v6[2]NFT/FT/FT/FT/FT/FT/Ft1t2u1u1u2u2u3u3NFT/FT/FT/FT/FT/FT/Ft1t2u1u1u2u2u3u3(a)(b) {u1, u2, u3, . . . , un} and E(H) = E ∪ {viui : i ∈ {1, 2, 3, . . . , n}}. Then H is a well-
covered graph where every maximal independent set has size n.
Proof. Observe that every maximal independent set I of H has a subset IG = I ∩ V .
Let U ⊆ {1, 2, 3, . . . , n} be the set of indices i such that vi ∈ I. Since I is maximal, the
set {ui : i ∈ {1, 2, 3, . . . , n} \ U} must be contained in I, so I = n.
Lemma 8. (3, 0)wc-g is NP-hard.
Proof. Let I = (U, C) be a 3-sat instance; let G = (V, E) be the graph obtained from I
in Stockmeyer's [34] NP-completeness proof for 3-coloring; and let H be the graph
obtained from G by the transformation described in Proposition 6. We prove that I is
satisfiable if and only if H is a (3, 0)-well-covered graph. Suppose that I is satisfiable.
Then by [34] we have that G is 3-colorable. Since a vertex v ∈ V (H) \ V (G) has just
one neighbor, there are 2 colors left for v to extend a 3-coloring of G, and so H is a
(3, 0)-graph. Hence, by Proposition 6, H is a (3, 0)-well-covered graph. Suppose that H
is a (3, 0)-well-covered graph. Then we have that G is a (3, 0)-graph. By [34], I is
satisfiable.
Note that Theorem 1 combined with Lemma 7 does not imply that (1, 3)wc-g is
NP-complete.
Lemma 9. (1, 3)wc-g is NP-complete.
Proof. As by Theorem 2 the Well-Covered Graph problem can be solved in poly-
nomial time on (1, 3)-graphs, by Fact 4 (1, 3)wc-g is in NP.
Let I = (U, C) be a 3-sat instance. Without loss of generality, I has more than two
clauses. We produce a (1, 3)wc-g instance H polynomial in the size of I, such that I is
satisfiable if and only if H is (1, 3)-well-covered.
Let G = (V, E) be the graph of Stockmeyer [34] obtained from I (see Figure 2(a)),
and let H be the graph obtained from G (the complement of the graph G) by adding one
pendant vertex pv for each vertex v of G. Note that V (H) = V (G) ∪ {pv : v ∈ V (G)},
E(H) = E(G) ∪ {pvv : v ∈ V (G)}, and NH (pv) = {v}.
First suppose that I is satisfiable. Then by [34], G is a (3, 0)-graph, and G is a
). Thus it follows that
(0, 3)-graph with partition into cliques V (G) = (K1
(S = {pv : v ∈ V (G)}, K1
G
In addition, from
Proposition 6 and by the construction of H, H is a well-covered graph. Hence H is
(1, 3)-well-covered.
) is a (1, 3)-partition of V (H).
, K2
G
, K2
G
, K3
G
G
, K3
G
Conversely, suppose that H is (1, 3)-well-covered, and let V (H) = (S, K1, K2, K3)
be a (1, 3)-partition for H. Then we claim that no vertex pv ∈ V (H) \ V (G) belongs to
Ki, i ∈ {1, 2, 3}. Indeed, suppose for contradiction that pv ∈ Ki for some i ∈ {1, 2, 3}.
14
Then, Ki ⊆ {pv, v}. Hence, H \ Ki is a (1, 2)-graph and G \ {v} is an induced subgraph
of a (2, 1)-graph. But by construction of G, G \ {v} (for any v ∈ V (G)) contains
at least one 2K3 (that is, two vertex-disjoint copies of K3) as an induced subgraph,
which is a contradiction given that 2K3 is clearly a forbidden subgraph for (2, 1)-graphs.
Therefore, {pv : v ∈ V (G)} ⊆ S, and since {pv : v ∈ V (G)} is a dominating set of H,
S = {pv : v ∈ V (G)}. Thus, G is a (0, 3)-graph with partition V (G) = (K1, K2, K3),
and therefore G is a (3, 0)-graph, i.e., a 3-colorable graph. Therefore, by [34], I is
satisfiable.
Corollary 4. If r ≥ 3 and (cid:96) = 0, then (r, (cid:96))wc-g is NP-hard. If r ∈ {0, 1} and (cid:96) ≥ 3,
then (r, (cid:96))wc-g is NP-complete.
Proof. (r, (cid:96))wc-g is NP-hard in all of these cases by combining Theorem 1, and Lem-
mas 7, 8 and 9. For r ∈ {0, 1} and (cid:96) ≥ 3, the Well-Covered Graph problem can be
solved in polynomial time on (r, (cid:96))-graphs, so by Fact 4 (r, (cid:96))wc-g is in NP.
Below we summarize the cases for which we have shown that wc-(r, (cid:96))g or (r, (cid:96))wc-g
is computationally hard.
Theorem 7. The following classification holds:
1. wc-(r, (cid:96))g with r ≥ 2 and (cid:96) ≥ 1 are coNP-complete;
2. (0, (cid:96))wc-g and (1, (cid:96))wc-g with (cid:96) ≥ 3 are NP-complete;
3. (2, 1)wc-g and (2, 2)wc-g are coNP-complete;
4. (r, (cid:96))wc-g with r ≥ 0 and (cid:96) ≥ 3 is NP-hard;
5. (r, (cid:96))wc-g with r ≥ 3 and (cid:96) ≥ 0 is NP-hard;
6. (r, (cid:96))wc-g with r ≥ 2 and (cid:96) ≥ 1 is coNP-hard.
Proof. Statement 1 follows from Proposition 5 and Theorem 1(i). Statement 2 follows
from Corollary 4. Statement 3 follows from Statement 1, Facts 2 and 3 and the fact that
recognizing (r, (cid:96))-graphs is in P if max{r, (cid:96)} ≤ 2 [1]. Statement 4 follows from Statement 2
and Theorem 1(ii)-(iii). Statement 5 follows from Lemma 8 and Theorem 1(ii)-(iii).
Finally, Statement 6 follows from Corollary 3 and Theorem 1(ii)-(iii).
15
4 Parameterized complexity of the problems
In this section we focus on the parameterized complexity of the Well-Covered Graph
problem, with special emphasis on the case where the input graph is an (r, (cid:96))-graph.
Recall that the results presented in Section 2 show that wc-(r, (cid:96))g is para-coNP-complete
when parameterized by r and (cid:96). Thus, additional parameters should be considered.
Henceforth we let α (resp. ω) denote the size of a maximum independent set (resp.
maximum clique) in the input graph G for the problem under consideration. Note that
wc-(r, (cid:96))g parameterized by r, (cid:96), and ω generalizes wc-(r, 0)g, whose complexity was
left open in the previous sections. Therefore, we focus on the complexity of wc-(r, (cid:96))g
parameterized by r, (cid:96), and α, and on the complexity of the natural parameterized version
of Well-Covered Graph, defined as follows:
k-Well-Covered Graph
Input: A graph G and an integer k.
Parameter:
k.
Question: Does every maximal independent set of G have size exactly k?
The next lemma provides further motivation to study of the wc-(0, (cid:96))g problem,
as it shows that k-Well-Covered Graph (on general graphs) can be reduced to the
wc-(0, (cid:96))g problem parameterized by (cid:96).
Lemma 10. The k-Well-Covered Graph problem can be fpt-reduced to the wc-
(0, (cid:96))g problem parameterized by (cid:96).
Proof. Consider an arbitrary input graph G with vertices u1, . . . , un. First, we find an
arbitrary maximal (with respect to set-inclusion) independent set I in G. Without loss
of generality we may assume that I = k and I = {u1, . . . , uk}. Let (cid:96) = k + 1.
We construct a (0, (cid:96))-graph G(cid:48) with vertex set {vi,j : i ∈ {1, . . . , (cid:96)}, j ∈ {1, . . . , n}}
as follows:
• For all i ∈ {1, . . . , (cid:96)} add edges to make Vi := {vi,j : j ∈ {1, . . . , n}} into a clique.
• For all j ∈ {1, . . . , n} add edges to make Wj := {vi,j : i ∈ {1, . . . , (cid:96)}} into a clique.
• For all pairs of adjacent vertices ua, ub in G, add edges between vi,a and vj,b for
all i, j ∈ {1, . . . , (cid:96)} (so that Va is complete to Vb).
Note that the sets Vi partition G(cid:48) into (cid:96) cliques, so G(cid:48) is indeed a (0, (cid:96))-graph, where
(cid:96) = k + 1.
16
The graph G(cid:48) has a maximal independent set of size k, namely {v1,1, . . . , vk,k}, so G(cid:48)
is well-covered if and only if every maximal independent set in G(cid:48) has size exactly k.
Every maximal independent set in G(cid:48) has at most one vertex in any set Vi and at
most one vertex in any set Wj, since Vi and Wj are cliques. As there are (cid:96) = k + 1
sets Vi, it follows that every independent set in G(cid:48) contains at most k + 1 vertices.
If G(cid:48) contains an independent set {vi1,j1, . . . , vix,jx} for some x then {uj1, . . . , ujx} is
an independent set in G. If G contains an independent set {uj1, . . . , ujx} for some x
then {v1,j1, . . . , vmin(x,k+1),jmin(x,k+1)
} is an independent set in G(cid:48). Therefore G contains a
maximal independent set smaller than k if and only if G(cid:48) contains a maximal independent
set smaller than k and G contains a (not necessarily maximal) independent set of size
at least k + 1 if and only G(cid:48) contains a maximal independent set of size exactly k + 1.
It follows that G(cid:48) is well-covered if and only if G is. As (cid:96) = k + 1, this completes the
proof.
Recall that the Well-Covered Graph problem is coNP-complete [3,33]. In order
to analyze the parameterized complexity of the problem, we will need the following
definition.
Definition 1. The class coW[2] is the class of all parameterized problems whose com-
plement is in W[2].
For an overview of parameterized complexity classes, see [11, 19].
We are now ready to show the next result.
Theorem 8. The wc-(0, (cid:96))g problem parameterized by (cid:96) is coW[2]-hard.
Proof. Red-Blue Dominating Set (RBDS) is a well-known W[2]-complete prob-
lem [14], which consists of determining whether a given bipartite graph G = (R ∪ B, E)
admits a set D ⊆ R of size k (the parameter) such that D dominates B (that is, every
vertex in B has a neighbor in D). To show the coW[2]-hardness of our problem, we
present an fpt-reduction from Red-Blue Dominating Set to the problem of deter-
mining whether a given (0, (cid:96))-graph is not well-covered, where (cid:96) = k + 1.
m}, . . . , Rk = {rk
From an instance (G, k) of RBDS we construct a (0, (cid:96))-graph G(cid:48) as follows.
m}, R2 =
Replace the set R = {r1, r2, . . . , rm} by k copies: R1 = {r1
{r2
m}, where each new vertex has the same neigh-
borhood as the corresponding vertex did in G. Add edges to make B, as well as
each Ri for 1 ≤ i ≤ k, induce a clique. For each i ∈ {1, . . . , k}, create a vertex si,
and add all possible edges between si and the vertices in Ri. Let G(cid:48) be the result-
ing graph. Note that the vertex set of G(cid:48) can be partitioned into (cid:96) = k + 1 cliques:
B, R1 ∪ {s1}, R2 ∪ {s2}, . . . , Rk ∪ {sk}.
1 , rk
2 , . . . , rk
1, r2
2, . . . , r2
1, r1
2, . . . , r1
17
Clearly, for every b ∈ B, the set {s1, s2, . . . , sk} ∪ {b} is an independent set of G(cid:48) of
size k + 1. Note that such an independent set is maximum, as it contains one vertex from
each of the k + 1 cliques that partition V (G(cid:48)). In addition, any maximal independent
set of G(cid:48) has size at least k, since every maximal independent set contains either si or a
vertex of Ri. At this point, we claim that G has a set D ⊆ R of size k which dominates B
if and only if G(cid:48) has a maximal independent set of size k (i.e., G(cid:48) is not well-covered).
If D = {ri1, ri2, . . . , rik} is a subset of R of size k which dominates B in G, then
} is a maximal independent set of G(cid:48), implying that G(cid:48) is not
, . . . , rk
ik
D(cid:48) = {r1
, r2
i2
well-covered.
i1
Conversely, if G(cid:48) is not well-covered then there exists in G(cid:48) a maximal independent
set D(cid:48) of size k. Note that D(cid:48) ∩ B = ∅ and each vertex in B has at least one neighbor
in D(cid:48), as otherwise D(cid:48) would not be a maximal independent set of size k. Therefore, by
letting D be the set of vertices in R that have copies in D(cid:48) ∩ {R1 ∪ R2 ∪ . . . ∪ Rk}, we
find that D is a subset of R of size at most k which dominates B in G.
From the previous theorem we immediately obtain the following corollaries.
Corollary 5. The k-Well-Covered Graph problem is coW[2]-hard.
Proof. This follows immediately from Lemma 10 and Theorem 8.
Corollary 6. Unless FPT = coW[2], the wc-(r, (cid:96))g problem cannot be solved in time
f (α + (cid:96))ng(r) for any computable function f.
Proof. This follows from the fact that an algorithm running in time f (α + (cid:96))ng(r), would
be an FPT-algorithm for wc-(0, (cid:96))g parameterized by (cid:96), and from the coW[2]-hardness
of the problem demonstrated in Theorem 8.
In contrast to Corollary 6, Lemma 11 shows that the wc-(r, (cid:96))g problem can be
solved in time 2rαnO((cid:96)).
Lemma 11. The wc-(r, (cid:96))g problem can be solved in time 2rαnO((cid:96)). In particular, it is
FPT when (cid:96) is fixed and r, α are parameters.
Proof. Note that each of the r independent sets S1, . . . , Sr of the given partition of V (G)
must have size at most α. On the other hand, any maximal independent set of G contains
at most one vertex in each of the (cid:96) cliques. The algorithm exhaustively constructs all
i=1 Si, and
then choose at most one vertex in each clique. For each choice, we just have to verify
whether the constructed set is a maximal independent set, and then check that all the
constructed maximal independent sets have the same size. The claimed running time
1≤i≤r Si,
maximal independent sets of G as follows: we start by guessing a subset of(cid:83)r
follows. In fact, in the statement of the lemma, one could replace rα with(cid:80)
which yields a stronger result.
18
Although wc-(1, (cid:96))g parameterized by (cid:96) is coW[2]-hard (see Theorem 8), Theorem 2
shows that the problem is in XP.
Corollary 7. The wc-(1, (cid:96))g problem can be solved in time nO((cid:96)). In other words, it is
in XP when parameterized by (cid:96).
Proof. This follows from Theorem 2 by considering (cid:96) to not be a constant.
Table 1 summarizes the results presented so far. Note that, by Ramsey's Theo-
rem [30], when both ω and α are parameters the input graph itself is a trivial kernel.
Table 1: Parameterized complexity of wc-(r,(cid:96))g.
Param.\Class
r
(cid:96)
r, (cid:96)
r, (cid:96), ω
r, (cid:96), α
ω, α
(0, (cid:96))
--
coW[2]-h
XP
coW[2]-h
XP
FPT
Trivial
coW[2]-h
XP
(1, (cid:96))
--
coW[2]-h
XP
coW[2]-h
XP
FPT
Trivial
coW[2]-h
XP
(r, (cid:96))
para-coNP-h
para-coNP-h
para-coNP-h
Open
(generalizes wc-(3,0)g)
coW[2]-h,
no f (α + (cid:96))ng(r) algo.
unless FPT=coW[2],
algo. in time 2rαnO((cid:96))
FPT
FPT
FPT
Ramsey's Thm.
Ramsey's Thm.
Ramsey's Thm.
4.1 Taking the neighborhood diversity as the parameter
Neighborhood diversity is a structural parameter based on a special way of partitioning
a graph into independent sets and cliques. Therefore, it seems a natural parameter to
consider for our problem, since an (r, (cid:96))-partition of a graph G is also a partition of its
vertex set into cliques and independent sets.
Definition 2 (Lampis [25]). The neighborhood diversity nd(G) of a graph G = (V, E)
is the minimum integer t such that V can be partitioned into t sets V1, . . . , Vt where for
every v ∈ V (G) and every i ∈ {1, . . . , t}, either v is adjacent to every vertex in Vi or it is
adjacent to none of them. Note that each part Vi of G is either a clique or an independent
set.
19
Another natural parameter to consider is the vertex cover number, because well-
covered graphs can be equivalently defined as graphs in which every minimal vertex
cover has the same size. However, neighborhood diversity is stronger than vertex cover,
in the sense that every class of graphs with bounded vertex cover number is also a class
of graphs with bounded neighborhood diversity, but the reverse is not true [25]. Thus,
for our analysis, it is enough to consider the neighborhood diversity as the parameter.
In addition, neighborhood diversity is a graph parameter that captures more precisely
than vertex cover number the property that two vertices with the same neighborhood
are "equivalent".
It is worth mentioning that an optimal neighborhood diversity decomposition of a
graph G can be computed in time O(n3); see [25] for more details.
Lemma 12. The Well-Covered Graph problem is FPT when parameterized by
neighborhood diversity.
Proof. Given a graph G, we first obtain a neighborhood partition of G with minimum
width using the polynomial-time algorithm of Lampis [25]. Let t := nd(G) and let
V1, . . . , Vt be the partition of V (G). As we can observe, for any pair u, v of non-adjacent
vertices belonging to the same part Vi, if u is in a maximal independent set S then v
also belongs to S, otherwise S cannot be maximum. On the other hand, if N [u] = N [v]
then for any maximal independent set Su such that u ∈ Su there exists another maximal
independent set Sv such that Sv = Su\{u}∪{v}. Hence, we can contract each partition Vi
that is an independent set into a single vertex vi with weight τ (vi) = Si, and contract
each partition Vi that is a clique into a single vertex vi with weight τ (vi) = 1, in order
to obtain a graph Gt with V (Gt) = t, where the weight of a vertex vi of Gt means that
any maximal independent set of G uses either none or exactly τ (v) vertices of Vi. At
this point, we just need to analyze whether all maximal independent sets of Gt have the
same weight (sum of the weights of its vertices), which can be done in time 2tnO(1).
Corollary 8. The Well-Covered Graph problem is FPT when parameterized by the
vertex cover number n − α.
4.2 Taking the clique-width as the parameter
In the 90's, Courcelle proved that for every graph property Π that can be formulated in
monadic second order logic (MSOL1), there is an f (k)nO(1) algorithm that decides if a
graph G of clique-width at most k satisfies Π (see [4 -- 6,8]), provided that a k-expression
is given.
LinEMSOL is an extension of MSOL1 which allows searching for sets of vertices
which are optimal with respect to some linear evaluation functions. Courcelle et al. [7]
20
showed that every graph problem definable in LinEMSOL is linear-time solvable on
graphs with clique-width at most k (i.e., FPT when parameterized by clique-width) if a
k-expression is given as input. Using a result of Oum [28], the same result follows even
if no k-expression is given.
Theorem 9. The Well-Covered Graph problem is FPT when parameterized by
clique-width.
Proof. Given S ⊆ V (G), first observe that the property "S is a maximal independent
set" is MSOL1-expressible. Indeed, we can construct a formula ϕ(G, S) such that "S is
a maximal independent set" ⇔ ϕ(G, S) as follows:
[(cid:64) u, v ∈ S : edge(u, v)] ∧ [(cid:64) S(cid:48) : (S ⊆ S(cid:48)) ∧ ((cid:64) x, y ∈ S(cid:48) : edge(x, y))]
Since ϕ(G, S) is an MSOL1-expression, the problem of finding goal(S) : ϕ(G, S)
for goal ∈ {max, min} is definable in LinEMSOL. Thus we can find max(S) and
min(S) satisfying ϕ(G, S) in time f (cw(G))nO(1). Finally, G is well-covered if and only
if max(S) = min(S).
Corollary 9. The Well-Covered Graph problem is FPT when parameterized by
treewidth.
Proof. This follows from the fact that graphs with treewidth bounded by k have clique-
width bounded by a function of k [9].
Corollary 10. For any fixed r and (cid:96), the (r,(cid:96))-Well-Covered Graph problem is
FPT when parameterized by clique-width.
Proof. As r and (cid:96) are constants, the problem of determining whether G is an (r, (cid:96))-graph
is also MSOL1-expressible.
Note that, since for every graph G we have cw(G) ≤ nd(G) + 1 [25], Lemma 12 is
also a corollary of Theorem 9. Nevertheless, the algorithm derived from the proof of
Lemma 12 is much simpler and faster than the one that follows from the meta-theorem
of Courcelle et al. [7].
5 Further research
Concerning the complexity of the (r, (cid:96))wc-g and wc-(r, (cid:96))g problems, note that the
only remaining open cases are wc-(r, 0)g for r ≥ 3 (see the tables in Section 1). We do
not even know if there exists some integer r ≥ 3 such that wc-(r, 0)g is coNP-complete,
although we conjecture that this is indeed the case.
21
As another avenue for further research, it would be interesting to provide a com-
plete characterization of well-covered tripartite graphs, as has been done for bipartite
graphs [16, 31, 37]. So far, only partial characterizations exist [21, 38].
Acknowledgement. We would like to thank the anonymous reviewers for helpful remarks that
improved the presentation of the manuscript.
References
[1] A. Brandstädt. Partitions of graphs into one or two independent sets and cliques.
Discrete Mathematics, 152(1-3):47 -- 54, 1996.
[2] A. Brandstädt, F. F. Dragan, H.-O. Le, and R. Mosca. New graph classes of bounded
clique-width. Theory of Computing Systems, 38(5):623 -- 645, 2005.
[3] V. Chvátal and P. J. Slater. A note on well-covered graphs. Annals of Discrete
Mathematics, 55:179 -- 181, 1993.
[4] B. Courcelle. The monadic second-order logic of graphs. I. recognizable sets of finite
graphs. Information and Computation, 85(1):12 -- 75, 1990.
[5] B. Courcelle. The expression of graph properties and graph transformations in
monadic second-order logic. Handbook of Graph Grammars, 1:313 -- 400, 1997.
[6] B. Courcelle and J. Engelfriet. Graph Structure and Monadic Second-Order Logic:
A Language-Theoretic Approach. Cambridge University Press, 1st edition, 2012.
[7] B. Courcelle, J. A. Makowsky, and U. Rotics. Linear time solvable optimiza-
tion problems on graphs of bounded clique-width. Theory of Computing Systems,
33(2):125 -- 150, 2000.
[8] B. Courcelle and M. Mosbah. Monadic second-order evaluations on tree-
decomposable graphs. Theoretical Computer Science, 109(1 -- 2):49 -- 82, 1993.
[9] B. Courcelle and S. Olariu. Upper bounds to the clique width of graphs. Discrete
Applied Mathematics, 101(1 -- 3):77 -- 114, 2000.
[10] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk,
M. Pilipczuk, and S. Saurabh. Parameterized Algorithms. Springer, 2015.
[11] R. de Haan and S. Szeider. Machine Characterizations for Parameterized Com-
plexity Classes Beyond Para-NP. In Proc. of the 41st International Conference on
Current Trends in Theory and Practice of Computer Science (SOFSEM), volume
8939 of LNCS, pages 217 -- 229, 2015.
22
[12] N. Dean and J. S. Zito. Well-covered graphs and extendability. Discrete Mathemat-
ics, 126(1 -- 3):67 -- 80, 1994.
[13] R. Diestel. Graph Theory, 4th Edition, volume 173 of Graduate texts in mathematics.
Springer, 2012.
[14] R. G. Downey and M. R. Fellows. Fundamentals of Parameterized Complexity. Texts
in Computer Science. Springer, 2013.
[15] J. Edmonds. Paths, trees and flowers. Canadian Journal of Mathematics, 17:449 --
467, 1965.
[16] O. Favaron. Very well-covered graphs. Discrete Mathematics, 42:177 -- 187, 1982.
[17] T. Feder, P. Hell, S. Klein, and R. Motwani. List partitions. SIAM Journal on
Discrete Mathematics, 16(3):449 -- 478, 2003.
[18] T. Feder, P. Hell, S. Klein, L. T. Nogueira, and F. Protti. List matrix partitions of
chordal graphs. Theoretical Computer Science, 349(1):52 -- 66, 2005.
[19] J. Flum and M. Grohe. Parameterized Complexity Theory. Springer, 2006.
[20] M. C. Golumbic and U. Rotics. On the clique-width of some perfect graph classes.
International Journal of Foundations of Computer Science, 11(03):423 -- 443, 2000.
[21] H. Haghighi. A generalization of Villarreal's result for unmixed tripartite graphs.
Bulletin of the Iranian Mathematical Society, 40(6):1505 -- 1514, 2014.
[22] M. Kamiński, V. V. Lozin, and M. Milanič. Recent developments on graphs of
bounded clique-width. Discrete Applied Mathematics, 157(12):2747 -- 2761, 2009.
[23] R. M. Karp. Reducibility among combinatorial problems. Complexity of Computer
Computations, pages 85 -- 103, 1972.
[24] S. Kolay, F. Panolan, V. Raman, and S. Saurabh. Parameterized algorithms on
perfect graphs for deletion to (r, (cid:96))-graphs. In Proc. of the 41st International Sym-
posium on Mathematical Foundations of Computer Science (MFCS), volume 58 of
LIPIcs, pages 75:1 -- 75:13, 2016.
[25] M. Lampis. Algorithmic meta-theorems for restrictions of treewidth. Algorithmica,
64(1):19 -- 37, 2012.
[26] M. Lesk, M. D. Plummer, and W. R. Pulleyblank. Equi-matchable graphs. Graph
Theory and Combinatorics. Academic Press, pages 239 -- 254, 1984.
23
[27] R. Niedermeier.
Invitation to Fixed-Parameter Algorithms, volume 31. Oxford
University Press, 2006.
[28] S.-I. Oum. Approximating rank-width and clique-width quickly. ACM Transactions
on Algorithms, 5(1):10:1 -- 10:20, 2008.
[29] M. D. Plummer. Some covering concepts in graphs. Journal of Combinatorial
Theory, 8(1):91 -- 98, 1970.
[30] F. P. Ramsey. On a problem of formal logic. Proceedings of the London Mathematical
Society, s2-30(1):264 -- 286, 1930.
[31] G. Ravindra. Well-covered graphs. Journal of Combinatorics, Information & System
Sciences, 2(1):20 -- 21, 1977.
[32] N. Robertson and P. D. Seymour. Graph minors. X. obstructions to tree-
decomposition. Journal of Combinatorial Theory, Series B, 52(2):153 -- 190, 1991.
[33] R. S. Sankaranarayana and L. K. Stewart. Complexity results for well-covered
graphs. Networks, 22(3):247 -- 262, 1992.
[34] L. Stockmeyer. Planar 3-colorability is polynomial complete. ACM SIGACT News,
5(3):19 -- 25, 1973.
[35] D. Tankus and M. Tarsi. Well-covered claw-free graphs. Journal of Combinatorial
Theory, Series B, 66(2):293 -- 302, 1996.
[36] J. Topp and L. Volkmann. Well covered and well dominated block graphs and
unicyclic graphs. Mathematica Pannonica, 1(2):55 -- 66, 1990.
[37] R. H. Villarreal. Unmixed bipartite graphs. Revista Colombiana de Matemáticas,
41(2):393 -- 395, 2007.
[38] R. Zaare-Nahandi and R. J. Golzari. Unmixed r-partite graphs. Bulletin of the
Iranian Mathematical Society, in press, 2017.
24
|
1606.00001 | 1 | 1606 | 2016-05-30T21:04:19 | Graph isomorphism testing boosted by path coloring | [
"cs.DS"
] | A method for improving the efficiency of graph isomorphism testing is presented. The method uses the structure of the graph colored by vertex hash codes as a means of partitioning vertices into equivalence classes, which in turn reduces the combinatorial burden of isomorphism testing. Unrolling the graph into a tree at each vertex allows structurally different regular graphs to be discriminated, a capability that the color refinement algorithm cannot do. | cs.DS | cs | Graph isomorphism testing boosted by path coloring
Thomas E. Portegys, Dialectek, [email protected]
Abstract
A method for improving the efficiency of graph isomorphism testing is presented. The
method uses the structure of the graph colored by vertex hash codes as a means of
partitioning vertices into equivalence classes, which in turn reduces the combinatorial
burden of isomorphism testing. Unrolling the graph into a tree at each vertex allows
structurally different regular graphs to be discriminated, a capability that the color
refinement algorithm cannot do.
Key words: graph isomorphism, graph hashing, color refinement, vertex partitioning,
equivalence classes.
Introduction
Numerous uses can be found for unique and concise graph identifiers: graphs could then
be counted, sorted, compared and verified more easily. For example, chemical compounds
could be specified by identifying their constituent molecules represented by graphs of
spatial and bonding relationships between atoms. However, a problem with developing a
method for identifying graphs is that graphs are very general objects. Uniquely identifying
vertices and edges solves the problem but begs the question, since the problem then
becomes how to arrive at these identifiers in a uniform fashion (Sayers and Karp, 2004).
A method developed by Portegys (2008) identifies vertices by computing an MD5 hash
(Rivest, 1992) for a tree of nodes rooted at each vertex. A vertex tree is composed by
unrolling reachable vertices. Once each vertex is hashed, the vertex hashes are sorted and
hashed to yield a hash for the graph, a technique similar to that used by Melnik and Dunham
(2001) and Bhat (1980).
The vertex hashing can be seen as a coloring process, along the lines of the well-known
color refinement algorithm (Arvind et al., 2015; Grohe et al., 2014). Given a graph G, the
color refinement algorithm (or naive vertex classification) iteratively computes a sequence
of colorings Ci of V (G). The initial coloring C0 is uniform.
Then,
Ci+1(u) = {{ Ci(a) : a ∈ N(u) }},
(1)
where {{. . .}} is a multiset operator. Note that C1(u) = C1(v) iff the two vertices have the
same degree.
Thus the coloring begins with a uniform coloring of the vertices of the graph and refines it
step by step so that, if two vertices have equal colors but differently colored neighborhoods
(with the multiplicities of colors counted), then these vertices get new different colors in
the next refinement step. The algorithm terminates as soon as no further refinement is
possible.
Graphs are isomorphic if there is a consistent mapping between their vertices (Karp, 1972).
For isomorphism testing of graphs G and H, the color refinement algorithm concludes that
G and H are non-isomorphic if the multisets of colors occurring in these graphs are
different. If this happens, the conclusion is correct. However, not all non-isomorphic
graphs are distinguishable, or amenable, to color refinement. The simplest example is given
by any two non-isomorphic regular graphs of the same degree with the same number of
vertices, such as those shown in Figure 1.
Figure 1 – Regular non-isomorphic graphs.
Graph isomorphism testing, a problem that has long been believed to be of non-polynomial
complexity (NP), has recently been the subject of renewed attention (Babai, 2015). Graph
isomorphism is also of practical use in a number of areas, including mathematical
chemistry and electronic design automation. A number of isomorphism testing algorithms
are in use, e.g. the Ullman (1976) and Schmidt-Druffel (1976) algorithms. The Ullmann
algorithm is one of the most commonly used for graph isomorphism because of its
generality and effectiveness (Cordella, et. al., 2001).
Graph coloring partitions vertices into equivalence classes of identical colors, which can
reduce the complexity of isomorphism testing significantly. While the color refinement
algorithm uses vertex degree as a shallow means of grouping vertices, using deep vertex
hashing as a coloring method produces a finer discrimination of structure such that non-
isomorphic regular graphs can be differentiated. For example, the hashes for the graphs
depicted in Figure 1 will be different.
Description
This section describes the hashing algorithm.
Graph format
Using a pseudo-C++ notation, the following define a graph vertex and edge:
Vertex
{
int label;
Edge edges[];
};
Edge
{
int label;
Vertex source;
Vertex target;
bool directed;
};
This general scheme allows for a number of graph variations: labeled/unlabeled (using null
labels), directed/undirected (for undirected, source and target are synonymous), and
multigraphs.
Algorithm
The following object is used to construct MD5 hash codes based on vertex graph
neighborhoods:
VertexCoder
{
Vertex vertex;
vector<Vertex *> vertexBranch;
unsigned char code[MD5_SIZE];
void encode(bool hashLabels);
void expand();
void contract();
};
The algorithm iteratively expands each vertex in the graph into tree of coder objects
representing the vertices and edges in its neighborhood. Branching terminates when a
duplicate vertex appears in a branch, at which point the terminal coder takes on the hashed
value of the distance of the first appearance of the vertex in the branch. The graph hash
code is then constructed by sorting and hashing the vertex codes.
// Encode graph.
// The boolean argument allows labels to be included in the
// hash calculation.
void encode(bool hashLabels)
{
if (vertex != NULL)
{
expand();
}
int numChildren = children.size();
for (i = 0; i < numChildren; i++)
{
children[i]->coder->encode(hashLabels);
children[i]->coder->contract();
}
sort(children);
input = new unsigned char[HASH_INPUT_SIZE];
if (vertex != NULL)
{
if (hashLabels)
{
append(input, vertex->label);
}
if (numChildren > 0)
{
for (i = 0; i < numChildren; i++)
{
edge = children[i]->edge;
if (hashLabels)
{
append(input, edge->label);
}
if (edge->directed)
{
if (edge->source == vertex)
{
append(input, 1);
}
else
{
append(input, 0);
}
}
else
{
append(input, 2);
}
}
}
else
{
for (i = 0; i < vertexBranch.size(); i++)
{
if (vertex == vertexBranch[i])
{
break;
}
}
i++;
append(input, i);
}
}
for (i = 0; i < numChildren; i++)
{
append(input, children[i]->coder->code);
}
code = MD5hash(input);
}
// Expand coder.
void expand()
{
vector<Vertex *> childVertexBranch;
Vertex *childVertex;
VertexCoder *child;
for (i = 0; i < vertexBranch.size(); i++)
{
if (vertex == vertexBranch[i])
{
return;
}
childVertexBranch.push_back(vertexBranch[i]);
}
childVertexBranch.push_back(vertex);
for (i = 0; i < vertex->edges.size(); i++)
{
if (vertex == vertex->edges[i]->source)
{
childVertex = vertex->edges[i]->target;
}
else
{
childVertex = vertex->edges[i]->source;
}
child = new VertexCoder(childVertex, childVertexBranch);
children.push_back(child);
}
}
// Contract coder.
void contract()
{
for (i = 0; i < children.size(); i++)
{
delete children[i]->coder;
delete children[i];
}
children.clear();
}
Since each of N vertices unrolls a tree of potentially N nodes, the algorithm complexity is
O(N2). A proof of the algorithm appears challenging, but is currently underway. These
initial results are provided with the hope of eliciting further insights.
Example
The method is illustrated through an example. Consider the simple directed graph shown
in Figure 2. The vertices and edges are labeled for illustrative purposes, but the algorithm
works for unlabeled vertices and edges as well as undirected edges.
Figure 2 – A simple directed graph.
Before the first call to encode(), the coder is configured as in Figure 3. This configuration
reveals nothing about the edges in the graph, and thus must always be expanded.
Figure 3 – Initial coder configuration.
After the first expansion, the coder appears as in Figure 4. The (f) and (b) notation on the
edges represent a directed edge in the source graph in the forward and backward direction
respectively. Note that for vertex 0, there are 2 forward edges to vertices 1 and 2. For vertex
1, there is a forward and backward edge to vertex 2, and a backward edge to vertex 0.
Vertex 2 has a forward and backward edge to vertex 1, and a backward edge to vertex 0.
Although Figure 4 shows expanded vertices concurrently, in actuality a vertex is removed
through contraction after its hash values is obtained by its parent. The expansion continues
until each branch reaches a duplicate vertex.
Figure 4 – First expansion.
Figure 5 depicts how the terminal coder values for two branches in the expansion tree are
assigned. On the left branch, the terminal is assigned a value of 1, as it is a duplicate of the
first coder (shown double-bordered), which 1 distant from the root. Likewise the right
terminal is a duplicate of the second coder and is assigned a value of 2. If the terminal coder
has no duplicate, it is assigned the length of the branch as a value.
Figure 5 – Branch terminal coder values.
Results
To highlight the potency of vertex partitioning using a coloring algorithm such as hashing,
a comparison with brute force isomorphism testing is given in Table 1. For each
isomorphism test, a graph with random edge connections and random vertex and edge
labels is generated. Its isomorph is created by adding to each vertex and edge label the
maximum label value of the original graph plus one. The number of search combinations
to test isomorphism was measured. The relatively small graphs rapidly explodes in
complexity for the brute force method, while the hashed method remains remarkably flat.
Vertices x edges
5x5
5x10
10x10
10x20
15x15
Brute force
Hashed
14.1
16.2
5465.9
1593.8
5108975
15.5
26.9
29.6
43.2
44.2
Table 1 – Isomorphism testing comparison.
The results presented in Portegys (2008) validated the ability of the algorithm to uniquely
hash unique graphs. Here we focused on the ability of the hash algorithm to discriminate
regular graphs in comparison to the color refinement algorithm. A graph generation
package (Johnsonbaugh and Kalin, 1991) was used to generate pairs of regular graphs with
varying number of vertices and degrees. Each pair consisted of graphs having the equal
quantities of vertices and equal degree. Some pairs were by chance isomorphic and others
were non-isomorphic. The color refinement algorithm, as expected, classified all the pairs
as isomorphic. The hash algorithm correctly distinguished all isomorphic and non-
isomorphic pairs.
Conclusion
A method for boosting the efficiency of graph isomorphism testing has been presented.
The method is able to discriminate graphs that elude the color refinement algorithm. The
method builds on a previously developed technique for identifying graphs using MD5
hashing.
The C++ code can be found here:
http://sourceforge.net/projects/graph-hashing/
References
V. Arvind, J. Köbler J., G. Rattan, O. Verbitsky (2015). Graph Isomorphism, Color
Refinement, and Compactness. ECCC TR15-032.
L. Babai (2015). Graph Isomorphism in Quasipolynomial Time.
http://arxiv.org/abs/1512.03547
K. V. S. Bhat (1980). Refined vertex codes and vertex partitioning methodology for
graph isomorphism testing. IEEE Trans. Systems Man Cybernet, 10(10) (1980) 610-615.
L. P. Cordella, P. Foggia, C. Sansone, and M. Vento (2001). An Improved
Algorithm for Matching Large Graphs. Proceedings of International
Workshop on Graph-based Representation in Pattern Recognition, Ischia,
Italy, pp. 149 - 159.
M. Grohe, K. Kersting, M. Mladenov, E. Selman (2014). Dimension Reduction via
Colour Refinement. In: Schulz, A.S., Wagner, D. (eds.) ESA 2014. LNCS, vol. 8737, pp.
505-516. Springer, Heidelberg.
R. Johnsonbaugh and M. Kalin (1991). A graph generation software package. Proceedings
of the twenty-second SIGCSE technical symposium on Computer science education. ACM
New York, NY, USA. pp. 151-154.
R. M. Karp (1972). Reducibility among combinatorial problems, in: R. E. Miller and J.
W. Thatcher (Eds.), Complexity of Computer Computations (Plenum, New York) 85-103.
S. Melnik (2001). RDF API draft: Cryptographic digests of RDF models and statements,
http://www-db.stanford.edu/~melnik/rdf/api.html#digest.
T. E. Portegys (2008). General Graph Identification by Hashing.
http://arxiv.org/abs/1512.07263
R. Rivest (1992). RFC 1321 The MD5 Message-Digest Algorithm,
http://tools.ietf.org/html/rfc1321.
C. Sayers and A. H. Karp (2004). RDF Graph Digest Techniques and Potential
Applications. Mobile and Media Systems Laboratory, HP Laboratories Palo Alto, HPL-
2004-95.
D. C. Schmidt and L. E. Druffel (1976). A Fast Backtracking Algorithm to Test Directed
Graphs for Isomorphism Using Distance Matrices. Journal of the Association for
Computing Machinery, 23, pp. 433-445.
J. R. Ullmann (1976). An Algorithm for Subgraph Isomorphism. Journal of the
Association for Computing Machinery, vol. 23, pp. 31-42.
|
1712.07504 | 1 | 1712 | 2017-12-20T14:53:56 | On Counting Perfect Matchings in General Graphs | [
"cs.DS"
] | Counting perfect matchings has played a central role in the theory of counting problems. The permanent, corresponding to bipartite graphs, was shown to be #P-complete to compute exactly by Valiant (1979), and a fully polynomial randomized approximation scheme (FPRAS) was presented by Jerrum, Sinclair, and Vigoda (2004) using a Markov chain Monte Carlo (MCMC) approach. However, it has remained an open question whether there exists an FPRAS for counting perfect matchings in general graphs. In fact, it was unresolved whether the same Markov chain defined by JSV is rapidly mixing in general. In this paper, we show that it is not. We prove torpid mixing for any weighting scheme on hole patterns in the JSV chain. As a first step toward overcoming this obstacle, we introduce a new algorithm for counting matchings based on the Gallai-Edmonds decomposition of a graph, and give an FPRAS for counting matchings in graphs that are sufficiently close to bipartite. In particular, we obtain a fixed-parameter tractable algorithm for counting matchings in general graphs, parameterized by the greatest "order" of a factor-critical subgraph. | cs.DS | cs |
On Counting Perfect Matchings in General Graphs
Daniel Stefankovic∗
Eric Vigoda†
John Wilmes†
December 21, 2017
Abstract
Counting perfect matchings has played a central role in the theory of counting problems. The per-
manent, corresponding to bipartite graphs, was shown to be #P-complete to compute exactly by Valiant
(1979), and a fully polynomial randomized approximation scheme (FPRAS) was presented by Jerrum,
Sinclair, and Vigoda (2004) using a Markov chain Monte Carlo (MCMC) approach. However, it has
remained an open question whether there exists an FPRAS for counting perfect matchings in general
graphs. In fact, it was unresolved whether the same Markov chain defined by JSV is rapidly mixing in
general. In this paper, we show that it is not. We prove torpid mixing for any weighting scheme on
hole patterns in the JSV chain. As a first step toward overcoming this obstacle, we introduce a new
algorithm for counting matchings based on the Gallai–Edmonds decomposition of a graph, and give an
FPRAS for counting matchings in graphs that are sufficiently close to bipartite. In particular, we obtain
a fixed-parameter tractable algorithm for counting matchings in general graphs, parameterized by the
greatest "order" of a factor-critical subgraph.
1 Introduction
Counting perfect matchings is a fundamental problem in the area of counting/sampling problems. For an
undirected graph G = (V, E), let P denote the set of perfect matchings of G. Can we compute (or estimate)
P in time polynomial in n = V ? For which classes of graphs?
A polynomial-time algorithm for the corresponding decision and optimization problems of determining
if a given graph contains a perfect matching or finding a matching of maximum size was presented by
Edmonds [2]. For the counting problem, a classical algorithm of Kasteleyn [9] gives a polynomial-time
algorithm for exactly computing P for planar graphs.
For bipartite graphs, computing P is equivalent to computing the permanent of n × n (0, 1)-matrices.
Valiant [14] proved that the (0, 1)-Permanent is #P-complete. Subsequently attention turned to the Markov
Chain Monte Carlo (MCMC) approach. A Markov chain where the mixing time is polynomial in n is said
to be rapidly mixing, and one where the mixing time is exponential in Ω(n) is referred to as torpidly mixing.
A rapidly mixing chain yields an FPRAS (fully polynomial-time randomized approximation scheme) for the
corresponding counting problem of estimating P [8].
For dense graphs, defined as those with minimum degree > n/2, Jerrum and Sinclair [6] proved rapid
mixing of a Markov chain defined by Broder [1], which yielded an FPRAS for estimating P. The Broder
chain walks on the collection Ω = P∪N of perfect matchings P and near-perfect matchings N ; a near-perfect
matching is a matching with exactly 2 holes or unmatched vertices. Jerrum and Sinclair [6], more generally,
proved rapid mixing when the number of perfect matchings is within a poly(n) factor of the number of
near-perfect matchings, i.e., P/N ≥ 1/poly(n). A simple example, referred to as a "chain of boxes" which
is illustrated in Figure 1, shows that the Broder chain is torpidly mixing. This example was a useful testbed
for catalyzing new approaches to solving the general permanent problem.
Jerrum, Sinclair and Vigoda [7] presented a new Markov chain on Ω = P ∪N with a non-trivial weighting
scheme on the matchings based on the holes (unmatched vertices). They proved rapid mixing for any bipartite
graph with the requisite weights used in the Markov chain, and they presented a polynomial-time algorithm
∗University of Rochester, USA. Email: [email protected]
†Georgia Institute of Technology, USA. Email: {vigoda,wilmesj}@gatech.edu
1
to learn these weights. This yielded an FPRAS for estimating P for all bipartite graphs. That is the current
state of the art (at least for polynomial-time, or even sub-exponential-time algorithms).
Could the JSV-Markov chain be rapid mixing on non-bipartite graphs? Previously there was no example
for which torpid mixing was established, it was simply the case that the proof in [7] fails. We present a
relatively simple example where the JSV-Markov chain fails for the weighting scheme considered in [7]. More
generally, the JSV-chain is torpidly mixing on our class of examples for any weighting scheme based on the
hole patterns, see Theorem 2.2 in Section 2 for a formal statement following the precise definition of the
JSV-chain.
A natural approach for non-bipartite graphs is to consider Markov chains that exploit odd cycles or
blossoms in the manner of Edmonds' algorithm. We observe that a Markov chain which considers all
blossoms for its transitions is intractable since sampling all blossoms is NP-hard, see Theorem 3.1. On the
other hand, a chain restricted to minimum blossoms is not powerful enough to overcome our torpid mixing
examples. See Section 3 for a discussion.
Finally we utilize the Gallai–Edmonds graph decomposition into factor-critical graphs [2, 3, 4, 12] to
present new algorithmic insights that may overcome the obstacles in our classes of counter-examples. In
Section 4, we describe how the Gallai–Edmonds decomposition can be used to efficiently estimate P, the
number of perfect matchings, in graphs whose factor-critical subgraphs have bounded order (Theorem 4.2),
as well as in the torpid mixing example graphs (Theorem 4.3).
Although all graphs are explicitly defined in the text below, figures depicting these graphs are deferred
to the appendix,
1.1 Markov Chains
Consider an ergodic Markov chain with transition matrix P on a finite state space Ω, and let π denote the
unique stationary distribution. We will usually assume the Markov chain is time reversible, i.e., that it
satisfies the detailed balance condition π(x)P (x, y) = π(y)P (x, y) for all states x, y ∈ Ω.
(cid:80)
For a pair of distributions µ and ν on Ω we denote their total variation distance as dTV(µ, ν) =
x∈Ω µ(x) − ν(x). The standard notion of mixing time Tmix is the number of steps from the worst
starting state X0 = i to reach total variation distance ≤ 1/4 of the stationary distribution π, i.e., we write
Tmix = maxi∈Ω min{t : dTV(P t(i,·), π) ≤ 1/4}.
We use conductance to obtain lower bounds on the mixing time. For a set S ⊂ Ω its conductance is
1
2
defined as:
Φ(S) =
x∈S,y /∈S π(x)P (x, y)
.
x∈S π(x)
(cid:80)
(cid:80)
Let Φ∗ = minS⊂Ω:π(S)≤1/2 Φ(S). Then (see, e.g., [13, 10])
Tmix ≥ 1
4Φ∗
.
(1)
1.2 Factor-Critical Graphs
A graph G = (V, E) is factor-critical if for every vertex v ∈ V , the graph induced on V \ {v} has a perfect
matching. (In particular, V is odd.)
Factor-critical graphs are characterized by their "ear" structure. The quotient G/H of a graph G by
a (not necessarily induced) subgraph H is derived from G by deleting all edges in H and contracting all
vertices in H to a single vertex vH (possibly creating loops or multi-edges). An ear of G relative a subgraph
H of G is simply a cycle in G/H containing the vertex vH .
C0 ∪ ··· ∪ Cr such that C0 is a single vertex, and Ci is an odd-length ear in G relative to (cid:83)
Theorem 1.1 (Lov´asz [11]). A graph G is factor-critical if and only if there is a decomposition G =
j<i Cj, for all
0 < i ≤ r.
Furthermore, if G is factor critical, there exists such a decomposition for every choice of vertex C0, and
the order r of the decomposition is independent of all choices.
2
Since the number of ears in the ear decomposition of a factor-critical graph depends only on the graph,
and not on the choice made in the decomposition, we say the order of the factor-critical graph G is the
number r of ears in any ear decomposition of G.
Factor-critical graphs play a central role in the Gallai–Edmonds structure theorem for graphs. We state
an abridged version of the theorem below.
Given a graph G, let D(G) be the set of vertices that remain unmatched in at least one maximum
matching of G. Let A(G) be the set of vertices not in D(G) but adjacent to at least one vertex of D(G).
And let C(G) denote the remaining vertices of G.
Theorem 1.2 (Gallai–Edmonds Structure Theorem). The connected components of D(G) are factor-critical.
Furthermore, every maximum matching of G induces a perfect matching on C(G), a near-perfect matching
on each connected component of D(G), and matches all vertices in A(G) with vertices from distinct connected
components of D(G).
2 The Jerrum–Sinclair–Vigoda Chain
Ω = P ∪(cid:83)
We recall the definition of the original Markov chain proposed by Broder [1]. The state space of the chain is
u,v N (u, v) where P is the collection of perfect matchings and N (u, v) are near-perfect matchings
with holes at u and v (i.e., vertices u and v are the only unmatched vertices). The transition rule for a
matching M ∈ Ω is as follows:
1. If M ∈ P, randomly choose an edge e ∈ M and transition to M \ {e}.
2. If M ∈ N (u, v), randomly choose a vertex x ∈ V . If x ∈ {u, v} and u is adjacent to v, transition
to M ∪ {(u, v)}. Otherwise, let y ∈ V be the vertex matched with x in M , and randomly choose
w ∈ {u, v}. If x is adjacent to w, transition to the matching M ∪ {(x, w)} \ {(x, y)}.
The chain XB is symmetric, so its stationary distribution is uniform. In particular, when P/Ω is at
least inverse-polynomial in n, we can efficiently generate uniform samples from P via rejection sampling,
given access to samples from the stationary distribution of XB.
In order to sample perfect matchings even when Ω/P is exponentially large, Jerrum, Sinclair, and
Vigoda [7] introduce a new chain XJSV that changes the stationary distribution of XB by means of a Metropo-
lis filter. The new stationary distribution is uniform across hole patterns, and then uniform within each hole
pattern, i.e., for every M ∈ Ω, the stationary probability of M is proportional to 1/N (u, v) if M ∈ N (u, v),
and proportional to 1/P if M ∈ P.
We define XJSV in greater detail. For M ∈ Ω, define the weight function
(cid:40) 1P
1
N (u,v)
w(M ) =
if M ∈ P
if M ∈ N (u, v)
(2)
Definition 2.1. The chain XJSV has the same state space as XB. The transition rule for a matching M ∈ Ω
is as follows:
1. First, choose a matching M(cid:48) ∈ Ω to which M may transition, according to the transition rule for XB
2. With probability min{1, w(M(cid:48))/w(M )}, transition to M(cid:48). Otherwise, stay at M .
In their paper, Jerrum, Sinclair, and Vigoda [7] in fact analyze a more general version of the chain XJSV
that allows for arbitrary edge weights in the graph. They show that the chain is rapidly mixing for bipartite
graphs G. (They also study the separate problem of estimating the weight function w, and give a "simulating
annealing" algorithm that allows the weight function w to be estimated by gradually adjusting edge weights
to obtain an arbitrary bipartite graph G from the complete bipartite graph.) Their analysis of the mixing
time uses a canonical paths argument that crucially relies on the bipartite structure of the graph. However,
it remained an open question whether a different analysis of the same chain XJSV, perhaps using different
canonical paths, might generalize to non-bipartite graphs. We rule out this approach.
In fact, we rule out a more general family of Markov chains for sampling perfect matchings. We say
a Markov chain is "of XJSV type" if it has the same state space as XJSV, with transitions as defined in
3
Definition 2.1, for some weight function w(M ) (not necessarily the same as in Eq. (2)) depending only the
hole pattern of the matching M .
Theorem 2.2. There exists a graph G on n vertices such that for any Markov chain X of XJSV type on G,
either the stationary probability of P is at most exp(−Ω(n)), or the mixing time of X is at least exp(Ω(n)).
The graph G of Theorem 2.2 is constructed from several copies of a smaller gadget H, which we now
define.
Definition 2.3. The chain of boxes gadget Bk of length k is the graph on 4k vertices depicted in Figure 1.
To construct Bk, we start with a path P2k−1 = v0, v1, . . . , v2k of length 2k − 1. Then, for every even edge
{v2i, v2i+1} on the path, we add two additional vertices ai, bi, along with edges to form a path v2i, ai, bi, v2i+1
of length 3.
Figure 1: The "chain of boxes" gadget Bk, which has 2k perfect matchings, but only a single matching in
N (v0, v2k+1).
Observation 2.4. The chain of boxes gadget Bk has 2k perfect matchings, but only one matching in
N (v0, v2k+1).
Definition 2.5. The torpid mixing gadget Hk is the graph depicted in Figure 2. To construct H, first
take a C12 and label two antipodal vertices as a and b. Add an edge between a and b, and label the two
vertices farthest from a and b as u and v. Label the neighbor of u closest to a as w1, and the other neighbor
of u as w2. Label the neighbor of v closest to a as z1 and the other neighbor of v as z2. Finally, add four
chain-of-boxes gadgets Bk, identifying the vertices v0 and v2k of the gadgets with w1 and a, with a and z1,
with w2 and b, and with b and z2, respectively.
Note that in Figures 2 and 3, one "box" from each copy of Bk in the torpid mixing gadget is left undrawn,
for visual clarity.
Lemma 2.6. The torpid mixing gadget H = Hk has 16k + 4 vertices and exactly 2 perfect matchings.
Furthermore, NH (u, v) = 1 and NH (x1, v) ≥ 2k.
Proof. A matching M ∈ NH (u, v) is depicted in Figure 2. We argue that M is the only matching in N (u, v).
First note that x1 must be matched with either w1 or a. Either choice forces the matching on the "chain
of boxes" above x1 remain identical to M . But then if x1 is matched with a, there are no vertices to which
w1 can be matched. So x1 must be matched with w1, and the choice of edge for x2, y1, and y2 is forced
symmetrically, giving the matching M .
Similarly, there are exactly two perfect matchings of H. Vertex u is matched with either w1 or w2, and
either choice determines all other edges. In particular, if u is matched with w1, then x1 must be matched
with a, and y1 with z1, and so on along the entire 12-cycle containing u and v. The edges on the four "chains
of boxes" are then also completely determined. The other case, when u is matched with w2, is symmetric.
We now argue that NH (x1, v) ≥ 2k. Starting from the matching M(cid:48) ∈ NH (x1, v) depicted in Figure 3,
each of the k copies of C4 in the chain of boxes above x1 can be independently alternated, giving 2k distinct
matchings in NH (x1, v).
The torpid mixing gadget already suffices on its own to show that the Markov chain XXJSV defined in [7]
is torpidly mixing. In particular, the conductance out of the set NH (x1, v) ⊆ Ω(H) is 2−Ω(k). In order to
prove the stronger claim of Theorem 2.2, that every Markov chain of XJSV-type fails to efficiently sample
perfect matchings, we construct a slightly larger graph from copies of the torpid mixing gadgets.
4
v0a0b0v1v2a1b1v3...v2kakbkv2k+1 Figure 2: The torpid mixing gadget Hk. The unique matching M ∈ N (u, v) is depicted with thick edges.
Definition 2.7. The counterexample graph Gk is the graph depicted in Figure 4.
It is defined by
replacing every third edge of the twelve-cycle C12 with the gadget Hk defined in Figure 2. Specifically, let
{ui, vi} be the 3i-th edge of C12 for i ∈ {1, . . . , 4}. We delete each edge {ui, vi} and replace it with a copy
of H, identifying the vertices u and v of H with vertices ui and vi of C12. The resulting graph is Gk. Thus,
of the 12 original vertices in C12, 8 of the corresponding vertices in Gk participate in a copy of the gadget
H, and 4 do not. These 4 vertices of Gk which do not participate in any copy of the gadget H are labeled
t1, . . . , t4 in cyclic order, and the copies of the gadget H are labeled H1, . . . H4 in cyclic order, with H1
coming between t1 and t2, and so on. Thus, t1 is adjacent to u1 and v4, ti is adjacent to ui and vi−1 for
i ∈ {2, . . . , 4}, and Hi contains both ui and vi.
In particular, Gk has 4V (H) + 4 = 64k + 8 vertices.
The perfect and near-perfect matchings of Gk are naturally divided into four intersecting families. For
i ∈ {1, . . . , 4} we define Si to be the collection of (perfect and near-perfect) matchings M ∈ Ω(Gk) such that
the restriction of M to Hi has two holes, at ui and vi, i.e., such that the vertices ui and vi either have holes
in M or are matched outside of Hi.
Lemma 2.8. The counterexample graph Gk has exactly 8 perfect matchings. Of these, 4 are in S1 ∩ S3 \
(S2 ∪ S4) and 4 are in S2 ∩ S4 \ (S1 ∪ S3).
Proof. The graph Gk has exactly 8 perfect matchings. To obtain a matching in S1 ∩ S3 \ (S2 ∪ S4), we
may without loss of generality start by matching the vertices in H1 and H3 according to a matching in
NH1 (u1, v1) or NH3(u3, v3), respectively. We must then match t1 with u1, t2 with v1, t3 with u3, and t4
with v3. Finally, we must match the remaining vertices according to a perfect matching on each of H2 and
H4. By Lemma 2.6, there are two perfect matchings on each of H2 and H4, and a unique matching in each
of NH1 (u1, v1) and NH3 (u3, v3), so indeed there are four matchings in (S1 ∩ S3) \ (S2 ∪ S4). Similarly, there
are exactly four matchings in (S2 ∩ S4) \ (S1 ∪ S3).
To see that there are no other perfect matchings, let M be an arbitrary perfect matching of Gk. Then
t1 is matched either with u1 or v4. Suppose t1 is matched with u1. Then v4 is matched within H4. Since
H4 has an even number of vertices, u4 must also be matched within H4, and hence M induces a perfect
matching on H4. Continuing in a similar fashion, M must also induce a perfect matching on H2. Then the
restriction of M to H1 or H3 has holes at u1 and v1, and at u3 and v3, respectively, so M ∈ S1∩S3\(S2∪S4).
Symmetrically, if t1 is matched with v4 then M ∈ S2 ∩ S4 \ (S1 ∪ S3).
In the proof below, we use the notation N (M ) denote the collection of matchings with the hole pattern
as M . That is, N (M ) = P if M ∈ P, and N (M ) = N (u, v) if M ∈ N (u, v).
5
uw1x1ax2w2y1z1vz2y2............ Figure 3: A matching M(cid:48) ∈ N (x1, v). There are exponentially many matchings with the same hole pattern,
obtained by alternating the 4-cycles above x1.
Figure 4: The "counterexample graph" Gk on which XJSV is torpidly mixing. The boxes labeled Hk represent
copies of the torpid mixing gadget of Definition 2.5.
Proof of Theorem 2.2. Let Gk be the counterexample graph of Definition 2.7. We will show that the set
S1 ∪ S3 ⊆ Ω(Gk) has poor conductance, unless the stationary probability of PGk is small. We will write
A = S1 ∪ S3 and A = Ω(Gk) \ (S1 ∪ S3).
Let M ∈ A and M(cid:48) ∈ A be such that P (M, M(cid:48)) > 0. We claim that neither M nor M(cid:48) are perfect
matchings. Assume without loss of generality that M ∈ S1. If M ∈ S1 is a perfect matching, then M ∈ P2
and so M ∈ S3. The only legal transitions from M to Ω\ S1 are those that introduce additional holes within
H1, but none of these transitions to a matching outside of S3. Hence, M cannot be perfect. But if M(cid:48) is
perfect, then M(cid:48) ∈ P1, and so M(cid:48) induces a perfect matching on S1. But then the transition from M to M(cid:48)
must simultaneously affect u1 and v1, and no such transition exists.
We denote by ∂A the set of matchings M(cid:48) ∈ A such that there exists a matching M ∈ A with P (M, M(cid:48)) >
0. We claim that for every matching M(cid:48) ∈ A, we have
N (M
(cid:48)) ∩ ∂A ≤ 2k−1N (M
(3)
Let M(cid:48) ∈ ∂A, and let M ∈ A be such that P (M, M(cid:48)) > 0. Suppose first that M ∈ S1. Label the vertices
of H1 as in Figure 2, identifying u1 with u and v1 with v. Let N be the matching on H = H1 induced
by M , and let N(cid:48) be the matching on H1 induced by M(cid:48). We have N ∈ NH (u1, v1). But by Lemma 2.6,
we have NH (u1, v1) = 1, i.e., N is exactly the matching depicted in Figure 2. The only transitions that
(cid:48)) .
6
x1v............t1u1v1t2u2v2t3u3v3t4u3v3HkHkHkHk remove the hole at u are the two that shift the hole to x1 or x2, and the only transitions that remove the
hole at v are the two that shift the hole to y1 or y2. So, without loss of generality, by the symmetry of
Gk, we have N(cid:48) ∈ NH (x1, v1). By Lemma 2.6, NH (x1, v1) ≥ 2k, but only one matching in NH (x1, v1) has
a legal transition to N . Therefore, if we replace the restriction of M(cid:48) to H1 with any other matching in
NH (x1, v1), we obtain another matching M(cid:48)(cid:48) ∈ N (M(cid:48)), but M(cid:48)(cid:48) has no legal transition to any matching in
N (M ). Hence, only a 2−k-fraction of N (M(cid:48)) has a legal transition to S1, and similarly only a 2−k-fraction
of N (M(cid:48)) has a legal transition to S3. In particular, we have proved Eq. (3).
From Eq. (3), it immediately follows that the stationary probability of ∂A is
(cid:88)
(cid:88)
π(∂A) =
(cid:48)) =
π(M
M(cid:48)∈∂A
M(cid:48)∈A
(cid:48))
π(M
N (M(cid:48)) ∩ ∂A
N (M(cid:48))
= 2−k+1π(A)
(4)
We now compute (cid:88)
M∈A,M
(cid:48)
P (M,M
(cid:48)∈A
)>0
π(M )P (M, M
(cid:48)) =
(cid:88)
M∈A,M
(cid:48)
P (M,M
(cid:48)∈A
)>0
(cid:48))P (M
(cid:48)
, M ) ≤ π(∂(A))
π(M
< 2−k+1π(A),
where we first use the detailed balance condition and then Eq. (4).
Now by (1) and the definition of conductance, we have
1
4τX
< Φ(A) < 2−k π(A)
π(A)
.
In particular, if τX < 2k/2−2, then π(A) > 2k/2+1π(A). Suppose this is the case. By Lemma 2.8, half
In particular, π(PGk ) ≤ 2π(A) < 2−k/2+2. Hence, either
of the perfect matchings of Gk belong to A.
the stationary probability of P is at most 2−k/2+2 = exp(−Ω(n)), or the mixing time of X is at least
2k/2−2 = exp(Ω(n)).
We remark that the earlier Markov chain studied by Broder [1] and Jerrum and Sinclair [6] is also torpidly
mixing on the counterexample graph of Definition 2.7, since the ratio of near-perfect matchings to perfect
matchings is exponential [6].
3 Chains Based on Edmonds' Algorithm
Given that Edmonds' classical algorithm for finding a perfect matching in a bipartite graph requires the
careful consideration of odd cycles in the graph, it is reasonable to ask whether a Markov chain for counting
perfect matchings should also somehow track odd cycles.
In this section, we briefly outline some of the
difficulties of such an approach.
A blossom of length k in a graph G equipped with a matching M is simply an odd cycle of length
2k + 1 in which k of the edges belong to M . Edmonds' algorithm finds augmenting paths in a graph by
exploring the alternating tree rooted at an unmatched vertex, and contracting blossoms to a vertex as they
are encountered. Given a blossom B containing an unmatched vertex u, there is an alternating path of even
length to every vertex v ∈ B. Rotating B to v means shifting the hole at u to v by alternating the u-v path
in B.
Adding rotation moves to a Markov chain in the style of XJSV is an attractive possible solution to the
obstacles presented in the previous section. Indeed, if it were possible to rotate the 7-cycle containing u and
a in the graph in Figure 2, it might be possible to completely avoid problematic holes at x1 or x2.
The difficulty in introducing such an additional move the Markov chain XJSV is in defining the set of
feasible blossoms that may be rotated, along with a probability distribution over such blossoms. In order to
be useful, we must be able to efficiently sample from the feasible blossoms at a given near-perfect matching
M . Furthermore, the feasible blossoms must respect time reversibility: if B is feasible when the hole is at
u ∈ B, then it must also be feasible after rotating the hole to v ∈ B; reversibility of the Markov chain is
7
Figure 5: After rotating the blossom so that the hole is moved from u to v, the blossom is no longer
"minimal".
needed so that we understand its stationary distribution. Finally, the feasible blossoms must be rich enough
to avoid the obstacles outlined in the previous section.
The set of "minimum length" blossoms at a given hole vertex u satisfies the first criterion of having an
efficient sampling algorithm. But it is easy to see that if only minimum length blossoms are feasible, then
the obstacles outlined in the previous section will still apply (simply by adding a 3-cycle at every vertex).
Moreover, families blossoms characterized by minimality may struggle to satisfy the second criterion of time-
reversibility. In Figure 5, there is a unique blossom containing u, but after rotating the hole to v, it is no
longer minimal.
On the other hand, the necessity of having an efficient sampling algorithm for the feasible blossoms
already rules out the simplest possibility, namely, the uniform distribution over all blossoms containing a
given hole vertex u. Indeed, if we could efficiently sample from the uniform distribution over all blossoms
containing a given vertex u, then by an entropy argument we could find arbitrarily large odd cycles in the
graph, which is NP-hard.
Theorem 3.1. Let Sampling Blossoms problem be defined as follows. The input is an undirected graph G
and a near-perfect matching M with holes at w, r ∈ V (G). The output is a uniform sample from the uniform
distribution of blossoms containing w. Unless NP=RP there is no randomized polynomial-time sampler for
Sampling Blossoms.
Proof. We reduce from the problem of finding the longest s-t-path in a directed graph H (ND29 in [5]). We
construct an instance of Sampling Blossoms, that is, G and M as follows. For every v ∈ V (H) we add two
vertices v0, v1 into V (G) and also add {v0, v1} into M . For every edge (u, v) ∈ E(H) we add edge {u1, v0}
into E(G). Finally we add w, r into V (H) and {w, s0},{t1, w} into E(H).
Note that there is one-to-one correspondence between blossoms that contain w in G and s-t-paths in H.
We now modify G to "encourage longer paths". We replace each {v0, v1} edge in G by a chain of boxes (with
(cid:96) boxes) and replace {v0, v1} in M by the unique perfect matching of the chain of boxes. In the modified
graph G for every s-t-path p in H there are now 2k(cid:96) blossoms that contain w in G, where k is the number
of vertices in p.
Taking (cid:96) = n2 a uniformly random blossom that contains w in G will with probability 1− o(1) correspond
to a longest s-t-path in H (the number of s-t-paths is bounded by (n+1)n = 2O(n log n) and hence the fraction
of blossoms corresponding to non-longest s-t-paths is 2O(n log n)2−n2
= o(1)).
4 A Recursive Algorithm
We now explore a new recursive algorithm for counting matchings, based on the Gallai–Edmonds decompo-
sition. In the worst case, this algorithm may still require exponential time. However, for graphs that have
additional structural properties, for example, those that are "sufficiently close to bipartite" in a sense that
will be made precise, our recursive algorithm runs in polynomial time. In particular, it will run efficiently
on examples similar to those used to prove torpid mixing of Markov chains in the previous section.
We now state the algorithm.
It requires as a subroutine an algorithm for computing the permanent
of the bipartite adjacency matrix of a bipartite graph G up to accuracy ε. We denote this subroutine by
Permanent(G, ε). The Permanent subroutine requires time polynomial in V (G) and 1/ε using the
algorithm of Jerrum, Sinclair, and Vigoda [7], but we use it as a black-box.
We first argue the correctness of the algorithm.
8
uv Algorithm 1 Recursive algorithm for approximately counting the number of perfect matchings in a graph
1: procedure Recursive-Count(G, ε)
2:
If V (G) = ∅, return 1.
Choose u ∈ V (G).
Compute the Gallai–Edmonds decomposition of G − u.
for all v ∈ D(G − u) do
Hv ← the connected component of G − u containing v
mv ← Recursive-Count(Hv − v, ε/(2n))
end for
mC ← Recursive-Count(C(G − u), ε/3)
Let X = A(G − u) ∪ {u}, and let Y be the set of connected components in D(G − u). Let G(cid:48) be the
bipartite graph on (X, Y ) defined as follows: for every x ∈ X and H ∈ Y , if x has any neighbors in H
in G(cid:48), add an edge {x, H} in G(cid:48) with weight
3:
4:
5:
6:
7:
8:
9:
10:
w(x, H) =
return mC ∗ Permanent(G(cid:48), ε/3)
11:
12: end procedure
(cid:88)
mv .
v∈N (x)∩H
Theorem 4.1. Algorithm 1 computes the number of perfect matchings in G to within accuracy ε.
Proof. We show that the algorithm is correct for graphs on n vertices, assuming it is correct for all graphs
on at most n − 1 vertices.
We claim that permanent of the incidence matrix of G(cid:48) defined on line 10 equals the number of perfect
matchings in G. Indeed, every perfect matching M of G induces a maximum matching Mu on G − u. By
the Gallai–Edmonds theorem, Mu matches each element of A(G(cid:48)) with a vertex from a distinct component
of D(G(cid:48)), leaving one component of D(G(cid:48)) unmatched. Vertex u must therefore be matched in M with a
vertex from the remaining component of D(G(cid:48)). Therefore, M induces a perfect matching M(cid:48) on G(cid:48). Now
let Hx ∈ Y be the vertex of G(cid:48) matched to x for each x ∈ X. Then the number of distinct matchings of G
inducing the same matching M(cid:48)(cid:48) on G(cid:48)(cid:48) is exactly
(cid:89)
with Hx defined as above, we obtain(cid:81)
x∈X
(cid:88)
v∈N (x)∩Hx
(cid:89)
x∈X
mv =
w(x, Hx)
which is the contribution of M(cid:48) to the permanent of G(cid:48). Similarly, from an arbitrary matching M(cid:48) of G(cid:48),
Hence, it suffices to to compute the permanent of the incidence matrix of G(cid:48) up to accuracy ε. We know
the entries of the incidence matrix up to accuracy ε/(2n), and (1 + ε/(2n))n/2 < 1 + ε/3 for ε sufficiently
small. Therefore, it suffices to compute the permanent of our approximation of the incidence matrix up to
accuracy ε/3 to get overall accuracy better than ε.
x∈X w(x, Hx) matchings of G, proving the claim.
The running time of Algorithm 1 is sensitive to the choice of vertex u on line 3. If u can be chosen so
that each component of D(G − u) is small, then the algorithm is an efficient divide-and-conquer strategy.
More generally, if u can be chosen so that each component of D(G− u) is in some sense "tractable", then an
efficient divide-and conquer strategy results. In particular, since it is possible to exactly count the number
of perfect matchings in a factor-critical graph of bounded order in polynomial time, we obtain an efficient
algorithm for approximately counting matchings in graphs whose factor-critical subgraphs have bounded
order. This is the sense in which Algorithm 1 is efficient for graphs "sufficiently close" to bipartite.
Theorem 4.2. Suppose every factor-critical subgraph of G has order at most k. Then the number of perfect
matchings in G can be counted to within accuracy ε in time 2O(k)poly(n, 1/ε).
The essential idea of the proof is to first observe that a factor-critical graph can be shrunk to a graph
with O(k) edges having the same number of perfect matchings after deleting any vertex. The number of
9
perfect matchings can then be counted by brute force in time 2O(k)poly(n). This procedure replaces the
recursive calls on line 6 of the algorithm.
Proof. We first observe that if H is a factor-critical graph of order k with n vertices, then the number of
perfect matchings in H − v can be counted exactly in time 2O(k)poly(n) for every vertex v. Writing du for
the degree of a vertex u, we have
(du − 2) = 2(k − 1),
(5)
(cid:88)
u∈V (H)
since adding one ear to a graph adds some number of vertices of degree 2, and increases the degree of two
existing vertices by one each, or one vertex by two. Fix v ∈ H, and suppose there is a vertex u of degree 2
in H − v, with neighbors w1 and w2. Let H(cid:48) denote the multigraph obtained from H − v by contracting the
edges from u to w1 and w2, so H(cid:48) has two fewer vertices than H, and has a vertex w with the same multiset
of neighbors as w1 and w2 (excluding v). Then there is a bijection between the perfect matchings of H(cid:48) and
of H − v; each perfect matching of H(cid:48) lifts to a matching of H − v with a hole at u and exactly one of w1
or w2, and each perfect matching of H − v projects to a perfect matching of H(cid:48) by ignoring the matched
edge at u. Hence, we may contract away all degree-2 vertices of H − v, and obtain a graph with the same
number of perfect matchings in which every vertex (save at most two of degree 1, the former neighbors of
v) has degree at least 3. Then since the contraction does not change the sum in Eq. (5), we have
(cid:48)) − 2) ≤ (cid:88)
3(V (H
du ≤ 2(k − 1) + 2V (H
(cid:48))
u∈V (H(cid:48))
and hence H(cid:48) has O(k) edges, and the perfect matchings of H(cid:48) can be enumerated in time 2O(k).
Now we modify Algorithm 1 to run in time 2O(k)poly(n, 1/ε). First, we delete all edges not appearing in
any perfect matching, call Recursive-Count(Gi, ε/(2n)) on each connected component Gi, and multiply
the results of all of these calls to estimate the number of perfect matchings in G. We have C(Gi − u) = ∅
for each such component Gi and every vertex u ∈ V (Gi), since edges leaving C(Gi − u) cannot appear
in any matching of Gi − u. Therefore, the recursive call on line 8 of the algorithm can be eliminated.
On line 6, instead of computing mv by a recursive call, we instead use the procedure described above to
compute it in time 2O(k). Hence, Algorithm 1 requires O(n) calls to a procedure that takes time 2O(k). The
other lines of Algorithm 1 require only polynomial time in n and 1/ε, so in all Algorithm 1 requires time
2O(k)poly(n, 1/ε).
We note that Theorem 4.2 is proved by eliminating recursive calls in the algorithm. Although the
recursive calls of Algorithm 1 can be difficult to analyze, they can also be useful, as we now demonstrate by
showing that Algorithm 1 runs as-is in polynomial time on the counterexample graph of Definition 2.7, for
appropriate choice of the vertex u on the line 3 of the algorithm.
Theorem 4.3. Algorithm 1 runs in polynomial time on the counterexample graph of Definition 2.7, for
appropriate choice of the vertex u on the line 3 of the algorithm
Proof. After deleting the vertices u and v from the torpid mixing gadget H in Figure 2, no odd cycles remain
the graph H. Let U denote the set of all four copies of the vertices u and v appearing in the counterexample
graph G, so U = 8. With every recursive call Recursive-Count(G(cid:48), ε(cid:48)), if U ∩ V (G(cid:48)) (cid:54)= ∅, we choose
u ∈ U ∩V (G(cid:48)). Hence, after 8 recursive calls, there are no odd cycles remaining in G(cid:48), and each factor-critical
subgraph is a single vertex. When U ∩ V (G(cid:48)) = ∅, we choose u so that A(G(cid:48) − u) = Ω(n)-for example
taking u at one end of a chain of boxes-so that the overall recursive depth is O(1).
Acknowledgements
This research was supported in part by NSF grants CCF-1617306, CCF-1563838, CCF-1318374, and CCF-
1717349. The authors are grateful to Santosh Vempala for many illuminating conversations about Markov
chains and the structure of factor-critical graphs.
10
References
[1] A. Z. Broder. How hard is it to marry at random? (On the approximation of the permanent), Proceed-
ings of the 18th Annual ACM Symposium on Theory of Computing (STOC), 50–58, 1986. Erratum in
Proceedings of the 20th Annual ACM Symposium on Theory of Computing, p. 551, 1988.
[2] J. Edmonds. Paths, trees, and flowers. Canadian Journal of Mathematics, 17:449–467, 1965.
[3] T. Gallai. Kritische Graphen II. Magyar Tud. Akad. Mat. Kutat´o Int. Kozl., 8:273–395, 1963.
[4] T. Gallai. Maximale systeme unabhangiger kanten. Magyar Tud. Akad. Mat. Kutat´o Int. Kozl., 9:401–
413, 1964.
[5] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of NP-
Completeness. W. H. Freeman & Co., New York, NY, 1979.
[6] M. Jerrum and A. Sinclair. Approximating the permanent. SIAM Journal on Computing, 18(6):1149–
1178, 1989.
[7] M. Jerrum, A. Sinclair, and E. Vigoda. A polynomial-time approximation algorithm for the permanent
of a matrix with non-negative entries. Journal of the ACM, 51(4):671–697, 2004.
[8] M. R. Jerrum, L. G. Valiant, and V. V. Vazirani. Random generation of combinatorial structures from
a uniform distribution. Theoret. Comput. Sci., 43(2-3):169–188, 1986.
[9] P. W. Kasteleyn. Graph theory and crystal physics. In Graph Theory and Theoretical Physics, pages
43–110. Academic Press, London, 1967.
[10] D. A. Levin, Y. Peres, and E. L. Wilmer. Markov Chains and Mixing Times. American Mathematical
Society, Providence, RI, 2009.
[11] L. Lov´asz. A note on factor-critical graphs. Studia Sci. Math. Hungar, 7(11):279–280, 1972.
[12] A. Schrijver. Theory of Linear and Integer Programming. Wiley, 1998.
[13] A. J. Sinclair. Algorithms for Random Generation and Counting: A Markov Chain Approach,
Birkhauser, 1988.
[14] L. G. Valiant. The complexity of computing the permanent. Theoretical Computer Science, 8(2):189–201,
1979.
11
|
1612.06962 | 2 | 1612 | 2017-10-11T03:24:57 | Stochastic Runtime Analysis of a Cross Entropy Algorithm for Traveling Salesman Problems | [
"cs.DS",
"cs.AI",
"cs.NE"
] | This article analyzes the stochastic runtime of a Cross-Entropy Algorithm on two classes of traveling salesman problems. The algorithm shares main features of the famous Max-Min Ant System with iteration-best reinforcement.
For simple instances that have a $\{1,n\}$-valued distance function and a unique optimal solution, we prove a stochastic runtime of $O(n^{6+\epsilon})$ with the vertex-based random solution generation, and a stochastic runtime of $O(n^{3+\epsilon}\ln n)$ with the edge-based random solution generation for an arbitrary $\epsilon\in (0,1)$. These runtimes are very close to the known expected runtime for variants of Max-Min Ant System with best-so-far reinforcement. They are obtained for the stronger notion of stochastic runtime, which means that an optimal solution is obtained in that time with an overwhelming probability, i.e., a probability tending exponentially fast to one with growing problem size.
We also inspect more complex instances with $n$ vertices positioned on an $m\times m$ grid. When the $n$ vertices span a convex polygon, we obtain a stochastic runtime of $O(n^{3}m^{5+\epsilon})$ with the vertex-based random solution generation, and a stochastic runtime of $O(n^{2}m^{5+\epsilon})$ for the edge-based random solution generation. When there are $k = O(1)$ many vertices inside a convex polygon spanned by the other $n-k$ vertices, we obtain a stochastic runtime of $O(n^{4}m^{5+\epsilon}+n^{6k-1}m^{\epsilon})$ with the vertex-based random solution generation, and a stochastic runtime of $O(n^{3}m^{5+\epsilon}+n^{3k}m^{\epsilon})$ with the edge-based random solution generation. These runtimes are better than the expected runtime for the so-called $(\mu\!+\!\lambda)$ EA reported in a recent article, and again obtained for the stronger notion of stochastic runtime. | cs.DS | cs |
Stochastic Runtime Analysis of a Cross-Entropy
Algorithm for Traveling Salesman Problems
Zijun Wua,1,∗, Rolf H. Mohringb,2,∗∗, Jianhui Laic,
aBeijing Institute for Scientific and Engineering Computing (BISEC), Pingle Yuan 100,
bBeijing Institute for Scientific and Engineering Computing (BISEC), Pingle Yuan 100,
cCollege of Metropolitan Transportation, Beijing University of Technology, Pingle Yuan
Beijing, P. R. China
Beijing, P. R. China
100, Beijing, P. R. China
Abstract
This article analyzes the stochastic runtime of a Cross-Entropy Algorithm
mimicking an Max-Min Ant System with iteration-best reinforcement.
It
investigates the impact of magnitude of the sample size on the runtime to
find optimal solutions for TSP instances.
For simple TSP instances that have a {1, n}-valued distance function and
a unique optimal solution, we show that sample size N ∈ ω(ln n) results in a
stochastically polynomial runtime, and N ∈ O(ln n) results in a stochastically
exponential runtime, where "stochastically" means with a probability of 1 −
In particular, for N ∈ ω(ln n),
n−ω(1), and n represents number of cities.
we prove a stochastic runtime of O(N · n6) with the vertex-based random
solution generation, and a stochastic runtime of O(N · n3 ln n) with the edge-
based random solution generation. These runtimes are very close to the best
known expected runtime for variants of Max-Min Ant System with best-so-
far reinforcement by choosing a small N ∈ ω(ln n). They are obtained for
∗Principal corresponding author
∗∗Corresponding author
Corresponding author
Email addresses: [email protected] (Zijun Wu), [email protected]
(Rolf H. Mohring), [email protected] (Jianhui Lai)
1The author is also affiliated with the School of Applied Mathematics and Physics at
Beijing University of Technology
2The author is a professor emeritus of mathematics at Berlin University of Technology
Preprint submitted to Theoretical Computer Science
October 12, 2017
the stronger notion of stochastic runtime, and analyze the runtime in most
cases.
We also inspect more complex instances with n vertices positioned on
an m × m grid. When the n vertices span a convex polygon, we obtain
a stochastic runtime of O(n4m5+) with the vertex-based random solution
generation, and a stochastic runtime of O(n3m5+) for the edge-based random
solution generation. When there are k ∈ O(1) many vertices inside a convex
polygon spanned by the other n − k vertices, we obtain a stochastic runtime
of O(n4m5+ + n6k−1m) with the vertex-based random solution generation,
and a stochastic runtime of O(n3m5+ + n3km) with the edge-based random
solution generation. These runtimes are better than the expected runtime
for the so-called (µ+λ) EA reported in a recent article, and again obtained
for the stronger notion of stochastic runtime.
Keywords: probabilistic analysis of algorithms, stochastic runtime analysis
of evolutionary algorithms, Cross Entropy algorithm, Max-Min Ant
System, (µ+λ) EA.
1. Introduction
The Cross Entropy (CE) algorithm is a general-purpose evolutionary al-
gorithm (EA) that has been applied successfully to many NP-hard combi-
natorial optimization problems, see e.g.
It
was initially designed for rare event simulation by Rubinstein [2] in 1997,
and thereafter formulated as an optimization tool for both continuous and
discrete optimization (see [3]).
the book [1] for an overview.
CE has much in common with the famous ant colony optimization (ACO,
see [4]) and the estimation of distribution algorithms (EDAs, see [5]). They
all belong to the so-called model-based search paradigm (MBS), see [6]. In-
stead of only manipulating solutions, which is very typical in traditional
heuristics like Genetic Algorithms [7] and Local Search [8] and others, MBS
algorithms attempt to optimize the solution reproducing mechanism. In each
iteration, they produce new solutions by sampling from a probabilistic dis-
tribution on the search space. The distribution is often called a model in the
literature (see e.g.
[6] and [9]). This model evolves iteratively by incorpo-
rating information from some elite solutions occurring in the search history,
so as to asymptotically model the spread of optimal solutions in the search
2
space. See the recent Thesis [9] for more details on MBS algorithms and
their mathematical properties.
An important issue for MBS algorithms is to determine a suitable size
for the sampling in each iteration. A large sample size makes each iteration
unwieldy, however a small sample size may mislead the underlying search
due to the randomness in the sampling. Sample size reflects the iterative
complexity (computational complexity in each iteration). Whether a large
sample size is harmful depends on the required optimization time (i.e., the
total number of iterations required to reach an optimal solution). This article
aims to shed a light on this issue by theoretically analyzing the relation
between sample size and optimization time for a CE variant that includes
also some essential features of the famous Max-Min Ant System (MMAS
[10]). To this end, a thorough runtime analysis is needed.
The theoretical runtime analysis of EAs has gained rapidly growing in-
terest in recent years, see e.g.
[11], [12], [13], [14], [15], [16], [17], [18], [19],
[20], [21], [22], [23], [24], and [25]. In the analysis, an oracle-based view of
computation is adopted, i.e., the runtime of an algorithm is expressed as
the total number of solutions evaluated before reaching an optimal solution.
Since the presence of randomness, the runtime of an EA is often conveyed
in expectation or with high probability. Due to the famous No Free Lunch
Theorem [26], the analysis must be problem-specific. The first steps towards
this type of analysis were made for the so-called (1+1) EA [11] on some test
problems that use pseudo boolean functions as cost functions, e.g., OneMax
[15], LeadingOnes [27] and BinVar [11]. Recent research addresses prob-
lems of practical importance, such as the computing a minimum spanning
trees (MST) [28], matroid optimization [29], traveling salesman problem [30],
the shortest path problem [23], the maximum satisfiability problem [31] and
the max-cut problem [32].
Runtime analysis generally considers two cases: expected runtime analysis
and stochastic runtime analysis. Expected runtime is the average runtime of
an algorithm on a particular problem, see, e.g., the runtime results of (1 + 1)
EA reported in [11]. Expected runtime reflects the oracle-based average per-
formance of an algorithm. A mature technique for expected runtime analysis
is the so-called drift analysis [12]. However, this technique requires that the
algorithm has a finite expected runtime for the underlying problem. By [33],
drift analysis is not applicable to the traditional CE [3].
An algorithm with a smaller expected runtime need not be more efficient,
see [34] for details. In contrast, stochastic runtime provides a better under-
3
standing of the performance of a (randomized) EA. Stochastic runtime is a
runtime result conveyed with an overwhelming probability guarantee (see,
e.g., the classic runtime result of 1-ANT in [15]), where an overwhelming
probability means a probability tending to 1 superpolynomially fast in the
problem size. It therefore reflects the efficiency of an algorithm for most cases
in the sense of uncertainty. This article is concerned with stochastic runtime
analysis, aiming to figure out the relation between stochastic runtime and
magnitude of the sample size.
Runtime analysis of CE algorithms has been initiated in [33], where Wu
and Kolonko proved a pioneering stochastic runtime result for the traditional
CE on the standard test problem LeadingOnes. As a continuation of the
study of [33], Wu et al [34] further investigated the stochastic runtime of
the traditional CE on another test problem OneMax. The runtime results
reported in [33] and [34] showed that sample size plays a crucial role in
efficiently finding an optimal solution. In particular, Wu et al [34] showed
that if the problem size n is moderately adapted to the sample size N , then
the stochastic runtime of the traditional CE on OneMax is O(n1.5+ 4
3 ) for
arbitrarily small > 0 and a constant smoothing parameter ρ > 0, which
beats the best-known stochastic runtime O(n2) reported in [13] for the classic
1-ANT algorithm, although 1-ANT employs a much smaller sample size (i.e.,
sample size equals one). Moreover, by imposing upper and lower bounds on
the sampling probabilities as was done in MMAS [10], Wu et al [34] showed
further that the stochastic runtime of the resulting CE can be significantly
improved even in a very rugged search space.
The present article continues the stochastic runtime analysis of [34], but
now in combinatorial optimization with a study of CE on the traveling sales-
man problem (TSP). We emphasize the impact of the magnitude of N on
the stochastic runtime, put ρ = 1, and consider a CE variant resembling an
MMAS with iteration-best reinforcement under two different random solu-
tion generation mechanisms, namely, a vertex-based random solution gener-
ation and an edge-based random solution generation.
Stochastic runtime for MMAS with iteration-best reinforcement on sim-
ple problems like OneMax has been studied in [20] and [25]. In particular,
Neumann et al [20] showed that to obtain a stochastically polynomial run-
time for OneMax, N/ρ ∈ Ω(ln n) is necessary. We shall not only extend
this to TSP for the case of ρ = 1, but also prove that N ∈ ω(ln n) is already
sufficient to guarantee a stochastically polynomial runtime for simple TSP
instances.
4
TSP is a famous NP-complete combinatorial optimization problem. It
concerns finding a shortest Hamiltonian cycle on a weighted complete graph.
Existing algorithms exactly solving TSP generally have a prohibitive com-
plexity. For instance, the Held-Karp algorithm [35] solves the problem with
a complexity of O(n22n). A well-known polynomial time approximation al-
gorithm for metric TSP is the so-called Christofides algorithm [36], which
finds a solution with a cost at most 3/2 times the cost of optimal solutions.
As mentioned in [37], this is still the best known approximation algorithm
for the general metric TSP so far. For Euclidean TSP there exists a famous
polynomial-time approximation scheme (PTAS) by Arora, see [38]. To de-
sign a superior approximation algorithm, researchers in recent years tend to
study TSP instances with particular structures, see, e.g., [39].
Due to the prohibitive running time of exact algorithms, heuristics are
frequently employed in practice so as to efficiently compute an acceptable
solution for a TSP problem, e.g., the Lin-Kernighan (LK) algorithm [40].
As a popular heuristic, CE has also been applied in practice to solve TSP
instances, see [41] and [3]. The implementation there shows that CE can also
efficiently compute an acceptable solution.
In view of the high complexity of general TSP, we consider in our analysis
two classes of TSP instances with a particular structure. The first kind of
instances has been used in [19] and [42] for analyzing the expected runtime of
some MMAS variants with best-so-far reinforcement. These TSP instances
have polynomially many objective function values and a unique optimal so-
lution. Moreover, on these TSP instances, solutions containing more edges
from the optimal solution have a smaller cost than those with fewer such
edges. For more details on these instances, see Section 5.
For these simple instances, we prove in Theorem 2 that with a probability
1 − e−Ω(n), the runtime is O(n6+) with the vertex-based random solution
generation, and O(n3+ ln n) with the edge-based random solution generation,
for any constant ∈ (0, 1) and N ∈ Ω(n). For the case of N ∈ ω(ln n), we
show that the runtimes (resp., O(n6N and n3(ln n)N ) are even smaller with
probability 1 − n−ω(1), see Corollary 1. These results are very close to the
known expected runtime O(n6 + n ln n
) for (1 + 1) MMAA reported in [19],
and the expected runtime O(n3 ln n + n ln
Arb reported in [42]
(where ρ ∈ (0, 1) is an evaporation rate), if N ∈ ω(ln n) is suitably small.
But they give the stronger guarantee of achieving the optimal solution in the
respective runtime with an overwhelming probability. Moreover, we show
ρ ) for MMAS∗
ρ
5
a stochastically exponential runtime for a suitable choice of N ∈ O(ln n),
see Theorem 3. This generalizes the finding in [20] for OneMax to TSP
instances. Therefore, N ∈ Ω(ln n) is necessary, and N ∈ ω(ln n) is sufficient
for a stochastically polynomial runtime for simple TSP instances.
We also inspect more complex instances with n vertices positioned on an
m×m grid, and the Euclidean distance as distance function. These instances
have been employed in [43] and [30] for analyzing the expected runtime of
(µ+λ) EA and randomized local search (RLS). When the n vertices span a
convex polygon without vertices in the interior of the polygon (so they are
the corners of that polygon), we prove a stochastic runtime of O(n4m5+)
for the vertex-based random solution generation, and a stochastic runtime
of O(n3m5+) for the edge-based random solution generation, see Theorem
4 for details. Similarly, the in the stochastic runtimes can be removed
by slightly decreasing the probability guarantee, see Corollary 2. When the
vertices span a convex polygon with k ∈ O(1) vertices in the interior, we
show a stochastic runtime of O(n4m5+ + n6k−1m) with the vertex-based
random solution generation, and a stochastic runtime of O(n3m5+ + n3km)
with the edge-based random solution generation, see Theorem 5 for details.
These runtimes are better than the expected runtime for the so-called (µ+λ)
EA and RLS reported in the recent paper [30].
The remainder of this paper is arranged as follows. Section 2 defines the
traditional CE and related algorithms, Section 3 defines the traveling sales-
man problem and provides more details of the used CE variants, Section 4
shows some important facts on the two random solution generation methods,
and Section 5 reports the stochastic runtime results on the TSP instances.
A short conclusion and suggestions for future work are given in Section 6.
Notations for runtime
Our analysis employs some commonly used notations from complexity
theory. We use O(f (n)) to denote the class of functions which are bounded
from above by the function f (n), i.e., those functions g(n) with g(n) ≤ c·f (n)
for large enough n and some constant c ≥ 0 not depending on n. Similarly,
Ω(f (n)) is the class of functions that are bounded from below by f (n), i.e.,
for any g(n) ∈ Ω(f (n)) there exists a constant c > 0 not depending on n
such that g(n) ≥ c · f (n) for large enough n. Class Θ(f (n)) is the inter-
section of Ω(f (n)) and O(f (n)). Class o(f (n)) is the class of functions g(n)
with g(n)/f (n) → 0 as n → ∞, and class ω(f (n)) is the class of functions
6
g(n) with g(n)/f (n) → +∞ as n → ∞. Obviously, o(f (n)) ⊂ O(f (n)) and
ω(f (n)) ⊂ Ω(f (n)).
2. The general cross entropy algorithm and related algorithms
We now introduce the traditional CE algorithm. The CE variant we will
analyze inherits the framework of this traditional version. To compare our
results with those in the literature, we shall give also details about some
related algorithms.
2.1. The traditional cross entropy algorithm
Algorithm 1 lists the traditional CE that was proposed in [3], adapted to
an abstract notion of combinatorial optimization problems. The algorithm
assumes a combinatorial minimization problem (S, f ), where S is a finite
search space of "feasible" solutions and f is the cost function. Every feasible
solution s ∈ S is composed of elements from a fixed finite set A, the ground
set of the problem, i.e., we assume S ⊆ An for some integer n ∈ N. Further-
more there is a product distribution on the product space An that induces a
distribution on S ⊆ An. The distribution on An can usually be represented as
a vector (or matrix) of real-valued probabilities. The convex combination of
the two distributions in Step 6 of Algorithm 1 then corresponds to a convex
combination of the two vectors (or matrices).
Specific to the TSP, the ground set A can be the set of nodes or edges, n
is the number of nodes, and a feasible solution is a sequence of elements from
A that forms a Hamiltonian cycle. The product distribution for the TSP is
represented as an n×n matrix.
When we consider the set of nodes as our ground set A, each row i
of the matrix is a marginal distribution that specifies choice probabilities
for all nodes following the current node i. A random Hamiltonian cycle
is sequentially constructed from the product distribution by allowing only
nodes not yet visited as continuations in each step, see Algorithm 2 for more
details.
When we consider the set of edges as A, marginals of the product distri-
bution will be represented by the same n×n matrix where the sum of the
(i, j)-th and (j, i)-th entries reflects the probability that the edge {i, j} oc-
curs in a random solution. A random Hamiltonian cycle is still constructed
sequentially and only edges leading to a feasible solution are taken in each
step, see Algorithm 3 for details.
7
Algorithm 1 The general Cross-Entropy algorithm
Input:
an initial distribution Π0 on the solution space, a fixed smoothing param-
eter ρ ∈ (0, 1], a sample size N ∈ N+, an elite size M ∈ N+ with M ≤ N
1: t = 0;
2: loop
3:
4:
5:
t
6:
7:
8: end loop
t
independently generate N random solutions X(1)
current distribution Πt;
sort these N solutions in non-decreasing order as f (X[1]
f (X[N ]
learn an empirical distribution Wt
X[1]
t , . . . , X[M ]
set Πt+1 = (1 − ρ)Πt + ρWt;
t = t + 1;
) according to the cost function f ;
;
t
, . . . , X(N )
t with the
t ) ≤ ··· ≤
from the M best solutions
Traditionally, CE sets a small elite ratio α ∈ (0, 1) and uses the best
(cid:98)α · N(cid:99) solutions in Step 5 to build the empirical distribution Wt. Here, we
use the elite size M instead. This does not intrinsically change the original
algorithm. Steps 3 and 5 depend on the detailed definition of the underlying
problem. We shall give details to them in Subsection 3.2.
Step 6 of Algorithm 1 plays a crucial role in the different theoretical anal-
yses of the algorithm, see, e.g., [44], [33], [45], [9], [34]. The occurrence of
good solutions are probabilistically enforced by incorporating the new infor-
mation Wt into Πt+1. This idea, somehow, coincides with the reinforcement
learning in [46]. The smoothing parameter ρ reflects the relative importance
of the new information Wt in the next sampling.
It balances global ex-
ploration and local exploitation to a certain degree. A larger ρ makes the
algorithm concentrate more on the particular area spanned by the elite solu-
tions X[1]
t , . . . , X[M ]
, while a smaller ρ gives more opportunities to solutions
t
outside that area.
However, balancing global exploration and local exploitation through tun-
ing ρ is ultimately limited. Wu and Kolonko [33] proved that the famous
"genetic drift" [47] phenomenon also happens in this algorithmic scheme,
i.e., the sampling (Step 3) eventually freezes at a single solution and that
solution needs not to be optimal. This means that the algorithm gradually
8
loses the power of global exploration.
As a compensation for global exploration, Wu et al [34] proved that a
moderately large sample size N might be helpful. The results there showed
that a moderately large N configured with a large ρ (e.g., ρ = 1) can make the
algorithm very efficient. Although a large N introduces a high computational
burden in each iteration, the total number of iterations required for getting
an optimal solution is considerably reduced.
Wu et al [34] also indicated another way to compensate the global ex-
ploration, i.e., imposing a lower bound πmin ∈ (0, 1) and an upper bound
πmax ∈ (0, 1) on the sampling distributions in each iteration. This idea is
originated from MMAS [10]. In each iteration t, after applying Step 6, the
entries of distribution Πt+1 that are out of the range [πmin, πmax] are reset to
that range by assigning to them the nearest bounds, see (6) in Section 3 for
more details. Wu et al [34] have proved that this can make CE more efficient
even in the case of a rugged search space.
To follow these theoretical suggestions made in [34], we shall in our
stochastic runtime analysis use a CE that modifies the traditional CE (Al-
gorithm 1) accordingly. We shall see that these modifications make the CE
very efficient for the considered TSP instances.
2.2. Related evolutionary algorithms
Related evolutionary algorithms for TSP whose runtime has been exten-
sively studied are RLS [28], (µ + λ) EA [30], and those theoretical abstrac-
tions of MMAS [10] including MMAS∗
bs [17], (1+1) MMAA [19]. We now
give algorithmic details of them. In order to facilitate the comparison, their
runtimes for TSP instances will be discussed in Section 5.
(µ + λ) EA is an extension of the famous (1 + 1) EA [11]. (µ + λ) EA
In each iteration,
randomly chooses µ solutions as the initial population.
(µ+λ) EA randomly chooses λ parents from current population, then produces
λ children by applying randomized mutation to each of the selected parents,
and forms the next population by taking the best µ solutions from these
µ + λ solutions in the end of current iteration. The expected runtime of
(µ + λ) EA on TSP instances is studied in [30], where Sutton et al uses a
Poisson distribution to determine the number of randomized mutations (2-
opt move or jump operation) should be taken by a selected parent in each
iteration.
RLS is a local search technique [48]. It employs a randomized neighbor-
hood. In each iteration, it randomly chooses a number of components of the
9
best solution found so far and then changes these components. The expected
runtime of RLS for TSP instances is also studied in [30], where the neighbor-
hood is taken to be a k-exchange neighborhood with k randomly determined
by a Poisson distribution.
(1+1) MMAA is a simplified version of the famous MMAS [10], where
the sample size is set to 1 and pheromones are updated only with the best
solution found so far (best-so-far reinforcement) in each iteration. In each
iteration of (1+1) MMAA, the ant which constructed the best solution found
so far deposits an amount πmax of pheromones on the traversed edges, and an
amount πmin of pheromones on the non-traversed edges, and the pheromones
are updated by linearly combining the old and these newly added pheromones
as in Algorithm 1. The expected runtime of (1+1) MMAA on simple TSP
instances is studied in [19]. The expected runtime of its variant MMAS∗
on simple TSP instances is studied in [42].
Arb
3. The traveling salesman problem and details of the CE variant
Now, we formally define TSP, and give more details of the CE variant we
will analyze.
3.1. The traveling salesman problem
and edge set E = (cid:8){i, j} i ∈ V, j ∈ V, i (cid:54)= j(cid:9). A Hamiltonian cycle is a
We consider an undirected graph G = (V, E) with vertex set V = {1, . . . , n}
sequence {{il, il+1} l = 1, . . . , n} of edges such that
a) i1 = in+1;
b) (i1, . . . , in) is a permutation of {1, 2, . . . , n}.
This definition actually considers E as the ground set A. As mentioned above,
we can also put A = V and represent Hamiltonian cycles in a more com-
pact way as permutations of V. Note that a Hamiltonian cycle corresponds
to n different permutations, whereas a permutation corresponds to a unique
Hamiltonian cycle. However, the two representations are intrinsically the
same. We shall use them interchangeably in the sequel. To facilitate our
discussion, we shall refer to a Hamiltonian cycle by just referring to one of
the n corresponding permutations, and denote by S the set of all possible
permutations. We employ the convention that two permutations are said
10
to be same iff they form the same underlying Hamiltonian cycle. The no-
tation {k, l} ∈ s shall mean that the edge {k, l} belongs to the underlying
Hamiltonian cycle of the solution (permutation) s.
Once a distance function d : E (cid:55)→ R+ is given, the (total traveling) cost
f (s) of a feasible solution s = (i1, i2, . . . , in) ∈ S is then calculated by
n−1(cid:88)
f (s) :=
d(ij, ij+1) + d(in, i1).
(1)
j=1
We denote by S∗ ⊆ S the set of feasible solutions (Hamiltonian cycles) that
minimize the cost (1).
3.2. Details of the CE variant
The CE variant we consider in the analysis completely inherits the struc-
ture of Algorithm 1, and additionally employs a component from MMAS.
We now formalize the sampling distribution, and define Steps 3 and 5 in
more detail. As mentioned, we represent a sampling distribution (a product
distribution on An) for the TSP by a matrix Π = (πi,j)n×n, such that
a) (cid:80)n
j=1 πi,j = 1, for all i = 1, . . . , n,
b) πi,i = 0 for all i = 1, . . . , n,
c) πi,j = πj,i for each edge {i, j} ∈ E.
For each edge {i, j} ∈ E, πi,j reflects the probability that a Hamilton cycle
continues with vertex j when it is in vertex i. In the sequel, we write the
sampling distribution Πt in iteration t as (πt
i,j)n×n, where the superscript t of
πt
i,j indicates the iteration. The initial distribution Π0 = (π0
i,j)n×n is, without
loss of generality, set to be the uniform distribution, i.e., π0
i,j = π0
n−1 for
all edges {i, j} ∈ E.
j,i = 1
We shall consider two random solution generation methods, a vertex-based
random solution generation and an edge-based random solution generation.
Algorithm 2 lists the vertex-based random solution generation method. This
method uses V as the ground set A. A product distribution of An is therefore
represented as a matrix Π = (πi,j)n×n satisfying a)-c) above, i.e., each row
of Π represents a sampling distribution on A = V. Directly sampling from
Π may produce infeasible solutions from An − S. To avoid that, Algorithm
2 starts with a randomly fixed initial node, and then sequentially extends
11
a partial solution with an unvisited vertex until a complete permutation is
obtained. This method is efficient and rather popular in practice, see, e.g.,
[41] and [4]. Here, "s + (v)" means that appends a vertex v to the end of a
partial solution s.
Algorithm 2 Vertex-based random solution generation
Input:
a distribution Π = (πi,j)n×n
a permutation of 1, 2, . . . , n
Output:
1: s = ∅, and Vunivisted = V ;
2: randomly select v from V, s = s + (v), and Vunvisited = Vunvisited − {v};
3: while (Vunvisited (cid:54)= ∅) do
select a random vertex v(cid:48) from Vunvisited with a probability
4:
;
πv,k
(2)
(cid:80)
P[v(cid:48) s] =
πv,v(cid:48)
k∈Vunvisited
set s = s+(v(cid:48)), Vunvisited = Vunvisited − {v(cid:48)};
v = v(cid:48);
5:
6:
7: end while
8: return s;
The edge-based random solution generation is listed in Algorithm 3. The
idea is from [42]. This method considers edge set E as the ground set A. A
feasible solution is then a sequence of edges that form a Hamiltonian cycle, i.e.
S ⊆ En. To unify the notation of feasible solutions, Algorithm 3 translates
its outcomes into permutations. As the actual ground set is E, a product
distribution is an n× n(n−1)
2 matrix such that each row is a marginal specifying
a sampling distribution on E. Algorithm 3 only considers those with identical
marginals, a product distribution can be therefore fully characterized by
one of its marginals and is therefore again represented by an n × n matrix
Π = (πi,j)n×n as above. An edge {i, j} ∈ E is then sampled from Π with
l=1 πk,l = 2πi,j/n since each row of Π sums
up to 1. A random sequence ∈ En is generated by independently sampling
from Π n times. To avoid infeasible solutions, Algorithm 3 considers in
every sampling only edges that are admissible by the edges selected before.
Given a set B of edges such that the subgraph (V,B) does neither contain a
probability (πi,j + πj,i)/(cid:80)n
(cid:80)n
k=1
12
cycle nor a vertex of degree ≥ 3, an edge e(cid:48) ∈ E is said to be admissible by
B if and only if the subgraph (V,B ∪ {e(cid:48)}) still does neither contain a cycle
nor a vertex of degree ≥ 3. We denote by Badmissible the set of edges /∈ B that
are admissible by B.
Algorithm 3 Edge-based random solution generation
Input:
a distribution Π = (πi,j)n×n
a permutation of 1, 2, . . . , n
Output:
1: B = ∅, Badmissible = E;
2: while (B ≤ n − 1) do
select an edge {i, j} from Badmissible with a probability
3:
P[e s] =
(cid:80){k,l}∈Badmissible
πi,j + πj,i
;
(3)
πk,l + πl,k
set B =B∪{{i, j}};
update Badmissible;
4:
5:
6: end while
7: let s = (1, i2, i3, . . . , in) with {1, i2},{ij, ij+1} ∈ B for j = 2, . . . , n−1;
8: return s;
The N random solutions X(1)
in iteration t are then generated
by N runs of Algorithm 2 or Algorithm 3 with the current distribution Πt =
(πt
i,j)n×n is then calculated from
the M elite solutions by setting
i,j)n×n. The empirical distribution Wt = (wt
, . . . , X(N )
t
t
(cid:80)M
k=1
1{e(cid:48)∈E e(cid:48)∈X[k]
t }({i, j})
wt
i,j =
where 1A(·) is the indicator function of set A = {e(cid:48) ∈ E e(cid:48) ∈ X[k]
{i, j} ∈ E. The next distribution Πt+1 = (πt+1
,
(4)
t } for each
i,j )n×n is therefore obtained as
M
i,j = (1 − ρ)πt
πt+1
i,j + ρwt
i,j
(5)
for each {i, j} ∈ E.
We continue with the suggestions made in [34]. In the CE variant, we
shall use a moderately large N and a large ρ = 1. To fully use the best elite
13
πmin
πt+1
i,j
πmax
πt+1
i,j =
if πt+1
if πt+1
if πt+1
i,j < πmin,
i,j ∈ [πmin, πmax],
i,j > πmax,
(6)
solutions, we take M = 1. To prevent premature convergence (i.e., a possible
stagnation at a non-optimal solution), we employ a feature from MMAS
[10], called max-min calibration, in the construction of Πt+1,. We choose
a lower bound πmin ∈ (0, 1) and an upper bound πmax ∈ (0, 1), and, after
applying (5), adjust Πt+1 by
for any edge {i, j} ∈ E. Note that the max-min calibration is the only step
that does not occur in the general CE (i.e., Algorithm 1).
This setting turns CE into an MMAS with iteration-best reinforcement,
i.e., only the iteration-best solution X[1]
is allowed to change the 'pheromones'
t
Πt. Stutzle and Hoos [10] indicated in an empirical study that the practi-
cal performance of iteration-best reinforcement is comparable to best-so-far
reinforcement for TSP instances. Thus, it should also be worthwhile to com-
pare the theoretical runtime of iteration-best reinforcement with the known
expected runtimes of best-so-far reinforcement for TSP instances presented
in, e.g., [19] and [42].
4. Properties of the random solution generation methods
Before we start with our runtime analysis, we shall discuss some relevant
properties of the two random solution generation methods, which concern the
probability of producing a k-exchange move of the iteration-best solution in
the next sampling.
Formally, a k-exchange move on a Hamiltonian cycle is an operation that
removes k edges from the cycle and adds k new edges to obtain again a cycle.
A k-opt move is a k-exchange move reducing the total travel cost. Figure
1a shows an example of a 2-exchange move, in which edges {i, j},{k, l} are
removed, and edges {i, l},{k, j} are added.Figure 1b shows an example of a
3-exchange move.
In our analysis, we shall consider only iteration-best reinforcement with
ρ = 1 and the max-min calibration (6). The empirical distribution Wt =
i,j)n×n for each iteration t ∈ N in this particular case therefore satisfies
(wt
πt+1
i,j = πt+1
j,i =
min{1, πmax} = πmax
max{0, πmin} = πmin
if edge {i, j} ∈ X[1]
t ,
otherwise,
(7)
(cid:40)
14
(a) A 2-exchange move
(b) A 3-exchange move
Figure 1: Examples for edge exchange moves
for every edge {i, j} ∈ E and iteration t ∈ N. Furthermore, Πt+1 = Wt.
Since Πt+1 is biased towards the iteration-best solution X[1]
t , k-exchanges
of X[1]
t with a large k are unlikely to happen among the N draws from Πt+1
by either of the two generation methods. Thus, an optimal solution is more
likely to be reached by a sequence of repeatedly k-exchange moves with small
k from iteration-best solutions. Therefore, it is necessary to estimate the
probabilities of producing a k-exchange of X[1]
in the two generation methods,
t
especially for the case of small k.
4.1. Probabilities of producing k-exchanges in the vertex-based random solu-
tion generation
The probability of producing k-exchanges with k = 2, 3 in the vertex-
based random solution generation has been studied in Zhou [19]. With πmin =
n2 and πmax = 1 − 1
1
n , Zhou [19] proved for (1 + 1) MMAA that with a
probability of Ω(1/n5), Algorithm 2 produces a random solution having more
t (the best solution found so far) provided that x∗
edges from s∗ than x∗
t is not
t (cid:54)= s∗, then there exists either
optimal. Zhou [19] actually showed that if x∗
a 2-opt move or a 3-opt move for x∗
t , and Algorithm 2 produces an arbitrary
2-exchange of x∗
t with a probability of Ω(1/n3), and an arbitrary 3-exchange
of x∗
n(n−2) and consider iteration-best reinforcement,
a similar result holds in our case. Claim 1 below gives a lower bound on the
probability of producing a k-exchange move of the iteration-best solution in
the next round with the vertex-based random solution generation.
t with a probability of Ω(1/n5).
Although we use πmin = 1
15
ijklijklwv Claim 1. Let M = 1, ρ = 1, and consider a k-exchange move of X[1]
for some
t
integer k = 2, 3, . . . , n. Then, Algorithm 2 produces the given k-exchange
move with a probability Ω(1/n2k−1) in every of the N draws in iteration t + 1.
i,j (or, equivalently, πt
Proof. Recall that in Algorithm 2, the probability (2) to select a continuing
edge {i, j} is always bounded from below by πt
j,i) for
each iteration t ∈ N, since each row of Πt sums up to 1. Given a k-exchange
move of X[1]
t , one possibility to generate it from Πt+1 = Wt by Algorithm 2
is that one of the new edges is added in the last step. This happens with a
e·n2k−1 , where e ≈ 2.71828
probability at least 1
is Euler's number, 1
n represents the probability to select the starting vertex,
n(n−2) is the common lower bound of the probability to select the remaining
k − 1 new edges, and 1 − 1
n is the common lower bound of the probability to
select one of the remaining n − k edges from X[1]
t .
(cid:3)k−1 ·(cid:0)1 − 1
(cid:1)n−k ≥ 1
n ·(cid:2)
1
1
n(n−2)
n
Because of Claim 1, every 2-exchange of X[1]
t
is produced from Πt+1 by
Algorithm 2 with a probability Ω(1/n3), and every 3-exchange is produced
by Algorithm 2 with a probability Ω(1/n5). Note that for any k = 2, 3, . . . ,
if a k-opt move of X[1]
t occurs among the N draws in the next sampling,
t+1) < f (X[1]
then f (X[1]
t ) must hold. Thus, if we take a moderately large
sample size, say N = Θ(n5+) for some > 0, with a probability 1 − (1 −
Ω(n−5))Ω(5+) = 1− e−Ω(n), f (X[1]
t ) will hold, provided that there
still exists a 2-opt or 3-opt move from X[1]
t .
t+1) < f (X[1]
Claim 2. Suppose that M = 1, ρ = 1. Then, for iteration t + 1, the probability
that Algorithm 2 produces a solution with a cost not larger than X[1]
in one
t
application is in Ω(1).
Proof. Observe that the probability that X[1]
is reproduced in one application
of Algorithm 2 is larger than (1 − 1/n)n−1 ∈ Ω(1), which implies that the
t
cost of the generated random solution is not larger than f (X[1]
t ).
Note that if X[1]
t
next sampling, then f (Xt+1)[1] ≤ f (X[1]
Ω(ln n), then f (X[1]
O(1/n). Particularly, when N ∈ Ω(n) for some > 0, f (X[1]
with an overwhelming probability 1 − e−Ω(n).
is reproduced at least once among the N draws in the
t ). Thus, if the sample size N ∈
t ) with a probability 1 − (1 − Ω(1))N = 1 −
t+1) ≤ f (X[1]
t )
t+1) ≤ f (X[1]
16
4.2. Probabilities of producing k-exchanges in the edge-based random solution
generation
i,j = πt
j,i = πmin or πt
i,j = πt
Recall that in each iteration t, either πt
The behavior of the edge-based random solution generation is comprehen-
sively studied in [42]. Kotzing et al [42] proved for MMAS∗
Arb and a constant
k ∈ O(1) that, with a probability of Ω(1), Algorithm 3 produces a random
solution that is obtained by a k-exchange move from the best solution found
so far.
j,i = πmax
for any edge {i, j} ∈ E. For convenience, we will call an edge {i, j} ∈ E with
i,j = πt
πt
j,i = πmax a high edge, and otherwise a low edge. Kotzing et al [42]
an arbitrary fixed step conditioned on the event that l ≤ √
showed the probability of the event that Algorithm 3 chooses a high edge in
n low edges have
been chosen in some l steps before this step is 1 − O(1/n). Our setting is
only slightly different with from theirs, i.e., we use πmin = 1
n(n−2) but they
put πmin = 1
n(n−1). Thus, the result should also hold here. Claim 3 below
formally asserts this, readers may also refer to [42] for a similar proof.
√
Claim 3. Assume M = 1, ρ = 1. Then, the probability of choosing a high
edge at any fixed step in Algorithm 3 is at least 1 − 12/n if at most
n low
edges have been chosen before that step and there exists at least one high
admissible edge to be added.
Proof. We now fix a step n−m for some m = 0, 1, . . . , n−1, and assume that
l ≤ √
n low edges have been chosen before this step. Obviously, we still need
to add m + 1 ≥ 1 edges to obtain a complete solution. We now estimate the
numbers of admissible high and low edges in this step. Note that every of
the l low edges blocks at most 3 of the m+l remaining high edges (at most
two which are incident to the end points of the low edge, and at most one
that may introduce a cycle). So at least m+l − 3l = m−2l ≥ m−3l high
edges are available for adding in this step. Of course, it may happen that
there is no admissible high edges in this step. However, we are not interested
in such a case. We consider only the case that there exists at least one
admissible high edge in this step, i.e. the number of admissible high edges
in this step is at least max{1, m − 3l}. Note also that the n−m edges added
before partition the subgraph of G = (V, E) with vertices V and edges from
the partial solution constructed so far into exactly m connected components
(here, we see an isolated vertex also as a connected component). For any two
of the components, there are at most 4 admissible edges connecting them.
17
(cid:1),(cid:0)n
Therefore, there are at most min{4(cid:0)m
(cid:1)} admissible low edges. Observing
l ≤ √
1− min{4(cid:0)m
(cid:1),(cid:0)n
n, the probability of choosing a high edge in this step is bounded from
below by
(cid:40)
≥
(8)
2
2
2
2
√
if m > 3
n,
√
if m ≤ 3
n,
max{1, m − 3l}
where the first inequality is obtained by observing that
πmin
πmax
(cid:1)}
(cid:18)m
(cid:19)
2m2
(m−3l)n(n−2) ≥ 1− 3
1−
(n−2)
1− 12
n−2
(cid:18)n
(cid:19)
} ≤ min{2m2,
(cid:18)n
(cid:19)
,
2
n(n−2), and 2m2
m−3l ≤ 2
m− 3
1
√
n
m2
} ≤ 2m2,
2
≤ 3n.
min{4
2
τmax = 1 − 1/n, πmin = 1
With Claim 3, we can show that, for any t ∈ N and any fixed k ∈ O(1),
the probability of the event that a k-exchange of X[1]
is produced by one
t
application of Algorithm 3 is Ω(1), see Claim 4. Here, we shall use a different
proof from the one presented by Kotzing et al [42], which appears to us as
problematic.
Claim 4. Let M = 1, ρ = 1. For any k ∈ O(1), with probability Ω(1), the
random solution produced by Algorithm 3 is a k-exchange of X[1]
t .
Proof. Let k ∈ O(1) be arbitrarily fixed, and M be the set of all k-element
subsets of {1, 2, . . . , n/2} (where we assume without loss of generality that
n is even). Obviously, M ∈ Θ(nk) since k ∈ O(1). Let M ∈ M be an arbi-
trarily fixed k-element subset. The probability of the event that Algorithm 3
selects k new edges (low edges) at steps i ∈ M and n − k edges (high edges)
from X[1]
t at other steps, is bounded from below by
(cid:0)1 − O(
)(cid:1)n−k(cid:89)
i∈M
1
n
((cid:0)n−i+1
(cid:1) − (n − i + k + 1))πmin
2
n(n − 1)πmin + (n − i + k + 1)πmax
≥ Θ(
1
nk ),
(9)
where 1 − O(1/n) is a lower bound for the probability of selecting an edge
t . In each step i ∈ M, the edges chosen before partition the graph
from X[1]
into n−i+1 connected components, and for any two of the components there
exists at least 2 edges connecting them without introducing a cycle. Hence,
that the number of admissible high edges in this case is at most n− i + k + 1
(cid:1) admissible edges in each step i ∈ M. Notice also
there are at least(cid:0)n−i+1
2
18
((n−i+1
)−(n−i+k+1))πmin
n(n−1)πmin+(n−i+k+1)πmax
(n − i+k+1 is the maximal number of high edges that have not been chosen
of (9) is just the lower
before). Therefore, each factor
bound of the probability for choosing an admissible edge not belonging to
X[1]
t
in a step i ∈ M.
As a result, the probability of the random event that Algorithm 3 pro-
t with k ∈ O(1) in any of the N independent draws
nk ) ∈ Ω(1),
duces a k-exchange of X[1]
in iteration t+1 is bounded from below by M·Θ( 1
since new edges can also be added in steps l ≥ n/2.
2
nk ) = Θ(nk)·Θ( 1
Notice that in the edge-based random solution generation, for any k =
2, 3, . . . , n, any two k-exchanges of X[1]
t are generated with the same proba-
bility, since the generation does not require adding the edges in a particular
order. Therefore, by Claim 4, for any k ∈ O(1), any specified k-exchange of
X[1]
t will be produced with a probability Θ(1/nk). Since reproducing X[1]
t can
be seen as a 0-exchange of X[1]
t , we can thus derive the following conclusion.
Claim 5. Let M = 1, ρ = 1. With probability Ω(1), the random solution
generated by Algorithm 3 has a cost not larger than that of X[1]
t .
for some k < nδ.
Claim 6 shows that it is unlikely that the random solution generated by
t . This
Algorithm 3 is "very" different from the last iteration-best solution X[1]
will be fundamental for deriving the runtime lower bound.
Claim 6. Let M = 1, ρ = 1. For any δ ∈ (0, 1], with an overwhelming
probability 1− e−ω(nmin{δ,1/4}/2), the random solution generated by Algorithm 3
is a k-exchange move from X[1]
t
Proof. Let δ ∈ (0, 1] be arbitrarily fixed, and put γ = min{δ, 1/4}. To prove
the claim, we just need to show that with an overwhelming probability, the
random solution generated by Algorithm 3 is a k-exchange of X[1]
for some
k ≤ nγ ≤ n1/4. This is again implied by the fact that with an overwhelming
t
probability, at most nγ/2 low edges are chosen within the first T := n − 3nγ
4
steps in Algorithm 3, since the best case nγ/2 + 3nγ
is still smaller than nγ.
By Claim 3, for any k ≤ nγ/2 and any m ≤ T, Algorithm 3 chooses
4
high edges with a probability at least 1 − 12/n at step m if at most k edges
have been chosen before step m, since there exist at least n − m − 3k ≥
4 − 3nγ/2 ≥ 3 admissible high edges at step m.
3nγ
19
Let P denote the probability of the random event that at most nγ/2 low
edges are chosen within T steps, and Q the probability of the random event
that at least nγ/2 + 1 low edges are chosen within the same T steps. Then
P = 1 − Q. We shall bound Q from above, which will give a lower bound for
P.
Let E be the random event that at least nγ/2 + 1 low edges are chosen
within T steps. Then Q = P[E]. For each l = 1, . . . , nγ/2 + 1, we define
a random variable vl denoting the first step m ≤ T such that l low edges
are chosen within m steps. Obviously, E implies the random event E1 that
v1 < v2 < ··· < vnγ/2+1 ≤ T. Thus, Q ≤ P[E1], and P ≥ 1 − P[E1].
Observe that
P[E1] =
(cid:88)
a1<a2<···<a
nγ/2+1
≤T
P[v1 = a1, . . . , vnγ/2+1 = anγ/2+1],
and v1 = a1, . . . , vnγ/2+1 = anγ/2+1 is equivalent to the random event that
before step a1 only high edges are chosen, that at any step between al and
al+1 only high edges are chosen for any l with 1 ≤ l ≤ nγ/2, and that at steps
a1, . . . , anγ/2+1 only low edges are chosen. Thus, we have by Claim 3 that
P[v1 = a1, . . . , vnγ/2+1 = anγ/2+1] ≤(cid:0)12
(cid:1)nγ/2+1,
n
since at each step al, there exists at least one admissible high edge and we
do not care about what happens after step vnγ/2+1.
There are at most (cid:0)
anγ/2+1. Therefore, P ≥ 1 − P[E1] ≥ 1 −(cid:0)
have(cid:0)
(cid:1) different combinations for a1 < a2 < ··· <
(cid:1)nγ/2+1 = e−ω(nγ/2). Hence, P ≥ 1 − e−ω(nγ/2) is overwhelm-
(cid:1)nγ/2+1.
By Stirling's formula, and observing that nγ/1 + 1 ∈ o(T ), T ∈ Θ(n), we
n
nγ/2+1
(cid:1)(cid:0) 12
T
T
nγ/2+1
T
(cid:1)(cid:0) 12
n
nγ/2+1
ingly large.
5. Main results
We shall now analyze the stochastic runtime of our two different random
solution generation methods for two classes of TSP instances that have been
well studied in the literature.
20
5.1. Stochastic runtime analysis for simple instances
following distance function d : E → R on a graph with n vertices.
We first consider a class of simple TSP instances that is defined by the
1
if {i, j} = {i, i + 1} for each i = 1, 2, . . . , n − 1,
if {i, j} = {n, 1},
(10)
d({i, j}) =
1
n otherwise.
Obviously, TSP instances with this distance function have a unique optimal
solution s∗ = (1, 2, . . . , n) (in the sense of the underlying Hamiltonian cycle),
and s∗ has a cost of n. The cost of an arbitrary feasible solution s equals
k + (n − k) · n, where k is the number of edges ∈ s that are also in s∗. We
shall refer to these instances as G1 in the sequel.
The class G1 has been used in [19] and [42] for analyzing the expected
runtime of variants of MMAS. Zhou [19] proved that the (1 + 1) MMAA
algorithm has an expected runtime of O(n6 + n ln n
) on G1 in the case of non-
visibility (i.e., without the greedy distance information in the sampling), and
has an expected runtime of O(n5 + n ln n
) in the case of visibility (i.e., with
considering the greedy distance information in the sampling). Kotzing et al
[42] continued the study in [19]. They investigated the expected runtime of
(1 + 1) MMAA and its variant MMAS∗
Arb on G1 and other TSP instances
on which both (1 + 1) MMAA and MMAS∗
Arb have exponential expected
runtime. MMAS∗
Arb differs with (1 + 1) MMAA only in the random solution
generation. MMAS∗
Arb uses Algorithm 3 as its random solution generation
method, while (1 + 1) MMAA used Algorithm 2. Kotzing et al [42] proved
that MMAS∗
Arb has an expected runtime of O(n3 ln n + n ln n
) on G1.
ρ
ρ
Theorem 1 shows a stochastic runtime of O(n6+) for the CE variant with
the add-on, i.e., Algorithm 1 with max-min calibration (6), the vertex-based
random solution generation, and a stochastic runtime of O(n4+) for the
edge-based random solution generation. These results are comparable with
the above known expected runtimes. Although we are not able to get strictly
superior runtimes, our results are actually stronger and more informative.
ρ
Theorem 1 (Stochastic runtime of Algorithm 1 with max-min calibration
on G1). Assume that we set M = 1, ρ = 1, and use Algorithm 1 with the
max-min calibration (6) for the values πmin = 1
n(n−2), πmax = 1 − 1
n. Then
a) if we use the vertex-based random solution generation method (Algo-
rithm 2), and take a sample size N ∈ Ω(n5+) for any constant ∈
21
(0, 1), then with a probability at least 1− e−Ω(N/n5) the optimal solution
s∗ can be found within n iterations;
b) if we use the edge-based random solution generation method (Algorithm
3), and take a sample size N ∈ Ω(n3+) for a constant ∈ (0, 1), then
with a probability at least 1 − e−Ω(N/n3), the optimal solution can be
found within n iterations.
Proof. We prove the Theorem by showing that the probability of the random
event that before the optimal solution is met, the number of edges shared by
the iteration-best and optimal solution strictly increases is overwhelmingly
large. This implies that the optimal solution is found within n iterations,
since the optimal solution has only n edges. Furthermore, the runtimes
presented in the Theorem hold. We only discuss the case of a), b) follows
with an almost identical argument.
By [19] (see also proof of Theorem 2), if X[1]
t
is not optimal, it has at
least either a 2-opt move or a 3-opt move. Note that for G1, any k-opt
move of the iteration-best solution increases the number of its edges shared
with the optimal solution. By Claim 1, any 2-opt move is generated by
Algorithm 2 with probability Ω(n−3), and any 3-opt move is generated with
probability Ω(n−5). Thus, if X[1]
t+1 shares more edges with
t
t with a probability at least 1 − (1 − n−5)N =
the optimal solution than X[1]
1−e−Ω(N/n5) ∈ 1−e−Ω(n) if N ∈ Ω(n5+) for any > 0. Thus, this repeatedly
happens within polynomially many number of iterations with overwhelming
probability 1 − e−Ω(N/n5). This completes the proof.
is not optimal, X[1]
The stochastic runtimes of Theorem 1 are derived for a relatively large
sample size, namely N = Ω(n5+) and N = Ω(n3+). Actually, Theorem 1
may still hold for a smaller sample size. Theorem 2 partially asserts this.
It states that the total number of iterations required to reach the optimal
solution for both generation schemes may increase considerably if a smaller
sample size is used. However, the stochastic runtime does not increase. In-
terestingly, one can obtain a smaller stochastic runtime with a small sample
size for the edge-based random solution generation.
Theorem 2 (Stochastic runtime of Algorithm 1 on G1 for a small sample
size). Assume the conditions in Theorem 1, but set N ∈ Ω(n) for any ∈
(0, 1). Then:
22
a) For the vertex-based random solution generation, Algorithm 1 finds the
optimal solution s∗ within n6 iterations with a probability of 1− e−Ω(N ).
b) For the edge-based random solution generation, Algorithm 1 finds the
optimal solution s∗ within n3 ln n iterations with a probability of 1 −
e−Ω(N ).
Proof of Theorem 2. The proof shares a similar idea with that of Theorem 1.
However, we consider here the random event that the number of edges shared
by the iteration-best and optimal solution does not decrease and strictly
increases enough times within a specified polynomial number of iterations.
For a), we shall consider the first n6 iterations. By Claim 2, the number
of edges shared by the iteration-best and optimal solution does not decrease
with a probability 1 −(cid:0)1 − Ω(1)(cid:1)N = 1 − e−Ω(N ) (N ∈ Ω(n)). Therefore,
(cid:81)n6
the number does not decrease within the first n6 iterations with probability
t=0(1−e−Ω(N )) = 1−e−Ω(N ). By Claim 1, for every consecutive n5 iterations,
1 −(cid:0)(1 − n−5)N(cid:1)n5
if the starting iteration-best solution is not optimal, then with probability
= 1 − e−Ω(N ), the number will strictly increase at least
once within these n5 iterations. Therefore, with overwhelming probability
1− e−Ω(N ), the optimal solution will be reached within the period of the first
n6 iterations, since there are n many consecutive n5 iterations within that
period.
b) can be proved by a similar way with a). We shall consider the first
n3 ln n iterations. By Claim 4, with probability 1− (1− Ω(1))N = 1− e−Ω(N ),
the number of shared edges does not decrease in consecutive two iterations.
To complete the proof, we need an extra fact on 2, 3-exchanges.
t found
so far has n−k edges from the optimal solution s∗, then the probability of the
t+1 has at least n−k+1 edges from s∗, is in Ω(k/n3). We shall use a
event that s∗
different but simpler proof to show that this also holds in our case of iteration-
best reinforcement. And with this fact, if X[1]
n, then with probability 1 − ((1 − k · n−3)N )n3/k = 1 − e−Ω(N ), the number
of edges shared by the iteration-best solution and s∗ will strictly increase
at least once within the period [t, t + n3/k]. This implies that s∗ is sampled
within the first n3 ln n iterations with overwhelming probability 1 − e−Ω(N ),
since n3 ln n iterations can be partitioned into n many consecutive phases
[0, n2), [n2, n2 +n3/(n−1)), [n2 +n3/(n−1), n2 +n3/(n−1)+n3/(n−2)), . . . .
We now prove that fact.
(cid:84) s∗ = n−k for some 0 < k ≤
Kotzing et al [42] showed for MMAS∗
Arb that if the best solution s∗
t
23
We first show that when X[1]
2-opt move or a 3-opt move for X[1]
t
that X[1]
t
e∗ = {i, i + 1} be an edge in s∗ but not in X[1]
graph is exactly incident to two edges of s∗ and X[1]
there exists an edge e0 ∈ X[1]
incident to i, an edge e(cid:48)
and e0, e(cid:48)
{i, v}, and e(cid:48)
t ∩ s∗ = n− k with k > 0, then there exists a
(see also [19] for a similar proof). Assume
contains exactly n − k edges from s∗ for some integer k > 0. Let
t . Note that each node of the
t , respectively. Therefore
0 ∈ X[1]
incident to i+1,
0 are not in s∗. Figure 2 shows an example, where e0 is either {i, u} or
0 = {i+1, w} or
0 is either {i+1, w} or {i+1, y}. If e0 = {i, u} and e(cid:48)
t
t
0 = {i+1, y}, then there exists a 2-opt move of X[1]
Figure 2: Demonstration of adding a new edge. The solid edges represent the cycle X[1]
t .
if e0 = {i, v} and e(cid:48)
t which
0 of distance n and adds e∗ and another edge (either {u, w} or
removes e0, e(cid:48)
{v, y}) of distance at most n + 1 together. If e0 = {i, u}, e(cid:48)
0 = {i+1, y} or
0 = {i+1, w}, there is a 3-opt move of X[1]
e0 = {i, v}, e(cid:48)
t which removes e0, e(cid:48)
0,
and an edge e1 /∈ s∗, and adds edge e∗ and another two edges, this replacing
3 edges of distance n by 3 edges of distance at most 2n+1 together. Here,
observe the fact that adding e∗ to X[1]
results
in graph containing a cycle, and there must be an edge e1 ∈ X[1]
t on that cycle
that does not belong to s∗. We choose this edge as the edge e1. Therefore,
for each e∗ of the k remaining edges in s∗ that are not in X[1]
t , there exists a
2-opt or 3-opt move of X[1]
By Claim 4, for any l ∈ O(1), the probability of producing an l-exchange
t
of the iteration-best solution X[1]
t by Algorithm 3 in iteration t + 1 is Ω(1).
Since any two l-exchanges are produced with the same probability, the prob-
ability of producing a particular l-exchange in iteration t + 1 is Ω(1/nl). As
a result, Algorithm 3 produces for each edge e∗ ∈ s∗ − X[1]
t a 2-opt or 3-opt
move of X[1]
t
that adds edge e∗ with probability at least Ω(1/n3).
t and removing e0, e(cid:48)
0 from X[1]
t
that adds e∗.
24
ii+1e∗uvwye1 Note that the generation of a 2-exchange (or a 3-exchange) with two newly
added edges e2, e3 by Algorithm 3 includes two mutually exclusive cases (3!
cases for a 3-exchange): e2 is chosen before e3, or e3 is chosen before e2. It
is not difficult to see that these two cases (3! cases for 3-exchange) have the
same probability. Therefore, the probability of the event that Algorithm 3
t that e∗ as one of the newly added edges
generates a 2-opt or 3-opt move of X[1]
and selects e∗ before the other newly added edges, is bounded from below
t has k such e∗ and the corresponding k
by Ω(1/3!n3) = Ω(1/n3). Since X[1]
events are also mutually exclusive, we obtain that the probability that X[1]
t+1
t has exactly n − k edges from s∗ for
has more edges from s∗ than X[1]
a constant k > 0 is Ω(k/n3)
if X[1]
t
Corollary 1 further improves the stochastic runtime for an even smaller
sample size. It can be proved by an argument similar to the proof of Theorem
1, where we observe that (1− (1− p(n))ω(ln n))nl = 1− n−ω(1) for any constant
l > 0 and probability p(n) ∈ Ω(1), and that 1 − e−ω(ln n) = 1 − n−ω(1).
Corollary 1. Assume the conditions in Theorem 1, but let N ∈ ω(ln n).
Then:
a) For the vertex-based random solution generation, Algorithm 1 finds the
optimal solution s∗ within n6 iterations with a probability of 1− n−ω(1).
Particularly, if N = (ln n)2, the runtime is n6(ln n)2 with probability
1 − n−ω(1).
b) For the edge-based random solution generation, Algorithm 1 finds the
optimal solution s∗ within n3 ln n iterations with a probability of 1 −
n−ω(1). Particularly, if N = (ln n)2, the runtime is n3(ln n)3 with prob-
ability 1 − n−ω(1).
Theorem 2 tells that, for any ∈ (0, 1), a sample size of N ∈ Θ(n) is
already sufficient for iteration-best reinforcement to efficiently find an optimal
solution of simple TSP instances with an overwhelming probability. Corollary
1 further shows that N ∈ ω(ln n) even leads to a better runtime with a slightly
smaller but still overwhelming probability. Theorem 3 below shows that with
an overwhelming probability, the runtime of iteration-best reinforcement will
be exponential if N ∈ O(ln n), even if the instances are as simple as those in
G1.
25
Theorem 3. Assume the conditions of Theorem 1, but set N < 1
220 ln n.
Then, with probability 1 − e−Ω(n1/200), Algorithm 1 with edge-based solution
generation does not find the optimal solution s∗ within eΘ(n1/300) iterations.
Proof. We prove the Theorem by inspecting the probability of the random
event that, before the optimal solution is found, the cost of the iteration-
best solution X[1]
t will oscillate for exponentially many iterations with an
overwhelming probability. We shall consider this in the last stages of the
optimization process.
Let T0 be the first iteration which samples a solution containing at least
n − n1/4 + n1/5 edges from the optimal solution. We show that with an
overwhelming probability, the number of common edges in the iteration-best
and optimal solution will drop below n− n1/4 + n1/5 and the optimal solution
is not sampled before that. This will imply the conclusion of Theorem 3,
since, with an overwhelming probability, this phenomenon can repeatedly
occur exponentially many times before optimal solution is found.
To that end, we need to show the following:
1) For any 1/4 > δ > 0, if X[1]
t
contains at least n − nδ edges from the
optimal solution, then with a probability O( 1√
n), the random solution
generated by Algorithm 3 will contain more edges from the optimal
solution than X[1]
t
in iteration t + 1;
2) For any 1/4 > δ > 0, if X[1]
t
contains at least n − nδ edges from
the optimal solution, then with a probability Ω(1) (at least e−5), the
random solution generated by Algorithm 3 will contain fewer edges
from the optimal solution than X[1]
t
in iteration t + 1.
T0
However, we first use these two facts and show them afterwards.
By Claim 6, with probability 1−e−ω(n1/20), X[1]
contains at most n−n1/4+
n1/5 + n1/10 edges from the optimal solution, since the random event that
the number of common edges from the iteration-best and optimal solution
increases more than n1/10 in one iteration implies an occurrence of a Ω(n1/10)-
exchange. Similarly, by Claim 6 again, with probability 1 − e−ω(n1/200), the
iteration-best solution contains k ∈ [n − n1/4 + n1/5 − n1/6+1/100, n − n1/4 +
n1/5 + n1/10 + n1/6+1/100] edges from the optimal solution in each iteration
t ∈ [T0, T0 + n1/6]. This means that the optimal solution is not found in the
period [T0, T0 + n1/6] with an overwhelming probability. With the help of
1) and 2), we are now to show that within this period, the number of edges
26
shared by the iteration-best and optimal solution is significantly reduced with
an overwhelming probability. This will complete the proof.
To facilitate our discussion, we call an iteration a successful iteration
if its iteration-best solution contains more edges from the optimal solution
than the last iteration-best solution, and an iteration a failure iteration if its
iteration-best solution contains fewer edges from the optimal solution than
the last iteration-best solution.
6− 1
44 ), since N < 1
n1/3 ), since N < 1
By 1) and the subsequent discussion, the expected number of success-
ful iterations within [T0, T0 + n1/6] is O( ln n
220 ln n. Thus,
by the Chernoff bound, with probability 1 − e−Ω(n1/6), at most n1/100 suc-
cessful iterations can occur within [T0, T0 + n1/6]. By 2) and the subse-
quent discussion, the expected number of failure iterations in [T0, T0 + n1/6]
is Ω(n 1
220 ln n. By the Chernoff bound, it happens that
with probability 1 − e−Ω(n1/6), at least n1/7 failure iterations will occur in
[T0, T0 + n1/6]. Since a successful iteration can add at most n1/100 edges from
the optimal solution with probability 1 − e−ω(n1/200), it totally adds at most
n1/100 × n1/100 = n1/50 edges from the optimal solution to the iteration-best
solution within [T0, T0 +n1/6] with probability 1−e−ω(n1/200). Note that within
[T0, T0 + n1/6], with probability 1 − e−Ω(n1/6), at least n1/7 × 1 = n1/7 "good"
edges are removed from the iteration-best solution. Therefore, with over-
whelming probability 1 − e−Ω(n1/200), X[1]
T0+n1/6 will contain at most
n − n1/4 + n1/5 + n1/10 − n1/7 + n1/50 < n − n1/4 + n1/5
contains at most n− n1/4 + n1/5 +
edges from the optimal solution, since X[1]
T0
n1/10 iterations with probability 1 − e−Ω(n1/20). As a result, with probability
1−e−Ω(n1/200), the number of common edges in the iteration-best and optimal
solution will again be smaller than n− n1/4 + n1/5 in some iteration after T0,
and the optimal solution is not found before that. And this will repeatedly
happen eΘ(n1/300) times with probability 1 − e−Ω(n1/200).
By taking k = 2 and considering the(cid:0)n
(cid:1) 2-exchanges that happen in the first
To finish the proof, we now formally prove 1) and 2). We first consider 2).
√
n − 3
n steps in the proof of Claim 4, one can show a tighter probability
√
e5 for producing 2-exchanges of X[1]
lower bound 1
t by Algorithm 3. Here, we
observe that the probability of choosing a high edge at a step before n− 3
n
is at least 1 − 3/(n − 2), see the proof of Claim 3.
2
Note that if 2-exchanges deleting 2 edges from the optimal solution hap-
pen N times in an iteration, then the iteration will be a failure iteration.
27
By the above and the fact that any two k-exchanges happen with the same
probability, a failure iteration then occurs with a probability at least
(cid:32)
(cid:1) (cid:33)N
(cid:0)n−nδ
(cid:1)
(cid:0)n
2
2
1
e5
(cid:32)
≥
1
e5
(cid:1) (cid:33) 1
(cid:0)n−nδ
(cid:1)
(cid:0)n
2
2
220 ln n
∈ Ω(n−1/44),
where δ ∈ (0, 1/4) and N < 1
220 ln n. This asserts 2).
1) follows with a similar discussion. Since X[1]
t
is assumed to contain
at least n − nδ edges from the optimal solution for some δ ∈ (0, 1/4), and
since Ω(nδ)-exchanges happen with an overwhelmingly small probability, we
need to consider only O(nδ)-exchanges when we estimate the probability of a
successful iteration. For each k ∈ Ω(nδ), the proportion of failure k-exchanges
is bounded from below by
(cid:0)n−nδ
(cid:1)
(cid:0)n
(cid:1) = e− 2knδ
k
n + o(1) ≥ e−2n−1/2 + o(1),
k
since 0 < δ < 1/4, and k-exchanges removing k edges shared by the iteration-
best and optimal solution are not "successful" k-exchanges. Since for any
k ∈ Ω(nδ), any two k-exchanges happen with the same probability, and
since the sum of the probabilities of successful and failure k-exchanges is
smaller than 1, we conclude that successful O(nδ)-exchanges happen with a
probability smaller than 1−e2n−1/2 ∈ O( 1√
n). Therefore, a successful iteration
happens with a probability 1−(1−O( 1√
n))N ∈ O( ln n√
n ) since N < 1
220 ln n.
Theorem 3 generalizes the finding of [20] to simple TSP instances.
It
formally states that for ρ = 1, N ∈ Ω(ln n) is necessary to efficiently find
an optimal solution to TSP. By Theorem 3, Theorem 1, Theorem 2 and its
Corollary 1, we have clearly analyzed the impact of the size of N on the
resulting stochastic runtime for the simple TSP instances in the case of that
ρ = 1. N ∈ ω(ln n) is sufficient to find the optimal solution in a stochastically
polynomial runtime, and the degree of the polynomial may increase with N ,
but the probability guaranteeing the runtime is also increasing with N .
5.2. Stochastic runtime analysis for grid instances
Now, we consider more general TSP instances. Herein, the n vertices are
positioned on an m × m grid for some integer m ∈ N+. The vertices are
positioned in a way that no three of them are collinear. Figure 3 gives an
28
Figure 3: A grid instance
example of such an instance where m = 5 and n = 8. The weight of an edge
{l, k} ∈ E in this case is defined as the usual Euclidean distance d(l, k)
between vertex l and vertex k for every l, k = 1, . . . , n. In this section, we
shall refer to these TSP instances as grid instances.
Grid instances have been studied in [43] and [30]. Sutton and Neumann
[43] investigated the expected runtime of (1+1) EA and RLS for these in-
stances. As a continuation of [43], Sutton et al [30] further proved that
the more extensive algorithm (µ + λ) EA finds an optimal solution for the
instances expectedly in
O((µ/λ)n3m5+nm5+(µ/λ)n4k(2k−1)!)
iterations if every of the λ selected parents is mutated by taking a random
number of consecutive 2-exchange moves, and expectedly in
O((µ/λ)n3m5+nm5+(µ/λ)n2k(k−1)!)
iterations with a mixed mutation operator, where k denotes the number of
vertices that are not on the boundary of the convex hull of V. Sutton et
al [30] also studied general Euclidean TSP instances (without collinearity)
and showed similar results in terms of the maximum distance value dmax,
the minimum distance value dmin, k and the minimum angle in the triangles
formed by the vertices.
Before we present our stochastic runtime, we summarize some structural
properties of grid instances (some just follow from properties of general Eu-
clidean instances). We say that two different edges {i, j} and {k, l} intersect
with each other if there exists a point p such that p /∈ {i, j, k, l} locates
on both of the two edges, see, e.g., Figure 4a. We say that a solution is
29
01234554321 intersection-free if the corresponding Hamiltonian cycle does not contain in-
tersections, see, e.g., Figure 4b.
(a) intersection
(b) intersection free
Figure 4: Example for intersections
Obviously, the triangle inequality [49] holds for grid instances. Therefore,
removing an intersection by a (unique) 2-exchange move in a solution strictly
reduces the total traveling cost, see Figure 4a. Lemma 1 states the well known
fact that an optimal solution of grid instances is intersection-free.
Lemma 1. Optimal solutions of grid instances are intersection-free.
Figure 5: Example for a 2-opt move
We now restrict 2-opt moves to 2-exchange moves that remove an inter-
section. For example, removing edges {i, j},{k, l} in Figure 5 and adding
new edges {i, l},{k, j} form such a 2-opt move. Lemma 2 below says that
for grid instances, removing one intersection may reduce the total traveling
cost Ω(m−4) if it is applicable. We omit the simple proof here. Interested
readers may refer to [30] for a proof.
Lemma 2. If a feasible solution to a grid instance contains intersections,
then removing the intersection can reduce the total traveling cost Ω(m−4).
30
ikljpikljijklp The convex hull Y(V ) of the vertex set V is the smallest convex set in R2
that contains V . Its boundary is a convex polygon spanned by some vertices
with possibly other vertices in the interior of that polygon. Let V b denote
the set of vertices on the boundary of Y(V ). Figure 6 illustrates this.
Figure 6: Example of a convex hull
Quintas and Supnick [50] proved that if a solution s is intersection-free,
then the solution respects the hull-order, i.e., any two vertices in the sub-
sequence of s induced by the boundary (the outer polygon) of Y(V ) are
consecutive in s if and only if they are consecutive on the boundary of Y(V ).
Therefore, if V b = V, i.e., all of the vertices are on the convex hull, then
every intersection-free solution is optimal.
Theorem 4 below analyzes the stochastic runtime of Algorithm 1 for grid
instances for the case that V = V b. It states that the stochastic runtime is
O(n4·m5+) for the vertex-based random solution generation, and O(n3·m5+)
for the edge-based random solution generation. Corollary 2 further improves
the runtime by sacrificing the probability guarantee. These stochastic run-
times are close to the expected runtime O(n3·m5) for RLS reported by Sutton
et al [43] and [30].
Theorem 4. Consider a TSP instance with n vertices located on an m × m
grid such that no three of them are collinear. Assume that V b = V , i.e., every
vertex in V is on the convex hull V b, that we apply the max-min calibration
n(n−2), ρ = 1, M = 1 and N ∈ Ω(m) for some
(6) with πmax = 1− 1
constant > 0. Then:
n, πmin = 1
a) With an overwhelming probability of 1 − e−Ω(N ), Algorithm 1 finds the
optimal solution within at most n4 · m5 iterations with the vertex-based
random solution generation.
b) With an overwhelming probability of 1 − e−Ω(N ), Algorithm 1 finds an
optimal solution within at most n3 · m5 iterations with edge-based ran-
dom solution generation.
31
Proof of Theorem 4. Note that under the conditions of Theorem 4, every
intersection free solution is optimal. By Lemma 2, we know that a 2-opt move
reduces the total traveling cost by Ω(m−4). Therefore, n·m5 consecutive 2-opt
moves turn a feasible solution into an optimal one, since the worst solution in
this case has a total traveling cost smaller than n·m and the optimal solution
has total traveling cost larger than n. Notice also that m≥ n/2, since the n
vertices are positioned on the m× m grid and no three of them are collinear.
With these facts, we prove the Theorem by a similar argument to the one
used in the proof of Theorem 2.
Again, we consider the random event that the cost of the iteration best
solution does not increase within a specified period of polynomially many
iterations and strictly decreases sufficiently many times within that period.
For a), we consider the first n4m5 iterations. For b), we consider the first
n3m5 iterations.
For a) : By Claim 2, with probability (1− (1− Ω(1))N )n4m5 = 1− e−Ω(N ),
the cost of the iteration-best solution does not increase within n4m5 itera-
tions. By Claim 1, for a phase consisting of consecutive n3 iterations, with
probability 1 − (1 − n−3)N·n3 = 1 − e−Ω(N ), in at least one iteration of that
phase an intersection is removed from the iteration-best solution, provided
the phase starts with an iteration-best solution containing at least one in-
tersection. Since the first n4m5 iterations can have nm5 such phases, a)
follows.
b) follows with an almost identical discussion. We therefore omit the
proof.
Corollary 2. Consider a TSP instance with n vertices located on an m× m
grid such that no three of them are collinear. Assume that V b = V , i.e., every
vertex in V is on the convex hull V b, that we apply the max-min calibration
(6) with πmax = 1 − 1
n(n−2), ρ = 1, M = 1 and N ∈ ω(ln m). Then:
a) With probability 1 − m−ω(1), Algorithm 1 finds the optimal solution
within at most n4 · m5 iterations with the vertex-based random solu-
tion generation.
n , πmin = 1
b) With probability 1−m−ω(1), Algorithm 1 finds an optimal solution within
at most n3 · m5 iterations with the edge-based random solution genera-
tion.
Now, we consider the more interesting case that V − V b = k ∈ O(1),
i.e., k vertices are not on the convex hull. Note that we can turn an arbitrary
32
intersection-free solution to an optimal solution only by rearranging the po-
sitions of those k interior points in that solution, and this requires at most k
consecutive jump moves (see [30] for a proof). A jump move δi,j transforms a
solution into another solution by shifting positions i, j as follows. Solution s
is transformed into solution δi,j(s) by moving the vertex at position i into po-
sition j while vertices at positions between i and j are shifted appropriately,
e.g.,
δ2,5(i1, i2, i3, i4, i5, i6, i7) = (i1, i3, i4, i5, i2, i6, i7) and
δ5,2(i1, i2, i3, i4, i5, i6, i7) = (i1, i5, i2, i3, i4, i6, i7).
It is not difficult to see that a jump move δi,j can be simulated by either a
2-exchange move (in the case that i−j = 1) or a 3-exchange move (in all
other cases). Therefore, we can actually turn an intersection-free solution
into an optimal one by a sequence of at most k consecutive 2-exchange or
3-exchange moves. Furthermore, a sequence of k consecutive 2-exchange or
3-exchange moves can be simulated by a κ-exchange move with an integer
κ ≤ 3k. This means that any intersection-free solution can be turned into an
optimal solution by a κ-exchange move with κ ≤ 3k. We shall call such a κ-
exchange move in the sequel a 3k-opt move, although κ may be smaller than
n6k−1 ) by
3k. Recall that a 3k-opt move is produced with a probability of Ω(
Algorithm 2 (see Claim 1), and with a probability of Ω( 1
n3k ) by Algorithm
3 (see Lemma 6 of [42], or Claim 4) in any of the N independent draws in
iteration t, if X[1]
t−1 is intersection-free and not optimal. As a result, we obtain
by a similar proof as above Theorem 5 below.
Theorem 5. Consider a TSP instance with n vertices located on an m × m
grid such that no three of them are collinear. Assume that V − V b = k ∈
O(1) (k vertices are not on the convex hull V b), that we apply the max-min
calibration 6 with πmax = 1 − 1
n(n−2), and set ρ = 1, M = 1, for
some constant > 0. Then:
n , πmin = 1
1
a) If we set N ∈ Ω(n3 · m), then with an overwhelming probability of
1 − e−Ω(N/n3), Algorithm 1 finds an optimal solution within at most n ·
m5 + n6k−4 iterations with the vertex-based random solution generation;
b) If we set N ∈ Ω(n2 · m), then with an overwhelming probability of
1 − e−Ω(N/n2), Algorithm 1 finds an optimal solution within at most
n·m5 +n3k−2 iterations with the edge-based random solution generation.
33
Proof of Theorem 5. We only prove a). b) can be derived by a very similar
argument. We define two random events as following:
E1 : for each t ≤ n · m5 + n6k−4, f (X[1]
E2 : for each t ≤ n· m5 + n6k−4, if X[1]
t−1 is not intersection-free, then a 2-opt
t−1) ≥ f (X[1]
t );
move happens in iteration t.
By a similar argument as the one for Theorem 4, we obtain that P[E1∩E2] ≥
1−e−Ω(N/n3). Let η be a random variable denoting the number of iterations for
t−1 is intersection-free. Notice that, conditioned on E1∩E2, η ≥ n·m5
which X[1]
implies that an optimal solution occurs within n · m5 + n6k−4 iterations.
Conditioned on E1 ∩ E2 and η < n · m5, there are at least Ω(n6k−4) itera-
tions in which Xt−1 is intersection-free, since each X[1]
t−1 is either intersection-
free or not intersection-free. Note also that in each iteration in which X[1]
t−1
intersection-free and not optimal, a 3k-opt move that turns X[1]
t−1 into an op-
timal solution happens with probability of at least 1 − (1 − Ω(
n6k−1 ))N . This
means for any fixed t ∈ N, if X[1]
t−1 is intersection-free, then the probability of
the event that X[1]
n6k−1 ))N .
t
Therefore, for any fixed Ω(n6k−4) iterations in which the iteration-best solu-
tion X[1]
t−1 is intersection-free and not optimal, the probability of the event
that the corresponding Ω(n6k−4) X[1]
t 's are still not optimal, is bounded from
above by (1 − Ω(
n6k−1 ))N·n6k−4 = e−Ω(N/n3). This means that, conditioned on
E1 ∩ E2 and η < n · m5, an optimal solution occurs within n · m5 + n6k−4
iterations with a probability of 1 − e−Ω(N/n3).
As a result, an optimal solution occurs within the first n · m5 + n6k−4
iterations with a probability of 1 − e−Ω(N/n3).
is optimal is bounded from below by 1− (1− Ω(
1
1
1
Theorem 5 shows a stochastic runtime of n3m5+ +n6k−1m for Algorithm
1 equipped with the vertex-based solution generation, and a stochastic run-
time of n3m5+ + n3km for Algorithm 1 equipped with edge-based solution
generation, in the case of that V − V b = k ∈ O(1). This is much better
than the expected runtime
O(µ · n3m5+nm5+µ · n4k(2k−1)!)
for (µ+λ) EA with sequential 2-opt mutations reported by Sutton et al [30].
However, we are not able to analyze the stochastic runtime in the case that
34
k ∈ ω(1), since k ∈ ω(1) interior points may require super-polynomially
many iterations to turn an intersection-free solution into an optimal solution
when a polynomial sample size is used.
6. Conclusion
We have analyzed the stochastic runtime of a CE algorithm on two classes
of TSP instances under two different random solution generation methods.
The stochastic runtimes are comparable with corresponding expected run-
times reported in the literature.
Our results show that the edge-based random solution generation method
makes the algorithm more efficient for TSP instances in most cases. More-
over, N ∈ Ω(ln n) is necessary for efficiently finding an optimal solution with
iteration-best reinforcement. For simple instances, N ∈ ω(ln n) is sufficient
to efficiently find an optimal solution with an overwhelming probability, and
N ∈ O(ln n) results in an exponential runtime with an overwhelming proba-
bility. However, for more difficult instances, one may need to use a relatively
large sample size.
Our stochastic runtimes are better than the expected runtimes of the
(µ + λ) EA on the grid instances. The EA randomly changes local structures
of some of its current solutions by a Poisson distributed number of consecu-
tive 2-exchange moves in every iteration, while our algorithm refrains from
local operations on current solutions and only refreshes solutions by sam-
pling from an evolving distribution. The solution reproducing mechanism in
the EA stays the same throughout the optimization, only the current solu-
tions in every iteration vary. However, the solution reproducing mechanism
(sampling distribution) of our algorithm also evolves. This is the essential
difference of MBS with traditional EAs. The comparison of our results with
the expected runtimes in [30] therefore show that using a self-adaptive dy-
namic solution reproducing mechanism is helpful (in efficiently finding an
optimal solution) when the search space becomes rugged. The stochastic
runtimes in Theorem 4 are only valid for instances with a bounded number
of interior points. In the future, it should be interesting to analyze the case
that V − V b ∈ ω(1). This might also give more insight to the problem of
RP v.s. P [51].
Our analysis is actually a kind of worst-case analysis, which is rather
pessimistic. We analyze the optimization progress by only checking some very
particular random events. This may not only underestimate the probability
35
of finding an optimal solution with our algorithm, but also overestimate the
required number of iterations. In the future, it should be of great interest to
consider a smoothed runtime analysis over an -neighborhood of the n nodes
in the real plane as has been done for the Simplex method by Spielman and
Teng in their famous paper [52].
Acknowledgment
We thank the anonymous reviewers for their numerous useful suggestions
on improving the scientific quality and English presentation of this article.
References
[1] R. Y. Rubinstein, D. P. Kroese, The cross-entropy method: a unified
approach to combinatorial optimization, Monte-Carlo simulation and
machine learning, Springer Science & Business Media, 2004.
[2] R. Y. Rubinstein, Optimization of computer simulation models with rare
events, European Journal of Operational Research 99 (1) (1997) 89–112.
[3] R. Y. Rubinstein, The cross-entropy method for combinatorial and con-
tinuous optimization, Methodology and computing in applied probabil-
ity 1 (2) (1999) 127–190.
[4] M. Dorigo, T. Stutzle, Ant colony optimization, Cambridge, Mas-
sachusetts: A Bradford Book, MIT Press, 2004.
[5] M. Hauschild, M. Pelikan, An introduction and survey of estimation
of distribution algorithms, Swarm and Evolutionary Computation 1 (3)
(2011) 111–128.
[6] M. Zlochin, M. Birattari, N. Meuleau, M. Dorigo, Model-based search
for combinatorial optimization: A critical survey, Annals of Operations
Research 131 (1-4) (2004) 373–395.
[7] D. Whitley, A genetic algorithm tutorial, Statistics and computing 4 (2)
(1994) 65–85.
[8] H. R. Louren¸co, O. C. Martin, T. Stutzle, Iterated local search, Springer,
2003.
36
[9] Z. Wu, Model-based heuristics for combinatorial optimization: a math-
ematical study of their asymptotic behavior, Ph.D. thesis, Institut fur
Angewandte Stochastik und Operations Research (IASOR), Technical
University of Clausthal (2015).
[10] T. Stutzle, H. H. Hoos, MAX-MIN ant system, Journal of Future Gen-
eration Computer Systems 16 (2000) 889–914.
[11] S. Droste, T. Jansen, I. Wegener, On the analysis of the (1+1) evolution-
ary algorithm, Theoretical Computer Science 276 (1-2) (2002) 51–81.
[12] J. He, X. Yao, Drift analysis and average time complexity of evolutionary
algorithms, Artificial Intelligence 127 (1) (2001) 57–85.
[13] F. Neumann, C. Witt, Runtime analysis of a simple ant colony optimiza-
tion algorithm, Tech. rep., Departmant of Computer Science, University
of Dortmund, Germany (2006).
[14] C. Witt, Runtime analysis of the (µ +1) ea on simple pseudo-boolean
functions, Evolutionary Computation 14 (1) (2006) 65–86.
[15] F. Neumann, C. Witt, Runtime analysis of a simple ant colony opti-
mization algorithm, Algorithmica 54 (2) (2009) 243–255.
[16] B. Doerr, F. Neumann, D. Sudholt, C. Witt, Runtime analysis of the 1-
ant ant colony optimizer, Theoretical Computer Science 412 (17) (2011)
1629–1644.
[17] W. J. Gutjahr, G. Sebastiani, Runtime analysis of ant colony optimiza-
tion with best-so-far reinforcement, Methodology & Computing in Ap-
plied Probability 10 (3) (2008) 409–433.
[18] Y. Zhou, J. He, A runtime analysis of evolutionary algorithms for con-
strained optimization problems, IEEE Transactions on Evolutionary
Computation 11 (5) (2007) 608–619.
[19] Y. Zhou, Runtime analysis of an ant colony optimization algorithm for
tsp instances, Evolutionary Computation IEEE Transactions on 13 (5)
(2009) 1083–1092.
37
[20] F. Neumann, D. Sudholt, C. Witt, A few ants are enough:aco with
iteration-best update, in: Genetic and Evolutionary Computation Con-
ference, GECCO 2010, Proceedings, Portland, Oregon, Usa, July, 2010,
pp. 63–70.
[21] P. S. Oliverto, C. Witt, Improved time complexity analysis of the simple
genetic algorithm, Theoretical Computer Science 605 (15) (2015) 21–41.
[22] D. Sudholt, C. Thyssen, Runtime analysis of ant colony optimization for
shortest path problems, Journal of Discrete Algorithms 10 (10) (2012)
165–180.
[23] A. Lissovoi, C. Witt, Runtime analysis of ant colony optimization on dy-
namic shortest path problems, Theoretical Computer Science 561 (2015)
73–85.
[24] Y. Chen, X. Zou, Runtime analysis of a multi-objective evolutionary al-
gorithm for obtaining finite approximations of pareto fronts, Information
Sciences 262 (2014) 62–77.
[25] D. Sudholt, C. Witt, Update strength in edas and aco: How to avoid
genetic drift, in: Genetic and Evolutionary Computation Conference,
2016, pp. 61–68.
[26] D. H. Wolpert, W. G. Macready, No free lunch theorems for optimiza-
tion, IEEE Transactions on Evolutionary Computation 1 (1) (1997) 67–
82.
[27] F. Neumann, D. Sudholt, C. Witt, Analysis of different mmas aco algo-
rithms on unimodal functions and plateaus, Swarm Intelligence 3 (2009)
35–68.
[28] F. Neumann, I. Wegener, Randomized local search, evolutionary algo-
rithms, and the minimum spanning tree problem, Theoretical Computer
Science 378 (2007) 32–40.
[29] J. Reichel, M. Skutella, Evolutionary algorithms and matroid optimiza-
tion problems, Algorithmica 57 (1) (2010) 187–206.
[30] A. M. Sutton, F. Neumann, S. Nallaperuma, Parameterized runtime
analyses of evolutionary algorithms for the planar euclidean traveling
salesperson problem, Evolutionary Computation 22 (4) (2014) 595–628.
38
[31] A. M. Sutton, J. Day, F. Neumann, A parameterized runtime analysis
of evolutionary algorithms for max-2-sat, in: Conference on Genetic &
Evolutionary Computation, 2012, pp. 433–440.
[32] Y. Zhou, X. Lai, K. Li, Approximation and parameterized runtime anal-
ysis of evolutionary algorithms for the maximum cut problem., IEEE
Transactions on Cybernetics 45 (8) (2015) 1491–1498.
[33] Z. Wu, M. Kolonko, Asymptotic properties of a generalized cross entropy
optimization algorithm, IEEE Transactions on Evolutionary Computa-
tion 18 (5) (2014) 658 – 673.
[34] Z. Wu, M. Kolonko, R. H. Mohring, Stochastic runtime analysis of the
cross entropy algorithm, IEEE Transactions on Evolutionary Computa-
tion, DOI: 10.1109/TEVC.2017.2667713.
[35] M. Held, R. M. Karp, A dynamic programming approach to sequencing
problems, Journal of the Society for Industrial and Applied Mathematics
10 (1) (1962) 196–210.
[36] N. Christofides, Worst-case analysis of a new heuristic for the travelling
salesman problem, Tech. rep., Graduate School of Industrial Adminis-
tration, CMU (1976).
[37] M. T. Goodrich, R. Tamassia, Algorithm Design and Applications, Wi-
ley, 2015.
[38] S. Arora, Polynomial time approximation schemes for Euclidean travel-
ing salesman and other geometric problems, Journal of the ACM 45 (5)
(1988) 753–782.
[39] J. S. B. Mitchell, A constant-factor approximation algorithm for tsp with
pairwise-disjoint connected neighborhoods in the plane, in: Twenty-
Sixth Symposium on Computational Geometry, 2010, pp. 183–191.
[40] S. Lin, B. W. Kernighan, An effective heuristic algorithm for the
traveling-salesman problem, Operations Research 21 (2) (1973) 498–516.
[41] P. T. D. Boer, D. P. Kroese, S. Mannor, R. Y. Rubinstein, A tutorial
on the cross-entropy method, Annals of Operations Research 134 (1)
(2005) 19–67.
39
[42] T. Kotzing, F. Neumann, H. Roglin, C. Witt, Theoretical analysis of two
aco approaches for the traveling salesman problem, Swarm Intelligence
6 (1) (2012) 1–21.
[43] A. M. Sutton, F. Neumann, A parameterized runtime analysis of evolu-
tionary algorithms for the euclidean traveling salesperson problem, in:
Proceedings of the Twenty-Sixth Conference on Artificial Intelligence
(AAAI'12), AAAI press, 2012, pp. 1105–1111.
[44] A. Costa, O. D. Jones, D. Kroese, Convergence properties of the cross-
entropy method for discrete optimization, Operations Research Letters
35 (5) (2007) 573–580.
[45] Z. Wu, M. Kolonko, Absorption in model-based search algorithms for
combinatorial optimization, in: Evolutionary Computation (CEC), 2014
IEEE Congress on, IEEE, 2014, pp. 1744–1751.
[46] M. Thomas, Machine learning, New Delhi: McGraw Hill Education In-
dia, 1997.
[47] H. Asoh, H. Muhlenbein, On the mean convergence time of evolutionary
algorithms without selection and mutation, in: Parallel Problem Solving
from Nature-PPSN III, Springer, 1994, pp. 88–97.
[48] M. Pirlot, General local search methods, European Journal of Opera-
tional Research 92 (3) (1996) 493–511.
[49] M. A. Khamsi, W. A. Kirk, An introduction to metric spaces and fixed
point theory, John Wiley,, 2001.
[50] L. V. Quintas, F. Supnick, On some properties of shortest Hamiltonian
circuits, American Mathematical Monthly 72 (9) (1965) 977–980.
[51] W. Gasarch, Classifying problems into complexity classes, Advances in
Computers 95 (2015) 239–292.
[52] D. A. Spielman, S. H. Teng, Smoothed analysis of algorithms: Why the
simplex algorithm usually takes polynomial time, Journal of the Acm
51 (3) (2004) 385–463.
40
|
1503.01203 | 2 | 1503 | 2015-04-02T08:31:44 | On the Number of Minimal Separators in Graphs | [
"cs.DS",
"cs.DM",
"math.CO"
] | We consider the largest number of minimal separators a graph on n vertices can have at most.
We give a new proof that this number is in $O( ((1+\sqrt{5})/2)^n n )$.
We prove that this number is in $\omega( 1.4521^n )$, improving on the previous best lower bound of $\Omega(3^{n/3}) \subseteq \omega( 1.4422^n )$.
This gives also an improved lower bound on the number of potential maximal cliques in a graph. We would like to emphasize that our proofs are short, simple, and elementary. | cs.DS | cs |
On the Number of Minimal Separators in
Graphs
Serge Gaspers1,2 and Simon Mackenzie1,2
1 The University of New South Wales, Australia,
{sergeg,simonwm}@cse.unsw.edu.au
2 NICTA, Australia
Abstract. We consider the largest number of minimal separators a
graph on n vertices can have at most.
-- We give a new proof that this number is in O (cid:16)(cid:16) 1+√5
-- We prove that this number is in ω (1.4521n), improving on the pre-
2 (cid:17)n
· n(cid:17).
vious best lower bound of Ω(3n/3) ⊆ ω(1.4422n).
This gives also an improved lower bound on the number of potential
maximal cliques in a graph. We would like to emphasize that our proofs
are short, simple, and elementary.
1
Introduction
For a graph G = (V, E), and two vertices a, b ∈ V , a vertex subset S ⊆ V \ {a, b}
is an (a, b)-separator if a and b are in different connected components of G − S,
the graph obtained from G by removing the vertices in S. An (a, b)-separator
is minimal if it does not contain another (a, b)-separator as a subset. A vertex
subset S ⊂ V is a minimal separator in G if it is a minimal (a, b)-separator for
some pair of distinct vertices a, b ∈ V .
By sep(G), we denote the number of minimal separators in the graph G. By
sep(n), we denote the maximum number of minimal separators, taken over all
graphs on n vertices.
Potential maximal cliques are closely related to minimal separators, especially
in the context of chordal graphs. A graph is chordal if every induced cycle has
length 3. A triangulation of a graph G is a chordal supergraph of G obtained by
adding edges. A graph H is a minimal triangulation of G if it is a triangulation
of G and G has no other triangulation that is a subgraph of H. A vertex set is
a potential maximal clique in G if it is a maximal clique in at least one minimal
triangulation of G.
By pmc(G), we denote the number of potential maximal cliques in the graph
G. By pmc(n), we denote the maximum number of potential maximal cliques,
taken over all graphs on n vertices.
Minimal separators and potential maximal cliques have been studied ex-
tensively [1,2,3,10,12,15,16,19,20,21]. Upper bounds on sep(n) are used to up-
per bound the running time of algorithms for enumerating all minimal sepa-
rators [1,16,20]. Bounds on both sep(n) and pmc(n) are used in analyses of
2
Serge Gaspers and Simon Mackenzie
a
b
...
Fig. 1. Melon graphs have Ω(3n/3) minimal separators.
algorithmic running times for computing the treewidth and minimum fill-in of
a graph [3,10,12], and for computing a maximum induced subgraph isomorphic
to a graph from a family of bounded treewidth graphs [11].
Our results. Fomin et al. [10] proved that sep(n) ∈ O (1.7087n). Fomin and
Villanger [12] improved the upper bound and showed that sep(n) ∈ O (ρn · n),
where ρ = 1+√5
2 = 1.6180 . . . 3. We prove the same upper bound with simpler
arguments.
As for lower bounds, it is known [10] that sep(n) ∈ Ω(3n/3); see Fig. 1.
We improve on this lower bound by giving an infinite family of graphs with
ω (1.4521n) minimal separators. This answers an open question raised numerous
times (see, e.g., [9,10]), for example by Fomin and Kratsch [9, page 100], who
state
It is an open question, whether the number of minimal separators in
every n-vertex graph is O∗(3n/3).
Here, the O∗-notation is similar to the O-notation, but hides polynomial factors.
As a corollary, we have that there is an infinite family of graphs, all with
ω(1.4521n) potential maximal cliques. This answers another open question on
lower bounds for potential maximal cliques. For example, Fomin and Villanger
[11] state
There are graphs with roughly 3n/3 ≈ 1.442n potential maximal
cliques [10]. Let us remind that by the classical result of Moon and
Moser [18] (see also Miller and Muller [17]) that the number of maxi-
mal cliques in a graph on n vertices is at most 3n/3. Can it be that the
right upper bound on the number of potential maximal cliques is also
roughly 3n/3? By Theorem 3.2, this would yield a dramatic improvement
for many moderate exponential algorithms.
3 The bound stated in [12] is O(1.6181n ), but this stronger bound can be derived from
their proof.
On the Number of Minimal Separators in Graphs
3
Preliminaries. We use standard graph notation from [4]. For an edge uv in a
graph G, we denote by G/uv the graph obtained from G by contracting the edge
uv, i.e., making u adjacent to NG({u, v}) and removing v.
2 Upper bound on the number of minimal separators
Measure and Conquer is a technique developed for the analysis of exponential
time algorithms [7]. Its main idea is to assign a cleverly chosen (sometimes, by
solving mathematical programs [5,13,14]) potential function to the instance -- a
so-called measure -- to track several features of an instance in solving it. While
developed in the realm of exponential-time algorithms, it has also been used to
upper bound the number of extremal vertex sets in graphs (see, e.g., [6,8]).
Our new proof upper bounding sep(n) uses a measure that takes into ac-
count the number of vertices of the graph and the difference in size between the
separated components of the graph. This simple trick allows us to avoid several
complications from [12], including the use of an auxiliary lemma (Lemma 3.1 in
[12]), fixing the size of the separators, the discussion of "full components", and
distinguishing between separators of size at most n/3 and at least n/3.
Theorem 1. sep(n) = O(ρn ·n), where ρ = 1+√5
2 = 1.6180... is the golden ratio.
Proof. Let G = (V, E) be any graph on n vertices with a ∈ V . For d ≤ V , an
[a, d]-separation is a partition (A, S, B) of V such that
-- a ∈ A,
-- G[A] is connected,
-- S is a minimal (a, b)-separator for some b ∈ B, and
-- A ≤ B − d.
Let sepa(G, d) denote the number of [a, d]-separations in G. By symmetry,
sepa(G, 0) upper bounds the number of minimal separators in G up to a fac-
tor O(V ). To upper bound sepa(G, d), we will use the measure
µ(G, d) = V − d.
The theorem will follow from the claim that sepa(G, d) ≤ ρµ(G,d) for 0 ≤ d ≤ V .
If µ(G, d) = 0, then d = V and sepa(G, d) = 0 since there is no A ⊆ V with
A ≤ 0 and a ∈ A. If dG(a) = 0, then there is at most one [a, d]-separation, which
is ({a}, ∅, V \ {a}). Therefore, assume µ(G, d) ≥ 1, a has at least one neighbor,
and assume the claim holds for smaller measures. Consider a vertex u ∈ N (a).
For every [a, d]-separation (A, S, B), either u ∈ S or u ∈ A. Therefore, we can
upper bound the [a, d]-separations (A, S, B) counted in sepa,b(G, d) with u ∈ S
by ρµ(G−{u},d) = ρµ(G,d)−1, and those with u /∈ S by ρµ(G/au,d+1) = ρµ(G,d)−2.
It remains to observe that ρµ(G,d)−1 + ρµ(G,d)−2 = ρµ(G,d).
⊓⊔
4
Serge Gaspers and Simon Mackenzie
3 Lower bound on the maximum number of minimal
separators
In the melon graph in Fig. 1, each horizontal layer implies a choice between
3 vertices. Each of those choices also 'costs' 3 vertices. The new construction
improves the bound by adding a vertical choice on top of the horizontal choice.
This is achieved by 'sacrificing' one of the horizontal choices. This allows us to
chose which layer to sacrifice, at the cost of 6 vertices. If it has more than 3·3 = 9
layers, then this will give a larger range of choices than if we hadn't eliminated
that layer.
Theorem 2. sep(n) ∈ ω(1.4521n).
Proof. We prove the theorem by exhibiting a family of graphs {G1, G2, . . . } and
lower bounding their number of minimal separators.
Let I = {1, . . . , 6} and J = {1, . . . , 24}. The graph G1 is constructed as
follows (see Fig. 2). It has vertex set V = {a, b} ∪ {vi,j : i ∈ I, j ∈ J}. We
denote by Vi the vertex set {vi,j : j ∈ J}. The edge set E of G1 is obtained by
first adding the paths (a, v1,j, v2,j, v3,j) and (v4,j, v5,j, v6,j, b) for all j ∈ J, and
then adding the edges {v3,jv4,k : j, k ∈ J and j 6= k}. The graph Gℓ, ℓ ≥ 2, is
obtained from ℓ disjoint copies of G1, merging the copies of a, and merging the
copies of b.
Let us now lower bound the minimal (a, b)-separators Sj in G1 that do not
contain any vertex from {v1,j, v2,j, v3,j, v4,j, v5,j, v6,j} for some j ∈ J. Each
such separator contains a vertex from {v1,k, v2,k, v3,k}, for k ∈ K \ {j}, since
(a, v1,k, v2,k, v3,k, v4,j, v5,j, v6,j, b) is a path in G1, and it contains a vertex from
{v4,k, v5,k, v6,k}, for k ∈ K \{j}, since (a, v1,j, v2,j, v3,j, v4,k, v5,k, v6,k, b) is a path
in G1. Due to minimality, the separators in Sj contain no other vertices. Thus,
we have that Sj = 32·(J−1). We also note that Sj ∩ Sk = ∅ if j 6= k. Therefore,
the number of minimal separators of G1 is at least4 J · 32·(J−1) > 2.1271 · 1023.
Minimal (a, b)-separators for Gℓ are obtained by taking the union of minimal
separators for the different copies of G1. Their number is therefore at least
6·J ∈ ω(1.4521n), where n = ℓ · 6 · J + 2 is
(J · 32·(J−1))ℓ = (J · 32·(J−1))
⊓⊔
the number of vertices of Gℓ.
n−2
Based on results from [2], Bouchitt´e and Todinca [3] observed that the num-
ber of potential maximal cliques in a graph is at least the number of minimal
separators divided by the number of vertices n. Therefore, we arrive at the fol-
lowing corollary of Theorem 2.
Corollary 1. pmc(n) ∈ ω(1.4521n).
4 There are also minimal (a, b)-separators that are completely contained in V1 ∪ V2 ∪ V3
or V4 ∪ V5 ∪ V6, but their number does not affect our bound in the first 10 decimal
digits in the base of the exponent.
On the Number of Minimal Separators in Graphs
5
a
b
...
...
Fig. 2. The graph G1 has 24 horizontal layers; only 4 are depicted.
4 Conclusion
We have given a simpler proof for the best known asymptotic upper bound on
sep(n), and we have improved the best known lower bound from Ω(3n/3) to
ω(1.4521n), thereby reducing the gap between the current best lower and upper
bound. Before our work, it seemed reasonable to believe that sep(n) could be
asymptotically equal to the best known lower bound. We showed that this is not
the case, and we believe there is room to further improve the lower bound.
Acknowledgments
NICTA is funded by the Australian Government through the Department of
Communications and the Australian Research Council through the ICT Cen-
tre of Excellence Program. Serge Gaspers is the recipient of an Australian
Research Council Discovery Early Career Researcher Award (project number
DE120101761) and a Future Fellowship (project number FT140100048).
References
1. Anne Berry, Jean Paul Bordat, and Olivier Cogis. Generating all the minimal
separators of a graph. International Journal of Foundations of Computer Science,
11(3):397 -- 403, 2000.
2. Vincent Bouchitt´e and Ioan Todinca. Treewidth and minimum fill-in: grouping the
minimal separators. SIAM Journal on Computing, 31:212 -- 232, 2001.
3. Vincent Bouchitt´e and Ioan Todinca. Listing all potential maximal cliques of a
graph. Theoretical Computer Science, 276(1-2):17 -- 32, 2002.
4. Reinhard Diestel. Graph Theory. Springer, 2010.
5. David Eppstein. Quasiconvex analysis of multivariate recurrence equations for
backtracking algorithms. ACM Transactions on Algorithms, 2(4):492 -- 509, 2006.
6. Fedor V. Fomin, Serge Gaspers, Artem V. Pyatkin, and Igor Razgon. On the
minimum feedback vertex set problem: Exact and enumeration algorithms. Algo-
rithmica, 52(2):293 -- 307, 2008.
7. Fedor V. Fomin, Fabrizio Grandoni, and Dieter Kratsch. A measure & conquer
approach for the analysis of exact algorithms. Journal of the ACM, 56(5):1 -- 32,
2009.
6
Serge Gaspers and Simon Mackenzie
8. Fedor V. Fomin, Fabrizio Grandoni, Artem V. Pyatkin, and Alexey A. Stepanov.
Combinatorial bounds via measure and conquer: Bounding minimal dominating
sets and applications. ACM Transactions on Algorithms, 5(1):1 -- 17, 2008.
9. Fedor V. Fomin and Dieter Kratsch. Exact Exponential Algorithms. Springer, 2010.
10. Fedor V. Fomin, Dieter Kratsch, Ioan Todinca, and Yngve Villanger. Exact
algorithms for treewidth and minimum fill-in. SIAM Journal on Computing,
38(3):1058 -- 1079, 2008.
11. Fedor V. Fomin and Yngve Villanger. Finding induced subgraphs via minimal
triangulations. In Proceedings of the 27th International Symposium on Theoretical
Aspects of Computer Science (STACS 2010), volume 5 of LIPIcs, pages 383 -- 394.
Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2010.
12. Fedor V. Fomin and Yngve Villanger. Treewidth computation and extremal com-
binatorics. Combinatorica, 32(3):289 -- 308, 2012.
13. Serge Gaspers. Exponential Time Algorithms: Structures, Measures, and Bounds.
VDM Verlag Dr. Mueller e.K., 2010.
14. Serge Gaspers and Gregory B. Sorkin. A universally fastest algorithm for Max
2-Sat, Max 2-CSP, and everything in between. Journal of Computer and System
Sciences, 78(1):305 -- 335, 2012.
15. Pinar Heggernes. Minimal triangulations of graphs: A survey. Discrete Mathemat-
ics, 306(3):297 -- 317, 2006.
16. Ton Kloks and Dieter Kratsch. Listing all minimal separators of a graph. SIAM
Journal on Computing, 27(3):605 -- 613, 1998.
17. R. E. Miller and D. E. Muller. A problem of maximum consistent subsets. IBM
Research Report RC-240, J. T. Watson Research Center, Yorktown Heights, NY,
1960.
18. John W. Moon and Leo Moser. On cliques in graphs. Israel Journal of Mathemat-
ics, 3:23 -- 28, 1965.
19. Andreas Parra and Petra Scheffler. Characterizations and algorithmic applica-
tions of chordal graph embeddings. Discrete Applied Mathematics, 79(1-3):171 --
188, 1997.
20. Hong Shen and Weifa Liang. Efficient enumeration of all minimal separators in a
graph. Theoretical Computer Science, 180(1-2):169 -- 180, 1997.
21. Yngve Villanger.
Improved exponential-time algorithms for treewidth and min-
imum fill-in.
In Proceedings of the 7th Latin American Theoretical Informatics
Symposium (LATIN 2006), volume 3887 of Lecture Notes in Computer Science,
pages 800 -- 811. Springer-Verlag, Berlin, 2006.
|
1808.08419 | 3 | 1808 | 2018-11-05T17:04:10 | The Complexity of $(\Delta + 1)$Coloring inCongested Clique, Massively Parallel Computation,and Centralized Local Computation | [
"cs.DS"
] | We present new randomized algorithms that improve the complexity of the classic $(\Delta+1)$-coloring problem, and its generalization $(\Delta+1)$-list-coloring, in three well-studied models of distributed, parallel, and centralized computation:
Distributed Congested Clique: We present an $O(1)$-round randomized algorithm for $(\Delta+1)$-list coloring in the congested clique model of distributed computing. This settles the asymptotic complexity of this problem. It moreover improves upon the $O(\log^\ast \Delta)$-round randomized algorithms of Parter and Su [DISC'18] and $O((\log\log \Delta)\cdot \log^\ast \Delta)$-round randomized algorithm of Parter [ICALP'18].
Massively Parallel Computation: We present a $(\Delta+1)$-list coloring algorithm with round complexity $O(\sqrt{\log\log n})$ in the Massively Parallel Computation (MPC) model with strongly sublinear memory per machine. This algorithm uses a memory of $O(n^{\alpha})$ per machine, for any desirable constant $\alpha>0$, and a total memory of $\widetilde{O}(m)$, where $m$ is the size of the graph. Notably, this is the first coloring algorithm with sublogarithmic round complexity, in the sublinear memory regime of MPC. For the quasilinear memory regime of MPC, an $O(1)$-round algorithm was given very recently by Assadi et al. [SODA'19].
Centralized Local Computation: We show that $(\Delta+1)$-list coloring can be solved with $\Delta^{O(1)} \cdot O(\log n)$ query complexity, in the centralized local computation model. The previous state-of-the-art for $(\Delta+1)$-list coloring in the centralized local computation model are based on simulation of known LOCAL algorithms. | cs.DS | cs |
The Complexity of (∆ + 1) Coloring in
Congested Clique, Massively Parallel Computation,
and Centralized Local Computation
Yi-Jun Chang
U. Michigan
Manuela Fischer
Mohsen Ghaffari
ETH Zurich
ETH Zurich
[email protected]
[email protected]
[email protected]
Jara Uitto
ETH Zurich & U. Freiburg
Yufan Zheng
U. Michigan
[email protected]
[email protected]
Abstract
In this paper, we present new randomized algorithms that improve the complexity of the
classic (∆+1)-coloring problem, and its generalization (∆+1)-list-coloring, in three well-studied
models of distributed, parallel, and centralized computation:
Distributed Congested Clique: We present an O(1)-round randomized algorithm for (∆ +
1)-list coloring in the congested clique model of distributed computing. This settles the
asymptotic complexity of this problem. It moreover improves upon the O(log∗ ∆)-round
randomized algorithms of Parter and Su [DISC'18] and O((log log ∆) · log∗ ∆)-round ran-
domized algorithm of Parter [ICALP'18].
Massively Parallel Computation: We present a (∆ + 1)-list coloring algorithm with round
complexity O(√log log n) in the Massively Parallel Computation (MPC) model with strongly
sublinear memory per machine. This algorithm uses a memory of O(nα) per machine, for
any desirable constant α > 0, and a total memory of eO(m), where m is the size of the
graph. Notably, this is the first coloring algorithm with sublogarithmic round complexity,
in the sublinear memory regime of MPC. For the quasilinear memory regime of MPC, an
O(1)-round algorithm was given very recently by Assadi et al. [SODA'19].
Centralized Local Computation: We show that (∆ + 1)-list coloring can be solved with
∆O(1)·O(log n) query complexity, in the centralized local computation model. The previous
state-of-the-art for (∆ + 1)-list coloring in the centralized local computation model are
based on simulation of known LOCAL algorithms. The deterministic O(√∆poly log ∆ +
log∗ n)-round LOCAL algorithm of Fraigniaud et al. [FOCS'16] can be implemented in the
centralized local computation model with query complexity ∆O(√∆poly log ∆)·O(log∗ n); the
randomized O(log∗ ∆) + 2O(√log log n)-round LOCAL algorithm of Chang et al. [STOC'18]
can be implemented in the centralized local computation model with query complexity
∆O(log∗ ∆) · O(log n).
1 Introduction, Related Work, and Our Results
In this paper, we present improved randomized algorithms for vertex coloring in three models
of distributed, parallel, and centralized computation: the congested clique model of distributed
computing, the massively parallel computation model, and the centralized local computation model.
We next overview these results in three different subsections, while putting them in the context of
the state of the art. The next section provides a technical overview of the known algorithmic tools
as well as the novel ingredients that lead to our results.
(∆ + 1)-coloring and (∆ + 1)-list Coloring. Our focus is on the standard ∆ + 1 vertex coloring
problem, where ∆ denotes the maximum degree in the graph. All our results work for the gener-
alization of the problem to (∆ + 1)-list coloring problem, defined as follows: each vertex v in the
graph G = (V, E) is initially equipped with a set of colors Ψ(v) such that Ψ(v) = ∆ + 1. The goal
is to find a proper vertex coloring where each vertex v ∈ V is assigned a color in Ψ(v) such that no
two adjacent vertices are colored the same.
1.1 Congested Clique Model of Distributed Computing
Models of Distributed Computation. There are three major models for distributed graph
algorithms, namely LOCAL, CONGEST, and CONGESTED-CLIQUE. In the LOCAL model [Lin92,
Pel00], the input graph G = (V, E) is identical to the communication network and each v ∈ V hosts
a processor that initially knows deg(v), a unique Θ(log n)-bit ID(v), and global graph parameters
n = V and ∆ = maxv∈V deg(v). Each processor is allowed unbounded computation and has access
to a stream of private random bits. Time is partitioned into synchronized rounds of communication,
in which each processor sends one unbounded message to each neighbor. At the end of the algorithm,
each v declares its output label, e.g., its own color. The CONGEST model [Pel00] is a variant of
LOCAL where there is an O(log n)-bit message size constraint. The CONGESTED-CLIQUE model,
introduced in [LPSPP05], is a variant of CONGEST that allows all-to-all communication: Each
vertex initially knows its adjacent edges of the input graph G = (V, E). In each round, each vertex
is allowed to transmit n − 1 many O(log n)-bit messages, one addressed to each other vertex.
model. It is worth noting that the CONGESTED-CLIQUE model has been receiving extensive atten-
tion recently, see e.g., [PST11, DLP12, BHP12, Len13, DKO14, Nan14, HPS14, HP15, CHKK+16,
HPP+15, BKKL17, Gal16, CLT18, Gha16, Gha17, GGK+18, PS18, Par18, BK18].
In this paper, our new distributed result is an improvement for coloring in the CONGESTED-CLIQUE
State of the Art for Coloring in LOCAL and CONGEST. Most prior works on distributed
coloring focus on the LOCAL model. The current state-of-the-art randomized upper bound for the
(∆ + 1)-list coloring problem is O(log∗ ∆) + O(Detd(poly log n)) = O(Detd(poly log n)) of [CLP18]
(which builds upon the techniques of [HSS18]), where Detd(n′) = 2O(√log log n′) is the determinis-
tic complexity of (deg +1)-list coloring on n′-vertex graphs [PS96].
In the (deg +1)-list coloring
problem, each v has a palette of size deg(v) + 1. This algorithm follows the graph shattering frame-
work [BEPS16, Gha16]. The pre-shattering phase takes O(log∗ ∆) rounds. After that, the remaining
uncolored vertices form connected components of size O(poly log n). The post-shattering phase then
applies a (deg +1)-list coloring deterministic algorithm to color all these vertices.
State of the Art for Coloring in CONGESTED-CLIQUE. Hegeman and Pemmaraju [HP15]
gave algorithms for O(∆)-coloring in the CONGESTED-CLIQUE model, which run in O(1) rounds
if ∆ ≥ Θ(log4 n) and in O(log log n) rounds otherwise. It is worth noting that O(∆) coloring is
1
a significantly more relaxed problem in comparison to ∆ + 1 coloring. For instance, we have long
known a very simple O(∆)-coloring algorithm in LOCAL-model algorithm with round complexity
2O(√log log n) [BEPS16], but only recently such a round complexity was achieved for ∆ + 1 coloring
[CLP18, HSS18].
Our focus is on the much more stringent ∆ + 1 coloring problem. For this problem, the LOCAL
model algorithms of [CLP18, HSS18] need messages of O(∆2 log n) bits, and thus do not extend
to CONGEST or CONGESTED-CLIQUE. For CONGESTED-CLIQUE model, the main challenge is
when ∆ > √n, as otherwise, one can simulate the algorithm of [CLP18] by leveraging the all-to-
all communication in CONGESTED-CLIQUE which means each vertex in each round is capable of
communicating O(n log n) bits of information. Parter [Par18] designed the first sublogarithmic-time
(∆+1) coloring algorithm for CONGESTED-CLIQUE, which runs in O(log log ∆ log∗ ∆) rounds. The
algorithm of [Par18] is able to reduce the maximum degree to O(√n) in O(log log ∆) iterations, and
each iteration invokes the algorithm of [CLP18] on instances of maximum degree O(√n). Once
the maximum degree is O(√n), the algorithm of [CLP18] can be implemented in O(log∗ ∆) rounds
in CONGESTED-CLIQUE. Subsequent to [Par18], the upper bound was improved to O(log∗ ∆)
in [PS18]. Parter and Su [PS18] observed that the algorithm of [Par18] only takes O(1) iterations
if we only need to reduce the degree to n1/2+ǫ, for some constant ǫ > 0, and they achieved this by
modifying the internal details of [CLP18] to reduce the required message size to O(∆8/5 log n).
Our Result. For the CONGESTED-CLIQUE model, we present a new algorithm for (∆ + 1)-list
coloring in the randomized congested clique model running in O(1) rounds. This improving on the
previous best known O(log∗ ∆)-round algorithm of Parter and Su [PS18] and settles the asymptotic
complexity of the problem.
Theorem 1.1. There is an O(1)-round algorithm that solves the (∆ + 1)-list coloring problem in
CONGESTED-CLIQUE, with success probability 1 − 1/poly(n).
The proof is presented in two parts: If ∆ ≥ log4.1 n, the algorithm of Theorem 3.2 solves the
(∆ + 1)-list coloring problem in O(1) rounds; otherwise, the algorithm of Theorem 4.7 solves the
problem in O(1) rounds.
1.2 Massively Parallel Computation
Model. The Massively Parallel Computation (MPC) model was introduced by Karloff et al. [KSV10],
as a theoretical abstraction for practical large-scale parallel processing settings such as MapRe-
duce [DG04], Hadoop [Whi12], Spark [ZCF+10], and Dryad [IBY+07], and it has been receiving
increasing more attention over the past few years [KSV10, GSZ11, LMSV11, BKS13, ANOY14,
BKS14, HP15, AG15, RVW16, IMS17, CLM+18, Ass17, ABB+19, GGK+18, HLL18, BFU18a,
ASW18, BEG+18b, ASS+18]. In the MPC model, the system consists of a number of machines,
each with S bits of memory, which can communicate with each other in synchronous rounds through
a complete communication network. Per round, each machine can send or receive at most S bits
in total. Moreover, it can perform some poly(S) computation, given the information that it has.
In the case of graph problems, we assume that the graph G is partitioned among the machines
using a simple and globally known hash function such that each machine holds at most S bits,
and moreover, for each vertex or potential edge of the graph, the hash function determines which
machines hold that vertex or edge. Thus, the number of machines is Ω(m/S) and ideally not too
much higher, where m denotes the number of edges. At the end, each machine should know the
output of the vertices that it holds, e.g., their color.
2
State of the Art for Coloring. The CONGESTED-CLIQUE algorithms discussed above can
be used to obtain MPC algorithms with the same asymptotic round complexity if machines have
memory of S = Ω(n log n) bits.
In particular, the work of Parter and Su [PS18] leads to an
O(log∗ ∆)-round MPC algorithm for machines with S = Ω(n log n) bits. However, this MPC algo-
rithm would have two drawbacks: (A) it uses Ω(n2 log n) global memory, and thus would require
(n2 log n)/S machines, which may be significantly larger than O(m)/S. This is basically because
the algorithm makes each vertex of the graph learn some eΘ(n) bits of information. (B) It is limited
to machines with S = Ω(n log n) memory, and it does not extend to the machines with strongly
sublinear memory, which is gaining more attention recently due to the increase in the size of graphs.
We note that for the regime of machines with super-linear memory, very recently, Assadi, Chen,
and Khanna [ACK19] gave an O(1)-round algorithm which uses only O(n log3 n) global memory.1
However, this algorithm also relies heavily on S = Ω(n log3 n) memory per machine and cannot be
run with weaker machines that have strongly sublinear memory.
Our Result. We provide the first sublogarithmic-time algorithm for (∆ + 1) coloring and (∆ + 1)-
list coloring in the MPC model with strongly sublinear memory per machine:
Theorem 1.2. There is an MPC algorithm that, in O(log∗ ∆ +√log log n) = O(√log log n) rounds,
w.h.p. computes a (∆ + 1) list-coloring of an n-vertex graph with m edges and maximum degree
∆ and that uses O(nα) memory per machine, for an arbitrary constant α > 0, as well as a total
memory of eO(m).
The proof is presented in Section 3.3.
1.3 Centralized LOCAL Computation
Model. This Local Computation Algorithms (LCA) model is a centralized model of computation
that was introduced in [RTVX11]; an algorithm in this model is usually called an LCA. In this
model, there is a graph G = (V, E) where the algorithm is allowed to make the following queries:
Degree Query: Given ID(v), the oracle returns deg(v).
Neighbor Query: Given ID(v) and an index i ∈ [1, ∆], if deg(v) ≤ i, the oracle returns ID(u),
where u is the ith neighbor of v; otherwise, the oracle returns ⊥.
It is sometimes convenient to assume that there is a query that returns the list of all neighbors of v.
This query can be implemented using one degree query and deg(v) neighbor queries. For randomized
algorithms, we assume that there is an oracle that given ID(v) returns an infinite-length random
sequence associated with the vertex v. Similarly, for problems with input labels (e.g., the color lists
in the list coloring problem), the input label of a vertex v can be accessed given ID(v). Given a
distributed problem P, an LCA A accomplishes the following. Given ID(v), the algorithm A returns
A(v) = the output of v, after making a small number of queries. It is required that the output of
A at different vertices are consistent with one legal solution of P.
The complexity measure for an LCA is the number of queries. It is well-known [PR07] that any
τ -round LOCAL algorithm A can be transformed into an LCA A′ with query complexity ∆τ . The
LCA A′ simply simulates the LOCAL algorithm A by querying all radius-τ neighborhood of the given
vertex v. See [LM17] for a recent survey about the state-of-the-art in the centralized local model.
1Here "global memory" refers to the memory used for communication. Of course we still need O(m) memory to
store the graph.
3
State of the Art LCA for Coloring The previous state-of-the-art for (∆ + 1)-list coloring in
the centralized local computation model are based on simulation of known LOCAL algorithms.
The deterministic O(√∆poly log ∆ + log∗ n)-round LOCAL algorithm of [FHK16, BEG18a]2 can be
implemented in the centralized local computation model with query complexity ∆O(√∆poly log ∆) ·
O(log∗ n); the randomized O(log∗ ∆) + 2O(√log log n)-round LOCAL algorithm of [CLP18] can be
implemented in the centralized local computation model with query complexity ∆O(log∗ ∆)·O(log n).
Our Result. We show that (∆ + 1)-list coloring can be solved with ∆O(1) · O(log n) query com-
plexity. Note that ∆O(1) · O(log n) matches a "natural barrier" for randomized algorithms based on
the graph shattering framework, as each connected component in the post-shattering phase has this
size ∆O(1) · O(log n).
Theorem 1.3. There is an centralized local computation algorithm that solves the (∆ + 1)-list
coloring problem with query complexity ∆O(1) · O(log n), with success probability 1 − 1/poly(n).
The proof is presented in Section 4.3.
2 Technical Overview: Tools and New Ingredients
In this section, we first review some of the known technical tools that we will use in our algorithms,
and then we overview the two new technical ingredients that lead to our improved results (in
combination with the known tools).
Notes and Notations. When talking about randomized algorithms, we require the algorithm to
succeed with high probability (w.h.p.), i.e., to have success probability at least 1 − 1/poly(n). For
each vertex v, we write N (v) to denote the set of neighbors of v. If there is an edge orientation,
N out(v) refers to the set of out-neighbors of v. We write N k(v) = {u ∈ V dist(u, v) ≤ k}. We
use subscript to indicate the graph G under consideration, e.g., NG(v) or N out
G (v). In the course of
our algorithms, we slightly abuse the notation to also use Ψ(v) to denote the set of available colors
of v. i.e., the subset of Ψ(v) that excludes the colors already taken by its neighbors in N (v). The
number of excess colors at a vertex is the number of available colors minus the number of uncolored
neighbors. Moreover, we make an assumption that each color can be represented using O(log n)
bits. This is without loss of generality (in all of the models under consideration in our paper),
since otherwise we can hash the colors down to this magnitude, as we allow a failure probability of
1/poly(n) for randomized algorithms.
2.1 Tools
Lenzen's Routing. The routing algorithm of Lenzen [Len13] for CONGESTED-CLIQUE allows
us to deliver all messages in O(1) rounds, as long as each vertex v is the source and the destination
of at most O(n) messages. This is a very useful (and frequently used) communication primitive for
designing CONGESTED-CLIQUE algorithms.
Lemma 2.1 (Lenzen's Routing). Consider a graph G = (V, E) and a set of point-to-point routing
requests, each given by the IDs of the corresponding source-destination pair. As long as each vertex
v is the source and the destination of at most O(n) messages, namely O(n log n) bits of information,
we can deliver all messages in O(1) rounds in the CONGESTED-CLIQUE model.
2Precisely, the complexity is O(√∆ log2.5 ∆ + log∗ n) in [FHK16], and this has been later improved to
O(√∆ log ∆ log∗ ∆ + log∗ n) in [BEG18a].
4
The Shattering Framework. Our algorithm follows the graph shattering framework [BEPS16],
which first performs some randomized process (known as pre-shattering) to solve "most" of the prob-
lem, and then performs some clean-up steps (known as post-shattering) to solve the remaining part
of the problem. Typically, the remaining graph is simpler in the sense of having small components
and having a small number of edges. Roughly speaking, at each step of the algorithm, we specify
an invariant that all vertices must satisfy in order to continue to participate. Those bad vertices
that violate the invariant are removed from consideration, and postponed to the post-shattering
phase. We argue that the bad vertices form connected components of size ∆O(1) · O(log n) with
probability 1 − 1/poly(n); we use this in designing LCA. Also, the total number of edges induced
by the bad vertices is O(n). Therefore, using Lenzen's routing, in CONGESTED-CLIQUE we can
gather all information about the bad vertices to one distinguished vertex v⋆, and then v⋆ can color
them locally. More precisely, we have the following lemma [BEPS16, FG17]; see Appendix B for
the proof.
Lemma 2.2 (The Shattering Lemma). Let c ≥ 1. Consider a randomized procedure that generates
a subset of vertices B ⊆ V . Suppose that for each v ∈ V , we have Pr[v ∈ B] ≤ ∆−3c, and this holds
even if the random bits not in N c(v) are determined adversarially. Then, the following is true.
1. With probability 1 − n−Ω(c′), each connected component in the graph induced by B has size at
most (c′/c)∆2c log∆ n.
2. With probability 1 − O(∆c) · exp(−Ω(n∆−c)), the number of edges induced by B is O(n).
Round Compression in CONGESTED-CLIQUE and MPC by Information Gathering. Sup-
pose we are given a τ -round LOCAL algorithm A on a graph of maximum degree ∆. A direct
simulation of A on CONGESTED-CLIQUE costs also τ rounds. However, if each vertex v already
knows all information in its radius-τ neighborhood, then v can locally compute its output in zero
rounds. In general, this amount of information can be as high as Θ(n2), since there could be Θ(n2)
edges in the radius-τ neighborhood of v. For the case of ∆τ = O(n), it is possible to achieve
an exponential speed-up in the round complexity in the CONGESTED-CLIQUE, compared to that
of LOCAL. In particular, in this case, each vertex v can learn its radius-τ neighborhood in just
O(log τ ) rounds in CONGESTED-CLIQUE. Roughly speaking, after k rounds, we are able to sim-
ulate the product graph G2k
, which is the graph where any two vertices with distance at most
2k in graph G are adjacent. This method is known as graph exponentiation [LW10], and it has
been applied before in the design of algorithms in CONGESTED-CLIQUE and MPC models, see e.g.,
[Gha17, GU19, PS18, Par18, ASS+18].
Round Compression via Opportunistic Information Gathering. Our goal is to achieve
the O(1) round complexity in CONGESTED-CLIQUE, so an exponential speed-up compared to the
LOCAL model will not be enough. Consider the following "opportunisitc" way of simulating a LOCAL
algorithm A in the CONGESTED-CLIQUE model. Each vertex u sends its local information (which
has O(∆ log n) bits) to each vertex v ∈ V with some fixed probability p = O(1/∆), independently,
and it hopes that there exists a vertex v ∈ V that gathers all the required information to calculate
the outcome of A at u. To ensure that for each u, there exists such a vertex v w.h.p., it suffices
that p∆τ
n . We note that a somewhat similar idea was key to the O(1)-round MST algorithm
of [JN18] for CONGESTED-CLIQUE.
≫ log n
Lemma 2.3, presented below, summarizes the criteria for this method to work; see Appendix C
for the proof of the lemma. Denote ℓin as the number of bits needed to represent the random bits
and the input for executing A at a vertex. Denote ℓout as the number of bits needed to represent
5
the output of A at a vertex. We assume that each vertex v initially knows a set N∗(v) ⊆ N (v) such
that throughout the algorithm A, each vertex v only receives information from vertices in N∗(v).
We write ∆∗ = maxv∈v N∗(v). Note that it is possible that u ∈ N∗(v) but v /∈ N∗(u). In this case,
during the execution of A, all messages sent via the edge {u, v} are from u to v. Denote N k
∗ (v) as
the set of all vertices u such that there is a path (v = w0, w1, . . . , wx−1 = u) such that x ≤ k and
wi ∈ N∗(wi−1) for each i ∈ [1, x − 1]. Intuitively, if A takes τ rounds, then all information needed
for vertex v ∈ V to calculate its output is the IDs and the inputs of all vertices in N τ
Lemma 2.3 (Opportunistic Speed-up). Let A be a τ -round LOCAL algorithm on G = (V, E).
There is an O(1)-round simulation of A in in CONGESTED-CLIQUE, given that (i) ∆τ
∗ log(∆∗ +
ℓin/ log n) = O(log n), (ii) ℓin = O(n), and (iii) ℓout = O(log n).
∗ (v).
2.2 Our New Technical Ingredients, In a Nutshell
The results in our paper are based on the following two novel technical ingredients, which are used
in combination with the known tools mentioned above: (i) a new graph partitioning algorithm for
coloring and (ii) a sparsification of the CLP coloring algorithm [CLP18]. We note that the first
ingredient suffices for our CONGESTED-CLIQUE result for graphs with maximum degree at least
poly(log n), and also for our MPC result. This ingredient is presented in Section 3. The second
ingredient, which is also more involved technically, is used for extending our CONGESTED-CLIQUE
result to graphs with smaller maximum degree, as well as for our LCA result. This ingredient is
presented in Section 4. Here, we provide a brief overview of these ingredients and how they get used
in our results.
Ingredient 1 -- Graph Partitioning for Coloring. We provide a simple random partitioning
that significantly simplifies and extends the one in [Par18, PS18]. The main change will be that,
besides partitioning the vertices randomly, we also partition the colors randomly. In particular, this
new procedure partitions the vertices and colors in a way that allows us to easily apply CLP in a
black box manner.
Concretely, our partitioning breaks the graph as well as the respective palettes randomly into
many subgraphs B1, . . . , Bk of maximum degree O(√n) and size O(√n), while ensuring that each
vertex in these subgraphs receives a random part of its palette with size close to the maximum
degree of the subgraph. The palettes for each part are disjoint, which allows us to color all parts
in parallel. There will be one left-over subgraph L, with maximum degree O(∆3/4), as well as
sufficiently large remaining palettes for each vertex in this left-over subgraph.
Application in CONGESTED-CLIQUE: Since each subgraph has O(n) edges, all of B1, . . . , Bk can
be colored, in parallel, in O(1) rounds, using Lenzen's routing (Lemma 2.1). The left-over part
L is handled by recursion. We show that when ∆ > log4.1 n, we are done after O(1) levels of
recursion.
Application in Low-memory MPC: We perform recursive calls on not only on L but also on
B1, . . . , Bk. After O(1) levels of recursion, the maximum degree can be made O(nβ), for any
given constant β > 0, which enables us to run the CLP algorithm on a low memory MPC.
We note that the previous partitioning approach [Par18, PS18] is unable to reduce the maximum
degree to below √n; this is a significant limitation that our partitioning overcomes.
6
Ingredient 2 -- Sparsification of the CLP Algorithm.
In general, to calculate the output of a
vertex v in a τ -round LOCAL algorithm A, the output may depend on all of the τ -hop neighborhood
of v and we may need to query ∆τ vertices. To efficiently simulate A in CONGESTED-CLIQUE or to
transform A to an LCA, a strategy is to "sparsify" the algorithm A so that the number of vertices a
vertex has to explore to decide its output is sufficiently small. This notion of sparsification is a key
idea behind some recent algorithms [Gha17, GU19]. In the present paper, a key technical ingredient
is providing such a sparsification for the (∆ + 1) coloring algorithm of CLP [CLP18].
The pre-shattering phase of the CLP algorithm [CLP18] consists of three parts: (i) initial
coloring, (ii) dense coloring, and (iii) color bidding. Parts (i) and (ii) take O(1) rounds;3 part (iii)
takes τ = O(log∗ ∆) rounds. In this paper, we sparsify the color bidding part of the CLP algorithm.
We let each vertex v sample O(poly log ∆) colors from its palette at the beginning of this procedure,
and we show that with probability 1− 1/poly(∆), these colors are enough for v to correctly execute
the algorithm. Based on the sampled colors, we can do an O(1)-round pre-processing step to let
each vertex v identify a subset of neighbors N∗(v) ⊆ N (v) of size ∆∗ = O(poly log ∆) neighbors
N∗(v) ⊆ N (v), and v only needs to receive messages from neighbors in N∗(v) in the subsequent
steps of the algorithm.
Application in CONGESTED-CLIQUE: For the case ∆ = O(poly log n), the parameters τ =
O(log∗ ∆) and ∆∗ = O(poly log ∆) = O(poly(log log n)) satisfy the condition for applying
the opportunistic speedup lemma (Lemma 2.3), and so the pre-shattering phase of the CLP
algorithm can be simulated in O(1) rounds in CONGESTED-CLIQUE.
Application in Centralized Local Computation: With sparsification, the pre-shattering phase
of the CLP algorithm can be transformed into an LCA with ∆O(1) · ∆τ
∗ = ∆O(1) queries.
The recent work [ACK19] on (∆ + 1)-coloring in MPC is also based on some form of palette
sparsification, as follows. They showed that if each vertex samples O(log n) colors uniformly at
random, then w.h.p., the graph still admits a proper coloring using the sampled colors. Since
we only need to consider the edges {u, v} where u and v share a sampled color, this effectively
reduces the degree to O(log2 n). For an MPC algorithm with O(n) memory per processor, the entire
sparsified graph can be sent to one processor, and a coloring can be computed there, using any
coloring algorithm, local or not. This sparsification is not applicable for our setting. In particular,
in our sparsified CLP algorithm, we need to ensure that the coloring can be computed by a LOCAL
algorithm with a small locality volume; this is because the final coloring is constructed distributedly
via the opportunistic speedup lemma (Lemma 2.3).
3 Coloring of High-degree Graphs via Graph Partitioning
In this section, we describe our graph partitioning algorithm, which is the first new technical
ingredient in our results. As mentioned in Section 2.2, this ingredient on its own leads to our
CONGESTED-CLIQUE result for graphs with ∆ = Ω(poly(log n)) and also our MPC result, as we
will explain in Section 3.2 and Section 3.3, respectively. The algorithm will be applied recursively,
but it is required that the failure probability is at most 1− 1/poly(n) in all recursive calls, where n
is the number of vertices in the original graph. Thus, in this section, n does not refer to the number
of vertices in the current subgraph G = (V, E) under consideration.
3In the preliminary versions (arXiv:1711.01361v1 and STOC'18) of [CLP18], dense coloring takes O(log∗ ∆) time.
This time complexity has been later improved to O(1) in a revised full version of [CLP18] (arXiv:1711.01361v2).
7
3.1 Graph Partitioning
The Graph Partitioning Algorithm. The graph partitioning is parameterized by two constants
γ and λ satisfying γ ≥ 2 and λ = 1
3γ+2 . Consider a graph G = (V, E) with maximum degree
∆. Note that G is a subgraph of the n-vertex original graph, and so n ≥ V . Each vertex v ∈ V
has a palette Ψ(v) of size Ψ(v) ≥ max{degG(v), ∆′} + 1, where ∆′ = ∆ − ∆λ. Denote G[S] as the
subgraph induced by the vertices S ⊆ V . For each vertex v ∈ V , denote degS(v) as N (v)∩ S. The
algorithm is as follows, where we set k = √∆.
2 + 2
Vertex Set: The partition V = B1 ∪ ··· ∪ Bk ∪ L is defined by the following procedure. Including
each v ∈ V to the set L with probability q = Θ(cid:16)q log n
∆1/4(cid:17). Each remaining vertex joins one of
B1, . . . , Bk uniformly at random. Note that Pr[v ∈ Bi] = p(1 − q), where p = 1/k = 1/√∆.
Palette: Denote C = Sv∈V Ψ(v) as the set of all colors. The partition C = C1 ∪ ··· ∪ Ck is
defined by having each color c ∈ C joins one of C1, . . . , Ck uniformly at random. Note that
Pr[c ∈ Ci] = p.
We require that with probability 1− 1/poly(n), the output of the partitioning algorithm satisfies
the following properties, assuming that ∆ = ω(logγ n).
i) Size of Each Part: It is required that E(G[Bi]) = O(V ), for each i ∈ [k]. Also, it is required
that L = O(qV ) = O(
√log n
∆1/4 ) · V .
ii) Available Colors in Bi: For each i ∈ {1, . . . , k} and v ∈ Bi, the number of available colors in
v in the subgraph Bi is gi(v) := Ψ(v) ∩ Ci. It is required that gi(v) ≥ max{degBi(v), ∆i −
∆λ
i } + 1, where ∆i := maxv∈Bi degBi(v).
iii) Available Colors in L: For each v ∈ L, define gL(v) := Ψ(v)− (degG(v)− degL(v)). It is re-
quired that gL(v) ≥ max{degL(v), ∆L−∆λ
L}+1 for each v ∈ L, where ∆L := maxv∈L degL(v).
Note that gL(v) represents a lower bound on the number of available color in v after all of
B1, . . . , Bk have been colored.
iv) Remaining Degrees: The maximum degrees of Bi and L are degBi(v) ≤ ∆i = O(√∆) and
√log n
∆1/4 ) · ∆. For each vertex individually, we have degBi(v) ≤
degL(v) ≤ ∆L = O(q∆) = O(
max{O(log n), O(1/√∆) · deg(v)} and degL(v) ≤ max{O(log n), O(q) · deg(v)}.
Intuitively, we will use this graph partitioning in the following way. First compute the de-
composition of the vertex set and the palette, and then color each Bi using colors in Ci. Since
E(G[Bi]) = O(V ) = O(n), in the CONGESTED-CLIQUE model we are able to send the entire
graph G[Bi] to a single distinguished vertex v⋆
i can compute a proper coloring of G[Bi]
locally. This procedure can be done in parallel for all i. If E(G[L]) = O(n), then similarly we
can let a vertex to compute a proper coloring of G[L]; otherwise we apply the graph partitioning
recursively on G[L], with the same parameter n.
Lemma 3.1. Suppose Ψ(v) ≥ max{degG(v), ∆′} + 1 with ∆′ = ∆− ∆λ, and V > ∆ = ω(logγ n),
where γ and λ are two constants satisfying γ ≥ 2 and λ = 1
3γ+2 . The two partitions V =
B1 ∪···∪ Bk ∪ L and C =Sv∈V Ψ(v) = C1 ∪···∪ Ck satisfy the required properties, with probability
1 − 1/poly(n).
i , and then v⋆
2 + 2
8
Proof. We prove that the properties i), ii), iii), and iv) hold with high probability. Note that for
some of the bounds, it is straightforward to observe that they hold in expectation.
i) Size of Each Part: We first show that E(G[Bi]) = O(V ), for each i ∈ [k], with probability
1 − 1/poly(n). To have E(G[Bi]) = O(V ), it suffices to have degBi(v) = O(p∆) for each v, and
Bi = O(pV ), since p = 1/√∆. Recall that we already have E[degBi(v)] ≤ (1 − q)p∆ < p∆ and
E[Bi] = (1 − q)pV < pV , so we only need to show that these parameters concentrate at their
expected values with high probability. This can be established by a Chernoff bound, as follows.
Note that we have ǫ1 < 1 and ǫ2 < 1. In particular, the inequality ǫ1 < 1 holds because of the
assumption ∆ = ω(logγ n) ≥ ω(log2 n).
Pr[degBi(v) ≤ (1 + ǫ1)(1 − q)p∆] = 1 − exp(−Ω(ǫ2
where ǫ1 = Θ s log n
(1 − q)p∆! = Θ s log n
p∆ ! .
1(1 − q)p∆)) = 1 − O(1/poly(n)),
2(1 − q)pV )) = 1 − O(1/poly(n)),
Pr[Bi ≤ (1 + ǫ2)(1 − q)pV ] = 1 − exp(−Ω(ǫ2
where ǫ2 = Θ s log n
∆1/4 (cid:17), where ∆L = maxv∈L degL(v). Similarly, we already have
(1 − q)pV ! = Θ s log n
pV ! .
Next, we show the analogous results for L, i.e., with probability 1 − 1/poly(n), both L/V
√log n
E[degL(v)] ≤ q∆ and E[L] = qV , and remember that q = O(
∆1/4 ), so we only need to show that
these parameters concentrate at their expected values with high probability, by a Chernoff bound.
and ∆L/∆ are O(q) = O(cid:16)√log n
Pr[degL(v) ≤ (1 + ǫ3)q∆] = 1 − exp(−Ω(ǫ2
where ǫ3 = Θ s log n
q∆ ! .
3q∆)) = 1 − O(1/poly(n)),
Pr[L ≤ (1 + ǫ4)qV ] = 1 − exp(−Ω(ǫ2
where ǫ4 = Θ slog n
qV ! .
4qV )) = 1 − O(1/poly(n)),
Similarly, we have ǫ3 < 1 and ǫ4 < 1. In particular, ǫ3 < 1 because ∆ = ω(logγ n) ≥ ω(log2 n).
ii) Available Colors in Bi: Now we analyze the number of available color for each set Bi. Recall
that for each v ∈ Bi, the number of available colors in v in the subgraph Bi is gi(v) := Ψ(v) ∩ Ci.
We need to prove the following holds with probability 1− 1/poly(n): (i) Ψ(v)∩ Ci ≥ degBi(v) + 1,
and (ii) Ψ(v)∩Ci ≥ ∆i−∆λ
i +1, where ∆i := maxv∈Bi degBi(v). We will show that with probability
1− 1/poly(n), we have Ψ(v)∩ Ci ≥ ∆i + 1 for each Bi and each v ∈ Bi, and this implies the above
(i) and (ii).
Recall that ∆′ = ∆(cid:0)1 − ∆−(1−λ)(cid:1), q = Θ(cid:16)√log n
selecting q ≥ 3ǫ1 = Θ(cid:16)√log n
∆1/4 (cid:17), we have
∆1/4 (cid:17) ≫ ∆−(1−λ),4 and ǫ1 = Θ(cid:16)√log n
∆1/4 (cid:17). By
(1 − ǫ1)p∆′ = (1 − ǫ1)(cid:16)1 − ∆−(1−λ)(cid:17) p∆ ≥ (1 + ǫ1)(1 − q)p∆ + 1.
4The assumptions γ ≥ 2 and λ = 1
2 + 2
3γ+2 imply that λ ∈ (1/2, 3/4], and so ∆−(1−λ) ≤ ∆−1/4 ≪ q.
9
We already know that ∆i ≤ (1 + ǫ1)(1 − q)p∆ with probability 1 − 1/poly(n). In order to have
Ψ(v) ∩ Ci ≥ ∆i + 1, we only need to show that Ψ(v) ∩ Ci ≤ (1 − ǫ1)p∆′ with probability
1− 1/poly(n). For the expected value, we know that E[Ψ(v)∩ Ci] = pΨ(v) ≥ p∆′. By a Chernoff
bound, we have
Pr[Ψ(v) ∩ Ci ≤ (1 − ǫ1)p∆′] = 1 − exp(−Ω(ǫ2
1p∆′)) = 1 − O(1/poly(n)).
iii) Available Colors in L: Next, we consider the number of available colors in L. We show that
with probability 1 − 1/poly(n), for each v ∈ L, we have gL(v) ≥ max{degL(v), ∆L − ∆λ
L} + 1,
where gL(v) = Ψ(v) − (degG(v) − degL(v)). It is straightforward to see that gL(v) ≥ degL(v) + 1,
since gL(v) = (Ψ(v) − degG(v)) + degL(v) ≥ 1 + degL(v). Thus, we only need to show that
gL(v) ≥ ∆L − ∆λ
In this proof, without loss of generality we assume degG(v) = Ψ(v) − 1 ≥ ∆′.5 Since
E[degL(v)] = q degG(v) ≥ q∆′, by a Chernoff bound, we have
L + 1.
Pr[degL(v) ≥ (1 − ǫ3)q∆′] = 1 − exp(−Ω(ǫ2
3q∆′)) = 1 − O(1/poly(n))
Remember that ǫ3 = Θ(cid:16)q log n
q∆′(cid:17), and we already know that ǫ3 < 1. Using this
concentration bound, the following calculation holds with probability 1 − 1/poly(n).
q∆ (cid:17) = Θ(cid:16)q log n
gL(v) ≥ (1 − ǫ3)q∆′
≥ q∆′ − O(cid:16)pq∆′ log n(cid:17)
≥ q∆ − q∆λ − O(cid:16)pq∆ log n(cid:17) .
Combining this with ∆L ≤ (1 + ǫ3)q∆ = q∆ + O(√q∆ log n), we obtain gL(v) ≥ ∆L − q∆λ −
O(√q∆ log n). Note that q∆λ + O(√q∆ log n) = o(cid:0)(q∆)λ(cid:1) = o(cid:0)∆λ
L(cid:1),6 and so we finally obtain
gL(v) ≥ ∆L − ∆λ
iv) Remaining Degrees: The degree upper bounds of ∆i and ∆L follow immediately from the
concentration bounds on degBi(v) and degL(v) calculated in the proof of i). The bounds degBi(v) ≤
max{O(log n), O(1/√∆) · deg(v)} and degL(v) ≤ max{O(log n), O(q) · deg(v)} can be derived by a
L + 1.
straightforward application of Chernoff bound.
3.2 Congested Clique Algorithm for High-Degree Graphs
In this section, we show that the (∆ + 1)-list coloring problem can be solved in O(1) rounds in
the CONGESTED-CLIQUE model when the degrees are assumed to be sufficiently high. The formal
statement is captured in Theorem 3.2. First, we show that the partitioning algorithm can indeed be
implemented in the CONGESTED-CLIQUE model. Then, we show how to color the parts resulting
from the graph partitioning efficiently. The proof of Theorem 3.2 is completed by showing that only
O(1) recursive applications of the partitioning are required.
5If this is not the case, we can increase the degree of v in a vacuous way by adding dummy neighbors to it. For
instance, we can add a clique of size ∆ next to v (to be simulated by v), remove a large enough matching from this
clique and instead connect the endpoints to v.
6The bound √q∆ log n ≪ (q∆)λ can be derived from the assumptions λ = 1
2 log1/2 n ≪ (q∆)
2 n) =⇒ √q∆ log n = (q∆)
2 n) = ω(log
4 log
4 γ+ 1
1
3
q∆ = Θ(∆
3
1
2 + 2
2 (q∆)
1
3γ+2 and ∆ = ω(logγ n), as follows:
1
2 ( 3
4 γ+ 1
2 )−1
= (q∆)λ.
10
Implementation of the Graph Partitioning. The partitions can be computed in O(1) rounds
on CONGESTED-CLIQUE. Partitioning the vertex set V is straightforward, as every vertex can
make the decision independently and locally, whereas it is not obvious how to partition C to make
all vertices agree on the same partition. Note that we can assume C ≤ (∆ + 1)V ; if C is greater
than (∆ + 1)V initially, then we can let each vertex decrease its palette size to ∆ + 1 by removing
some colors in its palette, and we will have C ≤ (∆ + 1)V after removing these colors.
A straightforward way of partitioning C is to generate Θ(C log n) random bits at a vertex v
locally, and then v broadcasts this information to all other vertices. Note that it takes O(log k) =
O(log V ) = O(log n) bits to encode which part of C1 ∪ ··· ∪ Ck each c ∈ C is in. A direct
implementation of the approach cannot be done in O(1) rounds, due to the message size constraint
of CONGESTED-CLIQUE, as each vertex can send at most Θ(n log n) bits in each round.
To solve this issue, observe that it is not necessary to use total independent random bits for
each c ∈ C, and Θ(log n)-wise independence suffices. More precisely, suppose X is the summation
of n K-wise independent 0-1 random variables with mean p, and so µ = E[X] = np. A Chernoff
bound with K-wise Independence [SSS95] guarantees that
Pr[X ≥ (1 + q)µ] ≤ exp(cid:0)− min{K, q2µ}(cid:1) .
In order to guarantee a failure probability of 1/poly(n) in all applications of Chernoff bound in
Lemma 3.1, it suffices that K = Θ(log n). Therefore, to compute the decomposition C = C1 ∪
···∪ Ck with K-wise independent random bits, we only need O(K · log(C log k)) = O(log2 n) total
independent random bits. Broadcasting O(log2 n) bits of information to all vertices can be done in
O(1) rounds via Lenzen's routing (Lemma 2.1).
The Algorithm of (∆ + 1)-list coloring on High-degree Graphs. We next present our
CONGESTED-CLIQUE-model coloring algorithm for high-degree graphs, using the partitioning ex-
plained above.
Theorem 3.2. Suppose ∆ = Ω(log4+ǫ n) for some constant ǫ > 0. There is an algorithm that solves
(∆ + 1)-list coloring in CONGESTED-CLIQUE in O(1) rounds.
Proof. We show that a constant-depth recursive applications of Lemma 3.1 suffices to give an O(1)-
round CONGESTED-CLIQUE (∆ + 1)-list coloring algorithm for graphs with ∆ = Ω(log4+ǫ n), for
any constant ǫ > 0. Consider the graph G = (V, E). First, we apply the graph partitioning
algorithm of Lemma 3.1 to partition vertices V into subsets B1, . . . , Bk, L with parameter n = V ,
and k = √∆. After that, let arbitrary k = √∆ vertices to be responsible for coloring each G[Bi].
Each of these k vertices, in parallel, gathers all information of G[Bi] from vertices Bi, and then
computes a proper coloring of G[Bi], where each vertex v ∈ Bi uses only the palette Ψ(v) ∩ Ci.
The existence of such a proper coloring is guaranteed by Property (ii). Using this approach, we can
color all vertices in V \ L in O(1) rounds using Lenzen's routing. Note that Property (i) guarantees
that E(G[Bi]) = O(n). Finally, each vertex v ∈ L removes the colors that have been taken by its
neighbors in V \L from its palette Ψ(v). In view of Property (iii), after this operation, the number
of available colors for each v ∈ L is at least gL(v) ≥ max{degL(v), ∆L − ∆λ
L} + 1. Now the subgraph
G[L] satisfies all conditions required to apply Lemma 3.1, so long as ∆L = ω(logγ n). We will see
that this condition is always met in our application.
We then recursively apply the algorithm of the lemma on the subgraph induced by vertices L
with the same parameter n. The recursion stops once we reach a point that E(G[L]) = O(n), and
so we can apply Lenzen's routing to let one vertex v gather all information of G[L] and compute
its proper coloring.
11
Now we analyze the number of iterations needed to reach a point that E(G[L]) = O(n). Here
we use γ = 2 and λ = 3/4.7 Define V1 = V and ∆1 = ∆ as the vertex set and the maximum degree
for the first iteration. Let V = B1 ∪ ··· ∪ Bk ∪ L be the outcome of the first iteration, and define
V2 = L and ∆2 = ∆L. Similarly, for i > 2, we define Vi and ∆i based on the set L in the outcome
of the graph partitioning algorithm for the (i − 1)th iteration. We have the following formulas.
∆1 = ∆
∆i = ∆i−1 · O √log n
i−1 !
Vi = Vi−1 · O √log n
i−1 !
V1 = n
∆1/4
∆1/4
by Property iv)
by Property i)
Let α > 0 be chosen such that ∆ = ∆1 = (log n)2+α, and assume α = Ω(1) and i = O(1). We can
calculate the value of ∆i and Vi as follows.
∆i = O(cid:16)(log n)2+α·(λ)i−1(cid:17)
Vi = n · O(cid:16)(log n)α((λ)i−1−1)(cid:17)
Thus, given that α = Ω(1) and i = O(1), the condition of ∆i = ω(logγ n) = ω(log2 n) for applying
Lemma 3.1 must be met.
Next, we analyze the number of iterations it takes to make ∆iVi sufficiently small.
In the
CONGESTED-CLIQUE model, if ∆iVi = O(n), then we are able to compute a proper coloring of Vi
in O(1) rounds by information gathering. Let us write ∆ = log2+α n, where α = 2 + β. The lemma
statement implies that β = Ω(1). Note that the condition for ∆iVi = O(n) can be re-written as
Combining this with α = 2 + β, a simple calculation shows that this condition is met when
2α(cid:0)1 − (λ)i−1(cid:1) ≥ 2 + α.
3β (cid:19) / log (4/3) .
i ≥ log(cid:18) 8(β + 2)
3β (cid:17) / log (4/3) = O(1), and so our algorithm takes only O(1)
Since β = Ω(1), we have log(cid:16) 8(β+2)
iterations. In particular, when β ≥ 10.8, i.e., ∆ = Ω(log12.8 n), we have ∆4V4 = O(n), and so 3
iterations suffice. Since each iteration can be implemented in CONGESTED-CLIQUE in O(1) rounds,
overall we get an algorithm with round complexity O(1).
Remark 3.3. Similar to the proof of Theorem 3.2, the graph partitioning algorithm also leads
memory. This gives an simple alternate proof (with a slightly worse memory size) of the main result
to an O(1)-round MPC coloring algorithm with S = eO(n) memory per processor and eO(m) total
of [ACK19] that (∆ + 1)-coloring can be solved with S = eO(n) memory per processor.
7We choose γ = 2 (the smallest possible) to minimize the degree requirement in Theorem 3.2.
12
3.3 Massively Parallel Computation with Strongly Sublinear Memory
We now show how to apply Lemma 3.1 as well as the CLP algorithm of [CLP18], as summarized in
the following lemma, to prove Theorem 1.2.
Lemma 3.4 ([CLP18, Par18]). Let G be an n-vertex graph with m edges and maximum degree
Then the list-coloring problem can be solved w.h.p. in O(√log log n) rounds of low-memory MPC
∆. Suppose any vertex v has a palette Ψ(v) that satisfies Ψ(v) ≥ max(cid:8)degG(v) + 1, ∆ − ∆3/5(cid:9).
with local memory O(nα) for an arbitrary constant α ∈ (0, 1) and total memory eO(cid:0)Pv degG(v)2(cid:1)
if ∆2 = O (nα).
The proof of Lemma 3.4 almost immediately follows from [CLP18, Par18]; there are only few
changes that have to be made in order to turn their CONGESTED-CLIQUE algorithm into a low-
memory MPC algorithm. The details are deferred to Appendix E.
Proof of Theorem 1.2. We present a recursive algorithm based on the randomized partitioning al-
gorithm of Lemma 3.1. If ∆ = poly(log n) then the conditions of Lemma 3.4 are satisfied trivially;
we can solve the problem in O(log∗ ∆ + √log log n) = O(√log log n) rounds of low-memory MPC
with total memory eO(n · ∆2) = eO(m). Otherwise, we execute the following algorithm.
Randomized Partitioning: Let G be the graph that we want to color. We apply the randomized
partitioning algorithm of Lemma 3.1 to G, which gives us sets B1, . . . , Bk and L, as well as color
sets C1, . . . , Ck. The goal is now to first color B1, . . . , Bk with colors from C1, . . . , Ck, respectively.
i=1 Bi. Then, for
every vertex in L, we remove all colors already used by neighbors in B from the palettes, leaving us
with a list-coloring problem of the graph induced by L with maximum degree ∆L.
Since the colors in the sets Ci are disjoint, this gives a proper coloring of B :=Sk
In the following, we first describe how to color each set Bi with colors in Ci, and then how to
solve the remaining list-coloring problem in L. For the parameters in Lemma 3.1, we use γ = 6 and
λ = 3/5.8
i = O(nα), then,
List-Coloring Problem in Bi:
by Lemma 3.1 ii), Bi satisfies the conditions of Lemma 3.4 We thus can apply the algorithm of
Lemma 3.4 to Bi. Otherwise, we recurse on Bi. Note that this is possible since, by Lemma 3.1 ii)
applied to G, Bi satisfies the conditions of Lemma 3.1.
If the maximum degree ∆i in Bi satisfies ∆2
L = O(nα), then,
List-Coloring Problem in L:
by Lemma 3.1 iii) applied to G, L satisfies the conditions of Lemma 3.4. We thus can apply the
algorithm of Lemma 3.4 to L. Otherwise, we recurse on L. Note that this is possible since by
Lemma 3.1 iii), L satisfies the conditions of Lemma 3.1.
If the maximum degree ∆L in L satisfies ∆2
Number of Iterations: Since the maximum degree in L reduces by a polynomial factor in every
step, after at most O(1/α) steps, the resulting graph has maximum degree at most O(nα/2), where
we satisfy the conditions of Lemma 3.4, and hence do not recurse further. Note that when recursing
on sets Bi, the degree drop is even larger, and hence the same reasoning applies to bound the
number of iterations.
8The choice λ = 3/5 is to ensure that the number of available colors for each vertex in each subgraph meets the
palette size constraint specified in Lemma 3.4.
13
Memory Requirements:
It is obvious that the recursive partitioning of the input graph G does
not incur any overhead in the memory, neither local nor global. Now, let H be the set of all graphs
H on which we apply the algorithm of Lemma 3.4. As we only apply this algorithm when the
maximum degree ∆H of H is O(nα/2) or poly(log n), we clearly have ∆2
H = O(nα), so the algorithm
Lemma 3.4 is guaranteed to run with local memory O(nα).
number of edges in the input graph G, as promised in Theorem 1.2. First, observe that due to the
First, assume that the graph G has been partitioned at least three times to get to H. By
Lemma 3.1 iv), the degree of any vertex v in H is either O(1) or at most
It remains to show how to guarantee the total memory requirement of eO(m), where m is the
specifications of Lemma 3.4, we can write the total memory requirement asPH∈HPv∈H (degH(v))2.
degG(v) · O(cid:16)∆− 1
4 )2(cid:17) = degG(v) · O(cid:16)∆−37/64(cid:17) < O(cid:16)pdegG(v)(cid:17) .
4(cid:17) · O(cid:16)∆− 1
4(cid:17) · O(cid:16)∆− 1
Note that in the above calculation we assume v always goes to the left-over part L in all three
iterations. If v goes to Bi, then the degree shrinks faster. Remember that we set q = O(∆−1/4).
Hence, we require a total memory of
4· 3
4·( 3
eO XH∈HXv∈H
(degH(v))2! = eO XH∈HXv∈H
degG(v)! = eO Xv∈G
degG(v)! = eO(m) .
Note that the algorithm can be easily adapted to always perform at least three partitioning steps
if ∆H is bounded from below by a sufficiently large poly(log n), because then the conditions of
Lemma 3.1 are satisfied. On the other hand, if ∆H = poly(log n), it is follows immediately that
eO(cid:0)Pv(degH(v))2(cid:1) = poly(log n) = eO(1). Put together, we havePH∈HPv∈H (degH(v))2 = eO(m).
4 Distributed Coloring with Palette Sparsification
In this section, we present our sparsification for the LOCAL-model coloring algorithm of CLP [CLP18],
which is the second novel technical ingredient in our results. As a consequence, this sparsifi-
cation gives us (i) an LCA solving (∆ + 1) list coloring with query complexity ∆O(1) · O(log n)
and (ii) an O(1)-round CONGESTED-CLIQUE algorithm solving (∆ + 1) list coloring for the case
∆ = O(poly log n), using the speedup lemma (Lemma 2.3).
The Chang-Li-Pettie Coloring Algorithm. We will not sparsify the entire algorithm of [CLP18].
The algorithm of [CLP18] is based on the graph shattering framework. Each vertex successfully col-
ors itself with probability 1 − 1/poly(∆) during the pre-shattering phase of [CLP18], and so by the
shattering lemma (Lemma 2.2), the remaining uncolored vertices VBad form connected components
of size ∆O(1)O(poly log n).9 The post-shattering phase then applies a deterministic (deg +1)-list
coloring algorithm to color them. Lemma 2.2 guarantees that the number of edges within VBad is
O(n), and so they can be colored in O(1) rounds in the CONGESTED-CLIQUE model. Similarly,
dealing with VBad only adds an ∆O(1) · O(log n)-factor overhead for LCA. Thus, we only need to
focus on the pre-shattering phase, which consists of the following three steps.
Initial Coloring Step: This step is an O(1)-round procedure that generates excess colors at ver-
tices that are locally sparse.
9In the analysis of [CLP18], this can also be made O(poly log n), regardless of ∆.
14
Dense Coloring Step: This step is an O(1)-round procedure that colors most of the locally dense
vertices.
Color Bidding Step: This step is an O(log∗ ∆)-round procedure that colors most of the remaining
uncolored vertices, using the property that these vertices have large number of excess colors.
For our LCA and CONGESTED-CLIQUE algorithms, the plan is to run the initial coloring step
and the dense coloring step by a direct simulation, which costs O(1) rounds. Then, we will give a
sparsified version of the color bidding step where each vertex v only need to receive the information
from O(poly log ∆) of its neighbors to decide its output.
In view of the above, we will use part of the algorithm
A Black Box Coloring Algorithm.
of [CLP18] as a black box. The specification of this black box is as follows. Consider an instance
of the (∆ + 1)-list coloring on the graph G = (V, E). The black box algorithm colors a subset of
V such that the remaining uncolored vertices are partitioned into three subsets VGood, VBad, and R
meeting the following conditions.
Good Vertices: The edges within VGood are oriented as a DAG, and each vertex v ∈ VGood is
associated with a parameter pv ≤ Ψ(v) − deg(v) satisfying the conditions p⋆ = minv∈V pv ≥
∆/ log ∆ and Pu∈N out(v) 1/pu ≤ 1/C, where C > 0 can be any specified constant.10 Recall
that N out(v) refers to the set of out-neighbors of v.
Bad Vertices: The probability that a vertex v ∈ V joins VBad is 1 − 1/poly(∆). In particular, in
view of Lemma 2.2, with probability 1 − 1/poly(n), they form connected components of size
∆O(1) · O(log n), and the number of edges within the bad vertices is O(n).
Remaining Vertices: The subgraph induced by R has a constant maximum degree.
Lemma 4.1 follows from [CLP18], after some minor modifications. For the sake of completeness
we show the details of how we obtain Lemma 4.1 from the results in [CLP18] in Appendix D. Note
that for the case of CONGESTED-CLIQUE, as long as ∆ = O(√n), Lemma 4.1 can be implemented
in O(1) rounds.
Lemma 4.1 ([CLP18]). Consider an instance of the (∆ + 1)-list coloring on the graph G = (V, E).
There is an O(1)-round LOCAL algorithm that colors a subset of vertices such that the remaining
uncolored vertices are partitioned into three subsets VGood, VBad, and R meeting the above conditions,
and the algorithm uses O(∆2 log n)-bit messages.
4.1 A Sparsified Color Bidding Algorithm
In view of Lemma 4.1, we focus on the subgraph induced by VGood, and denote it as G0 = (V0, E0).
The graph G0 is a directed acyclic graph. The set of available colors for v is denoted as Ψ0(v). Our
goal is to give a proper coloring of G0. An important property of G0 is that each vertex v ∈ V is
associated with a parameter pv ≤ Ψ0(v) − degG0(v) such thatPu∈N out(v) 1/pu ≤ 1/C0, where C0
can be any specified large constant. Intuitively, pv gives the lower bound of the number of excess
colors at vertex v. It is guaranteed that p⋆ = minv∈V0 pv ≥ ∆/ log ∆. Parameters C0 and p⋆ are
initially known to all vertices in V0.
10Here Ψ(v) is the set of available colors at v, i.e., the colors in the palette of v that have not been taken by v's
neighbors. Here deg(v) refers to the number of uncolored neighbors of v in VGood. We use outdeg(v) to refer to the
number of out-neighbors of v. Intuitively, pv ≤ Ψ(v) − deg(v) is a lower bound on the number of excess colors at v.
15
Review of the Color Bidding Algorithm. The above conditions might look a bit strange,
but it allows us to find a proper coloring in O(log∗ ∆) rounds in the LOCAL model by applying
O(log∗ ∆) iterations of the procedure ColorBidding [CLP18], as follows.
1. Each color c ∈ Ψ(v) is added to Sv with probability
2. If there exists a color c⋆ ∈ Sv that is not selected by any vertex in N out(v), v colors itself c⋆.
We give a very high-level explanation about how this works. For the first iteration we use C = C0.
Intuitively, for each color c ∈ Sv, the probability that c is selected by an out-neighbor of v is
independently.
2Ψ(v)
C
Xu∈N out(v)
C/(2Ψ(u)) ≤ Xu∈N out(v)
C/(2pu) ≤ 1/2.
In the calculation we use the inequalityPu∈N out(v) 1/pu ≤ 1/C0 that is guaranteed by Lemma 4.1.
The probability that v fails to color itself is roughly 1/2Sv, which is exponentially small in C0,
as in expectation Sv = C0/2. Thus, for the next iteration we may use a parameter C that is
exponentially small in C0, and so after O(log∗ ∆) iterations, we are done.
Parameters. Let β > 0 be a constant to be determined. Let p⋆ ∈ [∆/ log ∆, ∆] be the parameter
specified in the conditions for Lemma 4.1. The C-parameters used in the algorithms C0, . . . , Ck−1
are defined as follows. For the base case, C0 is the parameter C specified in the conditions for
Lemma 4.1. Given that Ci has been defined, we set
Ci+1 = 2(cid:24)(cid:18)min(cid:26) 1
2
exp(Ci/6)Ci, logβ p⋆(cid:27)(cid:19) /2(cid:25) − 2.
In other words, Ci+1 is the result of rounding min(cid:8) 1
2 exp(Ci/6)Ci, logβ p⋆(cid:9) down to the nearest even
number. The number of iterations k is chosen as the smallest index such that Ck−1 = 2(cid:6)logβ p⋆/2(cid:7)−
2. It is clear that k = O(log∗ ∆), as p⋆ ≤ ∆ + 1.
We will use this sequence C0, . . . , Ck−1 in our sparsified color bidding algorithm. This sequence
is slightly different than the one used in [CLP18]. The last number in the sequence used in [CLP18]
is set to be √p⋆, but here we set it to be O(poly log p⋆). Having a larger C-parameter leads to
a smaller failure probability, but it comes at a cost that we have to sample more colors, and this
means that each vertex needs to communicate with more neighbors to check for conflict.
Overview of the Proof. We first review the analysis of the multiple iterations of ColorBidding
in [CLP18], and then we discuss how we sparsify this algorithm. The proof in [CLP18] maintains an
invariant Ii(v) for each vertex v that is uncolored at the beginning of each iteration i, as follows.11
Ii(v) : Xu∈N out
G (v)
1/pu ≤ 1/Ci.
We will use the same pu because the number of excess colors of a vertex never decreases. By
Lemma 4.1, this invariant is met for i = 0. The vertices u not satisfying the invariant Ii(v) are
considered bad, and are removed from consideration. The analysis of [CLP18] shows that
11In this section, G refers the current graph under consideration, i.e., it excludes all vertices that have been colored
or removed in previous iterations. We use G0 to refer to the original graph.
16
1. Suppose all vertices u in G at the beginning of the ith iteration satisfy the Ii(u). Then at
end of this iteration, for each vertex u, with probability 1 − 1/poly(∆), either u has been
successfully colored, or Ii+1(u) is satisfied.
2. For the last iteration, Given that all vertices u in G satisfy Ik−1(u), then v is successfully
colored at iteration k with probability 1 − 1/poly(∆).
By the shattering lemma (Lemma 2.2), all vertices that remain uncolored at the end of the algorithm
induce a subgraph with O(n) edges. In particular, in CONGESTED-CLIQUE we are able to color
them in O(1) additional rounds.
To sparsify the algorithm, our strategy is to let each vertex sample the colors needed in all
iterations at the beginning of the algorithm.
It is straightforward to see that each vertex only
needs to use O(poly log ∆) colors throughout the algorithm, with probability 1 − 1/poly(∆). After
sampling the colors, if u finds that v ∈ N out(u) do not share any sampled color, then there is no need
for u to communicate with v. This effectively reduces the maximum degree to ∆′ = O(poly log ∆).
If ∆ = O(poly log n), then ∆′ = O(poly(log log n)), which is enough to apply the opportunistic
speedup lemma (Lemma 2.3).
There is one issue needed to be overcome. That is, verifying whether Ii(u) is met has to be done
on the original graph G, as we have to go over all vertices v ∈ N out
G (u), regardless of whether u and
v have shared sampled colors. One way to deal with this issue is to simply not remove the vertices
u violating Ii(u), but if we do it this way, then when we calculate the failure probability of a vertex
v, we have to apply a union bound over all vertices u within radius τ = O(log∗ ∆) to v that u does
not violate the invariant for each iteration. Due to this union bound, we can only upper bound the
size of the connected components of bad vertices by ∆O(log∗ ∆) · O(log n), so this does not lead to
an improved LCA.12 To resolve this issue, we observe that the invariant Ii(u) might be too strong
for our purpose, since intuitively if v ∈ N out(u) does not share any sampled colors with u, then v
should not be able to affect u throughout the algorithm.
In this paper, we will consider an alternate invariant I′i(u) that can be checked in the sparsified
graph. More precisely, in each iteration, each vertex v will do a two-stage sampling to obtain two
color sets Sv ⊆ Tv ⊆ Ψ(v). The set Sv has size C/2, and the set Tv has size logβ ∆, where β > 0 is
a constant to be determined. The alternate invariant I′i(v) is defined as
This invariant I′i(v) can be checked by having v communicating only with its neighbors that share
a sampled color with v. Intuitively, if Ii(v) holds, then I′i(v) holds with probability 1 − 1/poly(∆).
It is also straightforward to see that I′i(v) implies that v has a high probability of successfully
coloring itself in this iteration, as Sv is a size-(Ci/2) uniformly random subset of Tv. In subsequent
discussion, we say that v is rich if I′i(v) is met. Other than not satisfying I′i(v), there are two other
bad events that we need to consider:
• (Informally) v has too many neighbors that share a sampled color with v; in this case, we
say that v is overloaded. This is a bad event since the goal of the palette sparsification is to
reduce the number of neighbors that v needs to receive information from.
12We remark that this is only an issue for LCA, and this is not an issue for application in CONGESTED-CLIQUE.
In the shattering lemma (Lemma 2.2), for the parameters ∆ = O(poly log n) and c = O(log∗ ∆), we can still bound
the number of edges within the bad vertices B by O(n).
17
I′i(v) :(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Tv \ [u∈N out
G (v)
Su(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≥ Tv/3.
• Most of the sampled colors of v reserved for iteration i have already be taken by the neighbors
of v during the previous iterations 1, . . . , i − 1, so v does not have enough colors to correctly
run the algorithm for the ith iteration; in this case, we say that v is lazy.
The Sparsified Color Bidding Algorithm. We are now in a position to describe the sparsified
version of ColorBidding. For the sake of clarity we use the following notations to describe the palette
of a vertex u. Recall that Ψ0(u) refers to the palette of u initially in the original graph G0. At the
beginning of an iteration, we write Ψ+(u) to denote the set of available colors at u, and write Ψ−(u)
to denote the set of colors already taken by vertices in NG0(u). Note that Ψ+(u) = Ψ0(u) \ Ψ−(u).
The function SampleColors describe how we sample the colors Su and Tu in an iteration. Intu-
itively, we use k1 = C/2 and k2 = logβ p⋆ as the target set sizes. The set R represents a length-K
sequence of colors that u pre-sampled for the ith iteration, where K = log3+β p⋆, and R(j) repre-
sents the jth color of R. We will later see that R is generated in such a way that each R(j) is a
uniformly random color chosen from Ψ0(u), where Ψ0(u) is the set of available colors of v initially in
G0. The set S− represents the set Ψ−(u) which consists of the colors already taken by the vertices
in NG0(u) before iteration i.
Function SampleColors(k1, k2, S−, R)
T ← ∅;
for j ← 1 to k2 log3 p⋆ do
c ← R(j);
if c /∈ S− then
T ← T ∪ {c};
if T = k1 then
T1 ← T ;
if T = k2 then
return (T1, T );
return (∅, ∅);
end
The procedure SparsifiedColorBidding is the sparsified version of ColorBidding. In this proce-
dure, it is straightforward to verify that the outcome Sv ← T1 and Tv ← T of SampleColors(C/2,
logβ p⋆, Ψ−(v), Rv) satisfies either one of the following:
• Sv = ∅ and Tv = ∅. This happens when most of the pre-sampled colors for this iteration have
been taken by the neighboring vertices. We will later show that this occurs with probability
1/poly(∆).
• Given that each Rv(j) is a uniformly random color of Ψ0(v), we have: (i) Sv is a size-(C/2)
uniformly random subset of Tv, and (ii) Tv is a size-(cid:0)logβ p⋆(cid:1) uniformly random subset of
Ψ0(v) \ Ψ−(v). That is, these two sets Sv and Tv are sampled uniformly randomly from the
set of available colors of v, i.e., Ψ0(v) \ Ψ−(v).
The condition for v to be overloaded is defined in the procedure SparsifiedColoring. Intu-
R(i′)
itively, v is said to be overloaded at iteration i if the colors in R(i)
u ,
for too many neighbors u ∈ NG0(v); this is undesirable as we want the degree of the sparsified graph
v have appeared inS0≤i′≤i
18
to be small.
Procedure SparsifiedColorBidding(G, C, Ψ−, {Rv}v∈V0 )
for each vertex v ∈ V (G) do simultaneously
1. (Sv, Tv) ← SampleColors(C/2, logβ p⋆, Ψ−(v), Rv). If v is overloaded, reset
(Sv, Tv) ← (∅, ∅). We call v lazy if Sv = ∅.
2. v collects information about Su from all neighbors u ∈ N out
G (v) Su(cid:12)(cid:12)(cid:12) ≥ Tv/3, i.e., at most 2/3 of colors v sampled in Tv are
3. If(cid:12)(cid:12)(cid:12)Tv \Su∈N out
selected in Su of some neighbors u ∈ N out
or (ii) v is lazy, then v marks itself Bad and it skips the next step.
4. If there is a color c ∈ Sv that is not inSu∈N out
If v is lucky with c, v colors itself c. Tie is broken arbitrarily.
G (v) Su, we call v lucky with color c.
G (v), then we call v rich. If (i) v is not rich
G (v).
end
The procedure SparsifiedColoring represents the entire coloring algorithm, which consists of
k = O(log∗ ∆) iterations of SparsifiedColorBidding. The notation G[U ] refers to the subgraph
induced by U . Note that the set U does not include the vertices that are marked Bad, i.e., once a
vertex v marked itself Bad, it stops attempting to color itself; but a Bad vertex might still need
to provide information to other vertices in subsequent iterations.
Procedure SparsifiedColoring()
G ← G0;
K ← log3+β ∆;
for i ← 0 to k − 1 do /* Obviously k = O(log∗ p⋆ − log∗ C0) = O(log∗ ∆).
*/
1. G ← G[U ], where U consists of the yet uncolored vertices in G that are not Bad.
2. Each vertex v ∈ V (G0) generates a color sequence R(i)
following rule: for each j = 1, . . . , K, R(i)
random, independently.
v (j) is a color in Ψ0(v), chosen uniformly at
v (1), . . . , R(i)
v (K) by the
u }0≤i′≤i from each
3. Each vertex v ∈ V (G) gathers the information about {R(i′)
neighbor u ∈ NG0(v).
4. If there exist three indices i′ ∈ [0, i], j ∈ [1, K], and j′ ∈ [1, K] such that
v (j) = R(i′)
R(i)
more than K 2 log ∆ significant neighbors, we call v overloaded.
5. Each vertex v ∈ V (G) gathers the information about the colors that have been
taken by the vertices in NG0(v). Let Ψ−(v) be the set of these colors.
6. Call SparsifiedColorBidding(G, Ci, Ψ−, {R(i)
u (j′), we say u ∈ NG0(v) is a significant neighbor of v ∈ V (G). If v has
v }v∈V0 ).
end
It is straightforward to see that SparsifiedColoring can be implemented in such a way that
after an O(1)-round pre-processing step, each vertex v is able to identify O(poly log ∆) neighbors
such that v only need to receive information from these vertices during SparsifiedColoring. In
the pre-processing step, we let each vertex v sample the color sequences R(i)
for each 0 ≤ i ≤ k − 1,
v
and let each vertex v learn the set of colors sampled by NG0(v). Based on this information, before
the first iteration begins, v is able to identify at most K 2 log ∆ = O(poly log ∆) neighbors of v
for each iteration i such that v is sure that v does not need to receive information from all other
neighbors during the ith iteration. See Section 4.3 for details.
For the rest of Section 4.1, we focus on the analysis of SparsifiedColoring. For each iteration
i, recall that Ψ+(v) = Ψ0(v) \ Ψ−(v) is the set of available colors at v at the beginning of this
iteration. For a vertex v ∈ V0, and its neighbor u ∈ NG0(v), we say that u is a c-significant
19
neighbor of v in iteration i if c = R(i′)
u (j) for some i′ ∈ [1, i] and j ∈ [1, k].
Consider the beginning of the ith iteration of the for-loop in SparsifiedColoring. In the graph
G = (V, E) ← G[U ] under consideration in this iteration, we say that a vertex v ∈ V is (C, D)-honest
if the following two conditions are met.
G (v) 1/pu ≤ 1/C.
(i) Pu∈N out
iteration.
(ii) For each color c ∈ Ψ0(v), v has at most D c-significant neighbors u ∈ NG0(v) in the previous
Clearly all vertices are (C0, 0)-honest in G = G0 at the beginning of iteration i = 0. Lemma 4.2
shows that (C, D)-honest vertices are well-behaved.
Lemma 4.2. Consider the ith iteration of SparsifiedColorBidding in SparsifiedColoring. Let
U be the set of yet uncolored vertices after this iteration. Suppose a vertex v is (C, D)-honest, with
C ≤ logβ p⋆ and D ≤ 2K · k = O(K log∗ ∆), at the beginning of this iteration, then The following
holds.
i) Pr[v does not successfully color itself] ≤ exp(−C/6) + exp(−Ω(logβ ∆)).
ii) Pr[v marks itself Bad] ≤ exp(−Ω(logβ ∆)).
iii) Pr[at the beginning of the next iteration, v ∈ U or v is not (C′, D′)-honest]
≤ exp(−Ω(logβ ∆)), where C′ = min(cid:8) 1
2 exp(C/6)C, logβ p⋆(cid:9) and D′ = D + 2K.
The probability calculation only relies on the distribution of random bits generated in N 2
G0
iteration, i.e., {R(i)
N 2
G0
(v) in this
(v). In particular, the result holds even if random bits generated outside
(v) are determined adversarially.
u }u∈N 2
G0
Note that Lemma 4.2 only relies on the assumption that the vertex v under consideration is
(C, D)-honest, and the lemma works even many of neighboring of v are not (C, D)-honest. This is
in contrast to most of the analysis of graph shattering algorithms where the analysis relies on the
assumption that all vertices at the beginning of each iteration to satisfy certain invariants. Based on
Lemma 4.2, we show that SparsifiedColoring colors a vertex with a sufficiently high probability
that enables us to apply the shattering lemma.
Lemma 4.3. The algorithm SparsifiedColoring gives a partial coloring of G0 such that the
probability that a vertex v does not successfully color itself with a color in Ψ0(v) is
O(k) · exp(cid:16)−Ω(logβ ∆)(cid:17) ≪ 1/poly(∆),
and this holds even if the random bits generated outside N 2
G0
(v) are determined adversarially.
Proof. We consider the sequence D0 = 0 and Di+1 = Di + 2K. Suppose the algorithm does not
color a vertex v, then v must falls into one of the following cases.
• There is an index i ∈ [0, k − 2] such that v is (Ci, Di)-honest at the beginning of iteration i,
but v is not (Ci+1, Di+1)-honest at the beginning of iteration i + 1. By Lemma 4.2 (iii), this
• There is an index i ∈ [0, k − 1] such that v is (Ci, Di)-honest at the beginning of iteration i,
but v marks itself Bad in iteration i. By Lemma 4.2 (ii), this occurs with probability at most
occurs with probability at most (k − 1) · exp(cid:0)−Ω(logβ ∆)(cid:1).
k · exp(cid:0)−Ω(logβ ∆)(cid:1).
20
• For the last iteration i = k−1, the vertex v is (Ck−1.Dk−1)-honest at the beginning of iteration
k − 1, but v does not successfully colors itself by a color in its palette. in iteration k − 1. By
Lemma 4.2 (iii), this occurs with probability at most exp(cid:0)−Ω(logβ ∆)(cid:1).
Note that our analysis only relies on the distribution of random bits generated in N 2
(v), which
G0
is guaranteed by Lemma 4.2. That is, even if the adversary is able to decide the random bits of
vertices outside of N 2
(v) throughout the algorithm SparsifiedColoring, the probability that v
G0
does not successfully color itself is still at most O(k) · exp(cid:0)−Ω(logβ ∆)(cid:1).
4.2 Analysis for the Sparsified Color Bidding Algorithm
v
, Erich
v
v
, Elazy
, and Elucky
In this section, we prove Lemma 4.2. We focus on the ith iteration of the algorithm, where the
vertex v is (C, D)-honest, and there is no guarantee about the (C, D)-honesty of all other vertices.
For this vertex v, we write Eoverload
to denote the event that v is overloaded,
lazy, rich, and lucky. Note that a lucky vertex must be rich and not lazy, and an overloaded vertex
must be lazy. In this proof we frequently use this inequality ∆ + 1 ≥ Ψ0(v) ≥ Ψ+(v) ≥ pv ≥ p⋆ =
Ω(∆/ log ∆). Our analysis only considers the random bits generated by vertices within N 2
(v) in
G0
this iteration.
Claim 4.4. The probability that v has more than D′ = D + 2K c-significant neighbors u ∈ NG0(v)
for some color c ∈ Ψ0(v) in this iteration i is at most exp(−Ω(log3+β ∆)), and this implies that
Pr[Eoverload
v
v
] ≤ exp(−Ω(log3+β ∆)).
Proof. Since v is (C, D)-honest, our plan is to show that for each color c ∈ Ψ0(v), the number of
new c-significant neighbor u ∈ NG0(v) brought by the color sequences in the ith iteration R(i)
u , is
at most 2K with probability 1 − exp(−Ω(log3+β ∆)).
ur }, Y =P1≤r≤s Xj. Then
Y is an upper bound on the number of new c-significant neighbors. Since v is (C, D)-honest, the
total number of c-significant neighbors is at most D + Y . To prove this claim, it suffices to bound
the probability of Y > 2K. Note that X1, . . . , Xs are independent, and
Write NG0(v) = {u1, . . . , us}, and let Xr = 1{color c appears in R(i)
E[Y ] = X1≤r≤s
E[Xr] ≤ X1≤r≤s 1 −(cid:18)1 −
1
Ψ0(ur)(cid:19)K! ≤ s ·
K
∆ + 1 ≤
K∆
∆ + 1
< K.
By a Chernoff bound, we have Pr[Y ≥ 2K] ≤ exp(−Ω(K)) = exp(−Ω(log3+β ∆)). By a union
bound over all c ∈ Ψ0(v) we are done.
Given that v has no more than D′ c-significant neighbors in this iteration for every color c ∈
Ψ0(v), we infer that v has at most
R(i)
u D′ ≤ K · D′ = K · (D + 2K) ≪ K 2 log ∆
significant neighbors, which implies that v is not overloaded. Hence we also have Pr[Eoverload
exp(−Ω(log3+β ∆)).
Claim 4.5. Pr[Elazy
v
v
] ≤ exp(−Ω(log2+β ∆)).
] ≤
Proof. Remember that v is lazy if either (i) v is overloaded, or (ii) SampleColors gives (∅, ∅).
In view of Claim 4.4, we only need to show that with probability at most exp(−Ω(log2+β ∆)),
SampleColors gives (∅, ∅).
21
Consider the sampling process in SampleColors, and suppose that we are in the middle of the
process, and T is the current set of colors we have obtained. Suppose T = r currently, i.e., we
have selected r colors from Ψ+(v). The probability that the next color Rj we consider is different
from these r colors in T is at least
Ψ0(v) ≥ Ψ+(v) − logβ p⋆
Ψ+(v) − r
Ψ0(v)
≥
Ω(∆/ log ∆)
∆ + 1
= Ω(1/ log ∆).
Remember that Ψ+(v) ≥ pv = Ω(∆/ log ∆) and logβ p⋆ = O(logβ ∆). Also remember that
SampleColors gives (∅, ∅) if after we go over all k2 log3 p⋆ = log3+β p⋆ elements in the sequence R,
the size of T is still less than k2 = logβ p⋆. The probability that this event occurs is at most
Pr[Binomial(n′, p′) < t′] < Pr[Binomial(n′, p′) < n′p′/2],
where n′ = log3+β p⋆ = Θ(log3+β ∆), p′ = Ω(1/ log ∆), and t′ = logβ p⋆ = Θ(logβ ∆) ≪ n′p′. By a
Chernoff bound, this event occurs with probability at most exp(−Ω(n′p′)) = exp(−Ω(log2+β ∆)).
Claim 4.6. Pr[Erich
v
] ≤ exp(−Ω(logβ ∆)).
G (v) Su(cid:12)(cid:12)(cid:12) ≥ Tv/3.
automatically rich. Thus, in subsequent discussion we assume v is not lazy. We write N out
Proof. Recall that v is rich if (cid:12)(cid:12)(cid:12)Tv \Su∈N out
{u1, . . . , us} and let Xr = Tv ∩ Sur. To prove the lemma, it suffices to show that for Y =Ps
we have Pr[Y ≥ 2
We consider the random variable Xr = Tv ∩ Sur. For notational simplicity, we write u = ur. If
u is lazy, then Xr = 0. Suppose u is not lazy. The set Su ⊆ Ψ+(u) is the result of randomly choosing
distinct C/2 colors c1, . . . , cC/2 from Ψ+(u), one by one. For each j ∈ [1, C/2], define Zu,j as the
indicator random variable that cj ∈ Tv. Then Xr =PC/2
j=1 Zu,j. We have the following observation.
In the process, when we pick the jth color, regardless of the already chosen colors c1, . . . , cj−1, the
probability that the color picked is in Tj is at most
If v is lazy, then Tv = ∅, so v is
G (v) =
r=1 Xr,
3Tv] ≤ exp(−Ω(logβ p⋆))).
Tv
Tv
E[Zu,j] ≤
Tv
Ψ+(u) − (C/2) ≤
Ψ+(u)−(j−1) ≤
pu − (C/2) ≤
Tv
Ψ+(u)−(C/2) . Thus, we have
1.1Tv
,
pu
since C/2 ≤ (logβ p⋆)/2 = O(poly log ∆) and pu = Ω(∆/ log ∆).
Therefore, in order to bound Y = X1 + ··· + Xs from above, we can assume w.l.o.g. each Xr
is the sum of C/2 i.i.d. random variables, and each of them is a bernoulli random variable with
p = 1.1Tv
, and so Y is the summation of s · (C/2) independent 0-1 random variables. Since v is
G (v) 1/pu ≤ 1/C. The expected value of Y can be upper bounded as
pu
(C, D)-honest, we havePu∈N out
follows.
E[Y ] ≤
By a Chernoff bound, we obtain
PrhErich
v
i ≤ Pr(cid:20)Y ≥(cid:18) 1
1.1 ·
4
3(cid:19)(cid:18) 1.1
1.1
2 Tv.
C
G (u)
1.1Tv
pv ≤
2 Xv∈N out
2 Tv(cid:19)(cid:21) ≤ exp(cid:18)−Ω(cid:18)1.1
2 Tv(cid:19)(cid:19) ≤ exp(cid:16)−Ω(cid:16)logβ p⋆(cid:17)(cid:17) .
Using the above three claims, we now prove the three conditions specified in Lemma 4.2.
22
Proof of i). Conditioning on Erich
of Tv/3 specific colors from Tv ⊆ Ψ+(v). Remember that Erich
Tv/3, and if any one of them is in Sv, then v successfully colors itself. Also remember that Sv is a
size-(C/2) subset of Tv chosen uniformly at random. Thus,
implies that(cid:12)(cid:12)(cid:12)Tv \Su∈N out
, v is lucky with some color unless it fails to select any
G (v) Su(cid:12)(cid:12)(cid:12) ≥
v ∩ Elazy
v
v
Pr[Elucky
v
Erich
v ∩ Elazy
v
≤ exp(−C/6).
3Tv
C/2(cid:1)
] ≤ (cid:0) 2
3(cid:19)C/2
(cid:0)TvC/2(cid:1) ≤(cid:18) 2
By Claim 4.5 and Claim 4.6, we have:
] ≤ Pr[Elucky
≤ exp(−C/6) + exp(−Ω(logβ p⋆)).
v ∩ Elazy
Erich
Pr[Elucky
v
v
v
] + Pr[Elazy
v
] + Pr[Erich
v
]
Proof of ii). This also follows from By Claim 4.5 and Claim 4.6.
Pr[v marks itself Bad] = Pr[Erich
≤ Pr[Erich
v
v ∪ Elazy
v
]
] + Pr[Elazy
v
] ≤ exp(−Ω(logβ p⋆)).
ur
Proof of iii). Define Y as the summation of 1/pu over all vertices u ∈ N out
G (v) such that u /∈ U
in the next iteration. We prove that the probabilities of (a) Y ≤ 1/C′ and (b) v has more than D′
c-significant neighbors u ∈ NG0(v) in this iteration are both at most exp(−Ω(logβ ∆)).
G (v) =
ur ∪ Elazy
{u1, . . . , us}. Consider the event E∗r = Elucky
that ur does not join U in the next
iteration, i.e., ur successfully colors itself or marks itself Bad.
For (b), it follows from Claim 4.4. For the rest of the proof, we deal with (a). Write N out
∪ Erich
For each r ∈ [1, s], define the random variable Zr as follows. Let Zr = 0 if the event Elucky
∪
ur ∪ Elazy
Erich
r=1 Zr. Note that E[Y ] ≤
exp(−C/6) · (1/C), because
Pr[E∗r ] = Pr[Elucky
occurs, and Zr = 1/pur otherwise. Clearly we have Y =Ps
ur ∩ Elazy
as calculated above in the proof of Condition (i). Since v is (C, D)-honest, we havePu∈N out
C′ = min(cid:8) 1
G (v) 1/pu ≤
1/C. Combining these two inequalities, we obtain that E[Y ] ≤ exp(−C/6) · (1/C). Recall that
Next, we prove the desired concentration bound on Y . Each variable Zr is within the range
] ≤ exp(−C/6),
] ≤ Pr[Elucky
ur ∩ Elazy
∩ Erich
Erich
ur
ur
ur
ur
ur
ur
ur
[ar, br], where ar = 0 and br = 1/pur . We have
2 exp(C/6)C, logβ p⋆(cid:9), and so E[Y ] ≤ 1/(2C′).
u ≤ Xu∈N out
G (v)
Recall E[Y ] ≤ 1/(2C′). By Hoeffding's inequality, we obtain
Pr[Y ≥ 1/C′] ≤ exp(cid:18) −2/(2C′)2
(br − ar)2 ≤ Xu∈N out
sXr=1
1/p2
G (v)
r=1(br − ar)2(cid:19) .
Ps
1/(pu · p⋆) ≤ 1/(Cp⋆).
By assumptions specified in the lemma, (1/C′)2 = Ω(1/ log2β p⋆) and 1/Ps
Ω(Cp⋆) = Ω(p⋆/ logβ p⋆). Thus,
Pr[Y ≥ 1/C′] ≤ exp(−Ω(p⋆ log−3β p⋆)) ≤ exp(−Ω(pp⋆)) ≪ exp(−Ω(logβ p⋆)).
r=1(br − ar)2 =
23
There is a subtle issue regarding the applicability of Hoeffding's inequality. The variables
{X1, . . . , Xk} are not independent, but we argue that we are still able to apply Hoeffding's in-
equality. Assume that N out(v) = (u1, . . . , us) is sorted in reverse topological order, and so for each
1 ≤ a ≤ s, we have N out(ua)∩{ua, . . . , us} = ∅. We reveal the random bits in the following manner.
First of all, we reveal the set Tu for all vertices u. Now the event regarding whether a vertex is rich or
is lazy has been determined. Then, for r = 1 to s, we reveal the set {Su u = ur or u ∈ N out(ur)}.
This information is enough for us to decide the outcome of Zr. Note that in this process, condi-
tioning on arbitrary outcome of Z1, . . . , Zr−1 and all random bits revealed prior to revealing the set
Sur , The probability that E∗r occurs is still at most exp(−C/6).
4.3 Implementation for the Sparsified Color Bidding Algorithm
In this section, we present an implementation of SparsifiedColoring in the LOCAL model such
that after an O(1)-round pre-processing step, each vertex v is able to identify a O(poly log ∆)-size
subset N∗(v) ⊆ NG0(v) of neighboring vertices such that v only needs to receive information from
these vertices during SparsifiedColoring.
Fixing All Random Bits.
Instead of having each vertex v generate the color sequence R(i)
v at it-
eration i, we determined all of {R(i)
v }0≤i≤k−1 in the pre-processing step. After fixing these sequences,
we can regard SparsifiedColoring as a deterministic LOCAL algorithm, where {R(i)
v }v∈V0, 0≤i≤k−1
can be seen as the input for the algorithm. To gather this information, we need to use messages of
k · K · O(log n) = O(poly log ∆) · O(log n) bits, where k = O(log∗ ∆) is the number of iterations,
and K is the length of the color sequence R(i)
v
for an iteration.
Determining the Set N∗(v). We show how to let each vertex v determine a O(poly log ∆)-size
set N∗(v) ⊆ NG0(v) based on the following information
{R(i)
u }u∈NG0 (v), 0≤i≤k−1.
such that v only needs to receive messages from N∗(v) during the execution of SparsifiedColoring.
We make the following two observations.
1. In order for v to execute SparsifiedColorBidding at iteration i correctly, v does not need to
u }u∈NG0 (v), 0≤i′≤i do not overlap with
receive information from u ∈ NG0(v) if all colors in {R(i′)
the colors in R(i)
u . In other words, v only needs information from its significant neighbors.
2. If v is overloaded at iteration i, then v knows that it is lazy in this iteration, and so the
outcome of SparsifiedColorBidding at iteration i is that v sets Sv = Tv = ∅, and v marks
itself Bad.
The above two observations follow straightforwardly from the description of SparsifiedColoring.
Therefore, we can define the set N∗(v) as follows. Add u ∈ NG0(v) to N∗(v) if there exists an index
i ∈ [0, k − 1] such that (i) u is a significant neighbor of v at iteration i, and (ii) v is not overloaded
at iteration i.
By the definition of overloaded vertices, we know that if v is not overloaded at iteration i,
then v has at most K 2 log ∆ = O(poly log ∆) significant neighbors for iteration i. Thus, N∗(v) =
O(poly log ∆). Note that the set N∗(v) can be locally calculated at v during the pre-processing
step.
24
Summary. Algorithm SparsifiedColoring can be implemented in LOCAL in the following way.
Pre-Processing Step. This step is randomized, and it takes one round and uses messages of
O(poly log ∆) · O(log n) bits. After this step, each vertex has calculated a set N∗(v) with
N∗(v) = O(poly log ∆).
Main Steps. This is a deterministic O(log∗ ∆)-round procedure. During the procedure, each ver-
tex v only receives messages from N∗(v). The output of each vertex is a color (or a special
symbol ⊥ indicating that v is uncolored), which can be represented by ℓout = O(log n) buts.
The input of each vertex consists of its color sequences for all iterations, which can be repre-
sented in ℓin = O(poly log ∆) · O(log n) bits.
Using the above implementation of SparsifiedColoring, we show that there is an LCA that
solves (∆ + 1)-list coloring with ∆O(1) · O(log n) queries.
Proof of Theorem 1.3. Consider the following algorithm for solving (∆ + 1)-list coloring.
1. Run the O(1)-round algorithm of Lemma 4.1. After that, each vertex v has four possible
status: (i) v has been colored, (ii) v is in VGood, (iii) v is in VBad, or v is in R. This can be
done with ∆O(1) queries.
2. The set R induces a subgraph with constant maximum degree. The LCA for (deg +1)-list
coloring in [FHK16] implies that each v ∈ R only needs O(log∗ n) queries of vertices in R to
compute its color.
3. By Lemma 4.1, each connected component in VBad has size ∆O(1)·O(log n). We let each vertex
v ∈ VBad learns the component S it belongs to, and apply a deterministic algorithm to color
S.
4. All vertices in VGood run the algorithm SparsifiedColoring. This adds an (∆ · ∆k
∗)-factor
in the query complexity, where ∆∗ = maxv∈V0 N∗(v) = O(poly log ∆), and k = O(log∗ ∆) is
the number of iterations of SparsifiedColoring. Note that in SparsifiedColoring, when
we query a vertex v, the set N∗(v) can be calculated from the random bits in NG0(v) ∪ {v}.
5. By Lemma 4.3 and Lemma 2.2, the vertices left uncolored after SparsifiedColoring induces
connected components of size ∆O(1) · O(log n). Similarly, we let each uncolored vertex v learns
the component S it belongs to, and apply a deterministic algorithm to color S.
By the standard procedure for converting an LOCAL algorithm to an LCA, it is straightforward to
implement the above algorithm as an LCA with query complexity
∆O(1) · ∆k
∗ ·(cid:16)∆O(1) · O(log n)(cid:17) = ∆O(1) · O(log n).
Next, we show that by applying SparsifiedColoring with the speedup lemma (Lemma 2.3),
we can solve (∆ + 1)-list coloring in CONGESTED-CLIQUE in O(1) rounds when ∆ = O(poly log n).
Theorem 4.7. Suppose ∆ = O(poly log n). There is an algorithm that solves (∆ + 1)-list coloring
in CONGESTED-CLIQUE in O(1) rounds.
25
Proof. Recall from Lemma 2.1 that one round in LOCAL with messages of at most O(n log n) bits
can be simulated in O(1) rounds in CONGESTED-CLIQUE.
The first step of the algorithm is to run the black box algorithm for Lemma 4.1, which takes
O(1) rounds in LOCAL with messages of size O(∆2 log n) ≪ O(n log n). The set R trivially induces
a subgraph with O(n) edges. By Lemma 2.2, VBad induces a subgraph with O(n) edges. We use
Lemma 2.1 to color them in O(1) rounds.
Now we focus on VGood. We execute the algorithm SparsifiedColoring using the above
implementation. The pre-processing step takes O(1) rounds in LOCAL with messages of size
O(poly log ∆)·O(log n) ≪ O(n log n). For the main steps, we apply the speedup lemma (Lemma 2.3)
with τ = O(log∗ ∆), ℓout = O(log n), ℓin = O(poly log ∆) · O(log n), and ∆∗ = O(poly log ∆). Since
we assume ∆ = O(poly log n), this satisfies the criterion for Lemma 2.3, and so this procedure can
be executed on CONGESTED-CLIQUE in O(1) rounds.
The algorithm SparsifiedColoring does not color all vertices in VGood. However, by Lemma 4.3
and Lemma 2.2, we know that these uncolored vertices induces a subgraph with O(n) edges. We
use Lemma 2.1 to color them in O(1) rounds.
References
[ABB+19] Sepehr Assadi, MohammadHossein Bateni, Aaron Bernstein, Vahab Mirrokni, and Cliff Stein.
Coresets meet edcs: algorithms for matching and vertex cover on massive graphs. In Proceedings
30th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), 2019.
[ACK19] Sepehr Assadi, Yu Chen, and Sanjeev Khanna. Sublinear Algorithms for (∆ + 1) Vertex
Coloring. In Proceedings 30th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA),
2019.
[AG15] Kook Jin Ahn and Sudipto Guha. Access to data and number of iterations: Dual primal
algorithms for maximum matching under resource constraints. In Proc. SPAA, pages 202 -- 211,
2015.
[ANOY14] Alexandr Andoni, Aleksandar Nikolov, Krzysztof Onak, and Grigory Yaroslavtsev. Parallel
In Proc. Symposium on Theory of Computation
algorithms for geometric graph problems.
(STOC), pages 574 -- 583, 2014.
[Ass17] Sepehr Assadi.
Simple round compression for parallel vertex cover.
arXiv preprint
arXiv:1709.04599, 2017.
[ASS+18] Alexandr Andoni, Clifford Stein, Zhao Song, Zhengyu Wang, and Peilin Zhong. Parallel graph
connectivity in log diameter rounds. In Proceedings 59th IEEE Symposium on Foundations of
Computer Science (FOCS), pages 674 -- 685, 2018.
[ASW18] Sepehr Assadi, Xiaorui Sun, and Omri Weinstein. Massively parallel algorithms for finding
well-connected components in sparse graphs. arXiv preprint arXiv:1805.02974, 2018.
[BEG18a] Leonid Barenboim, Michael Elkin, and Uri Goldenberg. Locally-iterative distributed (∆+
1)-coloring below Szegedy-Vishwanathan barrier, and applications to self-stabilization and to
restricted-bandwidth models.
In Proceedings of the 2018 ACM Symposium on Principles of
Distributed Computing, pages 437 -- 446. ACM, 2018.
[BEG+18b] Mahdi Boroujeni, Soheil Ehsani, Mohammad Ghodsi, MohammadTaghi HajiAghayi, and Saeed
Seddighin. Approximating edit distance in truly subquadratic time: quantum and MapReduce.
In Proc. Symposium on Discrete Algorithms (SODA), pages 1170 -- 1189, 2018.
26
[BEPS16] Leonid Barenboim, Michael Elkin, Seth Pettie, and Johannes Schneider. The locality of dis-
tributed symmetry breaking. J. ACM, 63(3):20:1 -- 20:45, 2016.
[BFU18a] Sebastian Brandt, Manuela Fischer, and Jara Uitto. Breaking the Linear-Memory Barrier in
MPC: Fast MIS on Trees with nǫ Memory per Machine. arXiv preprint arXiv:1802.06748, 2018.
[BFU18b] Sebastian Brandt, Manuela Fischer, and Jara Uitto. Matching and MIS for uniformly sparse
graphs in MPC with low memory. manuscript, 2018.
[BHP12] Andrew Berns, James Hegeman, and Sriram V Pemmaraju. Super-fast distributed algo-
rithms for metric facility location. In Automata, Languages, and Programming, pages 428 -- 439.
Springer, 2012.
[BK18] Leonid Barenboim and Victor Khazanov. Distributed symmetry-breaking algorithms for con-
gested cliques. In Computer Science - Theory and Applications - 13th International Computer
Science Symposium in Russia, CSR 2018, Moscow, Russia, June 6-10, 2018, Proceedings, pages
41 -- 52, 2018.
[BKKL17] Ruben Becker, Andreas Karrenbauer, Sebastian Krinninger, and Christoph Lenzen. Near-
Optimal Approximate Shortest Paths and Transshipment in Distributed and Streaming Models.
In Andréa W. Richa, editor, 31st International Symposium on Distributed Computing (DISC
2017), volume 91 of Leibniz International Proceedings in Informatics (LIPIcs), pages 7:1 -- 7:16.
Schloss Dagstuhl -- Leibniz-Zentrum fuer Informatik, 2017.
[BKS13] Paul Beame, Paraschos Koutris, and Dan Suciu. Communication steps for parallel query pro-
cessing. In Proceedings of the 32Nd ACM SIGMOD-SIGACT-SIGAI Symposium on Principles
of Database Systems (PODS), pages 273 -- 284, 2013.
[BKS14] Paul Beame, Paraschos Koutris, and Dan Suciu. Skew in parallel query processing. In Pro-
ceedings of the 33rd ACM SIGMOD-SIGACT-SIGART Symposium on Principles of Database
Systems (PODS), pages 212 -- 223, 2014.
[CHKK+16] Keren Censor-Hillel, Petteri Kaski, Janne H. Korhonen, Christoph Lenzen, Ami Paz, and Jukka
Suomela. Algebraic methods in the congested clique. Distributed Computing, 2016.
[CLM+18] Artur Czumaj, Jakub Lacki, Aleksander Madry, Slobodan Mitrovic, Krzysztof Onak, and Piotr
In Proc. Symposium on
Sankowski. Round compression for parallel matching algorithms.
Theory of Computation (STOC), pages 471 -- 484, 2018.
[CLP18] Yi-Jun Chang, Wenzheng Li, and Seth Pettie. An optimal distributed (∆ + 1)-coloring algo-
rithm? In Proceedings of the 50th Annual ACM SIGACT Symposium on Theory of Computing,
STOC 2018, pages 445 -- 456, New York, NY, USA, 2018. ACM.
[CLT18] Keren Censor-Hillel, Dean Leitersdorf, and Elia Turner. Sparse matrix multiplication with
bandwidth restricted all-to-all communication. CoRR, abs/1802.04789, 2018.
[DG04] Jeffrey Dean and Sanjay Ghemawat. MapReduce: Simplified data processing on large clus-
In Proceedings of the 6th Conference on Symposium on Operating Systems Design &
ters.
Implementation (OSDI), pages 10 -- 10, Berkeley, CA, USA, 2004. USENIX Association.
[DKO14] Andrew Drucker, Fabian Kuhn, and Rotem Oshman. On the power of the congested clique
model. In Proc. Principles of Distributed Computing (PODC), pages 367 -- 376. ACM, 2014.
[DLP12] Danny Dolev, Christoph Lenzen, and Shir Peled. "tri, tri again": Finding triangles and small
subgraphs in a distributed setting. In Distributed Computing, pages 195 -- 209. Springer, 2012.
[DP09] Devdatt P. Dubhashi and Alessandro Panconesi. Concentration of Measure for the Analysis of
Randomized Algorithms. Cambridge University Press, 2009.
27
[DR98] Devdatt P. Dubhashi and Desh Ranjan. Balls and bins: A study in negative dependence.
J. Random Structures and Algs., 13(2):99 -- 124, 1998.
[FG17] Manuela Fischer and Mohsen Ghaffari. Sublogarithmic distributed algorithms for Lovász local
lemma with implications on complexity hierarchies. In Proceedings 31st International Sympo-
sium on Distributed Computing (DISC), pages 18:1 -- 18:16, 2017.
[FHK16] Pierre Fraigniaud, Marc Heinrich, and Adrian Kosowski. Local conflict coloring. In Proceedings
57th Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 625 -- 634,
2016.
[Gal16] Francois Le Gall. Further algebraic algorithms in the congested clique model and applications
to graph-theoretic problems. In DISC, 2016.
[GGK+18] Mohsen Ghaffari, Themis Gouleakis, Christian Konrad, Slobodan Mitrović, and Ronitt Rubin-
feld. Improved massively parallel computation algorithms for mis, matching, and vertex cover.
In Proc. Principles of Distributed Computing (PODC). arXiv:1802.08237, 2018.
[Gha16] Mohsen Ghaffari. An improved distributed algorithm for maximal independent set. In Pro-
ceedings 27th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 270 -- 277,
2016.
[Gha17] Mohsen Ghaffari. Distributed MIS via all-to-all communication. In Proc. Principles of Dis-
tributed Computing (PODC), pages 141 -- 149, 2017.
[GSZ11] Michael T. Goodrich, Nodari Sitchinava, and Qin Zhang. Sorting, searching, and simulation in
the MapReduce framework. In Proc. ISAAC, pages 374 -- 383. Springer, 2011.
[GU19] Mohsen Ghaffari and Jara Uitto. Sparsifying Distributed Algorithms with Ramifications in
Massively Parallel Computation and Centralized Local Computation. In Proceedings 30th An-
nual ACM-SIAM Symposium on Discrete Algorithms (SODA), 2019.
[HLL18] Nicholas J. A. Harvey, Christopher Liaw, and Paul Liu. Greedy and local ratio algorithms in
the MapReduce model. In Proceedings of the 30th on Symposium on Parallelism in Algorithms
and Architectures (SPAA), pages 43 -- 52, New York, NY, USA, 2018. ACM.
[Hoe63] Wassily Hoeffding. Probability inequalities for sums of bounded random variables. Journal of
the American Statistical Association, 58(301):13 -- 30, 1963.
[HP15] James W Hegeman and Sriram V Pemmaraju. Lessons from the congested clique applied to
MapReduce. Theoretical Computer Science, 608:268 -- 281, 2015.
[HPP+15] James W. Hegeman, Gopal Pandurangan, Sriram V. Pemmaraju, Vivek B. Sardeshmukh, and
Michele Scquizzato. Toward optimal bounds in the congested clique: Graph connectivity and
MST. In Proc. Principles of Distributed Computing (PODC), pages 91 -- 100. ACM, 2015.
[HPS14] James W Hegeman, Sriram V Pemmaraju, and Vivek B Sardeshmukh. Near-constant-time dis-
tributed algorithms on a congested clique. In Distributed Computing, pages 514 -- 530. Springer,
2014.
[HSS18] David G Harris, Johannes Schneider, and Hsin-Hao Su. Distributed (∆ + 1)-coloring in sublog-
arithmic rounds. J. ACM, 65(4):19:1 -- 19:21, 2018.
[IBY+07] Michael Isard, Mihai Budiu, Yuan Yu, Andrew Birrell, and Dennis Fetterly. Dryad: Distributed
data-parallel programs from sequential building blocks. SIGOPS Operating Systems Review,
41(3):59 -- 72, 2007.
28
[IMS17] Sungjin Im, Benjamin Moseley, and Xiaorui Sun. Efficient massively parallel methods for
dynamic programming. In Proc. Symposium on Theory of Computation (STOC), pages 798 --
811, 2017.
[JN18] Tomasz Jurdzinski and Krzysztof Nowicki. MST in O(1) rounds of congested clique. In Pro-
ceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2018, New Orleans, LA, USA, January 7-10, 2018, pages 2620 -- 2632, 2018.
[KSV10] Howard J. Karloff, Siddharth Suri, and Sergei Vassilvitskii. A model of computation for MapRe-
duce. In Proc. Symposium on Discrete Algorithms (SODA), pages 938 -- 948, 2010.
[Len13] Christoph Lenzen. Optimal deterministic routing and sorting on the congested clique.
In
Proceedings 33rd ACM Symposium on Principles of Distributed Computing (PODC), pages
42 -- 50, 2013.
[Lin92] Nathan Linial. Locality in distributed graph algorithms. SIAM J. Comput., 21(1):193 -- 201,
1992.
[LM17] Reut Levi and Moti Medina. A (centralized) local guide. Bulletin of EATCS, 2(122), 2017.
[LMSV11] Silvio Lattanzi, Benjamin Moseley, Siddharth Suri, and Sergei Vassilvitskii. Filtering: a method
for solving graph problems in MapReduce. In Proc. SPAA, pages 85 -- 94, 2011.
[LPSPP05] Zvi Lotker, Boaz Patt-Shamir, Elan Pavlov, and David Peleg. Minimum-weight spanning tree
construction in O(log log n) communication rounds. SIAM Journal on Computing, 35(1):120 --
131, 2005.
[LW10] Christoph Lenzen and Roger Wattenhofer. Brief announcement: Exponential speed-up of local
algorithms using non-local communication. In Proceedings of the 29th ACM SIGACT-SIGOPS
Symposium on Principles of Distributed Computing (PODC), pages 295 -- 296. ACM, 2010.
[Nan14] Danupon Nanongkai. Distributed approximation algorithms for weighted shortest paths. In
Proc. Symposium on Theory of Computation (STOC), 2014.
[Par18] Merav Parter. (∆+1) coloring in the congested clique model. In Proceedings of the International
Colloquium on Automata, Languages and Programming (ICALP), 2018.
[Pel00] David Peleg. Distributed Computing: A Locality-Sensitive Approach. SIAM, 2000.
[Pem01] Sriram V. Pemmaraju. Equitable coloring extends chernoff-hoeffding bounds. In Michel Goe-
mans, Klaus Jansen, José D. P. Rolim, and Luca Trevisan, editors, Approximation, Random-
ization, and Combinatorial Optimization: Algorithms and Techniques, pages 285 -- 296, Berlin,
Heidelberg, 2001. Springer Berlin Heidelberg.
[PR07] Michal Parnas and Dana Ron. Approximating the minimum vertex cover in sublinear time and
a connection to distributed algorithms. Theoretical Computer Science, 381(1):183 -- 196, 2007.
[PS96] Alessandro Panconesi and Aravind Srinivasan. On the complexity of distributed network de-
composition. J. Algor., 20(2):356 -- 374, 1996.
[PS18] Merav Parter and Hsin-Hao Su. (∆ + 1) coloring in O(log∗ ∆) congested-clique rounds.
In
Proceedings of the International Symposium on Distributed Computing (DISC), 2018.
[PST11] Boaz Patt-Shamir and Marat Teplitsky. The round complexity of distributed sorting.
In
Proc. Principles of Distributed Computing (PODC), pages 249 -- 256, 2011.
[RTVX11] Ronitt Rubinfeld, Gil Tamir, Shai Vardi, and Ning Xie. Fast local computation algorithms. In
Proceedings of the First Symposium on Innovations in Computer Science (ICS), pages 223 -- 238,
2011. See also CoRR abs/1104.1377.
29
[RVW16] Tim Roughgarden, Sergei Vassilvitskii, and Joshua R. Wang. Shuffles and circuits: (on lower
bounds for modern parallel computation). In Proc. SPAA, pages 1 -- 12, 2016.
[SSS95] Jeanette P. Schmidt, Alan Siegel, and Aravind Srinivasan. Chernoff-Hoeffding bounds for
applications with limited independence. SIAM Journal on Discrete Mathematics, 8(2):223 --
250, 1995.
[Whi12] Tom White. Hadoop: The Definitive Guide. O'Reilly Media, Inc., 2012.
[ZCF+10] Matei Zaharia, Mosharaf Chowdhury, Michael J. Franklin, Scott Shenker, and Ion Stoica.
In 2nd USENIX Workshop on Hot Topics in
Spark: Cluster computing with working sets.
Cloud Computing (HotCloud), 2010.
A Probabilistic Tools
In this section we review some probabilistic tools used in this paper.
Chernoff Bound. Let X be the summation of n independent 0-1 random variables with mean p.
Multiplicative Chernoff bounds give the following tail bound of X with mean µ = np.
Pr[X ≥ (1 + δ)µ] ≤(exp(−δ2µ
)
3
exp(−δµ
3 )
if δ ∈ [0, 1]
if δ > 1.
Note that these bounds hold even when X is the summation of n negatively correlated 0-1 random
variables [DR98, DP09] with mean p, i.e., total independent is not required. These bounds also
hold when µ > np is an overestimate of E[X].
Chernoff Bound with k-wise Independence. Suppose X is the summation of n k-wise in-
dependent 0-1 random variables with mean p. We have µ ≥ E[X] = np and the following tail
bound [SSS95].
In particular, when k = Ω(δ2µ), we obtain the same asymptotic tail bound as that of Chernoff
bound with total independence.
Pr[X ≥ (1 + δ)µ] ≤ exp(cid:0)− min{k, δ2µ}(cid:1) .
Chernoff Bound with Bounded Independence. Suppose X is the summation of n indepen-
i=1 Xi.
dent 0-1 random variables with bounded dependency d, and let µ ≥ E[X], where X = Pn
Then we have [Pem01]:
Pr[X ≥ (1 + δ)µ] ≤ O(d) · exp(−Ω(δ2µ/d)).
i=1 Xi, and each Xi is an indepen-
dent random variable bounded by the interval [ai, bi]. Let µ ≥ E[X]. Then we have the following
concentration bound [Hoe63].
Hoeffding's Inequality. Consider the scenario where X =Pn
Pr[X ≥ (1 + δ)µ] ≤ exp(cid:18) −2(δµ)2
i=1(bi − ai)2(cid:19) .
Pn
30
B Proof of the Shattering Lemma
In this section, we prove the shattering lemma (Lemma 2.2). Let c ≥ 1. Consider a randomized
procedure that generates a subset of vertices B ⊆ V . Suppose that for each v ∈ V , we have
Pr[v ∈ B] ≤ ∆−3c, and this holds even if the random bits not in N c(v) are determined adversarially.
Then, the following is true.
1. With probability at least 1 − n−Ω(c′), each connected component in the graph induced by B
has size at most (c′/c)∆2c log∆ n.
2. With probability 1 − O(∆c) · exp(−Ω(n∆−c)), the number of edges induced by B is O(n).
Proof. Statement (1) is well-known; see e.g., [BEPS16, FG17]. Here we provide a proof for State-
ment (2). For each edge e = {u, v}, write Xe to be the indicator random variable such that
Xe = 1 if u ∈ B and v ∈ B. Let X = Pe∈E Xe.
It is clear that Pr[Xe] ≤ 2∆−3c, and so
µ = E[X] ≤ n∆1−3c ≪ n. By a Chernoff bound with bounded dependence d = 2∆c the probability
that X > n is O(∆c) · exp(−Ω(n∆−c)).
C Fast Simulation of LOCAL Algorithms in CONGESTED-CLIQUE
In this section, we prove Lemma 2.3. Let A be a τ -round LOCAL algorithm on G = (V, E).
We show that there is an O(1)-round simulation of A in in CONGESTED-CLIQUE, given that (i)
∆τ
∗ log(∆∗ + ℓin/ log n) = O(log n), (ii) ℓin = O(n), and (iii) ℓout = O(log n).
Proof. Assume A is in the following canonical form. Each vertex first generates certain amount
of local random bits, and then collects all information in its τ -neighborhood. The information
includes not only the graph topology, but also IDs, inputs, and the random bits of these vertices.
After gathering this information, each vertex locally computes its output based on the information
it gathered.
Consider the following procedure in CONGESTED-CLIQUE for simulating A. In the first phase,
for each ordered vertex pair (u, v), with probability p to be determined, u sends all its local infor-
mation to v. The local information can be encoded in Θ(∆∗ log n + ℓin) bits. This includes the local
input of u, the local random bits needed for u to run A, and the list of IDs in N∗(u) ∪ {u}. In the
second phase, for each ordered vertex pair (u, v), if v has gathered all the required information to
calculate the output of A at u, then v sends to u the output of A at u.
At first sight, the procedure seems to take ω(1) rounds because of the O(log n)-bit message size
constraint of CONGESTED-CLIQUE. However, if we set p = Θ(cid:16)
1
∆∗+ℓin/ log n(cid:17), the expected number
of O(log n)-bit messages sent from or received by a vertex is np · Θ(∆∗ + ℓin/ log n) = O(n).
More precisely, let Xu be the number vertices v ∈ V to which u sends its local information in
the first phase; similarly, let Yv be the number of vertices u ∈ V sending their local information to
a v. We have E[Xu] = np, for each u ∈ V , and E[Yv] = np, for each v ∈ V . By a Chernoff bound,
so long as np = Ω(log n), with probability 1 − exp(−Ω(np)) = 1/poly(n), we have Xu = O(np), for
each u ∈ V , and Yv = O(np), for each v ∈ V . That is, the number of O(log n)-bit messages sent
from or received by a vertex is at most np · Θ(∆∗ + ℓin/ log n) = O(n), w.h.p.
We verify that np = Ω(log n). Condition (i) implicitly requires ∆∗ = O(log n), and Condition
(ii) requires ℓin = O(n). Therefore, np = Θ(cid:16)
n
∆+ℓin/ log n(cid:17) = Ω(log n). Thus, we can route all
messages in O(1) rounds using Lenzen's routing (Lemma 2.1), and so the first phase can be done
in O(1) rounds.
31
Denote Eu,v as the event that v ∈ V that receives messages from all vertices in N τ
Condition (iii) guarantees that ℓout = O(log n), and so the messages in the second phase can be
sent directly in O(1) rounds. What remains to do is to show that for each u ∈ V , w.h.p., there is
a vertex v ∈ V that receives messages from all vertices in N τ
∗ (u) during the first phase, and so v is
able to calculate the output of u locally.
∗ (u) during
the first phase, and denote Eu as the event that at least one of {Eu,v v ∈ V } occurs. We have
∗. Thus, Pr[Eu] ≥ 1 − (1 − p∆τ
Pr[Eu,v] ≥ p∆τ
Condition (i) guarantees that ∆τ log(∆∗ + ℓin/ log n) = O(log n). By setting p = ǫ/(∆∗ +
ℓin/ log n) for some sufficiently small constant ǫ, we haves ∆τ
2 log n, and it implies
∗ )n = 1 − exp(−Ω(√n)). Thus, the simulation gives
∗ ≥ 1/√n. Therefore, Pr[Eu] ≥ 1 − (1 − p∆τ
p∆τ
the correct output for all vertices w.h.p.
∗ log p ≥ − 1
∗ , since N τ
∗ (u) ≤ ∆τ
∗ )n.
We remark that the purpose of the condition ℓout = O(log n) is only to allow the messages in the
second phase to be sent directly in O(1) rounds. With a more careful analysis and using Lenzen's
routing , the condition ℓout = O(log n) can be relaxed to ℓout = O(n), though in our application we
only need ℓout = O(log n).
D Proof of Lemma 4.1
In this section, we briefly review the algorithm of [CLP18] and show how to obtain Lemma 4.1
from [CLP18]. The algorithm uses a sparsity sequence defined by ǫ1 = ∆−1/10, ǫi = √ǫi−1 for i > 1,
and ℓ = Θ(log log ∆) is the largest index such that 1
ǫℓ ≥ K for some sufficiently large constant K.
The algorithm first do an O(1)-round procedure (initial coloring step) to color a fraction of the
vertex set V , and denote V ⋆ as the set of remaining uncolored vertices. The set V ⋆ is decomposed
into ℓ + 1 subsets (V1, . . . , Vℓ, Vsp) according to local sparsity. The algorithm then applies another
O(1)-round procedure (dense coloring step) to color a fraction of vertices in V1 ∪ ··· ∪ Vℓ.
The remaining uncolored vertices in V ⋆ after the above procedure (initial coloring step and
dense coloring step) are partitioned into three subsets: U , R, and Vbad.13 The set R induces a
constant-degree graph. The set Vbad satisfies the property that each vertex is added to Vbad with
probability ∆−Ω(c), where c can be any given constant, independent on the runtime. The vertices
in U satisfy the following properties.
Excess Colors: We have V1 ∩ U = ∅. Each v ∈ Vi ∩ U , with i > 1, has Ω(ǫ2
i−1∆) excess colors.
Each v ∈ Vsp ∩ U has Ω(ǫ2
ℓ ∆) = Ω(∆) excess colors. The number of excess colors at a vertex
v is defined by the number of available colors of v minus the number of uncolored neighbors
of v.
Number of Neighbors: For each v ∈ U , and for each i ∈ [2, ℓ], the number of uncolored neighbors
i−1∆). The number of uncolored neighbors of v in Vsp ∩ U is
of v in Vi ∩ U is O(ǫ5
of course at most ∆ = O(ǫ2.5
i ∆) = O(ǫ2.5
ℓ ∆), since ǫℓ is a constant.
At this moment, the two sets Vbad and R satisfy the required condition specified in Lemma 4.1.
In what follows, we focus on U .
Orientation. We orient the graph induced by the uncolored vertices in U as follows. For any
edge {u, v}, we orient it as (u, v) if one of the following is true: (i) u ∈ Vsp but v /∈ Vsp, (ii) u ∈ Vi
13The algorithm in [CLP18] for coloring layer-1 large blocks has two alternatives. Here we always use the one that
puts the remaining uncolored vertices in one of R or Vbad, where each vertex is added to Vbad with probability ∆−Ω(c).
32
and v ∈ Vj with i > j, (iii) u and v are within the same part in the partition V ⋆ = V1 ∪ . . . Vℓ ∪ Vsp
and ID(v) < ID(u). This results in a directed acyclic graph. We write N out(v) to denote the set of
out-neighbors of v in this graph.
Lower Bound of Excess Colors.
In view of the above, there exist universal constants η > 0
and C > 0 such that the following is true. For each i ∈ [2, ℓ] and each uncolored vertex v ∈ Vi\ Vbad,
we set pv = ηǫ2
ℓ ∆. By selecting a sufficiently small η,
the number pv is always a lower bound on the number of excess colors at v.
i−1∆. For each v ∈ Vsp \ Vbad, we set pv = ηǫ2
j=2 O(ǫ2.5
).
ℓ
ℓ ∆), since ǫℓ is a constant.
number of excess colors in each out-neighbor.
The Number of Excess Colors is Large. Recall that to color the graph quickly we need the
number of excess colors to be sufficiently large with respect to out-degree.
If v ∈ Vi ∩ U with
i−1∆). In this case, pv/N out(v) = Ω(ǫ−0.5
i−1 ).
In this case,
i ≥ 2, it satisfies N out(v) =Pi
j−1∆) = O(ǫ2.5
If v ∈ Vsp ∩ U , then of course N out(v) ≤ ∆ = O(ǫ2
pv/N out(v) = Ω(ǫ−0.5
However, due to the high variation on the palette size in our setting, pv/N out(v) is not a good
measurement for the gap between the number of excess colors and out-degree at v. The inverse of
the expressionPu∈N out(v) 1/pu turns out to be a better measurement, as it takes into account the
There is a constant C > 0 such that for each uncolored vertex v ∈ V ⋆ \ (Vbad ∪ R), we have
Pu∈N out(v) 1/pu ≤ 1/C. The calculation is as follows.
iXj=2
ℓ+1Xj=2
then Xu∈N out(v)
then Xu∈N out(v)
If v ∈ Vi ∩ U (i > 1),
j−1∆
ǫ2
O ǫ2.5
O ǫ2.5
j−1∆! =
j−1∆! =
j−1∆
ǫ2
j−1) = O(ǫ0.5
i−1) < 1/C.
iXj=2
ℓ+1Xj=2
O(ǫ0.5
j−1) = O(ǫ0.5
ℓ ) < 1/C.
If v ∈ Vsp ∩ U ,
1/pu =
1/pu =
O(ǫ0.5
For a specific example, if v is an uncolored vertex in V2\Vbad, then pv = ηǫ2
bound on the number of excess colors at v, and v has out-degree N out(v) = O(ǫ2.5
and we have Pu∈N out(v) 1/pu = O(ǫ0.5
between the number of excess colors and the out-degree at v is Ω(∆0.05).
1∆ = η∆0.8 is the lower
1 ∆) = O(∆0.75),
1 ) = O(∆−0.05) < 1/C. Intuitively, this means that the gap
j ∆), but the number of out-neighbors of v is at most O(ǫ2.5
Summary. Currently the graph induced by U satisfies the following conditions. Each vertex v is
associated with a parameter pv = ηǫ2
j ∆ (for some j ∈ [1, ℓ]) such that the number of excess colors at
v is at least pv = Ω(ǫ2
j ∆). In particular,
we always havePu∈N out(v) 1/pu = O(ǫ0.5
j ) < 1/C, where C > 0 is a universal constant. The current
pv-values for vertices in U almost satisfy the required condition for VGood specified in Lemma 4.1.
Lower Bound of p⋆. Define p⋆ as the minimum pv-value among all uncolored vertices v ∈ V ⋆.
1∆ = η∆0.8, but in Lemma 4.1 it is required that p⋆ ≥
Currently we only have p⋆ ≥ ηǫ2
∆/ log ∆.
Lower Bound of C. Currently we have Pu∈N out(v) 1/pu ≤ 1/C for some universal constant C,
For the rest of the section, we show that there is an O(1)-round that is able to improve the
lower bound of p⋆ to p⋆ ≥ ∆/ log ∆ and increase the parameter C to any specified constant. The
procedure will colors a fraction of vertices in U and puts some vertices in U to the set Vbad. We
but in Lemma 4.1 it is required that C > 0 can be any given constant.
33
first consider improving the lower bound of p⋆. This is done by letting all vertices whose pv-
value are too small (i.e., less than ∆/ log ∆) to jointly run Lemma D.1. For these vertices, we
have Pu∈N out(v) 1/pu ≤ O(cid:16)log−1/4 ∆(cid:17),14 and so we can use C = Ω(cid:16)log1/4 ∆(cid:17) in Lemma D.1.
The algorithm of Lemma D.1 takes only O(1) rounds. All participating vertices that still remain
uncolored join Vbad.
Lemma D.1 ([CLP18]). Consider a directed acyclic graph, where vertex v is associated with a
parameter pv ≤ Ψ(v)− deg(v) We write p⋆ = minv∈V pv. Suppose that there is a number C = Ω(1)
such that all vertices v satisfy Pu∈N out(v) 1/pu ≤ 1/C. Let d⋆ be the maximum out-degree of the
graph. There is an O(log∗(p⋆) − log∗(C))-time algorithm achieving the following. Each vertex v
remains uncolored with probability at most exp(−Ω(√p⋆)) + d⋆ exp(−Ω(p⋆)). This is true even if
the random bits generated outside a constant radius around v are determined adversarially.
Now the lower bound on p⋆ is met. We show how to increase the C-value to any given constant
we like in O(1) rounds. We apply Lemma D.2 using the current p⋆ and C. After that, we can set
the new C-value to be C′ = C · exp(C/6)/(1 + λ), after putting each vertex v not meeting the
following condition to Vbad:
Sum of 1/pu over all remaining uncolored vertices u in N out(v) is at most 1/C′ =
1 + λ
exp(C/6)C
.
If λ is chosen as a small enough constant, we have C′ > C. After a constant number of iterations,
we can increase the C-value to any constant we like. Now, all conditions in Lemma 4.1 are met for
the three sets R, Vbad, and VGood ← U .
Lemma D.2 ([CLP18]). There is an one-round algorithm meeting the following conditions. Let v
be any vertex. Let d be the summation of 1/pu over all vertices u in N out(v) that remain uncolored
after the algorithm. Then the following holds.
Pr(cid:20)d ≥
1 + λ
exp(C/6)C(cid:21) ≤ exp(cid:0)−2λ2p⋆ exp(−C/3)/C(cid:1) + d⋆ exp(−Ω(p⋆)).
E The CLP Algorithm in the Low-Memory MPC Model
In this section, we show which changes have to be made to [CLP18] to get a low-memory MPC
algorithm, thus proving Lemma 3.4.
There are two main issues in the low-memory MPC model that we need to take care of. First,
the total memory of the system is limited to Θ(m + n), where m and n are the number of edges
and vertices in the input graph, respectively. Second, the local memory per machine is restricted
to O(nα), for an arbitrary constant α > 0. These two restrictions force us to be careful about the
amount of information sent between the machines. In particular, no vertex can receive messages
from more than O(nα) other vertices in one round (as opposed to the CONGESTED-CLIQUE, where
a vertex can receive up to O(n) messages per round).
The key feature of our partitioning algorithm is that we can reduce the coloring problem to
several instances of coloring graphs with maximum degree ∆ = O(nα/2). Given this assumption,
we can implement the CLP algorithm in the low-memory MPC model almost line by line as done
by Parter [Par18, Appendix A.2] for the CONGESTED-CLIQUE. Therefore, here we simply point
out the differences in the algorithm and refer the reader to the paper by Parter for further technical
details.
14Each vertex v is associated with a parameter pv = ηǫ2
j ∆, and we have Pu∈N out(v) 1/pu = O(ǫ0.5
j ) = O(p1/4
v
).
34
Dense Vertices. Put briefly, a vertex is γ-dense, if a (1 − γ)-fraction of the edges incident on it
belong to at least (1 − γ) · ∆ triangles. An γ-almost clique is a connected component of γ-dense
vertices that have at most γ · ∆ vertices outside the component. Each such component has a weak
diameter of at most 2. These components can be computed in 2 rounds by each vertex learning
its 2-hop neighborhood. This process is performed O(log log ∆) times in parallel which incurs a
factor of O(log log ∆) in the memory requirements, which is negligible. Furthermore, the algorithm
requires running a coloring algorithm within the dense components. Since the component size is at
most ∆ ≪ ∆2, we can choose one vertex in the component as a leader and the leader vertex can
locally simulate the coloring algorithm without breaking the local memory restriction.
Memory Bounds. Once the 2-hop neighborhoods of nodes have been learned, no more memory
overhead is required. Since we have ∆ ≪ nα/2, learning the 2-hop neighborhoods does not violate the
local memory restriction of O(nα). For the total memory bound, storing the 2-hop neighborhoods
requires at most eO(Pv(degG(v))2) memory.
Post-Shattering and Clean-up. Another step that we cannot use as a black box is a subroutine
that colors a graph that consists of connected components of O(poly log n) size. Regardless of the
component sizes being small, all vertices over all components might not fit the memory of a single
machine. Hence, similarly to the CLP algorithm in the LOCAL model, we use the best deterministic
list coloring algorithm to color the components. For general graphs, currently the best runtime in
the LOCAL model is obtained by applying the algorithm by Panconesi and Srinivasan [PS96] with
runtime of 2O(√log n′), where n′ = O(poly log n) is the maximum size of the small components. We
can improve this bound exponentially in the MPC model by using the known graph exponentiation
technique [Len13, BFU18b] and obtain a runtime of O(√log log n).
The graph exponentiation technique works as follows. Suppose that every vertex knows all the
vertices and the topology of its 2i−1-hop neighborhood in round i−1 for some integer i ≥ 0. Then, in
round i, every vertex can communicate the topology of its 2i−1-hop neighborhood to all the vertices
in its 2i−1-hop neighborhood. This way, every vertex learns its 2i-hop neighborhood in round i and
hence, every vertex can simulate any 2i-round LOCAL algorithm in i rounds. We observe that, in
the components of O(poly log n) size, the 2i-hop neighborhood of any vertex for any i fits into the
memory of a single machine since the number of vertices in the neighborhood is clearly bounded by
O(poly log n). The same observation yields that the total memory of O(m) suffices.
35
|
1904.12596 | 3 | 1904 | 2019-11-16T16:09:05 | Graph Planarity Testing with Hierarchical Embedding Constraints | [
"cs.DS"
] | Hierarchical embedding constraints define a set of allowed cyclic orders for the edges incident to the vertices of a graph. These constraints are expressed in terms of FPQ-trees. FPQ-trees are a variant of PQ-trees that includes F-nodes in addition to P- and to Q-nodes. An F-node represents a permutation that is fixed, i.e., it cannot be reversed. Let $G$ be a graph such that every vertex of $G$ is equipped with a set of FPQ-trees encoding hierarchical embedding constraints for its incident edges. We study the problem of testing whether $G$ admits a planar embedding such that, for each vertex $v$ of $G$, the cyclic order of the edges incident to $v$ is described by at least one of the FPQ-trees associated with~$v$. We prove that the problem is fixed-parameter tractable for biconnected graphs, where the parameters are the treewidth of $G$ and the number of FPQ-trees associated with every vertex of $G$. We also show that the problem is NP-complete if parameterized by the number of FPQ-trees only, and W[1]-hard if parameterized by the treewidth only. Besides being interesting on its own right, the study of planarity testing with hierarchical embedding constraints can be used to address other planarity testing problems. In particular, we apply our techniques to the study of NodeTrix planarity testing of clustered graphs. We show that NodeTrix planarity testing with fixed sides is fixed-parameter tractable when parameterized by the size of the clusters and by the treewidth of the multi-graph obtained by collapsing the clusters to single vertices, provided that this graph is biconnected. | cs.DS | cs |
Graph Planarity Testing with
Hierarchical Embedding Constraints
Giuseppe Liottaa, Ignaz Rutterb, Alessandra Tappinia,∗
aDipartimento di Ingegneria, Universit`a degli Studi di Perugia, Italy
bDepartment of Computer Science and Mathematics, University of Passau, Germany
Abstract
Hierarchical embedding constraints define a set of allowed cyclic orders for the
edges incident to the vertices of a graph. These constraints are expressed in
terms of FPQ-trees. FPQ-trees are a variant of PQ-trees that includes F-nodes
in addition to P- and to Q-nodes. An F-node represents a permutation that is
fixed, i.e., it cannot be reversed. Let G be a graph such that every vertex of G is
equipped with a set of FPQ-trees encoding hierarchical embedding constraints
for its incident edges. We study the problem of testing whether G admits a
planar embedding such that, for each vertex v of G, the cyclic order of the
edges incident to v is described by at least one of the FPQ-trees associated
with v. We prove that the problem is fixed-parameter tractable for biconnected
graphs, where the parameters are the treewidth of G and the number of FPQ-
trees associated with every vertex of G. We also show that the problem is
NP-complete if parameterized by the number of FPQ-trees only, and W[1]-hard
if parameterized by the treewidth only. Besides being interesting on its own
right, the study of planarity testing with hierarchical embedding constraints
can be used to address other planarity testing problems. In particular, we apply
our techniques to the study of NodeTrix planarity testing of clustered graphs.
We show that NodeTrix planarity testing with fixed sides is fixed-parameter
tractable when parameterized by the size of the clusters and by the treewidth of
the multi-graph obtained by collapsing the clusters to single vertices, provided
that this graph is biconnected.
Keywords: Graph Algorithms, Fixed Parameter Tractability, Planarity
Testing, Embedding Constraints, NodeTrix Planarity
∗Corresponding author
Email addresses: [email protected] (Giuseppe Liotta),
[email protected] (Ignaz Rutter), [email protected]
(Alessandra Tappini)
1
1. Introduction
The study of graph planarity testing and of its variants is at the heart of
graph algorithms and of their applications in various domains (see, e.g., [1]).
Among the most studied variants we recall, for example, upward planarity test-
ing, rectilinear planarity testing, clustered planarity testing, and HV-planarity
testing (see, e.g., [2, 3, 4, 5, 6, 7]). This paper studies a problem of graph
planarity testing subject to embedding constraints.
In its more general terms, graph planarity with embedding constraints ad-
dresses the problem of testing whether a graph G admits a planar embedding
where the cyclic order of the edges incident to (some of) its vertices is totally or
partially fixed. For example, Angelini et al. [8] and Jel´ınek et al. [9] study the
case when the planar embedding of a subgraph H of G is given as part of the
input. Angelini et al. [8] present a linear-time solution to the problem of testing
whether G admits a planar embedding that extends the given embedding of H.
Jel´ınek et al. [9] show that if the planarity test fails, an obstruction taken from a
collection of minimal non-planar instances can be produced in polynomial time.
A different planarity testing problem with embedding constraints is studied by
Dornheim [10], who considers the case that G is given with a distinguished set
of cycles and it is specified, for each cycle, that certain edges must lie inside
or outside the cycle. He proves NP-completeness in general and describes a
polynomial-time solution when the graph is biconnected and any two cycles
share at most one vertex. Da Lozzo and Rutter [11] give an approximation
algorithm for a restricted version of the problem.
The research in this paper is inspired by a seminal work of Gutwenger et al.
[12] who study the graph planarity testing problem subject to hierarchical em-
bedding constraints. Hierarchical embedding constraints specify for each vertex
v of G which cyclic orders of the edges incident to v are admissible in a con-
strained planar embedding of G. The term "hierarchical" reflects the fact that
these constraints describe ordering relationships both between sets of edges in-
cident to a same vertex and, recursively, between edges within a same set. For
example, Fig. 1 shows a vertex v, its incident edges, and a set of hierarchical
embedding constraints on these edges. The edges in the figure are partitioned
into four sets, denoted as E1, E2, E3, and E4; the embedding constraints al-
low only two distinct clockwise cyclic orders for these edge-sets, namely either
E1E2E3E4 (Fig. 1(a)) or E1E3E2E4 (Fig. 1(b)). Within each set, the hierar-
chical embedding constraints of Fig. 1 allow the edges of E1, E2, and E3 to be
arbitrarily permuted with one another, while the edges of E4 are partitioned
into three subsets E(cid:48)
4 must always appear between
4 and E(cid:48)(cid:48)(cid:48)
E(cid:48)
4 can be ar-
bitrarily permuted, while the edges of E(cid:48)(cid:48)
4 have only two
possible orders that are the reverse of one another.
4 in the clockwise order around v. Also, the edges of E(cid:48)
4 and the edges of E(cid:48)(cid:48)(cid:48)
4, E(cid:48)(cid:48)
4 , and E(cid:48)(cid:48)(cid:48)
4 such that E(cid:48)(cid:48)
Hierarchical embedding constraints can be conveniently encoded by using
FPQ-trees, a variant of PQ-trees that includes F-nodes in addition to P-nodes
and to Q-nodes. An F-node encodes a permutation that cannot be reversed. For
example, the hierarchical embedding constraints of Fig. 1 can be represented by
2
(a)
(b)
Figure 1: (a)-(b) Two examples of a vertex v with hierarchical embedding constraints and the
corresponding FPQ-trees.
two FPQ-trees denoted as T and T (cid:48) in Fig. 1(a) and 1(b), respectively. The
leaves of T and T (cid:48) are the elements of E1, E2, E3, E(cid:48)
4 . In the figure,
F-nodes are depicted as shaded boxes, Q-nodes as white boxes, and P-nodes as
circles. The F-node of the FPQ-tree T in Fig. 1(a) enforces the cyclic order
E1E2E3E4 around v, while the F-node of the FPQ-tree T (cid:48) in Fig. 1(b) enforces
the cyclic order E1E3E2E4. Both in T and in T (cid:48), the Q-node that is adjacent
to the F-node enforces E(cid:48)(cid:48)
in clockwise order
around v. The constraints by which the edges of E1, E2, E3, and E(cid:48)
4 can be
arbitrarily permuted around v are encoded by P-nodes in T and in T (cid:48).
4 to appear between E(cid:48)
4 and E(cid:48)(cid:48)(cid:48)
4, E(cid:48)(cid:48)
4 , and E(cid:48)(cid:48)(cid:48)
4
Gutwenger et al. [12] study the planarity testing problem with hierarchical
embedding constraints by allowing at most one FPQ-tree per vertex. In this
paper we generalize their study and allow more than one FPQ-tree associated
with each vertex. Besides being interesting on its own right, this generalization
can be used to model and study other graph planarity testing problems. As a
proof of concept, we apply our results to the study of NodeTrix planarity testing
of clustered graphs.
Before listing our results, we recall here that NodeTrix representations have
been introduced to visually explore flat clustered graphs by Henry et al. [13] in
one of the most cited papers of the InfoVis conference [14]. See also [13, 15,
16, 17]. A flat clustered graph G is a graph whose vertex set is partitioned into
3
E4E(cid:48)(cid:48)4E(cid:48)4E1E2E3E(cid:48)(cid:48)(cid:48)4E(cid:48)4E3E1E2E(cid:48)(cid:48)(cid:48)4E(cid:48)(cid:48)4F-nodeQ-nodeP-nodeLegend:TvE4E(cid:48)(cid:48)4E(cid:48)4E1E3E2E(cid:48)(cid:48)(cid:48)4E(cid:48)4E2E1E3E(cid:48)(cid:48)(cid:48)4E(cid:48)(cid:48)4T(cid:48)v (a)
(b)
Figure 2: (a) A non-planar flat clustered graph G. Clusters are highlighted in blue and green.
(b) A planar NodeTrix representation of G.
subsets called clusters. A NodeTrix representation of G represents its clusters
as adjacency matrices, while the edges connecting different matrices are rep-
resented as simple curves (see for example Figure 2). The NodeTrix planarity
testing problem asks whether G admits a NodeTrix representation without edge
crossings. The question can be asked both in the "fixed sides" scenario and
in the "free sides" scenario. The fixed sides scenario specifies, for each edge e
connecting two matrices M and M(cid:48), the sides (Top, Bottom, Left, Right) of M
and M(cid:48) to which e must be incident; in the free sides scenario the testing algo-
rithm can choose the sides to which e is incident. NodeTrix planarity testing is
known to be NP-complete in both scenarios [16, 17, 18]. Our main results are
the following.
• We show that FPQ-Choosable Planarity Testing is NP-complete
even if the number of FPQ-trees associated with each vertex is bounded
by a constant, and it remains NP-complete even if the FPQ-trees only
contain P-nodes. This contrasts with the result of Gutwenger et al. [12]
who prove that FPQ-Choosable Planarity Testing can be solved
in linear time when each vertex is equipped with at most one FPQ-tree.
We also prove that FPQ-Choosable Planarity Testing is W[1]-hard
parameterized by treewidth, and that it remains W[1]-hard even when the
FPQ-trees only contain P-nodes.
• The above results imply that FPQ-Choosable Planarity Testing is
not fixed-parameter tractable if parameterized by treewidth only or by the
number of FPQ-trees per vertex only. For a contrast, we show that FPQ-
Choosable Planarity Testing becomes fixed-parameter tractable for
biconnected graphs when parameterized by both the treewidth and the
number of FPQ-trees associated with every vertex.
• We show that there is a strict interplay between the FPQ-Choosable
Planarity Testing problem and the problem of testing whether a flat
clustered graph G is NodeTrix planar. Indeed, we prove that NodeTrix
4
v1v2v3v4v5v6v7v8v9v10v11v6v7v9v8v10v1v2v4v3v5v11 planarity testing with fixed sides is fixed-parameter tractable when pa-
rameterized by the size of the clusters of G and by the treewidth of the
multi-graph obtained by collapsing the clusters of G to single vertices, pro-
vided that this graph is biconnected. If we consider the vertex degree of G
as an additional parameter, the fixed-parameter tractability immediately
extends to NodeTrix planarity testing with free sides.
From a technical point of view, our algorithmic approach is based on a
combined usage of different data structures, namely SPQR-trees [19], FPQ-
trees, and sphere-cut decomposition trees [20, 21, 22]. It may be worth recalling
that a polynomial-time solution for NodeTrix planarity testing with fixed sides
was known only when the size of each cluster is bounded by a constant and the
treewidth of the graph obtained by collapsing the clusters to single vertices is
two [17].
The rest of the paper is organized as follows. Section 2 reports prelimi-
nary definitions. Section 3 introduces the FPQ-Choosable Planarity Test-
ing problem, Section 4 studies its computational complexity, in Section 5 we
describe a fixed-parameter tractability approach for FPQ-Choosable Pla-
narity Testing, and in Section 6 we analyze the interplay between FPQ-
Choosable Planarity Testing and NodeTrix Planarity testing. Concluding
remarks and open problems are given in Section 7.
2. Preliminaries
We assume familiarity with graph theory and algorithms, and we only briefly
recall some of the basic concepts that will be used extensively in the rest of the
paper (see also [23, 24]).
A PQ-tree is a tree-based data structure that represents a family of permu-
tations on a set of elements [25]. In a PQ-tree, each element is represented by
one of the leaf nodes, and each non-leaf node is a P-node or a Q-node. The
children of a P-node can be permuted arbitrarily, while the order of the children
of a Q-node is fixed up to reversal. Given a graph G together with a fixed com-
binatorial embedding, we can associate with each vertex v a PQ-tree Tv whose
leaves represent the edges incident to v. Tree Tv encodes a set of permutations
for its leaves: Each of these permutations is in a bijection with a cyclic order of
the edges incident to v. If there is a permutation πv of the leaves of Tv that is
in a bijection with a cyclic order σv of the edges incident to v, we say that Tv
represents σv, or equivalently that σv is represented by Tv.
An FPQ-tree is a PQ-tree where, for some of the Q-nodes, the reversal
of the permutation described by their children is not allowed. To distinguish
these Q-nodes from the regular Q-nodes, we call them F-nodes.
It may be
worth recalling that Gutwenger et al. [12] call this data structure "embedding
constraint", and that their "gc-nodes" correspond to P-nodes, "mc-nodes" to
Q-nodes, and "oc-nodes" to F-nodes.
Let G be a biconnected planar (multi-)graph. An SPQR-decomposition of G
describes the structure of G in terms of its triconnected components by means
5
(a)
(b)
(c)
Figure 3: (a) A biconnected planar graph G. (b) An SPQR-decomposition tree of G. The
skeletons of S-, P-, and R- nodes are inside gray boxes, while Q-nodes are depicted as letters.
(c) The embedding tree of v2.
of a tree called the SPQR-decomposition tree, and denoted as T (see, e.g., [19,
24]). Tree T can be computed in linear time and it has three types of internal
nodes that correspond to different arrangements of the components of G. If the
components are arranged in a cycle, they correspond to an S-node of T ; if they
share two vertices and are arranged in parallel, they correspond to a P-node
of T ; if they are arranged in a triconnected graph, they correspond to an R-
node of T . The leaves of T are Q-nodes, and each of them corresponds to an
edge of G. To simplify the description and without loss of generality, we shall
assume that every S-node of T has exactly two children. For each node µ of T ,
the skeleton of µ is an auxiliary graph that represents the arrangement of the
triconnected components of G corresponding to µ, and it is denoted by skel(µ).
Each edge of skel(µ) is called a virtual edge, and the end-points of a (possibly
virtual) edge are called poles. Every virtual edge corresponds to a subgraph of
G called the pertinent graph, that is denoted by Gµ. Tree T encodes all possible
planar combinatorial embeddings of G. These embeddings are determined by P-
and R-nodes, since the skeletons of S- and Q-nodes have a unique embedding.
Indeed, the skeleton of a P-node consists of parallel edges that can be arbitrarily
permuted, while the skeleton of an R-node is triconnected, and hence it has a
unique embedding up to a flip. Figure 3(a) shows a biconnected planar multi-
graph G and Figure 3(b) illustrates an SPQR-decomposition tree of G.
Note that the planar combinatorial embeddings that are given by the SPQR-
decomposition tree of a biconnected graph G give constraints on the cyclic
order of edges around each vertex of G. These constraints can be encoded by
associating a PQ-tree to each vertex v of G, called the embedding tree of v and
denoted as T
v (see, e.g., [26]). For example, Figure 3(c) shows the embedding
tree T
of the vertex v2 in Figure 3(a). Note that edges f and h (i and j, resp.)
v2
belong to an R-node (a P-node, resp.) in the SPQR-decomposition tree of G,
hence the corresponding leaves are connected to a Q-node (a P-node, resp.)
in T
v2
.
6
v1v2v3v4v5abcdefghijiv1v4v1v2v4v3v1v2v5v2v1v5v2bcafghedjPPPRSfheijTv2 3. The FPQ-choosable Planarity Testing Problem
Let G = (V, E) be a (multi-)graph, let v ∈ V , and let Tv be an FPQ-
tree whose leaf set is E(v), i.e., the set of the edges incident to v. We define
consistent(Tv) as the set of cyclic orders of the edges incident to v in E that are
represented by the FPQ-tree Tv.
An FPQ-choosable graph is a pair (G, D) where G = (V, E) is a (multi-)
graph, and D is a mapping that associates each vertex v ∈ V with a set D(v)
of FPQ-trees whose leaf set is E(v). Given a planar embedding E of G, we
denote by E(v) the cyclic order of edges incident to v in E. An assignment A
is a function that assigns to each vertex v ∈ V an FPQ-tree in D(v). We say
that A is compatible with G if there exists a planar embedding E of G such
that E(v) ∈ consistent(A(v)) for all v ∈ V . In this case, we also say that E is
consistent with A.
An FPQ-choosable graph (G, D) is FPQ-choosable planar if there exists
an assignment of FPQ-trees that is compatible with G. Figure 4(a) shows an
FPQ-choosable planar graph G, whose vertices are equipped with the following
sets of FPQ-trees: D(u1) = {Tα}, D(u2) = {Tβ, Tγ}, D(u3) = {Tδ}, and
D(u4) = {Tε}. There are two possible assignments that differ from one another
for the chosen FPQ-tree in the set D(u2). As illustrated in Figures 4(b) and 4(c),
the first assignment is compatible with G, while there is no planar embedding
that is consistent with the second assignment.
The FPQ-Choosable Planarity Testing problem receives as input an
FPQ-choosable graph (G, D) and it asks whether (G, D) is FPQ-choosable pla-
nar, i.e., it asks whether there exists an assignment that is compatible with G.
In the rest of the paper we are going to assume that G is a biconnected (multi-)
graph. Clearly G must be planar or else the problem becomes trivial. Also,
any assignment that is compatible with G must define a planar embedding of
G among those described by an SPQR-decomposition tree of G.
Therefore, a preliminary step for an algorithm that tests whether (G, D)
is FPQ-choosable planar is to intersect each FPQ-tree Tv ∈ D(v) with the
embedding tree T
v of v, so that the cyclic order of the edges incident to v
satisfies both the constraints given by Tv and the ones given by T
(See,
v .
e.g., [26] for details about the operation of intersection between two PQ-trees,
whose extension to the case of FPQ-trees is straightforward since F-nodes are
just a more constrained version of Q-nodes). Therefore, from now on we shall
assume that the FPQ-trees of D have been intersected with the corresponding
embedding trees and, for ease of notation, we shall still denote with D(v) the
set of FPQ-trees associated with v and resulting from the intersection. We also
remove the null-tree, which represents the empty set of permutations, from the
sets D(v). Clearly, a necessary condition for the FPQ-choosable planarity of
(G, D) is that D(v) is not the empty set for every v ∈ G.
7
(a)
(b)
(c)
Figure 4: (a) An FPQ-choosable planar graph (G, D). (b) A planar embedding of G that
is consistent with assignment {A(u1) = Tα, A(u2) = Tγ , A(u3) = Tδ, A(u4) = Tε}; the
assignment is compatible with G. (c) A non-planar embedding of G that is consistent with
assignment {A(u1) = Tα, A(u2) = Tβ , A(u3) = Tδ, A(u4) = Tε}; there is no planar embedding
that is consistent with A.
4. Complexity of FPQ-choosable Planarity Testing
As we are going to show, FPQ-Choosable Planarity Testing is fixed-
parameter tractable when parameterized by treewidth and number of FPQ-trees
per vertex. One may wonder whether the problem remains fixed-parameter
tractable if parameterized by the treewidth only or by the number of FPQ-
trees per vertex only. The following two theorems answer this question in the
negative.
Theorem 1. FPQ-Choosable Planarity Testing with a bounded number
of FPQ-trees per vertex is NP-complete. It remains NP-complete even when the
FPQ-trees have only P-nodes.
Proof. We denote with n the number of vertices of the input graph and we
assume that for each vertex v of the input, D(v) ∈ O(n). We generate all pos-
sible assignments by performing O(n log n) non-deterministic guess operations
and, for each assignment, we decide whether it is compatible with the input
graph by applying the linear-time algorithm of Gutwenger et al. [12]. It follows
that FPQ-Choosable Planarity Testing is in NP.
In order to show that FPQ-Choosable Planarity Testing is NP-hard,
we use a reduction from the problem of deciding whether a triconnected cubic
8
u2u3u4u1512344634132212515TδTαTεTβTγ656u2u3u4u1512344634132251TδTαTεTγ65665215Tβu2u3u4u1344634132TδTαTε1526 graph admits a 3-edge-coloring. The 3-edge-coloring problem for a cubic graph
asks whether it is possible to assign a color in the set {red, green, blue} to
each edge of the graph so that no two edges of the same color share a vertex.
The problem is known to be NP-complete for triconnected cubic non-planar
graphs [27]. Note that a triconnected cubic graph admits a 3-edge-coloring
if and only if it admits a 3-edge-coloring for any choice of rotation system and
outer face, hence we perform the construction starting from a triconnected cubic
graph with an arbitrary choice of rotation system and outer face, which makes
it possible to talk about edge crossings in the graph. For any given triconnected
cubic graph G we construct an FPQ-choosable graph (G(cid:48), D(cid:48)) with D(v(cid:48)) ≤ 6
for each vertex v(cid:48) of G(cid:48), that is FPQ-choosable planar if and only if G has a
3-edge-coloring. Since every vertex of (G(cid:48), D(cid:48)) is equipped with at most six
FPQ-trees, the statement will follow. See Figure 5 for an example.
The construction that maps any triconnected cubic graph G into an FPQ-
choosable graph (G(cid:48), D(cid:48)) is as follows. Each vertex v of G is associated with
a vertex v(cid:48) in G(cid:48), and each edge e = (u, v) of G is associated in G(cid:48) with three
parallel edges e1, e2, and e3, whose end-vertices are u(cid:48) and v(cid:48). See for example
Figure 6(a) and 6(b). Each crossing c of G(cid:48) is replaced with a dummy vertex vc.
Note that every vertex of G(cid:48) has either degree 4 or 9, since we can assume
that each crossing is the intersection of exactly two edges (otherwise a small
perturbation can be applied). Each vertex vc of G(cid:48) having degree 4 is equipped
with one FPQ-tree Tvc consisting of a P-node whose leaves represent the four
edges incident to vc. Each vertex v(cid:48) of G(cid:48) having degree 9 is equipped with a set
D(cid:48)(v(cid:48)) of FPQ-trees. Each FPQ-tree in D(cid:48)(v(cid:48)) consists of a P-node ρ connected
to three Q-nodes χe1 , χe2 , and χe3, which have three leaves each, denoted as
, p2
p1
. See for example Figure 6(c), that shows an FPQ-tree of the vertex
ei
ei
v(cid:48) in Figure 6(b).
Observe that every FPQ-tree in D(cid:48)(v(cid:48)) can be defined as the union of three
trees T1, T2, and T3, such that each Ti consists of node ρ, node χei, and the
three leaves of χei (1 ≤ i ≤ 3). For example, T1, T2, and T3 are highlighted
in Figure 6(c). Consider a Q-node χei and the cyclic order σi of its incident
, p3
If the leaves of Ti appear as p1
in σi, we say that Ti
edges in Ti.
ei
ei
, p2
has a red configuration; if they appear as p1
, we say that Ti has a
ei
ei
, p1
green configuration; if they appear as p2
, we say that Ti has a blue
ei
ei
configuration. For example, in Figure 6(c) T1 has a red configuration, T2 has a
green configuration, and T3 has a blue configuration.
Let e1, e2, and e3 be the three edges incident to a vertex v in the triconnected
cubic graph G and let v(cid:48) be its corresponding vertex in (G(cid:48), D(cid:48)). For each 3-
edge-coloring of G, there is a bijection between an FPQ-tree Tv(cid:48) in D(cid:48)(v(cid:48)) and
the colors of the three edges incident to v. Namely, for a 3-edge-coloring of G
where ei is red, we impose a red configuration to Ti in Tv(cid:48); if ei is green, we
impose a green configuration to Ti; if ei is blue, we impose a blue configuration
to Ti. We say that Ti matches the color of ei and that Tv(cid:48) matches the color
of the edges incident to v. For example, the FPQ-tree of Figure 6(c) matches
the color of the edges incident to v in Figure 6(a), because T1 matches the color
of e1, T2 matches the color of e2 and T3 matches the color of e3. Since there
, p2
ei
, p3
ei
, p3
ei
, p3
ei
9
(a)
(b)
(c)
Figure 5: (a) A triconnected cubic non-planar graph G with a proper 3-edge-coloring. (b)
The corresponding FPQ-choosable graph (G(cid:48), D(cid:48)); the dummy vertices are black squares, and
the FPQ-tree associated with vc is inside a gray circle. D(cid:48)(v(cid:48)) and D(cid:48)(u(cid:48)) are shown in (c).
10
wuvw(cid:48)u(cid:48)v(cid:48)abcdefghijklmnovcppllpdefghibacdefighcbadfegihbacdfeighabcedfgihcbaedfghiabcD(cid:48)(v(cid:48))jklghinmojklighonmjlkgihnmojlkighmnokjlgihonmkjlghimnoD(cid:48)(u(cid:48)) (a)
(b)
(c)
Figure 6: (a) A vertex u of a triconnected cubic graph G and its incident edges e1 = (u, v), e2,
and e3. (b) Three parallel edges of G(cid:48) that are associated with edge e1 of G. (c) An FPQ-tree
Tu(cid:48) associated with vertex u(cid:48): T1 has a red configuration, T2 has a green configuration, and
T3 has a blue configuration.
are six possible permutations of the three colors around v in G, we have that
D(cid:48)(v(cid:48)) = 6 in (G(cid:48), D(cid:48)).
We now prove that if G admits a 3-edge-coloring, (G(cid:48), D(cid:48)) is FPQ-choosable
planar. Let v be any vertex of G with incident edges e1, e2, e3, and let v(cid:48) be
the vertex that corresponds to v in (G(cid:48), D(cid:48)). We define an assignment A for
(G(cid:48), D(cid:48)) where A(v(cid:48)) is the FPQ-tree Tv(cid:48) ∈ D(cid:48)(v(cid:48)) that matches the color of the
edges incident to v. For every vertex vc of (G(cid:48), D(cid:48)) of degree 4, A(vc) is the
only FPQ-tree associated with vc, hence the cyclic order of the edges around
vc is totally free. We show that there exists a planar embedding of G(cid:48) that is
consistent with A. Since Tu(cid:48) matches the color of the edges incident to u in
G and Tv(cid:48) matches the color of the edges incident to v in G, the leaves of Tu(cid:48)
and the leaves of Tv(cid:48) representing the edges (possibly subdivided by dummy
vertices) connecting u(cid:48) and v(cid:48) can be ordered so to avoid edge crossings. If, for
example, edge e = (u, v) is red in G, we have that Tu(cid:48) has a subtree T (cid:48) and Tv(cid:48)
has a subtree T (cid:48)(cid:48) such that both T (cid:48) and T (cid:48)(cid:48) match the red color. The sets of
leaves of T (cid:48) and T (cid:48)(cid:48) represent the same set of edges, and they appear in reverse
order around u(cid:48) and around v(cid:48) in a planar embedding of G(cid:48). It follows that if
G admits a 3-edge-coloring, (G(cid:48), D(cid:48)) is FPQ-choosable planar.
Suppose for a converse that (G(cid:48), D(cid:48)) is FPQ-choosable planar. There exists
an assignment A that is compatible with G(cid:48). Assignment A defines the cyclic
order of the edges incident to each vertex in a planar embedding of G(cid:48). Recall
that for any two vertices u(cid:48) and v(cid:48) having degree 9, they are connected by three
parallel edges (possibly subdivided by dummy vertices), where e is the edge of
G in a bijection with these three edges. Since A is compatible with G(cid:48), the two
FPQ-trees Tu(cid:48) = A(u(cid:48)) and Tv(cid:48) = A(v(cid:48)) both contain two subtrees T (cid:48) and T (cid:48)(cid:48)
such that: (i) T (cid:48) and T (cid:48)(cid:48) have the same set of three leaves; (ii) these three leaves
11
e1e2e3uve11e12e13u(cid:48)e21e31e32e33v(cid:48)e23e22p1e1p2e1p3e1p1e2p3e2p2e2p1e3p2e3p3e3χe1χe2χe3ρT1T2T3 Figure 7: An FPQ-tree with only P-nodes associated with a vertex of degree 18 in G(cid:48).
represent edges of G(cid:48) that correspond to e1, e2, and e3; (iii) T (cid:48) and T (cid:48)(cid:48) have the
same red (green, blue) configuration. We color edges e1, e2, and e3 with the
red (green, blue) color depending on the color configuration of T (cid:48) and of T (cid:48)(cid:48).
By iterating this procedure over all triplets of edges we have that around every
vertex of degree 9 in G(cid:48) there are three consecutive triplets of edges such that
the edges of each triplet all have the same color and no two triplets have the
same color. A 3-edge-coloring of G is therefore obtained by giving every edge e
of G the same color as the one of the corresponding triplet e1, e2, and e3 in G(cid:48).
It follows that if (G(cid:48), D(cid:48)) is FPQ-choosable planar then G has a 3-edge-coloring.
In order to prove that the problem remains NP-complete if the FPQ-trees
associated with the vertices have only P-nodes, we construct an FPQ-choosable
graph (G(cid:48), D(cid:48)) in a slightly different way from the one described above.
In
particular, each edge of G is associated with six parallel edges in G(cid:48), and each
vertex v(cid:48) having degree 18 in G(cid:48) is equipped with six FPQ-trees as the one in
Figure 7. In this case each FPQ-tree associated with a vertex v(cid:48) of G(cid:48) having
degree 18 is the union of three trees T1, T2, and T3, such that each of their
three pairs of leaves are connected to a P-node, which enforces each pair of
leaves to appear consecutively. If Ti (1 ≤ i ≤ 3) has a red configuration, the
, p4
) must be consecutive, as well as the leaves (p3
, p2
two leaves (p1
), and the
ei
ei
ei
ei
, p6
); if Ti has a green configuration the two leaves (p1
leaves (p5
, p6
) must be
ei
ei
ei
ei
), and the leaves (p5
consecutive, as well as the leaves (p3
, p4
, p2
); if Ti has a
ei
ei
ei
ei
blue configuration the two leaves (p1
, p4
) must be consecutive, as well as the
ei
ei
, p2
leaves (p3
). This guarantees that any two adjacent
ei
ei
vertices u(cid:48) and v(cid:48) of G(cid:48) are such that if Tu(cid:48) and Tv(cid:48) match the same color, there
is a cyclic order represented by Tu(cid:48) and Tv(cid:48) such that the edges incident to u(cid:48)
and the edges incident to v(cid:48) do not cross. Conversely, if they match different
2
colors these edges must cross.
), and the leaves (p5
ei
, p6
ei
We remark that Theorem 1 also holds if the number of FPQ-trees per vertex
is bounded by a constant larger than six, indeed it is possible to associate each
edge of the given triconnected cubic graph G with a suitable number of parallel
12
p1e1p2e1p3e1p4e1p5e1p6e1p1e2p6e2p3e2p2e2p5e2p4e2p1e3p4e3p3e3p6e3p5e3p2e3χe1χe2χe3ρT1T2T3 edges and each vertex of G with a suitable number of FPQ-trees.
We now prove that FPQ-Choosable Planarity Testing parameterized
by treewidth is W[1]-hard.
Theorem 2. FPQ-Choosable Planarity Testing parameterized by treewidth
is W[1]-hard. It remains W[1]-hard even when the FPQ-trees have only P-nodes.
If h = 0, the triplet e(cid:48)
1, e(cid:48)
2, e(cid:48)
1, e(cid:48)
2, e(cid:48)
2, and e(cid:48)
3 is labeled with no color.
Proof. We use a parameterized reduction from the list coloring problem, which
is defined as follows. Given a graph G = (V, E) and a set L containing a list
L(v) of colors for each vertex v ∈ V , is there a proper vertex coloring with
c(v) ∈ L(v) for each v? We denote as c(v) the color of v in a proper vertex
coloring. The list coloring problem parameterized by treewidth is known to be
W[1]-hard even for planar graphs [28, Theorem 14.29].
For any given instance (G, L) of list coloring such that G is a planar graph
whose treewidth is at most t, we construct an FPQ-choosable graph (G(cid:48), D(cid:48))
that is FPQ-choosable planar if and only if (G, L) is a yes instance of list
coloring. Note that (G, L) is a yes instance of list coloring if and only if it is
a yes instance for any planar embedding of G, hence we perform the reduction
to FPQ-Choosable Planarity Testing starting from any instance (G, L)
with an arbitrary planar embedding of G. Starting from a planar embedding of
graph G, we construct a planarly embedded multi-graph G(cid:48) by replacing each
edge of G with bundles of edges as follows. Also refer to Figure 8.
Each vertex v of G becomes vertex v(cid:48) in G(cid:48), and each edge e = (u, v) of G
3 in G(cid:48). Let h = L(u) ∩ L(v) in
is split into three parallel edges e(cid:48)
1, e(cid:48)
(G, L).
If h = 1 the
triplet e(cid:48)
3 is labeled with the color in common between u and v. If h > 1,
we create 3(h − 1) additional parallel edges in G(cid:48) between u(cid:48) and v(cid:48), and we
label each of the 3h triplets of edges with one of the colors shared by u and v.
Observe that G(cid:48) is a planar multi-graph with a given planar embedding and its
treewidth is the same as the treewidth of G.
We construct the set D(cid:48) of FPQ-trees associated with the vertices of G(cid:48)
starting from the set L of colors associated with the vertices of G as follows.
Let v be a vertex of G, let L(v) be its color list, and let v(cid:48) be the corresponding
vertex in G(cid:48). Denote as deg(v(cid:48)) the degree of v(cid:48) in G(cid:48). We equip v(cid:48) in G(cid:48)
with L(v) FPQ-trees, each encoding a color of L(v) in (G, L); we denote as
Tc(v(cid:48)) the FPQ-tree associated with v(cid:48) in G(cid:48) and encoding color c ∈ L(v). Let
k = deg(v(cid:48))/3 (note that k is a positive integer since deg(v(cid:48)) mod 3 = 0). If
k = 1, deg(v) = 1 in G, and there is at most one color c such that c ∈ L(u)∩L(v),
where u is the neighbor of v in G. Each FPQ-tree Tc(v(cid:48)) consists of a node ρ
whose leaves represent the triplet of edges incident to v(cid:48). Node ρ is a Q-node if
v shares color c with its neighbor, otherwise ρ is a P-node (observe that there
are at least L(v)− 1 FPQ-trees associated with v(cid:48) with the same set of nodes).
For example, in Figure 8(c), for vertex z(cid:48) we have k = 1; the triplet of edges
incident to z(cid:48) is labeled with color 3, and the FPQ-tree T3(z(cid:48)) consists of a node
ρ with three leaves. Node ρ is a Q-node because color 3 ∈ L(z) ∩ L(z) in G.
If k > 1, each FPQ-tree of D(cid:48)(v(cid:48)) consists of a P-node ρ connected to k nodes
ν1, . . . , νk having three leaves each. The leaves of each νi (1 ≤ i ≤ k) represent a
13
(a)
(b)
(c)
Illustration of the reduction from list coloring to FPQ-Choosable Planarity
Figure 8:
Testing. (a) An instance (G, L) of list coloring. The circled colors indicate a valid coloring
of G. (b) The corresponding FPQ-choosable graph (G(cid:48), D(cid:48)); some FPQ-trees of D(cid:48) are shown
in (c).
14
wuvz12345671122233344455566772zw(cid:48)u(cid:48)v(cid:48)fedghijklpqrstαγbaz(cid:48)mno1222344567z(cid:48)sαD(cid:48)(u(cid:48))D(cid:48)(v(cid:48))prqghiprqghiprqghistαsαpqrstαpqrstαD(cid:48)(w(cid:48))fdeihglkjbγfdeiglkjabγahfdeiglkjbγahD(cid:48)(z(cid:48))onm132254312ρν1ν2ν3tt 2, e(cid:48)
1, e(cid:48)
triplet e(cid:48)
3 of edges connecting v(cid:48) to some other vertex u(cid:48) of G(cid:48); this triplet
either encodes a color in L(v) ∩ L(u) or it encodes no color if L(v) ∩ L(u) = ∅.
Also, if the color c associated with Tc(v(cid:48)) is such that c ∈ L(v)∩ L(u), node νi is
a Q-node; it is a P-node otherwise. For example, in Figure 8(c) we have k > 1
for vertex v(cid:48). The FPQ-tree T1(v(cid:48)) encodes the color 1 of L(v); v(cid:48) has three
triplets of incident edges and node ρ of T1(v(cid:48)) has three children whose leaves
represent these three triplets. Since color 1 belongs to both L(u) and L(v) in
Figure 8(a), the node ν1 of T1(v(cid:48)) whose leaves represent the triplet of edges
p, q, r of G(cid:48) is a Q-node. Conversely, the nodes ν2 and ν3 of T1(v(cid:48)) associated
with the triplets labeled with colors 2 and 3 of L(v) are P-nodes.
Note that D(cid:48)(v(cid:48)) = L(v) for each vertex v of G and each vertex v(cid:48) of G(cid:48),
thus we have that the size of (G(cid:48), D(cid:48)) is polynomial in the size of (G, L). We
now prove that (G, L) admits a proper vertex coloring with c(v) ∈ L(v) for each
v if and only if (G(cid:48), D(cid:48)) is FPQ-choosable planar.
Suppose first that (G, L) admits a proper vertex coloring. Let v be any
vertex of G, let c(v) be the chosen color for v, and let v(cid:48) be the image of v
in (G(cid:48), D(cid:48)). Assignment A for (G(cid:48), D(cid:48)) is defined such that A(v(cid:48)) = Tc(v)(v(cid:48)).
We show that there exists a planar embedding of G(cid:48) that is consistent with A.
Since any pair of adjacent vertices u and v in G is such that c(u) (cid:54)= c(v), the
two FPQ-trees A(u(cid:48)) = Tc(u)(u(cid:48)) and A(v(cid:48)) = Tc(v)(v(cid:48)) contain pairs of nodes
whose leaves correspond to triplets of edges connecting u(cid:48) and v(cid:48). Each of these
triplets are connected to a P-node either in A(u(cid:48)) or in A(v(cid:48)) (or in both), hence
they can be ordered so to avoid edge crossings in G(cid:48). It follows that if (G, L) is
a yes instance of list coloring, then (G(cid:48), D(cid:48)) is FPQ-choosable planar.
Suppose now that (G(cid:48), D(cid:48)) is FPQ-choosable planar. There exists an assign-
ment A that defines the cyclic order of the edges incident to each vertex in a
planar embedding of G(cid:48). Let u(cid:48) and v(cid:48) be any two adjacent vertices of G(cid:48). FPQ-
trees A(v(cid:48)) = Tc1(v(cid:48)) and A(u(cid:48)) = Tc2(u(cid:48)) are such that the edges represented
by their leaves can be drawn in G(cid:48) without crossings, hence they correspond
to different colors c1 and c2 for v and u, and thus c(v) (cid:54)= c(u). It follows that
if (G(cid:48), D(cid:48)) is FPQ-choosable planar, then (G, L) is a yes instance of list color-
ing. It follows that FPQ-Choosable Planarity Testing parameterized by
treewidth is W[1]-hard.
The proof that the problem remains W[1]-hard even if the FPQ-trees asso-
ciated with the vertices have only P-nodes is a slight variant of the argument
above. Namely, we construct an FPQ-choosable graph (G(cid:48), D(cid:48)) such that each
vertex v of G becomes vertex v(cid:48) in G(cid:48), and each edge e = (u, v) of G is split
into 6-tuples of parallel edges in G(cid:48). If h = L(u) ∩ L(v) = 0, the six paral-
lel edges between u and v are labeled with no color. If h = 1 the 6-tuple of
parallel edges is labeled with the color in common between u and v. If h > 1,
we create 6(h − 1) additional parallel edges between u(cid:48) and v(cid:48). Similarly tothe
previous case, we label each of these 6-tuples of edges with one of the colors
in L(u) ∩ L(v). Each vertex v(cid:48) in G(cid:48) is equipped with L(v) FPQ-trees, each
encoding a color of L(v). If k = deg(v(cid:48))/6 = 1, each FPQ-tree Tc(v(cid:48)) consists
of a P-node ρ connected to three P-nodes whose leaves represent the six edges
incident to v(cid:48); see for example Figure 9(c). If k > 1 each FPQ-tree associated
15
(a)
(b)
(c)
(d)
Figure 9: (a) An edge (u, v) of an instance of list coloring.
(b) The corresponding FPQ-
choosable graph. (c) The FPQ-tree T2(u(cid:48)) of D(cid:48)(u(cid:48)) associated with color 2. (d) The FPQ-tree
T2(v(cid:48)) of D(cid:48)(v(cid:48)) associated with color 2. Note that they contain only P-nodes.
with vertex v(cid:48) of G(cid:48) consists of a P-node ρ connected to k P-nodes ν1, . . . , νk,
each of which is connected to three P-nodes. Each of these three P-nodes has
two leaves; see for example Figure 9(d). If v shares a color c with an adjacent
vertex u, the FPQ-tree Tc(v(cid:48)) contains a P-node νl (1 ≤ l ≤ k) whose leaves
represent the 6-tuple of edges connecting v(cid:48) with u(cid:48) that is labeled with color c.
Each of these three pairs of leaves is connected to a P-node, which enforces each
pair of leaves to appear consecutively. In particular, in Tc(v(cid:48)) the two leaves
i ), and the leaves (e5
i ) must be consecutive, as well as the leaves (e3
i , e2
(e1
i , e6
i )
(1 ≤ i ≤ deg(v)), while in Tc(u(cid:48)) the two leaves (e5
j ) must be consecutive, as
j ) (1 ≤ j ≤ deg(u)). This guaran-
well as the leaves (e3
tees that two adjacent vertices v(cid:48) and u(cid:48) of G(cid:48) are such that if their FPQ-trees
encode the same color c1, the edges incident to v(cid:48) and the edges incident to
u(cid:48) must respect cyclic orders that do not allow to connect them without edge
crossings. On the other hand, in an FPQ-tree Tc2(u(cid:48)) encoding a color c2 dif-
ferent from c1, the pairs of leaves that must be consecutive are the same as the
ones of Tc1(v(cid:48)), which allows to connect the corresponding edges of G(cid:48) without
edge crossings. By this argument, we can conclude that the FPQ-Choosable
Planarity Testing is W[1]-hard even if the FPQ-trees associated with the
2
vertices have only P-nodes.
j ), and the leaves (e1
i , e4
j , e2
j , e6
j , e4
The results of this section imply the following.
Corollary 1. FPQ-Choosable Planarity Testing is not fixed-parameter
tractable if parameterized by treewidth only or by the number of FPQ-trees per
vertex only. It remains fixed-parameter tractable even if the FPQ-trees consist
of P-nodes.
16
ue1v122e2e11u(cid:48)e21e31v(cid:48)e41e51e61e12e22e32e42e52e62e51e41e21e31e61e112ρT2(u(cid:48))e11e61e21e31e41e51e52e42e22e32e62e122ρν1ν2T2(v(cid:48)) Figure 10: Illustration for the proof of Lemma 1.
5. Fixed Parameter Tractability of FPQ-choosable Planarity Testing
This section is organized as follows. We first introduce the notions of bound-
aries and of extensible orders, and state two technical lemmas. Next, we de-
fine the concepts of pertinent FPQ-tree, skeletal FPQ-tree and admissible tu-
ple, which are fundamental in the algorithm description. Finally, we present a
polynomial-time testing algorithm for FPQ-choosable graphs having bounded
branchwidth and such that the number of FPQ-trees associated with each ver-
tex is bounded by a constant. Note that, if a graph has bounded branchwidth
b it has treewidth at most(cid:4) 3
2 b(cid:5) − 1 [29].
Boundaries and Extensible Orders: Let T be an FPQ-tree, let yield(T )
denote the set of its leaves, and let L be a proper subset of yield(T ). We
denote by σ a cyclic order of the leaves of an FPQ-tree, and we say that σ ∈
consistent(T ) if the FPQ-tree T represents σ. We say that L is a consecutive
set if the leaves in L are consecutive in every cyclic order represented by T . Let
e be an edge of T , and let T (cid:48) and T (cid:48)(cid:48) be the two subtrees obtained by removing
e from T . If either yield(T (cid:48)) or yield(T (cid:48)(cid:48)) are a subset of a consecutive set L,
then we say that e is a split edge of L. The subtree that contains the leaves in
L is the split subtree of e. A split edge e is maximal if there exists no split edge
e(cid:48) such that the split subtree of e(cid:48) contains e.
Lemma 1. Let T be an FPQ-tree, let L be a consecutive proper subset of
yield(T ), and let S be the set of maximal split edges of L. Then either S = 1,
or S > 1 and there exists a Q-node (or an F-node) χ of T such that χ has
degree at least S + 2 and the elements of S appear consecutive around χ.
Proof. Assume that S > 1. Let e and f be two maximal split edges of L,
and let Te and Tf be the split subtrees of e and f , respectively. Let further χe
denote the endpoint of e that is not in Te. The endpoint χf is defined likewise.
Refer to Figure 10 for an illustration.
Suppose for the sake of contradiction that χe and χf are distinct. Let g
denote the first edge on the path from χe to χf . By the maximality of e and f ,
the edge g is not a split edge. It follows that there is an edge e(cid:48) incident to χe
that is different from g and that is not a split edge. Likewise, we find an edge
f(cid:48) incident to χf that is different from the first edge on the path from χf to χe
and that is not a split edge. But then g is an edge of a tree T such that one
of the two subtrees it separates has leaves in L and leaves that are not in L. It
17
χeχffegTeTfLL¬L¬L follows that L is not a consecutive set. This is a contradiction to the assumption
that χe and χf are distinct.
It follows that the edges in S are all incident to a single vertex χ. If χ has
degree S, then L is not a proper subset of the leaves, and if it has degree S+1,
then also its remaining edge is a split edge, which contradicts the maximality
of the split edges in S. Hence deg(χ) ≥ S + 2. If χ were a P-node, this would
2
contradict the assumption that L is a consecutive set.
If S = 1, the split edge in S is called the boundary of L. If S > 1, the
Q-node (or the F-node) χ defined in the statement of Lemma 1 is the boundary
of L. Since F-nodes are a more constrained version of Q-nodes, when we refer
to boundary Q-nodes we also take into account the case in which they are F-
nodes. Figure 11(a) shows an FPQ-choosable graph (G, D) and two FPQ-trees
Tu ∈ D(u) and Tv ∈ D(v). The three red edges b, c, and d of G define a
consecutive set Lu in Tu; the edges e and f define a consecutive set Lv in Tv.
The boundary of Lu in Tu is a Q-node, while the boundary of Lu in Tu is an
edge. We denote as B(L) the boundary of a set of leaves L. If B(L) is a Q-
node, we associate B(L) with a default orientation (i.e., a flip) that arbitrarily
defines one of the two possible permutations of its children. We call this default
orientation the clockwise orientation of B(L). The other possible permutation
of the children of B(L) corresponds to the counter-clockwise orientation.
Let L(cid:48) = L ∪ {(cid:96)}, where (cid:96) is a new element. Let σ ∈ consistent(T ), and
let σL(cid:48) be a cyclic order obtained from σ by replacing the elements of the
consecutive set yield(T )\L by the single element (cid:96). We say that a cyclic order σ(cid:48)
of L(cid:48) is extensible if there exists a cyclic order σ ∈ consistent(T ) with σL(cid:48) = σ(cid:48).
In this case, we say that σ is an extension of σ(cid:48). Note that if the boundary of
L is a Q-node χ, then any two extensions of σ(cid:48) induce the same clockwise or
counter-clockwise orientation of the edges incident to χ. An extensible order σ is
clockwise if the orientation of χ is clockwise; σ is counter-clockwise otherwise. If
the boundary of L is an edge, we consider any extensible order as both clockwise
and counter-clockwise.
Let L and L be two disjoint consecutive sets of leaves that have the same
boundary Q-node χ in T . Let σ and σ be two extensible orders of L and L,
respectively. We say that σ and σ are incompatible if one of them is clockwise
and the other one is counter-clockwise.
Lemma 2. Let T be an FPQ-tree, let L1 ∪ ··· ∪ Lk be a partition of yield(T )
into consecutive sets, and let σ1, . . . , σk be extensible orders of L1, . . . , Lk. There
exists an order Σ of yield(T ) represented by T such that ΣLi = σi if and only
if no pair σi, σj (1 ≤ i, j ≤ k) is incompatible.
Proof. The only-if direction is clear. For the if-direction, assume that no pair
is incompatible. Note that, since Li is consecutive, so is yield(T ) \ Li. We
denote by Ti the subtree of T that is obtained by replacing the consecutive
set yield(T ) \ Li by a single leaf (cid:96). Note that Ti (1 ≤ i ≤ k) is a subtree of T
and the set {T1, . . . , Tk} forms a partition of the edges of T . Observe that σi
18
(a)
(b)
(c)
Figure 11: (a) Two different types of boundaries: A boundary Q-node in Tu and a boundary
edge in Tv.
(c) The
skeletal FPQ-trees Skelµ(Tu) of Pertµ(Tu) and Skelµ(Tv) of Pertµ(Tv).
(b) The pertinent FPQ-trees Pertµ(Tu) of Tu and Pertµ(Tv) of Tv.
defines a cyclic order of the edges around each node of Ti. Moreover, if Ti and Tj
overlap, then they do so in the boundary of Li and Lj, which must hence be a
Q-node χ. Since no pair is incompatible, it follows that they induce the same
cyclic order Σ of the edges around χ. Thus, together the σi determine a unique
order in consistent(T ) such that ΣLi = σi.
2
Pertinent FPQ-trees, Skeletal FPQ-trees, and Admissible Tuples: Let
(G, D) be an FPQ-choosable graph, let T be an SPQR-decomposition tree of G
and let v be a pole of a node µ of T , let Tv ∈ D(v) be an FPQ-tree associated
with v, let Eext be the set of edges that are incident to v and not contained in
µ(v) = E(v) \ Eext. Note that there is a bijection between the
Gµ, and let E(cid:63)
edges E(v) of G and the leaves of Tv, hence we shall refer to the set of leaves of
Tv as E(v). Also note that E(cid:63)
µ(v) is represented by a consecutive set of leaves
in Tv, because in every planar embedding of G the edges in E(cid:63)
µ(v) must appear
consecutively in the cyclic order of the edges incident to v.
The pertinent FPQ-tree of Tv, denoted as Pertµ(Tv), is the FPQ-tree ob-
tained from Tv by replacing the consecutive set Eext with a single leaf (cid:96). In-
formally, the pertinent FPQ-tree of v describes the hierarchical embedding con-
straints for the pole v within the pertinent graph Gµ. For example, in Fig-
ure 11(b) a pertinent graph Gµ with poles u and v is highlighted by a shaded
region; the pertinent FPQ-tree Pertµ(Tu) of Tu and the pertinent FPQ-tree
19
kdabcefuvboundaryboundaryTughijhgTvijkqmnopkabcdqoppafemnpkdabcefuvabcdghija(cid:96)fekqmnopkqopmn(cid:96)Pertµ(Tu)Pertµ(Tv)Gµ(cid:96)1(cid:96)(cid:96)2(cid:96)3Skelµ(Tu)(cid:96)2(cid:96)(cid:96)1(cid:96)3Skelµ(Tv) νi
Pertµ(Tv) of Tv are obtained by the FPQ-trees Tu and Tv of Figure 11(a).
Let ν1, . . . , νk be the children of µ in T . Observe that the edges E(cid:63)
(v) of
each Gνi (1 ≤ i ≤ k) form a consecutive set of leaves of Aµ(v) = Pertµ(Tv).
The skeletal FPQ-tree of Pertµ(Tv), denoted by Skelµ(Tv), is the tree obtained
(v) (1 ≤ i ≤ k) by
from Pertµ(Tv) by replacing each of the consecutive sets E(cid:63)
νi
a single leaf (cid:96)i. See for example, Figure 11(c). Observe that each Q-node of
Skelµ(Tu) corresponds to a Q-node of Pertµ(Tu), and thus to a Q-node of Tu;
also, distinct Q-nodes of Skelµ(Tu) correspond to distinct Q-nodes of Pertµ(Tu),
and thus to distinct Q-nodes of Tu. For each Q-node χ of Tu that is a boundary
of µ or of one of its children νi, there is a corresponding Q-node in Skelµ(Tu)
that inherits its default orientation from Tu.
Let (G, D) be an FPQ-choosable graph, let T be an SPQR-decomposition
tree of G, let µ be a node of T , and let u and v be the poles of µ. We denote
with (Gµ, Dµ) the FPQ-choosable graph consisting of the pertinent graph Gµ
and the set Dµ that is defined as follows: Dµ(z) = D(z) for each vertex z of
Gµ that is not a pole, and Dµ(v) = {Pertµ(Tv) Tv ∈ D(v)} if v is a pole of
µ. A tuple (cid:104)Tu, Tv, ou, ov(cid:105) ∈ D(u) × D(v) × {0, 1} × {0, 1} is admissible for Gµ
if there exists an assignment Aµ of (Gµ, Dµ) and a planar embedding Eµ of Gµ
consistent with Aµ such that Aµ(u) = Pertµ(Tu), Aµ(v) = Pertµ(Tv), B(E(cid:63)
µ(u))
is clockwise (counter-clockwise) in Tu if ou = 0 (ou = 1), and B(E(cid:63)
µ(v)) is
clockwise (counter-clockwise) in Tv if ov = 0 (ov = 1). We say that a tuple
is admissible for µ if it is admissible for Gµ. We denote by Ψ(µ) the set of
admissible tuples for Gµ.
FPT Algorithm: In order to test if (G, D) is FPQ-choosable planar, we root
the SPQR-decomposition tree T at an arbitrary Q-node and we visit T from
the leaves to the root. At each step of the visit, we equip the currently visited
node µ with the set Ψ(µ). If we encounter a node µ such that Ψ(µ) = ∅, we
return that (G, D) is not FPQ-choosable planar; otherwise the planarity test
returns an affirmative answer. If the currently visited node µ is a leaf of T , we
set Ψ(µ) = D(u)× D(v)×{0, 1}×{0, 1}, because its pertinent graph is a single
edge. If µ is an internal node, Ψ(µ) is computed from the sets of admissible
tuples of the children of µ. The next lemmas describe how to compute Ψ(µ)
depending on whether µ is an S-, P-, or R-node.
Lemma 3. Let µ be an S-node with children ν1 and ν2. Given Ψ(ν1) and
Ψ(ν2), the set Ψ(µ) can be computed in O(D2
max log(Dmax)) time, where Dmax =
maxv∈V D(v).
Proof. Let u and v be the poles of µ, and let w be the pole in common between
Gν1 and Gν2 . We show that (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(µ) if and only if there exist
a tree Tw ∈ D(w) and an ow ∈ {0, 1}, such that (cid:104)Tu, Tw, ou, ow(cid:105) ∈ Ψ(ν1) and
(cid:104)Tw, Tv, ow, ov(cid:105) ∈ Ψ(ν2).
If (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(µ), then there exist an assignment Aµ of (Gµ, Dµ) and
a planar embedding Eµ of Gµ consistent with Aµ such that Aµ(u) = Pertµ(Tu),
Aµ(v) = Pertµ(Tv), B(E(cid:63)
µ(u)) is clockwise (counter-clockwise) in Tu if ou = 0
(ou = 1), and B(E(cid:63)
µ(v)) is clockwise (counter-clockwise) in Tv if ov = 0 (ov = 1).
20
ν1
ν1
ν1
ν1
ν1
ν2
(w) and E(cid:63)
ν2
(w)) and B(E(cid:63)
Let Eν1 and Eν2 be the embeddings of Gν1 and Gν2 induced by Eµ, respectively,
and let Tw = Aµ(w). Observe that E(cid:63)
(w) are disjoint consecutive
sets of Tw sharing the same boundary in Tw. Also, observe that Eµ(w) is an
ν1
extension of both Eν1(w) and Eν2 (w). By Lemma 2, Eν1 (w) and Eν2(w) are
not incompatible, and hence B(E(cid:63)
(w)) are both clockwise or
both counter-clockwise. We set ow = 0 if they are both clockwise, and ow = 1
otherwise. For every vertex x of Gν1 different from w, we set Aν1(x) = Aµ(x);
for w we set Aν1(w) = Pertν1(Tw). Since Eν1 is consistent with Aν1 and Eν1(u) =
Eµ(u), B(E(cid:63)
(u)) is clockwise (counter-clockwise) in Tu if ou = 0 (if ou = 1). By
observing that B(E(cid:63)
(w)) is clockwise (counter-clockwise) if ow = 0 (if ow = 1),
we have that (cid:104)Tu, Tw, ou, ow(cid:105) ∈ Ψ(ν1). The same argument can be used to show
that (cid:104)Tu, Tw, ou, ow(cid:105) ∈ Ψ(ν2). It follows that if (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(µ), there
exist a tree Tw ∈ D(w) and an ow ∈ {0, 1} such that (cid:104)Tu, Tw, ou, ow(cid:105) ∈ Ψ(ν1)
and (cid:104)Tw, Tv, ow, ov(cid:105) ∈ Ψ(ν2).
For the converse, assume that there exist a tree Tw ∈ D(w) and an ow ∈
{0, 1}, such that θ1 = (cid:104)Tu, Tw, ou, ow(cid:105) ∈ Ψ(ν1) and θ2 = (cid:104)Tw, Tv, ow, ov(cid:105) ∈
Ψ(ν2). By definition, there exist assignments Aν1 and Aν2 of (Gν1, Dν1 ) and
(Gν2, Dν2) respectively, and two planar embeddings Eν1 and Eν2 that are consis-
tent with Aν1 and Aν2 respectively, such that Aν1(u) = Pertν1 (Tu), Aν1(w) =
Pertν1(Tw), B(E(cid:63)
(u)) is clockwise (counter-clockwise) in Tu if ou = 0 (ou = 1),
B(E(cid:63)
(w)) is clockwise (counter-clockwise) in Tw if ow = 0 (ow = 1), Aν2(w) =
Pertν2(Tw), Aν2 (v) = Pertν2 (Tv), B(E(cid:63)
(w)) is clockwise (counter-clockwise) in
Tw if ow = 0 (ow = 1), and B(E(cid:63)
(v)) is clockwise (counter-clockwise) in Tv if
ov = 0 (ov = 1). We define an assignment Aµ and a planar embedding Eµ of Gµ
consistent with Aµ such that Aµ(u) = Pertµ(Tu), Aµ(v) = Pertµ(Tv), B(E(cid:63)
µ(u))
is clockwise (counter-clockwise) in Tu if ou = 0 (ou = 1), and B(E(cid:63)
µ(v)) is
clockwise (counter-clockwise) in Tv if ov = 0 (ov = 1). Embedding Eµ of Gµ
is obtained by merging Eν1 and Eν2 as follows. For every vertex x of Gν1 dif-
ferent from w, we set Eµ(x) = Eν1(x), for every vertex y of Gν2 different from
w, we set Eµ(y) = Eν2(y). For w, since ow has the same value in θ1 and in
θ2, hence B(E(cid:63)
(w)) are not incompatible. By Lemma 2, there
exists an order of the leaves of Tw that is an extension of both Eν1(w) and
Eν2 (w): Let Eµ(w) be this order. Assignment Aµ for (Gµ, Dµ) is defined as
follows. For every vertex x of Gν1 different from w, we set Aµ(x) = Aν1(x);
for every vertex y of Gν2 different from w, we set Aµ(y) = Aν2 (y); for w
we set Aµ(w) = Tw. Since Eµ is consistent with Aµ, Eµ(u) = Eν1 (u), and
Eµ(v) = Eν2(v), B(E(cid:63)
µ(u)) is clockwise (counter-clockwise) in Tu if ou = 0 (if
ou = 1), and B(E(cid:63)
µ(v)) is clockwise (counter-clockwise) in Tv if ov = 0. Fur-
thermore, since E(cid:63)
µ(v), Aµ(u) = Pertµ(Tu) and
It follows that if there exist a tree Tw ∈ D(w) and an
ν1
Aµ(v) = Pertµ(Tv).
ow ∈ {0, 1} such that (cid:104)Tu, Tw, ou, ow(cid:105) ∈ Ψ(ν1) and (cid:104)Tu, Tw, ou, ow(cid:105) ∈ Ψ(ν2),
then (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(µ).
Set Ψ(µ) is computed from Ψ(ν1) and Ψ(ν2) by looking for pairs of tuples
(cid:104)Tu, Tw, ou, ow(cid:105) ∈ Ψ(ν1), (cid:104)Tw, Tv, ow, ov(cid:105) ∈ Ψ(ν2) sharing the same Tw and
the same value of ow. By ordering Ψ(ν1) and Ψ(ν2), Ψ(µ) is computed in
O(D2
2
(w)) and B(E(cid:63)
(u) = E(cid:63)
µ(u) and E(cid:63)
ν2
(v) = E(cid:63)
ν1
ν2
ν2
ν2
max log(Dmax)) time.
21
Lemma 4. Let µ be a P-node with children ν1, ν2, . . . , νk. Given Ψ(ν1), Ψ(ν2),
max · n) time, where Dmax =
. . . , Ψ(νk), the set Ψ(µ) can be computed in O(D2
maxv∈V D(v).
Proof. Let u and v be the poles of µ. Let Skelµ(Tu) and Skelµ(Tv) be the
skeletal FPQ-trees of Pertµ(Tu) and of Pertµ(Tv), respectively. It can be proved
that a tuple (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(µ) if and only if the following two conditions are
satisfied: (i) There exists a planar embedding Eµ of skel(µ) and a pair of skeletal
FPQ-trees Skelµ(Tu) and Skelµ(Tv) such that Eµ(u) ∈ consistent(Skelµ(Tu)) and
Eµ(v) ∈ consistent(Skelµ(Tv)); (ii) For each child νi of µ (1 ≤ i ≤ k), there exist
an orientation ou of B(E(cid:63)
(v)) such that
(cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(νi).
(u)) and an orientation ov of B(E(cid:63)
νi
νi
Let u and v be the poles of µ. Let Skelµ(Tu) and Skelµ(Tv) be the skeletal
FPQ-trees of Pertµ(Tu) and of Pertµ(Tv), respectively. We first show that a
tuple (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(µ) if and only if the following two conditions are
satisfied:
(i) There exists a planar embedding Eµ of skel(µ) and a pair of skeletal FPQ-
trees Skelµ(Tu) and Skelµ(Tv) such that Eµ(u) ∈ consistent(Skelµ(Tu)) and
Eµ(v) ∈ consistent(Skelµ(Tv));
(ii) For each child νi of µ (1 ≤ i ≤ k), there exist an orientation ou of B(E(cid:63)
(v)) such that (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(νi).
and an orientation ov of B(E(cid:63)
νi
νi
(u))
νi
If (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(µ), then there exist an assignment Aµ of (Gµ, Dµ) and a
planar embedding Eµ of Gµ consistent with Aµ. Let Skelµ(Tu) and Skelµ(Tv)
be the skeletal FPQ-trees obtained from Aµ(u) and from Aµ(v), respectively.
By definition of skeletal FPQ-tree, the planar embedding Eµ and the pair of
skeletal FPQ-trees Skelµ(Tu) and Skelµ(Tv) satisfy Condition (i). Let Eνi be
the embedding of Gνi induced by Eµ (1 ≤ i ≤ k). E(cid:63)
(u) is a consecutive set
(v) is a consecutive set of Pertµ(Tv). Note that Eµ(u)
of Pertµ(Tu) and E(cid:63)
is an extension of Eνi(u) and that Eµ(v) is an extension of Eνi(v). We can
νi
therefore define an assignment Aνi for (Gνi, Dνi) as follows: For every vertex
w of Gνi different from the poles of Gνi, we set Aνi(w) = Aµ(w); for the poles
of Gνi we set Aνi(u) = Pertνi(Tu) and Aνi (v) = Pertνi(Tv). Note that Eνi is
consistent with Aνi . Thus, there exist values ou ∈ {0, 1} and ov ∈ {0, 1} such
that (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(νi) and hence Condition (ii) is satisfied. It follows that
if (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(µ), both Condition (i) and Condition (ii) are satisfied.
Suppose now that Condition (i) and Condition (ii) are satisfied. By Condi-
tion (i), the planar embedding Eµ and the pair of skeletal FPQ-trees Skelµ(Tu)
and Skelµ(Tv) describe how to arrange the children around u and v in a planar
embedding of skel(µ), since the union of all E(cid:63)
µ(u) and
µ(v) (1 ≤ i ≤ k). By Condition (ii)
νi
the union of all E(cid:63)
there exist an assignment Aνi of (Gνi, Dνi) and a planar embedding Eνi that is
νi
consistent with Aνi . A planar embedding Eµ of Gµ is obtained by merging all
the Eνi. More precisely, for every vertex w of Gνi different from the poles, we
set Eµ(w) = Eνi(w). Concerning the poles u and v, observe that there exists
(u) coincides with E(cid:63)
(v) coincides with E(cid:63)
22
an order of the leaves of Tu that is a common extension of all Eνi(u), and an
order of the leaves of Tv that is a common extension of all Eνi (v): Let Eµ(u)
and Eµ(v) be these orders. Also, for every vertex w of Gνi different from u and
v, we set Aµ(w) = Aνi(w). For the poles u and v we set Aµ(u) = Pertµ(Tu)
and Aµ(v) = Pertµ(Tv), respectively. Thus obtaining an embedding Eµ that
is consistent with Aµ.
It follows that if Condition (i) and Condition (ii) are
satisfied, then (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(µ).
We test these conditions by solving a 2SAT problem. We create a Boolean
variable xχ for each boundary Q-node χ of either Skelµ(Tu) or Skelµ(Tv) that
encodes the orientation of χ as clockwise or counter-clockwise. For ease of
notation, we also define xχ when χ is not a Q-node but an edge. In this case,
we simply treat this as a placeholder for true, i.e., both xχ and ¬xχ are true.
In the following, we identify the Q-nodes of Skelµ(Tu) with the Q-nodes of Tu
they correspond to. We claim that the two conditions can be encoded as 2SAT
formulas over the variables xχ.
Concerning Condition (i), we note that we seek for an ordering σ of the
virtual edges such that σ ∈ consistent(skelµ(Tu)) and its reversal σr satisfies
σr ∈ consistent(skelµ(Tv)). This can be modeled as an instances of Simulta-
neous PQ-Ordering [26] that has two nodes skelµ(Tu) and skelµ(Tv) and a
reversing arc (skelµ(Tu), skelµ(Tv)) with the identity as mapping. Then the so-
lutions to this instance are exactly the pairs of circular orderings represented by
the respective trees that are the reversal of each other. The existence of a cor-
responding 2SAT formula that describes the constraints on the orientations of
the Q-nodes then follows immediately from the work of Blasius and Rutter [26,
Lemma 4], who refer to these formulas as Q-constraints.
Concerning Condition (ii), consider a child νi and let χ and χ(cid:48) denote the
boundaries of νi in Tu and Tv, respectively. Observe that the subset of val-
ues (oχ, oχ(cid:48)) ∈ {0, 1}2 for which (cid:104)Tu, Tv, oχ, oχ(cid:48)(cid:105) ∈ Ψ(νi) is a subset of {0, 1}2,
and it can hence be encoded as the satisfying assignments of a 2SAT formula ϕi
over variables xχ and xχ(cid:48). Let now χ and χ(cid:48) denote the boundaries of µ in Tu
and Tv, respectively. It follows that (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(µ) if and only if there
i=1 ϕi such that xχ = ou and xχ(cid:48) = ov.
These values of ou and ov can be computed by using a linear-time 2SAT algo-
rithm. Since the number of virtual edges is O(n), so is the number of Q-nodes,
and therefore the number of variables. Therefore, for each pair of trees (Tu, Tv)
the 2SAT formula can be constructed and solved in O(n) time. This implies the
2
time complexity in the statement.
exists a satisfying assignment of ϕµ ∧(cid:86)h
Lemma 5. Let µ be an R-node with children ν1, ν2, . . . , νk. Given Ψ(ν1), Ψ(ν2),
µ) time, where
. . . Ψ(νk), the set Ψ(µ) can be computed in O(D
Dmax = maxv∈V D(v), b is the branchwidth of Gµ, and nµ is the number of
vertices of Gµ.
max · n2
3
2 b
µ + n3
Proof. Since µ is an R-node, skel(µ) has only two possible planar embeddings.
Let u and v be the poles of µ. Let νi (1 ≤ i ≤ k) be a child of µ that corre-
sponds to a virtual edge (x, y) of T and let Tx ∈ Dµ(x). Recall that E(cid:63)
(x) is a
νi
23
µ
νi
(x).
consecutive set of leaves in Tx. If B(E(cid:63)
(x)) in Tx is a Q-node χ, by Lemma 1
there are at least two edges incident to χ that do not belong to E(cid:63)
It
νi
follows that an orientation ox of χ determines an embedding of skel(µ). We
call the pair (Tx, ox) compliant with a planar embedding Eµ of skel(µ) if ei-
ther the boundary is an edge, or if the orientation of the boundary Q-node χ
determines the embedding Eµ of skel(µ). We denote by ΨEµ(νi) the subset of
tuples (cid:104)Tx, Ty, ox, oy(cid:105) ∈ Ψ(νi) such that Tx with orientation ox and Ty with
orientation oy are both compliant with Eµ. Similarly ΨEµ(µ) is the subset of tu-
ples (cid:104)Tu, Tv, ou, ov(cid:105) ∈ Ψ(µ) whose pairs (Tu, ou) and (Tv, ov) are both compliant
with Eµ.
We show how to compute ΨEµ(µ) from the sets ΨEµ(νi) of the children νi
of µ (1 ≤ i ≤ k). Set ΨE(cid:48)
(µ) is computed analogously. Note that the set
ΨEµ(νi) can be extracted by scanning Ψ(νi) and selecting only those admissible
tuples whose pairs (Tx, ox) and (Ty, oy) are both compliant with Eµ. Since Gµ
has branchwidth b, skel(µ) is planar, it has branchwidth at most b, and we
can execute a sphere-cut decomposition of width at most b [20] of the planar
embedding Eµ of skel(µ). Such a decomposition recursively divides skel(µ) into
two subgraphs, each of which is embedded inside a topological disc having at
most b vertices on its frontier. The decomposition is described by a rooted
binary tree, called the sphere-cut decomposition tree and denoted as Tsc. The
root of Tsc is associated with skel(µ); the leaves of Tsc are the edges of skel(µ);
any internal node β of Tsc is associated with the subgraph of skel(µ) induced by
the leaves of the subtree rooted at β. Tree Tsc is such that when removing any
of its internal edges, the two subgraphs induced by the leaves in the resulting
subtrees share at most b vertices. We denote as skel(β) the subgraph associated
with a node β of Tsc and with Dβ the topological disc that separates skel(β)
from the rest of skel(µ). Note that skel(β) has at most b vertices on the frontier
of Dβ. In particular, if β is the root of Tsc, skel(β) coincides with skel(µ) and
the vertices of skel(β) on the frontier of Dβ are exactly the poles u and v of µ.
We compute ΨEµ(µ) by visiting Tsc bottom-up. We equip each node β of Tsc
with a set of tuples ΨEµ(β), each one consisting of at most b pairs of elements
(Tx, ox) such that (Tx, ox) is compliant with Eµ, and (Tx, ox) belongs to some
ΨEµ (νi). The set of tuples associated with the root of Tsc is therefore the set
ΨEµ (µ). Let β be the currently visited node of Tsc. If β is a leaf, it is associated
with an edge representing a child νi of µ in T and ΨEµ(β) = ΨEµ (νi).
1, . . . , wi
1, w1
c , . . . , wr
If β is an internal node of Tsc, we compute ΨEµ (β) from the sets of tu-
ples ΨEµ(β1) and ΨEµ (β2) associated with the two children β1 and β2 of β.
c} be the set of vertices of skel(β1) that lie
Let B1 = {w1
c} be the set of
on the frontier of Dβ1 , and let B2 = {w1
2, w1
vertices of skel(β2) that lie on the frontier of Dβ2; see Figure 12 for an il-
2} be the set of vertices of
lustration. Let {w1
1, . . . , wi
B1 ∪ B2. Also, let B = {w1
c} be the set of vertices that lie on the fron-
tier of Dβ1 ∩ Dβ2; note that B consists of at most b vertices, i.e., r ≤ b, and
B ⊆ B1 ∪ B2. A tuple (cid:104)Tw1
(cid:105) ∈
, . . . , owi
) (1 ≤ l ≤ i, 1 ≤ h ≤ r)
ΨEµ (β1) consists of pairs (Twl
, . . . , Twr
) and pairs (Twh
1, w1
c , . . . , wr
, ow1
, owh
c , . . . , wr
c , w1
2, . . . , wj
2, . . . , wj
c , . . . , wr
, . . . , Twi
, ow1
c
, . . . , owr
c
1
, Tw1
c
1
1
, owl
1
1
c
1
c
c
24
Figure 12: An example illustrating two topological discs Dβ1 and Dβ2 containing two sub-
c},
graphs skel(β1) and skel(β2). B1 = {w1
c , . . . , wr
B = {w1
c}, B2 = {w1
2, . . . , wj
1, . . . , wi
c , . . . , wr
2, w1
1, w1
c}.
c , . . . , wr
c
c
c
2
2
2
2
2
, ow1
, owq
, Tw1
, . . . , owr
, . . . , Twj
, . . . , Twr
) and pairs (Twh
that are compliant with Eµ. Similarly, a tuple (cid:104)Tw1
,
, ow1
(cid:105) ∈ ΨEµ(β2) consists of pairs (Twq
. . . , owj
, owh
)
(1 ≤ q ≤ j, 1 ≤ h ≤ r) that are compliant with Eµ. We store the tuples of
ΨEµ (β1) in a table τ1 where each entry is a tuple and each column contains a
pair (Tx, ox). A table τ2 is built analogously to store the tuples ΨEµ(β2). We
sort τ1 and τ2 according to the columns associated with the pairs (Twh
, owh
)
(1 ≤ h ≤ r) and we obtain a new table τ by performing a join operation on
the columns that τ1 and τ2 have in common; we then select those tuples whose
) are compliant with Eµ. Finally, we compute the set ΨEµ(β), by
pairs (Twh
)
, owq
projecting τ on the columns associated with the pairs (Twl
(1 ≤ l ≤ i, 1 ≤ q ≤ j).
) and (Twq
, owh
, owl
2
c
c
c
1
1
2
2
c
c
c
c
max
max log(Db
Observe that τ1 consists of O(D(i+r)
Sorting of the two tables can be executed in O(Db
3
2 b
max) time. Also, selecting those tuples of τ for which (Twh
max ) tuples, and table τ2 consists of O(D(j+r)
max )
tuples. The join operation between τ1 and τ2 gives rise to a table τ that has
) tuples; since i + r ≤ b, j + r ≤ b, and i + j ≤ b, we have that
O(D(i+j+r)
2i + 2j + 2r ≤ 3b and thus i + j + r ≤ 3
2 b.
max)) time, since
i + r ≤ b, and j + r ≤ b. The join operation on the sorted tables can be executed
in O(D
) is
compliant with Eµ can be done in O(nµ) time per tuple by looking at the cyclic
c in Eµ (1 ≤ h ≤ r). It follows that the set ΨEµ (β)
order of the edges incident to wh
max · nµ) time. Since this procedure is
for a node β can be computed in O(D
repeated for every internal node of Tsc, since Tsc has O(nµ) nodes, and since
µ) time, we have that computing the set ΨEµ (µ)
Tsc can be constructed in O(n3
max · n2
3
2 b
µ + n3
can be executed in O(D
µ) time (see, e.g., [21, 22] for an algorithm to
compute Tsc). Since ΨE(cid:48)
(µ) is computed by an analogous procedure, the time
2
complexity in the statement follows.
, owh
3
2 b
µ
c
c
Theorem 3. Let (G, D) be a biconnected FPQ-choosable (multi-)graph such
that G = (V, E) and V = n. Let D(v) be the set of FPQ-trees associated
max · n2 + n3)-time algorithm to test
with vertex v ∈ V . There exists an O(D
3
2 b
25
Dβ1Dβ2wi1w11w12wj2w1cwrc whether (G, D) is FPQ-choosable planar, where b is the branchwidth of G and
Dmax = maxv∈V D(v).
Proof. While visiting T , we check the existence of the admissible tuples for a
node µ of T as shown by Lemmas 3, 4, or 5, depending on whether µ is an S-,
P-, or R-node. Recall that for any Q-node µ that is not the root of T and that
has poles u and v, we have Ψ(µ) = D(u)× D(v)×{0, 1}×{0, 1}. It follows that
the tuples that are admissible for a Q-node can be computed in O(D2
max) time
max · n) time for all Q-nodes of T . The admissible tuples for
and, hence, in O(D2
max log(Dmax)· n) time, the admissible
all S-nodes of T can be computed in O(D2
max· n2) time, and the admissible
tuples for all P-nodes can be computed in O(D2
max · n2 + n3) time. Recall that
3
2 b
tuples for all R-nodes can be computed in O(D
the SPQR-decomposition tree of a biconnected n-vertex graph can be computed
2
in O(n) time [19].
We remark that our algorithmic approach cannot be extended to simply
connected graphs, since it is based on the SPQR-decomposition of the input
graph G, that expects G to be biconnected.
6. FPQ-choosable Planarity Testing and NodeTrix Planarity Testing
The study of FPQ-Choosable Planarity Testing can be applied also to
address other planarity testing problems that can be modeled in terms of hier-
archical embedding constraints. As a proof of concept, in this section we study
the interplay between FPQ-Choosable Planarity Testing and NodeTrix
planarity testing.
A flat clustered graph is a graph for which subsets of its vertices are grouped
into clusters and no vertex belongs to two clusters. For example, Figure 2(a)
depicts a flat clustered graph with two clusters. In a NodeTrix representation,
each cluster is represented as an adjacency matrix, while the inter-cluster edges
are simple curves connecting the corresponding matrices [13, 15, 16, 17].
If
no inter-cluster edges cross, the NodeTrix representation is said to be planar.
For example, Figure 2(b) shows a planar NodeTrix representation of the flat
clustered graph of Figure 2(a).
A NodeTrix graph with fixed sides is a flat clustered graph G that admits
a NodeTrix representation where, for each inter-cluster edge e, the sides of the
matrices to which e is incident are specified as part of the input. If instead the
sides are not specified, G is a NodeTrix graph with free sides. If G admits a planar
NodeTrix representation, then we say that G is NodeTrix planar. NodeTrix
planarity testing is NP-complete both in the fixed sides scenario and in the free
sides scenario, even when the size of the matrices is bounded by a constant [16,
18]. On the positive side, it is proved in [17] that one can test in polynomial time
whether a flat clustered graph is NodeTrix planar with fixed sides if the size of
the matrices is bounded by a constant and if the graph obtained by collapsing
each cluster into a vertex has treewidth at most two. We extend this last result
to graphs having bounded treewidth (provided that the size of the clusters is
26
(a)
(b)
(c)
Figure 13: (a) A matrix Mi; (b) the matrix FPQ-tree TMi ; (c) the gadget W v replacing TMi .
bounded). To this aim we model NodeTrix planarity testing with fixed sides as
a problem of FPQ-Choosable Planarity Testing.
Let G be a NodeTrix graph with fixed sides and with clusters C1, . . . , CnC .
Each permutation of the vertices of Ci (1 ≤ i ≤ nC) corresponds to a matrix
Mi in some NodeTrix representation of G. Note that even if the side of Mi
to which each inter-cluster edge is incident to is fixed, it is still possible to
arbitrarily permute the edges incident to a same side and to a same vertex. For
example, we can permute the two edges f and g incident to the right side of
the matrix in Figure 13(a). It follows that all the possible cyclic orders of the
edges incident to Mi can be described by means of an FPQ-tree, that we shall
call the matrix FPQ-tree of Mi, denoted as TMi .
Namely, TMi consists of an F-node χc connected to 4Mi P-nodes represent-
ing the vertices of Ci; see, e.g., Figure 13(b). These P-nodes around χc appear
1, . . . , xρMi,
in the clockwise order that is defined by Mi, namely xτ
xβMi, . . . , xβ
1 , where τ , ρ, β, and λ represent the top, right, bot-
tom, and left side of Mi, respectively. Any inter-cluster edge incident to a
v (1 ≤ v ≤ Mi,
vertex v of Mi corresponds to a leaf of TMi adjacent to xs
s ∈ {τ, ρ, β, λ}).
1 , . . . , xτMi, xρ
1 , xλMi, . . . , xλ
The constraint graph of a NodeTrix graph with fixed sides G, denoted as GC,
is the FPQ-choosable multi-graph defined as follows. Graph GC has nC vertices,
each one corresponding to one of the clusters of G, and in GC there is an edge
(u, v) for each inter-cluster edge that connects the two clusters corresponding
to u and to v in G. Each vertex v of GC is associated with a set D(v) of Cv!
FPQ-trees. More precisely, for each permutation π of the vertices of Cv, let M π
v
be the matrix associated with Cv. For each such a permutation, we equip v
with the matrix FPQ-tree of M π
v .
Figure 14(a) shows a NodeTrix graph with fixed sides G whose constraint
graph is depicted in Figure 14(b). In Figure 14(b), each vertex vi of GC (1 ≤
i ≤ 4) represents a 3 × 3 matrix Mi of the graph G of Figure 14(a); hence, vi
is associated with six FPQ-trees, one for each possible permutation of the rows
27
aijhgfedck123hajigfedckχcxτ1xτ2xτ3xβ1xβ2xβ3xλ1xλ2xλ3xρ1xρ2xρ3kcdefghijat1t2t3r1r2r3b3b2b1l3l1l2 (a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Figure 14: (a) A NodeTrix graph with fixed sides G; (b) the constraint graph GC of G; (c)-(h)
the FPQ-trees associated with the vertex v1 of GC .
28
abcdefghijklmABCDEFGHIJKLM2M1M3M4abcfekgdhimjlv1v2v3v4aTα1fedcbaTβ1fdecbaTγ1efdbcaTδ1edfbcaTε1dfecbaTη1defbc and the columns of Mi. For example, the FPQ-trees of v1 are those depicted in
Figure 14(c)-(h).
Theorem 4. Let G be a flat clustered n-vertex graph whose clusters have size at
most k. Let t be the treewidth of G. If the constraint graph of G is biconnected,
4 t · n2 + n3)-time algorithm to test whether G is NodeTrix
there exists an O(k! 9
planar with fixed sides.
Proof. Let nC be the number of vertices of GC. We show that G is NodeTrix
planar with fixed sides if and only if GC is FPQ-choosable planar. This, together
with the observation that nC ∈ O(n), Theorem 3, and the fact that if a graph
has bounded branchwidth b it has treewidth at most(cid:4) 3
2 b(cid:5) − 1 [29], implies the
statement.
If GC is FPQ-choosable planar, there exists a tuple of FPQ-trees θnC that
is admissible for GC. Therefore, one can associate each vertex of GC with its
FPQ-tree in θnC , execute the embedding constrained planarity testing algorithm
by Gutwenger et al. [12] and obtain a positive answer. By this technique, each
FPQ-tree Tu is replaced by a gadget W u that is built as follows. Each F-node
χ is replaced with a wheel Hχ whose external cycle has a vertex for each edge
incident to χ. Each vertex of Hχ has an edge, called spoke, that is incident to it
and that is embedded externally to the wheel. For example, Figure 13(c) shows
the gadget corresponding to the FPQ-tree of Figure 13(b). Each P-node ρ of Tu
is represented in the gadget W u as a vertex vρ that has a spoke for each edge
of ρ. For example, the P-node ρ with incident edges f and g of Figure 13(b)
is represented in Figure 13(c) with a vertex vρ with two spokes f and g. By
performing this replacement for each FPQ-tree of θnC and by connecting the
spokes of the gadgets that correspond to the same edge, we obtain a graph GC.
Gutwenger et al. [12] show that GC is planar with the embedding constraints
if and only if GC is a planar graph.
In order to obtain a planar NodeTrix
representation, we compute a planar embedding of GC and replace each gadget
W u (corresponding to cluster Cu) by a matrix as follows. Let W u
x be a wheel
of W u, and let t1, t2, . . . , tCu, r1, r2, . . . , rCu, bCu, . . . , b2, b1, lCu, . . . , l2,
and l1 be the spokes that are encountered by walking clockwise along the cycle
of W u
x with a matrix Mu whose vertices are placed according
to the permutation v1, . . . , vCu. The spokes of W u that are adjacent to ti
(i = 1, . . . ,Cu) are connected to vi on the top side of Mu, analogously for
the spokes that are adjacent to ri, bi, and li, are connected to vi on the right,
bottom, or left side of Mu, respectively.
x . Replace W u
By performing this replacement for each gadget of GC, we obtain a planar
NodeTrix representation G of the FPQ-choosable planar graph GC. It follows
that, if GC is FPQ-choosable planar, G is NodeTrix planar with fixed sides.
We now show that if G is NodeTrix planar with fixed sides, then GC is FPQ-
choosable planar. Let Γ be a planar NodeTrix representation of G. Replace each
matrix Mv of Γ by a vertex v, and connect to it all the inter-cluster edges that
are incident to Mv. We obtain a planar drawing Γ(cid:48) such that the cyclic order
of the edges incident to each vertex v of Γ(cid:48) reflects the cyclic order of the edges
29
incident to matrix Mv in Γ. Such an order corresponds to one of the Cv! FPQ-
trees associated with v in GC (Cv is the number of rows and columns of Mv).
Therefore, GC is FPQ-choosable planar
2
Corollary 2. Let G be a flat clustered n-vertex graph whose clusters have size
at most k and whose vertices have degree at most d. Let t be the treewidth of G.
4 t·n2 +n3)-
If the constraint graph of G is biconnected, there exists an O((k!4kd) 9
time algorithm to test whether G is NodeTrix planar with free sides.
Proof. The number of possible configurations in which the inter-cluster edges
are incident to the matrices is k!4kd. Therefore, by Theorem 4 the statement
2
follows.
7. Concluding Remarks and Open Problems
In this paper we have studied the problem of testing when a graph G is pla-
nar subject to hierarchical embedding constraints. These constraints are given
as part of the input by equipping each vertex of G with a set of FPQ-trees.
While the problem is NP-complete even for sets of FPQ-trees having cardinal-
ity bounded by a constant and it is W[1]-hard parameterized by tree-with, for
biconnected graphs it becomes fixed-parameter tractable if parameterized by
both the treewidth and by the maximum number of FPQ-trees associated with
a vertex. Besides being interesting on its own right, FPQ-Choosable Pla-
narity Testing can be used to model and study other graph planarity testing
problems. As a proof of concept, we have applied our results to the study of
NodeTrix planarity testing of clustered graphs.
We mention three open problems that in our opinion are worth future studies.
• Theorem 1 is based on a reduction that associates six FPQ-trees to each
vertex of a suitable instance of FPQ-Choosable Planarity Testing.
It would be interesting to study the complexity of FPQ-Choosable Pla-
narity Testing when every vertex is associated with less than six FPQ-
trees. We recall that FPQ-Choosable Planarity Testing can be
solved in polynomial time if Dmax = 1 [12].
• It would be interesting to improve the time complexity stated by Theo-
rem 3.
• It would be interesting to extend Theorem 3 to simply connected graphs.
• It would be interesting to apply our approach to other problems of pla-
narity testing related with hybrid representations of clustered graphs in-
cluding, for example, intersection-link representations and (k, p)-planar
representations (see, e.g., [30, 31]).
Funding. This work was partially supported by: (i) MIUR, the Italian Min-
istry of Education, University and Research, under grant 20174LF3T8 AHeAD:
30
efficient Algorithms for HArnessing networked Data; (ii) Dipartimento di In-
gegneria dell'Universit`a degli Studi di Perugia, under grants RICBASE2017WD
and RICBA18WD: "Algoritmi e sistemi di analisi visuale di reti complesse e
di grandi dimensioni"; (iii) German Science Foundation (DFG), under grant
Ru 1903/3-1.
References
[1] M. Patrignani, Planarity testing and embedding, in: Handbook on Graph
Drawing and Visualization., 2013, pp. 1 -- 42.
[2] T. Blasius, I. Rutter, A new perspective on clustered planarity as a com-
binatorial embedding problem, Theor. Comput. Sci. 609 (2016) 306 -- 315.
doi:10.1016/j.tcs.2015.10.011.
[3] P. F. Cortese, G. Di Battista, Clustered planarity, in: Proceedings of the
21st ACM Symposium on Computational Geometry, Pisa, Italy, June 6-8,
2005, 2005, pp. 32 -- 34. doi:10.1145/1064092.1064093.
[4] W. Didimo, G. Liotta, M. Patrignani, Hv-planarity: Algorithms and com-
plexity, J. Comput. Syst. Sci. 99 (2019) 72 -- 90. doi:10.1016/j.jcss.
2018.08.003.
[5] Q. Feng, R. F. Cohen, P. Eades, Planarity for clustered graphs, in: Al-
gorithms - ESA '95, Third Annual European Symposium, Corfu, Greece,
September 25-27, 1995, Proceedings, 1995, pp. 213 -- 226. doi:10.1007/
3-540-60313-1\_145.
[6] A. Garg, R. Tamassia, Upward planarity testing, Order 12 (2) (1995) 109 --
133.
[7] A. Garg, R. Tamassia, On the computational complexity of upward and
rectilinear planarity testing, SIAM J. Comput. 31 (2) (2001) 601 -- 625. doi:
10.1137/S0097539794277123.
[8] P. Angelini, G. Di Battista, F. Frati, V. Jel´ınek, J. Kratochv´ıl, M. Pa-
trignani, I. Rutter, Testing planarity of partially embedded graphs, ACM
Trans. Algorithms 11 (4) (2015) 32:1 -- 32:42. doi:10.1145/2629341.
[9] V. Jel´ınek, J. Kratochv´ıl, I. Rutter, A Kuratowski-type theorem for pla-
narity of partially embedded graphs, Comput. Geom. 46 (4) (2013) 466 -- 492.
doi:10.1016/j.comgeo.2012.07.005.
[10] C. Dornheim, Planar graphs with topological constraints, J. Graph Algo-
rithms Appl. 6 (1) (2002) 27 -- 66. doi:10.7155/jgaa.00044.
[11] G. Da Lozzo, I. Rutter, Approximation algorithms for facial cycles in planar
embeddings, in: W. Hsu, D. Lee, C. Liao (Eds.), Proceedings of the 29th
International Symposium on Algorithms and Computation (ISAAC'18),
31
Vol. 123 of LIPIcs, Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik,
2018, pp. 41:1 -- 41:13. doi:10.4230/LIPIcs.ISAAC.2018.41.
[12] C. Gutwenger, K. Klein, P. Mutzel, Planarity testing and optimal edge
insertion with embedding constraints, J. Graph Algorithms Appl. 12 (1)
(2008) 73 -- 95. doi:10.7155/jgaa.00160.
[13] N. Henry, J. Fekete, M. J. McGuffin, NodeTrix: A hybrid visualization of
social networks, IEEE Trans. Vis. Comput. Graph. 13 (6) (2007) 1302 --
1309. doi:10.1109/TVCG.2007.70582.
[14] Citevis: Visualizing citations among infovis conference papers.
URL http://www.cc.gatech.edu/gvu/ii/citevis
[15] V. Batagelj, F. Brandenburg, W. Didimo, G. Liotta, P. Palladino, M. Pa-
trignani, Visual analysis of large graphs using (X,Y)-Clustering and hybrid
visualizations, IEEE Trans. Vis. Comput. Graph. 17 (11) (2011) 1587 -- 1598.
doi:10.1109/TVCG.2010.265.
[16] G. Da Lozzo, G. Di Battista, F. Frati, M. Patrignani, Computing Node-
Trix representations of clustered graphs, Journal of Graph Algorithms and
Applications 22 (2) (2018) 139 -- 176. doi:10.7155/jgaa.00461.
[17] E. Di Giacomo, G. Liotta, M. Patrignani, I. Rutter, A. Tappini, NodeTrix
planarity testing with small clusters, Algorithmica 81 (9) (2019) 3464 -- 3493.
doi:10.1007/s00453-019-00585-6.
[18] J. J. Besa Vial, G. Da Lozzo, M. T. Goodrich, Computing k-modal em-
beddings of planar digraphs, in: 27th Annual European Symposium on Al-
gorithms, ESA 2019, September 9-11, 2019, Munich/Garching, Germany,
2019, pp. 19:1 -- 19:16. doi:10.4230/LIPIcs.ESA.2019.19.
[19] G. Di Battista, R. Tamassia, On-line planarity testing, SIAM Journal on
Computing 25 (5) (1996) 956 -- 997. doi:10.1137/S0097539794280736.
[20] F. Dorn, E. Penninkx, H. L. Bodlaender, F. V. Fomin, Efficient exact
algorithms on planar graphs: Exploiting sphere cut decompositions, Algo-
rithmica 58 (3) (2010) 790 -- 810. doi:10.1007/s00453-009-9296-1.
[21] Q. Gu, H. Tamaki, Optimal branch-decomposition of planar graphs in
doi:
O(n 3) time, ACM Trans. Algorithms 4 (3) (2008) 30:1 -- 30:13.
10.1145/1367064.1367070.
[22] P. D. Seymour, R. Thomas, Call routing and the ratcatcher, Combinatorica
14 (2) (1994) 217 -- 241. doi:10.1007/BF01215352.
[23] S. Arumugam, A. Brandstadt, T. Nishizeki, K. Thulasiraman, Handbook
of graph theory, combinatorial optimization, and algorithms, Chapman and
Hall/CRC, 2016.
32
[24] G. Di Battista, P. Eades, R. Tamassia, I. G. Tollis, Graph Drawing, Pren-
tice Hall, Upper Saddle River, NJ, 1999.
[25] K. S. Booth, G. S. Lueker, Testing for the consecutive ones property, in-
terval graphs, and graph planarity using PQ-tree algorithms, J. Comput.
Syst. Sci. 13 (3) (1976) 335 -- 379. doi:10.1016/S0022-0000(76)80045-1.
[26] T. Blasius, I. Rutter, Simultaneous PQ-ordering with applications to con-
strained embedding problems, ACM Trans. Algorithms 12 (2) (2016) 16:1 --
16:46. doi:10.1145/2738054.
[27] I. Holyer, The NP-completeness of edge-coloring, SIAM J. Comput. 10 (4)
(1981) 718 -- 720. doi:10.1137/0210055.
[28] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk,
M. Pilipczuk, S. Saurabh, Parameterized Algorithms, Springer, 2015. doi:
10.1007/978-3-319-21275-3.
[29] N. Robertson, P. D. Seymour, Graph minors. X. obstructions to tree-
decomposition, J. Comb. Theory, Ser. B 52 (2) (1991) 153 -- 190. doi:
10.1016/0095-8956(91)90061-N.
[30] P. Angelini, G. Da Lozzo, G. Di Battista, F. Frati, M. Patrignani, I. Rutter,
Intersection-link representations of graphs, Journal of Graph Algorithms
and Applications 21 (4) (2017) 731 -- 755. doi:10.7155/jgaa.00437.
[31] E. Di Giacomo, W. J. Lenhart, G. Liotta, T. W. Randolph, A. Tappini, (k,
p)-planarity: A relaxation of hybrid planarity, in: WALCOM: Algorithms
and Computation - 13th International Conference, WALCOM 2019, Guwa-
hati, India, February 27 - March 2, 2019, Proceedings, 2019, pp. 148 -- 159.
doi:10.1007/978-3-030-10564-8\_12.
33
|
1204.4835 | 1 | 1204 | 2012-04-21T19:36:06 | Succinct Indices for Range Queries with applications to Orthogonal Range Maxima | [
"cs.DS"
] | We consider the problem of preprocessing $N$ points in 2D, each endowed with a priority, to answer the following queries: given a axis-parallel rectangle, determine the point with the largest priority in the rectangle. Using the ideas of the \emph{effective entropy} of range maxima queries and \emph{succinct indices} for range maxima queries, we obtain a structure that uses O(N) words and answers the above query in $O(\log N \log \log N)$ time. This is a direct improvement of Chazelle's result from FOCS 1985 for this problem -- Chazelle required $O(N/\epsilon)$ words to answer queries in $O((\log N)^{1+\epsilon})$ time for any constant $\epsilon > 0$. | cs.DS | cs |
Succinct Indices for Range Queries with
applications to Orthogonal Range Maxima(cid:63)
Arash Farzan1, J. Ian Munro2, and Rajeev Raman3
1 Max-Planck-Institut fur Informatik, Saarbucken, Germany.
2 University of Waterloo, Canada.
3 University of Leicester, UK.
Abstract. We consider the problem of preprocessing N points in 2D,
each endowed with a priority, to answer the following queries: given
a axis-parallel rectangle, determine the point with the largest prior-
ity in the rectangle. Using the ideas of the effective entropy of range
maxima queries and succinct indices for range maxima queries, we ob-
tain a structure that uses O(N ) words and answers the above query in
O(lg N lg lg N ) time. This is a direct improvement of Chazelle's result
from 1985 [11] for this problem -- Chazelle required O(N/) words to
answer queries in O((lg N )1+) time for any constant > 0.
1
Introduction
Range searching is one of the most fundamental problems in computer science
with important applications in areas such as computational geometry, databases
and string processing. The input is a set of N points in general position in Rd
(we focus on the case d = 2), where each point is associated with satellite data,
and an aggregation function defined on the satellite data. We wish to preprocess
the input to answer queries of the following form efficiently: given any 2D axis-
aligned rectangle R, return the value of the aggregation function on the satellite
data of all points in R. Researchers have considered range searching with respect
to diverse aggregation functions such as emptiness checking, counting, reporting,
minimum/maximum, etc. [18]. In this paper, we consider the problem of range
maximum searching (the minimum variant is symmetric), where the satellite
data associated with each point is a numerical priority, and the aggregation
function is "arg max", i.e., we want to report the point with the maximum
priority in the given query rectangle. This aggregation function is the canonical
one to study, among those of the "commutative semi-group" class [13, 11].
Our primary concern is the space requirement of the data structure -- we
aim for linear-space data structures, namely those that occupy O(N ) words --
and seek to minimize query time subject to this constraint. The space usage is
a fundamental concern in geometric data structures due to very large data vol-
umes; indeed, space usage is a main reason why range searching data structures
like quadtrees, which have poor worst-case query performance, are preferred in
(cid:63) Work done while Farzan was employed by, and Raman was visiting, MPI.
many practical applications over data structures such as range trees, which have
asymptotically optimal query performance. Space efficient solutions to range
searching date to the work of Chazelle [11] over a quarter century ago, and
Nekrich [19] gives a nice survey of much of this work. Recently there has been
a flurry of activity on various aspects of space-efficient range reporting, and for
some aggregation functions there has even been attention given to the constant
term within the space usage [5, 19].
We now formalize the problem studied by our paper, as well as those of [11, 9,
16]. We assume input points are in rank space: the x-coordinates of the n points
are {0, . . . , N − 1} = [N ], and the y-coordinates are given by a permutation
υ : [N ] → [N ], such that the points are (i, υ(i)) for i = 0, . . . , N − 1. The
priorities of the points are given by another permutation π such that π(i) is the
priority of the point (i, υ(i)). The reduction to rank space can be performed in
O(lg N ) time with a linear space structure even if the original and query points
are points in R2 [13, 11]. The query rectangle is specifed by two points from
[N ] × [N ] and includes the boundaries (see Fig. 1(R)). Analogous to previous
work, we also assume the word-RAM model with word size Θ (lg N ) bits4.
Fig. 1. 2-sided and 4-sided range maximum queries. The numbers with the points
represent their priorities, and the unshaded points are the answers.
Range maximum searching is a well-studied problem (see Table 1). Chazelle [11]
gave a few space/time tradeoffs covering a broad spectrum. To the best of our
knowledge, the solution with the lowest query time that uses only O(N ) words
is still that of Chazelle [11], who gave a query time that is polylogarithmic in
N . More precisely, he gave a data structure of size O( 1
N ) words with query
time O(lg1+ N ) for any fixed > 0. Other recent results on the range maximum
problem are as follows. Karpinski et al. [16] studied the problem of 3D five-sided
range emptiness queries which is closely related to range maximum searching in
2D. As observed in [9], their solution yields a query time of (lg lg N )O(1) with an
index of size N (lg lg N )O(1) words. Chan et al. [9] currently give the best query
time of O (lg lg N ), but this is at the expense of using O (N lg N ) words, for any
fixed > 0. However, there has been no improvement in the running time for
linear-space data structures. In this paper, we improve Chazelle's long-standing
4 lg x = log2x
894613275894613275 (cid:16)
Citation
Chazelle'88 [11]
Chan et al.'10 [9]
Karpinski et al.'09 [16] O
Chazelle'88 [11]
Chazelle'88 [11]
NEW
O (N )
Size (in words)
O (N lg N )
O (N lg N )
N (lg lg N )O(1)(cid:17)
N(cid:1)
O(cid:0) 1
O (N lg lg N )
Query time
O (lg N )
O (lg lg N )
O(cid:0)(lg lg N )2(cid:1)
O(cid:0)lg1+ N(cid:1)
O (lg N lg lg N )
O (lg N lg lg N )
Table 1. Space/time tradeoffs for 2D range maximum searching in the word RAM.
result by giving a data structure of O(N ) words and reducing the query time from
polylogarithmic to "almost" logarithmic, namely, O (lg N lg lg N ). Although our
primary focus is on 4-sided queries, which specify a rectangle that is bounded
from all sides, we also need to consider 2-sided and 3-sided queries, which are
"open" on two and one side respectively (thus a 2-sided query is specified by a
single point (i, j) -- see Fig. 1(L) -- and a 3-sided query by two points (i, j) and
(k, l) where either i = k or j = l). Our solution recursively divides the points
into horizontal and vertical slabs, and a query rectangle is decomposed into
smaller 2-sided, 3-sided, and 4-sided queries. A key intermediate result is the
data structure for 2-sided queries. The 2-sided sub-problems are partitioned into
smaller sub-problems, which are stored in a "compressed" format that is then
"decompressed" on demand. The "compression" uses the idea that to answer
2-sided range maxima queries on a problem of size m, one need not store the
entire total order of priorities using Θ (m lg m) bits: O (m) bits suffice, i.e., the
effective entropy [14] of 2-sided queries is low. This does not help immediately,
since Θ (m lg m) bits are needed to store the coordinates of the points compris-
ing these sub-problems: to overcome this bottleneck we use ideas from succinct
indices [2] -- namely, we separate the storage of point coordinates from the data
structure for answering range-maximum queries. The latter data structure does
not store point coordinates but instead obtains them as needed from a global
data structure.
We solve 3-sided and 4-sided subqueries by recursion, or by using structures
for range maximum queries on matrices [22, 6]. When recursing, we cannot afford
the space required to store the structures for rank space reduction for each
such subproblem: a further key idea is to use a single global structure to map
references to points within recursive subproblems back to the original points.
By reusing ideas from the data structure for 2-sided queries, we obtain two
stand-alone results on succinct indices for 2-sided range maxima queries. In a
succinct index, we wish to answer queries on some data, and it is assumed that
the succinct index is not charged for the space required to store the data itself,
but only for any additional space it requires to answer the queries. However, the
succinct index can access the data only through a specific interface (see [2] for a
discussion of the advantages of succinct indices). In our case, given N points in
rank space, together with priorities, we wish to answer 2-sided range-maximum
queries under the condition that the point coordinates are stored "elsewhere" --
the data structure is not "charged" for the space needed to store the points
"elsewhere" -- and are assumed to be available to the data structure in one of
two ways:
-- Through an orthogonal range reporting query. Here, we assume that a query
that results in k points being reported takes T (N, k) time. We assume that
T is such that T (N, O(k)) = O(T (N, k)).
-- As in Bose et al. [4], we assume that the point coordinates are stored in
read-only memory, permitting random access to the coordinates of the i-th
point. However, the ordering of the points is specified by the data structure,
which we call the permuted-point model.5
In both cases, we are able to achieve O(N )-bit indices with fast query time,
namely O(lg lg N · (lg N + T (N, lg N ))) time, and O(lg lg N ) time, respectively.
The paper is organized as follows. We first describe some building blocks used
in Section 2. Section 3 is devoted to our main result, and Section 4 describes the
succinct index results.
2 Preliminaries
Our result builds upon a number of relatively new results, all of which are
essential in one way or the other to obtain the final bound. In order to support
mapping between recursive sub-problems, we use the following primitives on a
set S of N points in rank space. A range counting query reports the number of
points within a query rectangle:
Lemma 1 ([15]). Given a set of N points in rank space in two dimensions,
there is a data structure with O(N ) words of space that supports range counting
queries in O (lg N/ lg lg N ) time.
A range reporting structure supports the operation of listing the coordinates of
all points within a query rectangle. We use the following consequence of a result
of Chan et al. [9]:
Lemma 2 ([9]). Given a set of N points in rank space in two dimensions,
there is a data structure with O(N ) words of space that supports range reporting
queries in O
(1 + k) lg1/3 N
time where k is the number of points reported.
(cid:16)
(cid:17)
The range selection problem is as follows: given an input array A of size N , to
preprocess it so that given a query (i, j, k), with 1 ≤ i ≤ j ≤ N , we return
an index i1 such that A[i1] is the k-th smallest of the elements in the subarray
A[i], A[i + 1], . . . , A[j].
Lemma 3 ([7]). Given an array of size N , there is a data structure with O(N )
words of space that supports range selection queries in O(lg N/ lg lg N ) time.
5 Clearly, a succinct index for this problem is interesting only if it uses o(N ) words
= o(N lg N ) bits of space, so if the points are stored in arbitrary order in read-only
memory, the succinct index cannot itself store the permutation that re-orders the
points.
3 The data structure
In this section we show our main result:
Theorem 1. Given N points in two-dimensional rank space, and their prior-
ities, there is a data structure that occupies O(N ) words of space and answers
range maximum queries in O(lg N lg lg N ) time.
We first give an overview of the data structure. We begin by storing all the
points (using their input coordinates) once each in the structures of Lemmas 1
and 2. We also store an instance of the data structure of Lemma 3 once each
for the arrays X and Y , where X[i] = ν(i) and Y [i] = ν−1(i) for i ∈ N (X
stores the y-coordinates of the points in order of increasing x-coordinate, and Y
the x-coordinates in order of increasing y-coordinate). These four "global" data
structures use O(N ) words of space in all.
We recursively decompose the problem `a la Afshani et al. [1]. Let n be the
recursive problem size (initially n = N ). Given a problem of size n, we divide
the problem into n/k mutually disjoint horizontal slabs of size n by k, and n/k
mutually disjoint vertical slabs of size k by n. A horizontal and vertical slab
intersect in a square of size k × k. We recurse on each horizontal or vertical slab:
observe that each horizontal or vertical slab has exactly k points in it, and is
treated as a problem of size k -- i.e. it is logically comprised of two permutations
υ and π on [k] (Fig. 2(L); Sec. 3.1). Clearly, given a slab in a problem of size n
containing k points, we require some kind of mapping between the coordinates
in the slab (which in one dimension will be from [n]) and the recursive problem
in order to view the slab as a problem of size k. A key to the space-efficiency of
our data structure is that this mapping is not explicitly stored, and is achieved
through a slab-rank operation. The slab-select problem is a generalization of the
inverse of the slab-rank problem (Sec. 3.2).
The given query rectangle is decomposed into a number of disjoint 2-sided,
3-sided and 4-sided queries, based upon the decomposition of the input into
slabs. There are three kinds of terminal queries which do not generate further
recursive problems: all 2-sided queries are terminal, 4-sided queries whose sides
are slab boundaries are terminal, and all queries that reach slabs at the bottom
of the recursion are terminal. The problems (or data structures) that involve
answering terminal queries (or are used to answer terminal queries) are also
termed terminal. Each terminal query produces some candidate points: the set
of all candidate points must contain the final answer. A key invariant needed to
achieve the space bound is that all terminal problems of size n -- except those at
the bottom of the recursion -- use space o(n lg n) bits (Sec. 3.1).
Clearly, since storing the input permutation representing the point sets in
terminal problems takes Θ(n lg n) bits, we do not store them explicitly. Instead,
the terminal data structures are succinct indices -- the points that comprise them
are accessed by means of queries to a single global data structure. The terminal
4-sided problems use an index due to [6], while the 2-sided problems reduce the
range maximum query to planar point location. Although there is a succinct
index for planar point location [4], this essentially assumes that the points that
comprise the planar subdivision are stored explicitly ("elsewhere") and permuted
in a manner specified by their data structure. In our case, if the points are repre-
sented explictly, they would occupy Θ(N lg N/ lg lg N ) words across all recursive
problems. A key to our approach is an implicit representation of the planar sub-
division in the recursive problems, relevant parts of which are recomputed from
a "compressed" representation at query time (Sec. 3.4); the other key step is to
note that O(n) bits suffice to encode the priority information needed to answer
2-sided queries in a problem of size n (Sec. 3.3).
3.1 A recursive formulation and its space usage
The recursive structure is as follows. Let L = lg N , and consider a recursive
problem of size n (at the top level n = N ). We assume that N is a power of 2,
as are a number of expressions which represent the size of recursive problems.
This can readily be achieved by replacing real-valued parameters x ≥ 1 by
2(cid:98)lg x(cid:99) or 2(cid:100)lg x(cid:101) without affecting the asymptotic complexity. Unless we have
reached the bottom of the recursion, we partition the input range [n] × [n] into
nL and also into mutually disjoint
mutually disjoint vertical slabs of width k =
horizontal slabs of height k =
nL -- each such slab contains k points and can
be logically viewed as a recursive problem of size k (see Fig. 2(L)). Observe that
the input is divided into (n/k)2 = n/L squares of size k × k, each representing
the intersection of a horizontal slab with a vertical one. We need to answer either
2-sided, 3-sided or 4-sided queries on this problem.
√
√
The data structures associated with the current recursive problem are:
data structure which uses O (n lg n) bits of space and has query time O(cid:0)(lg n)2(cid:1).
-- for problems at the bottom of the recursion, we store an instance of Chazelle's
-- For 2-sided queries (which are terminal) in non-terminal problems we use
√
the data structure with space usage O(n
lg n) bits described in Sec. 3.4.
-- 3- and 4-sided queries, all of whose sides are slab boundaries, are square-
aligned : they exactly cover a rectangular sub-array of squares. For such
queries, we use a O(n)-bit data structure comprising.
• A n/k × n/k matrix containing the (top-level) x and y coordinates and
priority of the maximum point (if any) in each square. This uses O(n/L·
L) = O(n) bits.
• The data structure of [22, 6] for answering 2D range maximum queries
on the elements in the above matrix. This also uses O(n) bits.
Finally, each recursive problem has O(lg N ) = O(L) bits of "header" informa-
tion, containing, e.g., the bounding box of the problem in the top-level coordinate
system. Ignoring the header information, the space usage is given by:
S(n) = 2(cid:112)n/LS(
√
nL) + O(n(cid:112)lg n),
(cid:113)
(cid:18)
) + O
2rN
(cid:19)
lg(N 1/2r L1−1/2r )
.
which after r levels of recursion becomes:
S(N ) = 2r N 1−1/2r
L1−1/2r S(N 1/2r
L1−1/2r
level, the problems are of size O(cid:0)(lg N )2(cid:1) and Ω(cid:0)(lg N )1.5(cid:1) and the second term
The recursion is terminated for the first level r where 2r ≥ lg N/ lg lg N . At this
in the space usage becomes O(N lg N ) bits. Applying S(n) = O(n lg n) for the
base case, we see that the first term is O((lg N/ lg lg N )· N · lg lg N ) = O(N lg N )
bits, and the space used by the header information is indeed negligible.
Fig. 2. The recursive decomposition of the input (L) and queries (R). In (R), shaded
problems are terminal problems. The 4-sided query is decomposed into a square-aligned
4-sided query in the middle and four recursive 3-sided queries, two in horizontal slabs
and two in vertical slabs. The 3-sided query is decomposed into two 2-sided queries
in vertical slabs, a square-aligned 3-sided query and one recursive 3-sided query in a
horizontal slab.
We now discuss the time complexity. In general, we have to answer either
2-sided, 3-sided or 4-sided queries on a slab. Note that (see Fig. 2(R)):
-- A 2-sided query is terminal and generates one candidate.
-- A 3-sided query results in at most one recursive 3-sided query on a slab
(generating no candidates) at most two 2-sided queries on slabs, and at
most one square-aligned 3-sided query (generating one candidate).
-- A 4-sided query either results in a recursive 4-sided query on a slab (gener-
ating no candidates) or generates at most one square-aligned 4-sided query
(generating one candidate), plus up to four 3-sided queries in slabs.
Since each 3-sided query only generates one recursive 3-sided query, the number
of recursive problems solved, and hence the number of candidates, is O(r) =
O(lg lg N ). The time complexity will depend on the cost of mapping query points
and candidates between the problems and their recursive sub-problems and the
cost of solving the terminal problems; the former is discussed next.
3.2 The slab-rank and slab-select problems
The input to each recursive problem of size n is given in local coordinates (i.e.
from [n] × [n]). Upon decomposing the query to this problem, we need to solve
the following slab-rank problem (with a symmetric variant for vertical slabs):
Given a point p = (i∗, j∗) in top-level coordinates, which is mapped
to (i, j) in a recursive problem of size n, such that (i, j) that lies in a
horizontal slab of size n× k, map (i, j) to the appropriate position (i(cid:48), j(cid:48))
in the size k problem represented by this slab.
We formalize the "inverse" slab-select problem as follows:
Given a rectangle R in the coordinate system of a recursive problem,
return the top-level coordinates of all points that lie within R.
The following lemma assumes and builds upon the four "global" data structures
mentioned after the statement to Theorem 1.
Lemma 4. The slab-rank problem can be solved in O(lg N/ lg lg N ) time, and
the slab-select problem in O(lg N/ lg lg N ) time as well, provided that R contains
at most O(cid:0)√
lg N(cid:1) points.
Proof. We first consider the slab-rank problem. Without loss of generality, as-
sume that the given (i, j) is in a horizontal n × k slab. To translate j to j(cid:48), we
only need to subtract the appropriate multiple of k in O(1) time. To map i to i(cid:48),
we need to count the number of points in the slab with x-coordinate smaller than
i (see Fig. 3(L)). Since the top-level coordinates of the point (i, j) are known,
and the top-level coordinates of the slab are also stored in the "header" of the
slab by assumption, top-level coordinates of all sides of the query are known. As
all input points are stored in a global instance of the data structure of Lemma 1,
counting the number of points can be performed by an orthogonal range counting
query in O(lg N/ lg lg N ) time.
Fig. 3. Slab-rank (left) and slab-select (right).
The slab-select problem is solved in a similar manner, and is most easily
explained with reference to Fig. 3(R). We are given a recursive sub-problem
P and a rectangle R within P , and we know P 's bounding box in the top-
level problem (as shown on the far right). Our aim is to retrieve the top-level
coordinates of the image of R in the top-level problem. Suppose that the local
RRPPA x and y coordinates of R are xl, xr, yb and yt. Then we perform a orthogonal
range count (Lemma 1) in the area A, which lies under P but within P 's x-
coordinates. This takes O(lg N/ lg lg N ) time, and let z be the value returned. We
then select the z +yb and z +yt-th smallest y-coordinates within X[xl], . . . , X[xr]
by Lemma 3, also taking O(lg N/ lg lg N ) time. The (top-level) y-coordinates of
the points returned (shown shaded in the middle) are R's boundaries in the
top-level coordinate system. A similar query on Y [yb], . . . , Y [yt] gives the other
lg N(cid:1) points in this
boundaries of R in the top-level coordinate system. The O(cid:0)√
rectangle are then retrieved in O(cid:0)(lg N )5/6(cid:1) = o(lg N/ lg lg N ) time by Lemma 2.(cid:117)(cid:116)
Remark 1. In all applications of the slab-rank result, the top-level coordinates
of (i, j) will be known, as (i, j) will either be a vertex of the original input query,
or is the result of intersecting a horizontal (or vertical) line through the vertex
of a recursive problem (whose top-level coordinates are inductively known) with
a vertical (or horizontal) line defining a recursive sub-problem.
3.3 Encoding 2-sided queries
In this section we show Lemma 5. Although the reduction of 2-sided range max-
ima queries at point q (RM Q(q) hereafter) to orthogonal planar point location
is not new, the observation about the amount of priority information needed to
answer RM Q is new (and essential). This lemma shows that although storing
the permutation π itself requires Θ(n lg n) bits, the "effective entropy" [14] of
the permutation with respect to 2-sided range maximum queries is much lower,
generalizing the equivalent statement regarding 1D range-maximum queries [12,
21].
Lemma 5. Given a set S of n points from R2 and relative priorities given as a
permutation π on [n], the query RM Q(q) can be reduced to point location of q in
a collection of at most n horizontal semi-open line segments, whose left endpoints
are points from S, and whose right endpoints have x-coordinate equal to the x-
coordinate of some point from S. Further, given at most 2(n − r) + lg(cid:0)n
bits of extra information, the collection of line segments can be reconstructed
from S, where r is the number of redundant points -- those that are never the
answer to any query -- in S, and this bound is tight.
(cid:1) ≤ 3n
r
Proof. Assume the points are in general position and that the 2-sided query is
open to the top and left. Associate each point p = (x(p), y(p)) ∈ S with a hor-
izontal semi-open line of influence, possibly of length zero, whose left endpoint
(included in the line) is p itself, and is denoted by Inf (p), and contains all points
q such that y(q) = y(p), x(q) ≥ x(p) and RM Q(q) = p. It can be seen that (see
e.g. [17]) the answer to RM Q(q) for any q ∈ R2 can be obtained by shooting a
vertical ray upward from q until the first line Inf (p) is encountered; the answer
to RM Q(q) is then p (if no line is encountered then there is no point in the
2-sided region specified by q). See Fig. 4 for an example.
Fig. 4. (Left) Example for Lemma 5. The horizontal lines are the lines of influence.
Vertical dotted lines show where a point has terminated the line of influence of another
point. The arrow shows how point location in the lines of influence answers the 2-sided
query with lower right hand corner at q, returning the point with priority 7. (Right)
Example showing tightness of space bound -- points along the diagonal are A(cid:96) and the
queries at p and q illustrate how 1D RMQ queries can be answered and redundancy of
elements tested, respecitvely.
The set Inf (S) = {Inf (p)p ∈ S} can be computed by sweeping a vertical line
from left to right. At any given position x = t of the sweep line, the sweep line will
intersect Inf (S(cid:48)) for some set S(cid:48) (initially S(cid:48) = ∅). If S(cid:48) = pi1 , . . . , pir such that
y(pi1 ) > . . . > y(pir ) then it follows that π(i1) < . . . < π(ir) (the current lines of
influence taken from top to bottom represent points with increasing priorities).
Upon reaching the next point ps such that y(pij ) < y(ps) < y(pij+1), either (i)
π(s) < π(ij) -- in this case Inf (ps) is empty -- or (ii) π(k) > π(ij). In the latter
case, it may be that π(k) > π(ij+1), . . . π(ij+k) for some k ≥ 0, which would
mean that Inf (pij+1), . . . , Inf (pij+k ) are terminated, with their right endpoints
being x(ps). To construct Inf (S), therefore, only O(n) bits of information are
needed: for each point, one bit is needed to indicate whether case (i) or (ii) holds,
and in the latter case, the value of k needs to be stored. However, k can be stored
in unary using k + 1 bits, and the total value of k, over the course of the entire
sweep, is at most (n − r), giving a total of at most 2(n − r) bits. The bit-string
that indicates whether case (i) or (ii) holds can be stored in
(cid:1)(cid:109) ≤ n bits.
(cid:108)
lg(cid:0) n
n−r
To show that this is tight, we consider the 1D range maximum with redundant
entries problem, defined as follows. Given an array A of size n, of which r
entries are redundant, answer the following two queries: firstly, given an index
i, state whether A[i] is redundant, and secondly, given indices 1 ≤ i ≤ j ≤ n,
return the index of the largest value in a query interval A[i], . . . , A[j] (ignoring
redundant values). Given n, r, it is easy to see that 2(n − r) + lg(cid:0)n
these queries without accessing A. This is because there are(cid:0)n
(cid:1) − O(lg n)
(cid:1) choices for the
bits are required to encode the answers to the above queries, namely, to answer
r
r
846132759q qp (cid:0)2(n−r)
(cid:1) partial orders among the
n−r
n−r+1
positions of the redundant elements, and for each choice of the positions of the
redundant elements there are Cn−r = 1
values in A that can be distinguished by 1D range maximum queries [12, 21]. We
associate the values in A with a set of n points A(cid:96) placed on a slanted line (in
increasing x-coordinate order), and give each point in A(cid:96) a priority equal to the
corresponding entry in A. The point associated with each redundant value in A
is given a priority of −∞. In addition we place an n + 1st point z that dominates
all of the points in A(cid:96) (and hence is included in any query that also includes a
point from A(cid:96)), whose priority is greater than −∞ but less than the smallest
priority associated with a non-redundant entry of A (see Fig. 4(R)). The 1D
range maximum query problem with redundant entries can can be solved via
2-sided 2D range maximum queries that include only z and the appropriate sub-
range of A(cid:96). Also, we can determine if an entry in A is redundant by making
a 2-sided query that includes only the corresponding point from A(cid:96) and z; the
(cid:117)(cid:116)
point is redundant iff the answer to such a query is z.
Remark 2. It can be shown that 2(n − r) + lg(cid:0)n
(cid:1) ≤ n lg 5 + o(n), which is at
r
most 2.33n bits for large n.
3.4 Data structures for 2-sided queries
In this section we show the following lemma:
Lemma 6. Given a recursive sub-problem of size n, we can answer 2-sided
queries on this problem in O(lg N ) time using O(n
lg n) bits of space.
√
√
This lemma assumes and builds upon the four "global" data structures metioned
after the statement to Theorem 1. We view the given problem of size n, on which
we need to support 2-sided queries, as point location in a collection of O(n)
horizontal line segments as in Lemma 5, but with a limited space budget of
O(n
lg n) bits. We therefore need to devise an implicit representation of these
problems. We begin with an overview of the process. Let T be the set of n points
in the sub-problem we are considering. We start with an explicit representation
of Inf (T ) and choose a parameter λ = Θ(cid:0)√
called the skeleton, requires O((n/λ) lg n) = O(cid:0)n
lg n(cid:1). We select O(n/λ) lines of
lg n(cid:1) bits. Furthermore, we
influence that partition the plane into rectangular regions with O(λ) points (from
T ) and parts of line segments (from Inf (T )) [3, 10] and store a standard point
location data structure on the selected lines of influence. This data structure,
√
store O(λ) bits of information with each region (including the O(λ)-bit encoding
of priority information from Lemma 5).
The query proceeds as follows. Given a query point q, we first perform a point
location query on the skeleton to determine the region R in which q lies. We now
need to reconstruct the original point location structure within R, and perform a
slab-select to determine the points of T that lie within this region. This, together
with the priority information, allows us to partially -- but not fully, since lines of
influence may originate from outside R -- reconstruct the point location structure
within R. To handle lines of influence starting outside R, we do a binary search
with O(lg λ) steps, where in each step we need to perform a slab-select, giving
the claimed bound. The details are as follows.
√
lg n and
Preprocessing. As noted above, we first create Inf (T ), take λ =
select a set of points T (cid:48) ⊆ T with the following properties: (a) T (cid:48) = O(n/λ); (b)
the vertical decomposition, whereby we shoot vertical rays upward and downward
from each endpoint of each segment in Inf (T (cid:48)) until they hit another segment
(see Fig. 5), of the plane induced by Inf (T (cid:48))6 decomposes the plane into O(n/λ)
rectangular regions each of which has at most O(λ) points from T and parts
of line segments from Inf (T ) in it. T (cid:48) always exists and can be found by plane
sweep [3, Section 3],[10, Section 4.3]. The skeleton is any standard point location
data structure on Inf (T (cid:48)).
Let R be any region, and let Left(R) (Right(R)) be the set of line segments
from Inf (T ) that intersect the left (right) boundaries of R, and let P (R) be the
set of points from T in R. We store the following bit strings for R:
1. For each line segment (cid:96) ∈ Left(R), ordered top-to-down by y-axis, a bit
that indicates whether the right endpoint of (cid:96) is in R or not; similarly for
(cid:96) ∈ Right(R), a bit indicating whether (cid:96) begins in R or not.
2. If the left boundary of R is adjacent to other regions R1, R2, . . . (taken
from top to bottom) and li ≥ 0 represents the number of line segments
from Left(R) that also intersect Ri, then we store a bit-string 0l1 10l21 . . .. A
similar bit-string is stored for the right boundary of R.
3. For each point in P (R) and each line segment in Left(R), a bit-string of
length P (R) + Left(R) whose i-th bit indicates whether the i-th largest
y-coordinate in P (R) ∪ Left(R) is from P (R) or L(R).
The purpose of (1) and (2) is to trace a line segment as it crosses multiple regions:
if a line segment crosses from a region R(cid:48) to a region R(cid:48)(cid:48) on its right, then given
its position in Right(R(cid:48)), we can deduce its position in Left(R(cid:48)(cid:48)) and vice-versa.
However, there is no useful bound on the number of regions a single line segment
may cross, so we store the following information:
4. Suppose that a line segment (cid:96) = Inf (p) for some p ∈ T crosses m ≥ λ
regions. Then, in every λth region that (cid:96) crosses, we explicitly store the
region containing p, and p's local coordinates. As in (1), for each region R,
we store one bit for each (cid:96) ∈ Right(R), (cid:96) = Inf (p) , indicating whether or
not R holds information about p.
5. Finally, for each point p ∈ P (R), we store the sequence of bits from Lemma 5,
which indicates whether p has a non-empty Inf (p) and if so, for how many
lines from Left(R) ∪ Inf (P (R)), p is a right endpoint (p cannot be a right
endpoint of any other line in Inf (T ), by the construction of the skeleton).
6 Note that the extent of a line segment in Inf (T (cid:48)) is defined, as originally, wrt points
in T , and not wrt the points in T (cid:48).
√
As noted previously, the skeleton takes O(n
lg n) bits, within our budget.
We now add up the space required for (1)-(5). By construction, the sum of
Left(R), Right(R) and P (R) is O(n) summed over all regions R. The space
bound for (1) and (3) is therefore O(n) bits. The number of 1s in the bit string
of (3), summed over all regions, is O(n/λ), as there are O(n/λ) regions and the
graph which indicates adjacency of regions is planar; the number of 0s is O(n)
lg n)
as before. The space used by (4) is O(n
portions of line segments in the regions we store O(lg n) bits. Finally, the space
used for (5) is O(n) bits by Lemma 5.
√
lg n) bits again, as for every O(
√
Query algorithm. Suppose that we are given a query point q in a sub-problem
of size n and need to answer RM Q(q) (assume that we have q's local and top-
level coordinates). The query algorithm proceeds as follows:
(a) Do a planar point location in the skeleton, and find a region R in which the
point q lies. Perform slab-select on R to get P (R).
(b) As we know how many segments from Left(R) lie vertically between any pair
of points in P (R), when we are given the data in (5) above, we are able to
determine whether the x-coordinate of a given point p in P (R) is the right
endpoint of a line from either Left(R) or Inf (P (R)). Thus, we have enough
information to determine Inf (p) for all p ∈ P (R) (at least until the right
boundary of R). Furthermore, for each line in Left(R) that terminates in R,
we also know (the top-level coordinates of) its right endpoint.
(c) Using the top-level coordinates of q, we determine the nearest segment from
Inf (P (R)) that is above q.
∗
∗
∗
∗
∗
∗
(d) Using the top-level coordinates of q we also find the set of segments from
(R).
Left(R) whose right endpoints are not to the left of q. Let this set be Left
(R) that is above q. Un-
We now determine the nearest segment from Left
(R) = O(λ), since the segments in Left
fortunately, although Left
(R)
originate in points outside R, we do not have their y-coordinates. Hence, we
need to perform the following binary search on Left
(d1) Take the line segment (cid:96) ∈ Left
(R) with median y-coordinate, and sup-
pose that (cid:96) = Inf (p). The first task is to find the region Rp containing
p, as follows. Use (2) to determine which of the adjacent regions of R (cid:96)
intersects, say this is R(cid:48). If (cid:96) ends in R(cid:48), or R(cid:48) = Rp and we are done.
Otherwise, use (1) to locate (cid:96) in Left(R(cid:48)) and continue.
(d2) Once we have found Rp, we perform a slab-select on R(cid:48) to determine
P (Rp), and sort P (Rp) by y-axis. Then we perform (c) above on P (Rp),
thus determining which points of P (Rp) have lines of influence that reach
the right boundary of Rp. Using this we can now determine the (top-
level) coordinates of p.
(R):
(d3) We compare the top-level y-coordinates of p and q and recurse.
(e) We take the lower of the lines found in (d) and (e) and use it to return a
candidate. Observe that we have the top-level coordinates of this candidate.
We now derive the time complexity of a 2-sided query. Step (a) takes O(lg n)
for the point location, and O(lg N/ lg lg N ) for the slab-select. Step (b) can
√
be done in O(lg n) = O(lg N ) time by running the plane sweep algorithm of
Lemma 5 (recall that P (R) = O(
lg n) -- a quadratic algorithm will suffice).
√
Step (c) likewise can be done by a simple plane sweep in O(lg n) time. Step (d1)
lg n) times before Rp is found since every λ-th region
is iterated at most O(
intersected by (cid:96) contains information about p. Each iteration of (d1) takes O(1)
time: operations on the bit-strings are done either by table lookup if the bit-
string is short (O(λ) bits), or else using rank and select operations [20], if the
bit string is long (as e.g. the bit-string in (2) may be) -- these entirely standard
tricks are not described in detail. Step (d2) takes O(lg N/ lg lg N ) time as before.
Steps (d1)-(d3) are performed O(lg λ) = O(lg lg N ) times, so this takes O(lg N )
time overall. Step (e) is trivial. We have thus shown Lemma 6.
3.5 Putting things together
As noted in Section 3.1, the space usage of our data structure is O(N ) words.
Coming to the running time, we solve O(lg lg N ) 2-sided queries using Lemma 6,
giving a time of O(lg N lg lg N ). The O(lg lg N ) square-aligned queries are solved
in O(1) time each. The O(1) terminal problems at the bottom of the recursion
are solved using Chazelle's algorithm in O((lg lg N )2) time. Any candidate given
in local coordinates is converted to top-level coordinates in O(lg N/ lg lg N ) time,
or O(lg N ) time overall. We simply sequentially scan all O(lg lg N ) candidates
to find the answer. This proves Theorem 1.
4 A succinct index for 2-sided queries
In this section, we give a succinct index for 2-sided range maxima queries over
N points in rank space. This is in essence a stand-alone variant of Lemma 6
and reuses its structure. As noted earlier, we consider the case where the point
coordinates are stored "elsewhere" and are assumed to be accessible in one of
two ways, repeated here for convenience:
-- Through an orthogonal range reporting query. Here, we assume that a query
that results in k points being reported takes T (N, k) time. We assume that
T is such that T (N, O(k)) = O(T (N, k)).
-- The permuted-point model of [4], where we assume that the point coor-
dinates are stored in read-only memory, permitting random access to the
coordinates of the i-th point. However, the ordering of the points is specified
by the data structure.
Note that the priority information, unlike the coordinates, is encoded within the
index.
4.1 Succinct index in the orthogonal range reporting model
Lemma 7. Let λ ≥ 2 be some parameter. There is a succinct index of size
O(N +(N lg N )/λ)) bits such that RM Q(q) queries can be answered in O(lg N +
lg λ(λ + T (N, λ)) time.
Proof. The proof follows closely the proof of Lemma 6, except that the distinc-
tion between local and top-level coordinates vanishes, and the slab-select oper-
ation is replaced by the assumed orthogonal range reporting query. The space
complexity of the skeleton and associated bit-strings is exactly as in Lemma 6.
For the time complexity, the planar point location to find the region R contain-
ing the query point q is O(lg N ) time. Finding P (R) takes T (N, λ) time, and
each of the O(lg λ) iterations of the binary search takes O(λ + T (N, λ)) time. (cid:117)(cid:116)
Choosing λ = lg N in the above, we get:
Corollary 1. There is a succinct index of O(N ) bits such that RM Q(q) queries
can be answered in O(lg lg N · (lg N + T (N, lg N ))) time.
4.2 Succinct index in the permuted-point model
Lemma 8. There is a succinct index of N lg 5 + o(N ) = 2.33N + o(N ) bits
for answering RM Q(q) queries on a set of N points in O(lg lg N ) time in the
permuted-point model.
Proof. (sketch) Let S be the set of input points. We again solve RM Q(q) queries
by planar point location on Inf (S). We consider the regions of the vertical de-
composition induced by Inf (S) as nodes in a planar graph (the dual graph of
the set of regions), with edges between adjacent cells (see Fig 5). Using the pla-
Fig. 5. Vertical decomposition of the plane induced by a collection of line segments,
and a part of the dual graph.
nar separator theorem, for any parameter λ ≥ 1, there is a collection of O(n/λ)
cells such that the removal of this collection of cells this graph can be decom-
posed into connected components of size O(λ2) each (this approach is used by
Bose et al. [4] and Chan and Patra¸scu [10] for example). As in [4] we use a
two-level decomposition, first decomposing the vertical decomposition of Inf (S)
using λ = (lg N )2, and then decomposing the connected components themselves
using λ(cid:48) = (lg lg N )2. The key difference to Lemma 6 is that the boundaries of
the cells can be relatively compactly described. For instance, for each separator
cell in the top-level decomposition, we can specify the points in S that define its
four sides using O(lg N ) bits, and still use only o(N ) bits. In the second-level
decomposition, this information is stored in O(lg lg N ) bits per separator cell
(again o(N ) bits overall), as the relevant point will either belong to the same
top-level connected component of size O((lg N )4), or else the relevant informa-
tion will be stored in one of the top-level separator cells that form its boundary.
The leading-order term comes from storing the N lg 5 + o(N ) bits of priority
information needed to answer queries within the connected components. As in
[4], we also permute the points in a manner aligned with the decomposition,
which allows us to reconstruct the appropriate part of the planar point location
rapidly. The planar point location in the top level is performed using Chan's
data structure [8] taking O(lg lg N ) time. Note that a third level of decomposi-
tion, this time into connected components of size O((lg lg lg N )3) is needed to
achieve "decompression" of the relevant parts of the point location structure in
(cid:117)(cid:116)
O(lg lg N ) time.
5 Conclusions
We have introduced a new approach to producing space-efficient data structures
for orthogonal range queries. The main idea has been to partition the problem
into smaller sub-problems, which are stored in a "compressed" format that are
then "decompressed" on demand. We applied this idea to give the first linear-
space data structure for 2D range maxima that improves upon Chazelle's 1985
linear-space data structure.
References
1. P. Afshani, L. Arge, and K. D. Larsen. Orthogonal range reporting: query lower
bounds, optimal structures in 3-D, and higher-dimensional improvements.
In
J. Snoeyink, M. de Berg, J. S. B. Mitchell, G. Rote, and M. Teillaud, editors,
Symposium on Computational Geometry, pages 240 -- 246. ACM, 2010.
2. J. Barbay, M. He, J. I. Munro, and S. S. Rao. Succinct indexes for strings, binary
relations and multi-labeled trees. In N. Bansal, K. Pruhs, and C. Stein, editors,
SODA, pages 680 -- 689. SIAM, 2007.
3. M. A. Bender, R. Cole, and R. Raman. Exponential structures for efficient cache-
oblivious algorithms.
In P. Widmayer, F. T. Ruiz, R. M. Bueno, M. Hennessy,
S. Eidenbenz, and R. Conejo, editors, ICALP, volume 2380 of Lecture Notes in
Computer Science, pages 195 -- 207. Springer, 2002.
4. P. Bose, E. Y. Chen, M. He, A. Maheshwari, and P. Morin. Succinct geometric
In C. Mathieu, editor, SODA, pages
indexes supporting point location queries.
635 -- 644. SIAM, 2009.
5. P. Bose, M. He, A. Maheshwari, and P. Morin. Succinct orthogonal range search
structures on a grid with applications to text indexing.
In F. K. H. A. Dehne,
M. L. Gavrilova, J.-R. Sack, and C. D. T´oth, editors, WADS, volume 5664 of
Lecture Notes in Computer Science, pages 98 -- 109. Springer, 2009.
6. G. S. Brodal, P. Davoodi, and S. S. Rao. On space efficient two dimensional range
minimum data structures. In M. de Berg and U. Meyer, editors, ESA (2), volume
6347 of Lecture Notes in Computer Science, pages 171 -- 182. Springer, 2010.
7. G. S. Brodal and A. G. Jørgensen. Data structures for range median queries. In
Y. Dong, D.-Z. Du, and O. H. Ibarra, editors, ISAAC, volume 5878 of Lecture
Notes in Computer Science, pages 822 -- 831. Springer, 2009.
8. T. M. Chan. Persistent predecessor search and orthogonal point location on the
word ram. In D. Randall, editor, SODA, pages 1131 -- 1145. SIAM, 2011.
9. T. M. Chan, K. G. Larsen, and M. Patra¸scu. Orthogonal range searching on the
ram, revisited. In Proceedings of the 27th annual ACM symposium on Computa-
tional geometry, SoCG '11, pages 1 -- 10, New York, NY, USA, 2011. ACM.
10. T. M. Chan and M. Patrascu. Transdichotomous results in computational geom-
etry, I: Point location in sublogarithmic time. SIAM J. Comput., 39(2):703 -- 729,
2009.
11. B. Chazelle. A functional approach to data structures and its use in multidimen-
sional searching. SIAM J. Comput., 17(3):427 -- 462, 1988. Prel. vers. FOCS 85.
12. J. Fischer and V. Heun. Space-efficient preprocessing schemes for range minimum
queries on static arrays. SIAM J. Comput., 40(2):465 -- 492, 2011.
13. H. N. Gabow, J. L. Bentley, and R. E. Tarjan. Scaling and related techniques
In Proc. 16th Annual ACM Symposium on Theory of
for geometry problems.
Computing, pages 135 -- 143. ACM, 1984.
14. M. J. Golin, J. Iacono, D. Krizanc, R. Raman, and S. S. Rao. Encoding 2d range
maximum queries.
In T. Asano, S.-I. Nakano, Y. Okamoto, and O. Watanabe,
editors, ISAAC, volume 7074 of Lecture Notes in Computer Science, pages 180 --
189. Springer, 2011.
15. J. J´aJ´a, C. W. Mortensen, and Q. Shi. Space-efficient and fast algorithms for
multidimensional dominance reporting and counting. In R. Fleischer and G. Trip-
pen, editors, ISAAC, volume 3341 of Lecture Notes in Computer Science, pages
558 -- 568. Springer, 2004.
16. M. Karpinski and Y. Nekrich. Space efficient multi-dimensional range reporting. In
H. Q. Ngo, editor, COCOON, volume 5609 of Lecture Notes in Computer Science,
pages 215 -- 224. Springer, 2009.
17. C. Makris and A. K. Tsakalidis. Algorithms for three-dimensional dominance
searching in linear space. Inf. Process. Lett., 66(6):277 -- 283, 1998.
18. D. P. Mehta and S. Sahni, editors. Handbook of Data Structures and Applications.
Chapman & Hall/CRC, 2009.
19. Y. Nekrich. Orthogonal range searching in linear and almost-linear space. Comput.
Geom., 42(4):342 -- 351, 2009.
20. N. Rahman and R. Raman. Rank and select operations on binary strings. In M.-Y.
Kao, editor, Encyclopedia of Algorithms. Springer, 2008.
21. J. Vuillemin. A unifying look at data structures. Communications of the ACM,
23(4):229 -- 239, 1980.
22. H. Yuan and M. J. Atallah. Data structures for range minimum queries in multi-
dimensional arrays. In M. Charikar, editor, SODA, pages 150 -- 160. SIAM, 2010.
|
1908.00491 | 1 | 1908 | 2019-08-01T16:36:43 | On Cycle Transversals and Their Connected Variants in the Absence of a Small Linear Forest | [
"cs.DS",
"cs.CC",
"cs.DM",
"math.CO"
] | A graph is $H$-free if it contains no induced subgraph isomorphic to $H$. We prove new complexity results for the two classical cycle transversal problems Feedback Vertex Set and Odd Cycle Transversal by showing that they can be solved in polynomial time on $(sP_1+P_3)$-free graphs for every integer $s\geq 1$. We show the same result for the variants Connected Feedback Vertex Set and Connected Odd Cycle Transversal. We also prove that the latter two problems are polynomial-time solvable on cographs; this was already known for Feedback Vertex Set and Odd Cycle Transversal. We complement these results by proving that Odd Cycle Transversal and Connected Odd Cycle Transversal are NP-complete on $(P_2+P_5,P_6)$-free graphs. | cs.DS | cs |
On Cycle Transversals and Their Connected Variants
in the Absence of a Small Linear Forest⋆
Konrad K. Dabrowski1, Carl Feghali2, Matthew Johnson1,
Giacomo Paesani1, Daniël Paulusma1, and Paweł Rzążewski3
1 Department of Computer Science, Durham University, UK
{konrad.dabrowski,matthew.johnson2,giacomo.paesani,daniel.paulusma}@durham.ac.uk
2 Department of Informatics, University of Bergen, Norway [email protected]
3 Faculty of Mathematics and Information Science, Warsaw University of Technology,
Warsaw, Poland [email protected]
Abstract. A graph is H-free if it contains no induced subgraph isomorphic
to H. We prove new complexity results for the two classical cycle transversal
problems Feedback Vertex Set and Odd Cycle Transversal by showing
that they can be solved in polynomial time on (sP1 + P3)-free graphs for every
integer s ≥ 1. We show the same result for the variants Connected Feedback
Vertex Set and Connected Odd Cycle Transversal. We also prove that
the latter two problems are polynomial-time solvable on cographs; this was
already known for Feedback Vertex Set and Odd Cycle Transversal.
We complement these results by proving that Odd Cycle Transversal and
Connected Odd Cycle Transversal are NP-complete on (P2 +P5, P6)-free
graphs.
1 Introduction
Graph transversal problems play a central role in Theoretical Computer Science. To
define the notion of a graph transversal, let H be a family of graphs, G = (V, E) be a
graph and S ⊆ V be a subset of vertices of G. The graph G − S is obtained from G by
removing all vertices of S and all edges incident to vertices in S. We say that S is an
H-transversal of G if G − S is H-free, that is, if G − S contains no induced subgraph
isomorphic to a graph of H. In other words, S intersects every induced copy of every
graph of H in G. Let Cr and Pr denote the cycle and path on r vertices, respectively.
Then S is a vertex cover, feedback vertex set, or odd cycle transversal if S is an H-
transversal for, respectively, H = {P2} (that is, G − S is edgeless), H = {C3, C4, . . .}
(that is, G − S is a forest), or H = {C3, C5, . . .} (that is, G − S is bipartite).
Usually the goal is to find a transversal of minimum size in some given graph. In
this paper we focus on the decision problems corresponding to the three transversals
defined above. These are the Vertex Cover, Feedback Vertex Set and Odd
Cycle Transversal problems, which are to decide whether a given graph has a
vertex cover, feedback vertex set or odd cycle transversal, respectively, of size at most k
for some given positive integer k. Each of these three problems is well studied and is
well known to be NP-complete.
We may add further constraints to a transversal. In particular, we may require
a transversal of a graph G to be connected, that is, to induce a connected subgraph
of G. The corresponding decision problems for the three above transversals are then
⋆ The research in this paper received support from the Leverhulme Trust (RPG-2016-
258). The last author was supported by Polish National Science Centre grant no.
2018/31/D/ST6/00062. An extended abstract of this paper appeared in the proceedings
of FCT 2019 [13].
called Connected Vertex Cover, Connected Feedback Vertex Set and Con-
nected Odd Cycle Transversal, respectively.
Garey and Johnson [15] proved that Connected Vertex Cover is NP-complete
even on planar graphs of maximum degree 4 (see, for example, [14,31,36] for NP-
completeness results for other graph classes). Grigoriev and Sitters [18] proved that
Connected Feedback Vertex Set is NP-complete even on planar graphs with
maximum degree 9. More recently, Chiarelli et al. [10] proved that Connected Odd
Cycle Transversal is NP-complete even on graphs of arbitrarily large girth and on
line graphs.
As all three decision problems and their connected variants are NP-complete, we
can consider how to restrict the input to some special graph class in order to achieve
tractability. Note that this approach is in line with the aforementioned results in the
literature, where NP-completeness was proven on special graph classes. It is also in
line with with, for instance, polynomial-time results for Connected Vertex Cover
by Escoffier, Gourvès and Monnot [12] (for chordal graphs) and Ueno, Kajitani and
Gotoh [35] (for graphs of maximum degree at most 3 and trees).
Just as in most of these papers, we consider hereditary graph classes, that is, graph
classes closed under vertex deletion. Hereditary graph classes form a rich framework
that captures many well-studied graph classes. It is not difficult to see that every
hereditary graph class G can be characterized by a (possibly infinite) set FG of for-
bidden induced subgraphs. If FG = 1, say F = {H}, then G is said to be monogenic,
and every graph G ∈ G is said to be H-free. Considering monogenic graph classes can
be seen as a natural first step for increasing our knowledge of the complexity of an
NP-complete problem in a systematic way. Hence, we consider the following research
question:
How does the structure of a graph H influence the computational complexity of a graph
transversal problem for input graphs that are H-free?
Note that different graph transversal problems may behave differently on some class of
H-free graphs. However, the general strategy for obtaining complexity results is to first
try to prove that the restriction to H-free graphs is NP-complete whenever H contains
a cycle or the claw (the 4-vertex star). This is usually done by showing, respectively,
that the problem is NP-complete on graphs of arbitrarily large girth (length of a
shortest cycle) and on line graphs, which form a subclass of claw-free graphs. If this
is the case, then we are left to consider the case when H does not contain a cycle,
implying that H is a forest, and does not contain a claw either, implying that H is a
linear forest, that is, the disjoint union of one or more paths.
1.1 The Graph H Contains a Cycle or Claw
It follows from Poljak's construction [30] that Vertex Cover is NP-complete on
graphs of arbitrarily large girth. Hence, Vertex Cover is NP-complete on H-free
graphs if H contains a cycle. However, Vertex Cover becomes polynomial-time solv-
able when restricted to claw-free graphs [25,32]. In contrast, the other five problems
Connected Vertex Cover, (Connected) Feedback Vertex Set and (Con-
nected) Odd Cycle Transversal are all NP-complete on graphs of arbitrarily
large girth and on line graphs; see Table 1. Hence, for these five problems, it remains
to consider only the case when H is a linear forest.
1.2 The Graph H Is a Linear Forest
In this paper, we focus on proving new complexity results for Feedback Vertex Set,
Connected Feedback Vertex Set, Odd Cycle Transversal and Connected
2
Odd Cycle Transversal on H-free graphs. It follows from Section 1.1 that we
may assume that H is a linear forest. Below we first discuss the known polynomial-
time solvable cases. As we will use algorithms for Vertex Cover and Connected
Vertex Cover as subroutines for our new algorithms, we include these two problems
in our discussion.
For every s ≥ 1, Vertex Cover (by combining the results of [1,34]) and Con-
nected Vertex Cover [10] are polynomial-time solvable on sP2-free graphs.4 More-
over, Vertex Cover is also polynomial-time solvable on (sP1 + P6)-free graphs, for
every s ≥ 0 [20], as is the case for Connected Vertex Cover on (sP1 + P5)-free
graphs [24]. Their complexity on Pr-free graphs is unknown for r ≥ 7 and r ≥ 6,
respectively.
Both Feedback Vertex Set and Odd Cycle Transversal are polynomial-
time solvable on permutation graphs [4], and thus on P4-free graphs. Recently, Okrasa
and Rzążewski [29] proved that Odd Cycle Transversal is NP-complete on P13-
free graphs. A small modification of their construction yields the same result for Con-
nected Odd Cycle Transversal. The complexity of Feedback Vertex Set
and Connected Feedback Vertex Set is unknown when restricted to Pr-free
graphs for r ≥ 5. For every s ≥ 1, both problems and their connected variants are
polynomial-time solvable on sP2-free graphs [10], using the price of connectivity for
feedback vertex set [2,21].5
1.3 Our Results
In Section 3 we prove that Connected Feedback Vertex Set and Connected
Odd Cycle Transversal are polynomial-time solvable on P4-free graphs, just as is
the case for Feedback Vertex Set and Odd Cycle Transversal. In Section 4
we prove that for every s ≥ 1, these four problems are all polynomial-time solvable
on (sP1 + P3)-free graphs; see also Table 1. Finally, in Section 5, we show that Odd
Cycle Transversal and Connected Odd Cycle Transversal are NP-complete
on (P2 + P5, P6)-free graphs, that is, graphs that are both (P2 + P5)-free and P6-free.
To prove our polynomial-time results, we rely on two proof ingredients. The first one is
that we use known algorithms for Vertex Cover and Connected Vertex Cover
restricted to H-free graphs as subroutines in our new algorithms. The second is that
we consider the connected variant of the transversal problems in a more general form.
For Connected Vertex Cover this variant is defined as follows:
Connected Vertex Cover Extension
Instance: a graph G = (V, E), a subset W ⊆ V and a positive integer k.
Question: does G have a connected vertex cover SW with W ⊆ SW and
SW ≤ k?
Note that Connected Vertex Cover Extension becomes the original problem if
W = ∅. We define the problems Connected Feedback Vertex Set Extension
and Connected Odd Cycle Transversal Extension analogously. We will prove
all our results for connected feedback vertex sets and connected odd cycle transversals
for the extension versions. These extension versions will serve as auxiliary problems
for some of our inductive arguments, but this approach also leads to slightly stronger
results.
4 The graph G + H is the disjoint union of graphs G and H and sG is the disjoint union of s
copies of G; see Section 2.
5 The price of connectivity concept was introduced by Cardinal and Levy [9] for vertex cover;
see also, for example, [6,7,8].
3
Vertex Cover
Feedback Vertex Set
Odd Cycle Transversal
girth p
NP-c [30] P
NP-c [30] NP-c [33]
NP-c [10] NP-c [10]
NP-c [28] NP-c [28]
NP-c [10] NP-c [10]
Con. Feedback Vertex Set
Con. Odd Cycle Transversal NP-c [10] NP-c [10]
Con. Vertex Cover
line graphs sP2-free P4-free sP1 + Pr-free
[25,32] P [1,34] P
P: s ≥ 0, r = 6 [20]
P [10] P [4] P: s ≥ 0, r = 3∗
P [10] P [4] P: s ≥ 0, r = 3∗
P [10] P
P [10] P∗
P [10] P∗
P: s ≥ 0, r = 5 [24]
P: s ≥ 0, r = 3∗
P: s ≥ 0, r = 3∗
Table 1: The complexities of the three connected transversal problems together with
the original transversal problems on graphs of girth at least p for every (fixed) constant
p ≥ 3, on line graphs, and on H-free graphs for various linear forests H. In particular,
Feedback Vertex Set can be shown to be NP-complete on graphs of arbitrarily
large girth by using Poljak's construction (see [3,26]). We also note that Munro [28]
showed that Feedback Vertex Set is NP-complete even on line graphs of planar
cubic bipartite graphs. Unreferenced results directly follow from other results in the
table, and results marked with ∗ are new results proven in this paper. Our two other
new results, namely that Odd Cycle Transversal and Connected Odd Cycle
Transversal are NP-complete on (P2 + P5, P6)-free graphs, are not included in the
table.
Remark 1. For any connected extension variant of these problems on H-transversals,
we may assume that the input graph G is connected. If it is not, then either all but
at most one connected component of G is H-free and does not intersect W , in which
case it need not be considered, or the answer is immediately no. It is easy to check
H-freeness for the three problems we consider.
Remark 2. Note that one could also define extension versions for any original
transversal problem (that is, where there is no requirement for the transversal to be
connected). However, such extension versions will be polynomially equivalent. Indeed,
we can solve the extension version on the input (G, W, k) by considering the original
problem on the input (G − W, max{0, k − W }) and adding W to the solution. How-
ever, due to the connectivity condition, we cannot use this approach for the connected
variants.
Remark 3. It is known that Vertex Cover is polynomial-time solvable on (P1+H)-
free graphs whenever this is the case on H-free graphs. This follows from a well-known
observation, see, for example, [27]: one can solve the complementary problem of finding
a maximum independent set in a (P1 + H)-free graph by solving this problem on each
H-free graph obtained by removing a vertex and all of its neighbours. However, this
trick does not work for Connected Vertex Cover. Moreover, it does not work
for Feedback Vertex Set and Odd Cycle Transversal and their connected
variants either.
2 Preliminaries
Let G = (V, E) be a graph. For a set S ⊆ V , we write G[S] to denote the subgraph
of G induced by S. We say that S is connected if G[S] is connected. We write G − S
to denote the graph G[V \ S]. A subset D ⊆ V is a dominating set of G if every
vertex of V \ D is adjacent to at least one vertex of D. An edge uv of a graph G =
(V, E) is dominating if {u, v} is a dominating set. The complement of G is the graph
G = (V, {uv uv 6∈ E and u 6= v}). The neighbourhood of a vertex u ∈ V is the set
NG(u) = {v uv ∈ E} and for U ⊆ V , we let NG(U ) = Su∈U N (u) \ U . We omit
4
the subscript when there is no ambiguity. We denote the degree of a vertex u ∈ V by
deg(u) = NG(u).
Let G = (V, E) be a graph and let S ⊆ V . Then S is a clique if the vertices of S are
pairwise adjacent and an independent set if the vertices of S are pairwise non-adjacent.
A graph is complete if its vertex set is a clique. We let Kr denote the complete graph
on r vertices. Let T ⊆ V with S ∩ T = ∅. Then S is complete to T if every vertex
of S is adjacent to every vertex of T , and S is anti-complete to T if there are no edges
between S and T . In the first case, we also say that S is complete to G[T ] and in the
second case anti-complete to G[T ].
A graph is bipartite if its vertex set can be partitioned into at most two independent
sets. A bipartite graph is complete bipartite if its vertex set can be partitioned into
two independent sets X and Y such that X is complete to Y . If X or Y has size 1,
the complete bipartite graph is said to be a star. Note that every edge of a complete
bipartite graph is dominating.
Let G1 and G2 be two vertex-disjoint graphs. The union operation creates the
disjoint union G1 + G2 of G1 and G2, that is, the graph with vertex set V (G1)∪V (G2)
and edge set E(G1) ∪ E(G2). We denote the disjoint union of r copies of G1 by rG1.
The join operation adds an edge between every vertex of G1 and every vertex of G2.
A graph G is a cograph if G can be generated from K1 by a sequence of join and union
operations. A graph is a cograph if and only if it is P4-free (see, for example, [5]).
The following lemma is well known, but we include a short proof for completeness.
Lemma 1. Every connected P4-free graph on at least two vertices has a spanning
complete bipartite subgraph which can be found in polynomial time.
Proof. Let G be a connected P4-free graph on at least two vertices. Then G is the join
of two graphs G[X] and G[Y ]. Hence, G has a spanning complete bipartite subgraph
with partition classes X and Y . Note that this implies that G is disconnected. In
order to find a (not necessarily unique) spanning complete bipartite subgraph of G
with partition classes X and Y in polynomial time, we put the vertices of one connected
⊓⊔
component of G in X and all the other vertices of G in Y .
Grzesik et al. [20] gave a polynomial-time algorithm for finding a maximum indepen-
dent set of a P6-free graph in polynomial time. As the complement V (G) \ I of every
independent set I of a graph G is a vertex cover, their result implies that Vertex
Cover is polynomial-time solvable on P6-free graphs. Using the folklore trick men-
tioned in Remark 3 (see also, for example, [24,27]) their result can also be formulated
as follows.
Theorem 1 ([20]). For every s ≥ 0, Vertex Cover can be solved in polynomial
time on (sP1 + P6)-free graphs.
We recall also that Connected Vertex Cover is polynomial-time solvable on
(sP1 + P5)-free graphs [24]. We will need the extension version of this result. Its proof
is based on a straightforward adaption of the proof for Connected Vertex Cover
on (sP1 + P5)-free graphs [24].6
Theorem 2 ([24]). For every s ≥ 0, Connected Vertex Cover Extension can
be solved in polynomial time on (sP1 + P5)-free graphs.
6 See Appendix A, where we include a proof for reviewing purposes.
5
3 The Case H = P4
Recall that Brandstädt and Kratsch [4] proved that Feedback Vertex Set and
Odd Cycle Transversal can be solved in polynomial time on permutation graphs,
which form a superclass of the class of P4-free graphs. Hence, we obtain the following
proposition.
Proposition 1 ([4]). Feedback Vertex Set and Odd Cycle Transversal can
be solved in polynomial time on P4-free graphs.
In this section, we prove that the (extension versions of the) connected variants of
Feedback Vertex Set and Odd Cycle Transversal are also polynomial-time
solvable on P4-free graphs. We make use of Proposition 1 in the proofs.
Theorem 3. Connected Feedback Vertex Set Extension can be solved in
polynomial time on P4-free graphs.
Proof. Let G = (V, E) be a P4-free graph on n vertices and let W be a subset of V . By
Remark 1, we may assume that G is connected. By Lemma 1, in polynomial time we
can find a spanning complete bipartite subgraph G′ = (X, Y, E′), and we note that,
by definition, every edge in G′ is dominating. Below, in Step 1, in polynomial time we
compute a smallest connected feedback vertex set of G that contains W and intersects
both X and Y . In Step 2, in polynomial time we compute a smallest connected feedback
vertex set of G that contains W and that is a subset of either X or Y (if such a set
exists). Then the smallest set found is a smallest connected feedback vertex set of G
that contains W .
Step 1. Compute a smallest connected feedback vertex set S of G such that W ⊆ S,
S ∩ X 6= ∅ and S ∩ Y 6= ∅.
We perform Step 1 as follows. Consider two vertices u ∈ X and v ∈ Y . We shall describe
how to find a smallest connected feedback vertex set of G that contains W ∪{u, v}. We
find a smallest feedback vertex set S′ in G − (W ∪ {u, v}). As G − (W ∪ {u, v}) is P4-
free, this takes polynomial time by Proposition 1. Then S′ ∪ W ∪ {u, v} is a smallest
feedback vertex set of G that contains W ∪ {u, v} and is connected, since uv is a
dominating edge. By repeating this polynomial-time procedure for all O(n2) possible
choices of u and v, we will find S in polynomial time.
Step 2. Compute a smallest connected feedback vertex set S of G such that S ⊆ X or
S ⊆ Y .
For Step 2 we describe only the S ⊆ X case, as the S ⊆ Y case is symmetric. Thus
we may assume that W ⊆ X, otherwise no such set exists. Clearly, we may also
assume that G[Y ] contains no cycles. If G[Y ] contains an edge it follows that S = X,
otherwise G − S would contain a triangle. Suppose instead that Y is an independent
set. If Y = 1, then X \ S must be an independent set, otherwise G − S contains a
triangle. So S is a smallest connected vertex cover of G[X] that contains W . As G[X]
is P4-free, we can find such an S in polynomial time by Theorem 2. If Y ≥ 2, then
X \ S ≤ 1, as otherwise G − S contains a 4-cycle. Thus, we check, in polynomial time,
if there exists a vertex x ∈ X \ W , such that X \ {x} is connected. If so, S = X \ {x}.
⊓⊔
Theorem 4. Connected Odd Cycle Transversal Extension can be solved in
polynomial time on P4-free graphs.
Proof. We only provide an outline, as the proof follows that of Theorem 3. We perform
the same two steps. In Step 1, we need to find a smallest odd cycle transversal S′ in
6
G − (W ∪ {u, v}) and can again apply Proposition 1. In Step 2, we again note that
if G[Y ] contains an edge, then S = X. Suppose that Y is an independent set. Then
G − S contains no odd cycles if and only if X \ S is independent, so S is a smallest
connected vertex cover of G[X] that contains W . (That is, the Y = 1 case from the
proof of Theorem 3 can be used for all values of Y , as we are no longer concerned
⊓⊔
with whether G − S might contain cycles of even length.)
4 The Case H = sP1 + P3
In this section, we will prove that Feedback Vertex Set and Odd Cycle
Transversal and their connected variants can be solved in polynomial time on
(sP1 +P3)-free graphs. We need three structural results. First, let us define a function c
on the non-negative integers by c(s) := max{3, 2s − 1}. We will use this function c
throughout the remainder of this section, starting with the following lemma.
Lemma 2. Let s ≥ 0 be an integer. Let G be a bipartite (sP1 + P3)-free graph. If G
has a connected component on at least c(s) vertices, then there are at most s − 1 other
connected components of G and each of them is on at most two vertices.
Proof. First note that the s = 0 case of the lemma is trivially true, as every connected
component of a bipartite P3-free graph has at most two vertices.
Suppose, for contradiction, that G has a connected component C1 on at least c(s)
vertices and a connected component C2 on at least three vertices. As C1 is bipartite
and contains at least 2s − 1 vertices, C1 contains a independent set of s vertices that
induce sP1. As C2 is bipartite and contains at least three vertices, C2 has a vertex v
of degree at least 2, and so v and two of its neighbours induce a P3. Thus G is not
(sP1 + P3)-free, a contradiction.
Similarly, if G contains a connected component C1 on at least c(s) ≥ 3 vertices,
then this component contains an induced P3. Since G is (sP1 + P3)-free, G can contain
at most s − 1 connected components other than C1.
⊓⊔
The internal vertices and leaves of a tree are the vertices of degree at least 2 and
degree 1, respectively.
Lemma 3. Let s ≥ 0 be an integer. Let T be an (sP1 + P3)-free tree. Then T has at
most 4s internal vertices.
Proof. Let U be the set of internal vertices of T . Suppose that U ≥ 4s + 1 ≥ 1. We
will show that this leads to a contradiction. As a path with at least 4s + 1 internal
vertices contains an induced sP1 + P3, we may assume that T is not a path and so has
at least three leaves. Hence V (T ) ≥ 4s + 4.
Let X and Y be the two bipartition sets of T , and assume without loss of generality
that X ≥ 2s + 2. For Z ∈ {X, Y }, let LZ and UZ be the leaves and internal vertices
of T that belong to Z. If there is a vertex in Y of degree at least 2 that is anti-complete
to a set of s vertices of X, then T contains an induced sP1 + P3, a contradiction.
Therefore we may assume that every vertex of Y either has degree at least X − s + 1
7
or is in LY . Then
X + UY + LY − 1 = X + Y − 1
= V (T ) − 1
= E(T )
= X
v∈Y
≥ X
v∈UY
deg(v)
(X − s + 1) + LY
= (X − s + 1)UY + LY
= XUY − sUY + UY + LY .
Thus we have X − 1 ≥ XUY − sUY and we rearrange to see that
UY ≤
X − 1
X − s
= 1 +
s − 1
X − s
.
Since X ≥ 2s + 2, we have that UY < 2. First suppose UY = 0. Then UX ≤ 1 and
LX = 0, or UX = 0 and LX ≤ 1. Both cases contradict the assumption that X
has at least 2s + 2 vertices. Now suppose UY = 1. Then, by our assumption that
U ≥ 4s + 1, we have that UX ≥ 4s and so LY ≥ UX ≥ 4s. Now it is easy to find
⊓⊔
an induced sP1 + P3 (see Fig. 1), and this contradiction completes the proof.
LY
z
LX
UY
UX
y
x
≥ 4s
Fig. 1: The structure of the tree T in the proof of Lemma 3 in the case when UY = 1.
The set LX is an independent set of vertices that each are adjacent to the unique
vertex y ∈ UY . The set LY is partitioned into independent sets of vertices that have
the same neighbour in UX . The vertices y, x, z, together with s vertices of Ly not
adjacent to x, induced an sP1 + P3 in T (which leads to the desired contradiction in
the proof).
The bound of 4s in Lemma 3 is not tight but, as we shall see later, it suffices for
our purposes.
8
Lemma 4. Let s ≥ 0 be an integer. Let G be a connected (sP1 + P3)-free graph, and
let U be a set of vertices in G. Then there is a set of vertices R in G such that G[R∪U ]
is connected and R ≤ 2s2 − 2s + 3.
Proof. If G[U ] is connected, then let R = ∅. Otherwise, since G cannot now be a
complete graph, it contains an induced path P on three vertices in G. The number
of connected components of G[U ] that do not contain a vertex that is either in P or
adjacent to a vertex of P in G is at most s−1, otherwise G contains an induced sP1+P3.
Let R contain the vertices of P and the internal vertices of shortest paths in G from P
to each set of vertices that induces a connected component of G[U ]. As at most s − 1
of these shortest paths have more than zero internal vertices, and as each contains at
most 2s internal vertices (any longer path contains an induced sP1 + P3), it follows
that R ≤ 3 + 2s(s − 1) = 2s2 − 2s + 3. As G[R ∪ U ] is connected, the lemma is
⊓⊔
proved.
We now prove our four results. For the connected variants, we consider the more
general extension versions.
Theorem 5. For every s ≥ 0, Feedback Vertex Set can be solved in polynomial
time on (sP1 + P3)-free graphs.
Proof. Let s ≥ 0 be an integer, and let G = (V, E) be an (sP1 + P3)-free graph. We
must show how to find a smallest feedback vertex set of G. We will in fact show how
to find a largest induced forest of G, the complement of a smallest feedback vertex
set. The proof is by induction on s. If s = 0, then we can use Proposition 1. We now
assume that s ≥ 1 and that we have a polynomial-time algorithm for finding a largest
induced forest in ((s − 1)P1 + P3)-free graphs. Our algorithm performs the following
two steps in polynomial time. Together, these two steps cover all possibilities.
Step 1. Compute a largest induced forest F such that every connected component of F
has at least c(s) vertices.
By Lemma 2 we know that F will be connected, and so by Lemma 3 F will be a tree
with at most 4s internal vertices. We consider every possible choice U of a non-empty
set of at most 4s vertices. There are O(n4s) choices. If U induces a tree, we will find
a largest induced tree whose internal vertices all belong to U . This can be found by
adding to U the largest possible set of vertices that are independent and belong to
the set R of vertices in G − U that each have exactly one neighbour in U . That is, we
need a largest independent set in G[R] and, by Theorem 1, such a set can be found in
polynomial time.
Step 2. Compute a largest induced forest F such that F has a connected component
with at most c(s) − 1 vertices.
We consider every possible choice of a non-empty set T of at most c(s) − 1 vertices
and discard those that do not induce a tree. There are O(nc(s)−1) choices for T . Let
U = N (T ), and let G′ = G − (T ∪ U ). Then G′ is ((s − 1)P1 + P3)-free. Thus we can
find a largest induced forest F ′ of G′ in polynomial time and F ′ + G[T ] is a largest
⊓⊔
induced forest of G among those that have G[T ] as a connected component.
Theorem 6. For every s ≥ 0, Connected Feedback Vertex Set Extension
can be solved in polynomial time on (sP1 + P3)-free graphs.
Proof. There are similarities to the proof of Theorem 5, but more arguments are
needed. Let s ≥ 0 be an integer, let G = (V, E) be a connected (sP1 + P3)-free
graph and let W be a subset of V . We must show how to find a smallest connected
feedback vertex set of G that contains W in polynomial time. We show how to solve
9
the complementary problem in polynomial time: how to find a largest induced forest F
of G that does not include any vertex of W and V \ F is connected. We will say that
an induced forest F is good if it has these two properties.
Our algorithm performs the following three steps in polynomial time. Together,
these three steps cover all possibilities.
Step 1. Compute a largest good induced forest F such that there is a connected com-
ponent of F that has at least c(s) vertices.
By Lemma 2 we know that F has exactly one connected component on at least c(s)
and there are at most s− 1 other connected components of F , each on at most two ver-
tices. By Lemma 3, the connected component on at least c(s) vertices has at most 4s
internal vertices. We consider O(n4s+2(s−1)) choices of a non-empty set U of at most 4s
vertices that induces a tree and a set U ′ of at most 2(s − 1) vertices that induces a
disjoint union of vertices and edges such that U ∪U ′ does not intersect W , U is disjoint
from U ′ and no vertex of U has a neighbour in U ′. Let R be the set of vertices that
each have exactly one neighbour in U and no neighbour in U ′, but do not belong to W .
We then add to U ∪ U ′ the largest possible set L of vertices that are independent and
belong to the set R such that G − (L ∪ U ∪ U ′) is connected. This is achieved by taking
the complement of the smallest connected vertex cover of G − (U ∪ U ′) that contains
V \ (R ∪ U ∪ U ′). By Theorem 2, this can be done in polynomial time.
Step 2. Compute a largest good induced forest F such that F has at most s − 1
connected components and each connected component has at most c(s) − 1 vertices.
Since the number of vertices in F is bounded by the constant (s − 1)(c(s) − 1), we can
simply check all sets containing at most that many vertices to see if they induce such
a good forest.
Step 3. Compute a largest good induced forest F such that F has at least s connected
components and each connected component has at most c(s) − 1 vertices.
We consider O(ns(c(s)−1)) choices of a non-empty set L of at most s(c(s) − 1) vertices.
We reject L unless G[L] is a good induced forest on s connected components with
no connected component of more than c(s) − 1 vertices. Assuming our choice of L is
correct, the connected components of G[L] will become connected components of G[F ].
Let U = N (L) and note that no vertex of U is in F . If G − U is a good forest,
then we are done. Otherwise we consider every set R of at most 2s2 − 2s + 3 vertices
of G − (L ∪ U ∪ W ) such that G[R ∪ U ∪ W ] is connected; see also Fig. 2. We note
that if there is a largest induced forest F such that the connected components of G[L]
are also connected components of G[F ], then Lemma 4 applied to G − F implies that
such a set R exists.
Let S = R ∪ U ∪ W . If G − S is a forest, then we are done. Otherwise note that
G − (L ∪ S) is the disjoint union of one or more complete graphs: G − (L ∪ S) cannot
contain an induced P3, as it is anti-complete to L which contains an induced sP1.
As G is connected, each of the complete graphs in G − (L ∪ S) contains at least
one vertex that is adjacent to some vertex of S. Hence in polynomial time we can find
a set S′ of vertices containing all but min{2, X} vertices from each of the complete
graphs X in such a way that G[S ∪ S′] is connected. Then G − (S ∪ S′) is a largest
good induced forest that contains L and no vertex of R ∪ U .
After considering each of the O(n2s2 −2s+3) choices for R, in polynomial time we find a
largest good induced forest that contains L and no vertex of U . After considering each
of the O(ns(c(s)−1)) choices for L, we find in polynomial time a largest good induced
forest that has at least s connected components, each with at most c(s)−1 vertices. ⊓⊔
10
W
s
L
U = N (L)
G − (L ∪ U ∪ W )
Fig. 2: The decomposition of the (sP1 + P3)-free graph G, as given in Step 3 of the
algorithm from the proof of Theorem 6.
Theorem 7. For every s ≥ 0, Odd Cycle Transversal can be solved in polyno-
mial time on (sP1 + P3)-free graphs.
Proof. Let s ≥ 0 be an integer, and let G = (V, E) be an (sP1 + P3)-free graph. We
must describe how to find a smallest odd cycle transversal of G. If s = 0, then we
can use Proposition 1. We now assume that s ≥ 1 and use induction. We will in fact
describe how to solve the complementary problem and find a largest induced bipartite
subgraph of G. The proof is by induction on s and our algorithm performs two steps
in polynomial time, which together cover all possibilities.
Step 1. Compute a largest induced bipartite subgraph B such that every connected
component of B has at least c(s) vertices.
By Lemma 2, we know that B will be connected. Hence, B has a unique bipartition,
which we denote {X, Y }. We first find a largest induced bipartite subgraph B that
is a star: we consider each vertex x and find a largest induced star centred at x by
finding a largest independent set in N (x). This can be done in polynomial time by
Theorem 1.
Next, we find a largest induced bipartite subgraph B that is not a star. We consider
each of the O(n2) choices of edges xy of G and find a largest induced connected
bipartite subgraph B such that x ∈ X and y ∈ Y and neither x nor y has degree 1
in B (since B is not a star, it must contain such a pair of vertices). Note that the
number of vertices in X non-adjacent to y is at most s − 1, otherwise B induces
an sP1 + P3. Similarly there are at most s − 1 vertices in Y non-adjacent to x. We
consider each of the O(n2s−2) possible pairs of disjoint sets X ′ and Y ′, which are each
independent sets of size at most s − 1 such that X ′ ∪ Y ′ is anti-complete to {x, y}.
We will find a largest induced bipartite subgraph with partition classes X and Y such
that {x} ∪ X ′ ⊆ X and {y} ∪ Y ′ ⊆ Y and every vertex in X \ X ′ is adjacent to y and
every vertex in Y \ Y ′ is adjacent to x. That is, we must find a largest independent
11
set in both N (x) \ N ({y} ∪ Y ′) and N (y) \ N ({x} ∪ X ′); see Fig. 3 for an illustration.
This can be done in polynomial time, again by applying Theorem 1.
N (y) \ N ({x} ∪ X ′)
≤ s − 1
x
y
X ′
Y ′
N (x) \ N ({y} ∪ Y ′)
≤ s − 1
Fig. 3: An illustration of Step 1 of the algorithm in the proof of Theorem 7. Full
and dotted lines indicate when two sets are complete or anti-complete to each other,
respectively. The absence of a full or dotted lines indicates that edges may or may not
exist between two sets.
Step 2. Compute a largest induced bipartite subgraph B such that B has a connected
component with at most c(s) − 1 vertices.
We consider each of the O(nc(s)−1) possible choices of a non-empty set L of at most
c(s) − 1 vertices and discard those that do not induce a bipartite graph. We will
find the largest B that has G[L] as a connected component. Let U = N (L), and let
G′ = G − (L ∪ U ). As G′ is ((s − 1)P1 + P3)-free, we can find a largest induced bipartite
subgraph B′ of G′ in polynomial time and B′ + G[L] is a largest induced bipartite
⊓⊔
subgraph among those that have G[L] as a connected component.
Theorem 8. For every s ≥ 0, Connected Odd Cycle Transversal Extension
can be solved in polynomial time on (sP1 + P3)-free graphs.
Proof. Let s ≥ 0 be an integer, let G = (V, E) be a connected (sP1 + P3)-free graph
and let W be a subset of V . We must describe how to find a smallest connected odd
cycle transversal of G that contains W . We will solve the complementary problem: how
to find a largest induced bipartite graph of G that does not include any vertex of W
and whose complement is connected. We will say that an induced bipartite graph B
is good if it has these two properties. Our algorithm consists of three steps, which can
each be performed in polynomial time and which together cover all the possible cases.
Step 1. Compute a largest good induced bipartite subgraph B such that B has a bi-
partition {X, Y } in which one set, say X, has size X ≤ s. (Note that this includes
the case when every connected component of B has at most two vertices and B has at
most s connected components.)
We consider O(ns) choices of an independent set X of at most s vertices of G that does
not intersect W . We wish to find Y , the largest possible independent set in G−(W ∪X)
such that G − (X ∪ Y ) is connected. By Theorem 2, we can do this in polynomial time
12
by computing a minimum connected vertex cover of G−X that contains W and taking
its complement (in G − X).
Step 2. Compute a largest good induced bipartite subgraph B such that B has at least s
connected components and each connected component has at most two vertices.
Note that 2 ≤ c(s) − 1. The algorithm mimics Step 3 of the algorithm in the proof of
Theorem 6, but checks for a good bipartite graph instead of a good forest.
Step 3. Compute a largest good induced bipartite subgraph B such that there is a
connected component of B that has at least three vertices and B has a bipartition
{X, Y } with X ≥ s + 1 and Y ≥ s + 1.
It is in this case that we must do most of the work in proving the theorem, and here
we will need ideas beyond those already met in this section.
As B contains a connected component on at least three vertices, it will contain
an induced P3 and so X ≥ 1 and Y ≥ 1. We consider O(n2s+2) choices of disjoint
independent sets X ′ and Y ′ that each contain s+1 vertices of G and do not intersect W .
If G[X ′ ∪ Y ′] contains an induced P3, our aim is to compute a largest good induced
bipartite graph B with bipartition {X, Y } such that X ′ ⊆ X and Y ′ ⊆ Y ; otherwise
we discard the choice of X ′, Y ′.
We define (see also Fig. 4) a partition of V \ (X ′ ∪ Y ′):
U = (N (X ′) ∩ N (Y ′)) ∪ W
VX = N (X ′) \ (Y ′ ∪ N (Y ′) ∪ W )
VY = N (Y ′) \ (X ′ ∪ N (X ′) ∪ W )
Z = V \ (X ′ ∪ Y ′ ∪ N (X ′) ∪ N (Y ′) ∪ W )
There are a number of steps where our procedure branches as we consider all
possible ways of choosing whether or not to add certain vertices to B. Note that
assuming our choice of X ′ and Y ′ is correct, no vertex of U can be in B. If we decide
that a vertex will not be in B, we will then add it to U .
X ′
VY
W
U
Z
N (X ′) ∩ N (Y ′)
Y ′
VX
Fig. 4: The decomposition of G in Step 3. Full and dotted lines indicate when two
sets are complete or anti-complete to each other, respectively. The absence of a full or
dotted lines indicates that edges may or may not exist between two sets. The circles
in VX and VY represent disjoint unions of complete graphs.
Step 3.1. Reduce Z to the empty set.
Notice that Z does not contain an independent set on more than s−1 vertices otherwise
13
G[X ′ ∪ Y ′ ∪ Z] would contain an induced sP1 + P3. We consider O(n2s−2) choices of
disjoint independent sets ZX and ZY that are each subsets of Z and each contain at
most s − 1 vertices. We move the vertices of ZX and ZY by adding them to X ′ and Y ′,
respectively. We move the vertices of Z \ (ZX ∪ ZY ) by adding them to U . If after this
process is complete there are vertices in VX ∪ VY with neighbours in both X ′ and Y ′,
we move these vertices by adding them to U . We note that now:
-- Z is the empty set,
-- VX still contains vertices with neighbours in X ′ but not in Y ′,
-- VY still contains vertices with neighbours in Y ′ but not in X ′, and
-- U contains vertices that will not be in B.
So our task is to decide how best to add vertices of VX to Y ′ and vertices of VY to X ′,
but first there is another step: as G − B must be connected, and G[U ] is a subgraph
of G − B, we choose some vertices that will not be in B, but will connect together the
connected components of G[U ]. This will not be possible if the vertices of U belong to
more than one connected component of G − (X ′ ∪ Y ′). Hence, in that case we discard
this choice of ZX , ZY .
Step 3.2. Make G[U ] connected.
We consider O(n2s2 −2s+3) choices of sets R of vertices of G − (X ′ ∪ Y ′) such that each
contains at most 2s2 − 2s + 3 vertices. If G[R ∪ U ] is connected, we move the vertices
of R by adding them to U , and so G[U ] becomes connected. Note that since all vertices
of U are in the same connected component of G − (X ′ ∪ Y ′), Lemma 4 implies that at
least one such set R can be found.
Step 3.3. Add vertices from VX to Y ′ and from VY to X ′.
We note that G[VX ] is P3-free, as no vertex of VX has a neighbour in Y ′, Y ′ ≥ s,
and G is (sP1 + P3)-free. By symmetry, G[VY ] is P3-free. Thus both G[VX ] and G[VY ]
are disjoint unions of complete graphs. Note that B can contain at most one vertex
from each of these complete graphs. We consider two subcases.
Step 3.3a. Compute a largest good induced bipartite subgraph B with bipartition
{X, Y } such that X ′ ⊆ X, Y ′ ⊆ Y and G − B contains no edges between VX and VY .
As G − B must be connected, each clique of VX and VY that contains at least two
vertices must contain a vertex adjacent to U (otherwise such a set B cannot exist).
Thus we can form X from X ′ by adding to X ′ one vertex from each clique of VY and
form Y by adding to Y ′ one vertex from each clique of VX in such a way that G − B
is connected. (If we do this, it is possible that G − B will contain an edge from VX
to VY , but then this solution is at least as large as one where such edges are avoided.)
Step 3.3b. Compute a largest good induced bipartite subgraph B with bipartition
{X, Y } such that X ′ ⊆ X, Y ′ ⊆ Y and G − B has an edge xy where x ∈ VX ,
y ∈ VY .
We consider O(n2) choices of an edge xy, x ∈ VX , y ∈ VY . Let vX ∈ X ′ be a neighbour
of x and note that vX , x and y induce a P3 in G. Therefore x must be complete to
all but at most s − 1 cliques of VY . By symmetry, y must be complete to all but at
most s − 1 cliques of VX . A clique in VX or VY is bad if it is not complete to y or x,
respectively. Note that the cliques containing x and y may be bad. We move x and y
to U .
We consider O(n2s−2) choices of a set S of at most 2s − 2 vertices that each belong
to a distinct bad clique and move each to X ′ or Y ′ if they are in VY or VX respectively.
We move the other vertices of the bad cliques to U . If the vertices of U are not in the
same connected component of G − (X ′ ∪ Y ′), we discard this choice of S. We consider
O(n2s2−2s+3) choices of sets R′ of vertices of G − (X ′ ∪ Y ′) such that each contains
14
at most 2s2 − 2s + 3 vertices. If G[R′ ∪ U ] is connected we move the vertices of R′
to U , so G[U ] becomes connected. Since the vertices of U are in the same connected
component of G − (X ′ ∪ Y ′), Lemma 4 implies that at least one such set R′ can be
found.
Note that some cliques might have been completely removed from VX and VY by
the choice of R′. It only remains to pick one vertex from each remaining clique of VX
and VY , and add these vertices to Y ′ or X ′, respectively to finally obtain B. As all
vertices in these cliques are adjacent to x or y we know that G − B will be connected.
⊓⊔
5 The Case H = P6
In this section we prove that Odd Cycle Transversal and Connected Odd Cy-
cle Transversal are NP-hard on (P2 + P5, P6)-free graphs. We do this by modifying
the construction used in [29] for proving that these two problems are NP-complete on
P13-free segment graphs.
Theorem 9. Odd Cycle Transversal and Connected Odd Cycle Transver-
sal are NP-complete on (P2 + P5, P6)-free graphs.
Proof. Both problems are readily seen to belong to NP. To prove NP-hardness we
reduce from Vertex Cover, which is known to be NP-complete [16]. Let (G, k) be
an instance of Vertex Cover. Let n and m be the number of vertices and edges,
respectively, in G. Let v1, . . . , vn be the vertices of G. We construct a graph G∗ from G
as follows.
1. For i ∈ {1, . . . , n} create vertices ai, bi, ci, xi and yi. Let A, B, C, X and Y be the
sets of, respectively, ai, bi, ci, xi and yi vertices.
2. For i, j ∈ {1, . . . , n}, add the edges xiyj and biyj (so we make Y complete to both
X and B).
3. For each i ∈ {1, . . . , n}, add edges xiai, xibi, aibi, bici, ciyi (a vertex gadget, see also
Fig. 5(a) and note that bi is adjacent to yi by the previous step).
4. For each edge vivj in G with i < j, add a vertex di,j adjacent to both xi and yj
(an edge gadget, see also Fig. 5(b)). Let D be the set of di,j vertices.
ai
bi
ci
di,j
xi
yi
xi
yj
(a) Vertex gadget
(b) Edge gadget
Fig. 5: The two gadgets used in the proof of Theorem 9.
We first claim that the following statements are equivalent:
(i) G has a vertex cover of size at most k;
(ii) G∗ has an odd cycle transversal of size at most n + k;
(iii) G∗ has a connected odd cycle transversal of size at most n + k.
15
The implication (iii) ⇒ (ii) is trivial. Below we prove (i) ⇒ (iii) and (ii) ⇒ (i).
(i) ⇒ (iii). Suppose that G has a vertex cover Q of size at most k. We define the set
S = [
vi∈Q
{xi, yi} ∪ [
vi /∈Q
{bi}
and observe that S = 2Q + (n − Q) = n + Q ≤ n + k and that S is connected. We
claim that S is an odd cycle transversal of G∗. This can be seen as follows. The only
induced odd cycles in G∗ are the three triangles in each vertex gadget and the triangle
in each edge gadget. By construction of S, for every i ∈ {1, . . . , n}, either S contains
both xi and yi or S contains bi, thus every triangle in every vertex gadget intersects S.
Furthermore, since Q is a vertex cover of G, for every edge gadget {xi, yj, di,j}, either
xi ∈ S or yj ∈ S. Therefore S intersects every odd cycle in G∗.
(ii) ⇒ (i). Suppose that G∗ has an odd cycle transversal S of size at most n+k. Consider
an edge gadget on {xi, yj, di,j}. If di,j ∈ S then S′ := (S \ {di,j}) ∪ {xi} is an odd cycle
transversal of G with S′ ≤ S. We may therefore assume that S contains no vertices
of D. For i ∈ {1, . . . , n}, the vertex bi intersects all odd cycles in the vertex gadget on
{ai, bi, ci, xi, yi}. If bi /∈ S then S ∩ {ai, bi, ci, xi, yi} ≥ 2 since S intersects all induced
odd cycles of the vertex gadget. Note that {xi, yi} intersects all odd cycles of the vertex
gadget. Therefore, if S ∩ {ai, bi, ci, xi, yi} ≥ 2, then S′ := (S \ {ai, bi, ci}) ∪ {xi, yi} is
an odd cycle transversal of G∗ with S′ ≤ S. We may therefore assume that for every
i ∈ {1, . . . , n}, either bi ∈ S or {xi, yi} ⊆ S and there are no other vertices in S. Let
BS = B ∩ S, XS = S ∩ X and YS = S ∩ Y . Then S = BS + SX + SY = n + SX .
Let Q = Sxi∈S{vi}. Then Q = SX = S − n ≤ n + k − n = k.
We claim that Q is a vertex cover of G. This can be seen as follows. Consider an
edge vivj of G (without loss of generality assume i < j). Then {xi, yj, di,j } ∩ S ≥ 1,
as S is an odd cycle transversal of G∗. By assumption on S, di,j /∈ S and if yj ∈ S
then xj ∈ S. It follows that xi ∈ S or xj ∈ S and so vi ∈ Q or vj ∈ Q. We conclude
that Q is a vertex cover of G of size at most k.
It only remains to show that G∗ is (P2 + P5, P6)-free. Suppose, for contradiction, that
H ∈ {P2 +P5, P6} is an induced subgraph of G∗. Every vertex in A∪C ∪D has degree 2
and its two neighbours are adjacent. Therefore no vertex in V (H) ∩ (A ∪ C ∪ D) is an
internal vertex of a path of H. That is, if x ∈ V (H) ∩ (A ∪ C ∪ D) then x has degree 1
in H. Furthermore, A ∪ C ∪ D is an independent set in G∗. Hence, if H = P2 + P5,
then at most one vertex of the P2 connected component of H can be in A ∪ C ∪ D.
We conclude that G∗[V (H) ∩ (B ∪ X ∪ Y )] contains an induced subgraph H ′ on four
vertices that is isomorphic to P1 + P3 if H = P2 + P5 or P4 if H = P6. Since Y is an
independent set and B ∪ X is a perfect matching, H ′ must contain at least one vertex
of B ∪ X and at least one vertex of Y . As Y is complete to B ∪ X, we find that H ′
contains either C4 or K1,3 as a (not necessarily induced) subgraph, a contradiction.
⊓⊔
This completes the proof.
The proof of Theorem 9 gives a slightly stronger result if we assume the Exponential
Time Hypothesis (ETH). The ETH is one of standard assumptions in complexity
theory which, along with the sparsification lemma, implies that 3-Sat with n variables
and m clauses cannot be solved in 2o(n+m) time [22,23]. The number of vertices in the
graph G∗ constructed in the proof of Theorem 9 is 5n + m. Thus an algorithm solving
(Connected) Odd Cycle Transversal on (P2 +P5, P6)-free graphs with n vertices
in time 2o(n) could be used to solve Vertex Cover on graphs with n vertices and m
edges in 2o(n+m) time. However, such a fast algorithm for Vertex Cover does not
exist unless the ETH fails [11]. Thus we get the following statement.
16
Corollary 1. Odd Cycle Transversal and Connected Odd Cycle Transver-
sal cannot be solved in 2o(n) time on (P2 + P5, P6)-free graphs with n vertices, unless
the ETH fails.
6 Conclusions
We proved polynomial-time solvability of Feedback Vertex Set and Odd Cycle
Transversal on H-free graphs when H = sP1 + P3 and polynomial-time solvability
of their connected variants on H-free graphs, when H = P4 or H = sP1 + P3; see also
Table 1, where we place these results in the context of known results for these problems
on H-free graphs. We also showed that Odd Cycle Transversal and Connected
Odd Cycle Transversal are NP-complete on (P2 + P5, P6)-free graphs.
Natural cases for future work are the cases when H = sP1 + P4 for s ≥ 1 and
H = P5 for all four problems (in particular the case when H = P5 is the only open
case for Odd Cycle Transversal and Connected Odd Cycle Transversal
restricted to Pr-free graphs). Note that Lemma 2 does not hold on (sP1 + P4)-free
graphs: the disjoint union of any number of arbitrarily large stars is even P4-free.
Recall that Vertex Cover and Connected Vertex Cover are polynomial-
time solvable even on (sP1 + P6)-free graphs [20] and (sP1 + P5)-free graphs [24],
respectively, for every s ≥ 0. In contrast to the case for Odd Cycle Transversal
and Connected Odd Cycle Transversal, it is not known whether there is an
integer r for which any of the problems Vertex Cover, Feedback Vertex Set or
their connected variants is NP-complete on Pr-free graphs. Determining whether such
an r exists is an interesting open problem.
We note that a similar complexity study has also been undertaken for the indepen-
dent variants of the problems Feedback Vertex Set and Odd Cycle Transver-
sal.7 In particular, Independent Feedback Vertex Set and Independent Odd
Cycle Transversal are polynomial-time solvable on P5-free graphs [3], but their
complexity status is unknown on P6-free graphs. It is not known whether there is an
integer r such that Independent Feedback Vertex Set or Independent Odd
Cycle Transversal is NP-complete on Pr-free graphs.
We conclude that in order to make any further progress, we must better understand
the structure of Pr-free graphs. This topic has been well studied in recent years, see
also for example [17,19]. However, more research and new approaches will be needed.
References
1. E. Balas and C. S. Yu. On graphs with polynomially solvable maximum-weight clique
problem. Networks, 19(2):247 -- 253, 1989.
2. R. Belmonte, P. van 't Hof, M. Kamiński, and D. Paulusma. The price of connectivity
for feedback vertex set. Discrete Applied Mathematics, 217:132 -- 143, 2017.
3. M. Bonamy, K. K. Dabrowski, C. Feghali, M. Johnson, and D. Paulusma. Independent
feedback vertex set for P5-free graphs. Algorithmica, 81(4):1342 -- 1369, 2019.
4. A. Brandstädt and D. Kratsch. On the restriction of some NP-complete graph problems
to permutation graphs. Proc. FCT 1985, LNCS, 199:53 -- 62, 1985.
5. A. Brandstädt, V. B. Le, and J. P. Spinrad. Graph Classes: A Survey, volume 3 of SIAM
Monographs on Discrete Mathematics and Applications. SIAM, 1999.
6. E. Camby. Price of connectivity for the vertex cover problem and the dominating set
problem: Conjectures and investigation of critical graphs. Graphs and Combinatorics,
35(1):103 -- 118, 2019.
7 Independent Vertex Cover can be seen as 2-Colouring, with the additional restric-
tion that one of the colours can be used at most k times. This problem is polynomial-time
solvable.
17
7. E. Camby, J. Cardinal, S. Fiorini, and O. Schaudt. The price of connectivity for vertex
cover. Discrete Mathematics & Theoretical Computer Science, 16(1):207 -- 224, 2014.
8. E. Camby and O. Schaudt. The price of connectivity for dominating set: Upper bounds
and complexity. Discrete Applied Mathematics, 177:53 -- 59, 2014.
9. J. Cardinal and E. Levy. Connected vertex covers in dense graphs. Theoretical Computer
Science, 411(26 -- 28):2581 -- 2590, 2010.
10. N. Chiarelli, T. R. Hartinger, M. Johnson, M. Milanič, and D. Paulusma. Minimum
connected transversals in graphs: New hardness results and tractable cases using the
price of connectivity. Theoretical Computer Science, 705:75 -- 83, 2018.
11. M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk, M. Pilipczuk,
and S. Saurabh. Parameterized Algorithms. Springer, 1st edition, 2015.
12. B. Escoffier, L. Gourvès, and J. Monnot. Complexity and approximation results for
the connected vertex cover problem in graphs and hypergraphs. Journal of Discrete
Algorithms, 8(1):36 -- 49, 2010.
13. C. Feghali, M. Johnson, G. Paesani, and D. Paulusma. On cycle transversals and their
connected variants in the absence of a small linear forest. Proc. FCT 2019, LNCS,
11651:258 -- 273, 2019.
14. H. Fernau and D. F. Manlove. Vertex and edge covers with clustering properties: Com-
plexity and algorithms. Journal of Discrete Algorithms, 7(2):149 -- 167, 2009.
15. M. R. Garey and D. S. Johnson. The rectilinear Steiner tree problem is NP-complete.
SIAM Journal on Applied Mathematics, 32(4):826 -- 834, 1977.
16. M. R. Garey, D. S. Johnson, and L. J. Stockmeyer. Some simplified NP-complete graph
problems. Theoretical Computer Science, 1(3):237 -- 267, 1976.
17. P. A. Golovach, M. Johnson, D. Paulusma, and J. Song. A survey on the computational
complexity of colouring graphs with forbidden subgraphs. Journal of Graph Theory,
84(4):331 -- 363, 2017.
18. A. Grigoriev and R. Sitters. Connected feedback vertex set in planar graphs. Proc. WG
2009, LNCS, 5911:143 -- 153, 2010.
19. C. Groenland, K. Okrasa, P. Rzążewski, A. Scott, P. Seymour, and S. Spirkl. H-colouring
Pt-free graphs in subexponential time. Discrete Applied Mathematics, in press.
20. A. Grzesik, T. Klimošová, M. Pilipczuk, and M. Pilipczuk. Polynomial-time algorithm for
maximum weight independent set on P6-free graphs. Proc. SODA 2019, pages 1257 -- 1271,
2019.
21. T. R. Hartinger, M. Johnson, M. Milanič, and D. Paulusma. The price of connectivity
for cycle transversals. European Journal of Combinatorics, 58:203 -- 224, 2016.
22. R. Impagliazzo and R. Paturi. On the complexity of k-SAT. Journal of Computer and
System Sciences, 62(2):367 -- 375, 2001.
23. R. Impagliazzo, R. Paturi, and F. Zane. Which problems have strongly exponential
complexity? Journal of Computer and System Sciences, 63(4):512 -- 530, 2001.
24. M. Johnson, G. Paesani, and D. Paulusma. Connected vertex cover for (sP1 + P5)-free
graphs. Algorithmica, in press.
25. G. J. Minty. On maximal independent sets of vertices in claw-free graphs. Journal of
Combinatorial Theory, Series B, 28(3):284 -- 304, 1980.
26. N. Misra, G. Philip, V. Raman, and S. Saurabh. On parameterized independent feedback
vertex set. Theoretical Computer Science, 461:65 -- 75, 2012.
27. R. Mosca. Stable sets for (P6, K2,3)-free graphs. Discussiones Mathematicae Graph
Theory, 32:387 -- 401, 2012.
28. A. Munaro. Boundary classes for graph problems involving non-local properties. Theo-
retical Computer Science, 692:46 -- 71, 2017.
29. K. Okrasa and P. Rzążewski. Subexponential algorithms for variants of homomorphism
problem in string graphs. Proc. WG 2019, LNCS, 11789, to appear.
30. S. Poljak. A note on stable sets and colorings of graphs. Commentationes Mathematicae
Universitatis Carolinae, 15:307 -- 309, 1974.
31. P. L. K. Priyadarsini and T. Hemalatha. Connected vertex cover in 2-connected planar
graph with maximum degree 4 is NP-complete. International Journal of Mathematical,
Physical and Engineering Sciences, 2(1):51 -- 54, 2008.
32. N. Sbihi. Algorithme de recherche d'un stable de cardinalité maximum dans un graphe
sans étoile. Discrete Mathematics, 29(1):53 -- 76, 1980.
18
33. E. Speckenmeyer. Untersuchungen zum Feedback Vertex Set Problem in ungerichteten
Graphen. PhD thesis, Universität Paderborn, 1983.
34. S. Tsukiyama, M. Ide, H. Ariyoshi, and I. Shirakawa. A new algorithm for generating all
the maximal independent sets. SIAM Journal on Computing, 6(3):505 -- 517, 1977.
35. S. Ueno, Y. Kajitani, and S. Gotoh. On the nonseparating independent set problem
and feedback set problem for graphs with no vertex degree exceeding three. Discrete
Mathematics, 72(1 -- 3):355 -- 360, 1988.
36. T. Watanabe, S. Kajita, and K. Onaga. Vertex covers and connected vertex covers in
3-connected graphs. Proc. IEEE International Sympoisum on Circuits and Systems 1991,
2:1017 -- 1020, 1991.
A The Proof of Theorem 2
This appendix is for reviewing purposes only. We will adapt, in a straightforward
way, the proof from [24] for showing that Connected Vertex Cover is polynomial-
time solvable on (sP1 + P5)-free graphs for every s ≥ 1.
We need the following definitions and lemmas. Let G = (V, E) be a graph. The
contraction of an edge uv ∈ E deletes the vertices u and v and replaces them by a
new vertex made adjacent to precisely those vertices that were adjacent to u or v
in G (without introducing self-loops or multiple edges). Recall that a linear forest is
the disjoint union of one or more paths. The following lemma is a straightforward
observation.
Lemma 5. Let H be a linear forest and let G be a connected H-free graph. Then the
graph obtained from G after contracting an edge is also connected and H-free.
We need the following lemmas given in [24].
Lemma 6 ([24]). Let s ≥ 0 and let G be a connected (sP1 + P5)-free graph. Then G
has a connected dominating set D that is either a clique or has size at most 2s2 + s+ 3.
Moreover, D can be found in O(n2s2+s+3) time.
Lemma 7 ([24]). Let J be an independent set in a connected graph G such that J
has a vertex y that is adjacent to every vertex of G − J. Let J ′ consist of those vertices
of J \ {y} that have two adjacent neighbours in G − J (or equivalently, in G). Then a
subset S of the vertex set of G is a connected vertex cover of G that contains J if and
only if S \ J ′ is a connected vertex cover of G − J ′ that contains J \ J ′.
We also need an auxiliary problem defined in [24]. Let G be a connected graph, let
J ⊆ VG be a subset of the vertex set of G and let y be a vertex of J. We call say that
a triple (G, J, y) is cover-complete if it has the following three properties:
(a) J is an independent set;
(b) y is adjacent to every vertex of G − J;
(c) the neighbours of each vertex in J \ {y} form an independent set in G − J.
This leads to the following optimization problem:
Connected Vertex Cover Completion
Instance: a cover-complete triple (G, J, y).
Question: find a smallest connected vertex cover S of G such that J ⊆ S.
We also need the following two lemmas.
Lemma 8 ([24]). Let (G, {y}, y) be a cover-complete triple, where G is an (sP1 +P5)-
free graph for some s ≥ 0. Then it is possible to compute a smallest connected vertex
cover of G that contains y in O(ns+14) time.
19
Lemma 9 ([24]). For every s ≥ 0, Connected Vertex Cover Completion
can be solved in O(n2s+19) time for cover-complete triples (G, J, y), where G is an
(sP1 + P5)-free graph.
We are now ready to prove Theorem 2, which we restate below. The proof mimics
the proof of [24] and as mentioned at the start of this section, we include it only for
reviewing purposes.
Theorem 2 (restated). For every s ≥ 0, Connected Vertex Cover Extension
can be solved in polynomial time on (sP1 + P5)-free graphs.
Proof. Let G be an (sP1 + P5)-free graph on n vertices for some s ≥ 0 and let W ⊆
V (G) be a subset of vertices of G. We may assume without loss of generality that G
is connected. By Lemma 6 we can first compute in O(n2s2+s+3) time a connected
dominating set D that either has size at most 2s2 + s + 3 or is a clique. We note that,
if D is a clique, any vertex cover of G contains all but at most one vertex of D. This
leads to a case analysis where we guess the subset D∗ ⊆ D \ W of vertices not in a
smallest connected vertex cover of G that contains W . That is, we choose a set of at
most one vertex if D is a clique and a set of at most D \ W vertices otherwise, and
eventually look at all such sets. As D \ W ≤ D ≤ 2s2 + s + 3 if D is not a clique,
the number of guesses is O(n2s2+s+3). For each guess of D∗, we compute a smallest
connected vertex cover SD∗ that contains all vertices of (D \ D∗) ∪ W and no vertex
of D∗. Then, at the end, we return one that has minimum size overall. In particular
we note that, since D is a connected dominating set of G, D ∪ W is also a connected
dominating set of G.
Let D∗ be a guess. Before we start our case analysis we first prove the following
claim.
Claim 1. We may assume, at the expense of an O(n16s3+4) factor in the running
time, that D \ D∗ is connected.
We prove Claim 1 as follows. Suppose D\D∗ is not connected. Recall that G[D] is either
a complete graph or has size at most 2s2+s+3. In the first case, G[D\D∗] is connected.
Hence, the second case applies so D has size at most 2s2 + s + 3. Let v ∈ D \ D∗.
As G is (sP1 + P5)-free, G is also P5+2s-free. Hence, for each u ∈ D \ (D∗ ∪ {v}),
every connected vertex cover of G contains a path of at most 5 + 2s − 1 vertices that
connects u to v. We will guess all these u − v-paths (using only vertices from G − D∗)
and add their vertices to D. As the number of paths is at most 2s2 + s + 2, this
branching adds an O(n(5+2s−3)(2s2+s+2)) = O(n16s3+4) factor to our running time and
increases our set D by at most 24s3 extra vertices. We have proven Claim 1.
We distinguish two cases.
Case 1. D∗ = ∅.
We compute a minimum vertex cover S′ of G − (D ∪ W ) in polynomial time by
Theorem 1. To be more precise, this takes O(ns+14) time by using the same arguments
as in the proof of Lemma 8 (see [24]). Clearly S′ ∪ D ∪ W is a vertex cover of G. As D
is a connected dominating set, S′ ∪ D ∪ W is even a connected vertex cover of G. Let
S∅ = S′ ∪ D ∪ W . As S′ is a minimum vertex cover of G − (D ∪ W ), S∅ is a smallest
connected vertex cover of G that contains all vertices of D ∪ W . We remember S∅.
Note that S∅ is found in O(ns+14) time.
Case 2. 1 ≤ D∗ ≤ D (recall that D ≤ 2s2 + s + 3).
Recall that we are looking for a smallest connected vertex cover of G that contains
every vertex of (D \ D∗) ∪ W , but does not contain any vertex of D∗. Hence D∗ must
be an independent set, disjoint from W , and G − D∗ must be connected (if one of
20
these conditions is false, then we stop considering the guess D∗). Moreover, a vertex
cover that contains no vertex of D∗ must contain all vertices of NG(D∗). Hence we
can safely contract not only any edge between two vertices of (D \ D∗) ∪ W , but also
any edge between two vertices in NG(D∗) or between a vertex of (D \ D∗) ∪ W and a
vertex in NG(D∗). We perform edge contractions recursively and as long as possible
while remembering all the edges that we contract. This takes O(n) time. Let G∗ be
the resulting graph.
Note that the set D∗ still exists in G∗, as we did not contract any edges with an
endpoint in D∗. By Claim 1, the set D \ D∗ in G corresponds to exactly one vertex
of G∗. We denote this vertex by y. The set W of G corresponds to an independent
set of G∗. We denote this set by W ∗. We observe the following equivalence, which is
obtained after uncontracting all the contracted edges.
Claim 2. Every smallest connected vertex cover of G∗ that contains {y}∪W ∗ and that
does not contain any vertex of D∗ corresponds to a smallest connected vertex cover
of G that contains (D \ D∗) ∪ W and that does not contain any vertex of D∗, and vice
versa.
As we obtained G∗ in O(n) time, and we can also uncontract all contracted edges
in O(n) time, Claim 2 tells us that we may consider G∗ instead of G. As G is connected
and (sP1 + P5)-free, G∗ is also connected and (sP1 + P5)-free by Lemma 5.
We write J ∗ = NG∗ (D∗) ∪ W ∗ and note that y belongs to NG∗ (D∗) ⊆ J ∗ as D is
connected in G. We now consider the graph G∗ − D∗. As G − D∗ is connected, G∗ − D∗
is connected. By Claim 2, our new goal is to find a smallest connected vertex cover of
G∗ − D∗ that contains J ∗. By our procedure, J ∗ is an independent set of G∗ − D∗.
As D dominates G, we find that D \ D∗ dominates every vertex of G − D∗ that is not
adjacent to a vertex of D∗. Hence the vertex y, which corresponds to the set D \ D∗,
is adjacent to every vertex of (G∗ − D∗) − J ∗ in the graph G∗ − D∗.
Let J ⊆ J ∗ consist of y and those vertices in J ∗ whose neighbourhood in G∗ −D∗ is
an independent set. As y is adjacent to every vertex of (G∗ − D∗) − J ∗ in G∗ − D∗, and
we can remember the set J ∗ \ J, we can apply Lemma 7 and remove J ∗ \ J. That is, it
suffices to find a smallest connected vertex cover of the graph G′ = (G∗ −D∗)−(J ∗ \J)
that contains J.
As J ∗ is an independent set of G∗ − D∗, we find that J is an independent set of G′.
By definition, y ∈ J. As y is adjacent to every vertex of (G∗ − D∗) − J ∗ in G∗ − D∗,
we find that y is adjacent to every vertex in G′ − J. By definition, the neighbours of
each vertex in J \ {y} form an independent set in G′ − J. Hence the triple (G′, J, y) is
cover-complete. This means that we can apply Lemma 9 to find in O(n2s+19) time a
smallest connected vertex cover S′ of G′ that contains J.
We translate S′ in constant time into a smallest connected vertex cover S∗ of
G∗ − D∗ that contains J ∗ by adding J ∗ \ J to S′. We translate S∗ in O(n) time
into a smallest connected vertex cover SD∗ of G that contains (D \ D∗) ∪ W but no
vertex of D∗ by uncontracting any contracted edges. It takes O(n2s+19) time to find
the set SD∗ .
As mentioned, at the end we pick a smallest set of the sets SD∗ . This set is then
a smallest connected vertex cover of G that contains W . As there are O(n2s2+s+3 ·
n16s3+4) such sets, each of which is found in O(n2s+19) time, the total running time
is O(n21s3+26). The correctness of our algorithm follows immediately from the above
⊓⊔
case analysis and the description of the cases.
Note that the algorithm given in Theorem 2 not only solves the decision problem,
but also finds a minimum connected vertex cover of a given (sP1 + P5)-free graph in
polynomial time.
21
|
1309.1645 | 1 | 1309 | 2013-09-06T14:06:14 | Fast ranking algorithm for very large data | [
"cs.DS"
] | In this paper, we propose a new ranking method inspired from previous results on the diffusion approach to solve linear equation. We describe new mathematical equations corresponding to this method and show through experimental results the potential computational gain. This ranking method is also compared to the well known PageRank model. | cs.DS | cs | Fast ranking algorithm for very large data
Dohy Hong
Alcatel-Lucent Bell Labs
Route de Villejust
91620 Nozay, France
{dohy.hong}@alcatel-lucent.com
3
1
0
2
p
e
S
6
]
S
D
.
s
c
[
1
v
5
4
6
1
.
9
0
3
1
:
v
i
X
r
a
Abstract—In this paper, we propose a new ranking method
inspired from previous results on the diffusion approach to
solve linear equation. We describe new mathematical equations
corresponding to this method and show through experimental
results the potential computational gain. This ranking method is
also compared to the well known PageRank model.
Keywords-Large sparse matrix, Iteration, Fixed point, PageR-
ank, Ranking.
I. INTRODUCTION
Inspired from the previous research results on the diffusion
approach [2], [4] to solve fixed point problem in linear algebra,
we propose here a new data ranking definition and algorithm.
This result can be seen as a mix of PageRank solution [5],
diffusion approach [2] and path diversity idea [3].
In Section II, we define the notations and the theoretical
framework. Section III show the first experimental results,
including the comparison to PageRank ranking.
II. ALGORITHM DESCRIPTION
A. Notations
We will use the following notations:
• P ∈ IRN ×N a real matrix;
• I ∈ IRN ×N the identity matrix;
• Ji the matrix with all entries equal to zero except for the
i-th diagonal term: (Ji)ii = 1;
• Ω = {1, .., N };
• I = {i1, i2, .., in, ...} the sequence of nodes for the
update (diffusion): ik ∈ Ω;
• E = (1, .., 1)T ∈ IRN ;
• L1-norm: if X ∈ IRN , X = PN
We assume that P is the matrix associated to the directed
the
graph on Ω, for instance,
transition matrix multiplied by the damping factor [5]).
the PageRank matrix (i.e.
i=1 xi;
B. Fast ranking algorithm
The proposed ranking algorithm is based on the iteration of
the double equations on history Hn and fluid Fn vectors:
H0 = 0
Hn = Hn−1 + Jin ((int)Fn−1)
and
F0 = α.(1, .., 1)T
Fn = Fn−1 + (P − I)Jin ((int)Fn−1).
(1)
(2)
The above equations can be easily interpreted as: we apply
exactly the algorithm of D-iteration [2], but we only diffuse
the integer part of fluids. The Jacobi iterations of the above
equations are defined by:
Hn = Hn−1 + (int)Fn−1
Fn = Fn−1 + (P − I)((int)Fn−1).
If P is a non negative matrix with spectral radius less than
unity, it is obvious to see that Fn and Hn converge (Hn is non-
decreasing bounded by PageRank vector) in a finite number
of steps (if not, Hn would be unbounded) and the proposed
ranking (FR) method is based on H∞+F∞. F∞ can be seen as
a tie-breaker, but one may also use H∞. One may also consider
a personalized PageRank flavour extension replacing F0 by
any other initial vector V . One of this approach’s advantage
is to be not very dependent on the choice of the damping
factor, both for the ranking and the computation speed (cf.
Table II).
To solve the above equations, we will apply the diffusion
approach [2]. This means in particular that this computation
method will be naturally suited for the asynchronous parallel
computation, as it was for D-iteration.
Note that if one would diffuse all fluid retained in F∞, H∞
would be exactly the PageRank vector. The motivations of
using (int)Fn−1 instead of Fn−1 are:
• ranking quality improvement: indeed, as it has been
shown in [3], we think that the original PageRank vector
may be too much influenced by what we could call self-
estimation. In presence of loops (self-loop or loops of
longer length), a part of scores that are inherited will
be returned to the sender, which is not necessarily the
desired property;
• computation/convergence acceleration cf. Figure 1;
• computation/convergence acceleration for the ranking up-
dates when the graph (or matrix P) evolves in time.
If α goes to infinity, the proposed ranking vector converges
to PageRank vector. Therefore, PageRank can be seen as a
particular case α → ∞ of our model where α tunes the desired
influence of loops on the ranking score. Finally, we have also
the following interesting bound on the error:
Theorem 1: (1−d)
αN H∞ is an approximation of the PageRank
vector with L1-norm error bounded by 1/(α − 1).
Proof: The proof is based on the monotone property of
the diffusion. If we denote by H(αE, β) the limit of FI with
initial condition αE and diffusion of the β-integer part of Fn:
(int)(Fn/β) × β, then we have H(αE, 1) = αH(E, 1/α) =
α2H(E/αE, 1/α2) etc. Let’s call X the PageRank vector.
Then, X can be obtained from the diffusion of F (αE, 1) plus
H(αE, 1). Note that the diffusion of F (αE, 1) can be denoted
by H(F (αE, 1), 0) (implying X = H((1−d)/N E, 0) = (1−
d)/N H(E, 0)). Now using H(F (αE, 1), 0) ≤ H(E, 0), we
have:
H(αE, 1) ≤
αN
(1 − d)
and
X = H(αE, 1) + H(F (αE, 1), 0)
H(αE, 1) ≤
αN
(1 − d)
X ≤ H(αE, 1) + H(E, 0).
Then applying FI in 1/α-integer part and the same inequality
recursively, we obtain:
H(αE, 1) ≤
αN
(1 − d)
X ≤H(αE, 1) + H(E, α−1)
+ H(α−1E, α−2) + ...
Therefore
H(αE, 1) ≤
αN
(1 − d)
And
X ≤H(αE, 1) (cid:0)1 + α−1 + α−2 + ...(cid:1)
=
α
α − 1
H(αE, 1)
0 ≤
αN
(1 − d)
X − H(αE, 1) ≤
1
α − 1
H(αE, 1)
which can be rewritten as:
0 ≤ X −
(1 − d)
αN
H(αE, 1) ≤
1
(1 − d)
α − 1
αN
H(αE, 1).
Since H(αE, 1) ≤ αN/(1 − d), we have:
X −
(1 − d)
αN
(cid:12)
(cid:12)
(cid:12)
(cid:12)
(cid:12)
(cid:12)
H(αE, 1)
(cid:12)
(cid:12)
≤
1
α − 1
.
This means that choosing α = 1000 would gives an error
very close to 0.1% for norm L1 but also for each coordinate
(the exact bound is 1/999 = 0.001001001...).
Note that we also have:
(1 − d)
(cid:12)
(cid:12)
X −
(cid:12)(cid:12)
≤
1
αN
−
(cid:12)
(cid:12)
(H(αE, 1) + F (αE, 1))
(cid:12)(cid:12)
1 − d
F (E, α−1).
α − 1
III. EXPERIMENTAL EVALUATION
N
For the experimental evaluation purpose, we took the web
graph imported from the dataset uk-2007-05 @1000000
(available on [1]) which has 41,247,159 links on 106 nodes.
Below we vary N from 103 to 106 extracting from the
the information on the first N nodes. Few graph
dataset
properties are summarized in Table I:
• L: number of non-null entries (links) of P ;
• D: number of dangling nodes (0 out-degree nodes);
N
103
104
105
106
L/N
12.9
12.5
31.4
41.2
D/N
0.041
0.008
0.027
0.046
E/N
0.032
0.145
0.016
0
O/N
0.236
0.114
0.175
0.204
maxin maxout
716
7982
34764
403441
130
751
3782
4655
TABLE I
EXTRACTED GRAPH: N = 103 TO 106.
d
0.85
0.9
0.99
0.999
Jacobi
26
36
330
5076
DI
12
17
101
548
FI (α = 1)
FI (α = 2)
FI (α = 10)
1.72
1.99
3.60
15.3
TABLE II
3.12
3.94
19.6
92.7
6.97
9.55
53.1
258
COMPUTATION COST: N = 103. IMPACT OF DAMPING FACTOR.
COMPUTATION COST: NUMBER OF USE OF COORDINATES OF P DIVIDED
BY L.
• E: number of 0 in-degree nodes: the 0 in-degree nodes
are defined recursively: a node i, having incoming links
from nodes that are all 0 in-degree nodes, is also a 0
in-degree node; from the diffusion point of view, those
nodes are those who converged exactly in finite steps;
• O: number of loop nodes (pii 6= 0);
• maxin = maxi #ini (maximum in-degree, the in-degree
of i is the number of non-null entries of the i-th line
vector of P );
• maxout = maxi #outi (maximum out-degree, the out-
degree of i is the number of non-null entries of the i-th
column vector of P ).
Table II shows the comparative evaluation of the computa-
tion cost in number of iterations (one iteration is here defined
as a use of L coordinates of P in the computation) with
a target precision of 1/N (for L1 norm). We compared the
Jacobi iteration, D-iteration (DI, cf. [2]) and the fast ranking
algorithm (FI) we propose in this paper. The convergence
becomes very slow when the damping factor d is being close
to 1 to compute the PageRank vector, whereas our ranking
vector can be obtained very efficiently whatever the choice of
d.
Figure 1 shows the convergence speed (in number of
iterations) of Jacobi, D-iteration (DI) and the proposed (FI)
methods: our approach reaches the limit in all cases in less
than 2.2 iterations. The convergence efficiency is simply not
comparable.
Figures 2 and 3 compare the ranking results obtained with
FR (using H + F ), LOC (local computation: rank equal to
the number of incoming links) to PageRank vector on the top
x%: on y-axis, it counts the number of common nodes in the
top x% between two ranking methods, then it is divided by
the number of compared elements (nodes). The fifth curve
shows the common elements proportion we observe between
two PageRank vectors using a damping factor of 0.9 and 0.8.
Our FR ranking vector can be seen as an approximation of
PageRank vector, since it converges to PageRank vector for
n
o
i
t
c
a
r
f
:
s
t
n
e
m
e
e
l
n
o
m
m
o
C
1
0.9
0.8
0.7
0.6
0.5
0.4
0
Top X%: FR (alpha=1) vs PR
Top X%: FR (alpha=2) vs PR
Top X%: FR (alpha=10) vs PR
Top X%: PR vs LOC
Top X%: PR(d=0.8) vs PR(d=0.9)
10
20
30
40
50
60
70
80
90
100
Top X%
Fig. 3. N = 106: proportion of common elements between FR (proposed)
and PR (PageRank) in top x% ranked sites.
Applying the diffusion method on this new ranking vector, we
showed that a very efficient computation can be obtained while
targeting a relevant ranking score as PageRank.
REFERENCES
[1] http://law.dsi.unimi.it/datasets.php.
[2] D. Hong. D-iteration method or how to improve gauss-seidel method.
arXiv, http://arxiv.org/abs/1202.1163, February 2012.
Statistical
[3] D. Hong.
reliability and path diversity based pagerank
algorithm improvements. arXiv, http://arxiv.org/abs/1202.2393, Feb 2012.
[4] D. Hong, F. Mathieu, and G. Burnside. Convergence of the d-iteration
algorithm: convergence rate and asynchronous distributed scheme. arXiv,
http://arxiv.org/abs/1301.3007, January 2013.
[5] L. Page, S. Brin, R. Motwani, and T. Winograd. The pagerank citation
ranking: Bringing order to the web. Technical Report Stanford University,
1998.
Jacobi: N=1000
Jacobi: N=10000
Jacobi: N=100000
Jacobi: N=1000000
FI: N=1000
FI: N=10000
FI: N=100000
FI: N=1000000
DI: N=1000000
10
1
0.1
0.01
r
o
r
r
E
0.001
0.0001
1e-05
1e-06
1e-07
0
10
20
30
40
Nb of iterations
50
60
70
Fig. 1. Convergence speed comparison.
large α: however, by its definition, it tends to eliminate self-
ranking aspects due to the presence of loops (a part of scores
that I give to children nodes is coming back to me). Therefore,
the parameter α is meant to tune the influence of loops in the
ranking score and PageRank can be seen as a particular case
α → ∞. Globally, we see that our ranking vector preserves
very well the top ranked web sites (for N = 106, we see
that α = 2 is close enough already to PageRank vector,
always above 92%), because they are likely to be pointed
by many different and relevant other web sites: FR ranking
vector includes by its definition features and ideas of the path
diversity mechanism proposed in [3], when α is closer to 1,
but with a computation cost that is greatly reduced (whereas
the ideas in [3] requires more computation cost than PageRank
vector computation).
Even though it is hard to justify theoretically, the author
believe that a choice of α between 1 and 2 are the most
appropriate in terms of the optimal compromise between
computation cost and ranking relevancy.
n
o
i
t
c
a
r
f
:
s
t
l
n
e
m
e
e
n
o
m
m
o
C
1
0.9
0.8
0.7
0.6
0.5
0.4
0
Top X%: FR (alpha=1) vs PR
Top X%: FR (alpha=2) vs PR
Top X%: FR (alpha=10) vs PR
Top X%: PR vs LOC
Top X%: PR(d=0.8) vs PR(d=0.9)
10
20
30
40
50
60
70
80
90
100
Top X%
Fig. 2. N = 103: proportion of common elements between FR (proposed)
and PR (PageRank) in top x% ranked sites.
IV. CONCLUSION
In this paper, we proposed a new data ranking method and
compared its efficiency to the computation of PageRank vector.
|
1011.2843 | 2 | 1011 | 2010-11-22T15:11:33 | Improved Minimum Cuts and Maximum Flows in Undirected Planar Graphs | [
"cs.DS"
] | In this paper we study minimum cut and maximum flow problems on planar graphs, both in static and in dynamic settings. First, we present an algorithm that given an undirected planar graph computes the minimum cut between any two given vertices in O(n log log n) time. Second, we show how to achieve the same O(n log log n) bound for the problem of computing maximum flows in undirected planar graphs. To the best of our knowledge, these are the first algorithms for those two problems that break the O(n log n) barrier, which has been standing for more than 25 years. Third, we present a fully dynamic algorithm that is able to maintain information about minimum cuts and maximum flows in a plane graph (i.e., a planar graph with a fixed embedding): our algorithm is able to insert edges, delete edges and answer min-cut and max-flow queries between any pair of vertices in O(n^(2/3) log^3 n) time per operation. This result is based on a new dynamic shortest path algorithm for planar graphs which may be of independent interest. We remark that this is the first known non-trivial algorithm for min-cut and max-flow problems in a dynamic setting. | cs.DS | cs | Improved Minimum Cuts
and Maximum Flows in Undirected Planar Graphs
Giuseppe F. Italiano∗
Piotr Sankowski†
November 3, 2018
Abstract
In this paper we study minimum cut and maximum flow problems on planar graphs, both
in static and in dynamic settings. First, we present an algorithm that given an undirected
planar graph computes the minimum cut between any two given vertices in O(n log log n) time.
Second, we show how to achieve the same O(n log log n) bound for the problem of computing
maximum flows in undirected planar graphs. To the best of our knowledge, these are the first
algorithms for those two problems that break the O(n log n) barrier, which has been standing
for more than 25 years. Third, we present a fully dynamic algorithm that is able to maintain
information about minimum cuts and maximum flows in a plane graph (i.e., a planar graph
with a fixed embedding): our algorithm is able to insert edges, delete edges and answer min-cut
and max-flow queries between any pair of vertices in O(n2/3 log3 n) time per operation. This
result is based on a new dynamic shortest path algorithm for planar graphs which may be of
independent interest. We remark that this is the first known non-trivial algorithm for min-cut
and max-flow problems in a dynamic setting.
0
1
0
2
v
o
N
2
2
]
S
D
.
s
c
[
2
v
3
4
8
2
.
1
1
0
1
:
v
i
X
r
a
∗Dipartimento
di
"Tor Vergata",
[email protected]. Work partially supported by the 7th Framework Programme of the EU (Net-
work of Excellence "EuroNF: Anticipating the Network of the Future - From Theory to Design") and by MIUR, the
Italian Ministry of Education, University and Research, under Project AlgoDEEP.
Informatica,
Produzione,
University
of Rome
Sistemi
e
†Institute of Informatics, University of Warsaw and Department of Computer and System Science, Sapienza
University of Rome, [email protected].
1
Introduction
Minimum cut and maximum flow problems have been at the heart of algorithmic research on graphs
for over 50 years. Particular attention has been given to solving those problems on planar graphs,
not only because they often admit faster algorithms than general graphs but also since planar
graphs arise naturally in many applications. The pioneering work of Ford and Fulkerson [3, 4],
which introduced the max-flow min-cut theorem, also contained an elegant algorithm for comput-
ing maximum flows in (s, t)-planar graphs (i.e., planar graphs where both the source s and the sink
t lie on the same face). The algorithm was implemented to work in O(n log n) time by Itai and
Shiloach [9]. Later, a simpler algorithm for the same problem was given by Hassin [6], who reduced
the problem to single-source shortest path computations in the dual graph. The time required
to compute single-source shortest paths in (s, t)-planar graphs was shown to be O(n√log n ) by
Frederickson [5] and later improved to O(n) by Henzinger et al. [8]. As a result, minimum cuts and
maximum flows can be found in O(n) time in (s, t)-planar graphs.
Itai and Shiloach [9] generalized their approach to the case of general planar (i.e., not only
(s, t)-planar) graphs, by observing that the minimum cut separating vertices s and t in a planar
graph G is related to the minimum cost cycle that separates faces fs and ft (corresponding to
vertices s and t) in the dual graph. The resulting algorithm makes O(n) calls to their original
algorithm for (s, t)-planar graphs and thus runs in a total of O(n2 log n) time. In the case of undi-
rected planar graphs, Reif [13] improved this bound by describing how to find the minimum cost
separating cycle with a divide-and-conquer approach using only O(log n) runs of the (s, t)-planar
algorithm: this yields an O(n log2 n) time to compute a minimum cut for undirected planar graphs.
Later on, Frederickson [5] improved the running time of Reif's algorithm to O(n log n). The same
result can be obtained by using more recent planar shortest path algorithms (see e.g., [8]). Hassin
and Johnson [7] extended the minimum cut algorithm of Reif to compute a maximum flow in only
O(n log n) additional time: this implies an undirected planar maximum flow algorithm that runs
in O(n log n) time as well. In summary, the best bound known for computing minimum cuts and
maximum flows in planar undirected graphs is O(n log n).
The first contribution of this paper is to improve to O(n log log n) the time for computing
minimum cuts in planar undirected graphs. To achieve this bound, we improve Reif's classical
approach [13] with several novel ideas. To compute a minimum s-t cut in a planar graph G, we first
identify a path π between face fs (corresponding to vertex s) and face ft (corresponding to vertex
t) in the dual graph GD. Next, we compute a new graph Gπ as follows. We cut GD along path π,
so that the vertices and edges of π are duplicated and lie on the boundary of a new face: another
copy of the same cut graph is embedded inside this face. We show that minimum separating cycles
in GD correspond to some kind of shortest paths in Gπ. Applying a divide-and-conquer approach
on the path π yields the same O(n log n) time bound as previously known algorithms [5, 7, 13].
However, our novel approach has the main advantage that it allows the use of any path π in the
dual graph GD, while previous algorithms were constrained to choose π as a shortest path. We
will exploit the freedom implicit in the choice of this path π to produce a faster O(n log log n) time
algorithm, by using a suitably defined cluster decomposition of a planar graph, combined with
the Dijkstra-like shortest path algorithm by Fakcharoenphol and Rao [2]. Our second contribu-
tion is to show that also maximum flows can be computed in undirected planar graphs within the
same O(n log log n) time bound. We remark that this is not an immediate consequence of our new
minimum cut algorithm: indeed the approach of Hassin and Johnson [7] to extend minimum cut
algorithms to the problem of computing maximum flows has a higher overhead of O(n log n). To
get improved maximum flow algorithms, we have to appropriately modify the original technique of
Hassin and Johnson [7]. To the best of our knowledge, the algorithms presented in this paper are
1
the first algorithms that break the O(n log n) long-standing barrier for minimum cut and maximum
flow problems in undirected planar graphs.
As our third contibution, we present a fully dynamic algorithm that is able to maintain infor-
mation about minimum cuts and maximum flows in a plane graph (i.e., a planar graph with a fixed
embedding): our algorithm is able to insert edges, delete edges and answer min-cut and max-flow
queries between any pair of vertices in O(n2/3 log3 n) time per operation. This result is based on
the techniques developed in this paper for the static minimum cut algorithm and on a new dynamic
shortest path algorithm for planar graphs which may be of independent interest. We remark that
this is the first known non-trivial algorithm for min-cut and max-flow problems in a dynamic setting.
2 Minimum Cuts in Planar Graphs
Let G = (V, E, c) be a planar undirected graph where V is the vertex set, E is the edge set and
c : E → R+ is the edge capacity function. Let the planar graph G be given with a certain embed-
ding. Using the topological incidence relationship between edges and faces of G, one can define the
dual graph GD = (F, ED, cD) as follows. Each face of G gives rise to a vertex in F . Dual vertices f1
and f2 are connected by a dual undirected edge eD whenever primal edge e is adjacent to the faces
of G corresponding to f1 and f2. The weight cD(eD) of the dual edge eD is equal to the weight
c(e) of the primal edge: cD(eD) is referred to as the length of edge eD. In other terms, the length
of the dual edge eD is equal to the capacity of the primal edge e. In the following, we refer to G as
the primal graph and to GD as its dual. Throughout the paper we will refer to vertices of the dual
graph GD interchangeably as (dual) vertices or faces. Note that GD can be embedded in the plane
by placing each dual vertex inside the corresponding face of G, and placing dual edges so that each
one crosses only its corresponding primal edge. Thus, the dual graph is planar as well, although
it might contain multiple edges and self loops. In simpler terms, the dual graph GD of a planar
embedded graph G is obtained by exchanging the roles of faces and vertices, and G and GD are
each other's dual. Figure 1 shows an embedded planar graph G and its dual GD.
Let s and t be any two vertices of G (not necessarily on the same face). We consider the
problem of finding a minimum cut in G between vertices s and t. Let C be a cycle of graph G: we
define the interior of C, denoted by int(C), to be the region inside C and including C in the planar
embedding of the graph G. We can define the exterior ext(C) of the cycle C in a similar fashion. A
cycle C in GD is said to be a cut-cycle if int(C) contains exactly s but not t. The following lemma
was proven by Johnson [10].
Lemma 1. A minimum s-t cut in G has the same cost as a minimum cost cut-cycle of GD.
The lemma follows by the observation that for any cut-cycle C the faces of GD inside int(C)
give a set of vertices S in G which defines a cut separating s and t. Note that Lemma 1 gives an
equivalence between min-cuts in the primal graph G and minimum cost cut-cycles in the dual graph
GD. By using a divide-and-conquer approach, this equivalence can be turned into an efficient al-
gorithm for finding flows in undirected planar graphs [13]. The resulting algorithm, combined with
more recent results on shortest paths in planar graphs [8], is able to work in a total of O(n log n)
time. However, this approach seems to inherently require O(n log n) time, and does not seem to
leave margin for improvements. In the next section, we will present a completely different and more
flexible approach, which will yield faster running times.
2.1 Computing Min-Cuts
Let fs and ft be arbitrary inner faces incident to s and to t respectively. Find any simple path π
from fs to ft in GD. The path π can be viewed as connecting special vertices in the dual graph
2
corresponding to s and t. Hence, any s-t cut needs to cross this path, because it splits s from t.
Let π traverse dual vertices f1, . . . , fk, where f1 = fs and fk = ft. Let us look at the path π as a
horizontal line, with fs on the left and ft on the right (see Figure 2(b)). An edge eD 6∈ π in GD such
that eD is incident to some face fi, 1 ≤ i ≤ k, can be viewed as connected to fi from below or from
above. We now define a new graph G′π, by cutting GD along path π, so that the vertices and edges of
π are duplicated and lie on the boundary of a new face. This is done as follows. Let π′ be a copy of
π, traversing new vertices f′1, f′2, . . . , f′k. Then G′π is the graph obtained from GD by reconnecting to
f′i edges entering fi from above, 1 ≤ i ≤ k (see Figure 2(c)). Let G′′π be a copy of G′π. Turn over the
graph G′′π to make the face defined by π and π′ to be the outer face (see Figure 2(d)). Now, identify
vertices on the path π′ (respectively π) in G′′π with the vertices on the path π (respectively π′) in
G′π. Denote the resulting graph as Gπ (see Figure 2(e)). Note that the obtained graph Gπ is planar.
We define an fi-cut-cycle to be a cut-cycle in GD that includes face fi and does not include any
face fj for j > i. The proof of the following lemma is immediate.
Lemma 2. Let Ci be a minimum fi-cut-cycle in GD for i = 1, . . . , k. Then Ci with minimum cost
is a minimum cut-cycle in GD.
A path ρ between fi and f′i in Gπ is said to be fi-separating if ρ contains neither fj nor f′j, for
j > i. We say that a cycle C in GD touches path π in face fi if two edges on C incident to fi go
both up or both down, whereas we say that C crosses π in face fi if one of these edges goes up,
whereas the other goes down.
Lemma 3. The cost of a minimum fi-cut-cycle in GD is equal to the length of a shortest fi-
separating path in Gπ.
Proof. Let C be some fi-cut-cycle in GD: we show that there must be some fi-separating path ρ in
Gπ having the same cost as C. Note that the fi-cut-cycle C must either cross or touch the path π
in face fi. First, assume that C crosses π in fi. Note that in this case C has to cross the path π an
even number of times (excluding fi), as otherwise C would not separate s from t (see Figure 3(a)).
We can go along C starting from fi in the graph Gπ and each time when C crosses π in GD, we
switch between G′π and G′′π in Gπ. Hence, due to parity of the number of crossings with π, this will
produce a resulting path ρ in Gπ which must end in f′i (see Figure 3(b)). Second, assume that C
touches π in fi. Then C has to cross path π an odd number of times (see Figure 3(c)). Again if we
trace C starting from fi in Gπ we will produce a path ρ in Gπ ending up in f′i (see Figure 3(d)).
Moreover, since C is a fi-cut-cycle in GD, it cannot contain by definition any face fj, for j > i:
consequently, in either case the resulting path ρ in Gπ will contain neither fj nor f′j, for j > i.
Conversely, let ρ be some fi-separating path in Gπ: we show that there must be some fi-cut-
cycle C in GD having the same cost as ρ. First, assume that ρ enters fi and f′i using edges from the
same graph, i.e., either G′π or G′′π. If this is the case, then the two edges must be such that one is
from above and the other is from below. By tracing ρ in GD we obtain a cycle C that crosses π in
fi. This cycle does not need to be simple, but it has to cross π an odd number of times (including
the crossing at fi), as each time when C crosses π the path ρ must switch between G′π and G′′π.
Hence, C must be a cut-cycle. Second, assume that ρ enters fi and f′i using edges from different
graphs, i.e., one edge from G′π and the other edge from G′′π. If this is the case, then the two edges
must be both from above or both from below. Consider first the case where ρ leaves fi from above
using an edge from G′′π, and enters f′i from above using an edge from G′π (see Figure 3(d)). This
time we get a cut-cycle C touching π in fi and intersecting π an odd number of times. The case
where ρ leaves fi from below using an edge from G′π, and enters f′i from below using an edge from
G′′π is completely analogous (see Figure 3(f)). Once again, since ρ is an fi-separating path in Gπ,
in any case the resulting cut cycle C will not contain fj, for j > i.
3
Note that, having constructed the graph Gπ, we can find the shortest fi-separating path by
simply running Dijkstra's algorithm on Gπ where we remove all faces fj and f′j for j > i. By Lem-
mas 1, 2 and 3, the linear-time implementation of Dijkstra's algorithm known for planar graphs [8]
implies an O(n2) time algorithm for computing minimum cuts in planar graphs. There is a more
efficient way of computing minimum cuts that goes along the lines of Reif's recursive algorithm [13].
Before describing this approach, we need to prove some non-crossing properties of fi-cut-cycles.
Lemma 4. For i < j, let Ci be a minimum fi-cut-cycle in GD, and let Cj be a minimum fj-cut-
cycle in GD. Then there exists a cut cycle C ⊆ int(Ci) ∩ int(Cj) in GD such that c(C) ≤ c(Ci).
Proof. If Ci ⊆ int(Cj) then C = Ci and the lemma follows trivially. On the other hand, it is
impossible that Cj ⊆ int(Ci) because in this case Ci would include some face fk on π for k ≥ j > i.
The only possibility left is that Ci * int(Cj) and Cj * int(Ci). In this case, there must exist
a subpath pi of Ci from fa to fb such that pi intersects int(Cj) only at fa and fb. Let pj be the
subpath of Cj going from fa to fb (see Figure 4). We claim that c(pj) ≤ c(pi). Indeed, suppose by
contradiction that c(pj) > c(pi), and let C′j be the cycle obtained from Cj after replacing path pj
with pi. Then the cycle C′j is shorter than Cj. Moreover, since Ci does not include any face fk for k >
i, also C′j cannot include any face fk for k > j > i. As a consequence, the cycle C′j is an fj-cut-cycle
in GD, with c(C′j ) < c(Cj), contradicting our assumption that Cj is a minimum fj-cut-cycle in GD.
Since c(pj) ≤ c(pi), replacing path pi on Ci with the path pj yields a cycle C′i, with c(C′i) ≤ c(Ci).
As long as C′i * int(Cj) and Cj * int(C′i) we can repeat this procedure. At the end, we will obtain
a cycle C ⊆ int(Ci) ∩ int(Cj), such that c(C) ≤ c(Ci). Moreover, the obtained cycle C will be a
cut-cycle, as int(Ci) ∩ int(Cj) contains s.
Lemma 5. For i < j, let Ci be a minimum fi-cut-cycle in GD, and let Cj be a minimum fj-cut-
cycle in GD. Then, for some i′ ≤ i, there exists a minimum fi′-cut-cycle Ci′ ⊆ int(Cj) such that
c(Ci′) ≤ c(Ci).
Proof. Consider the cut-cycle C contained inside int(Ci)∩ int(Cj) as given by Lemma 4, for which
we know that c(C) ≤ c(Ci). As C ⊆ int(Ci), C cannot contain any fk such that k > i. Hence, it is
an fi′-cut-cycle for some i′ ≤ i. The minimum fi′-cut-cycle Ci′ is shorter than C and consequently
it is shorter than Ci. If Ci′ ⊆ int(Cj), the lemma follows. Otherwise, we can apply Lemma 4 to
produce another minimum fi′′-cut-cycle contained in int(Cj) such that c(Ci′′) ≤ c(Ci), for some
i′′ ≤ i′ ≤ i.
The above lemma shows that each computed cut splits the graph into two parts, the interior
and the exterior part of the cut, which can be handled separately. Hence, we can use a divide
and conquer approach on the path π. We first find a minimum cut-cycle that contains the middle
vertex on the path π. Then we recurse on both parts of the path, so there will be O(log n) levels of
recursion in total. In this recursion we will compute minimum cost s-t cuts for all vertices on π. By
Lemma 3, we compute a minimum fi-cut-cycle for some fi by finding a shortest path in the planar
graph Gπ. Then we need to divide the graph into the inside and the outside of the cut. The vertices
on the minimum cut need to be included into both parts, so we need to take care that the total size
of the parts does not increase too much. This can be done in a standard way as described by Reif [13]
or by Hassin and Johnson [7]. Their technique guarantees that on each level of the recursion the
total size of the parts is bounded by O(n). Hence, using the O(n)-time algorithm [8] for shortest
paths we get an O(n log n) time algorithm for finding minimum cuts in planar graphs. Although
this approach yields the same time bounds as previously known algorithms [7, 13], it has the main
advantage to allow the use of any path π in GD, while previous algorithms were constrained to
choose π as a shortest path. As shown in the next section, we can even allow the path π to be
4
non-simple. In Section 3 we will show how to exploit the freedom implicit in the choice of the path
π to produce a faster O(n log log n) time algorithm for finding minimum cuts in planar graphs.
2.2 Using Non-simple Paths
Let π = (v1, . . . , vℓ) be a non-simple path and let v be a vertex appearing at least twice on the path
π, i.e., v = vi = vj for some i < j. We say that the path π is self-crossing in v if the edges incident
to v on π appear in the following circular order (vi−1, vi), (vj−1, vj), (vi, vi+1), (vj, vj+1) in the
embedding. We say that a path π is self-crossing if π is self-crossing in some vertex v. Otherwise
we say that π is non-crossing. If a vertex v appears at least twice on a non-crossing path π than we
say that π touches itself in v. In the previous section we assumed that the path π from fs to ft in
GD was simple, now we will show that we only need the weaker assumption that π is non-crossing.
In order to work with non-crossing paths we will modify the graph GD to make the path π
simple. Let v = vi be a face where π touches itself in GD. Note that there is no other edge from π
between the edges (vi−1, vi), (vi, vi+1) in the circular order around v given by the embedding. Take
all edges Ev incident to v that are embedded between and including the edges (vi−1, vi), (vi, vi+1).
Now, add a new face v′ to GD and make the edges Ev to be incident with v′ instead of v. Finally
connect v with v′ using an undirected edge of length zero (see Figure 5). Let π be a non-crossing
path: perform this vertex-splitting operation until π becomes a simple path. This produced a new
graph GD,π. Note that this transformation does not change the lengths of cut-cycles, so we get the
following observation.
Corollary 6. The lengths of minimum fi-cut-cycles in GD and GD,π are the same.
As a result, if π is a non-simple non-crossing path, instead of running our algorithm on GD, we
can compute the graph GD,π and run our algorithm on GD,π.
3 Cluster Partitions
Our algorithms are based on a particular cluster decomposition. We start by presenting the ideas
behind this cluster decomposition and then show how this decomposition can be effectively ex-
ploited for computing min-cuts. From now on, we assume that we are given a graph for which
the dual graph has vertex degree at most three. This is without loss of generality, since it can be
obtained by triangulating the primal graph with zero capacity edges.
Let n be the number of vertices of G. We first define a cluster partition of G into edge clusters
which will be used by our algorithm. In the cluster partition the graph is decomposed into a set
of cluster P such that each edge of G is included into exactly one cluster. Each cluster contains
two types of vertices: internal vertices, and border vertices. An internal vertex is adjacent only to
vertices in the same cluster, while a border vertex is adjacent to vertices in different clusters. A
hole in a cluster is any face (including the external face) containing only boundary vertices. We
denote by ∂P the set of border vertices of a cluster P . We define an r-partition of an n-node graph
to be a cluster partition that contains at most O( n
r ) clusters, each containing at most r vertices,
O(√r ) border vertices and a constant number of holes. The proof of the following lemma is in
Appendix A. It is based on ideas of Frederickson [5], who constructed a similar partition without
the bound on the number of holes.
Lemma 7. An r-partition of an n-node planar graph can be computed in O(n log r + n√r log n) time.
We use the r-partition to define a representation for shortest paths in a graph that has similar
number of edges, but fewer vertices. In order to achieve this, we use the notion of dense distance
5
graphs. A dense distance graph for a cluster C is defined to be a complete graph over the border
vertices of C where edge lengths correspond to shortest path distances in C. In order to compute
dense distance graphs for all clusters we use Klein's algorithm [11], who have shown that after
O(n log n) preprocessing time any distance from the external face can be computed in O(log n)
time. The proof of the following lemma is given in Appendix B.
Lemma 8. Given an r-partition P we can compute a dense distance graph for all clusters in P in
O(n log r) time.
The dense distance graphs can be used to speed up shortest path computations using Dijkstra's
algorithm. It was shown by Fakcharoenphol and Rao ([2], Section 3.2.2) that a Dijkstra-like algo-
rithm can be executed on a dense distance graph for a cluster P in O(∂P log2 P) time. Having
constructed the dense distance graphs, we can run Dijkstra in time almost proportional to the
number of vertices (rather than to the number of edges, as in standard Dijkstra). We use this
algorithm in graphs composed of dense distance graphs and a subset E′ of edges of the original
graph G = (V, E):
Corollary 9. Dijkstra can implemented in O(E′ log V + Pi ∂Gi log2 ∂Gi) time on a graph
composed of a set of dense distance graphs Gi and a set of edges E′ over the vertex set V .
Proof. In order to achieve this running time we use Fakcharoenphol and Rao [2] data-structure for
each Gi. Moreover, minimum distance vertices from each Gi and all endpoints of edges in E′ are
kept in a global heap.
In general, clusters may contain holes although in typical cases, e.g., in grid graphs, the obtained
clusters are holeless. In this section, in order to introduce the main ideas behind our algorithm,
we restrict ourselves to holeless r-partitions, i.e., r-partitions where each cluster contains one hole
(external face). We will show in Appendix D how to handle the general case.
Assume that we have computed the dense distance graphs for all clusters in a given r-partition
P of the dual graph GD. Recall that in our min-cut algorithm we are free to choose any path π
from fs to ft in GD, as long as π is non-crossing. We choose π to minimize the number of clusters
it crosses and to maximally use the dense representation of clusters.
We define a skeleton graph GP = (∂P, EP ) to be a graph over the set of border vertices in P.
The edge set EP is composed of infinite length edges connecting consecutive (in the order on the
hole) border vertices on each hole. By our holeless assumption, all border vertices in each cluster
lie on the external face of the cluster, so the graph GP is connected. We define a patched graph
to be G = GD ∪ GP . Note that this graph is still planar and the shortest distances do not change
after adding infinite length edges.
to be the graph composed of: (1) two clusters Ps and Pt that include fs and ft
respectively; (2) the dense distance graphs (represented by square matrices) for all other clusters;
r √r ) = O(r + n√r )
(3) the skeleton graph GP (see Figure 6). Note that G
r ) dense distance graphs each over O(√r ) vertices; at most 3r edges of GD from Ps
vertices; O( n
and Pt; additional O( n√r ) edges from the skeleton graph GP . Using Corollary 9 to run Dijkstra's
algorithm we get the following.
contains: O(2r + n
s,t
Define G
s,t
Corollary 10. The shortest paths in G
s,t
can be computed in O((r + n√r ) log2 n) time.
3.1 Recursive Division
Let bs and bt be any border vertices in clusters Ps and Pt respectively. We define π to be com-
posed of: a simple path from fs to bs in Ps; a simple path from bs to bt in GP ; and a simple path
6
from bt to ft in Pt. Note that the construction of the graph G
O(n log r + n√r log n) time by Lemmas 7 and 8.
s,t
and consequently of G
s,t
π takes
s,t
s,t
s,t
i>
<i
and G
:= GD ∩ int(Ci)
:= GD ∩ ext(Ci)
Let Ci be some fi-cut-cycle in GD. After finding Ci, we need to recurse on the graphs GD ∩
int(Ci) and GD ∩ ext(Ci). These graphs cannot be computed explicitly. However, we will show how
to determine G
. Let P be a cluster in
the partition other then Ps or Pt, and let GP be its dense distance graph. For a set of vertices X we
so parts
define GP ∩X to be the dense distance graph of P∩X. The fi-cut-cycle was found using G
of Ci that pass through P correspond to shortest paths. Hence, the shortest paths in P ∩int(Ci) and
P∩ext(Ci) cannot cross the cycle Ci. As a result, distances in GP ∩int(Ci) and GP ∩ext(Ci) between
border vertices of P that are not separated by Ci are the same as in GP . On the other hand, for
border vertices that are separated by Ci the distances are infinite (see Figure 7). We define G
∩ X
to be the graph obtained by taking GP ∩ X for every cluster in G. Note that we have the following.
Corollary 11. G
using G
= G
= G
<i
i>
s,t
s,t
∩ int(Ci) and G
∩ ext(Ci).
s,t
s,t
Using these equalities the construction of G
<i
i>
and G
takes time proportional to the size of G
s,t
s,t
i>
i>
<i
<i
and G
and a minimum cut Ci> in G
2⌋. Next, construct graphs G
only. Given a path π = f1, . . . , fk, the recursive algorithm for computing min-cuts works as follows.
First, remove from G
vertices with degree 2 by merging the two incident edges, and find an fi-cut-
cycle Ci for i = ⌊ k
, and recursively find a minimum cut C<i
in G
. Finally, return the smallest of the three cuts C<i, Ci and Ci>.
To achieve our promised bounds, we need to show that the above algorithm works in sublinear
time. In order to show that, we only need to prove that the total size of the graphs on each level
in the recursion tree is small. Take a graph G
and for each cluster P that contains more then
one border vertex add a new vertex vP and replace each dense distance graph by a star graph with
s,t
c . Obviously,
vP in a center. We call the resulting graph the contracted graph and denote it by G
the contracted graph has more vertices than the original graph: thus, we can bound the number
of vertices in the original graph by considering only contracted graphs. The proof of this lemma is
included in Appendix C.
s,t
Lemma 12. The total number of vertices in contracted graphs on each level in the recursion tree
is O(r + n√r ).
By the above lemma and Corollary 9, running Dijkstra for each level takes O((r + n√r ) log2 n)
time in total. On each level the length of the path π is halved, so there are at most log n recursion
levels. Hence we obtain the following theorem.
Theorem 13. Let G be a flow network with the source s and the sink t. If GD allows holeless
r-partition, then the minimum cut between s and t can be computed in O(n log r + (r + n√r ) log3 n)
time. By setting r = log8 n we obtain an O(n log log n) time algorithm.
Theorem 13 holds also for general r-partitions within the same O(n log log n) time bound. The
modifications needed to make the above algorithm work in the general case are presented in Ap-
pendix D.
4 Computing Maximum Flows
The standard ways of computing maximum flow in near linear time assume that we already have
computed its flow value f . It is given by the min-cut capacity and as we already know it can be
computed in O(n log log n) time. We will use the approach proposed by Hassin and Johnson [7],
7
but adopted to our case as we use a different family of cuts Ci. As argued in the following, their
approach uses only very basic properties of the cut-cycles and can be directly applied to our case.
Moreover, for the sake of brevity, we will assume that that we are given a holeless r-partition. The
general case can be handled using ideas presented in Appendix D, and it will be included in the
full version of this paper.
Let us define the graph −→G π to be the graph obtained from G′π by adding directed edges of length
−f from fi to f′i for all 1 ≤ i ≤ k. After Lemma 4.1 in Miller and Naor [12] we know the following.
Corollary 14. The graph −→G π does not contain negative length cycles.
Let r be such that Cr is the shortest of all Ci for 1 ≤ i ≤ k. The above corollary assures that
distance δ(v) from f′r to a vertex v is well defined in −→G π. The next lemma follows by Theorem 1
in [7] or Section 5.1 in [12].
Lemma 15. Let e = (u, v) be the edge in G and let eD = (fu, fv) be the corresponding edge in GD.
The face fu is defined to lie to the left when going from u to v. The function f (u, v) := δ(fu)−δ(fv)
defines the maximum flow in G.
By this lemma, in order to construct the flow function we only need to compute distances from
f′r in −→G π. The cycles Ci for i = 1, . . . , k divide −→G π into k + 1 subnetworks −→G 0, . . . ,−→G k where for
1 ≤ i ≤ k − 1, −→G i is the subnetwork bounded by and including Ci and Ci+1. The next lemma is a
restatement of Lemma 2 from [7].
Lemma 16. Let v be a vertex in Ni. Then if i < r then there exists a shortest (f′r, v)-path which is
j=0 −→G j. Similarly, if i ≥ r then there exists a shortest (f′r, v)-path which is contained
contained in Sr−1
in Sk
j=r −→G j.
Lemma 16 implies that the computation of δ(v) for v ∈ Sr−1
j=0 −→G j can be restricted to Sr−1
j=0 −→G j
j=r −→G j in a similar fashion.
only. We can restrict the computation for v ∈ Sk
Similarly to [7] let us define a normal path to be a simple (f′r, v)-path ρ(v) = ρr·. . .·ρq·. . .·ρ2q−i
such that, for j = r, . . . , q, subpath ρj is in −→G j, and, for j = q + 1, . . . , 2q − i, subpath ρj is in
−→G 2g−j and uses no edges of negative length. We require as well that q is minimal. Hassin and
Johnson [7] have shown that distances from f′r in an n-vertex graph can be computed in O(n log n)
time when for each vertex v there exists a shortest (f′r, v)-path that is normal (Theorem 2 in [7]).
Their computation is based on Dijkstra's algorithm and can be sped up using dense distance graphs
and Corollary 9.
Lemma 17. Distances from f′r in the graph −→G π can be computed in O(n log r + ( n√r + r) log2 n)
time when for each vertex v there exists a shortest (f′r, v)-path that is normal.
Proof. The algorithm works in two phases. First, we take −→G π and substitute each cluster P with
its dense distance graph. The resulting graph has O( n√r ) vertices. The Dijkstra-like computation
from [7] can be executed in O(( n√r + r) log2 n) time on this graph. In this way we obtain distances
δ(v) for border nodes v in all clusters in GD. Second, for each cluster separately we run Hassin and
Johnson's computation starting from border vertices only. The second phase works in O(r log r)
time for each cluster, which yields O(n log r) time in total.
In order to use the above lemma we only need the following result, which can be proven in the
same way as Lemma 3 from [7]. The proof in [7] uses only the fact that subpaths of cut cycles Ci
are shortest paths and this holds in our case as well.
8
Lemma 18. For any i = r, . . . , k, for every vertex v, in −→G i there exists a normal shortest path ρ(v).
Proof. Assume that a shortest path ρ(v) intersect some cycle Ci. If the previous intersected cycle
was Ci as well, then we can either short cut ρ(v) using the part of Ci or short cut Ci using part
of ρ(v) (see Figure 8). This contradicts either the minimality of ρ(v) or the minimality of Ci.
Moreover, after ρ(v) leaves Ci to go into −→G i−1 it cannot use a negative edge any more, as otherwise
it would cross itself. If it crosses itself then the resulting cycle could be removed as it cannot have
a negative weight.
Combining Lemma 15, Lemma 17 and Lemma 18 we obtain the main result of this section.
Theorem 19. The maximum flow in an undirected planar graph can be computed in O(n log r +
( n√r + r) log2 n) time. By setting r = log8 n we obtain an O(n log log n) time algorithm.
5 Dynamic Shortest Paths and Max Flows in Planar Graphs
Most of the ideas presented in this section are not entirely new, but nevertheless combined together
they are able to simplify and improve previously known approaches. Our dynamic algorithm builds
upon the r-partition introduced in Section 3. We first show how to maintain a planar graph GD
with positive edge weights under an intermixed sequence of the following operations 1:
insertD(x, y, c)
add to GD an edge of length c between vertex x and vertex y, provided that
the new edge preserves the planarity of GD;
deleteD(x, y)
delete from GD the edge between vertex x and vertex y;
shortest pathD(x, y) return a shortest path in GD from vertex x to vertex y.
In our dynamic algorithm we maintain the r-partition of GD together with dense distance graphs
for all clusters in the partition. This information will be recomputed every √r operations. We now
show how to handle the different operations. We start with operation insert: let (x, y) be the edge
to be inserted, and let Px (respectively Py) be the cluster containing vertex x (respectively vertex
y). If x and y are not already border vertices in clusters Px and Py, we make them border vertices
in both clusters and add edge (x, y) arbitrarily either to cluster Px or to cluster Py. Next, we
recompute the dense distance graphs of clusters Px and Py, as explained in Lemma 8. This requires
overall time O(r log r). Note that an insert operation may increase by a constant the number of
border vertices in at most two clusters and adds an edge to one cluster: since the partition into
clusters is recomputed every √r operations, this will guarantee that throughout the sequence of
operations each cluster will always have at most O(r) edges and O(√r) border vertices. To delete
edge (x, y), we remove this edge from the only cluster containing it, and recompute the dense dis-
tance graph of this cluster. Once again, this can be carried out in time O(r log r). Amortizing the
initialization over √r operations yields O(
+ r log r) time per update.
log n
n log r+ n√r
√r
In order to answer a shortest path(x, y) query, we construct the graph G
tion 3). Note that the distance from s to t in GD and G
shortest path from s to t can be computed in O((r + n√r ) log2 n) time.
s,t
(as defined in Sec-
are equal. Hence, by Corollary 10 the
s,t
In order to minimize the update time we set r = n2/3 and obtain the following theorem.
1We have chosen to work with the dual graph GD for consistency with the remaining parts of the paper.
9
Lemma 20. Given a planar graph G with positive edge weights, we can insert edges (allowing
changes of the embedding), delete edges and report shortest paths between any pair of vertices in
O(n2/3 log2 n) amortized time per operation. 2
We recall that we can check whether a new edge insertion violates planarity within the bounds
of Lemma 20: indeed the algorithm of Eppstein et al. [1] is able to maintain a planar graph subject
to edge insertions and deletions that preserve planarity, and allow to test whether a new edge
would violate planarity in time O(n1/2) per operation. Finally, we observe that not only we have
improved slightly the running time over the O(n2/3 log7/3 n) time algorithm by Fakcharoenphol and
Rao [2], but our algorithm is also more general. Indeed, the algorithm in [2] can only handle edge
cost updates and it is not allowed to change the embedding of the graph. On the contrary, our
algorithm can handle the full repertoire of updates (i.e., edge insertions and deletions) and it allows
the embedding of the graph to change throughout the sequence of updates.
We now turn our attention to dynamic max-flow problems. In particular, given a planar graph
G = (V, E) we wish to perform an intermixed sequence of the following operations:
insert(x, y, c)
delete(x, y)
max flow(s, t)
add to G an edge of capacity c between vertex x and vertex y, provided that
the embedding of G does not change;
delete from G the edge between vertex x and vertex y;
return the value of the maximum flow from vertex s to vertex t in G.
Note that insert operations are now more restricted than before, as they are not allowed to
change the embedding of the graph. To answer max flow queries in the primal graph G we need to
maintain dynamically information about the distances in the dual graph GD, with the bounds re-
ported in Lemma 20. Unfortunately, things are more involved as a single edge change in the primal
graph G may cause more complicated changes in the dual graph GD. In particular, inserting a new
edge into the primal graph G results in splitting into two a vertex in the dual graph GD, whereas
deleting an edge in the primal graph G implies joining two vertices of GD into one. However, as
edges are inserted into or deleted from the primal graph, vertices in the dual graph are split or
joined according to the embedding of their edges. To handle efficiently vertex splits and joins in
the dual graph, we do the following. Let f be a vertex of degree d in the dual graph: we maintain
vertex f as a cycle Cf of d edges, each of cost 0. The actual edges originally incident to f , are made
incident to one of the vertices in Cf in the exact order given by the embedding. It is now easy to
see that in order to join two vertices f1 and f2, we need to cut their cycles Cf1 and Cf2, and join
them accordingly. This can be implemented in a constant number of edge insertions and deletions.
Similarly, we can support vertex splitting with a constant number of edge insertions and deletions.
Additionally, for each cluster P , for each pair h, h′ of holes we need to compute the dense distance
and the minimum-cuts between bh and bh′. However, following Corollary 24, this
graphs of Pπh,h′
does not increase the running time of our dynamic algorithm. Note that this information is enough
π in O(r + n√r ) time. Moreover, by Theorem 13 and Theorem 26
to construct the graphs G
the min-cut algorithm can be executed on these graphs in O((r + n√r ) log3 n) time. Setting r = n2/3
this yields immediately the main result of this section.
and G
s,t
s,t
Lemma 21. Given a planar graph G with positive capacities, each operation insert, delete and
max flow can be implemented in O(n2/3 log3 n) amortized time. 3
2The same worst-case time bounds can be obtained using a global rebuilding technique.
3Again, the same worst-case time bounds can be obtained using a global rebuilding technique.
10
References
[1] David Eppstein, Zvi Galil, Giuseppe F. Italiano, and Thomas H. Spencer. Separator based
sparsification I. Planarity testing and minimum spanning trees. Journal of Computer and
System Sciences, 52(1):3 -- 27, February 1996.
[2] J. Fakcharoenphol and S. Rao. Planar graphs, negative weight edges, shortest paths, and
near linear time. J. Comput. Syst. Sci., 72(5):868 -- 889, 2006.
[3] L. R. Ford and D. R. Fulkerson. Maximal flow through a network. Canadian Journal of
Mathematics, pages 399 -- 404, 1956.
[4] L. R. Ford and D. R. Fulkerson. Flows in Network. Princeton Univ. Press, 1962.
[5] Greg N. Frederickson. Fast algorithms for shortest paths in planar graphs, with applications.
SIAM Journal on Computing, 16(6):1004 -- 1022, 1987.
[6] R. Hassin. Maximum flows in (s, t) planar networks. IPL, page 107, 1981.
[7] Refael Hassin and Donald B. Johnson. An O(n log2 n) algorithm for maximum flow in
undirected planar networks. SIAM J. Comput., 14(3):612 -- 624, 1985.
[8] Monika R. Henzinger, Philip Klein, Satish Rao, and Sairam Subramanian. Faster shortest-path
algorithms for planar graphs. J. Comput. Syst. Sci., 55(1):3 -- 23, 1997.
[9] Alon Itai and Yossi Shiloach. Maximum flow in planar networks. SIAM Journal on Computing,
8(2):135 -- 150, 1979.
[10] Donald B. Johnson. Parallel algorithms for minimum cuts and maximum flows in planar
networks. J. ACM, 34(4):950 -- 967, 1987.
[11] Philip N. Klein. Multiple-source shortest paths in planar graphs. In SODA '05: Proc. 16th
Annual ACM-SIAM Symposium on Discrete algorithms, pages 146 -- 155, 2005.
[12] Gary L. Miller and Joseph Naor. Flow in planar graphs with multiple sources and sinks
(extended abstract). In FOCS, pages 112 -- 117. IEEE, 1989.
[13] J. Reif. Minimum s-t cut of a planar undirected network in O(n log2 n) time. SIAM Journal
on Computing, 12:71 -- 81, 1983.
11
fs
s
t
ft
Figure 1: An embedded planar graph G and its dual graph GD. Vertices of G are shown as circles,
and edges of G are solid. Vertices of GD are shown as gray squares, and edges of GD are dashed. s
and t are two vertices in G, and fs and ft are arbitrary inner faces incident respectively to s and t.
12
fs
f '
s
fs
tf
tf '
tf
(a)
(c)
fs
tf
fs
'
fs
(b)
(d)
tf
'
tf
(e)
Figure 2:
(a) The dual graph GD of Figure 1: the path π from fs to ft is shown in bold. (b)
The dual graph GD embedded so that the path π is a horizontal line. (c) The graph G′π. (d) The
graph G′′π = G′π, embedded so that the face defined by π and π′ is the outer face. (e) The graph Gπ
obtained after identifying the path π′ (respectively π) in G′′π with the path π (respectively π′) in G′π.
13
C
fs
fi
C
fs
fs
C
(a)
(c)
(e)
fi
fi
ft
ft
ft
'
f
f
fs
fs
fs
fs
'
fi
fi
fi'
fi
fi'
fi
(b)
(d)
(f)
'
f
f
ft
ft
ft
ft
(b) The corresponding fi-separating path ρ in Gπ.
Figure 3: On the proof of Lemma 3. (a) The fi-cut-cycle C crosses the path π in fi in the graph
GD.
(c) The fi-cut-cycle C touches from
above the path π in fi in the graph GD. (d) The corresponding fi-separating path ρ in Gπ. (e)
The fi-cut-cycle C touches from below the path π in fi in the graph GD. (f) The corresponding
fi-separating path ρ in Gπ.
Ci
pj
fs
fa
pi
fb
fi
ft
fj
Cj
Figure 4: On the proof of Lemma 4.
14
vi+1
v=vi
vi-1
(a)
v=vi
vi+1
0
v'
vi-1
(b)
Figure 5: Dealing with non-simple paths. (a) A non-crossing path π of GD that touches itself at
face v = vi. (b) The vertex splitting transformation on face v.
Pt
ft
Ps
fs
Figure 6: The graph G
clusters, and the skeleton graph GP . The path π from fs to ft is shown in bold.
containing clusters Ps and Pt, the dense distance graphs for all other
s,t
Ci
w
u
v
P
Figure 7: Computing GP ∩ int(Ci).
It two border vertices u and v are not separated by the
fi-cut-cycle Ci, then their distance in GP ∩ int(Ci) is equal to their distance in GP . On the
other hand, if border vertices v and w are separated by the fi-cut-cycle Ci, then their distance in
GP ∩ int(Ci) is infinite.
15
'fr
fr
'fr
fr
y
v
ft
ft
x
(a)
(b)
Ci
Ci
Figure 8: On the proof of Lemma 18. (a) The cycle Ci can be short cut using the subpath between
x and y in ρ(v). (b) A valid normal path.
16
A Computation of r-Partition: Proof of Lemma 7
In order to prove Lemma 7 we will combine Frederickson's [5] technique with the following impli-
cation of Section 3.1 and Section 5.1 from [2].
Corollary 22. An r-partition of an n-node planar graph that already contains O( n√r ) border ver-
tices can be computed in O(n log n) time.
Proof. Actually, Fakcharoenphol and Rao [2] construct in the above time bound a recursive decom-
position and in order to get an r-partition we only need to run their algorithm until the size of the
clusters drops below O(r).
Moreover, the recursive decomposition of Fakcharoenphol and Rao [2] assumes that border ver-
tices are present in the decomposed graphs or clusters. Moreover, each time a cluster is split into
smaller clusters, the border vertices are split into asymptotically equal parts. The O( n√r ) border
vertices we start with will be distributed equally into O( n
r ) clusters. Hence, each cluster in the
obtained partition will have additional O( n√r × r
n ) = O(√r ) border vertices.
Now in order to the find the r-partition quickly we use the following algorithm:
• Generate a spanning tree T of the graph;
• Find connected subtrees of T containing O(√r ) vertices using a procedure from [5];
• Contract the graph on these subsets, to obtain a simple planar graph Gs with O( n√r ) vertices;
• Using Corollary 22, find an r-division in Gs with O( n
• Expand Gs back to G. In G there are O(n/r) clusters P1 of size O(√r ) resulting from boundary
vertices in Gs, and O( n
r3/2 ) clusters P2 of size O(r3/2) resulting from the interior vertices of Gs;
• Apply Corollary 22 to find an r-division for clusters in P2 taking into account the border vertices
r3/2 ) clusters of size O(r);
already present in the clusters.
It is easy to see that the above procedure requires O(n log r + n√r log n) time. We only need to show
that the result is a valid r-division. The clusters P1 correspond exactly to the connected subtrees
of T of size O(√r ). They cannot have more then O(√r ) border vertices and all the border vertices
lie on the external face, i.e., they contain one hole. Hence, clusters in P1 satisfy the properties of
an r-partition.
Consider the clusters in P2 before we have applied Corollary 22. Each cluster is obtained by
expanding a cluster in Gs that had √r border vertices. Consider the process of expanding to a
subtree T a border vertex bh lying on a hole h in a cluster P . Let h′ be the hole obtained from h
by removing bh from P . The edges of T were not present in the cluster so the process of expanding
T can be seen as gluing T at the side of h′. Note that by the connectivity of T no new hole is
created and only the hole h′ becomes "smaller". Moreover, not all of the vertices of T become
border vertices of the expanded piece, as some of them do not lie on the side of the new hole (see
Figure 9). Nevertheless as T contains O(√r ) vertices at most O(√r ) vertices can become border
vertices. Hence, in total the pieces in P2 will have O(r) border vertices. This number satisfies the
assumption of Corollary 22, so pieces obtained by using it satisfy the assumption of r-partition.
This completes the proof of Lemma 7.
B Computation of Dense Distance Graphs: Proof of Lemma 8
Proof. In order to compute dense distance graphs we use the following result by Klein [11].
Theorem 23 (Klein [11]). Given an n-node planar graph with non-negative edge lengths, it takes
O(n log n) time to construct a data structure that supports queries of the following form in O(log n)
17
a)
b)
bh
h
T
h'
Figure 9: a) Before expanding the border vertex bh in hole h; b) after expanding bh to T no new
hole is created.
time: given a destination vertex t on the boundary of the external face, and given a start vertex s
anywhere, find the distance from s to t.
A cluster C in the r-partition has r vertices, O(√r ) border vertices and a constant number of
holes. Border vertices in C lie on one of the holes. In order to compute distances from all border
vertices to a given hole H we will apply Theorem 23. We simply find an embedding of the graph
such that H becomes the external face, and query the distances from all border vertices to H.
There are a constant number of holes in each cluster, and so Klein's data structure will be used
a constant number of times for each cluster. On the other hand, there are O(r) pairs of border
vertices in each cluster, so we will ask O(r) queries. The time needed to process each cluster is
hence O(r log r), which gives n
r × O(r log r) = O(n log r) time in total.
C Size of Contracted Graphs: Proof of Lemma 12
<i
i>
and G
s,t
is split along Ci into G
that has O(r + n√r ) vertices and
.
s,t
Proof. At the top level of the recursion we have one graph G
c
s,t
O(r + n√r ) faces. Now consider the case when the graph G
s,t
Observe that the cycle Ci can be traced in G
c when we replace each edge in dense distance graphs
for a cluster P by a two-edge path going through the center vertex vP . In such a case we have
<i
c ∩ int(Ci) and G
c = G
c ∩ ext(Ci). In other words each time we recurse we split the
G
s,t
c along some cycle and then remove degree-two vertices. Note that after splitting the graph
graph G
s,t
G
c along the cycle Ci, the number of faces in the union of the resulting graphs increases by exactly
one (see Figure 10). There are at most O(r + n√r ) recursive calls, so the union all contracted graphs
on a given level in the recursive tree has O(r+ n√r ) faces as well. Moreover, the vertices in this union
graph have degree at least three, because all degree-two vertices are removed. Let v, e and f be the
number of vertices, edges and respectively faces in the union graph. Now, by Euler's formula the
claim of the lemma follows: v = 3v−2v = 3v−2(2+e−f ) ≤ 2e−2(2+e−f ) ≤ 2f−4 = O(r+ n√r ).
i>
c = G
s,t
D General Clusters
When holes are present in the r-partition the skeleton graph GP is not connected. Hence, the path
connecting bs and bt cannot use border vertices only. We need to modify the algorithm to allow
such paths. We do this as follows.
18
Pt
ft
Ps
fs
(a)
Pt
ft
(b)
Ps
fs
(c)
s,t
(a) The graph G
c
Figure 10:
the cycle Ci is shown with dashed edges. (b) The graph G
G
and the cycle Ci. Cluster boundaries are shown in bold, and
c ∩ ext(Ci). (c) The graph
contains one more face than G
i>
c = G
s,t
<i
c = G
s,t
c ∩ int(Ci). Note that the union of G
i>
c and G
<i
c
s,t
c .
19
For each hole h in P , we fix a border vertex bh. For each pair of holes h, h′ in P , we fix a path
πh,h′ that starts from bh and ends in bh′, goes through bh′′ for all holes h′′ in P , and for all bh′′ on
the path walks around the hole h′′ passing through all its border vertices (see Figure 11). These
paths are used to do some additional preprocessing for each cluster P in the partition. For each
pair of holes h, h′ in P we compute dense distance graph for Pπh,h′
, and find the minimum cut Ch,h′
between bh and bh′ in P .
h
b
h
bh'
h'
bh''
h''
Figure 11: The path πh,h′.
Corollary 24. The additional preprocessing takes O(n log r) time.
Proof. For each cluster P and each pair of holes the dense distance graph can be computed in the
same manner as in Lemma 8. On the other hand, minimum-cuts can be found in O(r log r) time
using the algorithm by Henzinger et al. [8]. Hence, over all clusters we need a total of O(n log r)
time.
Now in order to connect fs and ft we will use paths πh,h′ whenever we need to pass between
two different holes h, h′ in a piece P . Let Pπ be the set of all such pieces on π. The resulting path
is no longer simple, but it will be non-crossing. As shown in Section 2.2, our min-cut algorithm can
be executed on non-crossing paths as well. We can make the following observation (see Figure 12).
Corollary 25. A minimum s-t cut C either contains a vertex in ∂π or is fully contained in one
of the pieces Pπ.
Proof. If the cycle C contains a vertex in ∂π, we are done. Assume that it does not contain any
vertex from ∂π. In order to be a cut, it has to cross path π and it can do it so in one of πh,h′.
Then it has to be fully contained in the corresponding P as by the construction of πh,h′ or border
vertices of P lie on π.
Using Corollary 25, we can find the minimum cut in two phases. First, let Ci be the smallest
of the cuts Ch,h′ in Pπ for bh, b′h ∈ π. Second, run the algorithm from previous section on a path
∂π in G
to find a cut Cb. Finally, return the smallest of the cuts Ci and Cb. The running time
of the above algorithm is the same as in Theorem 13.
s,t
Theorem 26. Let G be a flow network with source s and sink t. The minimum cut between s and
t in G can be computed in O(n log r + (r + n√r ) log3 n) time. By setting r = log8 n, we obtain an
O(n log log n) time algorithm.
20
Figure 12: On illustrating Corollary 25. A cut can be either fully contained in one of the pieces
Pπ (such as cut Ci) or it must contain one of the border vertices in ∂π (such as cut Cj).
21
|
1804.10673 | 1 | 1804 | 2018-04-27T20:19:39 | Buffered Count-Min Sketch on SSD: Theory and Experiments | [
"cs.DS"
] | Frequency estimation data structures such as the count-min sketch (CMS) have found numerous applications in databases, networking, computational biology and other domains. Many applications that use the count-min sketch process massive and rapidly evolving datasets. For data-intensive applications that aim to keep the overestimate error low, the count-min sketch may become too large to store in available RAM and may have to migrate to external storage (e.g., SSD.) Due to the random-read/write nature of hash operations of the count-min sketch, simply placing it on SSD stifles the performance of time-critical applications, requiring about 4-6 random reads/writes to SSD per estimate (lookup) and update (insert) operation.
In this paper, we expand on the preliminary idea of the Buffered Count-Min Sketch (BCMS) [15], an SSD variant of the count-min sketch, that used hash localization to scale efficiently out of RAM while keeping the total error bounded. We describe the design and implementation of the buffered count-min sketch, and empirically show that our implementation achieves 3.7x-4.7x the speedup on update (insert) and 4.3x speedup on estimate (lookup) operations.
Our design also offers an asymptotic improvement in the external-memory model [1] over the original data structure: r random I/Os are reduced to 1 I/O for the estimate operation. For a data structure that uses k blocks on SSD, was the word/counter size, r as the number of rows, M as the number of bits in the main memory, our data structure uses kwr/M amortized I/Os for updates, or, if kwr/M >1, 1 I/O in the worst case. In typical scenarios, kwr/M is much smaller than 1. This is in contrast to O(r) I/Os incurred for each update in the original data structure. | cs.DS | cs | Buffered
Count-Min Sketch on SSD: Theory and Experiments
Mayank Goswami
Queens College, City University of New York
[email protected]
Dzejla Medjedovic
International University of Sarajevo
[email protected]
Emina Mekic
Sarajevo School of Science and Technology
[email protected]
Prashant Pandey
Stony Brook University, New York
[email protected]
Abstract
Frequency estimation data structures such as the count-min sketch (CMS) have found numerous
applications in databases, networking, computational biology and other domains. Many applications
that use the count-min sketch process massive and rapidly evolving datasets. For data-intensive ap-
plications that aim to keep the overestimate error low, the count-min sketch may become too large
to store in available RAM and may have to migrate to external storage (e.g., SSD.) Due to the
random-read/write nature of hash operations of the count-min sketch, simply placing it on SSD stifles
the performance of time-critical applications, requiring about 4-6 random reads/writes to SSD per
estimate (lookup) and update (insert) operation.
In this paper, we expand on the preliminary idea of the Buffered Count-Min Sketch (BCMS) [15],
an SSD variant of the count-min sketch, that used hash localization to scale efficiently out of RAM
while keeping the total error bounded. We describe the design and implementation of the buffered
count-min sketch, and empirically show that our implementation achieves 3.7×-4.7× the speedup on
update (insert) and 4.3× speedup on estimate (lookup) operations.
Our design also offers an asymptotic improvement in the external-memory model [1] over the
original data structure: r random I/Os are reduced to 1 I/O for the estimate operation. For a data
structure that uses k blocks on SSD, w as the word/counter size, r as the number of rows, M as the
number of bits in the main memory, our data structure uses kwr/M amortized I/Os for updates, or,
if kwr/M >1, 1 I/O in the worst case. In typical scenarios, kwr/M is much smaller than 1. This is
in contrast to O(r) I/Os incurred for each update in the original data structure.
Lastly, we mathematically show that for the buffered count-min sketch, the error rate does not
substantially degrade over the original count-min sketch due to hash localization. Specifically, we prove
that for any query q, our data structure provides the guarantee: Pr[Error(q)≥ n(1+o(1))]≤ δ+o(1),
which, up to o(1) terms, is the same guarantee as that of a count-min sketch.
2012 ACM Subject Classification Theory of computation → Data structures and algorithms for
data management, Theory of computation → Streaming models, Theory of computation → Database
query processing and optimization (theory),
Keywords and phrases Streaming model, Count-min sketch, Counting, Frequency, External memory,
I/O efficiency, Bloom filter, Counting filter, Quotient filter.
Digital Object Identifier 10.4230/LIPIcs.ESA.2018.
8
1
0
2
r
p
A
7
2
]
S
D
.
s
c
[
1
v
3
7
6
0
1
.
4
0
8
1
:
v
i
X
r
a
© Mayank Goswami and Dzejla Medjedovic and Emina Mekic and Prashant Pandey;
licensed under Creative Commons License CC-BY
European Symposium on Algorithms (ESA 2018).
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
XX:2
Buffered Count-Min Sketch on SSD: Theory and Experiments
1
Introduction
Applications that generate and process massive data streams are becoming pervasive [3,16,20,22,28]
across many domain in computer science. Common examples of streaming datasets include fin-
ancial markets, telecommunications, IP traffic, sensor networks, textual data, etc [3,8,11,29].
Processing fast-evolving and massive datasets poses a challenge to traditional database systems,
where commonly the application stores all data and subsequently does queries on it. In the
streaming model [4], the dataset is too large to be completely stored in the available memory,
so every data item is seen and processed once - an algorithm in this model performs only one
scan of data, and uses sublinear local space.
with the goal of reporting all items whose frequency is at least n/k, n=PT
The streaming scenario exhibits some limitations on the types of problems we can solve with
such strict time and space constraints. A classic example is the heavy hitter problem HH(k) on the
stream of pairs (at,ct), where at is the item identifier, and ct is the count of the item at timeslot t,
t=1ct. The general version
of the problem, with the exception of when k is a small constant1, can not be exactly solved in the
streaming model [25,29], but the approximate version of the problem, -HH(k), where all items of
the frequency at least n/k−n are reported, and an item with larger error might be reported with
small probability δ, is efficiently solved with the count-min sketch [12,21] data structure. Count-min
sketch accomplishes this in O(ln(1/δ)/) space, usually far below linear space in most applications.
Count-min sketch [12,21] has been extensively used to answer heavy hitters, top k queries and
other popularity measure queries that represent the central problem in the streaming context, where
we are interested in extracting the essence from an impractically large amount of data. Common
applications include displaying the list of bestselling items, the most clicked-on websites, the hottest
queries on the search engine, most frequently occurring words in a large text, and so on [22,27,30].
Count-min sketch (CMS) is a hashing-based, probabilistic and lossy representation of a multiset,
that is used to answer the count of a query q (number of times q appears in a stream). It has
two error parameters: 1) , which controls the overestimation error, and 2) δ, which controls the
failure probability of the algorithm. The CMS provides the guarantee that the estimation error
for any query q is more than n with probability at most δ. If we set r=ln(1/δ) and c=e/, the
CMS is implemented using r hash functions as a 2D array of dimensions r x c.
When and δ are constants, the total overestimate grows proportionately with n, the size of
the count-min sketch remains small, and the data structure easily fits in smaller and faster levels
of memory. For some applications, however, the allowed estimation error of n is too high when
is fixed. Consider an example of n=230, where δ=0.01 and =2−26, hence the overestimate
is 16, and the total data structure size of 3.36GB, provided each counter uses 4 bytes. However,
if we double the dataset size, then the total overestimate also doubles to 32 if stays the same.
On the other hand, if we want to maintain the fixed overestimate of 16, then the data structure
size doubles to 6.72GB.
In this paper, we expand on the preliminary idea of Buffered Count-Min Sketch (BCMS) [15],
an SSD variant of the count-min sketch data structure, that scales efficiently to large datasets
while keeping the total error bounded. Our work expands on the previous work by introducing
detailed design, implementation and experiments, as well as mathematical analysis of the new
data structure (our original paper [15], which, to the best of our knowledge is the only attempt
thus far to scale count-min sketch to SSD, contains only the outline of the data structure).
To demonstrate the issues arising from a growing count-min sketch and storing it in lower
levels of memory, we run a mini in-RAM experiment for count-min sketch sizes 4KB-64MB. In
1 When k≈2 this problem goes by the name of majority element.
M. Goswami, D. Medjedovic, E. Mekic, and P. Pandey
XX:3
Figure 1 The effect of increasing count-min sketch size on the update operation cost in RAM.
Figure 1, we see that to maintain the same error, the cost of update will increase as the data
structure is being stored in the lower levels of memory, even though we keep the number of hash
functions fixed for all data structure sizes. The appropriate peak in the cost is visible at the
border of L2 and L3 cache (at 3MB).
Asymptotically, storing the unmodified count-min sketch on SSD or a disk is inefficient, given
that each estimate and update operation needs r hashes, which results in O(r) random reads/writes
to SSD, far below the desired throughput for most time-critical streaming applications.
Another context where we see CMS becoming large even when is fixed is in some text
applications,where the number of elements inserted in the sketch is quadratic in the original text
size. For instance, [19] uses CMS to record distributional similarity on the web, where each pair
of words is inserted as a single item into the CMS, and 90GB of text requires a CMS of 8GB.
1.1 Results
1. We describe the design and implementation of buffered count-min sketch, and empirically
show that our implementation achieves 3.7-4.7x the speedup on update (insert) and 4.3x
speedup on estimate (lookup) operations.
2. Our design also offers an asymptotic improvement in the external-memory model [1] over the
original data structure: O(r) random I/Os are reduced to 1 I/O for estimate. For a data
structure that uses k blocks on SSD, w as the word/counter size, r as the number of rows,
M as the number of bits in main memory, our data structure uses kwr/M amortized I/Os for
updates, or, if kwr/M >1, 1 I/O in the worst case. In typical scenarios, kwr/M <<1. This
is in contrast to O(r) I/Os incurred for each update in the original data structure.
3. We mathematically show that for buffered count-min sketch, the error rate does not substan-
tially degrade over the original count-min sketch. Specifically, we prove that for any query
q, our data structure provides the following guarantee:
Pr[Error(q)≥n(1+o(1))]≤δ+o(1).
We focus on scenarios where the allowed estimation error is sublinear in n. For example, what
√
if we want the estimation error to be no larger than n/logn, or
n? These scenarios correspond
√
to = 1/logn or 1/
n, and now for even moderately large values of n, the count-min sketch
becomes too large to fit in main memory. Even given more modest condition, such as =o(1/M),
where the memory is of size M, the count-min sketch is unlikely to fit in memory. We will assume
ESA 2018
XX:4
Buffered Count-Min Sketch on SSD: Theory and Experiments
that 1/n≤ <<1/M. Higher values of do not require the count-min sketch to be placed on
disk, and lower values of mean exact counts are desired.
2
Related Work
The streaming model represents many real-life situations where the data is produced rapidly
and on a constant basis. Some of the applications include sensor networks [22], monitoring web
traffic [26], analyzing text [19], and monitoring satellites orbiting the Earth [18].
Heavy hitters, top k queries, iceberg queries, and quantiles [3,22,28] are some of the most
central problems in the streaming context, where we wish to extract the general trends from a
massive dataset. Count-Min sketch has proved useful in such contexts for its space-efficiency and
providing accurate counts [12,20].
Count-Min sketch can be well illustrated using its connection to the Bloom filter [6,7,9]. Both
data structures are lossy and space-efficient representations of sets, used to reduce disk accesses in
time-critical applications. Bloom filter answers membership queries and can have false positives,
while Count-Min sketch answers frequency queries, and can overestimate the actual frequency
count. Both data structures are hashing-based, and suffer from similar issues when placed directly
to SSD or a magnetic disk.
There has been earlier attempts to scale Bloom filters to SSD using buffering and hash
localization [10,13]. Our paper employs similar methods to those in [10,13]. The improvement,
both in our case and in the case of Buffered Bloom filter [10] is achieved at the expense of having
an extra hash function that helps determine to which page each element is going to hash.
There has also been work in designing cache-efficient equivalents for Bloom filters such as quo-
tient filter and write-optimized on-disk quotient filter such as Cascade filter (CQF) [5,14,23]. An im-
portant distinction to make between these data structures and count-min sketch is that CQF gives
exact counts of most of the elements given that the errors caused by false positives are usually very
small. However, since the errors are independent, the CQF doesn't offer any guarantees on the over-
estimate. For example, two highly occurring elements in a multi-set can collide with each other and
both will have large overcounts. On the other hand, the CMS does not give exact counts of elements
due to multiple hashes and its size (width of the CMS is smaller than the number of slots in a CQF).
But the CMS can offer a guarantee that overestimate will be smaller than n with a probability of δ.
2.1 Count-Min Sketch: Preliminaries
In the streaming model, we are given a stream A of pairs (ai,ci), where ai denotes the item
identifier (e.g., IP address, stock ID, product ID), and ci denotes the count of the item (e.g., the
number of bytes sent from the IP address, the amount by which a stock has risen/fallen or the
number of sold items). Each pair Xi =(ai,ci) is an item within a stream of length T, and the
goal is to record total sum of frequencies for each particular item ai.
For a given estimation error rate and failure probabiltity δ, define r=ln(1/δ) and c=e/.
The Count-Min Sketch is represented via 2D table with c buckets (columns), r rows, implemented
using r hash functions (one hash function per row).
CMS has two operations: UPDATE(ai)and ESTIMATE(ai), the respective equivalents of in-
sert and lookup, and they are performed as follows:
UPDATE(ai)inserts the pair by computing r hash functions on ai, and incrementing appropriate
slots determined by the hashes by the quantity ci. That is, for each hash function hj, 1≤j≤r,
we set CMS[j][hj(ai)]=CMS[j][hj(ai)]+ci. Note that in this paper, we use ci=1, so every
time an item is updated, it is just incremented by 1.
M. Goswami, D. Medjedovic, E. Mekic, and P. Pandey
XX:5
ESTIMATE(ai)reports the frequency of ai which can be an overestimate of the true frequency.
It does so by calculating r hashes and taking the minimum of the values found in appropriate
cells. In other words, we return min1≤j≤r(CMS[j][hj(ai)]). Because different elements can
hash to the same cells, the count-min sketch can return the overestimated (never underestim-
ated) value of the count, but in order for this to happen, a collision needs to occur in each row.
The estimation error is bounded; the data structure guarantees that for any particular item,
the error is within the range n, with probability at least 1−δ, i.e., P r[Error(q)≥n]≤δ.
3
Buffered Count-Min Sketch
In this section, we describe Buffered Count-Min Sketch, an adaptation of CMS to SSD. The
traditional CMS, when placed on external storage, exhibits performance issues due to random-write
nature of hashing. Each update operation in CMS requires c=ln(1/δ) writes to different rows and
columns of CMS. On a large data structure, these writes become destined to different pages on
disk, causing the update to perform O(ln(1/δ)) random SSD page writes. For high-precision CMS
scenarios where δ=0.001%−0.01%, this can be between 5-7 writes to SSD, which is unacceptable
in a high-throughput scenario.
To solve this problem, we implement, analyze and empirically test the data structure presented
in [15] that outlines three adaptations to the original data structure:
1. Partitioning CMS into pages and column-first layout: We logically divide the CMS on SSD
into pages of block size B. CMS with r rows, c columns, cell size w, and a total of S =cr w-bit
counters, contains k pages P1,P2,P3,...,Pk, where k = S/B and each page spans contiguous
B/r columns: Pi spans columns [B(i−1)/r+1,Bi/r]. To improve cache-efficiency, CMS is laid
out on disk in column-first fashion, which allows each logical page to be laid out sequentially
in memory. Thus, each read/write of a logical page requires at most 2 I/Os.
2. Hash localization: We direct all hashes of each element to a single logical page of CMS that is
determined by an additional hash function h0:[1,k]. The subsequent r hash functions map to
the columns inside the corresponding logical page, i.e., the range of h1,h2,...,hr for an element
e is [B(h0(e)−1)/r+1,Bh0(e)/r]. This way, we direct all updates and reads related to one
element to one logical page.
3. Buffering: When an update operation occurs, the hashes produced for an element are first
stored inside an in-memory buffer. The buffer is partitioned into sub-buffers of equal size
S1,S2,...,Sk, and they directly correspond to logical pages on disk in that Si stores the hashes
for updates destined for page Pi. Each element first hashes using h0, which determines in
which sub-buffer the hashes will be temporarily stored for this element. Once the sub-buffer
Si becomes full, we read the page Pi from the CMS, apply all updates destined for that page,
and write it back to disk. The capacity of a sub-buffer is M/k hashes, which is equivalent
to M/kwr elements so the cost of an update becomes kwr/M <<1 I/O.
The pseudocode for UPDATE(ai)is shown in Algorithm 1, and for ESTIMATE(ai)in Algorithm 2.
We use murmurhash as our hashing algorithm due to its efficiency and simplicity [2]. Unlike
UPDATE(ai), ESTIMATE(ai)operation is not buffered. In a related work [10] that implements a
buffered Bloom filter on SSD, the data structure buffers lookups. However in the count-min
sketch scenario, buffering for ESTIMATE(ai)is unproductive given that even if the item is found
in the buffer, we still need to check the CMS page to obtain the correct count. Therefore,
our ESTIMATE(ai) is optimized for the worst-case single lookup scenario and works for solely
insert/lookup as well as mixed workloads. The ESTIMATE(ai) also first computes the correct
sub-buffer using h0, and flushes the corresponding sub-buffer to SSD page in case some updates
ESA 2018
XX:6
Buffered Count-Min Sketch on SSD: Theory and Experiments
Figure 2 UPDATE operation on Buffered Count-Min Sketch. Updates are stored in RAM, and all
updates are destined for the same block on disk.
Algorithm 1 Buffered Count-Min Sketch - UPDATE function
Require : key , r
subbufferIndex i := murmur0( key );
for i :=1 to r do
hashes [i] := murmur i( key );
end for
AppendToBuffer ( hashes , subbufferIndex );
if isSubbufferFull ( subbufferIndex ) then
bcmsBlock := readDiskPage ( subbufferIndex );
for each entry in Subbuffer [ subbufferIndex ] do
for each index in entry do
pageStart := calculatePageStart ( subbufferIndex );
offset := pageStart + entry [ index ];
bcmsBlock [ offset ][ index ]++;
end for
end for
writeBcmsPageBackToDisk ( bcmsBlock );
clearBuffer ( subbufferIndex );
end if
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
were present. Once it applies the necessary changes to the page, it reads the corresponding CMS
cells specified by r hashes and returns the minimum value.
4
Analysis of Buffered Count-Min Sketch
In this section, we show that the buffering and hash localization do not substantially degrade the
error guarantee of the buffered count-min data structure. Fix a failure probability 0<δ <1 and
let 0<(n)<1 be the function of n controlling the estimation error. Let r=ln(1/δ) and c=e/.
The traditional count-min sketch uses S =rc=(e/)ln(1/δ) counters/words of space. Recall that
for our purposes, 1/n≤(n)<<1/M.
Let k=S/B be the number of blocks occupied by the buffered count-min sketch. We assume
a block can hold B counters. Our analysis will assume the following mild conditions:
Assumption 1: n is sufficiently larger than the number of blocks k, n=ω(k(logk)3) suffices. Since
k depends inversely on (n), this assumption essentially means that (n)=ω(1/n). Assumption
M. Goswami, D. Medjedovic, E. Mekic, and P. Pandey
XX:7
Algorithm 2 Buffered Count-Min Sketch - ESTIMATE function
Require : key , k
subbufferIndex i := murmur0( key );
pageStart := calculatePageStart ( subbufferIndex );
bcmsBlock := readDiskPage ( subbufferIndex );
if isSubbufferNotEmpty ( subbufferIndex ) then
for each entry in Subbuffer [ subbufferIndex ] do
for each index in entry do
offset := pageStart + entry [ index ];
bcmsBlock [ offset ][ index ]++;
end for
end for
clearBuffer ( subbufferIndex );
end if
for i :=1 to k do
value := murmur i( key );
offset := pageStart + value ;
estimation := bcmsBlock [ offset ][i - 1];
estimates [i] := estimation ;
end for
writeBcmsPageBackToDisk ( bcmsBlock );
return min ( estimates )
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
2: limn→∞(n)=0.
Both conditions are satisfied, e.g., when (n)=1/logn or 1/nc for any c<1.
For brevity, we will drop the dependence of (n) on n, and write the error rate as just ,
however it is important to note that is not a constant.
(cid:73) Theorem 1. The Buffered-Count-Min-Sketch is a data structure that uses k blocks of space
on disk and for any query q,
returns ESTIMATE(q) in 1 I/O and performs updates in kwr/M I/Os amortized, or, if
kwr/M >1, in one I/O worst case.
Let Error(q) = ESTIMATE(q) - TrueFrequency(q). Then for any C≥1,
P r[Error(q)≥n(1+p(2(C+1)klogk)/n)]≤δ+O((B/e)C).
Remark: By assumption 1,p(2(C+1)klogk)/n is o(1) (in fact, it is o(1/logk)). By assumption 2,
(B/e)C is o(1). Thus we claim that the buffered count-min-sketch gives almost the same guarantees
as a traditional count-min sketch, while obtaining a factor r speedup in queries.The guarantee
for estimates taking 1 I/O is apparent from construction, as only one block needs to be loaded2.
The proof is a combination of the classical analysis of CMS and the maximum load of balls
in bins when the number of bins is much smaller than the number of balls. Also, note that unlike
the traditional CMS, the errors for a query q in different rows are no longer independent (in fact,
they are positively correlated: a high error in one row implies more elements were hashed by h0
to the same bucket as q).
2 In practice, we may need 2 I/Os sometimes due to block-page alignment, but never more than 2
ESA 2018
XX:8
Buffered Count-Min Sketch on SSD: Theory and Experiments
The hash function h0 maps into k buckets, each having size B (and so we will also call
them blocks). Each bucket can be thought of as a r×B/r matrix. Note that r =ln(1/δ), and
B/r = e/(k). We assume that h0 is a perfectly random hash function, and, abusing notation,
identify a bucket/block with a bin, where h0 assigns elements (balls) to one of the k buckets (bins).
In this scenario we use Lemma 2(b) from [24] and adapt it to our setting.
(cid:73) Lemma 2.
q=1−p. If t=np+o((pqn)2/3) and x:= t−np√
[24] Let B(n,p) denote a Binomial distribution with parameters n and p, and
pqn tends to infinity, then
P r[B(n,p)≥t]=e−x2/2−logx− 1
Let M(n,k) denote the maximum number of elements that fall into a bucket, when hashed by h0.
2 logπ+o(1).
q
(cid:73) Lemma 3. Let C≥1 and t=n/k+
P r[M(n,k)≤t]≥1−1/kC.
2(C+1) nlogk
k . Then
q
Proof. We first check that t satisfies the conditions of Lemma 2. Since h0 is uniform, p=1/k
(i.e., each bucket is equally probable), and np= n/k. We need to check that the extra term in
is o((n(1−1/k)/k)2/3). This is precisely the condition that n = ω(k(logk)3)
t,
(assumption 1).
2(C+1) nlogk
k
s
Next we apply Lemma 2. In our case,
=p2(C+1)logk(1+1/k−1),
x=
2(C+1)nlogk/k
n(1−1/k)/k
2 logπ+o(1)≤e−x2/2. Putting in x=q2(C+1)logk(1+ 1
Now by assumption 2, (n) goes to zero as n goes to infinity, and so k ∝ 1/(n) goes to
infinity, and therefore x goes to infinity as n goes to infinity. Thus we have that the number of
elements in any particular bucket (which follows a B(n,1/k) distribution) is larger than t with
probability e−x2/2−logx− 1
k−1), we get x2/2=
(C+1)logk(1+1/(k−1)≥(C+1)logk, and thus the probability is at most e−(C+1)logk =1/kC+1.
Thus the probability that the maximum number of balls in a bin is more than t is bounded
(cid:74)
Now that we know that with probability as least 1−1/kC, no bucket has more than t elements,
we observe that a bucket serves as a "mini" CMS for the elements that hash to it. In other words,
let n(q) be the number of elements that hash to the same bucket as q under h0. The expected
error in the ith row of the mini-CMS for q (the entry for which is contained inside the bucket
of q), is E[Errori(q)]=n(q)/(B/r)=n(q)k/e.
(by the union bound) by k.1/kC+1=1/kC, and the lemma is proved.
By Markov's inequality Pr[Errori(q)≥n(q)k]≤1/e.
Let α= tk/e=(n/k+p(2(C+1)nlogk)/k)k/e=(n/e)(1+p(2(C+1)klogk)/n). We now
compute the bound on the final error (after taking the min) as follows.
P r(Error(q)≥eα) = P r(Errori(q)≥eα ∀i∈{1,···,r})
(cid:19)r
(cid:18)1
= P r(Errori(q)≥eα ∀i n(q)≤t)P r(n(q)≤t)
+ P r(Errori(q)≥eα ∀i n(q)≥t)P r(n(q)≥t)
≤
= δ+1/kC,
1+1(1/kC)
e
M. Goswami, D. Medjedovic, E. Mekic, and P. Pandey
XX:9
where the second last equality follows from Markov's inequality on Errori(q) and Lemma 3.
Finally, by observing that for a fixed δ, k=O(e/B), the proof of the theorem is complete.
5
Evaluation
In this section, we evaluate our implementation of the buffered count-min sketch. We compare
the buffered count-min sketch against the (traditional) count-min sketch. We evaluate each data
structure on two fundamental operations, insertions and queries. We evaluate queries for set of
elements chosen uniformly at random.
In our evaluation, we address the following questions about how the performance of buffered
count-min sketch compares to the count-min sketch:
How does the insertion throughput in buffered count-min sketch compare to count-min sketch
on SSD?
How does the query throughput in buffered count-min sketch compare to count-min sketch
on SSD?
How does the hash localization in buffered count-min sketch affect the overestimates compared
to the overestimates in count-min sketch?
5.1 Experimental setup
To answer the above questions, we evaluate the performance of the buffered count-min sketch and
the count-min sketch on SSD by scaling the sketch out-of-RAM. For SSD benchmarks, we use
four different RAM-size-to-sketch-size ratios, 2, 4, 8, and 16. The RAM-size-to-sketch-size ratio
is the ratio of the size of the available RAM and the size of the sketch on SSD. We fix the size
of the available RAM to be ≈64MB and change the size of the sketch to change the RAM-size-to-
sketch-size ratio. The page size in all our benchmarks was set to 4096B. In all the benchmarks, we
measure the throughput (operations per second) to evaluate the insertion and query performance.
To measure the insertion throughput, we first calculate the number of elements we can insert
in the sketch using calculations described in Section 5.2. During an insert operation, we first
generate a 64-bit integer from a uniform-random distribution and then add that integer to the
sketch. This way, we do not use any extra memory to store the set of integers to be added to
the sketch. We then measure the total time taken to insert the given set of elements in the sketch.
Note that for the buffered count-min sketch, we make sure to flush all the remaining inserts from
the buffer to the sketch on SSD at the end and include the time to do that in the total time.
To measure the query throughput, we query for elements drawn from a uniform-random
distribution and measure the throughput. The reason for the query benchmark is to simulate
a real-world query workload where some elements may not be present in the sketch and the query
will terminate early thereby requiring fewer I/Os.
For all the query benchmarks, we first perform the insertion benchmark and write the sketch
to SSD. After the insertion benchmark, we flush all caches (page cache, directory entries, and
inodes). We then map the sketch back into RAM and perform queries on the sketch. This way
we make sure that the sketch is not already cached in kernel caches from the insertion benchmark.
We compare the overestimates in buffered count-min sketch and count-min sketch for all
the four sketch sizes for which we perform insertion and query benchmarks. To measure the
overestimates, we first perform the insertion benchmark. However, during the insertion benchmark,
we also store each inserted element in a multiset. Once insertions are done, we iterate over the
multiset and query for each element in the sketch. We then take the difference of the count
returned from the sketch and the actual count of the element to calculate the overestimate.
ESA 2018
XX:10
Buffered Count-Min Sketch on SSD: Theory and Experiments
Size Width Depth #elements
3355444
128MB
256MB
6710887
512MB 13421773
1GB
26843546
9875188
19750377
39500754
79001508
5
5
5
5
Table 1 Size, width, and depth of the sketch and the number of elements inserted in count-min
sketch and buffered count-min sketch in our benchmarks (insertion, query, and overestimate calculation).
Figure 3 Insert throughput of count-min sketch and buffered count-min sketch with increas-
ing sizes. The available RAM is fixed to ≈ 64MB. With increasing sketch sizes (on x-axis) the
RAM-size-to-sketch-size is also increasing 2, 4, 8, and 16. (Higher is better)
For SSD-based experiments, we allocate space for the sketch by mmap-ing it to a file on SSD.
We then control the available RAM to the benchmarking process using cgroups. We fix the RAM
size for all the experiments to be ≈67MB. We then increase the size of the sketch based on the
RAM-size-to-sketch-size ratio of the particular experiment. For the buffered count-min sketch, we
use all the available RAM as the buffer. Paging is handled by the operating system based on the
disk accesses. The point of these experiments is to evaluate the I/O efficiency of sketch operations.
All benchmarks were performed on a 64-bit Ubuntu 16.04 running Linux kernel 4.4.0-98-generic.
The machine has Intel Skylake CPU U (Core(TM) i7-6700HQ CPU @ 2.60GHz with 4 cores and
6MB L3 cache) with 32 GB RAM and 1TB Toshiba SSD.
5.2 Configuring the sketch
In our benchmarks, we take as input δ, overestimate O (n), and the size of the sketch to
configure the sketch S. The depth of the count-min sketch D is dln 1
δe. The number of cells C
is S/CELL_SIZE. And width of the count-min sketch is de/e.
Given these parameters, we calculate the number of elements n to be inserted in the sketch
as C×O
D×e . In all our experiments, we set δ to 0.01 and maximum overestimate to 8 and change
the sketch size. Table 1 shows dimensions of the sketch and number of elements inserted based
on the size of the sketch.
M. Goswami, D. Medjedovic, E. Mekic, and P. Pandey
XX:11
Figure 4 Query throughput of count-min sketch and buffered count-min sketch with increas-
ing sizes. The available RAM is fixed to ≈ 64MB. With increasing sketch sizes (on x-axis) the
RAM-size-to-sketch-size is also increasing 2, 4, 8, and 16. (Higher is better)
Figure 5 Maximum overestimate reported by count-min sketch and buffered count-min sketch for any
inserted element for different sketch sizes. The blue line represents the average overestimate reported by
count-min sketch and buffered count-min sketch for all the inserted elements. The average overestimate
is same for both count-min sketch and buffered count-min sketch.
5.3 Insert Performance
Figure 3 shows the insert throughput of count-min sketch and buffered count-min sketch with
changing RAM-size-to-sketch-size ratios. buffered count-min sketch is 3.7×–4.7× faster compared
to the count-min sketch in terms of insert throughput on SSD.
The buffered count-min sketch performs less than one I/O per insert operation because all the
hashes for a given element are localized to a single page on SSD. However, in the count-min sketch
the hashes for a given element are spread across the whole sketch. Therefore, the insert throughput
of the buffered count-min sketch is 3.7× when the sketch is twice the size of the RAM. And the
difference in the throughput increases as the sketch gets bigger and RAM size stays the same.
5.4 Query Performance
Figure 4 shows the query throughput of count-min sketch and buffered count-min sketch with
changing RAM-size-to-sketch-size ratios. buffered count-min sketch is ≈4.3× faster compared
to the count-min sketch in terms of query throughput on SSD.
ESA 2018
XX:12
Buffered Count-Min Sketch on SSD: Theory and Experiments
The buffered count-min sketch performs a single I/O per query operation because all the hashes
for a given element are localized to a single page on SSD. In comparison, count-min sketch may have
to perform as many as h I/Os per query operation, where h is the depth of the count-min sketch.
5.5 Overestimates
In Figure 5 we empirically compare overestimates returned by the count-min sketch and buffered
count-min sketch for all the four sketch sizes for which we performed insert and query benchmarks.
And we found that the average and the maximum overestimate returned from count-min sketch
and buffered count-min sketch are exactly the same. This shows that empirically hash localization
in buffered count-min sketch does not have any major effect on the overestimates.
6
Conclusion
In this paper we described the design and implementation of the Buffered count-min sketch, and
empirically showed that our implementation achieves 3.7×–4.7× the speedup on update (insert)
and 4.3× speedup on estimate (lookup) operations. Queries take 1 I/O, which is optimal in the
worst case if not allowed to buffer. However, we do not know whether the update time is optimal.
To the best of our knowledge, no lower bounds on the update time of such a data structure are
known (the only known upper bounds are on space, e.g., in [17]). We leave the question of deriving
update lower bounds and/or a SSD-based data structure with faster update time for future work.
References
1 Alok Aggarwal and S. Vitter, Jeffrey.
and related problems.
http://doi.acm.org/10.1145/48529.48535, doi:10.1145/48529.48535.
The input/output complexity of
Commun. ACM, 31(9):1116–1127, September 1988.
sorting
URL:
2 Austin Appleby. 32-bit variant of murmurhash3, 2011. URL: https://sites.google.com/
site/murmurhash/.
Brian Babcock, Shivnath Babu, Mayur Datar, Rajeev Motwani, and Jennifer Widom. Models
and issues in data stream systems.
In Proceedings of the Twenty-first ACM SIGMOD-
SIGACT-SIGART Symposium on Principles of Database Systems, PODS '02, pages 1–16,
New York, NY, USA, 2002. ACM. URL: http://doi.acm.org/10.1145/543613.543615,
doi:10.1145/543613.543615.
Brian Babcock, Shivnath Babu, Mayur Datar, Rajeev Motwani, and Jennifer Widom. Models
and issues in data stream systems. In Proceedings of the twenty-first ACM SIGMOD-SIGACT-
SIGART symposium on Principles of database systems, pages 1–16. ACM, 2002.
3
4
6
5 Michael A. Bender, Martin Farach-Colton, Rob Johnson, Russell Kraner, Bradley C. Kuszmaul,
Dzejla Medjedovic, Pablo Montes, Pradeep Shetty, Richard P. Spillane, and Erez Zadok. Don't
thrash: How to cache your hash on flash. Proc. VLDB Endow., 5(11):1627–1637, July 2012.
URL: http://dx.doi.org/10.14778/2350229.2350275, doi:10.14778/2350229.2350275.
Burton H. Bloom. Space/time trade-offs in hash coding with allowable errors. Commun.
ACM, 13(7):422–426, July 1970.
URL: http://doi.acm.org/10.1145/362686.362692,
doi:10.1145/362686.362692.
Flavio Bonomi, Michael Mitzenmacher, Rina Panigrahy, Sushil Singh, and George Var-
ghese. An improved construction for counting bloom filters.
In Proceedings of the 14th
Conference on Annual European Symposium - Volume 14, ESA'06, pages 684–695, Lon-
don, UK, UK, 2006. Springer-Verlag. URL: http://dx.doi.org/10.1007/11841036_61,
doi:10.1007/11841036_61.
7
M. Goswami, D. Medjedovic, E. Mekic, and P. Pandey
XX:13
8
Lee Breslau, Pei Cao, Li Fan, Graham Phillips, and Scott Shenker. Web caching and
zipf-like distributions: Evidence and implications. In INFOCOM '99. Eighteenth Annual Joint
Conference of the IEEE Computer and Communications Societies, pages 126–134, 1999.
9 Andrei Broder, Michael Mitzenmacher, and Andrei Broder I Michael Mitzenmacher. Network
applications of bloom filters: A survey. In Internet Mathematics, pages 636–646, 2002.
13
10 Mustafa Canim, George A. Mihaila, Bishwaranjan Bhattacharjee, Christian A. Lang,
and Kenneth A. Ross.
In Rajesh
Bordawekar and Christian A. Lang, editors, ADMS@VLDB, pages 1–8, 2010. URL:
http://dblp.uni-trier.de/db/conf/vldb/adms2010.html#CanimMBLR10.
Buffered bloom filters on solid state storage.
11 Aiyou Chen, Yu Jin, Jin Cao, and Li Erran Li.
Tracking long duration flows in
network traffic.
the 29th Conference on Information Communica-
tions, INFOCOM'10, pages 206–210, Piscataway, NJ, USA, 2010. IEEE Press. URL:
http://dl.acm.org/citation.cfm?id=1833515.1833557.
In Proceedings of
12 Graham Cormode and S. Muthukrishnan. An improved data stream summary: The
count-min sketch and its applications. J. Algorithms, 55(1):58–75, April 2005. URL: http:
//dx.doi.org/10.1016/j.jalgor.2003.12.001, doi:10.1016/j.jalgor.2003.12.001.
Biplob Debnath, Sudipta Sengupta, Jin Li, David J. Lilja, and David H. C. Du. Bloom-
flash: Bloom filter on flash-based storage.
In Proceedings of the 2011 31st International
Conference on Distributed Computing Systems, ICDCS '11, pages 635–644, Washington, DC,
USA, 2011. IEEE Computer Society. URL: http://dx.doi.org/10.1109/ICDCS.2011.44,
doi:10.1109/ICDCS.2011.44.
Sourav Dutta, Ankur Narang, and Suman K. Bera.
Streaming quotient filter: A near
optimal approximate duplicate detection approach for data streams. Proc. VLDB En-
dow., 6(8):589–600, June 2013. URL: http://dx.doi.org/10.14778/2536354.2536359,
doi:10.14778/2536354.2536359.
Ehsan Eydi, Dzejla Medjedovic, Emina Mekic, and Elmedin Selmanovic. Buffered count-min
sketch. In Mirsad Hadžikadić and Samir Avdaković, editors, Advanced Technologies, Systems,
and Applications II, pages 249–255, Cham, 2018. Springer International Publishing.
and Shonali Krishnaswamy.
ing data streams:
34(2):18–26, June 2005.
http://doi.acm.org/10.1145/1083784.1083789, doi:10.1145/1083784.1083789.
Sumit Ganguly. Lower bounds on frequency estimation of data streams.
Computer Science Symposium in Russia, pages 204–215. Springer, 2008.
16 Mohamed Medhat Gaber, Arkady Zaslavsky,
SIGMOD Rec.,
In International
A review.
Min-
URL:
15
17
14
18 Michael Gertz, Quinn Hart, Carlos Rueda, Shefali Singhal, and Jie Zhang. A data and
query model for streaming geospatial image data. In Torsten Grust, Hagen Höpfner, Arantza
Illarramendi, Stefan Jablonski, Marco Mesiti, Sascha Müller, Paula-Lavinia Patranjan, Kai-Uwe
Sattler, Myra Spiliopoulou, and Jef Wijsen, editors, Current Trends in Database Technology –
EDBT 2006, pages 687–699, Berlin, Heidelberg, 2006. Springer Berlin Heidelberg.
19 Amit Goyal, Jagadeesh Jagarlamudi, Hal Daumé, III, and Suresh Venkatasubramanian.
Sketch techniques for scaling distributional similarity to the web.
In Proceedings of the
2010 Workshop on GEometrical Models of Natural Language Semantics, GEMS '10, pages
51–56, Stroudsburg, PA, USA, 2010. Association for Computational Linguistics. URL:
http://dl.acm.org/citation.cfm?id=1870516.1870524.
20 Gurmeet Singh Manku and Rajeev Motwani.
over data streams.
Large Data Bases, VLDB '02, pages 346–357. VLDB Endowment, 2002.
http://dl.acm.org/citation.cfm?id=1287369.1287400.
frequency counts
the 28th International Conference on Very
URL:
In Proceedings of
Approximate
21 Muthu Muthukrishnan and Graham Cormode. Approximating data with the count-min sketch.
volume 29, pages 64–69, Los Alamitos, CA, USA, 10 2011. IEEE Computer Society Press.
URL: doi.ieeecomputersociety.org/10.1109/MS.2011.127, doi:10.1109/MS.2011.127.
ESA 2018
XX:14
Buffered Count-Min Sketch on SSD: Theory and Experiments
22
23
Suman Nath, Phillip B. Gibbons, Srinivasan Seshan, and Zachary R. Anderson. Synopsis
diffusion for robust aggregation in sensor networks.
In Proceedings of the 2Nd Interna-
tional Conference on Embedded Networked Sensor Systems, SenSys '04, pages 250–262, New
York, NY, USA, 2004. ACM.
URL: http://doi.acm.org/10.1145/1031495.1031525,
doi:10.1145/1031495.1031525.
Prashant Pandey, Michael A. Bender, Rob Johnson, and Robert Patro. A general-purpose
counting filter: Making every bit count.
In Proceedings of the 2017 ACM International
Conference on Management of Data, SIGMOD Conference 2017, Chicago, IL, USA, May
14-19, 2017, pages 775–787, 2017. URL: http://doi.acm.org/10.1145/3035918.3035963,
doi:10.1145/3035918.3035963.
24 Martin Raab and Angelika Steger. "balls into bins"-a simple and tight analysis. Randomization
and Approximation Techniques in Computer Science, pages 159–170, 1998.
25 Tim Roughgarden and Gregory Valiant. Cs168: The modern algorithmic toolbox lecture #2:
Approximate heavy hitters and the count-min sketch, 2018.
26 Tamás Sarlós, Adrás A. Benczúr, Károly Csalogány, Dániel Fogaras, and Balázs Rácz. To
randomize or not to randomize: Space optimal summaries for hyperlink analysis. In Proceedings
of the 15th International Conference on World Wide Web, WWW '06, pages 297–306, New
York, NY, USA, 2006. ACM.
URL: http://doi.acm.org/10.1145/1135777.1135823,
doi:10.1145/1135777.1135823.
Stuart Schechter, Cormac Herley, and Michael Mitzenmacher. Popularity is everything: A new
approach to protecting passwords from statistical-guessing attacks. In Proceedings of the 5th
USENIX Conference on Hot Topics in Security, HotSec'10, pages 1–8, Berkeley, CA, USA,
2010. USENIX Association. URL: http://dl.acm.org/citation.cfm?id=1924931.1924935.
28 David P. Woodruff. New algorithms for heavy hitters in data streams. CoRR, abs/1603.01733,
27
2016. URL: http://arxiv.org/abs/1603.01733.
29 Yin Zhang, Sumeet Singh, Subhabrata Sen, Nick Duffield, and Carsten Lund. Online identi-
fication of hierarchical heavy hitters: Algorithms, evaluation, and applications. In Proceedings
of the 4th ACM SIGCOMM Conference on Internet Measurement, IMC '04, pages 101–114,
New York, NY, USA, 2004. ACM. URL: http://doi.acm.org/10.1145/1028788.1028802,
doi:10.1145/1028788.1028802.
30 Qi (George) Zhao, Mitsunori Ogihara, Haixun Wang, and Jun (Jim) Xu. Finding global
icebergs over distributed data sets.
In Proceedings of the Twenty-fifth ACM SIGMOD-
SIGACT-SIGART Symposium on Principles of Database Systems, PODS '06, pages 298–307,
New York, NY, USA, 2006. ACM. URL: http://doi.acm.org/10.1145/1142351.1142394,
doi:10.1145/1142351.1142394.
|
1006.1117 | 1 | 1006 | 2010-06-06T17:03:09 | On the hardness of distance oracle for sparse graph | [
"cs.DS"
] | In this paper we show that set-intersection is harder than distance oracle on sparse graphs. Given a collection of total size n which consists of m sets drawn from universe U, the set-intersection problem is to build a data structure which can answer whether two sets have any intersection. A distance oracle is a data structure which can answer distance queries on a given graph. We show that if one can build distance oracle for sparse graph G=(V,E), which requires s(|V|,|E|) space and answers a (2-\epsilon,c)-approximate distance query in time t(|V|,|E|) where (2-\epsilon) is a multiplicative error and c is a constant additive error, then, set-intersection can be solved in t(m+|U|,n) time using s(m+|U|,n) space. | cs.DS | cs |
On the hardness of distance oracle for sparse
graph
Hagai Cohen and Ely Porat⋆
Department of Computer Science, Bar-Ilan University, 52900 Ramat-Gan, Israel
{cohenh5,porately}@cs.biu.ac.il
Abstract. In this paper we show that set-intersection is harder than
distance oracle on sparse graphs. Given a collection of total size n which
consists of m sets drawn from universe U , the set-intersection problem
is to build a data structure which can answer whether two sets have
any intersection. A distance oracle is a data structure which can answer
distance queries on a given graph. We show that if one can build distance
oracle for sparse graph G = (V, E), which requires s(V , E) space and
answers a (2 − ǫ, c)-approximate distance query in time t(V , E) where
2 − ǫ is a multiplicative error and c is a constant additive error, then,
set-intersection can be solved in t(m + U , n) time using s(m + U , n)
space.
1
Introduction and Related Work
Let G = (V, E) be a graph. The all-pairs shortest paths problem (APSP) requires
to construct a data structure for a given graph G so that the exact distance be-
tween every two vertices on that graph can be retrieved efficiently. This problem
is one of the most fundamental graph problems of computer science. Despite the
importance of this problem, there is still no efficient solution for it using less
than O(V 2) space. When the graph is dense, i.e., when E = O(V 2), this
space is not much. But, for sparse graphs where E = O(V ) this is extremely
a lot of space.
Thorup and Zwick [1] explored an alternative for the APSP problem. They
introduced a solution for the approximate distance oracle, which is a data struc-
ture that answers approximate distance queries in a graph. They achieved that,
for any integer k ≥ 1, an undirected weighted graph with n vertices and m edges
can be preprocessed in expected O(kmn1/k) time to construct a data structure
of size O(kn1+1/k) that can answer any (2k − 1)-approximate distance query in
O(k) time. This means that the distance oracle answers distance queries with
multiplicative error of 2k − 1.
In this paper we show by a reduction from the set intersection problem that
it is hard to build a (2 − ǫ, c)-approximate distance oracle where 2 − ǫ is a
multiplicative error and c is a constant additive error.
⋆ This work was supported by BSF and ISF
In the set intersection problem, we are given a collection of sets which we
can preprocess. Then, given two sets we need to answer quickly whether there
is any intersection between the sets. This is a common problem in many fields,
especially in retrieval algorithms and search engines. The formal definition of
the problem is as follows:
Definition 1. Let D be a database consisting of a collection of m sets drawn
from universe U , S1, . . . , Sm ⊆ U . Denote n to be the input size, i.e., n =
Pm
i=1 Si. The set intersection problem is to build a data structure that given a
query of two indices i, j ≤ m, can answer if sets Si and Sj have any intersection.
Cohen and Porat [2] showed how the set intersection problem can be solved
in O(√n) query time using O(n) space. Their solution is based on dividing the
sets in the database D to large and non-large sets, where they define a large set
to be a set which has more than √n elements. They construct a set intersection
matrix for the large sets in D, which is a matrix saving for each pair of sets if
there is any intersection between them. They showed that the number of large
sets is at most √n, thus, this matrix costs √n×√n = O(n) bits space. Moreover,
for each set in D they store a static hash table to retrieve in O(1) time if an
element belongs to that set or not.
Given a query consisting of two indices i, j, if both Si and Sj are large
sets, the answer can be retrieved from the set intersection matrix in O(1) time.
Otherwise, one of the sets is a non-large set, i.e., it has less than √n elements.
On this case, the answer can be retrieved by going over all the elements of the
smaller set, checking for each one of them if it belongs to the other set in O(1)
time. Because non-large sets have at most O(√n) elements, this takes at most
O(√n) time.
This solution can be easily extended to a tunable solution. If we define a
large set to be a set with more than t elements, the number of large sets can be
at most n
t2 ) space. Hence, this
problem can be answered in O(t) query time using O( n2
t sets. Thus, the set intersection matrix costs O( n2
t2 ) space.
In this paper we show a reduction from the set intersection problem to dis-
tance oracle on sparse graphs. In Sect. 2 we show that if one can build a dis-
tance oracle using s(V ,E) space with t(V ,E) query time, which answers
(2 − ǫ)-approximate distance queries, the set intersection problem can be solved
in t(m + U, n) query time using s(m + U, n) space. In Sect. 3 we extend the
reduction to a (2 − ǫ)-approximate distance oracle with constant additive error.
2 Set Intersection Reduction
In the next theorem we claim that if one can build a distance oracle that answers
(2− ǫ)-approximate distance queries, the set intersection problem can be solved.
Theorem 1. Let G = (V, E) be a sparse graph. Given a distance oracle that an-
swers (2− ǫ)-approximate distance queries using s(V ,E) space with t(V ,E)
query time, we can solve the set intersection problem using s(m + U, n) space
with t(m + U, n) query time.
i=1 Si.
Proof. For the set intersection problem we are given a database D consisting of
m sets drawn from universe U , S1, . . . , Sn ⊆ U . We denote n to be the input
size, i.e., n = Pm
We construct a bipartite graph with two disjoint sets of vertices: V1 with
vertices for each set in D and V2 with vertices for each element in U . Hence,
V1 = m and V2 = U. The edges between V1 and V2 are simple, if an element
e ∈ U belongs to a set s then there is an edge between the corresponding vertices
in the bipartite graph. Because this graph is a bipartite graph it is simple that the
distance between each two vertices must be even. We can see that if two sets have
any intersection between them, the distance between the corresponding vertices
is 2. The number of edges on this graph is bounded by n. We construct a distance
oracle for this graph which answers (2 − ǫ)-approximate distance queries using
s(m + U, n) space.
Given two sets Si, Sj we would like to calculate if there is any intersection
between them. To answer that we retrieve the approximate distance between
the corresponding vertices of Si and Sj in the bipartite graph. Because both the
vertices are in V1 if the approximate distance is less than 4−ǫ, the exact distance
must be 2 because the distance must be even. This means that there is an
element e that has an edge to either vi and vj, therefore, there is an intersection
between Si and Sj. Otherwise, there is no such an element, hence, there is
no intersection between the sets. Therefore, we can answer the set intersection
problem in t(m + U, n) query time using s(m + U, n) space.
⊓⊔
3 Distance Oracle with Constant Additive Error
In this section we extend the reduction to (2 − ǫ)-approximate distance oracle
with constant additive error. We prove that set intersection is harder than ap-
proximate distance oracle even for distance oracle with constant additive error.
We denote a distance oracle with multiplicative error d and additive error c as
(d, c)-approximation distance oracle.
Theorem 2. Let G = (V, E) be a sparse graph. Given a distance oracle that an-
swers (2−ǫ, c)-approximate distance queries using s(V ,E) space with t(V ,E)
query time, we can solve the set intersection problem using s(m + U, n) space
with t(m + U, n) query time.
Proof. We build a bipartite graph as in the proof of Theorem 1. But now each
edge between V1 and V2 will be a path of 2
ǫ (c− 1) vertices. This adds a constant
number of vertices and edges, hence, the space cost of a distance oracle for the
bipartite graph is still s(m + U, n) space.
Given two sets Si, Sj we would like to calculate if there is any intersection
between them. To answer that we retrieve the approximate distance between the
corresponding vertices of Si and Sj in the bipartite graph.
If the approximated distance is less than 4c
ǫ , because the distance oracle is
(2 − ǫ, c)-approximate distance oracle, it means that the exact distance has to
be less than 4c
ǫ − 4c + c. The minimal distance between vertex
ǫ × (2 − ǫ) + c = 8c
in V1 and vertex in V2 is 2
ǫ (c − 1), hence, the distance is exactly 4
ǫ (c − 1) and
therefore there is an intersection. If the approximated distance is greater than
4c
ǫ , there would be no intersection because the distance is too high.
s(m + U, n) space.
By that we solved the set intersection problem in t(m + U, n) time using
⊓⊔
4 Conclusions
In this paper we showed that set intersection is harder than distance oracle on
sparse graphs. We showed how the set intersection problem can be solved using
(2 − ǫ)-approximate distance oracle with constant additive error.
References
1. Thorup, M., Zwick, U.: Approximate distance oracles. In: STOC '01: Proceedings
of the thirty-third annual ACM symposium on Theory of computing, New York,
NY, USA, ACM (2001) 183 -- 192
2. Cohen, H., Porat, E.: Fast set intersection and two-patterns matching. In L´opez-
Ortiz, A., ed.: LATIN. Volume 6034 of Lecture Notes in Computer Science., Springer
(2010) 234 -- 242
|
1809.05791 | 1 | 1809 | 2018-09-16T01:42:13 | Constant factor FPT approximation for capacitated k-median | [
"cs.DS"
] | Capacitated k-median is one of the few outstanding optimization problems for which the existence of a polynomial time constant factor approximation algorithm remains an open problem. In a series of recent papers algorithms producing solutions violating either the number of facilities or the capacity by a multiplicative factor were obtained. However, to produce solutions without violations appears to be hard and potentially requires different algorithmic techniques. Notably, if parameterized by the number of facilities $k$, the problem is also $W[2]$ hard, making the existence of an exact FPT algorithm unlikely. In this work we provide an FPT-time constant factor approximation algorithm preserving both cardinality and capacity of the facilities. The algorithm runs in time $2^{\mathcal{O}(k\log k)}n^{\mathcal{O}(1)}$ and achieves an approximation ratio of $7+\varepsilon$. | cs.DS | cs |
Constant factor FPT approximation for capacitated k-median
Marek Adamczyk ∗
University of Warsaw
Jaros law Byrka †
University of Wroc law
Jan Marcinkowski ‡
University of Wroc law
Syed M. Meesum §
University of Wroc law
Micha l W lodarczyk ¶
University of Warsaw
Abstract
Capacitated k-median is one of the few outstanding optimization problems for which the
existence of a polynomial time constant factor approximation algorithm remains an open prob-
lem. In a series of recent papers algorithms producing solutions violating either the number
of facilities or the capacity by a multiplicative factor were obtained. However, to produce
solutions without violations appears to be hard and potentially requires different algorithmic
techniques. Notably, if parameterized by the number of facilities k, the problem is also W [2]
hard, making the existence of an exact FPT algorithm unlikely. In this work we provide an
FPT-time constant factor approximation algorithm preserving both cardinality and capacity of
the facilities. The algorithm runs in time 2O(k log k)nO(1) and achieves an approximation ratio
of 7 + ε.
1
Introduction
For many years approximation algorithms and FPT algorithms were developed in parallel. Recently
the two paradigms are being combined and provide intriguing discoveries in the intersection of
the two worlds. It is particularly interesting in the case of problems for which we fail to progress
improving the approximation ratios in polynomial time. An excellent example of such a combination
is the FPT approximation algorithm for the k-Cut problem by Gupta et al. [15].
In this work we focus on the Capacitated k-Median problem, whose approximability attracted
attention of many researchers. Unlike in the case of the k-Cut problem, it is still not clear what
approximation is possible for Capacitated k-Median in polynomial time. As shall be discussed
in more detail in the following section, the best true approximation known is O(log k) based on tree
embedding of the underlying metric. The other algorithms either violate the bound on the number
of facilities or the capacity constraints.
Our main result is a (7 + ǫ)-approximation algorithm for the Capacitated k-Median problem
running in FPT(k) time, that exploits techniques from both -- approximation and FPT -- realms.
The algorithm builds on the idea of clustering the clients into ℓ = O(k · (log n)/ε) locations, which
∗[email protected]
†[email protected]
‡[email protected]
§[email protected]
¶[email protected]
1
is similar to the approach from the O(log k)-approximation algorithm, where one creates O (k)
clusters. This is followed by guessing the distribution of the k facilities inside these ℓ clusters.
Having such a structure revealed, we simplify the instance further by rounding particular distances
and reduce the problem to linear programming over a totally unimodular matrix.
1.1 Problems overview and previous work
In the Capacitated k-Median problem (CKM), we are given a set F of facilities, each facility
f with a capacity uf ∈ Z>0, a set C of clients, a metric d over F ∪ C and an upper bound k on
the number of facilities we can open. A solution to the CKM problem is a set S ⊆ F of at most
k open facilities and a connection assignment φ : C → S of clients to open facilities such that
(cid:12)(cid:12)φ−1(f )(cid:12)(cid:12) 6 uf for every facility f ∈ S. The goal of the problem is to find a solution that minimizes
the connection cost Pc∈C d(c, φ(c)).
In the case when all the facilities can serve at most u clients, for some integer u, we obtain the
Uniform CKM problem.
Uncapacitated k-median The standard k-median problem, where there is no restriction on the
number of clients served by a facility, can be approximated up to a constant factor [9, 2]. The
current best is the (2.675 + ǫ)-approximation algorithm of Byrka et al. [4], which is a result of
optimizing a part of the algorithm by Li and Svensson [21].
Approximability of CKM As already stressed, Capacitated k-Median is among few re-
maining fundamental optimization problems for which it is not clear if there exist polynomial time
constant factor approximation algorithms. All the known algorithms violate either the number
of facilities or the capacities. In particular, already the algorithm of Charikar et al. [9] gave 16-
approximate solution for the uniform capacitated k-median violating the capacities by a factor
of 3. Then Chuzhoy and Rabani [10] considered general capacities and gave a 50-approximation
algorithm violating capacities by a factor of 40.
The difficulty appears to be related to the unbounded integrality gap of the standard LP re-
laxation. To obtain integral solutions that are bounded with respect to the fractional solution to
the standard LP, one has to either allow the integral solution to open twice as much facilities or to
violate the capacities by a factor of two. LP-rounding algorithms essentially matching these limits
have been obtained [1, 3].
Subsequently, Li broke this integrality gap barrier by giving a constant factor algorithm for the
capacitated k-median by opening (1 + ε) · k facilities [19, 20]. Afterwards analogous results, but
violating the capacities by a factor of (1 + ε) were also obtained [5, 12].
The algorithms with (1 + ε) violations are all based on strong LP relaxations containing ad-
ditional constraints for subsets of facilities. Notably, it is not clear if these relaxations can be
solved exactly in polynomial time, still they suffice to construct an approximation algorithm via
the "round-or-separate" technique that iteratively adds consistency constraints for selected subsets.
Although while spectacularly breaking the standard LP integrality bound, these techniques appear
insufficient to yield a proper approximation algorithm that does not violate constraints.
The only true approximation for CKM known is a folklore O (log k) approximation algorithm
that can be obtained via the metric tree embedding with expected logarithmic distortion [13]. To
the best of our knowledge, this result has not been explicitly published, but it can be obtained
similarly to the O (log k)-approximation for Uncapacitated KM by Charikar [7]. For the sake of
2
completeness and since it follows easily from our framework, we give its proof in Section 4 without
claiming credit for it.
This O (log k)-approximation is in contrast with other capacitated clustering problems such as
facility location and k-center, for which constant factor approximation algorithms are known [17, 11].
1.2 Parameterized Complexity
A parameterized problem instance is created by associating an input instance with an integer
parameter k. We say that a problem is fixed parameter tractable (FPT) if any instance (I, k) of the
problem can be solved in time f (k) · IO(1), where f is an arbitrary computable function of k. We
say that a problem is FPT if it is possible to give an algorithm that solves it in running time of the
required form. Such an algorithm we shall call a parameterized algorithm.
To show that a problem is unlikely to be FPT, we use parameterized reductions analogous to
those employed in the classic complexity theory. Here, the concept of W-hardness replaces the
one of NP-hardness, and we need not only to construct an equivalent instance in FPT time, but
also ensure that the size of the parameter in the new instance depends only on the size of the
parameter in the original instance. In contrast to the NP-hardness theory, there is a hierarchy of
classes FPT = W[0] ⊆ W[1] ⊆ W[2] ⊆ . . . and these containments are believed to be strict.
If
there exists a parameterized reduction transforming a problem known to be W[t]-hard for t > 0 to
another problem Π, then the problem Π is W[t]-hard as well. This provides an argument that Π is
unlikely to admit an algorithm with running time f (k) · IO(1).
We begin with an argument that allowing FPT time for (even uncapacitated) k-Median should
not help in finding the optimal solution and we still need to settle for approximation.
Fact 1. The Uncapacitated k-Median problem is W[2]-hard when parameterized by k, even on
metrics induced by unweighted graphs.
Proof. Consider an instance of the Dominating Set problem, which is W[2]-hard when param-
eterized by the solution size. A dominating set of size at most k exists in graph G if and only if
we can find a vertex set S of size k, such that all other vertices are at distance 1 from S. This is
equivalent to the solution to Uncapacitated k-Median on the metric induced by G being of size
exactly V (G) − k.
Parameterized Approximation In recent years new research directions emerged in the inter-
section of the theory of approximation algorithms and the FPT theory. It turned out that for some
problems that are intractable in the exact sense, parameterization still comes in useful when we
want to reduce the approximation ratio. Some examples are (2−ε)-approximation for k-Cut [15] or
f (F )-approximation for Planar-F Deletion [14] for some implicit function f . The dependency
on F was later improved, leading to O(log k)-approximations for, e.g., k-Vertex Separator[18]
and k-Treewidth Deletion [16].
On the other hand some problems parameterized by the solution size have been proven resistant
to such improvements. Chalermsook et al. [6] observed that under the assumption of Gap-ETH there
can be no parametrized approximation with ratio o(k) for k-Clique and none with ratio f (k) for
k-Dominating Set (for any function f ). Subsequently Gap-ETH has been replaced with a better
established hardness assumption FPT 6= W[1] for k-Dominating Set [23].
3
1.3 Organization of the paper
Our main result is stated in Theorem 11 (Section 3.3), where we present a (7 + ǫ)-approximation
algorithm for the Non-Uniform CKM problem running in FPT(k) time.
To obtain this result we need two ingredients. First is a metric embedding that reduces the
problem to a simpler instance, called ℓ-centered, what is described in Section 2. This reduction
provides a richer structure, which can be exploited to obtain an O (log k)-approximation via tree
embeddings [13]. As already mentioned, similar approach was presented by Charikar et al. [7] in
their algorithm for the uncapacitated setting. We present this result for the sake of completeness
in Section 4, after the main result.
The second ingredient is a parameterized algorithm for the ℓ-centered instances. Since it is
simpler in the uniform setting, we solve it in Section 3.2 as a warm up before the main result. This
way the new ideas are being revealed gradually to the reader.
2
ℓ-Centered instances
Suppose we work with a graph on nodes F ∪ C, on which we are given a metric d. In our con-
siderations the set F ∪ C will be fixed throughout, however we will be modifying the metric over
it. Consider an algorithm ALG which produces a solution ALG (d) for a metric d. This solution
can be seen as a mapping which we explicitly denote by φALG(d). Its cost in the metric d′ equals
Pc∈C d′(cid:0)c, φALG(d)(cid:1) which we shall briefly denote by cost(cid:0)φALG(d), d′(cid:1). The second argument is
useful, when an algorithm ALG produces a solution (mapping) ALG (d) with respect to metric d,
but later on we may be interested in its cost over a different metric. Also, let OP T (d) denote the
optimum solution for the CKM problem on metric d.
In order to solve CKM, we shall invoke an algorithm for Uncapacitated KM as a subroutine.
Let ALGℓ
unc (d) be a relaxed solution that opens up to ℓ ≥ k facilities and can break the capacity
constraints. It induces a mapping which, for consistency, we shall denote by φALGℓ
unc(d). Observe
that in this mapping every client can be connected to the closest open facility. Since Uncapac-
itated KM admits constant approximation algorithms, we can work with solutions satisfying:
cost(cid:16)φALGℓ
constant we will be able to achieve in the relation above.
unc(d), d(cid:17) = O(cid:0)cost(cid:0)φOP T (d), d(cid:1)(cid:1). The larger ℓ we allow in the relaxation, the smaller
Using such an algorithm for Uncapacitated KM as a subroutine, we can find a simpler metric
unc (d)) be the set
unc (d). For each such a facility f we create a copy vertex sf , which is
to work with. First we build a graph which will induce the metric. Let F (ALGℓ
of facilities opened by ALGℓ
at distance 0 from f . We denote the set of copies by S, i.e., S =(cid:8)sf(cid:12)(cid:12)f ∈ F (ALGℓ
that we demand the distance from f to sf to be 0, we can naturally extend the metric d to the set
C ∪ F ∪ S. To distinguish facilities from F (ALGℓ
unc (d)) from their copies S, we shall call each copy
s ∈ S a center.
unc (d))(cid:9). Given
We build a complete graph on S and preserve the metric d therein. For every node v 6∈ S, be
it either a client from C or a facility from F , we place an edge to the closest (according to the
extended d) center sv ∈ S and set its length to d (v, sv). We call such a graph ℓ-centered and refer
to its induced metric as dℓ.
Definition 2. An instance of CKM is called ℓ-centered if the metric, which we shall denote by dℓ,
is induced by a weighted graph G(F ∪ C ∪ S, E) such that
1. S 6 ℓ,
4
Figure 1: An ℓ-centered instance. In the upper layer there is a set S of l vertices connected as a
clique. The rest of vertices are divided into separate clusters. Vertices in a single cluster are only
connected to their center in the set S.
2. (cid:0)S
2(cid:1) ⊆ E, i.e., S forms a clique,
3. for every v ∈ C ∪ F there is only one edge incident to v in E, and it connects v to some
sv ∈ S.
For a center s ∈ S we shall say that all nodes from F ∪ C that are connected to s form a cluster
of s. If we consider only nodes from F , then we talk about an f -cluster of s, denoted F (s).
In the following lemma we relate the cost of embedding the optimum solution OP T (d) from a
metric d to dl.
Lemma 3 (Embedding d into l-centered metric dl). Let ALGℓ
unc (d) be a solution for the Unca-
pacitated KM problem on metric d from which we construct the ℓ-centered instance. Optimal
solution OP T (d) can be embedded into an l-centered metric dl with the cost relation being
cost(cid:16)φOP T (d), d(cid:17) 6 cost(cid:16)φOP T (d), dl(cid:17) 6 3 · cost(cid:16)φOP T (d), d(cid:17) + 4 · cost(cid:16)φALGℓ
unc(d), d(cid:17) .
Proof. Let c be a client connected to facility fc in the optimal solution OP T (d). Let sc be the
center closest to c within S (the ℓ-center), and let sfc be the center closest to fc. First let us
note that dl (c, fc) = d (c, sc) + d(cid:0)sc, sfc(cid:1) + d(cid:0)sfc , fc(cid:1). Now let us bound the terms d(cid:0)fc, sfc(cid:1) and
d(cid:0)sc, sfc(cid:1) separately.
Fact 4. For every client c and its facility fc from OP T we have d(cid:0)fc, sfc(cid:1) 6 d (fc, c) + d (c, sc).
Proof. Since sfc is the closest ℓ-center to the facility fc, we have that d(cid:0)fc, sfc(cid:1) 6 d (fc, sc). At the
Fact 5. For each c we have d(cid:0)sc, sfc(cid:1) 6 2 (d (fc, c) + d (c, sc)).
same time, from the triangle inequality it follows that d (fc, sc) 6 d (fc, c) + d (c, sc).
5
c
fc
sc
sfc
Figure 2: Situation in Lemma 3. In the optimum solution to the CKM instance, client c is connected
to the facility fc. In the ℓ-centered instance c resides in a cell, where sc is a center. The center of
fc is sfc.
Proof. From the triangle inequality we know that
d(cid:0)sc, sfc(cid:1) 6 d (sc, c) + d (c, fc) + d(cid:0)fc, sfc(cid:1) .
From Fact 4 we also know that d(cid:0)fc, sfc(cid:1) 6 d (fc, c) + d (c, sc), and combining the two inequalities
we get d(cid:0)sc, sfc(cid:1) 6 d (sc, c) + d (c, fc) + d(cid:0)fc, sfc(cid:1) 6 2 (d (fc, c) + d (c, sc)).
These facts imply
dl (c, fc) = d (c, sc) + d(cid:0)sc, sfc(cid:1) + d(cid:0)sfc , fc(cid:1)
6 d (c, sc) + d(cid:0)sc, sfc(cid:1) + (d (fc, c) + d (c, sc))
6 d (c, sc) + 2 (d (fc, c) + d (c, sc)) + (d (fc, c) + d (c, sc))
= 3 · d (fc, c) + 4 · d (c, sc) ,
(from Fact 4)
(from Fact 5)
which implies the second inequality from the statement of Lemma 3. The first one directly comes
from the triangle inequality
completing the whole proof.
d (c, fc) 6 d (c, sc) + d(cid:0)sc, sfc(cid:1) + d(cid:0)sfc , fc(cid:1) = dl (c, fc) ,
Another lemma is quite simple. Its proof just comes from the fact that metric dl dominates the
metric d, i.e., dl (u, v) > d (u, v) for all pairs of vertices u, v ∈ C ∪ F .
Lemma 6 (Going back from l-centered metric dl to d). Any solution for the l-centered metric dl
can be embedded back into d without any loss:
cost(cid:16)φALG(dl), dl(cid:17) > cost(cid:16)φALG(dl), d(cid:17) .
6
Blending together Lemmas 3 and 6 we can state the following Lemma about reducing the CKM
problem to ℓ-centered instances.
Lemma 7. Suppose we are given a solution ALGℓ
on metric d which opens ℓ centers, but β-approximates the optimum solution OP T k
unc (d) for the Uncapacitated KM problem
unc (d) for Un-
capacitated KM problem with k centers, i.e., cost(cid:0)ALGℓ
Suppose we are given an α-approximation algorithm for the CKM problem on ℓ-centered instances.
If so, then we can construct an α·(3 + 4β)-approximation algorithm for CKM on general instances.
unc (d) , d(cid:1) 6 β · cost(cid:0)OP T k
unc (d) , d(cid:1).
Proof. Suppose that we have an α-approximation solution for the ℓ-centered instance with metric
dl, i.e., ALG(dl) such that
Since OP T (d) is some solution for the ℓ-centered instance with metric dl we have
cost(cid:16)φALG(dl), dl(cid:17) 6 α · cost(cid:16)φOP T (dl), dl(cid:17) .
And from Lemma 2 we have that
cost(cid:16)φALG(dl), dl(cid:17) 6 α · cost(cid:16)φOP T (dl), dl(cid:17) 6 α · cost(cid:16)φOP T (d), dl(cid:17) .
cost(cid:16)φALG(dl), dl(cid:17) 6 α · cost(cid:16)φOP T (dl), dl(cid:17)
6 α · cost(cid:16)φOP T (d), dl(cid:17)
6 α(cid:16)3 · cost(cid:16)φOP T (d), d(cid:17) + 4 · cost(cid:16)φALGℓ
unc (d) β-approximates the optimal solution OP T k
unc(d), d(cid:17)(cid:17) .
Since solution ALGℓ
KM with k centers on metric d, we have that
unc (d) for Uncapacitated
cost(cid:16)φALGℓ
unc(d), d(cid:17) 6 β · cost(cid:16)φOP T k
unc(d), d(cid:17) 6 β · cost(cid:16)φOP T (d), d(cid:17) .
The second inequality cost(cid:16)φOP T k
unc(d), d(cid:17) 6 cost(cid:0)φOP T (d), d(cid:1) follows from an obvious fact that
uncapacitated version of the problem is easier than the capacitated. Hence
cost(cid:16)φALG(dl), dl(cid:17) 6 α(cid:16)3 · cost(cid:16)φOP T (d), d(cid:17) + 4 · cost(cid:16)φALGℓ
unc(d), d(cid:17)(cid:17)
6 α(cid:16)3 · cost(cid:16)φOP T (d), d(cid:17) + 4β · cost(cid:16)φOP T (d), d(cid:17)(cid:17)
6 α (3 + 4β) · cost(cid:16)φOP T (d), d(cid:17) .
Since without any loss we can embed the solution ALG(dl) for the ℓ-centered metric dl into the initial
metric d (Lemma 3) we obtain an α · (3 + 4β)-approximation algorithm. The claim follows.
3 Constant factor approximation
In this section we present the main result of the paper which is a (7 + ε)-approximation algorithm
for the Non-Uniform CKM problem. We precede it with a (7 + ε)-approximation algorithm
for the Uniform CKM problem to introduce the ideas gradually. Both algorithms enumerate
configurations of open facilities' locations, and as a subroutine we need to use an algorithm which,
for a fixed configuration of k open facilities, finds the optimal assignment of clients to facilities.
This subroutine is presented in the following subsection.
7
3.1 Optimal mapping subroutine
We are given an ℓ-centered metric instance (F ∪ C ∪ S, dℓ) of the k-median problem. Suppose that
we have already decided to open a fixed subset F open ⊆ F of the facilities and we look for a mapping
φ : C → F open. In the uncapacitated case we can just assign each client to the closest facility in
F open. It turns out that even in the capacitated setting we can find the mapping φ optimally in
polynomial time for a given F open. We state the problem of finding the optimal φ as an integer
program:
subject to
minimize Xc∈C Xf ∈F open
Xf ∈F open
Xc∈C
dℓ(c, f ) · xc,f
(MAPPING-IP)
xc,f = 1
∀c ∈ C,
xc,f 6 uf
∀f ∈ F open,
xc,f ∈ {0, 1}.
In the above program xc,f = 1 represents the fact that φ(c) = f .
Lemma 8. We can find an optimal solution to the (MAPPING-IP) in polynomial time.
Proof. The proof follows from the fact that the relaxation of the above integer program -- a program
which differs from (MAPPING-IP) only with the xc,f > 0 constraints instead of xc,f ∈ {0, 1} -- has
an optimal solution which is integral. To see this, observe that the linear program is a formulation of
the transportation problem. For such a linear program, the constraint matrix is totally unimodular,
which implies the integrality of an extremal solution. See [24] for a reference.
3.2 Uniform case
As a warm up, we begin with a parameterized algorithm for the uniform case. It is a bit simpler
than the general case, because once we know the number of facilities to open in f -cluster F (s),
then we can choose them greedily.
Lemma 9. Uniform CKM can be solved exactly in time ℓk · nO(1) on ℓ-centered instances.
Proof. Let (F ∪ C ∪ S, dℓ) be the ℓ-centered metric. Note that the f -clusters partition the whole set
of facilities, i.e., ∪s∈SF (s) = F . Let OP T (dℓ) be an optimal solution for the CKM problem on dℓ.
Every facility f ∈ F belongs to exactly one f -cluster F (s). Hence, the f -clusters partition the set
of k facilities opened by OP T (dℓ). Let us look at all the facilities from a particular f -cluster F (s)
opened by OP T (dℓ), and suppose that OP T (dℓ) opens ks of facilities in F (s). Since we consider
a uniform capacity case, we can assume without loss that these ks open facilities from F (s) are
exactly the ones that are closest to s.
Therefore, if we know what is the number of facilities that OP T (dℓ) opens in each f -cluster,
then we would know what the exact set of open facilities in OP T (dℓ) is due to the greediness in
each f -cluster. To find out this allocation we can simply enumerate over all possibilities. We just
need to scan over all configurations (ks)s∈S where Ps ks = k. Since there are k facilities to open,
and each of them can belong to one of ℓ f -clusters F (s), there are at most ℓk possible configurations.
8
Of course some configurations may not be feasible since it may happen that ks > F (s), but these
can be simply ignored.
For each configuration (ks)s∈S we need to find the optimal mapping of clients to the set of
by configuration (ks)s∈S, that is, where we greedily open ks facilities in f -cluster F (s). Given
open facilities that preserves their capacities. Let F (cid:0)(ks)s∈S(cid:1) be the set of open facilities induced
F (cid:0)(ks)s∈S(cid:1), to find the optimal mapping we use the polynomial time exact algorithm from Lemma 8
with F open = F (cid:0)(ks)s∈S(cid:1).
Once we know the optimal assignment for each configuration, we can simply take the cheapest
one, knowing that it is the optimal one. This proves the lemma.
This lemma suffices to obtain a (7 + ε)-approximation for Uniform CKM with a reasoning
that we will present in Theorem 11 in full generality.
3.3 Non-uniform case
Lemma 10. Non-Uniform CKM can be solved with approximation ratio (1 + ǫ) in time
ε ln n
(cid:0)O(cid:0)ℓ · 1
ε(cid:1)(cid:1)k
nO(1) on ℓ-centered instances.
Proof. We begin with guessing the largest distance in dℓ between a client and a facility that would
appear in the optimal solution -- let us denote this quantity as D. There are at most O(n2) choices
for D, and from now we assume that it is guessed correctly. Note that D ≤ cost (OP T (dℓ), dℓ) and
D ≥ d(f, sf ) for all facilities opened by OP T (dℓ).
Consider the set of facilities F (s) in the cluster of a center s. We can remove all facilities f such
that d(s, f ) > D, because they cannot be a part of the optimal solution. Let us partition remaining
facilities from F (s) into buckets F0 (s) , F1 (s) , ..., F⌈ log1+ε
ε ⌉ (s), such that
n
Fi (s) =
nf ∈ F (s)(cid:12)(cid:12)(cid:12)
nf ∈ F (s)(cid:12)(cid:12)(cid:12)
d (s, f ) ∈h(1 + ε)−(i+1) D, (1 + ε)−i Dio for i <(cid:6) log1+ε
d (s, f ) ∈h0, (1 + ε)−⌈ log1+ε
for i =(cid:6) log1+ε
n
n
ε(cid:7)
ε(cid:7)
n
The number of buckets equals log1+ε
n
ε = 1
ln(1+ε) ln n
ℓ(s, f ) = (1 + ε)−i D for f ∈ Fi (s). The distances within S remain untouched. Observe
ε ln n
ε(cid:1). We modify the metric again
ε ⌉ Dio
ε = O(cid:0) 1
by setting d′
that the distances can only increase.
We shall guess the structure of the solution OP T (d′
ℓ) similarly as in Lemma 9. For each of the
k facilities, we can choose its location as follows: first we choose one of the ℓ-centers s (ℓ choices),
and then we choose one of the Fi (s) partitions (O(cid:0) 1
facilities in a particular partition Fi (s) as ks,i. We can assume that ks,i ≤ Fi (s) because otherwise
we know that the guess was incorrect. Since d′
ℓ(s, f ) is the same for all f ∈ Fi (s), we can assume
the optimal solution opens ks,i facilities with the biggest capacities.
ε(cid:1) choices). Let us denote the number of
ε ln n
Once we establish the set of facilities to open, we can find the optimal assignment in metric d′
ℓ
using the polynomial time exact subroutine from Lemma 8. The total time complexity of solving
the problem exactly over d′
ℓ equals the running time of the subroutine times the number of possible
configurations, which is (cid:0)O(cid:0)ℓ · 1
ε ln n
ε(cid:1)(cid:1)k
nO(1).
9
It remains to prove that the algorithm yields a proper approximation. We will show that for
any solution SOL it holds that
(1)
ℓ of cost at
cost(cid:0)φSOL, dℓ(cid:1) 6 cost(cid:0)φSOL, d′
ℓ(cid:1) 6 (1 + ε) · cost(cid:0)φSOL, dℓ(cid:1) + ε · D.
By substituting SOL = OP T (dℓ) we learn that there exists a solution over metric d′
most (1 + ε) · cost(cid:0)φOP T (dℓ), dℓ(cid:1) + ε · D ≤ (1 + 2ε) · cost(cid:0)φOP T (dℓ), dℓ(cid:1) for correctly guessed D.
Therefore the cost of the solution found by our algorithm cannot be larger. Finally we substitute
this solution as SOL to see that its cost cannot increase when returning to metric dℓ. The claim
will follow by adjusting ε.
The first inequality in (1) is straightforward because d′
ℓ dominates dℓ. Consider now a pair
ℓ(c, f ) ≤ (1 + ε) · dℓ(c, f ),
so the cost of connecting such pairs increases at most by a multiplicative factor (1 + ε) during the
n . Since there are at most n such pairs, the
(c, f = φSOL(c)), where f ∈ Fi (s). If i < (cid:6) log1+ε
metric switch. If i = (cid:6) log1+ε
ε(cid:7) , then dℓ(c, f ) ≤ d′
total additive cost increase is bounded by ε · D.
ε(cid:7) , then d′
ℓ(s, f ) = εD
n
n
Theorem 11. Non-Uniform CKM can be solved with approximation ratio (7 + ǫ) in time
(k/ǫ)O(k)nO(1).
Proof. From Lemma 10 we know that we can get a (1 + ε) −approximation algorithm for the Non-
nO(1). We shall use
the (1 + ε)-approximation for Uncapacitated KM by Lin and Vitter [22], that opens at most
Uniform CKM problem on ℓ-centered instances in time (cid:0)O(cid:0)ℓ · 1
ℓ =(cid:0)1 + 1
together with Lemma 10, we obtain a (7 + ε) −approximation algorithm for the general Non-
Uniform CKM problem with running time
ε(cid:1) k · (ln n + 1) facilities. By plugging this subroutine to find ℓ-centers into the Lemma 7
O (cid:18)(cid:18)1 +
ε(cid:19)k! nO(1) = O (cid:18) 1
εO(1) k ln2 n(cid:19)k! nO(1).
ε(cid:19) k · (ln n + 1) ·
ε(cid:1)(cid:1)k
ε ln n
1
ε
ln
n
1
ln n
Finally, we use standard arguments to show that (ln n)2k ≤ max(n, kO(k)). Consider two cases. If
2 ln ln n ≤ k, then by inverting we know that ln n = O (k ln k), and so (ln n)2k = kO(k). Suppose now
that
2 ln ln n > k. In this case
ln n
(ln n)2k < (ln n)
ln n
ln ln n = 2ln ln n· ln n
ln ln n = n.
4 O (log k)-approximation
In this section present a polynomial-time O (log k)-approximation algorithm for CKM. A constant-
factor approximation algorithms for Uncapacitated KM exist [9], and so it is a clear conse-
quence of the Lemma 7 with β being constant that it is sufficient for us to construct an O (log k)-
approximation algorithm for the k-centered instances.
A standard tool to provide such a guarantee is the Probabilistic Tree Embedding by [13]. This
makes our algorithm a randomized one, but if needed, it is possible to derandomize it using the
ideas from [8].
10
Definition 12. A set of metric spaces T together with a probability distribution πT over T
probabilistically α-approximates the metric space (X, d) if
1. Every metric τ ∈ T dominates (X, d), that is, d(x, y) 6 τ (x, y) ∀x, y ∈ X.
2. For every pair of points x, y ∈ X its expected distance is not expanded by more then α, i.e.,
Eτ ∼πT [τ (x, y)] 6 α · d(x, y).
It is a well-known fact, that any metric (X, d), can be probabilistically O(log X)-approximated
by a distribution of tree metrics, such that the points in X are the leaves in the resulting tree [13].
As described in Definition 2, our k-centered metric dk is induced by a graph composed of two
layers -- the set S of k vertices connected in a clique, and the rest of vertices, F ∪C, each connected
to only one vertex in S. Let T be a random tree embedding of the set S (with a metric function
dT ). A modified instance GT of our problem is created by replacing the clique S with its tree
approximation T .
Lemma 13. An optimum solution for CKM on the instance GT is in expectation at most O (log k)
times larger than the optimum for the metric dk.
Proof. OP T (dk) denotes the optimum mapping of clients to k facilities in the k-centered metric
dk. Consider client c and facility f = φOP T (dk) (c). Let now sc be the center of c and sf the center
of f . The cost of connecting client c to f amounts to
in the metric dk.
dk (c, f ) = dk (c, sc) + dk(cid:0)sc, sf(cid:1) + dk(cid:0)sf , f(cid:1)
The guarantee of tree embeddings gives us an upper bound on a cost of applying the same
mapping in the instance GT ,
E [dT (c, f )] = dk (c, sc) + E(cid:2)dT (cid:0)sc, sf(cid:1)(cid:3) + dk(cid:0)sf , f(cid:1)
6 dk (c, sc) + O (log k) · dk(cid:0)sc, sf(cid:1) + dk(cid:0)sf , f(cid:1)
6 O (log k) · dk (c, f ) .
Which means that E(cid:2)cost(cid:0)φOP T (dk), dT(cid:1)(cid:3) 6 O (log k)·cost(cid:0)φOP T (dk), dGk(cid:1). Moreover, OP T (dk)
might not be the optimal solution for the metric dT , yet its optimal solution can only have smaller
cost:
cost(cid:16)φOP T (dT ), dT(cid:17) 6 cost(cid:16)φOP T (dk), dT(cid:17)
Theorem 14. The CKM problem admits an O(log k)-approximation algorithm with polynomial
running time.
Proof. After applying the probabilistic tree embedding to the graph inducing dk -- as presented
in Lemma 13 -- we obtain a tree instance GT . It should come as no surprise that the problem is
polynomially solvable on trees and we explain how to find the optimum solution on GT in Lemma 16.
The assignment φOP T (dT ), which yields the minimum cost on the tree GT , can be now used to match
clients to facilities in the original instance. It does not incur any additional cost, as
cost(cid:16)φOP T (dT ), dT(cid:17) > cost(cid:16)φOP T (dT ), dk(cid:17) > cost(cid:16)φOP T (dT ), d(cid:17)
from the property (1) of Definition 12 and Lemma 6. Combining this with a bound on
E(cid:2)cost(cid:0)φOP T (dk), dT(cid:1)(cid:3) from Lemma 13 finishes the proof.
11
et
t
(a)
et
et1
hk′ , bi
et2
hk′
1 , b1 i
hk′
2 , b2 i
t1
t2
(b)
Figure 3: (a) A subtree t with some open facilities (squares) and a number of clients (circles).
If the total capacity of facilities opened in t exceeds the needs of the clients, we may decide to
connect some clients from outside through edge et.
If the capacity of open clients is too small,
we may connect some clients with outside facilities through et. It never makes sense to do both.
(b) A dynamic programming step. To compute D(t, k′, b) we need to find the cheapest solutions
for subtrees, such that k′
2 = k′ and b1 + b2 = b. The additional cost incurred is d(et) · b.
1 + k′
4.1 CKM on a tree
The second ingredient to the O(log k)-approximation for CKM is solving the problem exactly on
trees. We will now describe a simple, exact, polynomial algorithm for that special case. In our
algorithm we can assume, that all the clients and facilities reside in leaves, but the principle is easy
to extend to the general problem on trees.
Imagine we have a subtree t of the tree instance, hanging on an edge et. Once we have decided,
which facilities to open inside the subtree t, we know if their total capacity is sufficient to serve all
the clients inside t. If not, then we need to route some clients' connections to the facilities outside
through the edge et. However, if the facilities we have opened in t have enough total capacity to
serve some b clients from the outside, we will connect them through the edge et (see Figure 3). This
insight lays out the dynamic algorithm for us. We first turn the tree into a complete binary tree by
adding dummy vertices and edges of length 0 (which may double its size). Then, for every subtree
t, numbers k′ and b, we compute D(t, k′, b).
Definition 15. D(t, k′, b), for subtree t, number k′ ∈ {0, . . . , k} of facilities and balance b ∈
{−n, . . . , n}, is the minimum cost of opening exactly k′ facilities in t and routing exactly b clients
down through et (b < 0 would mean that we are routing −b clients up). The cost of routing is
counted to the top endpoint of et.
Lemma 16. The CKM problem on trees admits a polynomial time exact algorithm.
Proof. Computing D(t, k′, b) on t with two children t1 and t2 amounts to finding k′
1, k′
2, b1 and b2
12
that minimize
D(t1, k′
1, b1) + D(t2, k′
2, b2),
such that b1 + b2 = b and k′
pair hk′, bi. Once k′
1, k′
2, b1 and b2 are found, we set
1 + k′
2 = k′. They can be trivially found in O (k · n) time for a single
D(t, k′, b) = D(t1, k′
1, b1) + D(t2, k′
2, b2) + d(et) · b,
where d(e) is the length of the edge in our tree. For a leaf l, D(l, k′, b) is defined naturally, depending
on whether the leaf holds a client or a facility. Note, that for a leaf with a facility, D(l, 1, b) is finite
also for b smaller than the capacity of the facility, as the optimal solution might not use it entirely.
Finally, the optimum solution to the CKM problem on the entire tree T is equal to
min
k′∈{1,...,k}
D(T, k′, 0).
5 Conclusions and open problems
We have presented a (7 + ε)-approximation algorithm for the CKM problem, which consists of
three building blocks: approximation for Uncapacitated KM, metric embedding into a simpler
structure, and a parameterized algorithm working on ℓ-centered instances.
Whereas the first and the last ingredient are almost lossless from the approximation point of
view, the embedding procedure seems to be the main bottleneck for obtaining a better approxi-
mation guarantee. One can imagine that a different technique would allow to obtain a (1 + ε)-
approximation in FPT time. We believe that finding such an algorithm or ruling out its existence
is an interesting research direction.
Another avenue for improvement is processing k-centered instances in time 2O(k)nO(1). Such
a routine would reduce the running time of the whole algorithm to single exponential. In order
to do so, one could replace the subroutine for Uncapacitated KM by Lin and Vitter [22] with
a standard approximation algorithm that opens exactly k facilities, what would moderately increase
the constant in approximation ratio.
Finally, whereas we have used the framework of ℓ-centered instances to devise an FPT ap-
proximation, it might be possible to explore the structure of special instances further and find
a polynomial time approximation algorithm. This could yield an improvement over the O(log k)-
approximation ratio for CKM, which remains a major open problem.
13
References
[1] K. Aardal, P. L. van den Berg, D. Gijswijt, and S. Li. Approximation algorithms for hard capac-
itated k-facility location problems. European Journal of Operational Research, 242(2):358 -- 368,
2015. 1.1
[2] V. Arya, N. Garg, R. Khandekar, A. Meyerson, K. Munagala, and V. Pandit. Local search
heuristics for k-median and facility location problems. SIAM Journal on Computing, 33(3):544 --
562, 2004. 1.1
[3] J. Byrka, K. Fleszar, B. Rybicki, and J. Spoerhase. Bi-factor approximation algorithms for
hard capacitated k-median problems. In Proceedings of the Twenty-Sixth Annual ACM-SIAM
Symposium on Discrete Algorithms (SODA), pages 722 -- 736. SIAM, 2015. 1.1
[4] J. Byrka, T. Pensyl, B. Rybicki, A. Srinivasan, and K. Trinh. An improved approximation for
k-median, and positive correlation in budgeted optimization. In Proceedings of the Twenty-
Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 737 -- 756. SIAM, 2015.
1.1
[5] J. Byrka, B. Rybicki, and S. Uniyal. An approximation algorithm for uniform capacitated
k-median problem with 1 + ǫ capacity violation. In Integer Programming and Combinatorial
Optimization - 18th International Conference, IPCO 2016, Li`ege, Belgium, June 1-3, 2016,
Proceedings, pages 262 -- 274, 2016. 1.1
[6] P. Chalermsook, M. Cygan, G. Kortsarz, B. Laekhanukit, P. Manurangsi, D. Nanongkai, and
L. Trevisan. From gap-eth to fpt-inapproximability: Clique, dominating set, and more. In
Foundations of Computer Science (FOCS), 2017 IEEE 58th Annual Symposium on, pages
743 -- 754. IEEE, 2017. 1.2
[7] M. Charikar, C. Chekuri, A. Goel, and S. Guha. Rounding via trees: Deterministic approxima-
tion algorithms for group steiner trees and k -median. In Proceedings of the Thirtieth Annual
ACM Symposium on the Theory of Computing, Dallas, Texas, USA, May 23-26, 1998, pages
114 -- 123, 1998. 1.1, 1.3
[8] M. Charikar, C. Chekuri, A. Goel, S. Guha, and S. A. Plotkin. Approximating a finite metric
by a small number of tree metrics. In 39th Annual Symposium on Foundations of Computer
Science, FOCS '98, November 8-11, 1998, Palo Alto, California, USA, pages 379 -- 388, 1998. 4
[9] M. Charikar, S. Guha, ´E. Tardos, and D. B. Shmoys. A constant-factor approximation algo-
rithm for the k-median problem. In Proceedings of the thirty-first annual ACM symposium on
Theory of computing, pages 1 -- 10. ACM, 1999. 1.1, 1.1, 4
[10] J. Chuzhoy and Y. Rabani. Approximating k-median with non-uniform capacities. In Proceed-
ings of the sixteenth annual ACM-SIAM symposium on Discrete algorithms, pages 952 -- 958.
Society for Industrial and Applied Mathematics, 2005. 1.1
[11] M. Cygan, M. Hajiaghayi, and S. Khuller. Lp rounding for k-centers with non-uniform hard
capacities. In Foundations of Computer Science (FOCS), 2012 IEEE 53rd Annual Symposium
on, pages 273 -- 282. IEEE, 2012. 1.1
14
[12] H. G. Demirci and S. Li. Constant approximation for capacitated k-median with (1 + ǫ)-
capacity violation. In 43rd International Colloquium on Automata, Languages, and Program-
ming, ICALP 2016, July 11-15, 2016, Rome, Italy, pages 73:1 -- 73:14, 2016. 1.1
[13] J. Fakcharoenphol, S. Rao, and K. Talwar. A tight bound on approximating arbitrary metrics
by tree metrics. In Proceedings of the 35th Annual ACM Symposium on Theory of Computing,
June 9-11, 2003, San Diego, CA, USA, pages 448 -- 455, 2003. 1.1, 1.3, 4, 4
[14] F. V. Fomin, D. Lokshtanov, N. Misra, and S. Saurabh. Planar f-deletion: Approximation,
kernelization and optimal fpt algorithms. In Proceedings of the 2012 IEEE 53rd Annual Sympo-
sium on Foundations of Computer Science, FOCS '12, pages 470 -- 479, Washington, DC, USA,
2012. IEEE Computer Society. 1.2
[15] A. Gupta, E. Lee, and J. Li. An fpt algorithm beating 2-approximation for k-cut. In Proceedings
of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 2821 -- 2837.
Society for Industrial and Applied Mathematics, 2018. 1, 1.2
[16] A. Gupta, E. Lee, J. Li, P. Manurangsi, and M. Wlodarczyk. Losing treewidth by separating
subsets. CoRR, abs/1804.01366, 2018. 1.2
[17] M. R. Korupolu, C. G. Plaxton, and R. Rajaraman. Analysis of a local search heuristic for
facility location problems. Journal of algorithms, 37(1):146 -- 188, 2000. 1.1
[18] E. Lee. Partitioning a graph into small pieces with applications to path transversal. Mathe-
matical Programming, 2018. Preliminary version in SODA 2017. 1.2
[19] S. Li. On uniform capacitated k-median beyond the natural LP relaxation. In Proceedings
of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages
696 -- 707. SIAM, 2015. 1.1
[20] S. Li. Approximating capacitated k-median with (1 + ǫ)k open facilities. In Proceedings of the
Twenty-Seventh Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), To Appear,
2016. 1.1
[21] S. Li and O. Svensson. Approximating k-median via pseudo-approximation. In proceedings of
the forty-fifth annual ACM symposium on theory of computing, pages 901 -- 910. ACM, 2013.
1.1
[22] J.-H. Lin and J. S. Vitter. ε-approximations with minimum packing constraint violation (ex-
tended abstract). In Proceedings of the Twenty-fourth Annual ACM Symposium on Theory of
Computing, STOC '92, pages 771 -- 782, New York, NY, USA, 1992. ACM. 3.3, 5
[23] K. C. S., B. Laekhanukit, and P. Manurangsi. On the parameterized complexity of approximat-
ing dominating set. In Proceedings of the 50th Annual ACM SIGACT Symposium on Theory
of Computing, STOC 2018, pages 1283 -- 1296, New York, NY, USA, 2018. ACM. 1.2
[24] A. Schrijver. Combinatorial Optimization - Polyhedra and Efficiency. Springer, 2003. 3.1
15
|
1204.4997 | 1 | 1204 | 2012-04-23T08:25:37 | Optimal Orthogonal Graph Drawing with Convex Bend Costs | [
"cs.DS",
"cs.DM"
] | Traditionally, the quality of orthogonal planar drawings is quantified by either the total number of bends, or the maximum number of bends per edge. However, this neglects that in typical applications, edges have varying importance. Moreover, as bend minimization over all planar embeddings is NP-hard, most approaches focus on a fixed planar embedding.
We consider the problem OptimalFlexDraw that is defined as follows. Given a planar graph G on n vertices with maximum degree 4 and for each edge e a cost function cost_e : N_0 --> R defining costs depending on the number of bends on e, compute an orthogonal drawing of G of minimum cost. Note that this optimizes over all planar embeddings of the input graphs, and the cost functions allow fine-grained control on the bends of edges.
In this generality OptimalFlexDraw is NP-hard. We show that it can be solved efficiently if 1) the cost function of each edge is convex and 2) the first bend on each edge does not cause any cost (which is a condition similar to the positive flexibility for the decision problem FlexDraw). Moreover, we show the existence of an optimal solution with at most three bends per edge except for a single edge per block (maximal biconnected component) with up to four bends. For biconnected graphs we obtain a running time of O(n T_flow(n)), where T_flow(n) denotes the time necessary to compute a minimum-cost flow in a planar flow network with multiple sources and sinks. For connected graphs that are not biconnected we need an additional factor of O(n). | cs.DS | cs |
Optimal Orthogonal Graph Drawing with Convex Bend Costs∗
Thomas Bläsius, Ignaz Rutter, Dorothea Wagner
Karlsruhe Institute of Technology (KIT)
[email protected]
Abstract
Traditionally, the quality of orthogonal planar drawings is quantified by either the total
number of bends, or the maximum number of bends per edge. However, this neglects that in
typical applications, edges have varying importance. Moreover, as bend minimization over all
planar embeddings is NP-hard, most approaches focus on a fixed planar embedding.
We consider the problem OptimalFlexDraw that is defined as follows. Given a planar
graph G on n vertices with maximum degree 4 and for each edge e a cost function coste : N0 −→ R
defining costs depending on the number of bends on e, compute an orthogonal drawing of G of
minimum cost. Note that this optimizes over all planar embeddings of the input graphs, and
the cost functions allow fine-grained control on the bends of edges.
In this generality OptimalFlexDraw is NP-hard. We show that it can be solved efficiently
if 1) the cost function of each edge is convex and 2) the first bend on each edge does not cause any
cost (which is a condition similar to the positive flexibility for the decision problem FlexDraw).
Moreover, we show the existence of an optimal solution with at most three bends per edge
except for a single edge per block (maximal biconnected component) with up to four bends. For
biconnected graphs we obtain a running time of O(n· Tflow(n)), where Tflow(n) denotes the time
necessary to compute a minimum-cost flow in a planar flow network with multiple sources and
sinks. For connected graphs that are not biconnected we need an additional factor of O(n).
Introduction
1
Orthogonal graph drawing is one of the most important techniques for the human-readable visu-
alization of complex data. Its aesthetic appeal derives from its simplicity and straightforwardness.
Since edges are required to be straight orthogonal lines -- which automatically yields good angular
resolution and short links -- the human eye may easily adapt to the flow of an edge. The readability
of orthogonal drawings can be further enhanced in the absence of crossings, that is if the underlying
data exhibits planar structure. Unfortunately, not all planar graphs have an orthogonal drawing in
which each edge may be represented by a straight horizontal or vertical line. In order to be able to
visualize all planar graphs nonetheless, we allow edges to have bends. Since bends obfuscate the
readability of orthogonal drawings, however, we are interested in minimizing the number of bends
on the edges.
In this paper we consider the problem OptimalFlexDraw whose input consists of a planar
graph G with maximum degree 4 and for each edge e a cost function coste : N0 −→ R defining costs
depending on the number of bends on e. We seek an orthogonal drawing of G with minimum cost.
Garg and Tamassia [9] show that it is NP-hard to decide whether a 4-planar graph admits an
∗Part of this work was done within GRADR -- EUROGIGA project no. 10-EuroGIGA-OP-003.
1
orthogonal drawing without any bends. Note that this directly implies that OptimalFlexDraw
is NP-hard in general. For a special case, namely planar graphs with maximum degree 3 and
series-parallel graphs, Di Battista et al. [4] give an algorithm minimizing the total number of bends
optimizing over all planar embeddings. They introduce the concept of spirality that is similar to
the rotation we use (see Section 2.3 for a definition). Bläsius et al. [2] show that the existence
of a planar 1-bend drawing can be tested efficiently. More generally, they consider the problem
FlexDraw, where each edge has a flexibility specifying its allowed number of bends. For the case
that all flexibilities are positive, they give a polynomial-time algorithm for testing the existence of
a valid drawing.
As minimizing the number of bends for 4-planar orthogonal drawings is NP-hard, many results
use the topology-shape-metrics approach initially fixing the planar embedding. Tamassia [15] de-
scribes a flow network for minimizing the number of bends. This flow network can be easily adapted
to also solve OptimalFlexDraw even for the case where the first bend may cause cost, however,
the planar embedding has to be fixed in advanced. Biedl and Kant [1] show that every plane graph
can be embedded with at most two bends per edge except for the octahedron. Morgana et al. [12]
give a characterization of plane graphs that have an orthogonal drawing with at most one bend per
edge. Tayu et al. [17] show that every series-parallel graph can be drawn with at most one bend
per edge. All these results and the algorithm we present here have the requirement of maximum
degree 4 in common. Although this is a strong restriction it is important to consider this case since
algorithms dealing with higher-degree vertices (drawing them as boxes instead of single points) rely
on algorithms for graphs with maximum degree 4 [16, 8, 11].
Even though fixing an embedding allows to efficiently minimize the total number of bends (with
this embedding), this neglects that the choice of a planar embedding may have a huge impact on
the number of bends in the resulting drawing. The result by Bläsius et al. [2] concerning the
problem FlexDraw takes this into account and additionally allows the user to control the final
drawing, for example by allowing few bends on important edges. However, if such a drawing does
not exist, the algorithm solving FlexDraw does not create a drawing at all and thus it cannot
be used in a practical application. Thus, the problem OptimalFlexDraw, which generalizes the
corresponding optimization problem, is of higher practical interest, as it allows the user to take
control of the properties of the final drawing within the set of feasible drawings. Moreover, it allows
a more fine-grained control of the resulting drawing by assigning high costs to bends on important
edges.
Contribution and Outline. Our main result is the first polynomial-time bend-optimization
algorithm for general 4-planar graphs optimizing over all embeddings. Previous work considers
only restricted graph classes and unit costs. We solve OptimalFlexDraw if 1) all cost functions
are convex and 2) the first bend is for free. We note that convexity is indeed quite natural, and
that without condition 2) OptimalFlexDraw is NP-hard, as it could be used to minimize the
total number of bends over all embeddings, which is known to be NP-hard [9].
In particular, our algorithm allows to efficiently minimize the total number of bends over all
planar embeddings, where one bend per edge is free. Note that this is an optimization version
of FlexDraw where each edges has flexibility 1, as a drawing with cost 0 exists if and only
if FlexDraw has a valid solution. Moreover, as it is known that every 4-planar graph has an
orthogonal representation with at most two bends per edge [1], our result can also be used to create
such a drawing minimizing the number of edges having two bends by setting the costs for three or
more bends to ∞.
To derive the algorithm for OptimalFlexDraw, we show the existence of an optimal solution
2
Figure 1: (a) Two parallel edges, the thin has one bend for free, every additional bend costs 1, the
thick edge has two bends for free, every additional bend costs 2. Whether embedding E1 or E2 is
better depends on the number of bends. The minimum (marked by gray boxes) yields a non-convex
cost function. (b) The non-convexity in (a) does not rely on multiple edges, the thick edge could
be replaced by the shown gadget where each edge of the gadget has one bend for free and every
additional bend costs 2. (c) This example has a non-convex cost function even if every edge has
one bend for free and each additional bend costs 1.
with at most three bends per edge except for a single edge per block with up to four bends,
confirming a conjecture of Rutter [14].
Our strategy for solving OptimalFlexDraw for biconnected graphs optimizing over all planar
embedding is the following. We use dynamic programming on the SPQR-tree of the graph, which
is a data structure representing all planar embeddings of a biconnected graph. Every node in
the SPQR-tree corresponds to a split component and we compute cost functions for these split
components determining the cost depending on how strongly the split component is bent. We
compute such a cost function from the cost functions of the children using a flow network similar
to the one described by Tamassia [15]. As computing flows with minimum cost is NP-hard for
non-convex costs we need to ensure that not only the cost functions of the edges but also the
cost functions of the split components we compute are convex. However, this is not true at all,
see Figure 1 for an example. This is not even true if every edge can have a single bend for free
and then has to pay cost 1 for every additional bend, see Figure 1(c). To solve this problem, we
essentially show that it is sufficient to compute the cost functions on the small interval [0, 3]. We
can then show that the cost functions we compute are always convex on this interval.
We start with some preliminaries in Section 2. Afterwards, we first consider the decision problem
FlexDraw for the case that the planar embedding is fixed in Section 3. In this restricted setting
we are able to prove the existence of valid drawings with special properties. Bläsius et al. [2] show
that "rigid" graphs do not exist in this setting in the sense that a drawing that is bent strongly can
be unwound under the assumption that the flexibility of every edge is at least 1. In other words this
shows that graphs with positive flexibility behave similar to single edges with positive flexibility.
We present a more elegant proof yielding a stronger result that can then be used to reduce the
number of bends of every edge down to three (at least for biconnected graphs and except for a single
edge on the outer face). In Section 4 we extend the term "bends", originally defined for edges, to
split components and show that in a biconnected graph the split components corresponding to the
nodes in its SPQR-tree can be assumed to have only up to three bends. In Section 5 we show that
these results for the decision problem FlexDraw can be extended to the optimization problem
OptimalFlexDraw. With this result we are able to drop the fixed planar embedding (Section 6).
We first consider biconnected graphs in Section 6.1 and compute cost functions on the interval [0, 3],
3
notconvex012345123456E1E2cost00bends1234502581236(a)(b)(c) which can be shown to be convex on that interval, bottom up in the SPQR-tree. In Section 6.2 we
extend this result to connected graphs using the BC-tree (see Section 2.2 for a definition).
2 Preliminaries
In this section we introduce some notations and preliminaries.
2.1 FlexDraw
The original FlexDraw problem asks for a given 4-planar graph G = (V, E) with a function
flex: E −→ N0 ∪ {∞} assigning a flexibility to every edge whether an orthogonal drawing of
G exists such that every edge e ∈ E has at most flex(e) bends. Such a drawing is called a
valid drawing of the FlexDraw instance. The problem OptimalFlexDraw is the optimization
problem corresponding to the decision problem FlexDraw and is defined as follows. Let G =
(V, E) be a 4-planar graph together with a cost function coste : N0 −→ R ∪ {∞} associated with
every edge e ∈ E having the interpretation that ρ bends on the edge e cause coste(ρ) cost. Then the
cost of an orthogonal drawing of G is the total cost summing over all edges. A drawing is optimal if
it has the minimum cost among all orthogonal drawings of G. The task of the optimization problem
OptimalFlexDraw is to find an optimal drawing of G.
Since OptimalFlexDraw contains the NP-hard problem FlexDraw, it is NP-hard itself.
However, FlexDraw is efficiently solvable for instances with positive flexibility, that is instances in
which the flexibility of every edge is at least 1. To obtain a similar result for OptimalFlexDraw
we have to restrict the possible cost functions slightly.
For a cost function coste(·) we define the difference function ∆ coste(·) to be ∆ coste(ρ) =
coste(ρ + 1) − coste(ρ). A cost function is monotone if its difference function is greater or equal
to 0. We say that the base cost of the edge e with monotone cost function is be = coste(0). The
flexibility of an edge e with monotone cost function is defined to be the largest possible number
of bends ρ for which coste(ρ) = be. As before, we say that an instance G of OptimalFlexDraw
has positive flexibility if all cost functions are monotone and the flexibility of every edge is positive.
Unfortunately, we have to restrict the cost functions further to be able to solve OptimalFlexDraw
efficiently. The cost function coste(·) is convex, if its difference function is monotone. We call an
instance of OptimalFlexDraw convex, if every edge has positive flexibility and each cost function
is convex. Note that this includes that the cost functions are monotone. We provide an efficient
algorithm solving OptimalFlexDraw for convex instances.
2.2 Connectivity, BC-Tree and SPQR-Tree
A graph is connected if there exists a path between any pair of vertices. A separating k-set is a set
of k vertices whose removal disconnects the graph. Separating 1-sets and 2-sets are cutvertices and
separation pairs, respectively. A connected graph is biconnected if it does not have a cut vertex
and triconnected if it does not have a separation pair. The maximal biconnected components of a
graph are called blocks. The cut components with respect to a separation k-set S are the maximal
subgraphs that are not disconnected by removing S.
The block-cutvertex tree (BC-tree) B of a connected graph is a tree whose nodes are the blocks
and cutvertices of the graph, called B-nodes and C-nodes, respectively. In the BC-tree a block B
and a cutvertex v are joined by an edge if v belongs to B. If an embedding is chosen for each block,
these embeddings can be combined to an embedding of the whole graph if and only if B can be
4
Figure 2: The unrooted SPQR-tree of a biconnected planar graph. The nodes µ1, µ3 and µ5 are
P-nodes, µ2 is an R-node and µ4 is an S-node. The Q-nodes are not shown explicitely.
rooted at a B-node such that the parent of every other block B in B, which is a cutvertex, lies on
the outer face of B.
We use the SPQR-tree introduced by Di Battista and Tamassia [5, 6] to represent all planar
embeddings of a biconnected planar graph G. The SPQR-tree T of G is a decomposition of G into
its triconnected components along its split pairs where a split pair is either a separation pair or an
edge. We first define the SPQR-tree to be unrooted, representing embeddings on the sphere, that
is planar embeddings without a designated outer face. Let {s, t} be a split pair and let H1 and H2
be two subgraphs of G such that H1 ∪ H2 = G and H1 ∩ H2 = {s, t}. Consider the tree containing
the two nodes µ1 and µ2 associated with the graphs H1 +{s, t} and H2 +{s, t}, respectively. These
graphs are called skeletons of the nodes µi, denoted by skel(µi) and the special edge {s, t} is said
to be a virtual edge. The two nodes µ1 and µ2 are connected by an edge, or more precisely, the
occurrence of the virtual edges {s, t} in both skeletons are linked by this edge. Now a combinatorial
embedding of G uniquely induces a combinatorial embedding of skel(µ1) and skel(µ2). Furthermore,
arbitrary and independently chosen embeddings for the two skeletons determine an embedding of
G, thus the resulting tree can be used to represent all embeddings of G by the combination of all
embeddings of two smaller planar graphs. This replacement can of course be applied iteratively
to the skeletons yielding a tree with more nodes but smaller skeletons associated with the nodes.
Applying this kind of decomposition in a systematic way yields the SPQR-tree as introduced by
Di Battista and Tamassia [5, 6]. The SPQR-tree T of a biconnected planar graph G contains four
types of nodes. First, the P-nodes having a bundle of at least three parallel edges as skeleton
and a combinatorial embedding is given by any ordering of these edges. Second, the skeleton of
an R-node is triconnected, thus having exactly two embeddings [18], and third, S-nodes have a
simple cycle as skeleton without any choice for the embedding. Finally, every edge in a skeleton
representing only a single edge in the original graph G is formally also considered to be a virtual
edge linked to a Q-node in T representing this single edge. Note that all leaves of the SPQR-tree T
are Q-nodes. Besides from being a nice way to represent all embeddings of a biconnected planar
graph, the SPQR-tree has only size linear in G and Gutwenger and Mutzel [10] showed how to
compute it in linear time. Figure 2 shows a biconnected planar graph together with its SPQR-tree.
Often the SPQR-tree T of a biconnected planar graph G is assumed to be rooted in a Q-node
representing all planar embeddings with the corresponding edge on the outer face. In contrast to
previous results, we assume the SPQR-tree T to be rooted in some node τ, which may be a Q-node
or an inner node. In the following we describe the interpretation of the SPQR-tree with root τ.
Every node µ, apart form τ itself, has a unique parent and thus its skeleton skel(µ) contains a
virtual edge corresponding to this parent. We refer to this virtual edge as the parent edge. A
planar embedding E of G is represented by T with root τ if the embedding induced on the skeleton
skel(µ) of every node µ 6= τ has the parent edge on the outer face. The embedding of skel(τ) is not
restricted, thus the choice of the outer face makes a difference for the root.
5
µ2v1v2v3v4v5abcdefghijv5v2ijv5v2v1dv1v2ev1v4v4v1v2v3bcafghµ5µ4µ3µ1 For every node µ in the SPQR-tree T apart from the root τ we define the pertinent graph of µ,
denoted by pert(µ), as follows. The pertinent graph of a Q-node is the edge associated to it. The
pertinent graph of an inner node µ is recursively defined to be the graph obtained by replacing
all virtual edges apart from the parent edge by the pertinent graphs of the corresponding children
in T . The expansion graph of a virtual edge ε in skel(µ) is the pertinent graph of µ0 where µ0 is
the child of µ corresponding to the virtual edge ε with respect to the root µ.
2.3 Orthogonal Representation
Two orthogonal drawings of a 4-planar graph G are equivalent, if they have the same topology, that
is the same planar embedding, and the same shape in the sense that the sequence of right and left
turns is the same in both drawings when traversing the faces of G. To make this precis, we define
orthogonal representations, originally introduced by Tamassia [15], as equivalence classes of this
equivalence relation between orthogonal drawings. To ease the notation we first only consider the
biconnected case.
Let Γ be an orthogonal drawing of a biconnected 4-planar graph G. In the planar embedding E
induced by Γ every edge e is incident to two different faces, let f be one of them. When traversing
f in clockwise order (counter-clockwise if f is the outer face) e may have some bends to the right
and some bends to the left. We define the rotation of e in the face f to be the number of bends to
the right minus the number of bends to the left and denote the resulting value by rot(ef). Similarly,
every vertex v is incident to several faces, let f be one of them. Then we define the rotation of v
in f, denoted by rot(vf), to be 1, −1 and 0 if there is a turn to the right, a turn to the left and
no turn, respectively, when traversing f in clockwise direction (counter-clockwise if f is the outer
face). The orthogonal representation R belonging to Γ consists of the planar embedding E of G
and all rotation values of edges and vertices, respectively. It is easy to see that every orthogonal
representation has the following properties.
(I) For every edge e incident to the faces f1 and f2 the equation rot(ef1) = − rot(ef2) holds.
(II) The sum over all rotations in a face is 4 for inner faces and −4 for the outer face.
(III) The sum of rotations around a vertex v is 2 · (deg(v) − 2).
Tamassia showed that the converse is also true [15], that is R is an orthogonal representation
representing a class of orthogonal drawings if the rotation values satisfy the above properties. He
moreover describes a flow network such that every flow in the flow network corresponds to an
orthogonal representation. A modification of this flow network can also be used to solve Op-
timalFlexDraw but only for the case that the planar embedding is fixed.
In some cases we
also write rotR(·) instead of rot(·) to make clear to which orthogonal representation we refer to.
Moreover, the face in the index is sometimes omitted if it is clear which face is meant.
When extending the term orthogonal representation to not necessarily biconnected graphs there
are two differences. First, a vertex v with deg(v) = 1 may exist. Then v is incident to a single face
f and we define the rotation rot(vf) to be −2. Note that the rotations around every vertex v still
sum up to 2·(deg(v)−2). The second difference is that the notation introduced above is ambiguous
since edges and vertices may occur several times in the boundary of the same face. For example
a bridge e is incident to the face f twice, thus it is not clear which rotation is meant by rot(ef).
However, it will always be clear from the context, which incidence to the face f is meant by the
index f. Thus, we use for connected graphs the same notation as for biconnected graphs.
Let G be a 4-planar graph with orthogonal representation R and two vertices s and t incident
to a common face f. We define πf(s, t) to be the unique shortest path from s to t on the boundary
of f, when traversing f in clockwise direction (counter-clockwise if f is the outer face). Let
6
Figure 3: On the left three tight orthogonal drawings are stacked together. This is not possible on
the right side, since the black vertices have angles larger than 90◦ in internal faces.
s = v1, . . . , vk = t be the vertices on the path πf(s, t). The rotation of π(s, t) is defined as
rot(π(s, t)) =
rot({vi, vi+1}) +
rot(vi) ,
k−1X
i=1
k−1X
i=2
where all rotations are with respect to the face f.
Note that it does not depend on the particular drawing of a graph G how many bends each edge
has but only on the orthogonal representation. Thus we can continue searching for valid and optimal
orthogonal representations instead of drawings to solve FlexDraw and OptimalFlexDraw,
respectively.
Let G be a 4-planar graph with positive flexibility and valid orthogonal representation R and
let {s, t} be a split pair. Let further H be a split component with respect to {s, t} such that the
orthogonal representation S of H induced by R has {s, t} on the outer face f. The orthogonal
representation S of H is called tight with respect to the vertices s and t if the rotations of s and
t in internal faces are 1, that is s and t form 90◦-angles in internal faces of H. Bläsius et al. [2,
Lemma 2] show that S can be made tight with respect to s and t, that is there exists a valid tight
orthogonal representation of H that is tight. Moreover, this tight orthogonal representation can
be plugged back into the orthogonal representation of the whole graph G. We call an orthogonal
representation R of the whole graph G tight, if every split component having the corresponding
split pair on its outer face is tight with respect to its split pair. It follows that we can assume
without loss of generality that every valid orthogonal representation is tight. This has two major
advantages. First, if we have for example a chain of graphs and orthogonal representations of
each graph in the chain, we can combine these orthogonal representations by simply stacking them
together; see Figure 3. Note that this may not be possible if the orthogonal representations are not
tight. Second, the shape of the outer face f of a split component with split pair {s, t} is completely
determined by the rotation of πf(s, t) and the degrees of s and t, since the rotation at the vertices s
and t in the outer face only depends on their degrees. In the following we assume every orthogonal
representation to be tight.
2.4 Flow Network
A cost flow network (or flow network for short) is a tuple N = (V, A, COST, dem) where (V, A) is
a directed (multi-)graph, COST is a set containing a cost function costa : N0 −→ R ∪ {∞} for each
arc a ∈ A and dem: V −→ Z is the demand of the vertices. A flow in N is a function φ: A −→ N0
assigning a certain amount of flow to each arc. A flow φ is feasible, if the difference of incoming
and outgoing flow at each vertex equals its demand, that is
The cost of a given flow φ is the total cost of the arcs caused by the flow φ, that is
φ(v, u) for all v ∈ V.
dem(v) = X
(u,v)∈A
φ(u, v) − X
cost(φ) = X
(v,u)∈A
costa(φ(a)).
a∈A
7
? A feasible flow φ in N is called optimal if cost(φ) ≤ cost(φ0) holds for every feasible flow φ0.
If the cost function of an arc a is 0 on an interval [0, c] and ∞ on (c,∞), we say that a has
A flow network N is called convex if the cost functions on its arcs are convex.
capacity c.
In the flow
networks we consider, every arc a ∈ A has a corresponding arc a0 ∈ A between the same vertices
pointing in the opposite direction. A flow φ is normalized if φ(a) = 0 or φ(a0) = 0 for each of these
pairs. Since we only consider convex flow networks a normalized optimal flow does always exist.
Thus we assume without loss of generality that all flows are normalized. We simplify the notation
as follows. If we talk about an amount of flow on the arc a that is negative, we instead mean the
same positive amount of flow on the opposite arc a0. In many cases minimum-cost flow networks are
only considered for linear cost functions, that is each unit of flow on an arc causes a constant cost
defined for that arc. Note that the cost functions in a convex flow network N are piecewise linear
and convex according to our definition. Thus, it can be easily formulated as a flow network with
linear costs by splitting every arc into multiple arcs, each having linear costs. It is well known that
flow networks of this kind can be solved in polynomial time. The best known running time depends
on additional properties that N may satisfy. We use an algorithm computing a minimum-cost flow
in the network N as black box and denote the necessary running time by Tflow(N). In Section 6.3
we have a closer look on which algorithm to use.
Let u, v ∈ V be two nodes of the convex flow network N with demands dem(u) and dem(v).
The parameterized flow network with respect to the nodes u and v is defined the same as N but
with a parameterized demand of dem(u)− ρ for u and dem(v)+ ρ for v where ρ is a parameter. The
cost function costN(ρ) of the parameterized flow network N is defined to be cost(φ) of an optimal
flow φ in N with respect to the parameterized demands determined by ρ. Note that increasing
ρ by 1 can be seen as pushing one unit of flow from u to v. We define the optimal parameter ρ0
to be the parameter for which the cost function is minimal among all possible parameters. The
correctness of the minimum weight path augmentation method to compute flows with minimum
costs implies the following theorem [7].
Theorem 1. The cost function of a parameterized flow network is convex on the interval [ρ0,∞],
where ρ0 is the optimal parameter.
Proof. Let N = (V, A, COST, dem) be a parameterized flow network and let φ0 be a minimum-cost
flow in N with respect to the optimal parameter ρ0. To simplify notation, we assume ρ0 = 0. The
residual network R0 with respect to φ0 is the graph (V, A) with a constant cost cost0(a) assigned to
every arc a such that cost0(a) is the amount of cost in N that has to be payed to push an additional
unit of flow along a, with respect to the given flow φ0. Note that this cost may be negative. It is
well known that an optimal flow φ1 with respect to the parameter 1 can be computed by pushing
one unit of flow along a path from u to v with minimum weight in R0 [7]. Moreover, we can
continue and compute an optimal flow φk+1 by augmenting φk along a minimum weight path in
the residual network Rk with respect to the flow φk. Assume we augment φk along the path πk
causing cost costk(πk) to obtain an optimal flow φk+1 with respect to the parameter k + 1 and
then we augment along a path πk+1 in Rk+1 with cost costk+1(πk+1) to obtain an optimal flow
φk+2 with respect to the parameter k + 2. To obtain the claimed convexity we have to show that
costk(πk) ≤ costk+1(πk+1) holds.
If πk and πk+1 contain an arc a in the same direction, then costk(a) ≤ costk+1(a) holds by the
convexity of the cost function of a. If πk contains the arc a and πk+1 contains the arc a0 in the
opposite direction then costk(a) = − costk+1(a0) holds. Assume πk and πk+1 share such an arc in
the opposite direction. Then we remove this arc in both directions, splitting each of the paths πk
and πk+1 into two subpaths. We define two new paths π and π0 by concatenating the first part of πk
8
Figure 4: Since a strictly directed path from t to s has a lower bound for its rotation this yields
upper bounds for paths from s to t (Lemma 1).
with the second part of πk+1 and vice versa, respectively. This can be done iteratively, thus we can
assume that π and π0 do not share arcs in the opposite direction. We consider the cost of π and π0 in
the residual network Rk. Obviously, for an arc a that is exclusively contained either in π or in π0 we
have costk(a) = costk+1(a). For an arc that is contained in π and π0 we have costk(a) ≤ costk+1(a).
Moreover, for every pair of arcs a and a0 that was removed we have costk(a) = − costk+1(a0). This
yields the inequality costk(πk) + costk+1(πk+1) ≥ costk(π) + costk(π0). Since πk was a path with
smallest possible weight in Rk we have costk(πk) ≤ costk(π) and costk(πk) ≤ costk(π0). With the
above inequality this yields costk+1(πk+1) ≥ costk(πk).
3 Valid Drawings with Fixed Planar Embedding
In this section we consider the problem FlexDraw for the case that the planar embedding is fixed.
We show that the existence of a valid orthogonal representation implies the existence of a valid
orthogonal representation with special properties. We first show the following. Given a biconnected
4-planar graph with positive flexibility and an orthogonal representation R such that two vertices s
and t lie on the outer face f, then the rotation along πf(s, t) can be reduced by 1 if it is at least 0.
This result is a key observation for the algorithm solving the decision problem FlexDraw [2]. It
in a sense shows that "rigid" graphs that have to bent strongly do not exists. This kind of graphs
play an important role in the N P-hardness proof of 0-embeddability by Garg and Tamassia [9].
Moreover, we show the existence of a valid orthogonal representation R0 inducing the same planar
embedding and having the same angles around vertices as R such that every edge has at most three
bends in R0, except for a single edge on the outer face with up to five bends. If we allow to change
the embedding slightly, this special edge has only up to four bends.
Let G be a 4-planar graph with positive flexibility and valid orthogonal representation R, and
let e be an edge. If the number of bends of e equals its flexibility, we orient e such that its bends
are right bends. Otherwise, e remains undirected. We define a path π = (v1, . . . , vk) in G to be a
directed path, if the edge {vi, vi+1} (for i ∈ {1, . . . , k − 1}) is either undirected or directed from vi
to vi+1. A path containing only undirected edges can be seen as directed path for both possible
directions. The path π is strictly directed, if it is directed and does not contain undirected edges.
These terms directly extend to (strictly) directed cycles. Given a (strictly) directed cycle C the
terms left(C) and right(C) denote the set of edges and vertices of G lying to the left and right of C,
respectively, with respect to the orientation of C. A cut (U, V \ U) is said to be directed from U to
V \ U, if every edge {u, v} with u ∈ U and v ∈ V \ U is either directed from u to v or undirected.
According to the above definitions a cut is strictly directed from U to V \ U if it is directed and
contains no undirected edges. Before we show how to unwind an orthogonal representation that is
bent strongly we need the following technical lemma.
Lemma 1. Let G be a graph with positive flexibility and vertices s and t such that G + st is
biconnected and 4-planar. Let further R be a valid orthogonal representation with s and t incident
9
stπ(s,t)π(t,s)ststπ(s,t)π(t,s)(a)(b)(c)(d)(e)stststt0t0(f) to the common face f such that πf(t, s) is strictly directed from t to s. Then the following holds.
(1) rotR(πf(s, t)) ≤ −3 if f is the outer face and G does not consist of a single path
(2) rotR(πf(s, t)) ≤ −1 if f is the outer face
(3) rotR(πf(s, t)) ≤ 5
Proof. We first consider the case where f is the outer face (Figure 4(a)), that is cases (1) and (2).
Due to the fact that πf(t, s) is strictly directed from t to s and the flexibility of every edge is positive,
each edge on πf(t, s) has rotation at least 1. Moreover, the rotations at vertices along the path
πf(t, s) are at least −1 since πf(t, s) is simple as G+ st is biconnected. Since the number of internal
vertices on a path is one less than the number of edges this yields rot(πf(t, s)) ≥ 1; see Figure 4(b).
If G consists of a single path this directly yields rot(πf(s, t)) ≤ −1 and thus concludes case (2). For
case (1) first assume that the degrees of s and t are not 1 (Figure 4(b)), that is rot(sf), rot(tf) ∈
{−1, 0, 1} holds. Since f is the outer face the equation rot(πf(s, t))+rot(tf)+rot(πf(t, s))+rot(sf) =
−4 holds and directly implies the desired inequality rot(πf(s, t)) ≤ −3.
In the case that for
example t has degree 1 (and deg(s) > 0), we have rot(tf) = −2 and rot(sf) ∈ {−1, 0, 1}, thus the
considerations above only yield rot(πf(s, t)) ≤ −2. However, in this case there necessarily exists a
vertex t0 where the paths πf(s, t) and πf(t, s) split, as illustrated in Figure 4(c). More precisely,
let t0 be the first vertex on πf(s, t) that also belongs to πf(t, s). Obviously, the degree of t0 is
at least 3 and thus rot(t0
f) (with respect to the path πf(t, s)) is at least 0. Hence we obtain the
stronger inequality rot(πf(t, s)) ≥ 2 yielding the desired inequality rot(πf(s, t)) ≤ −3. If s and t
both have degree 1 we cannot only find the vertex t0 but also the vertex s0 where the paths πf(s, t)
and πf(t, s) split. Since G + st is biconnected these two vertices are distinct and the estimation
above works, finally yielding rot(πf(s, t)) ≤ −3.
If f is an internal face (Figure 4(d)), that is case (3) applies, we start with the equation
rot(πf(s, t)) + rot(tf) + rot(πf(t, s)) + rot(sf) = 4. First we consider the case that neither t nor s
have degree 1. Thus, rot(tf), rot(sf) ∈ {−1, 0, 1}. With the same argument as above we obtain
rot(πf(t, s)) ≥ 1 and hence rot(πf(s, t)) ≤ 5; see Figure 4(e). Now assume that t has degree 1 and s
has larger degree. Then rot(tf) = −2 holds and the above estimation does not work anymore.
Again, at some vertex t0 the paths πf(t, s) and πf(s, t) split as illustrated in Figure 4(f). Obviously,
f) is at least 0. This yields rot(πf(t, s)) ≥ 2
the degree of t0 needs to be greater than 2 and thus rot(t0
in the case that deg(t) = 1, compensating rot(tf) = −2 (instead of rot(tf) ≥ −1 in the other case).
To sum up, we obtain the desired inequality rot(πf(s, t)) ≤ 5. The case deg(s) = deg(t) = 1 works
analogously.
The flex graph G×
R of G with respect to a valid orthogonal representation R is defined to be the
dual graph of G such that the dual edge e? is undirected if e is undirected, otherwise it is directed
from the face right of e to the face left of e. Figure 5(a) shows an example graph with an orthogonal
drawing together with the corresponding flex graph. Assume we have a simple directed cycle C
in the flex graph. Then bending along this cycle yields a new valid orthogonal representation R0
which is defined as follows. Let e? = (f1, f2) be an edge contained in C dual to e. Then we decrease
rot(ef1) and increase rot(ef2) by 1. It can be easily seen that the necessary properties for R0 to be
an orthogonal representation are satisfied. Obviously, rotR0(ef1) = − rotR0(ef2) holds and rotations
at vertices did not change. Moreover, the rotation around a face f does not change since f is either
not contained in C or it is contained in C, but then it has exactly one incoming and exactly one
outgoing edge. Note that bending along a cycle in the flex graph preserves the planar embedding
of G and for every vertex the rotations in all incident faces. The following lemma shows that a
high rotation along a path πf(s, t) for two vertices s and t sharing the face f can be reduced by 1
using a directed cycle in the flex graph.
10
Figure 5: (a) An orthogonal representation and the corresponding flex graph where every edge has
flexibility 1. (b, c, d) Illustration of Lemma 2.
Lemma 2. Let G be a biconnected 4-planar graph with positive flexibility, a valid orthogonal rep-
resentation R and s and t on a common face f. The flex graph G×
R contains a directed cycle C
such that f ∈ C, s ∈ left(C) and t ∈ right(C), if one of the following conditions holds.
(1) rotR(πf(s, t)) ≥ −2, f is the outer face and πf(s, t) is not strictly directed from t to s
(2) rotR(πf(s, t)) ≥ 0 and f is the outer face
(3) rotR(πf(s, t)) ≥ 6
Proof. Figure 5(b) shows the path πf(s, t) together with the desired cycle C. Due to the duality
of a cycle in the dual and a cut in the primal graph a directed cycle C in G×
R having s and t to
the left and to the right of C, respectively, induces a directed cut in G that is directed from s to
t and vice versa. Recall that directed cycles and cuts may also contain undirected edges. Assume
for contradiction that such a cycle C does not exist.
Claim 1. The graph G contains a strictly directed path π from t to s.
Every cut (S, T) with T = V \ S, s ∈ S and t ∈ T separating s from t must contain an edge that is
directed from T to S, otherwise this cut would correspond to a cycle C in the flex graph that does
not exist by assumption. Let T be the set of vertices in G that can be reached by strictly directed
paths from t. If T contains s we found the path π strictly directed from t to s. Otherwise, (S, T)
with S = V \ T is a cut separating S from T and there cannot be an edge that is directed from a
vertex in T to a vertex in S which is a contradiction, and thus the path π strictly directed from t
to s exists, which concludes the proof of the claim.
Let G0 be the subgraph of G induced by the paths π and πf(s, t) together with the orthogonal
representation R0 induced by R.
We first consider case (1). Let f0 be the outer face of the orthogonal representation R0. Obvi-
ously, πf0(s, t) = πf(s, t) and π = πf0(t, s) holds, see Figure 5(c). Moreover, the graph G0 + st is
biconnected and G0 does not consist of a single path since πf0(s, t) and πf0(t, s) are different due to
the assumption that πf(s, t) is not strictly directed from t to s. Since πf0(t, s) is strictly directed
from t to s we can use Lemma 1(1) yielding rotR0(πf0(s, t)) ≤ −3 and thus rotR(πf(s, t)) ≤ −3,
which is a contradiction.
For case (2) exactly the same argument holds except for the case where the strictly directed
path π is the path πf(s, t) strictly directed from t to s. In this case we have to use Lemma 1(2)
instead of Lemma 1(1) yielding rotR(πf(s, t)) ≤ −1, which is again a contradiction.
In case (3) the subgraph G0 of G induced by the two paths π and πf(s, t) again contains s and t
on a common face f0, which may be the outer or an inner face, see Figure 5(c) and Figure 5(d),
respectively. In both cases we obtain rotR(πf(s, t)) ≤ 5 due to Lemma 1(3), which is a contradiction.
Lemma 2 directly yields the following corollary, showing that graphs with positive flexibility
behave very similar to single edges with positive flexibility.
11
π(s,t)fCststπ(s,t)π=π(t,s)stπ=π(t,s)π(s,t)(a)(b)(c)(d)f0f0 Corollary 1. Let G be a graph with positive flexibility and vertices s and t such that G + st is
biconnected and 4-planar. Let further R be a valid orthogonal representation with s and t on the
outer face f such that ρ = rotR(πf(s, t)) ≥ 0. For every rotation ρ0 ∈ [−1, ρ] there exists a valid
orthogonal representation R0 with rotR0(πf(s, t)) = ρ0.
Proof. For the case that G itself is biconnected, the claim follows directly from Lemma 2(2), since we
can reduce the rotation along πf(s, t) stepwise by 1, starting with the orthogonal representation R,
until we reach a rotation of −1. For the case that G itself is not biconnected we add the edge {s, t}
to the orthogonal representation R such that the path πf(s, t) does not change, that is πf(t, s)
consists of the new edge {s, t}. Again Lemma 2(2) can be used to reduce the rotation stepwise
down to −1.
As edges with many bends imply the existence of paths with high rotation, we can use Lemma 2
to successively reduce the number of bends of every edge down to three, except for a single edge
on the outer face. Since we only bend along cycles in the flex graph, neither the embedding nor
the angles around vertices are changed.
Theorem 2. Let G be a biconnected 4-planar graph with positive flexibility, having a valid orthogo-
nal representation. Then G has a valid orthogonal representation with the same planar embedding,
the same angles around vertices and at most three bends per edge, except for at most one edge on
the outer face with up to five bends.
Proof. In the following we essentially pick an edge with more than three bends, reduce the number
of bends by one and continue with the next edge. After each of these reduction steps we set the
flexibility of every edge down to max{ρ, 1}, where ρ is the number of bends it currently has. This
ensures that in the next step the number of bends of each edge either is decreased, remains as it is
or is increased from zero to one.
We start with an edge e = {s, t} that is incident to two faces f1 and f2 and has more than
three bends. Due to the fact that we traverse inner faces in clockwise and the outer face in counter-
clockwise direction, the edge e forms in one of the two faces the path from s to t and in the other face
the path from t to s. Assume without loss of generality that πf1(t, s) and πf2(s, t) are the paths on
the boundary of f1 and f2, respectively, that consist of e. Note that rot(πf1(t, s)) = − rot(πf2(s, t))
holds and we assume that rot(πf1(t, s)) is not positive. As e was assumed to have more than three
bends, the inequality rot(πf1(t, s)) ≤ −4 holds. We distinguish between the two cases that f1 is an
inner or the outer face. We first consider the case that f1 is an inner face; Figure 6(a) illustrates
this situation for the case where e has four bends. Then the rotations around the face f1 sum up
to 4. As the rotations at the vertices s and t can be at most 1, we obtain rot(πf1(s, t)) ≥ 6. Thus
we can apply Lemma 2(3) to reduce the rotation of πf1(s, t) by bending along a cycle in the flex
graph that contains f1 and separates s from t. Obviously, this increases the rotation along πf1(t, s)
by 1 and thus reduces the number of bends of e by 1.
For the case that f1 is the outer face we first ignore the case where e has four or five bends and
show how to reduce the number of bends to five; Figure 6(b) shows the case where e has six bends.
Thus the inequality rot(πf1(t, s)) ≤ −6 holds. As the rotations around the outer face f1 sum up
to −4 and the rotations at the vertices s and t are at most 1, the rotation along πf1(s, t) must be
at least 0. Thus we can apply Lemma 2(2) to reduce the rotation of πf1(s, t) by 1, increasing the
rotation along πf1(t, s), and thus reducing the number of bends of e by one.
Finally, we obtain an orthogonal representation having at most three bends per edge except for
some edges on the outer face with four or five bends having their negative rotation in the outer face.
If there is only one of these edges left we are done. Otherwise let e = {s, t} be one of the edges with
12
Figure 6: Reducing the number of bends on edges (Theorem 2)
rot(πf(t, s)) ∈ {−5,−4}, where f is the outer face. Then the inequality rot(πf(s, t)) ≥ −2 holds by
the same argument as before and we can apply Lemma 2(1) to reduce the rotation, if we can ensure
that πf(s, t) is not strictly directed from t to s. To show that, we make use of the fact that πf(s, t)
contains an edge e0 = {u, v} with at least four bends due to the assumption that e was not the
only edge with more than three bends. Assume without loss of generality that u occurs before v
on πf(s, t), thus πf(s, t) splits into the three parts πf(s, u), πf(u, v) and πf(v, t). Recall that
rot(πf(s, t)) ≥ −2 holds and thus rot(πf(s, u))+rot(u)+rot(πf(u, v))+rot(v)+rot(πf(v, t)) ≥ −2.
As the rotation at the vertices u and v is at most 1 and the rotation of πf(u, v) at most −4 it
follows that rot(πf(s, u)) + rot(πf(v, t)) ≥ 0. Figure 6(c) illustrates the situation for the case
where e and e0 have four bends and rot(πf(s, u)) = rot(πf(v, t)) = 0. Note that at least one of the
two paths is not degenerate in the sense that s 6= u or v 6= t, otherwise the total rotation around the
outer face would be at most −6, which is a contradiction. Assume without loss of generality that
rot(πf(s, u)) ≥ 0. It follows that πf(s, u) cannot be strictly directed from u to s and since πf(s, u)
is a subpath of πf(s, t) the path πf(s, t) cannot be strictly directed from t to s. This finally shows
that we can use part (1) of Lemma 2 implying that we can find a valid orthogonal representation
such that at most a single edge with four or five bends remains, whereas all other edges have at
most three bends.
If we allow the embedding to be changed slightly, we obtain an even stronger result. Assume
the edge e lying on the outer face has more than three bends. If e has five bends, we can reroute
it in the opposite direction around the rest of the graph, that is we can choose the internal face
incident to e to be the new outer face. In the resulting drawing e has obviously only three bends.
Thus the following result directly follows from Theorem 2.
Corollary 2. Let G be a biconnected 4-planar graph with positive flexibility having a valid orthogo-
nal representation. Then G has a valid orthogonal representation with at most three bends per edge
except for possibly a single edge on the outer face with four bends.
Note that Corollary 2 is restricted to biconnected graphs. For general graphs it implies that
each block contains at most a single edge with up to four bends. Figure 7 illustrates an instance
of FlexDraw with linearly many blocks and linearly many edges that are required to have four
bends, showing that Corollary 2 is tight.
Theorem 2 implies that it is sufficient to consider the flexibility of every edge to be at most 5,
or in terms of costs we want to optimize, it is sufficient to store the cost function of an edge only
in the interval [0, 5]. However, there are two reasons why we need a stronger result. First, we want
to compute cost functions of split components and thus we have to limit the number of "bends"
they can have (see the next section for a precise definition of bends for split components). Second,
as mentioned in the introduction (see Figure 1) the cost function of a split component may already
be non-convex on the interval [0, 5]. Fortunately, the second reason is not really a problem since
13
f1steπf1(s,t)πf1(t,s)(a)(b)stf1eπf1(t,s)πf1(s,t)stuvπ(t,s)π(u,v)π(s,u)π(v,t)ee0(c) Figure 7: An instance of FlexDraw requireing linearly many edges to have four bends. Flexibilites
are 1 except for the thick edges with flexibility 4.
there may be at most a single edge with up to five bends, all remaining edges have at most three
bends and thus we only need to consider their cost functions on the interval [0, 3].
In the following section we focus on dealing with the first problem and strengthen the results so
far presented by extending the limitation on the number of bends to split components. Note that
a split pair inside an inner face of G with a split component H having a rotation less than −3 on
its outer face implies a rotation of at least 6 in some inner face of G. Thus, we can again apply
Lemma 2(3) to reduce the rotation showing that split components and single edges can be handled
similarly. However, by reducing the rotation for one split component, we cannot avoid that the
rotation of some other split component is increased. For single edges we did that by reducing the
flexibility to the current number of bends. In the following section we extend this technique by
defining a flexibility not only for edges but also for split components. We essentially show that all
results we presented so far still apply, if we allow this kind of extended flexibilities.
4 Flexibility of Split Components and Nice Drawings
Let G be a biconnected 4-planar graph with SPQR-tree T and let T be rooted at some node τ.
Recall that we do not require τ to be a Q-node. Let µ be a node of T that is not the root τ. Then µ
has a unique parent and skel(µ) contains a unique virtual edge ε = {s, t} that is associated with
this parent. We call the split-pair {s, t} a principal split pair and the pertinent graph pert(µ) with
respect to the chosen root a principal split component. The vertices s and t are the poles of this
split component. Note that a single edge is also a principal split component except for the case
that its Q-node is chosen to be the root. A planar embedding of G is represented by T with the
root τ if the embedding of each skeleton has the edge associated with the parent on the outer face.
Let R be a valid orthogonal representation of G such that the planar embedding of R is
represented by T rooted at τ. Consider a principal split component H with respect to the split
pair {s, t} and let S be the orthogonal representation of H induced by R. Note that the poles
s and t are on the outer face f of S. We define max{ rotS(πf(s, t)), rotS(πf(t, s))} to be the
number of bends of the split component H. Note that this is a straightforward extension of the
term bends as it is used for edges. With this terminology we can assign a flexibility flex(H) to
a principal split component H and we define the orthogonal representation R of G to be valid if
and only if H has at most flex(H) bends. We say that the graph G has positive flexibility if the
flexibility of every principal split component is at least 1, which is straightforward extension of the
original notion.
We define a valid orthogonal representation of G to be nice if it is tight and if there is a root τ
of the SPQR-tree such that every principal split component has at most three bends and the edge
corresponding to τ in the case that τ is a Q-node has at most five bends. The main result of this
section will be the following theorem, which directly extends Theorem 2.
Theorem 3. Every biconnected 4-planar graph with positive flexibility having a valid orthogonal
representation has an orthogonal representation with the same planar embedding and the same
14
Figure 8: Augmentation of G by the safety edges eH(s, t) and eH(t, s).
angles around vertices that is nice with respect to at least one node chosen as root of its SPQR-tree.
Before we prove Theorem 3 we need to make some additional considerations. In particular we
need to extend the flex-graph such that it takes the flexibilities of principal split components into
account. The extended version of the flex graph can then be used to obtain a result similar to
Lemma 2, which was the main tool to proof Theorem 2. Another difficulty is that it depends on
the chosen root which split components are principal split components. For the moment we avoid
this problem by choosing an arbitrary Q-node to be the root of the SPQR-tree T . Thus we only
have to care about the flexibilities of the principal split components with respect to the chosen
root. One might hope that the considerations we make for the flex-graph in the case of a fixed root
still work, if we consider the principal split components with respect to all possible roots at the
same time. However, this fails as we will see later, making it necessary to consider internal vertices
as the root.
Assume that the SPQR-tree T of G is rooted at the Q-node corresponding to an arbitrary
chosen edge. Let H be a principal split component with respect to the chosen root with the poles s
and t. In the embedding of G the outer face f of H splits into two faces f1 and f2, where the
path πf(s, t) is assumed to lie in f1 and πf(t, s) is assumed to lie in f2, that is πf1(s, t) = πf(s, t)
and πf2(t, s) = πf(t, s). We augment G by inserting the edge {s, t} twice, embedding one of them
in f1 and the other in f2. We denote the edge {s, t} inserted into the face f1 by eH(s, t) and the
edge inserted into f2 by eH(t, s). Figure 8 illustrates this process and shows how the dual graph
of G changes. We call the new edges eH(s, t) and eH(t, s) safety edges and define the extended flex
graph G× as before, ignoring that some edges have a special meaning. To simplify notation we often
use the term flex graph, although we refer to the extended flex graph. Note that every cycle in the
flex graph that separates s from t and thus crosses π(s, t) and π(t, s) needs to also cross the safety
edges eH(s, t) and eH(t, s). Thus we can use the safety edges to ensure that the flex graph respects
the flexibility of H by orienting them if necessary. More precisely, we orient the safety edge eH(s, t)
from t to s if rot(π(s, t)) = − flex(H) and similarly eH(t, s) from s to t if rot(π(t, s)) = − flex(H).
This ensures that the rotations along π(s, t) and π(t, s) cannot be reduced below − flex(H) by
bending along a cycle in the flex graph. Moreover, rot(π(s, t)) cannot be increased above flex(H)
as otherwise rot(π(t, s)) has to be below − flex(H) and vice versa. To sum up, we insert the safety
edges next to the principal split component H and orient them if necessary to ensure that bending
along a cycle in the flex graph respects not only the flexibilities of single edges but also the flexibility
of the principal split component H.
Since adding the safety edges for the graph H is just a technique to respect the flexibility
of H by bending along a cycle in the flex graph, we do not draw them. Note that the augmented
graph does not have maximum degree 4 anymore but this is not a problem since we do not draw
the safety edges. However, we formally assign an orthogonal representation to the safety edges
by essentially giving them the shape of the paths they "supervise". More precisely, the edges
eH(s, t) and eH(t, s) have the same rotations as the paths π(s, t) and π(t, s) on the outer face of
15
Hπ(s,t)π(t,s)f1f2HeH(s,t)eH(t,s)stst H, respectively. Moreover, the angles at the vertices s and t are also assumed to be the same as
for these two paths.
As we do not only want to respect the flexibility of a single split component, we add the safety
edges for each of the principal split components at the same time. Note that the augmented graph
remains planar as we only add the safety edges for the principal split components with respect
to a single root. It follows directly that the considerations above still work, which would fail if
the augmented graph was non-planar. This is the reason why we cannot consider the principal
split components with respect to all roots at the same time. The following lemma directly extends
Lemma 2 to the case where the extended flex graph is considered.
Lemma 3. Let G be a biconnected 4-planar graph with positive flexibility, a valid orthogonal rep-
resentation R and s and t on a common face f. The extended flex graph G×
R contains a directed
cycle C such that f ∈ C, s ∈ left(C) and t ∈ right(C), if one of the following conditions holds.
(1) rotR(πf(s, t)) ≥ −2, f is the outer face and πf(s, t) is not strictly directed from t to s
(2) rotR(πf(s, t)) ≥ 0 and f is the outer face
(3) rotR(πf(s, t)) ≥ 6
Proof. As in the proof of Lemma 2 we assume for contradiction that the cycle C does not exists,
yielding a strictly directed path from t to s in G. This directly yields the claim, if we can apply
Lemma 1 as before. The only difference to the situation before is that the directed path from t
to s may contain some of the safety edges. However, by definition a safety edge eH(u, v) is directed
from v to u if and only if rot(π(u, v)) = − flex(H). As flex(H) is positive rot(π(u, v)) has to be
negative and thus the rotation along eH(u, v) when traversing it from v to u is at least 1. Thus,
it does not make a difference whether the directed path from t to s consists of normal edges or
may contain safety edges. Hence, Lemma 1 extends to the augmented graph containing the safety
edges, which concludes the proof.
Now we are ready to prove Theorem 3. To improve readability we state it again.
Theorem 3. Every biconnected 4-planar graph with positive flexibility having a valid orthogonal
representation has an orthogonal representation with the same planar embedding and the same
angles around vertices that is nice with respect to at least one node chosen as root of its SPQR-tree.
Proof. Let R be a valid orthogonal representation of G. We assume without loss of generality
that R is tight. Since the operations we apply to R in the following do not affect the angles
around vertices, the resulting orthogonal representation is also tight. Thus it remains to enforce
the more interesting condition for orthogonal representations to be nice, that is reduce the number
of bends of principal split components down to three. As mentioned before, the SPQR-tree T of
G is initially rooted at an arbitrary Q-node. Let eref be the corresponding edge. As in the proof
of Theorem 2 we start with an arbitrary principal split component H with more than three bends.
Then one of the two paths in the outer face of H has rotation less than −3 and we have the same
situation as for a single edge, that is we can apply Lemma 3 to reduce the rotation of the opposite
site and thus reduce the number of bends of H by one. Afterwards, we can set the flexibility of H
down to the new number of bends ensuring that it is not increased later on. However, this only
works if the negative rotation of the split component H lies in an inner face of G. On the outer
face we can only increase to a rotation of −5 yielding an orthogonal representation such that every
principal split component has at most three bends, or maybe four or five bends, if it has its negative
rotation in the outer face. Note that this is essentially the same situation we also had in the proof
of Theorem 2. In the following we show similarly that the number of bends can be reduced further,
16
Figure 9: The path between the new and the old root in the SPQR-tree containing µ (left). The
whole graph G containing the principal split component H0 corresponding to µ with respect to the
new root and the principal split component H of the new root with respect to the old root (right).
until either a unique innermost principal split component (where innermost means minimal with
respect to inclusion) or the reference edge eref may have more than three bends.
First assume that eref has more than three, that is four or five, bends and that there is a
principal split component H with more than three bends having its negative rotation on the outer
face. Let {s, t} be the corresponding split pair and let without loss of generality πf(t, s) be the path
along H with rotation less than −3 where f is the outer face. Then the path πf(s, t) contains the
edge eref = {u, v}, otherwise H would not be a principal split component. Moreover, rot(πf(t, s)) ≤
−4 implies that rot(πf(s, t)) ≥ −2 holds. As in the proof of Theorem 2 (compare with Figure 6(c))
the path πf(s, t) splits into the paths πf(s, u), πf(u, v) and πf(v, t). Since πf(u, v) consists of the
single edge eref with more than three bends rot(πf(u, v)) ≤ −4 holds, implying that the rotation
along πf(s, u) or πf(v, t) is greater or equal to 0. This shows that πf(s, t) cannot be strictly directed
from t to s and thus we can apply Lemma 3(1) to reduce the number of bends H has. Finally,
there is no principal split component with more than three bends left and the reference edge eref
has at most five bends, which concludes this case.
In the second case, eref has at most three bends. We show that if there is more than one
principal split component with more than three bends, then they hierarchically contain each other.
Assume that the number of bends of no principal split component that has more than three bends
can be reduced further. Assume further there are two principal split components H1 and H2 with
respect to the split pairs {s1, t1} and {s2, t2} that do not contain each other, that is without loss of
generality the vertices t1, s1, t2 and s2 occur in this order around the outer face f when traversing
it in counter-clockwise direction and πf(t1, s1) and πf(t2, s2) belong to H1 and H2 respectively.
Analogous to the case where eref has more than three bends we can show that Lemma 3(1) can be
applied to reduce the number of bends of H1, which is a contradiction. Thus, either H1 is contained
in H2 or the other way round. This shows that there is a unique principal split component H that
is minimal with respect to inclusion having more than three bends. Due to the inclusion property,
all nodes in the SPQR-tree corresponding to the principal split components with more than three
bends lie on the path between the current root and the node corresponding to H. We denote the
node corresponding to H by τ and choose τ to be the new root of the SPQR-tree T . Since the
principal split components depend on the root chosen for T some split components may no longer
be principal and some may become principal due to rerooting. Our claim is that all principal split
components with more than three bends are no longer principal after rerooting and furthermore
that all split components becoming principal can be enforced to have at most three bends.
First note that the principal split component corresponding to a node µ in the SPQR-tree
changes if and only if µ lies on the path between the old and the new root, that is between τ
and the Q-node corresponding to eref. Since all principal split components (with respect to the
old root) that have more than three bends also lie on this path, all these split components are
no longer principal (with respect to the new root).
It remains to deal with the new principal
split components corresponding to the nodes on this path. Note that the new root τ itself has no
principal split component associated with it. Let µ 6= τ be a node on the path between the new
17
s0t0tsτµerefHH0rot≤−4tss0t0G and the old root and let H0 be the new principal split component corresponding to µ with the
poles s0 and t0. Recall that H is the former principal split component corresponding to the new
root τ with the poles s and t. Note that H of course is still a split component, although it is not
principal anymore. Figure 9 illustrates this situation. Now assume that H0 has more than three
bends. Then there are two possibilities, either it has its negative rotation on the outer face or in
some inner face. If only the latter case arises we can easily reduce the number of bends down to
three as we did before. In the remaining part of the proof we show that the former case cannot arise
due to the assumption that the number of bends of H cannot be reduced anymore. Assume H0
has its negative rotation in the outer face f, that is without loss of generality the path πf(t, s)
belongs to H0 and has rotation at most −4. Thus we have again the situation that the two split
components H0 and H both have a rotation of at most −4 in the outer face. Moreover, these two
split components do not contain or overlap each other since s and t are not contained in H0 as τ
is the new root and H does not contain s0 or t0 since µ is an ancestor of τ with respect to the old
root. Thus we could have reduced the number of bends of H before we changed the root, which is
a contradiction to the assumption we made that the number of bends of principal split components
with more than three bends cannot be reduced anymore. Hence, all new principal split components
either have at most three bends or they have their negative rotation in some inner face. Finally, we
obtain a valid orthogonal representation with at most three bends per principal split component
with respect to τ.
5 Optimal Drawings with Fixed Planar Embedding
All results from the previous sections deal with the case where we are only interested in the decision
problem of whether a given graph has a valid drawing or not. More precisely, we always assumed to
have a valid orthogonal representation of an instance of FlexDraw and showed that this implies
that there exists another valid orthogonal representation with certain properties. In this section,
we consider convex instances of the optimization problem OptimalFlexDraw. The following
generic theorem shows that the results for FlexDraw that we presented so far can be extended
to OptimalFlexDraw.
Theorem 4. If the existence of a valid orthogonal representation of an instance of FlexDraw
with positive flexibility implies the existence of a valid orthogonal representation with property P,
then every convex instance of OptimalFlexDraw has an optimal drawing with property P.
Proof. Let G be a convex instance of OptimalFlexDraw. Let further R be an optimal orthogonal
representation. We can reinterpret G as an instance of FlexDraw with positive flexibility by
setting the flexibility of an edge with ρ bends in R to max{ρ, 1}. Then R is obviously a valid
orthogonal representation of G with respect to these flexibilities. Thus there exists another valid
orthogonal representation R0 having property P. It remains to show that cost(R0) ≤ cost(R) holds
when going back to the optimization problem OptimalFlexDraw. However, this is clear for the
following reason. Every edge e has as most as many bends in R0 as in R except for the case where e
has one bend in R0 and zero bends in R. In the former case the monotony of coste(·) implies that
the cost did not increase. In the latter case e causes the same amount of cost in R as in R0 since
coste(0) = coste(1) = be holds for convex instances of OptimalFlexDraw. Note that this proof
still works, if the cost functions are only monotone but not convex.
It follows that every convex 4-planar graph has an optimal drawing that is nice since Theorem 4
shows that Theorem 3 can be applied. Thus, it is sufficient to consider only nice drawings when
searching for an optimal solution, as there exists a nice optimal solution. This is a fact that we
18
Figure 10: Split components with as few bends as possible.
crucially exploit in the next section since although the cost function of a principal split component
may be non-convex, we can show that it is convex in the interval that is of interest when only
considering nice drawings.
6 Optimal Drawings with Variable Planar Embedding
All results we presented so far were based on a fixed planar embedding of the input graph G. In
this section we present an algorithm that computes an optimal drawing of G in polynomial time,
optimizing over all planar embeddings of G. Our algorithm crucially relies on the existence of a
nice drawing among all optimal drawings of G. For biconnected graphs (Section 6.1) we present a
dynamic program that computes the cost function of all principal split components bottom-up in
the SPQR-tree with respect to a chosen root. To compute the optimal drawing among all drawings
that are nice with respect to the chosen root, it remains to consider the embeddings of the root itself.
If we choose every node to be the root once, this directly yields an optimal drawing of G taking
all planar embeddings into account. In Section 6.2 we extend our results to connected graphs that
are not necessarily biconnected. To this end we first modify the algorithm for biconnected graphs
such that it can compute an optimal drawing with the additional requirement that a specific vertex
lies on the outer face. Then we can use the BC-tree to solve OptimalFlexDraw for connected
graphs. We use the computation of a minimum-cost flow in a network of size n as a subroutine and
denote the consumed running time by Tflow(n). In Section 6.3 we consider which running time we
actually need.
6.1 Biconnected Graphs
In this section we always assume G to be a biconnected 4-planar graph forming a convex instance
of OptimalFlexDraw. Let T be the SPQR-tree of G. As defined before, an orthogonal represen-
tation is optimal if it has the smallest possible cost. We call an orthogonal representation τ-optimal
if it has the smallest possible cost among all orthogonal representation that are nice with respect to
the root τ. We say that it is (τ,E)-optimal if it causes the smallest possible amount of cost among
all orthogonal representations that are nice with respect to τ and induce the planar embedding E on
skel(τ). In this section we concentrate on finding a (τ,E)-optimal orthogonal representation with
respect to a root τ and a given planar embedding E of skel(τ). Then a τ-optimal representation
can be computed by choosing every possible embedding of skel(τ). An optimal solution can then
be computed by choosing every node in T to be the root once.
In Section 4 we extended the terms "bends" and "flexibility", which were originally defined for
single edges, to arbitrary principal split components with respect to the chosen root. We start out by
making precise what we mean with the cost function costH(·) of a principal split component H with
poles s and t. Recall that the number of bends of H with respect to an orthogonal representation
S with s and t on the outer face f is defined to be max{ rotS(πf(s, t)), rotS(πf(t, s))}. Assume
S is the nice orthogonal representation of H that has the smallest possible cost among all nice
orthogonal representations with ρ bends. Then we essentially define costH(ρ) to be the cost of
19
deg2deg3deg4deg4deg5deg60−10−1−1−2−10−1−1−2−2 Figure 11: A single vertex can be replaced by a split component with three bends.
S. However, with this definition the cost function of H is not defined for all ρ ∈ N0 since H
does not have an orthogonal representation with zero bends at all, if deg(s) > 1 or deg(t) > 1, as
at least one of the paths πf(s, t) and πf(t, s) has negative rotation in this case. More precisely,
if deg(s) + deg(t) > 2, then H has at least one bend, and if deg(s) + deg(t) > 4, then H has
at least two bends. Figure 10 shows for each combination of degrees a small example with the
smallest possible number of bends. In these two cases we formally set costH(0) = costH(1) and
costH(0), costH(1) = costH(2), respectively. Thus, we only need to compute the cost functions for at
least d(deg(s)+deg(t)−2)/2e bends. We denote this lower bound by 'H = d(deg(s)+deg(t)−2)/2e.
Hence, it remains to compute the cost function costH(ρ) for ρ ∈ ['H , 3]. For more than three bends
we formally set the cost to ∞. Note that the definition of the cost function only considers nice
orthogonal representations (including that they are tight). As a result of this restriction the cost for
an orthogonal representation with ρ bends might be less than costH(ρ). However, due to Theorem 3
in combination with Theorem 4 we know that optimizing over nice orthogonal representations is
sufficient to find an optimal solution.
As for single edges, we define the base cost bH of the principal split component H to be costH(0).
We will see that the cost function costH(·) is monotone and even convex in the interval [0, 3] (except
for a special case) and thus the base cost is the smallest possible amount of cost that has to be
payed for every orthogonal drawing of H. The only exception is the case where deg(s) = deg(t) = 3.
In this case H has at least two bends and thus the cost function costH(·) needs to be considered
only on the interval [2, 3]. However, it may happen that costH(2) > costH(3) holds in this case.
Then we set the base cost bH to costH(3) such that the base cost bH is really the smallest possible
amount of cost that need to be payed for every orthogonal representation of H. We obtain the
following theorem.
Theorem 5. If the poles of a principal split component do not both have degree 3, then its cost
function is convex on the interval [0, 3].
Before showing Theorem 5 we just assume that it holds and moreover we assume that the cost
function of every principal split component is already computed. We first show how these cost
functions can then be used to compute an optimal drawing. To this end, we define a flow network
on the skeleton of the root τ of the SPQR-tree, similar to Tamassias flow network [15]. The cost
functions computed for the children of τ will be used as cost functions on arcs in the flow network.
As we can only solve flow networks with convex costs we somehow have to deal with potentially
non-convex cost functions for the case that both endvertices of a virtual edge have degree 3 in its
expansion graph. Our strategy is to simply ignore these subgraphs by contracting them into single
vertices. Note that the resulting vertices have degree 2 since the poles of graphs with non-convex
cost functions have degree 3. The process of replacing the single vertex in the resulting drawing by
the contracted component is illustrated in Figure 11. The following lemma justifies this strategy.
Lemma 4. Let G be a biconnected convex instance of OptimalFlexDraw with τ-optimal orthog-
onal representation R and let H be a principal split component with non-convex cost function and
base cost bH. Let further G0 be the graph obtained from G by contracting H into a single vertex
and let R0 be a τ-optimal orthogonal representation of G0. Then cost(R) = cost(R0) + bH holds.
20
Figure 12: (a) The structure of the flow network NE for the case that τ is an R-node with skel(τ) =
K4. The outer face is split into several gray boxes to improve readability. (b) A flow together with
the corresponding orthogonal representation. The numbers indicate the amount of flow on the arcs.
Undirected edges imply 0 flow, directed arcs without a number have flow 1.
Proof. Assume we have a τ-optimal orthogonal representation R of G inducing the orthogonal
representation S on H. As H has either two or three bends we can simply contract it yielding
an orthogonal representation R0 of G with cost(R0) = cost(R) − cost(S) ≤ cost(R) − bH. The
opposite direction is more complicated. Assume we have an orthogonal representation R0 of G0,
then we want to construct an orthogonal representation R of G with cost(R) = cost(R0) + bH.
Let S be an orthogonal representation of H causing only bH cost. Since costH(·) was assumed to
be non-convex, S needs to have three bends. It is easy to see that R0 and S (or S0 obtained from S
by mirroring the drawing) can be combined to an orthogonal representation of G if the two edges
incident to the vertex v in G0 corresponding to H have an angle of 90◦ between them. However, this
can always be ensured without increasing the costs of R0. Let e1 and e2 be the edges incident to v
and assume they have an angle of 180◦ between them in both faces incident to v. If neither e1 nor e2
has a bend, the flex graph contains the cycle around v due to the fact that e1 and e2 have positive
flexibilities. Bending along this cycles introduces a bend to each of the edges, thus we can assume
without loss of generality that e1 has a bend in R0. Moving v along the edge e1 until it reaches this
bend decreases the number of bends on e1 by one and ensures that v has an angle of 90◦ in one of its
incident faces. Thus we can replace v by the split component H with orthogonal representation S
having cost bH yielding an orthogonal representation R of G with cost(R) = cost(R0) + bH.
When computing a (τ,E)-optimal orthogonal representation of G we make use of Lemma 4 in
the following way. If the expansion graph H corresponding to a virtual edge ε in skel(τ) has a non-
convex cost function, we simply contract this virtual edge in skel(τ). Note that this is equivalent
to contracting H in G. We can then make use of the fact that all remaining expansion graphs have
convex cost functions to compute a (τ,E)-optimal orthogonal representation of the resulting graph
yielding a (τ,E)-optimal orthogonal representation of the original graph G since the contracted
expansion graphs can be inserted due to Lemma 4. Note that expansion graphs with non convex
cost functions can only appear if the root is a Q- or an S-node. In the skeletons of P- and R-nodes
every vertex has degree at least three, thus the poles of an expansion graph cannot have degree 3
since G has maximum degree 4.
Now we are ready to define the flow network NE on skel(τ) with respect to the fixed embedding E
of skel(τ); see Figure 12(a) for an example. For each vertex v, each virtual edge ε and each face f
in skel(τ) the flow network NE contains the nodes v, ε and f, called vertex node, edge node and
face node, respectively. The network NE contains the arcs (v, f) and (f, v) with capacity 1, called
vertex-face arcs, if the vertex v and the face f are incident in skel(τ). For every virtual edge ε we
21
vertexnodefacenodeedgenodeskel(µ)edge-facearcvertex-facearc11232211(a)(b) be bτ =P
add edge-face arcs (ε, f) and (f, ε), if f is incident to ε. We use costH(·) − bH as cost function
of the arc (f, ε), where H is the expansion graph of the virtual edge ε. The edge-face arcs (ε, f)
in the opposite direction have infinite capacity with 0 cost. It remains to define the demand of
every node in NE. Every inner face has a demand of 4, the outer face has a demand of −4.
An edge node ε stemming from the edge ε = {s, t} with expansion graph H has a demand of
degH(s) + degH(t) − 2, where degH(v) denotes the degree of v in H. The demand of a vertex
node v is 4 − degG(v) − degskel(τ)(v).
In the flow network NE the flow entering a face node f using a vertex-face arc or an edge-
face arc is interpreted as the rotation at the corresponding vertex or along the path between the
poles of the corresponding child, respectively; see Figure 12(b) for an example. Incoming flow is
positive rotation and outgoing flow negative rotation. Let bH1, . . . , bHk be the base costs of the
expansion graphs corresponding to virtual edges in skel(τ). We define the total base costs of τ to
i bHi. Note that the total base costs of τ are a lower bound for the costs that have to be
paid for every orthogonal representation of G. We show that an optimal flow φ in NE corresponds
to a (τ,E)-optimal orthogonal representation R of G. Since the base costs do not appear in the
flow network, the costs of the flow and its corresponding orthogonal representation differ by the
total base costs bτ, that is cost(R) = cost(φ) + bτ. We obtain the following lemma.
Lemma 5. Let G be a biconnected convex instance of OptimalFlexDraw, let T be its SPQR-
tree with root τ and let E be an embedding of skel(τ). If the cost function of every principal split
component is known, a (τ,E)-optimal solution can be computed in O(Tflow( skel(τ))) time.
Proof. As mentioned before, we want to use the flow network NE to compute an optimal orthogonal
representation. To this end we show two directions. First, given a (τ,E)-optimal orthogonal
representation R, we obtain a feasible flow φ in NE such that cost(φ) = cost(R) − bτ, where bτ
are the total base costs. Conversely, given an optimal flow φ in NE, we show how to construct an
orthogonal representation R such that cost(R) = cost(φ) + bτ. As the flow network NE has size
O( skel(τ)), the claimed running time follows immediately.
Let R be a (τ,E)-optimal orthogonal representation of G. As we only consider nice and thus
only tight drawings we can assume the orthogonal representation R to be tight. Recall that being
tight implies that the poles of the expansion graph of every virtual edge have a rotation of 1 in
the internal faces. We first show how to assign flow to the arcs in NE. It can then be shown that
the resulting flow is feasible and causes cost(R) − bτ cost. For every pair of vertex-face arcs (f, v)
and (v, f) in NE there exists a corresponding face f in the orthogonal representation R of G and we
set φ((v, f)) = rot(vf). Let ε = {s, t} be a virtual edge in skel(µ) incident to the two faces f1 and f2.
Without loss of generality let πf1(s, t) be the path belonging to the expansion graph of ε. Then
πf2(t, s) also belongs to H. We set φ((ε, f1)) = rotR(πf1(s, t)) and φ((ε, f2)) = rotR(πf2(t, s)). For
the resulting flow φ we need to show that the capacity of every arc is respected, that the demand
of every vertex is satisfied, and that cost(φ) = cost(R) − bτ holds.
First note that the flow on the vertex-face arcs does not exceed the capacities of 1 since every
vertex has degree at least 2. Since no other arc has a capacity, it remains to deal with the demands
and the costs.
For the demands we consider each vertex type separately. Let f be a face node. The total
incoming flow entering f is obviously equal to the rotation in R around the face f. As R is an
orthogonal representation this rotation equals to 4 (−4 for the outer face), which is exactly the
demand of f. Let ε be an edge node corresponding to the expansion graph H with poles s and t.
Recall that dem(ε) = degH(s) + degH(t) − 2 is the demand of ε. Figure 13(a) illustrates the
demand of a virtual edge. Let S be the orthogonal representation induced on H by R and let f
be the outer face of S. Clearly, the flow leaving ε is equal to rotR(πf1(s, t)) + rotR(πf2(t, s)) =
22
Figure 13: (a) Illustration of the demand of virtual edges. (b) Rotation of poles in the outer face,
depending on the degree.
rotS(πf(s, t)) + rotS(πf(t, s)). Since f is the outer face of H, the total rotation around this faces
sums up to −4. The rotation of the pole s in the outer face f is degH(s)−3, see Figures 13(b), and
the same holds for t. Thus we have rotS(πf(s, t))+rotS(πf(t, s))+degH(s)−3+degH(t)−3 = −4.
This yields for the outgoing flow rotS(πf(s, t)) + rotS(πf(t, s)) = 2 − degH(s) − degH(t), which is
exactly the negative demand of ε. It remains to consider the vertex nodes. Let v be a vertex node,
recall that dem(v) = 4 − degG(v) − degskel(τ)(v) holds. The outgoing flow leaving v is equal to
the summed rotation of v in faces not belonging to expansion graphs of virtual edges in skel(τ).
As R is an orthogonal representation, the total rotation around every vertex v is 2 · (degG(v) − 2).
Moreover, v is incident to degskel(τ)(v) faces that are not contained in expansion graphs of virtual
edges of skel(τ). Thus there are degG(v) − degskel(τ)(v) faces incident to v belonging to expansion
graphs. As we assumed that the orthogonal representation of every expansion graph is tight, the
rotation of v in each of these faces is 1. Thus the rotation of v in the remaining faces not belonging
to expansion graphs is 2·(degG(v)−2)−(degG(v)−degskel(τ)(v)). Rearrangement yields a rotation,
and thus an outgoing flow, of degG(v) + degskel(τ)(v) − 4, which is the negative demand of v.
To show that cost(φ) = cost(R)−bτ holds it suffices to consider the flow on the edge-face arcs as
no other arcs cause cost. Let ε be a virtual edge and let f1 and f2 the two incident faces. The flow
entering f1 or f2 does not cause any cost, as (ε, f1) and (ε, f2) have infinite capacity with 0 cost.
Thus only flow entering ε over the arcs (f1, ε) and (f2, ε) may cause cost. Assume without loss of
generality that the number of bends ρ the expansion graph H of ε has is determined by the rotation
along πf1(s, t), that is ρ = − rotR(πf1(s, t)). Let ρ0 = − rotR(πf2(t, s)) be the negative rotation
along the path πf2(t, s) in the face f2. Note that φ((f1, ε)) = ρ and φ((f2, ε)) = ρ0. Obviously, the
flow on (f1, ε) causes the cost costH(ρ) − bH. We show that the cost caused by the flow on (f2, ε)
is 0. If ρ0 ≤ 0 this is obviously true, as there is no flow on the edge (f2, ε). Otherwise, 0 < ρ0 ≤ ρ
holds. It follows that the smallest possible number of bends 'H every orthogonal representation
of H has lies between ρ0 and ρ. It follows from the definition of costH(·) and from the fact that
all cost functions are convex that costH(ρ0) = bH. To sum up, the total cost on edge-face arcs
incident to the virtual edge ε is equal to the cost caused by its expansion graph H with respect to
the orthogonal representation R minus the base cost bH. As neither φ nor R have additional cost
we obtain cost(φ) = cost(R) − bτ.
It remains to show the opposite direction, that is given an optimal flow φ in NE, we can
construct an orthogonal representation R of G such that cost(R) = cost(φ) + bτ. This can be done
by reversing the construction above. The flow on edge-face arcs determines the number of bends for
the expansion graphs of each virtual edge. The cost functions of these expansion graphs guarantee
the existence of orthogonal representations with the desired rotations along the paths between the
poles, thus we can assume to have orthogonal representations for all children. We combine these
orthogonal representations by setting the rotations between them at common poles as specified by
the flow on vertex-face arcs. It can be easily verified that this yields an orthogonal representation
of the whole graph G by applying the above computation in the opposite direction.
The above results rely on the fact that the cost functions of principal split components are
23
(a)012111−2−1021t131st242st2tttssss(b) convex as stated in Theorem 5 and that they can be computed efficiently. In the following we show
that Theorem 5 really holds with the help of a structural induction over the SPQR-tree. More
precisely, the cost functions of principal split components corresponding to the leaves of T are the
cost functions of the edges and thus they are convex. For an inner node µ we assume that the
pertinent graphs of the children of µ have convex cost functions and show that H = pert(µ) itself
also has a convex cost function. The proof is constructive in the sense that it directly yields an
algorithm to compute these cost functions bottom up in the SPQR-tree.
Note that we can again apply Lemma 4 in the case that the cost function of the expansion
graph of one of the virtual edges in skel(µ) is not convex due to the fact that both of its poles
have degree 3. This means that we can simply contract such a virtual edge (corresponding to
a contraction of the expansion graph in H), compute the cost function for the remaining graph
instead of H and plug the contracted expansion graph into the resulting orthogonal representations.
Thus we can assume that the cost function of each of the expansion graphs is convex, without any
exceptions.
The flow network NE that was introduced to compute an optimal orthogonal representation
in the root of the SPQR-tree can be adapted to compute the cost function of the principal split
component H corresponding to a non-root node µ. To this end we have to deal with the parent
edge, which does not occur in the root of T , and we consider a parameterization of NE to compute
several optimal orthogonal representations with a prescribed number of bends, depending on the
parameter in the flow network. Before we describe the changes in the flow network we need to make
some considerations about the cost function. By the definition of the cost function it explicitely
optimizes over all planar embeddings of skel(µ). Moreover, as the cost function costH(ρ) depends
on the number of bends ρ a graph H has, it implicitly allows to flip the embedding of H since
the number of bends is defined as max{ rot(π(s, t)), rot(π(t, s))}. However, the flow network NE
can only be used to compute the cost function for a fixed embedding. Thus we define the partial
H(ρ) of H with respect to the planar embedding E of skel(µ) to be the smallest
cost function costE
possible cost of an orthogonal representation inducing the planar embedding E on skel(µ) with ρ
bends such that the number of bends is determined by πf(s, t), that is rot(πf(s, t)) = −ρ, where f
H(·),
is the outer face. Note that the minimum over the partial cost functions costE
where E0 is obtained by flipping the embedding E of skel(µ) yields a function describing the costs
of H with respect to the embedding E of skel(µ) depending on the number of bends H has (and
not on the rotation along πf(s, t) as the partial cost function does). Obviously, minimizing over all
partial cost functions yields the cost function of H.
The flow network NE is defined as before with the following modifications. The parent edge of
skel(µ) does not have a corresponding edge node. Let f1 and f2 be the faces in skel(µ) incident
to the parent edge. These two faces together form the outer face f of H, thus we could merge
them into a single face node. However, not merging them has the advantage that the incoming
flow in f1 and f2 corresponds to the rotations along πf(s, t) and πf(t, s), respectively (it might be
the other way round but we can assume this situation without loss of generality). Thus, we do not
merge f1 and f2, which enables us to control the number of bends of H by setting the demands
of f1 and f2. This is also the reason why we remove the vertex-face arcs between the poles and the
two faces f1 and f2. Before we describe how to set the demands of f1 and f2, we fit the demands
of the poles to the new situation. As we only consider tight orthogonal representations we know
that the rotation at the poles s and t in all inner faces is 1. Thus, we set dem(s) = 2− degskel(µ)(s)
and dem(t) = 2 − degskel(µ)(t) as this is the number of faces incident to s and t, respectively, after
removing the vertex-face arcs to f1 and f2. With these modifications the only flow entering f1 and f2
comes from the paths πf(s, t) and πf(t, s), respectively. As the total rotation around the outer face
is −4 and the rotation at the vertices s and t is degH(s)−3 and degH(t)−3, respectively, we have to
H(·) and costE0
24
ensure that dem(f1) + dem(f2) = 2− degH(s)− degH(t). As mentioned before, we assume without
loss of generality that πf(s, t) belongs to the face f1 and πf(t, s) belongs to f2. Then the incoming
flow entering f1 corresponds to rot(πf(s, t)) of an orthogonal representation. We parameterize NE
with respect to the faces f1 and f2 starting with dem(f1) = 0 and dem(f2) = 2−degH(s)−degH(t).
It obviously follows that an optimal flow in NE with respect to the parameter ρ corresponds to
an optimal orthogonal representation of H that induces E on skel(µ) and has a rotation of −ρ
along πf(s, t). Thus, up to the total base costs bµ, the cost function of the flow network equals
to the partial cost function of H on the interval ['H , 3], that is costNE(ρ) + bµ = costE
H(ρ) for
'H ≤ ρ ≤ 3. To obtain the following lemma it remains to show two things for the case that
deg(s) + deg(t) < 6. First, costNE(ρ) and thus each partial cost function is convex for 'H ≤ ρ ≤ 3.
Second, the minimum over these partial cost functions is convex.
Lemma 6. If Theorem 5 holds for each principal split component corresponding to a child of the
node µ in the SPQR-tree, then it also holds for pert(µ).
Proof. As mentioned before, we can use the flow network NE to compute the partial cost function
costE
H(ρ) = costNE(ρ) + bµ holds on this interval. In the following
we only consider the case where degH(s) + degH(t) < 6 holds for the poles s and t. For the case
degH(s) = degH(t) = 3 we do not need to show anything. To show that the partial cost function is
convex we do the following. First, we show that costE
H(ρ) is minimal for ρ = 'H. This implies that
the cost function costNE(ρ) of the flow network is minimal for ρ = ρ0 ≤ 'H. Then Theorem 1 can
be applied showing that costNE(ρ) is convex for ρ ∈ [ρ0,∞] yielding that the partial cost function
H(ρ) is convex for ρ ∈ ['H , 3]. Thus, it remains to show that costE
costE
H(ρ) is minimal for ρ = 'H to
obtain convexity for the partial cost functions.
Let S be an orthogonal representation of H with ρ ∈ ['H , 3] bends such that πf(s, t) determines
the number of bends, that is rotS(πf(s, t)) = −ρ, where f is the outer face of H. We show the
existence of an orthogonal representation S0 with rotS0(πf(s, t)) = −'H and cost(S0) ≤ cost(S).
Since we assume S to be tight, the rotations at the poles rotS(sf) and rotS(tf) only depend on the
degree of s and t. More precisely, we have rotS(sf) = degH(s) − 3 and the same holds for t. Since
the total rotation around the outer face f is −4 the following equation holds.
H(ρ) for 'H ≤ ρ ≤ 3 since costE
rotS(πf(t, s)) = ρ + 2 − degH(s) − degH(t)
(1)
In the following we show that rotS(πf(t, s)) ≥ 0 holds if the number of bends ρ exceeds 'H. Then
Corollary 1 in combination with Theorem 4 can be used to reduce the rotation along πf(t, s)
and thus reduce the number of bends by 1, yielding finally an orthogonal representation with 'H
bends determined by πf(s, t). Recall that the lower bound for the number of bends was defined
as 'H = d(deg(s) + deg(t) − 2)/2e. First consider the case that degH(s) + degH(t) is even (and of
course less than 6). Then Equation (1) yields rotS(πf(t, s)) = ρ − 2'H. If ρ is greater than 'H this
yields rotS(πf(t, s)) > −'H. Since 'H is at most 1 in the case that deg(s) + deg(t) is even and less
than 6, this yields rotS(πf(t, s)) > −1. The case that degH(s) + degH(t) is odd works similarly.
Then Equation (1) yields rotS(πf(t, s)) = ρ−2'H +1. As before ρ is assumed to be greater than 'H
yielding rotS(πf(t, s)) > −'H + 1. As 'H is at most 2 we again obtain rotS(πf(t, s)) > −1, which
concludes the proof that the partial cost functions are convex.
It remains to show that the minimum over the partial cost functions is convex. First assume
that µ is an R-node. Then its skeleton has only two embeddings E and E0 where E0 is obtained by
flipping E. We have to show that the minimum over the two partial cost functions costE
H(·) and
H(·) remains convex. For the case that deg(s) + deg(t) = 5 the equation 'H = 2 holds and thus
costE0
we only have to show convexity on the interval [2, 3]. Obviously, costH(·) is convex on this interval
25
H('H) = costE0
H(0) = costE
H(·) and costE0
H('H) = costE0
H(1) holds by definition, if deg(s) + deg(t) > 2.
if and only if costH(2) ≤ costH(3). As this is the case for both partial cost functions, it is also
true for the minimum. For deg(s) + deg(t) < 5 we first show that costE
H('H) holds.
For the case that deg(s) + deg(t) is even this is clear since mirroring an orthogonal representa-
tion S with rotS(πf(s, t)) = −'H inducing E on skel(µ) yields an orthogonal representation S0 with
rotS0(πf(s, t)) = −'H inducing E0 on skel(µ). For the case that deg(s) + deg(t) = 3, the orthogonal
representation S with rotation −1 along πf(s, t) can also be mirrored yielding S0 with rotation 0
along πf(s, t). By Corollary 1 this rotation can be reduced to −1 without causing any additional
cost. As this construction also works in the opposite direction we have costE
H('H)
for all cases. Moreover, costE
If
deg(s) = deg(t) = 1 this equation is also true as the rotation along πf(s, t) of an orthogonal
representation can be reduced by 1 if it is 0, again due to Corollary 1. Thus it remains to show
that the cost function costH(·) defined as the minimum of costE
H(·) is convex on the
interval [1, 3].
Assume for a contradiction that costH(ρ) is not convex for ρ ∈ [1, 3], that is ∆ costH(1) >
∆ costH(2). Assume without loss of generality that costH(3) = costE
H(3) holds. As we showed before
costH(1) = costE
H(2) we
additionally have costH(2) ≤ costE
H(1) ≥ ∆ costH(1)
and ∆ costE
H(ρ) is not convex
for ρ ∈ [1, 3], which is a contradiction. Thus costH(·) is convex.
The case that µ is a P-node works similar to the case that µ is an R-node. If µ has only two
children, its skeleton has only two embeddings E and E0 obtained from one another by flipping.
Thus the same argument as for R-nodes applies. If µ has three children, then deg(s) = deg(t) = 3
holds and thus we do not have to show convexity. Note that in the case deg(s) = deg(t) = 3 the
resulting cost function can be computed by taking the minimum over the partial cost functions with
respect to all embeddings of skel(µ), although it may by non-convex. If µ is an S-node, we have
a unique embedding and thus the partial cost function with respect to this embedding is already
the cost function of H. Note that considering only the rotation along πf(s, t) for the partial cost
function is not a restriction, as S-nodes are completely symmetric.
H(2). This implies that the inequalities ∆ costE
H(2) ≤ ∆ costH(2) hold, yielding that the partial cost function costE
H(1) also holds. Since costH(2) is the minimum over costE
H(2) and costE
Lemma 6 together with the fact that the cost function of every edge is convex shows that
Theorem 5 holds, that is the cost functions of all principal split components are convex on the
interesting interval [0, 3] except for the special case where both poles have degree 3. However,
this special case is easy to handle as principal split components of this type with non-convex
cost functions can be simply contracted to a single vertex by Lemma 4. Moreover, the proof is
constructive in the sense that it shows how the cost functions can be computed efficiently bottom
up in the SPQR-tree. For each node µ we have to solve a constant number of minimum-cost flow
problems in a flow network of size O( skel(µ)). As the total size of all skeletons in T is linear
in the number n of vertices in G, we obtain an overall O(Tflow(n)) running time to compute the
cost functions with respect to the root τ. Finally, Lemma 5 can be applied to compute an optimal
orthogonal representation with respect to a fixed root and a fixed embedding of the root's skeleton
in O(Tflow( skel(τ))) time. To compute an overall optimal solution, we have to compute a (τ,E)-
optimal solution for every root τ and every embedding E of skel(τ). The number of embeddings of
skel(τ) is linear in the size of skel(τ) (since P-nodes have at most degree 4) and the total size of all
skeletons is linear in n. We obtain the following theorem.
Theorem 6. OptimalFlexDraw can be solved in O(n · Tflow(n)) time for convex biconnected
instances.
26
6.2 Connected Graphs
In this section we extend the result obtained in Section 6.1 to the case that the input graph G
contains cutvertices. Let B be the BC-tree of G rooted at some B-node β. Then every Block except
for β has a unique cutvertex as parent and we need to find optimal orthogonal representations with
the restriction that this cutvertex lies on the outer face. We claim that we can then combine these
orthogonal representations of the blocks without additional cost.
Unfortunately, with the so far presented results we cannot compute the optimal orthogonal
representation of a biconnected graph considering only embeddings where a specific vertex v lies
on the outer face. We may restrict the embeddings of the skeletons we consider when traversing
the SPQR-tree bottom up to those who have v on the outer face. However, we can then no longer
assume that the cost functions we obtain are symmetric. To deal with this problem, we present a
modification of the SPQR-tree, that can be used to represent exactly the planar embeddings that
have v on the outer face and are represented by the SPQR-tree rooted at a node τ.
Let τ be the root of the SPQR-tree T . If v is a vertex of skel(τ), then restricting the embeddings
of skel(τ) to those who have v on the outer face of skel(τ) forces v to be on the outer face of the
resulting embedding of G. Otherwise, v is contained in the expansion graph of a unique virtual
edge ε in skel(τ), we say that v is contained in ε. Obviously, ε has to be on the outer face of the
embedding of skel(τ). However, this is not sufficient and it depends on the child µ of τ corresponding
to ε whether v lies on the outer face of the resulting embedding of G. Let Eτ be an embedding
of skel(τ) having ε on the outer face and let s and t be the endpoints of ε. Then there are two
possibilities, either ε = {s, t} has the outer face to the left or to the right, where the terms "left"
and "right" are with respect to an orientation from t to s. Assume without loss of generality that
the outer face lies to the right of ε and consider the child µ of τ corresponding to ε. As T is
rooted, we consider only embeddings of skel(µ) that have the parent edge {s, t} on the outer face.
As the choice of the outer face of skel(µ) does not have any effect on the resulting embedding, we
can assume that {s, t} lies to the left of skel(µ), that is the inner face incident to {s, t} lies to the
right of {s, t} with respect to an orientation from t to s. A vertex contained in skel(µ) then lies
obviously on the outer face of the resulting embedding of G if and only if it lies on the outer face
of the embedding of skel(µ). Thus, if v is contained in skel(µ), restricting the embedding choices
such that v lies on the outer face of skel(µ) forces v to be on the outer face of G. Note that in this
case µ is either an R- or an S-node. For S-nodes there is no embedding choice and every vertex in
skel(µ) lies on the outer face in this embedding. If µ is an R-node, there are only two embeddings
and either v lies on the outer face of exactly one of them or in none of them. In the latter case the
SPQR-tree with respect to the root τ does not represent an embedding of G with v on the outer
face at all.
Assume that v is not contained in skel(µ). Then it is again contained in a single virtual edge ε0
and it is necessary that ε0 lies on the outer face of the embedding of skel(µ). Moreover, it depends
on the child of µ corresponding to ε0 whether v really lies on the outer face. Note that fixing ε0
on the outer face completely determines the embedding of skel(µ) if it is not a P-node. If µ is a
P-node, the virtual edge ε0 has to be the rightmost, whereas the order of all other virtual edges can
be chosen arbitrarily. If this is the case we split the P-node into two parts, one representing the
fixed embedding of ε0, the other representing the choices for the remaining edges; see Figure 14(a).
More precisely, we split µ into two P-nodes, the first one containing the parent edge {s, t}, the
edge ε0 and a new virtual edge corresponding to the second P-node, which is inserted as child. The
skeleton of the second P-node contains a parent edge corresponding to the first P-node and the
remaining virtual edges that were contained in skel(µ) but are not contained in the first P-node.
The children of µ are attached to the two P-nodes depending on where the corresponding virtual
27
Figure 14: (a) Splitting a P-node into two P-nodes, the vertex v fixed to the outer face is contained
in the thick edges. (b) Contracting the path from the root to the node containing v in its skeleton.
edges are. Note that by splitting the P-node µ, the virtual edge ε0 can no longer be in between two
other virtual edges in µ. However, this is a required restriction, thus we do not loose embeddings
that we want to represent. Moreover, the new P-node containing the virtual edge ε0 that need
to be fixed to the outer face contains only two virtual edges (plus the parent edge) and thus the
embedding of its skeleton is completely fixed by requiring ε0 to be on the outer face.
To sum up, if skel(τ) contains v, then we simply have to choose an embedding of skel(τ) with v
on the outer face. Otherwise, we have to fix the virtual edge containing v to the outer face and
additionally have to consider the child of τ corresponding to this virtual edge. For the child we
then have essentially the same situation. Either v is contained in its skeleton, then the embedding
is fixed to the unique embedding having v on the outer face or v is contained in some virtual
edge. However, then the embedding of the skeleton is again completely fixed (P-nodes have to be
split up first) and we can continue with the child corresponding to the virtual edge containing v.
This yields a path of nodes starting with the root τ having a completely fixed embedding only
depending on the embedding Eτ chosen for skel(τ). As the nodes on the path do not represent any
embedding choices, we can simply contract the whole path into a single new root node, merging
the skeletons on the path, such that the embedding of the new skeleton of the root is still fixed.
This contraction is illustrated in Figure 14(b). More precisely, let τ be the root and let ε be the
edge containing v, corresponding to the child µ. Then we merge τ and µ by replacing ε in τ by the
skeleton of µ without the parent edge. The children of µ are of course attached to the new root
τ0 since skel(τ0) contains the corresponding virtual edges. As mentioned before, the embedding of
skel(µ) was fixed by the requirement that v is on the outer face, thus the new skeleton skel(τ0) has
a unique embedding Eτ0 inducing Eτ on skel(τ) and having v or the new virtual edge containing v
on the outer face. The procedure of merging the root with the child corresponding to the virtual
edge containing v is repeated until v is contained in the skeleton of the root. We call the resulting
tree the restricted SPQR-tree with respect to the vertex v and to the embedding Eτ of the root.
To come back to the problem OptimalFlexDraw, we can easily apply the algorithm presented
in Section 6.1 to the restricted SPQR-tree. All nodes apart from the root are still S-, P-, Q- or
R-nodes and thus the cost functions with respect to the corresponding pertinent graphs can be
computed bottom up. The root τ may have a more complicated skeleton, however, its embedding
is fixed, thus we can apply the flow algorithm as before, yielding an optimal drawing with respect
to the chosen root τ and to the embedding Eτ of skel(τ) with the additional requirement that v
lies on the outer face. Since the restricted SPQR-tree can be easily computed in linear time for
a chosen root τ and a fixed embedding E of skel(τ), we can compute a (τ,E)-optimal orthogonal
representation with the additional requirement that v lies on the outer face in Tflow(n) time, yielding
the following theorem.
28
µ1µ2µ3µ4ststµ1µ2µ3stµ4tsts(a)µ1µ6µ7µ2µ3µ4µ5µ2µ1µ6µ7µ3µ4µ5(b)vv Theorem 7. OptimalFlexDraw with the additional requirement that a specific vertex lies on
the outer face can be solved in O(n · Tflow(n)) time for convex biconnected instances.
As motivated before, we can use the BC-tree to solve OptimalFlexDraw for instances that
are not necessarily biconnected. We obtain the following theorem.
Theorem 8. OptimalFlexDraw can be solved in O(n2 · Tflow(n)) time for convex instances.
Proof. Let G be a convex instance with positive flexibility of OptimalFlexDraw and let B be its
BC-tree rooted at some B-node β. We show how to find an optimal drawing of G, optimizing over all
embeddings represented by B with respect to the root β. Then we can simply choose every B-node
in B to be the root once, solving OptimalFlexDraw. The algorithm consumes O(n·Tflow(n)) time
for each root β and thus the overall running time is O(n2 · Tflow(n)). For the block corresponding
to the root β we use Theorem 6 to find the optimal orthogonal representation. For all other blocks
we use Theorem 7 to find the optimal orthogonal representation with the cutvertex corresponding
to the parent in B on the outer face. It remains to stack these orthogonal representations together
without causing additional cost. This can be easily done, if a cutvertex that is forced to lie on the
outer face has all free incidences in the outer face and every other cutvertex has all free incidences
in a single face. The former can be achieved as we can assume orthogonal representations to be
tight. If the latter condition is violated by a cutvertex v, then v has two incident edges e1 and e2
and the rotation of v is 0 in both incident faces. If both edges e1 and e2 have zero bends, we bend
along a cycle around v in the flex graph and thus we can assume without loss of generality that e1
has a bend. Moving v along e1 to this bend yields an orthogonal representation where v has both
free incidences in the same face. Thus given the orthogonal representations for the blocks, we can
simply stack them together without causing additional cost.
6.3 Computing the Flow
In the previous sections we used Tflow(n) as placeholder for the time necessary to compute a
minimum-cost flow in a flow network of size n. Most minimum-cost flow algorithms do not consider
the case of multiple sinks and sources. However, this is not a real problem as we can simply
add a supersink connected to all sinks and a supersource connected to all sources. Unfortunately,
the resulting flow network is no longer planar. Orlin gives a strongly polynomial time minimum-
cost flow algorithm with running time O(m log n(m + n log n)), where n is the number of vertices
and m the number of arcs [13]. Since our flow network is planar (plus supersink and supersource)
the number of arcs is linear in the number of nodes. Thus with this flow algorithm we have
Tflow(n) ∈ O(n2 log2 n).
Cornelsen and Karrenbauer give a minimum-cost flow algorithm for planar flow networks with
multiple sources and sinks consuming O(√
χ n log3 n) time [3], where χ is the cost of the resulting
flow. Since the cost functions in an instance of OptimalFlexDraw may define exponentially large
costs in the size of the input, we cannot use this flow algorithm in general to obtain a polynomial
time algorithm. However, in practice it does not really make sense to have exponentially large costs.
Moreover, in several interesting special cases an optimal solution has cost linear in the number of
vertices. We obtain the following results.
Corollary 3. A convex instance G of OptimalFlexDraw can be solved in O(n4 log2 n) and
O(√
χ n3 log3 n) time, where χ is the cost of an optimal solution. The running time can be improved
by a factor of O(n) for biconnected graphs.
29
7 Conclusion
We presented an efficient algorithm for the problem OptimalFlexDraw that can be seen as the
optimization problem corresponding to FlexDraw. As a first step, we considered biconnected
4-planar graphs with a fixed embedding and showed that they always admit a nice drawing, which
implies at most three bends per edge except for a single edge on the outer face with up to four
bends.
Our algorithm for optimizing over all planar embeddings requires that the first bend on every
edge does not cause any cost as the problem becomes NP-hard otherwise. Apart from that restric-
tion we allow the user to specify an arbitrary convex cost function independently for each edge.
This enables the user to control the resulting drawing. For example, our algorithm can be used
to minimize the total number of bends, neglecting the first bend of each edge. This special case
is the natural optimization problem arising from the decision problem FlexDraw. As another
interesting special case, one can require every edge to have at most two bends and minimize the
number of edges having more than one bend. This enhances the algorithm by Biedl and Kant [1]
generating drawings with at most two bends per edge with the possibility of optimization. Note
that in both special cases the cost of an optimal solution is linear in the size of the graph, yielding
a running time in O(n
5
2 log3 n) if the graph is biconnected).
7
2 log3 n) (O(n
References
[1] T. Biedl and G. Kant. A Better Heuristic for Orthogonal Graph Drawings. Comput. Geom.,
9(3):159 -- 180, 1998.
[2] T. Bläsius, M. Krug, I. Rutter, and D. Wagner. Orthogonal Graph Drawing with Flexibility
Constraints. In Graph Drawing (GD'10), volume 6502 of LNCS, pages 92 -- 104. Springer, 2011.
[3] S. Cornelsen and A. Karrenbauer. Accelerated Bend Minimization. In Graph Drawing (GD'11),
volume 7034 of LNCS, pages 111 -- 122. Springer, 2012.
[4] G. Di Battista, G. Liotta, and F. Vargiu. Spirality and Optimal Orthogonal Drawings. SIAM
J. Comput., 27(6):1764 -- 1811, 1998.
[5] G. Di Battista and R. Tamassia. On-Line Maintenance of Triconnected Components with
SPQR-Trees. Algorithmica, 15(4):302 -- 318, 1996.
[6] G. Di Battista and R. Tamassia. On-Line Planarity Testing. SIAM J. Comput., 25(5):956 -- 997,
1996.
[7] J. Edmonds and R. M. Karp. Theoretical Improvements in Algorithmic Efficiency for Network
Flow Problems. J. ACM, 19:248 -- 264, 1972.
[8] U. Fössmeier and M. Kaufmann. Drawing High Degree Graphs with Low Bend Numbers. In
Graph Drawing (GD'95), LNCS, pages 254 -- 266. Springer, 1995.
[9] A. Garg and R. Tamassia. On the Computational Complexity of Upward and Rectilinear
Planarity Testing. SIAM J. Comput., 31(2):601 -- 625, 2001.
[10] C. Gutwenger and P. Mutzel. A Linear Time Implementation of SPQR-Trees.
Drawing (GD'00), volume 1984 of LNCS, pages 77 -- 90. Springer, 2001.
In Graph
30
[11] G. W. Klau and P. Mutzel. Quasi-Orthogonal Drawing of Planar Graphs. Technical report,
Max-Planck-Institut für Informatik, Saarbrücken, Germany, 1998.
[12] A. Morgana, C. P. de Mello, and G. Sontacchi. An Algorithm for 1-bend Embeddings of Plane
Graphs in the Two-Dimensional Grid. Discr. Appl. Math., 141(1-3):225 -- 241, 2004.
[13] J. B. Orlin. A Faster Strongly Polynomial Minimum Cost Flow Algorithm. Oper. Res., 41:338 --
350, 1993.
[14] I. Rutter. The Many Faces of Planarity -- Matching, Augmentation, and Embedding Algorithms
for Planar Graphs -- . PhD thesis, Fakultät für Informatik, Karlsruher Institut für Technologie
(KIT), 2011.
[15] R. Tamassia. On Embedding a Graph in the Grid with the Minimum Number of Bends. SIAM
J. Comput., 16(3):421 -- 444, 1987.
[16] R. Tamassia, G. Di Battista, and C. Batini. Automatic Graph Drawing and Readability of
Diagrams. IEEE Trans. Syst. Man Cybern., 18(1):61 -- 79, 1988.
[17] S. Tayu, K. Nomura, and S. Ueno. On the Two-Dimensional Orthogonal Drawing of Series-
Parallel Graphs. Discr. Appl. Math., 157(8):1885 -- 1895, 2009.
[18] H. Whitney. Non-Separable and Planar Graphs. Transactions of the American Mathematical
Society, 34(2):339 -- 338, April 1932.
31
|
1902.08742 | 1 | 1902 | 2019-02-23T06:04:30 | Optimal Algorithm to Reconstruct a Tree from a Subtree Distance | [
"cs.DS"
] | This paper addresses the problem of finding a representation of a subtree distance, which is an extension of the tree metric. We show that a minimal representation is uniquely determined by a given subtree distance, and give a linear time algorithm that finds such a representation. This algorithm achieves the optimal time complexity. | cs.DS | cs |
Optimal Algorithm to Reconstruct a Tree from a Subtree
Distance
Takanori Maehara
RIKEN Center for Advanced Intelligence Project
[email protected]
Kazutoshi Ando
Department of Mathematical and Systems Engineering, Shizuoka University
[email protected]
Abstract
This paper addresses the problem of finding a representation of a subtree distance,
which is an extension of the tree metric. We show that a minimal representation
is uniquely determined by a given subtree distance, and give a linear time algo-
rithm that finds such a representation. This algorithm achieves the optimal time
complexity.
Keywords: graph algorithm, phylogenetic tree, tree metric, subtree distance
1. Introduction and the results
A phylogenetic tree represents an evolutionary relationship among the species
that are investigated. Estimating a phylogenetic tree from experimental data is a
fundamental problem in phylogenetics [8]. One of the commonly used approaches
to achieve this task is the use of a distance-based method. In this approach, we first
compute the dissimilarity (i.e., a nonnegative and symmetric function) between the
species by, e.g., the edit distance between the genome sequences. Then, we find a
weighted tree having the shortest path distance that best fits the given dissimilarity.
The most popular method for this approach is the neighbor-joining method [6].
A weighted tree T is specified by the set of vertices V(T ), the set of edges E(T ),
and the nonnegative edge weight l : E(T ) → R+. Let us consider the case in which a
given dissimilarity d : X × X → R exactly fits some weighted tree; i.e., there exists
1
a weighted tree T and a mapping ψ : X → V(T ) such that
d(x, y) = dT (ψ(x), ψ(y))
(x, y ∈ X),
(1.1)
where dT (u, v) is the distance between u and v in T for u, v ∈ V(T ). In this case,
the dissimilarity d : X × X → R is called a tree metric, and the pair (T , ψ) is called
a representation of d. It is known that a dissimilarity d is a tree metric if and only
if it satisfies an inequality called the four-point condition [9, 2], which is given by
d(x, y) + d(z, w) ≤ max{d(x, z) + d(y, w), d(x, w) + d(y, z)}
(1.2)
for any x, y, z, w ∈ X. For any tree metric d, there exists a unique minimal represen-
tation (T , ψ) of d [2]. Here, a representation is minimal if there is no representation
(T ′, ψ′) of d, such that T ′ is obtained by removing some vertices and edges and/or
by contracting some edges of T (i.e., T ′ is a proper minor of T ). Furthermore, such
a representation is constructed in O(n2) time [3], where n = X.
In some applications, we are interested in the distance between groups of species
(e.g., genus, tribe, or family). In such a case, we aim to identify a group as a con-
nected subgraph in a phylogenetic tree. The subtree distance, which was introduced
by Hirai [5], is an extension of the tree metric that can be adopted for use in such
situations. A function d : X × X → R is called a subtree distance if there exists a
weighted tree T and a mapping φ : X → 2V(T ) such that φ(x) induces a subtree of
T (i.e., a connected subgraph of T ) for x ∈ X and equations
d(x, y) = dT (φ(x), φ(y))
(x, y ∈ X),
(1.3)
hold, where dT (U, W ) = min{dT (u, w) u ∈ U, w ∈ W } for U, W ⊆ V(T ). We
say that a pair (T , φ) is a representation of d. Note that a subtree distance is not
necessarily a metric because it may not satisfy the non-degeneracy (d(x, y) > 0 for
x 6= y) and the triangle inequality (d(x, z) ≤ d(x, y) + d(y, z)). Hirai proposed a
characterization of subtree distances in which a dissimilarity d : X × X → R is a
subtree distance if and only if it satisfies an inequality called the extended four-point
condition, which is given by
d(x, y) + d(z, w) ≤ max
d(x, z) + d(y, w), d(x, w) + d(y, z),
d(x, y), d(z, w),
d(x,y)+d(y,z)+d(z,x)
, d(x,y)+d(y,w)+d(w,x)
,
, d(y,z)+d(z,w)+d(w,y)
2
2
d(x,z)+d(z,w)+d(w,x)
2
2
(1.4)
for any x, y, z, w ∈ X. The extended four-point condition yields an O(n4) time
algorithm to recognize a subtree distance.
2
This paper addresses the following problem, which we call the subtree distance
reconstruction problem.
Problem 1. Given a subtree distance d : X × X → R on a finite set X, find a
representation (T , φ) of d.
Ando and Sato [1] proposed an O(n3) time algorithm for this problem. Their al-
gorithm consists of three steps: (1) identify a subset V0 = {x ∈ X d(y, z) ≤
d(x, y) + d(x, z) (y, z ∈ X)}, (2) find a representation (T , φ) for the restriction of
d onto V0, and (3) for x ∈ X \ V0, locate φ(x) in T by examining a connected
components of T \ φ(x).
In this study, we propose the following theorems. We define a minimal represen-
tation for a subtree distance in the same manner as a minimal representation for a
tree metric.
Theorem 2. For a subtree distance d : X ×X → R, a minimal representation (T , φ)
is uniquely determined by d.
Theorem 3. There exists an O(n2) time algorithm that finds, for any subtree dis-
tance d : X × X → R, its unique minimal representation, where n = X.
The proof of Theorem 3 is constructive. Similar to [1], our algorithm consists
of three parts: (1) identify the set of objects L that are mapped to the leaves, (2)
find the minimal representation (T , φ) for the restriction of d onto L, and (3) for
x ∈ X \ L, locate φ(x) in T by measuring the distances from the leaves. Since Steps
1 and 3 can be implemented with a time complexity of O(n2), and there is an O(n2)
time algorithm for Step 2 [3], the total time complexity of the algorithm is O(n2).
Note that even if we know d : X × X → R is a tree metric, Ω(n2) time is required to
reconstruct a tree [4]. Therefore, our algorithm achieves the optimal time complexity.
This algorithm can also be used to recognize a subtree distance by checking the
failure or inconsistency during the process and by verifying equations (1.3) after the
reconstruction.
Corollary 4. There exists an O(n2) time algorithm that determines whether a given
input d : X × X → R is a subtree distance or not, where n = X.
2. Proofs
We assume that there are no objects x, y ∈ X such that d(x, z) = d(y, z) for all
z ∈ X. This assumption is satisfied by removing such elements after lexicographic
sorting, which requires O(n2) time [7]. Clearly, this preprocessing does not change
3
the minimal representation. We also assume that X ≥ 3. Otherwise, the theorems
trivially hold.
First, we prove Theorem 2. We identify the properties of a minimal representa-
tion.
Lemma 5. Let (T , φ) be a minimal representation of a subtree distance d : X ×X →
R. Then, the following properties hold.
1. For each edge e ∈ E(T ), the length of e is positive.
2. For each leaf vertex u ∈ V(T ), there exists x ∈ X such that φ(x) = {u}.
Proof. 1. If there is an edge e of zero length, we can contract e from the represen-
tation.
2. Let u be a leaf vertex of T . If there is no x ∈ X with u ∈ φ(x), we can remove
u from the representation to obtain a smaller representation. Suppose that, for all
x ∈ X with u ∈ φ(x), φ(x) contains at least two elements. Then, these φ(x)s contain
the unique adjacent vertex v of u. Since d(v, w) ≤ d(u, w) for all w ∈ V(T ) \ {u}, u
does not contribute any shortest paths in the tree. Therefore, we can remove u from
the representation.
Motivated by Property 2 in Lemma 5, we introduce the following definition. For
a minimal representation (T , φ), an object x ∈ X is a leaf object if φ(x) = {u} for
some leaf u ∈ V(T ).
We defined a leaf object by specifying a minimal representation. However, as
shown below, the set of leaf objects is uniquely determined by d. First, we show that
there exists an object that is a leaf object for any minimal representation.
Lemma 6. Let (r, r′) ∈ argmaxx,y∈X d(x, y). Then, for any minimal representation,
r and r′ are leaf objects.
Proof. For any tree, the farthest pair is attained by a pair of leaves.
Next, we show that the leaf objects are characterized by d and a leaf object r.
Lemma 7. Let (T , φ) be a minimal representation of subtree distance d : X × X →
R, and let r ∈ X be a leaf object. An object x ∈ X \ {r} is a leaf object if and only
if d(y, r) < d(x, y) + d(x, r) for all y ∈ X \ {x, r}.
Proof. (The "if" part). Suppose x is not a leaf object. Then, there is another leaf
object y ∈ X \ {x, r} such that the path from φ(r) to φ(y) intersects φ(x). By
considering distances among φ(r), φ(y) and φ(x) on this path, we have d(r, y) ≥
d(x, r) + d(x, y).
4
(The "only if" part). Suppose x is a leaf object. Then, for any y ∈ X \ {x, r}
the shortest path from φ(y) to φ(r) never intersects φ(x). Thus, we have d(y, r) <
d(x, y) + d(x, r). Here, we used the assumption that there is no x, y ∈ X such that
d(x, z) = d(y, z) for all z ∈ X.
Since we can take a leaf object r universally by Lemma 6, and the condition
in Lemma 7 is described without specifying the underlying representation, we can
conclude that the set of leaf objects is uniquely determined by d. Thus, we obtain
the following corollary.
Corollary 8. Any minimal representation of a subtree distance has the same leaf
objects.
Now, we observe that the dissimilarity dL : L × L → R obtained by restricting
d : X × X → R on the set of leaf objects L ⊆ X forms a tree metric because the leaf
objects are mapped to singletons. Since the minimal representation of a tree metric
is unique [2], any minimal representation of d has the same topology as the minimal
representation of dL.
The remaining issue is to show that each non-leaf object x ∈ X \ L is uniquely
located in the minimal representation (T , φ) of dL. This is clear because any con-
nected subgraph in a tree is uniquely identified by the distances from the leaves.
More precisely, we obtain the following explicit representation.
We first consider T as a continuous object. We fix a leaf object r ∈ L. For each
leaf object x ∈ L \ {r}, there exists a unique path path(φ(r), φ(x)) from φ(r) to φ(x)
in T . Let I(r, a; x, b) be the interval on the path having a distance of at least a from
φ(r) and at least b from φ(x), i.e.,
I(r, a; x, b) = {u ∈ path(φ(r), φ(x)) : dT (φ(r), u) ≥ a, dT (φ(x), u) ≥ b}.
(2.1)
For U ⊆ V(T ), we denote by U the subgraph of T induced by U. Note that both
I(r, a; x, b) and U are continuous objects. By using these notations, we obtain the
following.
Lemma 9. Let d : X ×X → R be a subtree distance and L be the set of leaf objects.
Let (T , φ) be the minimal representation of dL. Fix a leaf object r ∈ L. Then, we
have for each non-leaf object z ∈ X \ L
φ(z) = [
x∈L\{r}
I(r, d(r, z); x, d(x, z)).
(2.2)
5
Proof. Since φ(z) is a connected subgraph, the intersection of φ(z) and the path from
φ(r) to φ(x) is the interval I(r, d(r, z); x, d(x, z)). Since any tree is covered by the
paths from a fixed leaf φ(r) and the other leaves φ(x) for x ∈ L \ {r}, we have
φ(z) =
[
= [
x∈L\{r}
x∈L\{r}
path(φ(r), φ(x))
∩ φ(z)
(cid:16)path(φ(r), φ(x)) ∩ φ(z)(cid:17)
= [
x∈L\{r}
I(r, d(r, z); x, d(x, z)).
(2.3)
By placing the vertices on the boundaries of φ(z) for z ∈ X \ L and then let-
ting φ(z) be the vertices intersecting φ(z) for z ∈ X \ L, we obtain the minimal
representation of d. This completes the proof of Theorem 2. Note that the number
of the vertices of the minimal representation is O(n2) since each φ(z) has at most
L = O(n) boundaries.
Now, we prove Theorem 3. The above proof of Theorem 2 is constructive, and it
provides the following algorithm:
1. Identify the set L of leaf objects by Lemmas 6 and 7.
2. Find the minimal representation of dL by the existing algorithm.
3. Locate the non-leaf objects by Lemma 9.
We evaluate the time complexity of this algorithm. Step 1 is conducted in O(n2) time
for finding a leaf object r ∈ X and O(n2) time for finding other leaf objects. Step 2
is performed in O(n2) time by using Culberson and Rudnicki's algorithm [3]. Also,
Step 3 is performed in O(n2) time by equation (2.2) since, for each z, it processes
each edge at most once. Hence, Theorem 3 is proved.
Acknowledgments
The authors thank Hiroshi Hirai and anonymous referees for helpful comments.
This work was supported by the Japan Society for the Promotion of Science, KAK-
ENHI Grant Number 15K00033.
6
References
[1] Kazutoshi Ando and Koki Sato. An algorithm for finding a representation of a
subtree distance. Journal of Combinatorial Optimization, 36(3):742 -- 762, 2018.
[2] Peter Buneman. The recovery of trees from measures of dissimilarity. In D.G.
Kendall and P. Tautu, editors, Mathematics in the Archeological and Historical
Sciences, pages 387 -- 395. Edinburgh University Press, 1971.
[3] Joseph C. Culberson and Piotr Rudnicki. A fast algorithm for constructing trees
from distance matrices. Information Processing Letters, 30(4):215 -- 220, 1989.
[4] Jotun J. Hein. An optimal algorithm to reconstruct trees from additive distance
data. Bulletin of Mathematical Biology, 51(5):597 -- 603, 1989.
[5] Hiroshi Hirai. Characterization of the distance between subtrees of a tree by the
associated tight span. Annals of Combinatorics, 10(1):111 -- 128, 2006.
[6] Naruya Saitou and Masatoshi Nei. The neighbor-joining method: a new method
for reconstructing phylogenetic trees. Molecular Biology and Evolution, 4(4):406 --
425, 1987.
[7] Juraj Wiedermann. The complexity of lexicographic sorting and searching. In
Proceedings of the 8th International Symposium on Mathematical Foundations of
Computer Science, pages 517 -- 522, 1979.
[8] Edward O. Wiley and Bruce S. Lieberman. Phylogenetics: Theory and Practice
of Phylogenetic Systematics. John Wiley & Sons, 2011.
[9] K. A. Zaretskii. Constructing a tree on the basis of a set of distances between
the hanging vertices. Uspekhi Matematicheskikh Nauk, 20(6):90 -- 92, 1965.
7
|
1711.02120 | 1 | 1711 | 2017-11-06T19:14:43 | Small Resolution Proofs for QBF using Dependency Treewidth | [
"cs.DS",
"cs.LO"
] | In spite of the close connection between the evaluation of quantified Boolean formulas (QBF) and propositional satisfiability (SAT), tools and techniques which exploit structural properties of SAT instances are known to fail for QBF. This is especially true for the structural parameter treewidth, which has allowed the design of successful algorithms for SAT but cannot be straightforwardly applied to QBF since it does not take into account the interdependencies between quantified variables.
In this work we introduce and develop dependency treewidth, a new structural parameter based on treewidth which allows the efficient solution of QBF instances. Dependency treewidth pushes the frontiers of tractability for QBF by overcoming the limitations of previously introduced variants of treewidth for QBF. We augment our results by developing algorithms for computing the decompositions that are required to use the parameter. | cs.DS | cs |
Small Resolution Proofs for QBF using
Dependency Treewidth
Eduard Eiben, Robert Ganian, and Sebastian Ordyniak
Algorithms and Complexity Group, TU Wien, Vienna, Austria
[email protected], [email protected], [email protected]
Abstract
In spite of the close connection between the evaluation of quantified Boolean formulas (QBF) and
propositional satisfiability (SAT), tools and techniques which exploit structural properties of SAT
instances are known to fail for QBF. This is especially true for the structural parameter treewidth,
which has allowed the design of successful algorithms for SAT but cannot be straightforwardly
applied to QBF since it does not take into account the interdependencies between quantified
variables.
In this work we introduce and develop dependency treewidth, a new structural parameter
based on treewidth which allows the efficient solution of QBF instances. Dependency treewidth
pushes the frontiers of tractability for QBF by overcoming the limitations of previously introduced
variants of treewidth for QBF. We augment our results by developing algorithms for computing
the decompositions that are required to use the parameter.
Digital Object Identifier 10.4230/LIPIcs...
1
Introduction
The problem of evaluating quantified Boolean formulas (QBF) is a generalization of the
propositional satisfiability problem (SAT) which naturally captures a range of computational
tasks in areas such as verification, planning, knowledge representation and automated rea-
soning [11, 19, 23, 24]. QBF is the archetypical PSpace-complete problem and is therefore
believed to be computationally harder than NP-complete problems such as SAT [17, 20, 30].
In spite of the close connection between QBF and SAT, many of the tools and techniques
which work for SAT are known not to help for QBF, and dynamic programming based on
the structural parameter treewidth [2, 31] is perhaps the most prominent example of this
behavior. Treewidth is a highly-established measure of how "treelike" an instance is, and in
the SAT setting it is known that n-variable instances of treewidth at most k can be solved
in time at most f(k) · n [31] for a computable function f. Algorithms with running time in
this form (i.e., f(k) · nO(1), where k is the parameter and the degree of the polynomial of n
is independent of k) are called fixed-parameter algorithms, and problems which admit such
an algorithm (w.r.t. a certain parameter) belong to the class FPT. Furthermore, in the SAT
setting, treewidth allows us to do more than merely solve the instance:
it is also possible
to find a so-called resolution proof [7, 5]. If the input was a non-instance, such a resolution
proof contains additional information on "what makes it unsatisfiable" and hence can be
more useful than outputting a mere Reject in practical settings.
In the QBF setting, the situation is considerably more complicated. It is known that
QBF instances of bounded treewidth remain PSpace-complete [2], and the intrinsic reason
for this fact is that treewidth does not take into account the dependencies that arise between
variables in QBF. So far, there have been several attempts at remedying this situation by
introducing variants of treewidth which support fixed-parameter algorithms for QBF: prefix
pathwidth (along with prefix treewidth) [12] and respectful treewidth [2], along with two other
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl -- Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
XX:2
Small Resolution Proofs for QBF using Dependency Treewidth
parameters [1, 6] which originate from a different setting but can also be adapted to obtain
fixed-parameter algorithms for QBF. We refer to Subsection 3.2 for a comparison of these pa-
rameters. Aside from algorithms with runtime guarantees, it is worth noting that empirical
connections between treewidth and QBF have also been studied in the literature [21, 22].
In this work we introduce and develop dependency treewidth, a new structural param-
eter based on treewidth which supports fixed-parameter algorithms for QBF. Dependency
treewidth pushes the frontiers of tractability for QBF by overcoming the limitations of both
the previously introduced prefix and respectful variants. Compared to the former, this new
parameter allows the computation of resolution proofs analogous to the case of classical
treewidth for SAT instances. Prefix pathwidth relies on entirely different techniques to
solve QBF and does not yield small resolution proofs. Moreover, the running time of the
fixed-parameter algorithm which uses prefix pathwidth has a triple-exponential dependency
on the parameter k, while dependency treewidth supports a considerably more efficient
O(32knk)-time algorithm for QBF.
Unlike respectful treewidth and its variants, which only take the basic dependencies
between variables into account, dependency treewidth can be used in conjunction with the
so-called dependency schemes introduced by Samer and Szeider [25, 27], see also the work of
Biere and Lonsing [3]. Dependency schemes allow an in-depth analysis of how the assignment
of individual variables in a QBF depends on other variables, and research in this direction has
uncovered a large number of distinct dependency schemes with varying complexities. The
most basic dependency scheme is called the trivial dependency scheme [25], which stipulates
that each variable depends on all variables with distinct quantification which precede it in
the prefix. Respectful treewidth in fact coincides with dependency treewidth when the trivial
dependency scheme is used, but more advanced dependency schemes allow us to efficiently
solve instances which otherwise remain out of the reach of state-of-the-art techniques.
Crucially, all of the structural parameters mentioned above require a so-called decom-
position in order to solve QBF; computing these decompositions is typically an NP-hard
problem. A large part of our technical contribution lies in developing algorithms to com-
pute decompositions for dependency treewidth. Without such algorithms, it would not be
possible to use the parameter unless a decomposition were supplied as part of the input (an
unrealistic assumption in practical settings). It is worth noting that all of these algorithms
can also be used to find respectful tree decompositions, where the question of finding suitable
decompositions was left open [2]. We provide two algorithms for computing dependency tree
decompositions, each suitable for use under different situations.
The article is structured as follows. After the preliminaries, we introduce the parameter
and show how to use it to solve QBF (assuming a decomposition has already been computed).
This section also contains an in-depth overview and comparison of previous work in the
area. A separate section then introduces other equivalent characterizations of dependency
treewidth. The last technical section contains our algorithms for finding dependency tree
decompositions, after which we provide concluding notes and remarks.
Preliminaries
2
For i ∈ N, we let [i] denote the set {1, . . . , i}. We refer to the book by Diestel [9] for standard
graph terminology. Given a graph G, we denote by V (G) and E(G) its vertex and edge set,
respectively. We use ab as a shorthand for the edge {a, b}. For V 0 ⊆ V (G), the guards of V 0
(denoted δ(V 0)) are the vertices in V (G) \ V 0 with at least one neighbor in V 0.
We refer to the standard textbooks [10, 14] for an in-depth overview of parameterized
Eduard Eiben, Robert Ganian, and Sebastian Ordyniak
XX:3
complexity theory. Here, we only recall that a parameterized problem (Q, κ) is a problem
Q ⊆ Σ∗ together with a parameterization κ: Σ∗ → N, where Σ is a finite alphabet. A
parameterized problem (Q, κ) is fixed-parameter tractable (w.r.t. κ), in short FPT, if there
exists a decision algorithm for Q, a computable function f : N → N, and a polynomial
function p: N → N, such that for all x ∈ Σ∗, the running time of the algorithm on x
is at most f(κ(x)) · p(x). Algorithms with this running time are called fixed-parameter
algorithms.
2.1 Quantified Boolean Formulas
For a set of propositional variables K, a literal is either a variable x ∈ K or its negation ¯x.
A clause is a disjunction over literals. A propositional formula in conjunctive normal form
(i.e., a CNF formula) is a conjunction over clauses. Given a CNF formula φ, we denote the
set of variables which occur in φ by var(φ). For notational purposes, we will view a clause
as a set of literals and a CNF formula as a set of clauses.
A quantified Boolean formula is a tuple (φ, τ) where φ is a CNF formula and τ is a
sequence of quantified variables, denoted var(τ), which satisfies var(τ) ⊇ var(φ); then φ is
called the matrix and τ is called the prefix. A QBF (φ, τ) is true if the formula τ φ is true.
A quantifier block is a maximal sequence of consecutive variables with the same quantifier.
An assignment is a mapping from (a subset of) the variables to {0, 1}.
The primal graph of a QBF I = (φ, τ) is the graph GI defined as follows. The vertex set
of GI consists of every variable which occurs in φ, and st is an edge in GI if there exists a
clause in φ containing both s and t.
2.2 Dependency Posets for QBF
Before proceeding, we define a few standard notions related to posets which will be used
throughout the paper. A partially ordered set (poset) V is a pair (V,≤V ) where V is a set
and ≤V is a reflexive, antisymmetric, and transitive binary relation over V . A chain W of
V is a subset of V such that x ≤V y or y ≤V x for every x, y ∈ W. A chain partition of
V is a tuple (W1, . . . , Wk) such that {W1, . . . , Wk} is a partition of V and for every i with
1 ≤ i ≤ k the poset induced by Wi is a chain of V. An anti-chain A of V is a subset of
V such that for all x, y ∈ A neither x ≤V y nor y ≤V x. The width (or poset-width) of a
poset V, denoted by width(V), is the maximum cardinality of any anti-chain of V. A poset
of width 1 is called a linear order. A linear extension of a poset P = (P,≤P ) is a relation
(cid:22) over P such that x (cid:22) y whenever x ≤P y and the poset P∗ = (P,(cid:22)) is a linear order. A
subset A of V is downward-closed if for every a ∈ A it holds that b ≤V a =⇒ b ∈ A. A
reverse of a poset is obtained by reversing each relation in the poset. For brevity we will
often write ≤V to refer to the poset V := (V,≤V ).
(cid:73) Proposition 1 ([13]). Let V be a poset. Then in time O(width(V)·kVk2), it is possible to
compute both width(V) = w and a corresponding chain partition (W1, . . . , Ww) of V.
We use dependency posets to provide a general and formal way of speaking about the
various dependency schemes introduced for QBF [25]. It is important to note that depen-
dency schemes in general are too broad a notion for our purposes; for instance, it is known
that some dependency schemes do not even give rise to sound resolution proof systems. Here
we focus solely on so-called permutation dependency schemes [28], which is a general class
containing all commonly used dependency schemes that give rise to sound resolution proof
XX:4
Small Resolution Proofs for QBF using Dependency Treewidth
systems. This leads us to our definition of dependency posets, which allow us to capture all
permutation dependency schemes.
Given a QBF I = (φ, τ), a dependency poset V = (var(φ),≤I) of I is a poset over var(φ)
with the following properties:
1. for all x, y ∈ var(φ), if x ≤I y, then x is before y in the prefix, and
2. given any linear extension (cid:22) of V, the QBF I0 = (φ, τ(cid:22)), obtained by permutation of the
prefix τ according the (cid:22), is true iff I is true.
The trivial dependency scheme is one specific example of a permutation dependency
scheme. This gives rise to the trivial dependency poset, which sets x ≤ y whenever x, y
are in different quantifier blocks and x is before y in the prefix. However, more refined
permutation dependency schemes which give rise to other dependency posets are known
to exist and can be computed efficiently [25, 28]. In particular, it is easy to verify that a
dependency poset can be computed from any permutation dependency scheme in polynomial
time (by computing the transitive closure).
To illustrate these definitions, consider the following QBF:
∃a∀b∃c(a ∨ c) ∧ (b ∨ c)
Then the trivial dependency poset would set a ≤ b ≤ c. However, for instance the resolu-
tion path dependency poset (arising from the resolution path dependency scheme [32, 26])
contains a single relation b ≤ c (in this case, a is incomparable to both b and c).
2.3 Q-resolution
Q-resolution is a sound and complete resolution system for QBF [16]. Our goal here is to
formalize the required steps for the Davis Putnam variant of Q-resolution.
We begin with a bit of required notation. For a QBF I = (φ, τ) and a variable x ∈ var(φ),
let φx be the set of all clauses in φ containing the literal x and similarly let φ¯x be the set
of all clauses containing literal ¯x. We denote by res(I, x) the QBF I0 = (φ0, τ0) such that
τ0 = τ \ {x} and φ0 = φ \ (φx ∪ φ¯x) ∪ {(D \ {x}) ∪ (C \ {¯x})D ∈ φx; C ∈ φ¯x}; informally,
the two clause-sets are pairwise merged to create new clauses which do not contain x. For
a QBF I = (φ, τ) and a variable x ∈ var(φ) we denote by I \ x the QBF I = (φ0, τ \ {x}),
where we get φ0 from φ by removing all occurrences of x and ¯x.
(cid:73) Lemma 2. Let I = (φ, τ) and x ∈ var(φ) be the last variable in τ. If x is existentially
quantified, then I is true if and only if res(I, x) is true.
Proof. Assume I is true and let F be a winning strategy for existential player in I [16].
We will show that F is also a winning existential strategy in res(I, x). Assume that the
existential player played according to F in res(I, x), but there is a clause B that is not
satisfied at the end of the game. Clearly B ∈ {(D \ {x}) ∪ (C \ {¯x})D ∈ φx; C ∈ φ¯x},
otherwise B is also a clause of I and hence it has to be satisfied due to the existential player
using F. In particular, B = (D \ {x}) ∪ (C \ {¯x}) for some D ∈ φx and C ∈ φ¯x. Now if
F assigns x to 1, since F is a winning strategy it follows that C must be satisfied by some
other literal, and hence B must also be satisfied -- a contradiction. A symmetric argument
also leads to a contradiction if F assigns x to 0.
Assume now that res(I, x) is true and let F be a winning strategy for the existential
player in res(I, x). Now suppose that after all variables of res(I, x) have been assigned
according to the strategy F, there is some D ∈ φx such that D \ {x} is false. Since
(D \ {x}) ∪ (C \ {¯x}) is true for all C ∈ φ¯x, it means that all C ∈ φ¯x are true before we
Eduard Eiben, Robert Ganian, and Sebastian Ordyniak
XX:5
assign x, and our strategy can assign x to 1. On the other hand if D \ {x} is true for all D
(cid:74)
in φx, our strategy assigns x to 0 and again satisfies all clauses of I.
(cid:73) Lemma 3. Let I = (φ, τ) and x ∈ var(φ) be the last variable in τ. If x is universally
quantified, then I is true if and only if I \ x is true.
Proof. We will prove an equivalent statement: I \ x is false if and only if I is false. It is easy
to see that if F is a winning strategy for the universal player in I, then if he plays according
F, then when the universal should assign the last variable x there is either a clause that
is already false and does not contain x, or a clause that contains x and is false after an
assignment of x according to F. In both cases I \ x contains a clause that is false.
On the other hand, assume F is a winning strategy for the universal player in I \ x and
the universal plays according to F in I until all variables but x are assigned. Clearly, this
strategy leads to an assignment of variables such that there is a clause B in I \ x, which
is false under this assignment. From the definition of I \ x, it is easy to observe that I
contains a clause B0, which is equal to one of the following: B, B ∪ {x}, or B ∪ {¯x}. It is
straightforward to extend F in a way that whenever F falsifies a clause B, then the new
(cid:74)
strategy falsifies the clause B0.
2.4 Treewidth
Here we will introduce three standard characterizations of treewidth [18]: tree decomposi-
tions, elimination orderings, and cops and robber games. These will play an important role
later on, when we define their counterparts in the dependency treewidth setting and use
these in our algorithms.
following properties hold: (T1)S
Tree decomposition: A tree decomposition of a graph G is a pair (T, χ), where T is
a rooted tree and χ is a function from V (T) to subsets of V (G), called a bag, such that the
t∈V (T ) χ(t) = V (G), (T2) for each uv ∈ E(G) there exists
t ∈ V (T) such that u, v ∈ χ(t), and (T3) for every u ∈ V (G), the set Tu = {t ∈ V (T) : u ∈
χ(t)} induces a connected subtree of T.
To distinguish between the vertices of the tree T and the vertices of the graph G, we will
refer to the vertices of T as nodes. The width of the tree decomposition T is maxt∈T χ(t)−1.
The treewidth of G, tw(G), is the minimum width over all tree decompositions of G. For
a node t ∈ V (T), we denote by Tt the subtree of T rooted at t. The following fact will be
useful later on:
(cid:73) Proposition 4. Let T = (T, χ) be a tree decomposition of a graph G and t ∈ V (T) a node
with parent p in T. Then χ(p) ∩ χ(t) separates χ(Tt) \ χ(p) from the rest of G.
Elimination ordering: An elimination ordering of a graph is a linear order of its
vertices. Given an elimination ordering φ of the graph G, the fill-in graph H of G w.r.t. φ
is the unique minimal graph such that:
V (G) = V (H).
E(H) ⊇ E(G).
If 0 ≤ k < i < j ≤ n and vi, vj ∈ NH(vk), then vivj ∈ E(H).
The width of elimination ordering φ is the maximum number of neighbors of any vertex v
that are larger than v (w.r.t. φ) in H.
(Monotone) cops and robber game: The cops and robber game is played between
two players (the cop-player and the robber-player) on a graph G. A position in the game is
a pair (C, R) where C ⊆ V (G) is the position of the cop-player and R is a (possibly empty)
XX:6
Small Resolution Proofs for QBF using Dependency Treewidth
connected component of G \ C representing the position of the robber-player. A move from
position (C, R) to position (C0, R0) is legal if it satisfies the following conditions:
CM1 R and R0 are contained in the same component of G \ (C ∩ C0),
CM2 δ(R) ⊆ C0.
A play P is a sequence (∅, R0), . . . , (Cn, Rn) of positions such that for every i with 1 ≤ i < n
it holds that the move from (Ci, Ri) to (Ci+1, Ri+1) is legal; the cop-number of a play is
maxi≤n Ci. A play P is won by the cop-player if Rn = ∅, otherwise it is won by the
robber-player. The cop-number of a strategy for the cop player is maximum cop-number
over all plays that can arise from this strategy. Finally, the cop-number of G is the minimum
cop-number of a winning strategy for the cop player.
(cid:73) Proposition 5 ([18]). Let G be a graph. The following three claims are equivalent:
G has treewidth k,
G has an elimination ordering of width k,
G has cop-number k.
3 Dependency Treewidth for QBF
We are now ready to introduce our parameter. We remark that in the case of dependency
treewidth, it is advantageous to start with a counterpart to the elimination ordering char-
acterization of classical treewidth, as this is used extensively in our algorithm for solving
QBF. We provide other equivalent characterizations of dependency treewidth (representing
the counterparts to tree decompositions and cops and robber games) in Section 4; these
are not only theoretically interesting, but serve an important role in our algorithms for
computing the dependency treewidth.
Let I = (φ, τ) be a QBF instance with a dependency poset P. An elimination ordering
of GI is compatible with P if it is a linear extension of the reverse of P; intuitively, this
corresponds to being forced to eliminate variables that have the most dependencies first.
For instance, if P is a trivial dependency poset then a compatible elimination ordering must
begin eliminating from the rightmost block of the prefix. We call an elimination ordering of
GI that is compatible with P a P-elimination ordering (or dependency elimination ordering).
The dependency treewidth w.r.t. P is then the minimum width of a P-elimination ordering.
3.1 Using dependency treewidth
Our first task is to show how dependency elimination orderings can be used to solve QBF.
(cid:73) Theorem 6. There is an algorithm that given
1. a QBF I with n variables and m clauses,
2. a dependency poset P for I, and
3. a P-elimination ordering π of width k,
decides whether I is true in time O(32kkn). Moreover, if I is false, then the algorithm
outputs a Q-resolution refutation of size O(3kn).
Proof. Let I = (φ, τ) and let x1, . . . , xn denote the variables of φ such that xi ≤π xi+1
for all 1 ≤ i < n. From the definition of the dependency poset and the fact that π is a
dependency elimination ordering, it follows that the QBF instance I0 = (φ, τ0), where τ0 is
the reverse of π, is true if and only if I is true.
Eduard Eiben, Robert Ganian, and Sebastian Ordyniak
XX:7
From now on let H be the fill-in graph of the primal graph of I with respect to π, and
To solve I we use a modification of the Davis Putnam resolution algorithm [7]. We start
with instance I0 and recursively eliminate the last variable in the prefix using Lemmas 2
and 3 until we either run out of variables or we introduce as a resolvent a non-tautological
clause that is either empty or contains only universally quantified variables. We show that
each variable we eliminate has the property that it only shares clauses with at most k other
variables, and in this case we introduce at most 3k clauses of size at most k at each step.
let us define Ii = (φi, τ i) for 1 ≤ i ≤ n as follows:
1. I1 = I0,
2. Ii+1 = Ii \ xi if xi is universally quantified, and
3. Ii+1 = res(Ii, xi), if xi is existentially quantified.
Note that xi is always the last variable of the prefix of Ii and it follows from Lemmas 2
and 3 that Ii+1 is true if and only if Ii is true. Moreover, In only contains a single variable,
and hence can be decided in constant time. In the following, we show by induction that Ii+1
contains at most 3k new clauses, i.e., clauses not contained in Ii. To this end, we show and
use the fact that both φi
x and φi¯x contain at most 3k clauses, and this is sufficient to ensure
a small Q-resolution refutation if the instance is false.
(cid:73) Claim 1. The instance Ii+1 contains at most 3k clauses not contained in Ii. Furthermore,
if the primal graph of Ii is a subgraph of H, then so is the primal graph of Ii+1.
Proof of the Claim. We distinguish two cases: either xi is universal, or existential. In the
former case, it is easily observed that the primal graph of Ii \ xi is a subgraph of the primal
graph of Ii obtained by removing xi from the graph. Hence, in this case the primal graph
of Ii+1 is a subgraph of H as well. Moreover, we do not add any new clauses, only remove
xi from already existing ones.
In the later case, let xi be an existentially quantified variable. Let xpxq be an edge in
the primal graph of res(Ii, xi), but not in the primal graph of Ii. Clearly, xp and xq are in
a newly added clause B = (D \ {x}) ∪ (C \ {¯x}) such that D ∈ φi
. But that
means that both xp and xq are in a clause with xi in Ii. Hence, xpxi and xqxi are both
edges in the primal graph of Ii and also in H. But, p > i and q > i, since otherwise xp and
xq will not appear in Ii+1. Since H is the fill-in graph w.r.t. π, H contains xpxq as well and
H is a supergraph of the primal graph of Ii. Moreover, as π is an elimination ordering of
the primal graph of I of width k, there are at most k variables in φi that appear with xi in
a clause in Ii. Hence there are at most 3k different clauses containing these variables, and
so φi
contain at most 3k clauses of size at most k + 1. Finally, the set of new clauses
xi
} contains at most 3k clauses too, and the claim
{(D \ {xi}) ∪ (C \ { ¯xi})D ∈ φi
(cid:74)
follows.
and C ∈ φi¯xi
xi
, φi¯xi
; C ∈ φi¯xi
xi
Since I is equivalent to In, it is easy to see that if I is false, then In either contains the empty
clause, or two clauses {xn} and { ¯xn}, or xn is universally quantified. In these cases it is
easy to see that the union of clauses of all Ii (including the empty clause, if it is not already
in In) is a Q-resolution refutation. Moreover, it follows from Claim 1 that this Q-resolution
refutation has at most m + 3kn clauses.
For the runtime analysis, the time necessary to compute a variable-clause adjacency list
data structure (which will be useful for the next step) is upper-bounded by O(mk) due to
clauses having size at most k (since the width of the elimination ordering is at most k).
If x is universal, then I \ x can be computed in time O(3k), since x occurs in at most 3k
clauses. If x is existential, then we need to compute the new clauses which takes time at
XX:8
Small Resolution Proofs for QBF using Dependency Treewidth
most O(32k · k) since there are at most 32k pairs of clauses containing x and each such clause
has size at most k.
Finally, observe that the total number of clauses cannot exceed n · k (because the
treewidth of the matrix is also bounded by k), and so both O(mk) and m are superseded
(cid:74)
by the other term in the runtime and the resolution refutation size.
3.2 A Comparison of Decompositional Parameters for QBF
As was mentioned in the introduction, two dedicated decompositional parameters have pre-
viously been introduced specifically for evaluating quantified Boolean formulas: prefix path-
width (and, more generally, prefix treewidth) [12] and respectful treewidth [2]. The first task
of this section is to outline the advantages of dependency treewidth compared to these two
parameters.
Prefix pathwidth is based on bounding the number of viable strategies in the classical
two-player game characterization of the QBF problem [12]. As such, it decomposes the
dependency structure of a QBF instance beginning from variables that have the least de-
pendencies (i.e., may appear earlier in the prefix). On the other hand, our dependency
treewidth is based on Q-resolution and thus decomposes the dependency structure begin-
ning from variables that have the most dependencies (i.e., may appear last in the prefix).
Lemma 7 shows that both approaches are, in principle, incomparable. That being said,
dependency treewidth has two critical advantages over prefix treewidth/pathwidth:
1. dependency treewidth outputs small resolution proofs, while it is not at all clear whether
the latter can be used to obtain such resolution proofs;
2. dependency treewidth supports a single-exponential fixed-parameter algorithm for QBF
(Theorem 6), while the latter uses a prohibitive triple-exponential algorithm [12].
(cid:73) Lemma 7. Let us fix the trivial dependency poset. There exist infinite classes A,B of
QBF instances such that:
a. A has unbounded dependency treewidth but prefix pathwidth at most 1;
b. B has unbounded prefix pathwidth (and prefix treewidth) but dependency treewidth at
most 1.
Proof. a. Let
Ai = ∃x1, . . . , xi∀y∃x(y ∨ x) ∧ i^
(xj ∨ x).
j=1
The trivial dependency poset Pi for Ai would be {x1, . . . , xi} ≤ y ≤ x. Hence every Pi-
elimination ordering must start with x, and then the width of such an ordering would be i+1.
On the other hand, one can observe [12] that the path decomposition Q = (Q1, . . . , Qi+1),
where Qj = {xj, x} for 1 ≤ j ≤ i and Qi+1 = {y, x}, is a prefix path-decomposition w.r.t.
Pi of width 1.
b. Consider the following formula with alternating prefix:
Bi = ∃x1∀x2 . . .∀x2i∃x2i+1
((xj ∨ x2j) ∧ (xj ∨ x2j+1).
2i−1^
j=1
Since the quantifiers in the prefix of Bi alternate, the trivial dependency poset Pi for Bi
would be the linear order x1 ≤ x2 ≤ ··· ≤ x2i. It is readily observed that the primal graph
of Bi is a balanced binary tree of depth i, and it is known that the pathwidth of such trees is
Eduard Eiben, Robert Ganian, and Sebastian Ordyniak
XX:9
i−1 [8]. From the fact that pathwidth is a trivial lower bound for prefix pathwidth together
with previous work on prefix treewidth [12, Theorem 6], it follows that i−1 is a lower bound
on the prefix treewidth of Bi.
On the other hand, since Pi is linear order, the only elimination ordering compatible
with Pi is the reverse of Pi. Moreover, from the definition of Bi, it is easily seen that xj
has at most 1 neighbor that is smaller w.r.t. Pi, namely xbj/2c. Therefore, the dependency
(cid:74)
treewidth of Bi is 1.
Respectful treewidth coincides with dependency treewidth when the trivial dependency
scheme is used, i.e., represents a special case of our measure. Unsurprisingly, the use of more
advanced dependency schemes (such as the resolution path dependency scheme [32, 28])
allows the successful deployment of dependency treewidth on much more general classes
of QBF instances. Furthermore, dependency treewidth with such dependency schemes will
always be upper-bounded by respectful treewidth, and so algorithms based on dependency
treewidth will outperform the previously introduced respectful treewidth based algorithms.
(cid:73) Lemma 8. There exists an infinite class C of QBF instances such that C has unbounded
dependency treewidth with respect to the trivial dependency poset but dependency treewidth
at most 1 with respect to the resolution-path dependency poset.
Proof. Recall the previous example:
Ai = ∃x1, . . . , xi∀y∃x(y ∨ x) ∧ i^
(xj ∨ x).
j=1
We have already established that Ai has dependency treewidth i + 1 when the trivial
dependency poset is used. However, e.g., the resolution-path dependency poset [32, 26])
contains a single relation y ≤ x. Since the primal graph of Ai is a star, it is then easy to
verify that Ai has dependency treewidth 1 when the resolution-path dependency poset is
(cid:74)
used.
Finally, we note that the idea of exploiting dependencies among variables has also given
rise to similarly flavored structural measures in the areas of first-order model checking (first
order treewidth) [1] and quantified constraint satisfaction (CD-width1) [6]. Even though
the settings differ, Theorem 5.5 [1] and Theorem 5.1 [6] can both be translated to a basic
variant of Theorem 6. We note that this readily-obtained variant of Theorem 6 would not
account for dependency schemes. We conclude this subsection with two lemmas which show
that there are classes of QBF instances that can be handled by our approach but are not
covered by the results of Adler, Weyer [1] and Chen, Dalmau [6].
Before we compare the dependency treewidth with the CD-width of Chen and Dalmau,
we will first define their parameter (or, more specifically, provide a translation into the QBF
setting). CD-width is also based on an elimination ordering, however with a slight modifi-
cation: they also use the fact that we can eliminate universal variables without introducing
new clauses (see Lemma 3) and therefore we can eliminate the last universal variable even
if it appears together with many variables. Formally the elimination ordering by Chen and
Dalmau is defined as follows:
1 We remark that in their paper, the authors refer to their parameter simply as "the width". For
disambiguation, here we call it CD-width (shorthand for Chen-Dalmau's width).
XX:10
Small Resolution Proofs for QBF using Dependency Treewidth
(cid:73) Definition 9. Let I = (φ, τ) be a QBF instance. Given a linear ordering (cid:22) of var(φ), the
CD-fill-in graph H(cid:22) of I is the unique minimal graph such that:
V (GI) = V (H(cid:22)).
E(H(cid:22)) ⊇ E(GI).
If u (cid:22) v (cid:22) w, u is existentially quantified, and v, w ∈ NH(cid:22)(u), then vw ∈ E(H(cid:22)).
A CD-elimination ordering of a QBF instance I = (φ, τ) is a linear order (cid:22) of var(φ) such
that for each existentially quantified x and each universally quantified y in var(φ):
If y is before x in τ, then x (cid:22) y.
If x is before y in τ and there is an edge xy in H(cid:22), then y (cid:22) x.
The width of a CD-elimination ordering (cid:22) is the maximum number of neighbors of an
existentially quantified vertex v that are larger than v (w.r.t. (cid:22)) in H(cid:22).
We remark that one could also use the less restrictive form of elimination orderings con-
sidered above with any dependency poset, obtaining a more general variant of Theorem 6.
However, such a notion would lose many of the nice structural properties used in our algo-
rithms for finding the decompositions; for instance, the result is no longer a restriction of
treewidth and does not have any immediate cops-and-robber game characterization. Hence
finding such an ordering of small width would become a more challenging problem.
(cid:73) Lemma 10. There exist infinite classes D, E of QBF instances such that:
a. D has unbounded CD-width but dependency treewidth at most 1 w.r.t. the resolution-path
b. E has unbounded dependency treewidth w.r.t. any dependency poset but CD-width at
dependency poset.
most 1.
Proof. a. Recall the previous example:
Ai = ∃x1, . . . , xi∀y∃x(y ∨ x) ∧ i^
(xj ∨ x).
j=1
We have already established that Ai has dependency treewidth 1 when the resolution-path
dependency poset is used. To establish the other directions it suffices to observe that Ex-
ample 3.6 of Chen and Dalmau [6] uses the class of prefixed graphs obtained from A by
taking primal graph of instances Ai and keeping the same prefix as an example of class with
unbounded elimination width.
b. Let
Ei = ∀x1∀x2 ···∀xi
(xp ∨ xq).
^
1≤p<q≤i
As Ei does not contain any existentially quantified variable, it is easy to observe that CD-
width of Ei is 0. However, the primal graph of Ei is a clique and hence dependency treewidth
is i − 1 w.r.t. every possible dependency poset.
(cid:74)
Similarly as above, before we compare first order treewidth to dependency treewidth, we
give some necessary definitions.
Let (cid:67) be a binary relation on the variables of some QBF instance I = (φ, τ). Then two
variables x and y are entangled with respect to (cid:67) and I, if x occurs in a clause with some
variable z such that y (cid:67) z and y occurs in a clause with some variable z such that x (cid:67) z.
(cid:73) Definition 11. Let I = (φ, τ) be a QBF instance. Then (cid:22)FO
to ⊆) binary relation on var(φ), such that the following holds:
is the minimal (with respect
I
Eduard Eiben, Robert Ganian, and Sebastian Ordyniak
XX:11
I
I
is reflexive.
is transitive.
(1) (cid:22)FO
(2) (cid:22)FO
(3) If x is before y in τ, x and y have different quantifiers in τ, and there is a sequence
x = z0, . . . , zn = y of variables such that for all 0 ≤ i < n we have that zi and zi+1 are
entangled w.r.t. (cid:22)I and I and that x (cid:22)I zi or y (cid:22)FO
zi, then x (cid:22)FO
y.
I
I
(cid:73) Definition 12. Let I = (φ, τ) be a QBF instance and x ∈ var(φ). The essential alternation
depth of x in I, denoted by eadI(x), is the maximum over all (cid:22)FO
-paths P ending in x of the
number of quantifier changes in P, adding +1 in case the first variable on P is existentially
quantified and +2 if it is universally quantified.
For a QBF instance I = (φ, τ), let VFO
and only if eadI(x) ≤ eadI(y). Notice that x (cid:22)FO
VFO
ordering that is compatible with VFO
ead) be the poset such that x ≤I
ead y if
y implies eadI(x) ≤ eadI(y) and hence
. The first order treewidth is the minimal width of an elimination
is an extension of (cid:22)FO
I = (var(φ),≤I
.
I
I
I
I
Now we are ready to prove the following lemma.
I
(cid:73) Lemma 13. There exists an infinite class F of QBF instances such that F has unbounded
first order treewidth but dependency treewidth at most 2 with respect to the resolution-path
dependency poset.
Proof. Let
Fi = ∀x2iy2i∃x2i−1y2i−1 ···∀x2y2∃x1y1∀z
(z ∨ x1 ∨ y1)
2i−1^
[(xj ∨ xj+1) ∧ (yj ∨ yj+1)] ∧ i^
j=1
j=1
(x2j−1 ∨ y1).
The resolution-path dependency poset would give us following relations: a chain y2i ≤
··· ≤ y2 ≤ y1 ≤ z and a chain x2i ≤ ··· ≤ x2 ≤ x1 ≤ z. It is readily observed that the
elimination ordering zx1x2 . . . x2iy1y2 . . . y2i with width 2 is compatible with this dependency
poset.
On the other hand, one can observe that x2j (cid:22)FO
entangled w.r.t. (cid:22)FO
and hence also x2j (cid:22)FO
elimination ordering that is compatible with VFO
and the width of such elimination ordering is at least i − 1.
Finally, we remark that if one were to show that VFO
Fi
Fi
Fi
x2j−1 and that y1 and x2j−1 are
y1 for all 1 ≤ j ≤ i.
It follows that any
I must have y1 before all xj for 2 ≤ j ≤ 2i
(cid:74)
is a dependency poset, then this
would imply that first order treewidth is a special case of dependency treewidth (for this
dependency poset). However, proving such a claim goes beyond the scope of this paper.
I
4 Dependency Treewidth: Characterizations
In this section we obtain other equivalent characterizations of dependency treewidth. The
purpose of this endeavor is twofold. From a theoretical standpoint, having several natural
characterizations (corresponding to the characterizations of treewidth) is not only interesting
but also, in some sense, highlights the solid foundations of a structural parameter. From
a practical standpoint, the presented characterizations play an important role in Section 5,
which is devoted to algorithms for finding optimal dependency elimination orderings.
Dependency tree decomposition: Let I be a QBF instance with primal graph G and
dependency poset P and let (T, χ) be a tree decomposition of G. Note that the rooted tree
XX:12
Small Resolution Proofs for QBF using Dependency Treewidth
T naturally induces a partial order ≤T on its nodes, where the smallest element is the root
and leaves form maximal elements. For a vertex v ∈ V (G), we denote by Fv(T) the unique
≤T -minimal node t of T with v ∈ χ(t), which is well-defined because of Properties (T1) and
(T3) of a tree decomposition. Let <T be the partial ordering of V (G) such that u <T v if
and only if Fu(T) <T Fv(T) for every u, v ∈ V (G). We say that (T, χ) is a dependency tree
decomposition if it satisfies the following additional property:
(T4) <T is compatible with ≤P, i.e., for every two vertices u and v of G it holds that
whenever Fu(T) <T Fv(T) then it does not hold that v ≤P u.
(cid:73) Lemma 14. A graph G has a P-elimination ordering of width at most ω if and only
if G has a dependency tree decomposition of width at most ω. Moreover, a P-elimination
ordering of width ω can be obtained from a dependency tree decomposition of width ω in
polynomial-time and vice versa.
Proof. For the forward direction we will employ the construction given by Kloks in [18],
which shows that a normal elimination ordering can be transformed into a tree decomposition
of the same width. We will then show that this construction also retains the compatibility
with P. Let ≤φ= (v1, . . . , vn) be a dependency elimination ordering for G of width ω and let
H be the fill-in graph of G w.r.t. ≤φ. We will iteratively construct a sequence (T0, . . . ,Tn−1)
such that for every i with 0 ≤ i < n, Ti = (Ti, χi) is dependency tree decompositions of
the graph Hi = H[{vn−i, . . . , vn}] of width at most ω. Because Tn−1 is a dependency tree
decomposition of Hn−1 = H of width at most ω, this shows the forward direction of the
lemma. In the beginning we set T0 to be the trivial tree decomposition of H0, which contains
merely one node whose bag consists of the vertex vn. Moreover, for every i with 0 < i < n,
Ti is obtained from Ti−1 as follows. Note that because NHi(vn−i) induces a clique in Hi−1,
Ti−1 contains a node that covers all vertices in NHi(vn−i). Let t be any such bag, then
is Ti is obtained from Ti−1 by adding a new node t0 to Ti−1 making it adjacent to t and
setting χi(t0) = NHi[vn−i]. It is known [18] that Ti satisfies the Properties (T1) -- (T3) of
a tree decomposition and it hence only remains to show that Ti satisfies (T4). Since, by
induction hypothesis, Ti−1 is a dependency tree decomposition, Property (T4) already holds
for every pair u, v ∈ V (Hi−1). Hence it only remains to consider pairs u and vn−i for some
u ∈ V (Hi−1). Because the only node containing vn−i in Ti is a leaf, we can assume that
Fu(T) <T Fvn−i(T) and because vn−i ≤φ u it cannot hold that vn−i ≤P u, as required.
For the reverse direction, let T = (T, χ) be a P-tree decomposition of G of width at most
ω. It is known [18] that any linear extension of <T is an elimination ordering for G of width
at most ω. Moreover, because of Property (T4), <T is compatible with ≤P and hence there
is a linear extension of <T , which is also a linear extension of the reverse of ≤P.
(cid:74)
Dependency cops and robber game: Recalling the definition of the (monotone) cops
and robber game for treewidth, we define the dependency cops and robber game (for a QBF
instance I with dependency poset P) analogously but with the additional restriction that
legal moves must also satisfy a third condition:
CM3 C0 \ C is downward-closed in R, i.e., there is no r ∈ R \ C0 with r ≤P c for any
c ∈ C0 \ C.
Intuitively, condition CM3 restricts the cop-player by forcing him to search vertices (vari-
ables) in an order that is compatible with the dependency poset.
To formally prove the equivalence between the cop-number for this restricted game and
dependency treewidth, we will need to also formalize the notion of a strategy. Here we
will represent strategies for the cop-player as rooted trees whose nodes are labeled with
Eduard Eiben, Robert Ganian, and Sebastian Ordyniak
XX:13
positions for the cop-player and whose edges are labeled with positions for the robber-
player. Namely, we will represent winning strategies for the cop-player on a primal graph
G by a triple (T, α, β), where T is a rooted tree, α : V (T) → 2V (G) is a mapping from the
nodes of T to subsets of V (G), and β : E(T) → 2V (G), satisfying the following conditions:
CS1 α(r) = ∅ and for every component R of G, the root node r of T has a unique child c
with β({r, c}) = R, and
CS2 for every other node t of T with parent p it holds that: the move from position
(α(p), β({p, t})) to position (α(t), β({t, c})) is legal for every child c of t and moreover
for every component R of G \ α(t) contained in β({p, t}), t has a unique child c with
β({t, c}) = R.
Informally, the above properties ensure that every play consistent with the strategy is win-
ning for the cop-player and moreover for every counter-move of the robber-player, the
strategy gives a move for the cop-player. The width of a winning strategy for the cop-
player is the maximum number of cops simultaneously placed on G by the cop-player, i.e.,
maxt∈V (T ) α(t). The cop-number of a graph G is the minimum width of any winning strat-
egy for the cop-player on G. We are now ready to show the equivalence between dependency
tree decompositions and winning strategies for the cop-player.
(cid:73) Lemma 15. For every graph G the width of an optimal dependency tree decomposition
plus one is equal to the cop-number of the graph. Moreover, a dependency tree decomposition
of width ω can be obtained from a winning strategy for the cop-player of width ω + 1 in
polynomial-time and vice versa.
Proof. Let T = (T, χ) be a dependency tree decomposition of G of width ω. We start
by showing that T can be transformed into a dependency tree decomposition of width ω
satisfying:
(*) χ(r) = ∅ for the root node r of T and for every node t ∈ V (T) with child c ∈ V (T) in
To ensure that T satisfies χ(r) = ∅ it is sufficient to add a new root vertex r0 to T and
set χ(r0) = ∅. We show next that starting from the root of T we can ensure that for every
node t ∈ V (T) with child c the set χ(Tc) \ χ(t) is a component of G \ χ(t). Let t be a
node with child c in T for which this does not hold. Because of Proposition 4, we have
that χ(Tc) \ χ(t) is a set of components, say containing C1, . . . , Cl, of G \ χ(t). For every
i with 1 ≤ i ≤ l, let Ti = (Ti, χi) be the dependency tree decomposition with Ti = Tc and
χi(t0) = χ(t0) ∩ (Ci ∪ χ(t)) and root ri = c. Then we replace the entire sub dependency
tree decomposition of T induced by Tc in T with the tree decompositions T1, . . . ,Tl such
that t now becomes adjacent to the roots r1, . . . , rl. It is straightforward to show that the
result of this operation is again a dependency tree decomposition of G of width at most ω
and moreover the node t has one child less that violates (*). By iteratively applying this
operation to every node t of T we eventually obtain a dependency tree decomposition that
satisfies (*).
Hence w.l.o.g. we can assume that T satisfies (*). We now claim that (T, α, β) where:
α(t) = χ(t) for every t ∈ V (T),
for a node t ∈ V (T) with parent p ∈ V (T), β({p, t}) = χ(Tt) \ χ(p).
T the set χ(Tc) \ χ(t) is a component of G \ χ(t).
is a winning strategy for ω + 1 cops. Observe that because T satisfies (*), it holds that
α(r) = ∅ and for every t ∈ V (T) with parent p ∈ V (T), the pair (α(p), β({p, t}) is a position
in the visible P-cops and robber game on G.
We show next that for every t, p as above and every child c of t in T, it holds that the move
from (α(p), β({p, t}) to (α(t), β({t, c}) is valid. Because β({t, c}) = χ(Tc) \ χ(t) ⊆ χ(Tt) \
XX:14
Small Resolution Proofs for QBF using Dependency Treewidth
χ(p) = β({p, t}) and β({p, t}) is connected in G\α(p) (and hence also in G\(α(p)∩α(t))), it
follows that β({p, t}) and β({t, c}) are contained in the same component of G\(α(p)∩ α(t)),
which shows CM1. Because of Proposition 4, it holds that α(p)∩α(t) and hence in particular
α(t) separates β({p, t}) from the rest of the graph, which shows CM2, i.e, δ(β({p, t})) ⊆ α(t).
Towards showing CM3 suppose for a contradiction that there is a r ∈ β({p, t}) \ α(t) with
r ≤P c for some c ∈ α(t) \ α(p). Because r ∈ β({p, t}) \ α(t) and c ∈ α(t), we obtain that
c <T r and because of (T4) it follows that r ≤P c does not hold. Note that CS1 and the
second part of CS2, i.e., for every node t with parent p and every component R of G \ α(t),
t has a unique child c with β({t, c}) = R, both hold because T satisfies T1 and T3 (T3 is
only needed to show that the child is unique).
On the other hand, let S = (T, α, β) be a winning strategy for the cop-player in the
visible P-cops and robber game on G using ω cops. Observe that S can be transformed into
a winning strategy for the cop-player using ω cops satisfying:
(a) for every node t of T with parent p it holds that α(t) ⊆ δ(β({p, t})) ∪ β({p, t}).
Indeed; if (a) is violated, then one can simply change α(t) to α(t)∩ (δ(β({p, t}))∪ β({p, t}))
without violating any of CS1 or CS2. Hence we can assume that S satisfies (a).
We now claim that T = (T, α) is a dependency tree decomposition of G of width ω − 1.
Towards showing T1, let v ∈ V (G). Because of CS1, it holds that either v ∈ α(r) for the
root r of T or there is a child c of r in T with v ∈ β({r, c}). Moreover, due to CS2 we have
that either v ∈ α(c) or v ∈ β({c, c0}) for some child c0 of c in T. By proceeding along T, we
will eventually find a node t ∈ V (T) with v ∈ α(t). Towards showing T2, let {u, v} ∈ E(G).
Again because of CS1, it holds that either {u, v} ⊆ α(r), or {u, v} ⊆ δ(β({r, c})) ∪ β({r, c})
for some child c of r in T. Because of CM2, we obtain that δ(β({r, c})) ⊆ α(c) and together
with CS2, we have that either {u, v} ⊆ α(c) or {u, v} ⊆ δ(β({c, c0})) ∪ β({c, c0}) for some
child c0 of c in T. By proceeding along T, we will eventually find a node t ∈ V (T) with
{u, v} ∈ α(t).
Next, we prove that T3 holds by showing that for any three nodes
t1, t2, and t3 in T such that t2 lies on the unique path from t1 to t3 in T, it holds that
α(t1) ∩ α(t2) ⊆ α(t3). We will distinguish two cases: (1) t1 is not an ancestor of t3 and vice
versa, and (2) t1 and t2 lie on the unique path from the root of T to t3.
(b) for every node t with child c in T it holds thatS
Before proceeding, we need to establish that S satisfies the following property:
t0∈V (Tc) α(t0) ⊆ δ(β({t, c}) ∪ β({t, c}).
Because of CM2 we have that β({t, c}) ⊆ β({p, t}) for every three nodes p, t, and c such
that p is the parent of t which in turn is the parent of c in T. Moreover, because of (a) we
have that α(t) ⊆ δ(β({p, t}))∪ β({p, t}) for every node t with parent p in T. Applying these
two facts iteratively along a path from t to any of its descendants t0 in T, we obtain that
α(t0) ⊆ δ(β({p, t})) ∪ β({p, t}), as required.
Towards showing case (1) assume that this is not the case, i.e., there are t1, t2, and t3
as above and a vertex v ∈ α(t1) ∩ α(t3) but v /∈ α(t2). W.l.o.g. we can assume that t2 is
the least common ancestor of t1 and t2 in T, otherwise we end up in case (2). Let c1 and
c2 be the two children of t2 such that t1 ∈ V (Tc1) and t2 ∈ V (Tc2). Then because v ∈ α(ti)
we obtain from (b) that v ∈ δ(β({t2, ci})) ∪ β({t2, ci})) for any i ∈ {1, 2}. Because the sets
β({t2, c1}) and β({t2, c2}) are disjoint, it follows that v ∈ δ(β({t2, ci})) and thus v ∈ α(t2)
contradicting our assumption that this is not the case.
Towards showing case (2) assume that this is not the case, i.e., there are t1, t2, and t3 as
above and a vertex v ∈ α(t1)∩ α(t3) but v /∈ α(t2). W.l.o.g. we can assume that t2 is a child
of t1 in T. Because v ∈ α(t3) we obtain from (b) that α(t3) ⊆ δ(β({t1, t2}) ∪ β({t1, t2}).
Hence either v ∈ δ(β({t1, t2})) or v ∈ β({t1, t2}). In the former case it follows from CM2
that v ∈ α(t2) a contradiction to our assumption that v /∈ α(t2) and in the later case, we
obtain that v /∈ α(t1) a contradiction to our assumption that v ∈ α(t1).
Eduard Eiben, Robert Ganian, and Sebastian Ordyniak
XX:15
S
i.e., there are u, v ∈ V (G)
Towards showing T4 assume that this is not the case,
with u <T v but v ≤P u. Let p be the parent of Fu(T) in T. Then because u, v ∈
t0∈V (TFu(T ))) α(t0) and (b), we obtain that u, v ∈ δ(β({p, Fu(T)})) ∪ β({p, Fu(T)}). More-
over since neither u or v are in α(p), we have that u, v ∈ β({p, Fu(T)}). Finally because
u ∈ α(Fu(T)) but not v ∈ α(Fu(T)), we obtain from CM3 that u ≤P v contradicting our
assumption that v ≤P u.
(cid:74)
5
Computing Dependency Treewidth
In this section we will present two exact algorithms to compute dependency treewidth. The
first algorithm is based on the characterization of dependency treewidth in terms of the
cops and robber game and shows that, for every fixed ω, determining whether a graph has
dependency treewidth at most ω, and in the positive case also computing a dependency
tree decomposition of width at most ω, can be achieved in polynomial time. The second
algorithm is based on a chain partition of the given dependency poset and shows that if
the width of the poset is constant, then an optimal dependency tree decomposition can be
constructed in polynomial time.
Before proceeding to the algorithms, we would like to mention here that the fixed-
parameter algorithm for computing first order treewidth [1] can also be used for computing
dependency treewidth in the restricted case that the trivial dependency poset is used.
Below we provide our first algorithm for the computation of dependency treewidth.
(cid:73) Theorem 16. There is an algorithm running in time O(V (G)2ω+2) that, given a graph G
and a poset P = (V (G),≤P) and ω ∈ N, determines whether ω cops have a winning strategy
in the dependency cops and robber game on G and P, and if so outputs such a winning
strategy.
Proof. The idea is to transform the cops and robber game on G into a much simpler two
player game, the so-called simple two player game, which is played on all possible positions
of the cops and robber game on G.
A simple two player game is played between two players, which in association to the
cops and robber game, we will just call the cops and the robber player [15]. Both players
play by moving a token around on a so-called arena, which is a triple A = (VC, VR, A) such
that ((VC ∪ VR), A) is a bipartite directed graph with bipartition (VC, VR). The vertices in
VC are said to belong to the cop-player and the vertices in VR are said to belong to the
robber-player. Initially, one token is placed on a distinguished starting vertex s ∈ VC ∪ VR.
From then onward the player who owns the vertex, say v, that currently contains the token,
has to move the token to an arbitrary successor (i.e., out-neighbor) of v in A. The cop-
player wins if the robber-player gets stuck, i.e., the token ends up in a vertex owned by the
robber-player that has no successors in A, otherwise the robber-player wins. It is well-known
that strategies in this game are deterministic and memoryless, i.e., strategies for a player are
simple functions that assign every node owned by the player one of its successors. Moreover,
the winning region for both players as well as their corresponding winning strategy can
be computed in time O(VC ∪ VR + A) by the following algorithm. The algorithm first
computes the winning region WC, as follows.
Initially all vertices owned by the robber-player which do not have any successors in A
are placed in WC. The algorithm then iteratively adds the following vertices to WC:
all vertices owned by the cop-player that have at least one successor WC,
all vertices owned by the robber-player for which all successors are in WC.
XX:16
Small Resolution Proofs for QBF using Dependency Treewidth
Once the above process stops, the set WC is the winning region of the cop-player in A and
(VC ∪ VR)\ WC is the winning region for the robber-player. Moreover, the winning strategy
for both players can now be obtained by choosing for every vertex a successors that is in the
winning region of the player owning that vertex (if no such vertex exists, then an arbitrary
successor must be chosen).
Given a graph G, a poset P = (V (G),≤P), and an integer ω, we construct an arena
A = (VC, VR, A) and a starting vertex s ∈ VR such that ω cops have a winning strategy in
the P-cops and robber game on G iff the cop-player wins from s in the simple two player
game on A as follows:
We set VC to be the set of all pairs (C, R) such that (C, R) is a position in the P-cops
and robber game on G using at most ω cops (C ≤ ω),
We set VR to be the set of all triples (C, C0, R) such that:
(C, R) is a position in the P-cops and robber game on G using at most ω cops (C ≤
ω), and
C0 ⊆ V (G) is a potential new cop-position for at most ω cops from (C0, R0), i.e.,
δ(R) ⊆ C0 and C, R, and C0 satisfy CM3.
From every vertex (C, R) ∈ VC we add an arc to all vertices (C, C0, R) ∈ VR.
From every vertex (C, C0, R) ∈ VR we add an arc to all vertices (C0, R0) ∈ VC such that
the move from (C, R) to (C0, R0) is legal.
Additionally VR contains the starting vertex s that has an outgoing arc to every vertex
(∅, R) ∈ VC such that R is a component of G.
By construction VC ≤ V (G)ω+1 and VR ≤ V (G)2ω+1. Moreover, because every
vertex in VC has at most V (G)ω successors and every vertex in VR has at most V (G)
successors, we obtain that A ≤ V (G)2ω+2. Let us now analyze the running time required
to construct A. We can construct all vertices (C, R) ∈ VC in time O(V (G)ωE(G)) by
computing the set of all components of G \ C for every cop-position C. Note that within
the same time we can additionally compute and store the guards δ(R) for every component
R. For each vertex in (C, R) ∈ VC we can then compute the associated vertices (C, C0, R) ∈
VR and add the necessary arcs by enumerating all sets C0 ⊆ V (G) with δ(R) ⊆ C0 and
checking for each of those whether R and C0 satisfy CM3. Enumerating the sets C0 ⊆ V (G)
with δ(R) ⊆ C0 can be achieved in time O(V (G)ω) (using the fact that we stored δ(R)
for every component R). Moreover, determining whether R and C0 satisfy CM3 can be
achieved in time O(ωR) = O(ωV (G)) by going over all vertices r ∈ R and verifying
that r ∈ C0 or that it is not smaller than any vertex in C0. Hence computing all vertices
of A and all arcs from vertices in VC to vertices in VR can be achieved in time at most
O(V (G)ωE(G) + V (G)2ω+1ω), which for the natural assumption that ω > 0 is at most
O(V (G)2ω+2). Finally, we need to add the arcs between vertices in (C, C0, R) ∈ VR and
the vertices (C0, R0) ∈ VC. Note that there is an arc from (C, C0, R) ∈ VR to (C0, R0) ∈ VC if
and only if R0 ⊆ R and moreover for every component R0 of G\ C0 either it is a subset of R
or it is disjoint with R, which can be checked in constant time. Hence the total time required
for this last step is equal to the number of vertices in VR times V (G) which is at most
O(V (G)2ω+2). It follows that the time required to construct A is at most O(V (G)2ω+2).
Once the arena is constructed the winning regions as well as the winning strategies for both
players can be computed in time O(VC ∪ VR + A) ∈ O(V (G)2ω+2).
(cid:74)
The next theorem summarizes our second algorithm for computing dependency treewidth.
The core distinction here lies in the fact that the running time does not depend on the de-
pendency treewidth, but rather on the poset-width. This means that the algorithm can
Eduard Eiben, Robert Ganian, and Sebastian Ordyniak
XX:17
precisely compute the dependency treewidth even when this is large, and it will perform
better than Theorem 16 for formulas with "tighter" dependency structures (e.g., formulas
which utilize the full power of quantifier alternations).
(cid:73) Theorem 17. There is an algorithm running in time O((V (G)kk2) that, given a graph
G and a poset P = (V (G),≤P) of width k and ω ∈ N, determines whether G and P admit a
dependency elimination ordering of width at most ω, and if yes outputs such a dependency
elimination ordering.
Proof. To decide whether G has a dependency elimination ordering of width at most ω, we
first build an auxiliary directed graph H as follows.
The vertex set of H consists of all pairs (D, d) such D ⊆ V (G) is a downward closed set
and d ∈ D is a maximal element of D such that NG[D\d](d) ≤ ω. Additionally, H contains
the vertices (V (G),∅) and (∅,∅). Furthermore, there is an arc from (D, d) to (D0, d0) of H if
and only if D0 = D ∪ {d0} or D = D0 = V (G) and d0 = ∅. This completes the construction
of H. It is immediate that G has a dependency elimination ordering of width at most ω if
and only if there is a directed path in H from (∅,∅) to (V (G),∅). Hence, given H we can
decide whether G has a dependency elimination ordering of width at most ω (and output
it, if it exists) in time O(V (H) log(V (H)) + E(H)) (e.g., by using Dijkstra's algorithm).
It remains to analyze the time required to construct H (as well as its size).
Let k be the width of the poset P. Due to Proposition 1, we can compute a chain
partition C = (W1, . . . , Wk) of width k in time O(k · V (G)2). Note that every downward
closed D set can be characterized by the position of the maximal element in D on each
of the chains W1, . . . , Wk, we obtain that there are at most V (G)k downward closed sets.
Hence, H has at most O(V (G)k(k + 1)) vertices its vertex set can be constructed in time
O(V (G)k(k + 1)). Since every vertex (D, d) of H has at most k + 1 possible out-neighbors,
we can construct the arc set of H in time O(V (G)kk2).
Hence, the total time required to construct H is O((V (G)kk2) which dominates the
(cid:74)
time required to find a shortest path in H, and so the runtime follows.
Concluding Notes
6
Dependency treewidth is a promising decompositional parameter for QBF which overcomes
the key shortcomings of previously introduced structural parameters; its advantages include
a single-exponential running time, a refined and flexible approach to variable dependencies,
and the ability to compute decompositions. It also admits several natural characterizations
that show the robustness of the parameter and allows the computation of resolution proofs.
The presented algorithms for computing dependency elimination orderings leave open
the question of whether this problem admits a fixed-parameter algorithm (parameterized by
dependency treewidth). We note that the two standard approaches for computing treewidth
fail here. In particular, the well-quasi-ordering approach with respect to minors does not
work since the set of ordered graphs can be observed not to be well-quasi ordered w.r.t. the
ordered minor relation [29]. On the other hand, the records used in the second approach [4]
do not provide sufficient information in our ordered setting.
6.0.0.1 Acknowledgments
The authors wish to thank the anonymous reviewers for their helpful comments. Eduard
Eiben acknowledges support by the Austrian Science Fund (FWF, projects P26696 and
W1255-N23). Robert Ganian is also affiliated with FI MU, Brno, Czech Republic.
3 Armin Biere and Florian Lonsing. Integrating dependency schemes in search-based QBF
solvers. In Theory and Applications of Satisfiability Testing - SAT 2010, volume 6175 of
LNCS, pages 158 -- 171. Springer, 2010.
4 Hans L. Bodlaender. A linear time algorithm for finding tree-decompositions of small
In Proceedings of the Twenty-Fifth Annual ACM Symposium on Theory of
treewidth.
Computing, May 16-18, 1993, San Diego, CA, USA, pages 226 -- 234, 1993.
Florent Capelli. Structural restrictions of CNF-formulas: applications to model counting
and knowledge compilation. PhD thesis, Université Paris Diderot, 2016.
6 Hubie Chen and Víctor Dalmau. Decomposing quantified conjunctive (or disjunctive) for-
5
mulas. SIAM J. Comput., 45(6):2066 -- 2086, 2016.
7 M. Davis and H. Putnam. A computing procedure for quantification theory. J. of the ACM,
8 Reinhard Diestel. Graph minors 1: A short proof of the path-width theorem. Combina-
torics, Probability & Computing, 4:27 -- 30, 1995.
9 Reinhard Diestel. Graph Theory, 4th Edition, volume 173 of Graduate texts in mathematics.
7(3):201 -- 215, 1960.
Springer, 2012.
XX:18
Small Resolution Proofs for QBF using Dependency Treewidth
1
References
Isolde Adler and Mark Weyer. Tree-width for first order formulae. Logical Methods in
Computer Science, 8(1), 2012.
2 Albert Atserias and Sergi Oliva. Bounded-width QBF is pspace-complete. J. Comput. Syst.
Sci., 80(7):1415 -- 1429, 2014.
10 Rodney G. Downey and Michael R. Fellows. Fundamentals of Parameterized Complexity.
Texts in Computer Science. Springer Verlag, 2013.
12
13
11 Uwe Egly, Thomas Eiter, Hans Tompits, and Stefan Woltran. Solving advanced reasoning
tasks using quantified boolean formulas. In Proceedings of the Seventeenth National Con-
ference on Artificial Intelligence and Twelfth Conference on on Innovative Applications of
Artificial Intelligence, July 30 - August 3, 2000, Austin, Texas, USA., pages 417 -- 422, 2000.
Eduard Eiben, Robert Ganian, and Sebastian Ordyniak. Using decomposition-parameters
for QBF: mind the prefix! In Proceedings of the Thirtieth AAAI Conference on Artificial
Intelligence, February 12-17, 2016, Phoenix, Arizona, USA., pages 964 -- 970, 2016.
S. Felsner, V. Raghavan, and J. Spinrad. Recognition algorithms for orders of small width
and graphs of small dilworth number. Order, 20(4):351 -- 364, 2003.
Jörg Flum and Martin Grohe. Parameterized Complexity Theory, volume XIV of Texts in
Theoretical Computer Science. An EATCS Series. Springer Verlag, Berlin, 2006.
Erich Grädel, Wolfgang Thomas, and Thomas Wilke, editors. Automata, Logics, and In-
finite Games: A Guide to Current Research [outcome of a Dagstuhl seminar, February
2001], volume 2500 of Lecture Notes in Computer Science. Springer, 2002.
16 Hans Kleine Büning and Uwe Bubeck. Theory of quantified boolean formulas. In A. Biere,
M. J. H. Heule, H. van Maaren, and T. Walsh, editors, Handbook of Satisfiability, volume
185 of Frontiers in Artificial Intelligence and Applications, chapter 23, pages 735 -- 760. IOS
Press, 2009.
17 Hans Kleine Büning and Theodor Lettman. Propositional logic: deduction and algorithms.
15
14
Cambridge University Press, Cambridge, 1999.
18 T. Kloks. Treewidth: Computations and Approximations. Springer Verlag, Berlin, 1994.
19 Charles Otwell, Anja Remshagen, and Klaus Truemper. An effective QBF solver for plan-
ning problems.
In Proceedings of the International Conference on Modeling, Simulation
& Visualization Methods, MSV '04 & Proceedings of the International Conference on Al-
gorithmic Mathematics & Computer Science, AMCS '04, June 21-24, 2004, Las Vegas,
Nevada, USA, pages 311 -- 316. CSREA Press, 2004.
20 Christos H. Papadimitriou. Computational Complexity. Addison-Wesley, 1994.
Eduard Eiben, Robert Ganian, and Sebastian Ordyniak
XX:19
21
22
23
Luca Pulina and Armando Tacchella. A structural approach to reasoning with quantified
boolean formulas. In Craig Boutilier, editor, IJCAI 2009, Proceedings of the 21st Interna-
tional Joint Conference on Artificial Intelligence, Pasadena, California, USA, July 11-17,
2009, pages 596 -- 602, 2009.
Luca Pulina and Armando Tacchella. An empirical study of QBF encodings: from treewidth
estimation to useful preprocessing. Fundam. Inform., 102(3-4):391 -- 427, 2010.
J. Rintanen. Constructing conditional plans by a theorem-prover. J. Artif. Intell. Res.,
10:323 -- 352, 1999.
24 Ashish Sabharwal, Carlos Ansótegui, Carla P. Gomes, Justin W. Hart, and Bart Selman.
QBF modeling: Exploiting player symmetry for simplicity and efficiency. In Theory and
Applications of Satisfiability Testing - SAT 2006, 9th International Conference, Seattle,
WA, USA, August 12-15, 2006, Proceedings, pages 382 -- 395, 2006.
25 Marko Samer and Stefan Szeider. Backdoor sets of quantified Boolean formulas. Journal
of Autom. Reasoning, 42(1):77 -- 97, 2009.
Friedrich Slivovsky and Stefan Szeider. Computing resolution-path dependencies in linear
time. In Alessandro Cimatti and Roberto Sebastiani, editors, Theory and Applications of
Satisfiability Testing - SAT 2012, volume 7317 of Lecture Notes in Computer Science, pages
58 -- 71. Springer Verlag, 2012.
Friedrich Slivovsky and Stefan Szeider. Variable dependencies and q-resolution. In Theory
and Applications of Satisfiability Testing - SAT 2014 - 17th International Conference, Held
as Part of the Vienna Summer of Logic, VSL 2014, Vienna, Austria, July 14-17, 2014.
Proceedings, volume 8561 of Lecture Notes in Computer Science, pages 269 -- 284. Springer,
2014.
Friedrich Slivovsky and Stefan Szeider. Quantifier reordering for QBF. J. Autom. Reason-
ing, 56(4):459 -- 477, 2016.
26
27
28
30
31
29 Daniel A Spielman and Miklós Bóna. An infinite antichain of permutations. Electron. J.
Combin, 7:N2, 2000.
L. J. Stockmeyer and A. R. Meyer. Word problems requiring exponential time. In Pro-
ceedings of the 5th Annual ACM Symposium on Theory of Computing, April 30 - May 2,
1973, Austin, Texas, USA, pages 1 -- 9. ACM, 1973.
Stefan Szeider. On fixed-parameter tractable parameterizations of SAT.
In Enrico
Giunchiglia and Armando Tacchella, editors, Theory and Applications of Satisfiability, 6th
International Conference, SAT 2003, Selected and Revised Papers, volume 2919 of Lecture
Notes in Computer Science, pages 188 -- 202. Springer Verlag, 2004.
32 Allen Van Gelder. Variable independence and resolution paths for quantified boolean for-
mulas.
In Jimmy Lee, editor, Principles and Practice of Constraint Programming - CP
2011, volume 6876 of Lecture Notes in Computer Science, pages 789 -- 803. Springer Verlag,
2011.
|
1811.00710 | 1 | 1811 | 2018-11-02T02:24:40 | On subexponential running times for approximating directed Steiner tree and related problems | [
"cs.DS"
] | This paper concerns proving almost tight (super-polynomial) running times, for achieving desired approximation ratios for various problems. To illustrate, the question we study, let us consider the Set-Cover problem with n elements and m sets. Now we specify our goal to approximate Set-Cover to a factor of (1-d)ln n, for a given parameter 0<d<1. What is the best possible running time for achieving such approximation? This question was answered implicitly in the work of Moshkovitz [Theory of Computing, 2015]: Assuming both the Projection Games Conjecture (PGC) and the Exponential-Time Hypothesis (ETH), any ((1-d) ln n)-approximation algorithm for Set-Cover must run in time >= 2^{n^{c d}}, for some constant 0<d<1.
We study the questions along this line. First, we show that under ETH and PGC any ((1-d) \ln n)-approximation for Set-Cover requires 2^{n^{d}}-time. This (almost) matches the running time of 2^{O(n^d)} for approximating Set-Cover to a factor (1-d) ln n by Cygan et al. [IPL, 2009]. Our result is tight up to the constant multiplying the n^{d} terms in the exponent. This lower bound applies to all of its generalizations, e.g., Group Steiner Tree (GST), Directed Steiner (DST), Covering Steiner Tree (CST), Connected Polymatroid (CP). We also show that in almost exponential time, these problems reduce to Set-Cover: We show (1-d)ln n approximation algorithms for all these problems that run in time 2^{n^{d \log n } poly(m).
We also study log^{2-d}n approximation for GST. Chekuri-Pal [FOCS, 2005] showed that GST admits (log^{2-d}n)-approximation in time exp(2^{log^{d+o(1)}n}), for any 0 < d < 1. We show the lower bound of GST: any (log^{2-d}n)-approximation for GST must run in time >= exp((1+o(1)){log^{d-c}n}), for any c>0, unless the ETH is false. Our result follows by analyzing the work of Halperin and Krauthgamer [STOC, 2003]. The same lower and upper bounds hold for CST. | cs.DS | cs | On subexponential running times for approximating directed
Steiner tree and related problems
Marek Cygan∗
Guy Kortsarz†
Bundit Laekhanukit‡
November 5, 2018
Abstract
This paper concerns proving almost tight (super-polynomial) running times, for achieving desired
approximation ratios for various problems. To illustrate the question we study, let us consider the
Set-Cover problem with n elements and m sets. Now we specify our goal to approximate Set-Cover
to a factor of (1 − α) ln n, for a given parameter 0 < α < 1. What is the best possible running
time for achieving such approximation ratio? This question was answered implicitly in the work of
Moshkovitz [Theory of Computing, 2015]: Assuming both the Projection Games Conjecture (PGC) and
the Exponential-Time Hypothesis (ETH), any ((1 − α) ln n)-approximation algorithm for Set-Cover
must run in time at least 2nc·α
, for some small constant 0 < c < 1.
We study the questions along this line. Our first contribution is in strengthening the above result. We
show that under ETH and PGC the running time requires for any ((1 − α) ln n)-approximation algorithm
for Set-Cover is essentially 2nα
. This (almost) settles the question since our lower bound matches the
best known running time of 2O(nα) for approximating Set-Cover to within a factor (1 − α) ln n given
α terms in the
by Cygan et al.
exponent.
[IPL, 2009]. Our result is tight up to the constant multiplying the n
The lower bound of Set-Cover applies to all of its generalization, e.g., Group-Steiner-Tree,
Directed-Steiner-Tree, Covering-Steiner-Tree and Connected-Polymatroid. We show that,
surprisingly, in almost exponential running time, these problems reduce to Set-Cover. Specifically, we
complement our lower bound by presenting an (1−α) ln n approximation algorithm for all aforementioned
problems that runs in time 2nα ·log n
· poly(m).
We further study the approximation ratio in the regime of log2−δ n for Group-Steiner-Tree and
Covering-Steiner-Tree. Chekuri and Pal [FOCS, 2005] showed that Group-Steiner-Tree admits
n)-approximation in time exp(2logα+o(1) n), for any parameter 0 < α < 1. We show the run-
(log2−α
ning time lower bound of Group-Steiner-Tree: any (log2−α
n)-approximation algorithm for Group-
Steiner-Tree must run in time at least exp((1 + o(1))logα−ǫ
n), for any constant ǫ > 0, unless the
ETH is false. Our result follows by analyzing the hardness construction of Group-Steiner-Tree due
to the work of Halperin and Krauthgamer [STOC, 2003]. The same lower and upper bounds hold for
Covering-Steiner-Tree.
8
1
0
2
v
o
N
2
]
S
D
.
s
c
[
1
v
0
1
7
0
0
.
1
1
8
1
:
v
i
X
r
a
∗Department of Math and information, University of Warsaw, Warsaw, Porland. Email: [email protected]
†Computer Science Department, Rutgers University -- Camden, Camden NJ, USA. Email: [email protected]
‡Institute for Theoretical Computer Science, Shanghai University of Finance and Economics, Shanghai, China. Email:
[email protected]
1
1
Introduction
The traditional study of approximation algorithms concerns designing algorithms that run in polynomial
time while producing a solution whose cost is within a factor α away from the optimal solution. Once the
approximation guarantees meet the barrier, a natural question is to ask whether the approximation ratio
can be improved if the algorithms are given running time beyond polynomial. This has been a recent trend
in designing approximation algorithms that allows ones to break through the hardness barrier; see, e.g.,
[1, 5, 22, 6, 6, 7, 7, 15, 14].
While ones ask for improving the approximation ratio, another interesting question is to ask the converse:
Suppose the approximation ratio has been specified at the start, what is the smallest running time required
to achieve such approximation ratio? This question has recently been an active subject of study; see, e.g.,
[9, 3, 4].
To answer the above question, ones need complexity assumptions stronger than P 6= NP as this standard
assumption does not precisely specify the running times besides polynomial versus super-polynomial. The
most popular and widely believed assumption is the Exponential-Time Hypothesis (ETH), which states that
3-SAT admits no 2o(n)-time algorithm. This together with the almost linear size PCP theorems [16, 33]
yields many running time lower bounds for approximation algorithms [9, 3, 4]. Let us give an example of
the results of this type:
Example: Consider the Maximum Clique problem, in which the goal is to find a clique of maximum size
in a graph G = (V, E) on n vertices. This problem is known to admit no n1−ǫ-approximation, for any ǫ > 0,
unless P = NP [26, 35]. Now, let us ask for an α-approximation algorithm, for α ranging from constant to
√n. There is a trivial 2n/αpoly(n)-time approximation algorithm, which is obtained by partitioning vertices
of G into to α parts and finding a maximum clique from each part separately. Clearly, the maximum clique
amongst these solutions is an α-approximate solution, and the running time is 2n/αpoly(n). The question is
whether this is the best possible running-time. Chalermsook et al. [9] showed that such a trivial algorithm
is almost tight1. To be precise, under the ETH, there is no α-approximation algorithm that runs in time
2n1−ǫ/α1+ǫ
, for any constant ǫ > 0, unless the ETH is false.
In this paper, we consider the question along this line. We wish to show the tight lower and upper bounds
on the running times of polylogarithmic approximation algorithms for Set-Cover, Group-Steiner-Tree
and Directed-Steiner-Tree (which we will define in the next section) and related problems.
• For any constant 0 < α < 1, what is the best possible running times for (1 − α) ln n-approximation
algorithms for Set-Cover and Directed-Steiner-Tree.
• For any constant 0 < α < 1, what is the best possible running time for log2−α-approximation algorithms
for Group-Steiner-Tree.
In fact, one of our ultimate goals is to find an evidence on which ranges of running-times that the
Directed-Steiner-Tree problem admits poly-logarithmic approximations. To be precise, we would like to
partially answer the question of whether Directed-Steiner-Tree admits polylogarithmic approximations
in polynomial-time, which is a big open problem in the area. While we are far from solving the above
question, we aim to prove possibly tight running-time for Directed-Steiner-Tree in the logarithmic range
in a very fine-grained manner, albeit assuming two strong assumptions, the Exponential-Time Hypothesis
(ETH) [27, 28] and the Projection Game Conjectures (PGC) [32], simultaneously.
1.1 The problems studied in this paper
1.1.1 The Set-Cover problem and its extensions
In the weighted Set-Cover problem, the input is a universe U of size n and a collection S of m subsets of
U . Each set s ∈ S has a cost c(s). The goal is to select a minimum cost subcollection S′ ⊆ S such that the
union of the sets in S′ spans the entire universe U .
1 Recently, Bansal et al. [1] showed that Maximum Clique admits α-approximation in time 2n/ O(α log2 α)poly(n).
2
The more general Submodular-Cover problem admits as input a universe U with cost c(x) on every
x ∈ U . A function is submodular if for every S ⊆ T ⊆ V and for every x ∈ U \ T , f (S + x) − f (S) ≥
f (T + x) − f (T ). Let f : 2U 7→ R be a submodular non-decreasing function. The goal in the submodular
cover problem is to minimize c(S) subject to f (S) = f (U ). This problem strictly generalizes the weighted
Set-Cover problem.
The Connected-Polymatroid problem is the case that the elements in U are leaves of a tree, and
both the elements and tree edges have costs. The goal is to select a set S so that f (S) = f (U ) and that
c(S) + c(T (S)) is minimized, where T (S) is the unique tree rooted at r spanning S.
1.1.2 The Group-Steiner-Tree problem
In the Group-Steiner-Tree problem, the input consists an undirected graph with cost c(e) on each edge
e ∈ E, a collection of subsets g1, g2, . . . , gk ⊆ V (called group) and a special vertex r ∈ V . The goal is
is to find a minimum cost tree rooted at r that contains at least one vertex from every group gi. In the
Covering-Steiner-Tree problem, there is a demand di for every gi and di vertices of gi must be spanned
in the tree rooted by r This Group-Steiner-Tree problem strictly contains the Set-Cover problem.
Every result for Group-Steiner-Tree holds also for the Covering-Steiner-Tree problem given that
there is a reduction from Covering-Steiner-Tree to Group-Steiner-Tree [19, 24].
1.1.3 The Directed-Steiner-Tree problem
In the Directed-Steiner-Tree problem, the input consists of a directed graph with costs c(e) on edges,
a collection S of terminals, and a designated root r ∈ V . The goal is to find a minimum cost directed graph
rooted at r that spans S. This problem has Group-Steiner-Tree as a special case.
1.2 Related work
The Set-Cover problem is a well-studied problem. The first logarithmic approximation, to the best of
our knowledge, is traced back to the early work of Johnson [29]. Many different approaches have been
proposed to approximate Set-Cover, e.g., the dual-fitting algorithm by Chv´atal [13]; however, all algorithms
yield roughly the same approximation ratio. The more general problem, namely, the Submodular-Cover
problem was also shown to admit O(log n)-approximation in the work of Wolsey [34]. The question of why
all these algorithms yield the same approximation ratio was answered by Lund and Yannakakis [31] who
showed that the approximation ratio Θ(log n) is essentially the best possible unless NP ⊆ DTIME(nlog log n).
Subsequently, Feige [21] showed the more precise lower bound that Set-Cover admits no (1 − ǫ) ln n-
approximation, for any ǫ > 0, unless NP ⊆ DTIME(npolylog(n)); this assumption has been weaken to P 6= NP
by the recent work of Dinur and Steurer [17]. These lower bounds are, however, restricted to polynomial-time
algorithms. In the regime of subexponential-time, Cygan, Kowalik and Wykurz [15] showed that Set-Cover
admits an approximation ratio of (1− α) ln n in 2O(nα+polylog(n)) time. On the negative side, Moshkovitz [32]
introduced the Projection Games Conjecture (PGC) to prove the approximation hardness of Set-Cover.
Originally, the conjecture was introduced in an attempt to show the (1 − ǫ) ln n-hardness of Set-Cover
under P 6= NP (which is now proved by Dinur and Steurer [17]). It turns out that this implicitly implies
that Set-Cover admits no (1 − α) ln-approximation algorithm in 2nO(α)
The generalization of the Set-Cover problem is the Group-Steiner-Tree problem. Garg, Konjevod
and Ravi [23] presented a novel LP rounding algorithm to approximate Group-Steiner-Tree on trees to
within a factor of O(log2 n). Using the probabilistic metric-tree embedding [2, 20], this implies an O(log3 n)-
approximation algorithm for Group-Steiner-Tree in general graphs. On the negative side, Halperin and
Krauthgamer showed the lower bound of log2−ǫ n for any ǫ > 0 for approximating Group-Steiner-Tree
on trees under the assumption that NP 6⊆ ZPTIME(npolylog(n)). This (almost) matches the upper bound
given by the algorithm by Garg et al. For the related problem, the Connected Polymatroid problem was
given a polylogarithmic approximation algorithm by C´alinescu and Zelikovsky [8]; their algorithm is based
time under PGC and ETH.
3
on the work of Chekuri, Even and Kortsarz [11], which gave a combinatorial polylog(n) approximation for
Group-Steiner-Tree on trees.
The problem that generalizes all the above problems is the Directed-Steiner-Tree problem. The
best known approximation ratio for this problem is nǫ for any constant ǫ > 0 [10, 30] in polynomial-time.
In quasi-polynomial-time, Directed-Steiner-Tree admits an O(log3 n)-approximation algorithm. The
question of whether Directed-Steiner-Tree admits a polylogarithmic approximation in polynomial-time
has been a long standing open problem.
2 Our results
We show that under the combination of ETH and PGC, the running time for approximating Set-Cover to
within a factor of (1− α) ln n must be at least 2nα
, where 0 < α < 1 is a given parameter. This improves the
work of Moshkovitz who (implicitly) showed the running time lower bound of 2nO(α)
. We complement this
by showing that Directed-Steiner-Tree admits a (1− α) ln n approximation algorithm that runs in time
2nα·log n time. Since Directed-Steiner-Tree is the generalization of Set-Cover, Group-Steiner-Tree
and Directed-Steiner-Tree, the lower bounds apply to all the aforementioned problems. Hence, up to
a small factor of log n in the exponent, we get tight running time lower bounds for approximating all these
problems to within (1 − α) ln n. Essentially, the same algorithm and proof give the same result for the
Connected-Polymatroid problem.
We also investigate the work of Chekuri and Pal [12] who showed that, for any constant 0 < δ < 1,
Group-Steiner-Tree admits a log2−δ n approximation algorithm that runs in time exp(2(1+o(1)) logδ n). We
show that, for any constant ǫ > 1, there is no log2−δ−ǫ n approximation algorithm for Group-Steiner-Tree
(and thus Covering-Steiner-Tree) that runs in time exp(2(1+o(1)) logδ−ǫ n). This lower bound is nearly
tight. We note that a reduction from Covering-Steiner-Tree to Group-Steiner-Tree was given in [19].
Thus, any approximation algorithm for Group-Steiner-Tree also applies for Covering-Steiner-Tree.
3 Formal definition of our two complexity assumptions
Definition 3.1. In the Label-Cover problem with the projection property (a.k.a., the Projection game),
we are given a bipartite graph G(A, B, E), two alphabet sets (also called labels) ΣA and ΣB, and for any
edge (also called query) e ∈ E, there is a function φe : ΣA 7→ ΣB. A labeling (σA, σB) is a pair of functions
σA : A 7→ ΣA and σB : B 7→ ΣB assigning labels to each vertices of A and B, respectively. An edge e = (a, b)
is covered by (σA, σB) if φe(σA(a)) = σB(b). The goal in Label-Cover is to find a labeling (σA, σB) that
covers as many edges as possible.
In the context of the Two-Provers One-Round game (2P1R), every label is an answer to some "question"
a sent to the Player A and some question b sent to the Player B, for a query (a, b) ∈ E. The two answers
make the verifier accept if a label x ∈ Σx assigned to a and a label y ∈ ΣB assigned to b satisfy φ(x) = y.
Since any label x ∈ ΣA has a unique label in ΣB that causes the verifier to accept, y is called the projection
of x into b.
We use two conjectures in our paper. The first is the Exponential Time Hypothesis (ETH), which asserts
that an instance of the 3-SAT problem on n variables and m clauses cannot be solved in 2o(n)-time. This was
later showed by Impagliazzo, Paturi and Zane [27] that any 3-SAT instance can be sparsified in 2o(n)-time
to an instance with m = O(n) clauses. Thus, ETH together with the sparsification lemma [28] implies the
following:
Exponential-Time Hypothesis combined with the Sparsification Lemma: Given a boolean 3-CNF
formula φ on n variables and m clauses, there is no 2o(n+m)-time algorithm that decides whether φ is
satisfiable. In particular, 3-SAT admits no subexponential-time algorithm.
The following was proven by Moshkovitz and Raz [33].
4
Theorem 3.2 ([33]). There exists c > 0, such that for every ǫ ≥ 1/nc, 3-SAT on inputs of size n can be
efficiently reduced to Label-Cover of size N = n1+o(1)poly(1/ǫ) over an alphabet of size exp(1/ǫ) that has
soundness error ǫ. The graph is bi-regular (namely, every two questions on the same side participate in the
same number of queries).
There does not seem to be an inherent reason that the alphabet would be so large. This leads to the
following conjecture posed by Moshkovitz [32].
Conjecture 3.3 (The Projection Games Conjecture [32]). There exists c > 0, such that for every ǫ ≥ 1/nc,
3-SAT on inputs of size n can be efficiently reduced to Label-Cover of size N = n1+o(1)poly(1/ǫ) over an
alphabet of size poly(1/ǫ) that has soundness error ǫ. Moreover, the graph is bi-regular (namely, every two
questions on the same side participate in the same number of queries).
The difference between Theorem 3.2 and Conjecture 3.3 is in the size of the alphabet.
For our purposes, we only need soundness ǫ = 1/polylog(n), and we know that the degree and alphabet
size of the graph in Conjecture 3.3 are always polylog(n) (which are inverse of the soundness). Hence, we
may assume the (slightly) weaker assumption (obtained by setting ǫ = 1/polylog(n) in Conjecture 3.3) as
below.
Conjecture 3.4 (Projection Games Conjecture, a variant). There exists c > 0, such that for every ǫ =
1/polylog(n), 3-SAT on inputs of size n can be efficiently reduced to Label-Cover of size N = n1+o(1)poly(1/ǫ)
where the graph is bi-regular and all degrees are bounded by polylog(n). The size of the alphabet is
polylog(n) and the soundness is 1/polylog(n). and the completeness is 1.
We need to inspect very carefully and slightly change the proof of [32] since we do not want the
Label-Cover instance to grow by a lot by the modification in [32]. Hence, in fact we have to go over
all steps of [32] and bound the size more carefully in all steps that require that.
4 First part of the proof
We start with the same definition as in [32].
Definition 4.1 (Total disagreement). Let (G = (A, B, E), ΣA, ΣB, Φ) be a Label-Cover instance. Let
φA : A → ΣA be an assignment to the A-vertices. We say that the A-vertices totally disagree on a vertex
b ∈ B, if there are no two neighbors a1, a2 ∈ A of b, for which
πe1 (φA(a1)) = πe2 (φA(a2)) ,
where e1 = (a1, b), e2 = (a2, b) ∈ E.
The above simply states that for a given assignment φA and a vertex b ∈ B, no matter which label we
assign to the vertex b, we will satisfy only one edge incident to it.
Definition 4.2 (Agreement soundness). Let G = (G = (A, B, E), ΣA, ΣB, Φ) be a Label-Cover for de-
ciding whether a Boolean formula φ is satisfiable. We say that G has agreement soundness error ǫ, if for
unsatisfiable φ, for any assignment φA : A → ΣA, the A-vertices are in total disagreement on at least 1 − ǫ
fraction of the b ∈ B.
For a Yes-Instance (of 3-SAT), a standard argument implies that you can label the vertices so that every
edge is covered. The usual condition of soundness required is that the number of edges covered is a small
fraction of the edges, for every label assignment. The total disagreement is stronger than that. It states that
for any assignment φA, no matter how we set φB almost all of vertices of B will have at most one incident
edge satisfied.
In the rest of this subsection the goal is to show (list) agreement soundness error of bounded degree
Label-Cover instances. First, we use the following lemma (we do not alter its proof).
5
Lemma 4.3 (Combinatorial construction). For 0 < ǫ < 1, for a prime power D, and ∆ that is a power of
D, there is an explicit construction of a regular bipartite graph H = (U, V, E) with U = n, V -degree D, and
V ≤ nO(1) that satisfies the following. For every partition U1, . . . , Uℓ of U into sets such that Ui ≤ ǫU,
for i = 1, . . . , ℓ, the fraction of vertices v ∈ V with more than one neighbor in any single set Ui, is at most
ǫD2.
It is rather trivial to show the above lemma by a probabilistic method. Moshkovitz showed in [32] that
such graphs can be constructed deterministically via a simple and elegant construction.
In the next lemma, we show how to take a Label-Cover instance with standard soundness and convert
it to a Label-Cover instance with total disagreement soundness, by combining it with the graph from
Lemma 4.3. Here (as opposed to [32]) we have to bound the size of the created instance more carefully ([32]
only states that the size is raised to a constant power).
Lemma 4.4. Let D ≥ 2 be a prime power and let ∆ be a power of D. Let ǫ > 0. From a Label-Cover
instance with soundness error ǫ2D2 and B-degree n, we can construct a Label-Cover instance with agree-
ment soundness error 2ǫD2 and B-degree D. The transformation preserves the alphabets, and the size of the
created instance is increased by a factor poly(∆), namely by polynomial in the original B-degree.
Proof. Let G = (G = (A, B, E), ΣA, ΣB, Φ) be the original Label-Cover from the Projection Game Con-
jecture. Let H = (U, V, EH ) be the graph from Lemma 4.3, where ∆, D and ǫ are as given in the current
lemma. Let us use U to enumerate the neighbors of a B-vertex, i.e., there is a function E← : B × U → A
that, given a vertex b ∈ B and u ∈ U , gives us the A-vertex which is the u neighbor of b.
We create a new Label-Cover (G = (A, B × V, E′), ΣA, ΣB, Φ′). The intended assignment to every
vertex a ∈ A is the same as its assignment in the original instance. The intended assignment to a vertex
hb, vi ∈ B × V is the same as the assignment to b in the original game. We put an edge e′ = (a,hb, vi) if
there exist u ∈ U such that E←(b, u) = a and (u, v) ∈ EH . We define πe′ = π(a,b).
If there is an assignment to the original instance that satisfies c fraction of its edges, then the corresponding
assignment to the new instance satisfies c fraction of its edges (this follows from the regularity of the graph
H).
Suppose there is an assignment for the new instance φA : A → ΣA in which more than 2ǫD2 fraction of
Let us say that b ∈ B is good if for more than an ǫD2 fraction of the vertices in {b} × V the A-vertices
Focus on a good b ∈ B. Consider the partition of U into ΣB sets, where the set corresponding to σ ∈ ΣB
the vertices in B × V do not have total disagreement.
do not totally disagree. Note that the fraction of good b ∈ B is at least ǫD2.
is:
Uσ = {u ∈ Ua = E←(b, u) ∧ e = (a, b) ∧ πe(φA(a)) = σ} .
champion for b.
b arbitrary values. The fraction of edges that φA, φB satisfy in the original instance is at least ǫ2D2.
By the goodness of b and the property of H, there must be σ ∈ ΣB such that Uσ > ǫU. We call σ the
We define an assignment φB : B → ΣB that assigns good vertices b their champions, and other vertices
The new instance is bigger by a factor V , which is poly(∆).
Next we consider a variant of Label-Cover that is relevant for the reduction to Set-Cover. In this
variant, the prover is allowed to assign each vertex ℓ values, and an agreement is interpreted as agreement
on one of the assignments in the list.
Definition 4.5 (List total disagreement [32]). Let (G = (A, B, E), ΣA, ΣB, Φ) be a Label-Cover. Let
ℓ ≥ 1. Let φA : A → (cid:0)ΣA
ℓ (cid:1) be an assignment that assigns each A-vertex ℓ alphabet symbols. We say that the
A-vertices totally disagree on a vertex b ∈ B if there are no two neighbors a1, a2 ∈ A of b, for which there
exist σ1 ∈ φA(a1), σ2 ∈ φA(a2) such that
where e1 = (a1, b), e2 = (a2, b) ∈ E.
πe1 (σ1) = πe2 (σ2) ,
6
Definition 4.6 (List agreement soundness [32]). Let (G = (A, B, E), ΣA, ΣB, Φ) be a Label-Cover for
deciding membership whether a Boolean formula φ is satisfiable. We say that G has list-agreement soundness
error (ℓ, ǫ), if for unsatisfiable φ, for any assignment φA : A → (cid:0)ΣA
on at least 1 − ǫ fraction of the b ∈ B.
ℓ (cid:1), the A-vertices are in total disagreement
If a PCP has low error ǫ, then even when the prover is allowed to assign each A-vertex ℓ values, the game
is still sound. This is argued in the next corollary.
Lemma 4.7 (Lemma 4.7 of [32]). Let ℓ ≥ 1, 0 < ǫ′ < 1. Any instance of Label-Cover with agreement
soundness error ǫ′ has list-agreement soundness error (ℓ, ǫ′ℓ2).
The following corollary summarizes this subsection.
Corollary 4.8. For any ℓ = ℓ(n) = polylog(n), for any constant prime power D and constant 0 < α < 1,
3-SAT on input of size n can be reduced to a Label-Cover instance of size N = n1+o(1) with alphabet size
polylog(n), where the B-degree is D, and the list-agreement soundness error is (ℓ, α).
Proof. Our starting point is the Label-Cover instance from 3.4 with soundness error ǫ, such that 2√ǫ· l2 ≤
α. Note that the B-degree of the instance is ∆ = polylog(n). The corollary then follows by invoking
Lemma 4.4 and Lemma 4.7.
4.1 From Label-Cover to Set-Cover
Lemma 4.9 (Partition System [32]). For natural numbers m, D, and 0 < α < 1, for all u ≥ (DO(log D) log m)1/α,
there is an explicit construction of a universe U of size u and partitions P1, . . . ,Pm of U into D sets that
satisfy the following: there is no cover of U with ℓ = D lnU(1 − α) sets Si1, . . . , Siℓ, 1 ≤ i1 < . . . < iℓ ≤ m,
such that each set Sij belongs to the partition Pij .
We will use the contrapositive of the lemma:
if U has a cover of size at most ℓ, then this cover must
contain at least two sets from the same partition. Next follows the reduction, which is almost the same as
in [32], where the only difference is the parameter setting.
We take a Label-Cover instance G from Corollary 4.8 and transform it into an instance of Set-Cover.
In order to do so, we invoke Lemma 4.9 with m = ΣB and D which is the B-degree of G. The parameter
u will be determined later. Let U be the universe, and Pσ1 , . . . ,Pσm be the partitions of U , where the
partitions are indexed by symbols of ΣB. The elements of the Set-Cover instance are B × U , i.e., for each
vertex b ∈ B there is a copy of U . Covering {b} × U corresponds to satisfying the edges that touch b. There
are m ways to satisfy the edges that touch b -- one for every possible assignment σ ∈ ΣB to b. The different
partitions covering U correspond to those different assignments.
For every vertex a ∈ A and an assignment σ ∈ ΣA to a, we have a set Sa,σ in the Set-Cover instance.
Taking Sa,σ to the cover would correspond to assigning σ to a. Notice that a cover might consist of several
sets of the form Sa,· for the same a ∈ A, which is the reason we consider list agreement. The set Sa,σ
is a union of subsets, one for every edge e = (a, b) touching a. Suppose e is the i-th edge coming into b
(1 ≤ i ≤ D), then the subset associated with e is {b} × S, where S is the i-th subset of the partition PΦe(σ).
If we have an assignment to the A-vertices such that all of the neighbors of b agree on one value for b,
then the D subsets corresponding to those neighbors and their assignments form a partition that covers bs
universe. On the other hand, if one uses only sets that correspond to totally disagreeing assignments to the
neighbors, then by the definition of the partitions, covering U requires ≈ lnU times more sets. The formal
claim proved by Moshkovitz is as follows.
Claim 4.10 (Claim 4.10 of [32]). The following holds
• Completeness: If all the edges in G can be satisfied, then the created instance admits a set cover of size
A.
• Soundness: Let ℓ := D lnU(1 − α) be as in Lemma 4.9. If G has agreement soundness (ℓ, α), then
every set cover of the created instance is of size more than A lnU(1 − 2α).
7
The following is our main theorem, where we fine-tune the parameters to get the best possible (and thus
almost tight) running time lower bound.
Theorem 4.11. Fix a constant γ > 0 and ǫ > 0. Assuming PGC there is an algorithm that given an
instance φ of 3-SAT of size n one can create an instance I of Set-Cover with universe of size n1+o(1) · u
such that if φ is satisfiable, then I has a set cover of size x, while if φ is not satisfiable, then I does not
admit a set cover of size at most x lnu(1 − ǫ).
Proof. Given a sparsified 3-CNF formula φ of size n we transform it into a Label-Cover instance G, by
Corollary 4.8, obtaining a list-agreement soundness error (ℓ, α), where we set α = ǫ/2 and ℓ = D lnU(1−α).
Next, we perform the reduction from this section and by Claim 4.10 we have the following:
• If φ is satisfiable, then there exists a solution of size A (where A is one side of G).
• If φ is not satisfiable, then any set cover has size more than A lnU(1 − 2α) = A lnU(1 − ǫ).
By setting the value of U = u appropriately we get a tradeoff between the approximation ratio and
running time in the following lower bound obtained directly from Theorem 4.11.
Corollary 4.12. Unless the ETH fails, for any 0 < α < 1 and ǫ > 0 there is no (1 − α) ln n approximation
for Set-Cover with universe of size n and m sets in time 2nα−ǫ
Proof. Set u = φ1/α−1, then the created instance has at most φ1/α+o(1) elements, which fits the desired
running time in the lower bound. It remains to analyze the approximation ratio. Note that A ≤ uα/(1−α),
hence
poly(m).
(1 − α) ln(A · u) ≤ (1 − α)(α/(1 − α) + 1) ln u = ln u .
5 Approximating Directed Steiner Tree
In this section, we present a (1− ǫ)· ln n-approximation algorithm for Directed-Steiner-Tree running in
time 2O(nα log n).
Lemma 5.1. For any rooted tree T with ℓ leaves, there exists a set X ⊆ V (T ) of O(nα) vertices together
with a family of edge disjoint trees T1, . . . , Tq, such that:
• the trees are edge (but not vertex) disjoint
• each Ti is a subtree of T ,
• the root of each Ti belongs to X,
• each leaf of T is a leaf of exactly one Ti,
• each Ti has more than nα but less than 2nα leaves.
Proof. As long as the tree has more than nα leaves do the following: pick the lowest vertex v in the tree, the
subtree rooted at which has more than nα leaves. This implies that all its children contain strictly less than
nα leaves. Accumulation subtrees gives at most 2nα leaves since before the last iteration there were less that
nα leaves, and the last iteration adds a tree of at most nα leaves. Remove the collected tree, but do not
remove their root (namely this root may later participate in other trees). Note that after the accumulated
trees are removed, the tree rooted by our chosen root may still have more than nα leaves. This gives Θ(nα)
edge disjoint trees with Θ(nα) leaves each. Thus, there is a tree with Θ(nα) leaves, and density (cost over
the number of leaves) no larger than the optimum density.
8
For simplicity, we make sure that the number of leaves in each tree is exactly nα by discarding leaves.
Since the trees are edge disjoint there must be a tree whose density: cost over the number of leaves is not
worse (up to a factor of 2) than the optimum density opt/ℓ.
Let (G, K, r) be an instance of Directed-Steiner-Tree. Our algorithm enumerates guesses the
(roughly) nα leaves L′ in the tree whose density is no worse than the optimal density, and also guesses
the subset XL′ ⊆ V (G) that behaves as Steiner vertices. Assuming the graph went via a transitive closure,
the size of XL′ is at most the size of L′ of size O(nα). For a fixed set X, the algorithm first finds an optimum
directed Steiner tree T0. We note that assuming that the graph went via transitive closure, we may assume
that the number of Steiner vertices is less than the number of leaves, and so we may guess the Steiner
vertices at time n√n as well. It is known that given the Steiner vertices and the leaves of the tree, we can, in
polynomial time, find the best density tree with these leaves and these X vertices. The first such algorithm
is due to Dreyfus and Wagner [18]. The algorithm is quite non trivial and that uses dynamic programming.
The running time of the algorithm is O(3n) time which is negligible in our context.
We iterate adding more trees in this way. Each time we find the best density tree rooted at some vertex
of X which covers nα leaves, and add the edges to S. Each time it requires 2O(nα log n) time. Finally, when
there are less than (e2 + 1)nα unconnected terminals left, we find an optimum directed Steiner tree for those
vertices.
Lemma 5.2. The approximation ratio of the above algorithm is at most (1 − α) ln n.
Proof. Let T be an optimum Steiner tree spanning K, let opt be its cost and let X be the set from Lemma 5.1
for the tree T . Let us analyze the algorithm in the iteration when it chooses the set X properly, that is picks
the same set as Lemma 5.1. Note that since vertices of X belong to T the cost of the first tree T0 found by
the algorithm is at most opt. The last property of Lemma 5.1 guarantees, that our algorithm always finds
a tree with at least as good density as opt/r, where r is the number of not yet connected terminals. By the
standard set-cover type analysis we can bound the cost of all the best density trees found by the algorithm
by
e2·21−αX
i=n
opt
i
= opt · (Hn − He2·21−α) =(1 − α) ln n · opt.
Finally, the last tree is of cost at most opt, and it can be found in time 2O(nα log n).
Now we observe that the same theorem applies for the Connected-Polymatroid problem. Since
the function is both submodular and increasing for every collection of pairwise disjoint sets {Si}i=1k ,
Pk
i=1 f (Si) ≥ f (Sk
i=1 Si). Thus for a given α, at iteration i there exists a collection of leaves Si so that
f (Si)/c(Si) ≥ f (U )/c(U ). We can guess Si in time exp(nα · log n) and its set of Steiner vertices Xi in time
O(3nα
). Using the algorithm of [18], we can find a tree of density at most 2 · opt/nα. The rest of the proof
is identical.
6 Hardness for Group Steiner Tree under the ETH
In this section, we show that the approximation hardness of the group Steiner problem under the ETH,
which implies that the subexponential-time algorithm for Group-Steiner-Tree of Chekuri and Pal [12] is
nearly tight. This hardness result is implicitly in the work Halperin and Krauthgamer [25]. More precisely,
the following is a corollary of Theorem 1.1 in [25].
Theorem 6.1 (Corollary of Theorem 1.1 in [25]). Unless the ETH is false, for any parameter 0 < δ < 1,
there is no exp(2logδ−ε N )-time log2−δ−ε k-approximation algorithm for Group-Steiner-Tree, for any 0 <
ε < δ.
Proof Sketch of Theorem 6.1. We provide here the parameter translation of the reduction in [25], which will
prove Theorem 6.1.
9
A, Σℓ
A = Σℓ
The Reduction of Halperin and Krauthgamer. We shall briefly describe the reduction of Halperin and
Krauthgamer. The starting point of their reduction is the Label-Cover instance obtained from ℓ rounds of
parallel repetition. In the first step, given a d-regular Label-Cover instance G = (G = (A, B, E), ΣA, ΣB, φ)
completeness 1 and soundness γ, they apply ℓ rounds of parallel repetition to get a dℓ-regular instance of
Label-Cover G′ = (Gℓ = (Aℓ, Bℓ, E′), Σℓ
B, φ′). To simplify the notation, we let m = A = B, σ =
Σℓ
B be the number of vertices and the alphabet size of the Label-Cover instance G, respectively.
Then we have that the number of vertices and the alphabet size of G′ is 2mℓ and σℓ, respectively. In the
second step, they apply a recursive composition to produce an instance I of Group-Steiner-Tree on a
complete (2 · mℓ · σℓ)-ary tree bT of height H on k = dℓmℓ·H groups. Moreover, if G is a Yes-Instance (i.e.,
there is a labeling that covers all the constraints), then there is a feasible solution to Group-Steiner-Tree
with cost H with high probability, and if G is a No-Instance (i.e., every labeing covers at most γ fraction of
the edges), then there is no solution with cost less than βH 2 log k, for some sufficiently small constant β > 0.
In short, the above reduction gives an instance of Group-Steiner-Tree on a tree with N = O((σm)ℓH )
vertices, k = O(dℓmℓH ) groups, and with approximation hardness gap Ω(H log k). Additionally, the reduc-
tion in [25] requires ℓ > c0(log H + log log m + log log d) for some sufficiently large constant c0 > 0.
Subexponential-Time Approximation-Hardness. Now we derive the subexponential-time approxi-
mation hardness for Group-Steiner-Tree. We start by the nearly linear-size PCP theorem of Dinur [16],
which gives a reduction from 3-SAT of size n (the number of variables plus the number of clauses) to a
label cover instance G = (G = (A, B, E), ΣA, ΣB, φ) with completeness 1, soundness γ for some 0 < γ < 1,
A,B ≤ n · polylog(n), degree d = O(1) and alphabet sets ΣA,ΣB = O(1).
For every parameter 0 < δ < 1, we choose H = log1/δ−1 n, which then forces us to choose ℓ = Θ((1/δ −
1) log log n). Note that we may assume that δ ≪ n since it is a fixed parameter. Plugging in these parameter
settings, we have an instance of Group-Steiner-Tree on a tree with N vertices and k groups such that
N = (cid:16)O(1) · n1+o(1)(cid:17)Θ((1/δ−1) log log n)·log1/δ−1 n
= exp(cid:16)log1/δ+o(1) n(cid:17)
and
k = O(1)Θ((1/δ−1) log log n)(cid:16)n1+o(1)(cid:17)Θ((1/δ−1) log log n)·log1/δ n
= exp(cid:16)log1/δ+o(1) n(cid:17)
Observe that H ≥ log1−δ−o(1) k. Thus, the hardness gap is Ω(H log k) = Ω(log2−δ−o(1) k). This means that
any algorithm for Group-Steiner-Tree on this family of instances with approximation ratio log2−δ−ε k,
for any constant ε > 0, would be able to solve 3-SAT.
Now suppose there is an exp(2logδ−ε N )-time log2−δ−ε k-approximation algorithm for Group-Steiner-
Tree, for some 0 < ε < δ. We apply such algorithm to solve an instance of Group-Steiner-Tree derived
from 3-SAT as above. Then we have an algorithm that runs in time
exp(2logδ−ε N ) = exp(2(log1/δ+o(1) n)δ−ε
) = exp(2(log
(1+o(1))(δ−ε)
δ
n)) = exp(2o(log n)) = 2o(n)
This implies a subexponential-time algorithm for 3-SAT, which contradicts the ETH. Therefore, unless the
ETH is false, there is no exp(2logδ−ε N )-time log2−δ−ε k-approximation algorithm for Group-Steiner-Tree,
thus proving Theorem 6.1
Since we take log from the expression, the above is also true if we replace k by N . Combined with [12]
and [19], we have the following corollary which shows almost tight running-time lower and upper bounds for
approximating Group-Steiner-Tree and Covering-Steiner-Tree to a factor log2−δ N .
Corollary 6.2. The Group-Steiner-Tree and Covering-Steiner-Tree problems on graphs with n ver-
tices admit log2−δ n-approximation algorithms for any constant δ < 1 that runs in time exp(2(1+o(1)) logδ n).
In addition, for any constant ǫ > 1 there is no log2−δ−ǫ n approximation algorithm for Group-Steiner-Tree
and Covering-Steiner-Tree that runs in time exp(2(1+o(1)) logδ−ǫ n).
10
Remark: We omit the algorithm for the Connected-Polymatroid problem, as its similar to the
algorithm for Directed-Steiner-Tree. The lower bound holds because Connected-Polymatroid has
Set-Cover as a special case.
Acknowledgments. The work of Marek Cygan is part of a project TOTAL that has received funding
from the European Research Council (ERC) under the European Unions Horizon 2020 research and innova-
tion programme (grant agreement No 677651). Guy Kortsarz was partially supported by NSF grant 1540547.
Bundit Laekhanukit was partially supported by ISF grant no. 621/12, and I-CORE grant no. 4/11 and
by the 1000-youth award from the Chinese Government. Parts of the work were done while Guy Kortsarz
and Bundit Laekhanukit were at the Weizmann Institute of Science, and parts of it were done while Bundit
Laekhanukit was at the Max-Plank-Institute for Informatics.
References
[1] N. Bansal, P. Chalermsook, B. Laekhanukit, D. Nanongkai, and J. Nederlof. New tools and connections
for exponential-time approximation. CoRR, abs/1708.03515, 2017.
[2] Y. Bartal. Probabilistic approximations of metric spaces and its algorithmic applications.
In 37th
Annual Symposium on Foundations of Computer Science, FOCS '96, Burlington, Vermont, USA, 14-
16 October, 1996, pages 184 -- 193, 1996.
[3] E. Bonnet, B. Escoffier, E. J. Kim, and V. T. Paschos. On subexponential and fpt-time inapproxima-
bility. Algorithmica, 71(3):541 -- 565, 2015. Preliminary version in IPEC'13.
[4] ´E. Bonnet, M. Lampis, and V. T. Paschos. Time-approximation trade-offs for inapproximable problems.
J. Comput. Syst. Sci., 92:171 -- 180, 2018.
[5] ´E. Bonnet and V. T. Paschos. Sparsification and subexponential approximation. Acta Inf., 55(1):1 -- 15,
2018.
[6] N. Bourgeois, B. Escoffier, and V. T. Paschos. Efficient approximation of min set cover by moderately
exponential algorithms. Theor. Comput. Sci., 410(21-23):2184 -- 2195, 2009.
[7] N. Bourgeois, B. Escoffier, and V. T. Paschos. Approximation of max independent set, min vertex
cover and related problems by moderately exponential algorithms. Discrete Applied Mathematics,
159(17):1954 -- 1970, 2011.
[8] G. Calinescu and A. Zelikovsky. The polymatroid steiner problems. J. Comb. Optim., 9(3):281 -- 294,
2005. Preliminary version in ISAAC'04.
[9] P. Chalermsook, B. Laekhanukit, and D. Nanongkai. Independent set, induced matching, and pricing:
Connections and tight (subexponential time) approximation hardnesses. In 54th Annual IEEE Sym-
posium on Foundations of Computer Science, FOCS 2013, 26-29 October, 2013, Berkeley, CA, USA,
pages 370 -- 379, 2013.
[10] M. Charikar, C. Chekuri, T. Cheung, Z. Dai, A. Goel, S. Guha, and M. Li. Approximation algorithms
for directed steiner problems. J. Algorithms, 33(1):73 -- 91, 1999. Preliminary version in SODA'98.
[11] C. Chekuri, G. Even, and G. Kortsarz. A greedy approximation algorithm for the group steiner problem.
Discrete Applied Mathematics, 154(1):15 -- 34, 2006.
[12] C. Chekuri and M. P´al. A recursive greedy algorithm for walks in directed graphs. In 46th Annual
IEEE Symposium on Foundations of Computer Science (FOCS 2005), 23-25 October 2005, Pittsburgh,
PA, USA, Proceedings, pages 245 -- 253, 2005.
11
[13] V. Chv´atal. A greedy heuristic for the set-covering problem. Math. Oper. Res., 4(3):233 -- 235, 1979.
[14] M. Cygan, L. Kowalik, M. Pilipczuk, and M. Wykurz. Exponential-time approximation of hard prob-
lems. CoRR, abs/0810.4934, 2008.
[15] M. Cygan, L. Kowalik, and M. Wykurz. Exponential-time approximation of weighted set cover. Inf.
Process. Lett., 109(16):957 -- 961, 2009.
[16] I. Dinur. The PCP theorem by gap amplification. J. ACM, 54(3):12, 2007. Preliminary version in
STOC'06.
[17] I. Dinur and D. Steurer. Analytical approach to parallel repetition.
In Symposium on Theory of
Computing, STOC 2014, New York, NY, USA, May 31 - June 03, 2014, pages 624 -- 633, 2014.
[18] S. E. Dreyfus and R. A. Wagner. The Steiner problem in graphs. Networks, 1(3):195 -- 207, 1971.
[19] G. Even, G. Kortsarz, and W. Slany. On network design problems: Fixed cost flows and the covering
steiner problem. In Algorithm Theory - SWAT 2002, 8th Scandinavian Workshop on Algorithm Theory,
Turku, Finland, July 3-5, 2002 Proceedings, pages 318 -- 327, 2002.
[20] J. Fakcharoenphol, S. Rao, and K. Talwar. A tight bound on approximating arbitrary metrics by tree
metrics. J. Comput. Syst. Sci., 69(3):485 -- 497, 2004. Preliminary version in STOC'03.
[21] U. Feige. A threshold of ln n for approximating set cover. J. ACM, 45(4):634 -- 652, 1998. Preliminary
in STOC'96.
[22] D. Fotakis, M. Lampis, and V. T. Paschos. Sub-exponential approximation schemes for csps: From
dense to almost sparse. In 33rd Symposium on Theoretical Aspects of Computer Science, STACS 2016,
February 17-20, 2016, Orl´eans, France, pages 37:1 -- 37:14, 2016.
[23] N. Garg, G. Konjevod, and R. Ravi. A polylogarithmic approximation algorithm for the group steiner
In Proceedings of the Ninth Annual ACM-SIAM Symposium on Discrete Algorithms,
tree problem.
25-27 January 1998, San Francisco, California., pages 253 -- 259, 1998.
[24] A. Gupta and A. Srinivasan. On the covering steiner problem.
In FST TCS 2003: Foundations of
Software Technology and Theoretical Computer Science, 23rd Conference, Mumbai, India, December
15-17, 2003, Proceedings, pages 244 -- 251, 2003.
[25] E. Halperin and R. Krauthgamer. Polylogarithmic inapproximability. In STOC, pages 585 -- 594, 2003.
[26] J. Hastad. Clique is hard to approximate within n1-epsilon. In 37th Annual Symposium on Foundations
of Computer Science, FOCS '96, Burlington, Vermont, USA, 14-16 October, 1996, pages 627 -- 636, 1996.
[27] R. Impagliazzo and R. Paturi. On the complexity of k-sat. J. Comput. Syst. Sci., 62(2):367 -- 375, 2001.
Preliminary version in CCC'99.
[28] R. Impagliazzo, R. Paturi, and F. Zane. Which problems have strongly exponential complexity? J.
Comput. Syst. Sci., 63(4):512 -- 530, 2001. Preliminary version in FOCS'98.
[29] D. S. Johnson. Approximation algorithms for combinatorial problems. J. Comput. Syst. Sci., 9(3):256 --
278, 1974. Preliminary version in STOC'73.
[30] G. Kortsarz and D. Peleg. Approximating the weight of shallow steiner trees. Discrete Applied Mathe-
matics, 93(2-3):265 -- 285, 1999. Preliminary version in SODA'97.
[31] C. Lund and M. Yannakakis. On the hardness of approximating minimization problems. J. ACM,
41(5):960 -- 981, 1994.
12
[32] D. Moshkovitz. The projection games conjecture and the np-hardness of ln n-approximating set-cover.
Theory of Computing, 11:221 -- 235, 2015. Preliminary version in APPROX'12.
[33] D. Moshkovitz and R. Raz. Two-query PCP with subconstant error. J. ACM, 57(5):29:1 -- 29:29, 2010.
Preliminary version in FOCS'08.
[34] L. A. Wolsey. An analysis of the greedy algorithm for the submodular set covering problem. Combina-
torica, 2(4):385 -- 393, 1982.
[35] D. Zuckerman. Linear degree extractors and the inapproximability of max clique and chromatic number.
Theory of Computing, 3(1):103 -- 128, 2007. Preliminary version in STOC'06.
13
|
1807.07719 | 1 | 1807 | 2018-07-20T07:16:22 | On Euclidean Methods for Cubic and Quartic Jacobi Symbols | [
"cs.DS"
] | We study the bit complexity of two methods, related to the Euclidean algorithm, for computing cubic and quartic analogs of the Jacobi symbol. The main bottleneck in such procedures is computation of a quotient for long division. We give examples to show that with standard arithmetic, if quotients are computed naively (by using exact norms as denominators, then rounding), the algorithms have $\Theta(n^3)$ bit complexity. It is a "folk theorem" that this can be reduced to $O(n^2)$ by modifying the division procedure. We give a self-contained proof of this, and show that quadratic time is best possible for these algorithms (with standard arithmetic or not).
We also address the relative efficiency of using reciprocity, as compared to Euler's criterion, for testing if a given number is a cubic or quartic residue modulo an odd prime. Which is preferable depends on the number of residue tests to be done.
Finally, we discuss the cubic and quartic analogs of Eisenstein's even-quotient algorithm for computing Jacobi symbols in ${\bf Z}$. Although the quartic algorithm was given by Smith in 1859, the version for cubic symbols seems to be new. As far as we know, neither was analyzed before. We show that both algorithms have exponential worst-case bit complexity. The proof for the cubic algorithm involves a cyclic repetition of four quotients, which may be of independent interest. | cs.DS | cs | On Euclidean Methods for Cubic and Quartic Jacobi Symbols
Eric Bach and Bryce Sandlund
Computer Sciences Dept.
University of Wisconsin
Madison, WI 53706
July 18, 2018
8
1
0
2
l
u
J
0
2
]
S
D
.
s
c
[
1
v
9
1
7
7
0
.
7
0
8
1
:
v
i
X
r
a
Abstract. We study the bit complexity of two methods, related to the Euclidean algo-
rithm, for computing cubic and quartic analogs of the Jacobi symbol. The main bottleneck
in such procedures is computation of a quotient for long division. We give examples to
show that with standard arithmetic, if quotients are computed naively (by using exact
norms as denominators, then rounding), the algorithms have Θ(n3) bit complexity. It is
a "folk theorem" that this can be reduced to O(n2) by modifying the division procedure.
We give a self-contained proof of this, and show that quadratic time is best possible for
these algorithms (with standard arithmetic or not).
We also address the relative efficiency of using reciprocity, as compared to Euler's
criterion, for testing if a given number is a cubic or quartic residue modulo an odd prime.
Which is preferable depends on the number of residue tests to be done.
Finally, we discuss the cubic and quartic analogs of Eisenstein's even-quotient algo-
rithm for computing Jacobi symbols in Z. Although the quartic algorithm was given by
Smith in 1859, the version for cubic symbols seems to be new. As far as we know, neither
was analyzed before. We show that both algorithms have exponential worst-case bit com-
plexity. The proof for the cubic algorithm involves a cyclic repetition of four quotients,
which may be of independent interest.
Presented as a poster at the 13th Algorithmic Number Theory Symposium (ANTS-XIII),
Madison, WI, July 16-20, 2018.
Bryce Sandlund is now at the University of Waterloo. E-mail addresses for authors:
[email protected], [email protected]
1
1. Introduction.
Let F be a finite field, of order q. When e divides q − 1, we can test whether x is
an e-th power in F using Euler's criterion. This is an "exactness" result, which asserts
that, on F∗
q, the image of the e-th power map is the kernel of the (q − 1)/e-th power map.
It therefore reduces the "hard" search for a y making ye = x to an "easy" evaluation of
x(q−1)/e.
It is natural to ask whether exponentiation can be avoided. It definitely can be, when
e = 2 and q = p, a prime of Z. This is because we can determine the quadratic character
by computing a Jacobi symbol, using a successive division process similar to the Euclidean
algorithm. The correctness of this algorithm rests on the quadratic reciprocity law, as
extended by Jacobi to include composite moduli. For several such algorithms, see [21].
In this paper we focus on e = 3, 4. Since the results and methods are very similar in
both cases, we will first restrict to e = 3, and then take up e = 4 toward the end of the
paper.
When e = 3, it is natural to move to the ring Z[ρ], with ρ = −1/2+√−3/2 a primitive
cube root of unity. When Fq is given to us as Z[ρ]/(π) we can replace quadratic by cubic
reciprocity, and produce an analogous test that is free of exponentiation. As far as we know,
the first complete description of such a test appeared in 1977, in a paper by Williams and
Holte [26]. They said nothing about its complexity, but later Williams [25] pointed out that
the number of [arithmetic] operations is logarithmic. (See also Muller [19].). Later (2005),
Damgard and Frandsen [5] pointed out that the bit complexity is O(nM (n)), assuming we
can multiply n-bit numbers with M (n) bit operations. If standard arithmetic is used, this
bound is O(n3).
Before describing our contributions we briefly mention some related work. Damgard
and Frandsen [5] gave Jacobi symbol algorithms for e = 3, 4, related to the binary algorithm
for the gcd in Z, that use O(n2) bit operations. Wikstrom [24] did the same for e = 8.
Weilert found an algorithm for the quartic Jacobi symbol with bit complexity n1+o(1), and
suggested that a similar result could be proved for the cubic symbol [23, p. 145]. (We do
not know if this was ever done.) Jacobi symbol methods for prime e ≤ 19, related to the
Euclidean algorithm, were presented by Scheidler and Williams [20]. (Special attention
was paid to e = 5.)
One goal of our paper is to give a complete self-contained discussion of the bit com-
plexity of the Williams-Holte algorithm. As far as we know this has never been done. In
particular, we make the following observations. First, the Damgard-Frandsen bound is
tight up to constant factors: we exhibit a sequence of input pairs that forces Ω(nM (n))
bit operations. Second, following ideas already used by Collins [3] to compute the gcd in
Z[i], approximate quotients can be used in the Williams-Holte algorithm to obtain O(n2)
bit complexity. (Earlier, Kaltofen and Rolletschek [16] had computed gcd ideals in rings
of quadratic integers in quadratic time using a related but more complicated process.)
Finally, even if division could be done for free, the Williams-Holte algorithm would still
need Ω(n2) bit operations in the worst case, just to write down the remainders.
Our original motivation for this work was to investigate the preference, expressed
by Muller and others, for the cubic Jacobi symbol algorithm over Euler's criterion.
In
conjunction with this, we note that there are really two different computational problems:
2
a) Given p ∈ Z, determine if a is a cube mod p, and b) Given π ∈ Z[ρ], determine if α is a
cube mod π. The situation is more complicated than appears at first glance, and we will
discuss it further in Section 6.
In Section 7, we discuss the cubic and quartic analogs of an elegant but flawed algo-
rithm of Eisenstein [7] for computing the Jacobi symbol in Z. The quartic version was
given by Smith about 15 years after Eisenstein, but the cubic version may well be new.
We show that both of these algorithms have exponential bit complexity. The proof for
the cubic algorithm involves some interesting dynamics:
for certain input pairs, there
is a repeating cycle of quotients, which has the effect of subtracting constants from the
pair. Consequently, these intermediate numbers decrease much more slowly than with the
Williams-Holte algorithm.
2. Mathematical Background.
Let Z denote the integers. The ring of Eisenstein integers is Z[ρ], where ρ = −1/2 +
√−3/2 is a primitive cube root of unity. It is the ring of algebraic integers of the imaginary
quadratic field Q(ρ). Since ρ2 + ρ + 1 = 0, any Eisenstein integer can be written as a + bρ,
where a, b ∈ Z. The norm of α = a + bρ is
N (α) = α ¯α = a2 − ab + b2.
When thinking of α as a complex number, we will write α = pN (α).
We can also think of the numbers a + bρ, where a and b are real, as a 2-dimensional
vector space over R. Let α2 = a2 + b2. From examining the quadratic form a2 − ab + b2,
we can derive the explicit norm-comparison inequality
2
3
N (α) ≤ α2 ≤ 2N (α).
(1)
A strong version of long division holds in Z[ρ]: Given α and nonzero β, there are
integral q, γ for which
α = qβ + γ,
N (γ) ≤
3
4
N (β).
(2)
We can take q = ⌊α/β⌉, where the brackets indicate each coefficient is rounded to a nearest
integer [13, p. 13]. Consequently, Z[ρ] is a principal ideal domain. The units of this ring,
the sixth roots of unity, are ±1, ±ρ, ±ρ2 (or, ±(1 + ρ)).
When P is a prime ideal of Z[ρ], distinct from (1 − ρ), the cubic symbol (cid:0) α
the unique element of {0, 1, ρ, ρ2} for which
A prime p of Z is ramified, split, or inert when it is 0, 1, or 2 mod 3, respectively.
P(cid:1) is defined as
(cid:16) α
P(cid:17) ≡ α(NP −1)/3
(mod P ).
Here, N P denotes the norm of P , that is, the size of Z[ρ]/P . This extends to an analog of
the Jacobi symbol, as follows. Let P = Q1 · · · Qr be a factorization into such prime ideals,
with repetitions allowed. Then,
P(cid:17) =
(cid:16) α
r
Yi=1
(cid:18) α
Qi(cid:19).
3
For convenience, we also allow r = 0 (that is, P = (1)), in which case the symbol is 1.
(This follows [2].) This cubic symbol is asymmetric, as its arguments are a number and
an ideal. To make it symmetric, we let (cid:16) α
β(cid:17) denote (cid:16) α
(β)(cid:17).
The cubic symbol obeys a reciprocity law for elements, but it must be stated carefully,
since multiplying the upper argument by a unit can change the symbol's value. Accord-
ingly, we will say that α is primary if α ≡ ±1 mod 3. Note that a primary number cannot
be divisible by 1 − ρ, and the only primary units are ±1. Some authors make a more
restrictive definition; to encompass this we let 2-primary mean ≡ 2 mod 3. Since Z[ρ]/(3)∗
is cyclic of order 6, each ideal prime to 3 has exactly one 2-primary generator.
The cubic reciprocity law consists of a main theorem and two supplementary results.
First, suppose that α, β are primary and relatively prime. Then
(cid:18) α
β(cid:19) = (cid:18) β
α(cid:19)
Further, suppose β = c + dρ, with c = 3c′ ± 1 and d = 3d′. Then
(cid:18) 1 − ρ
β (cid:19) = ρ±c′
= ρ−(c2−1)/3.
and
β(cid:19) = ρ∓(c′+d′) = ρ(c2−cd−1)/3
(cid:18) ρ
(3)
(4)
(5)
(6)
(The sign in (4) matches the one in the definition of c′, and the sign in (5) is its opposite.)
For completeness we note that
(cid:18)−1
β (cid:19) = 1,
under the same condition.
As Collison [4] recounts, the history of cubic reciprocity makes for a complicated tale.
(See also [10, 18, 22].) Suffice it to say that (3)–(5) were first proved in print by Eisenstein
in 1844 [6, 8]. For a modern exposition of these proofs, see Halter-Koch [12, §7.2].
We end this section with a few remarks on arithmetic in Z[ρ]. M (m, n) stands for the
number of bit operations (bops) used to multiply an m-bit integer and an n-bit integer.
We can assume that this depends only on input length. Following [1, p. 280], we require
M (n) = M (n, n) to satisfy a regularity condition: if a ≥ 1, then
M (n) ≤ M (an) ≤ a2M (n).
(7)
The discrete binary logarithm (bit length function) is given by
lg n = (cid:26) 1,
⌊log2 n⌋ + 1,
if n = 0;
if n 6= 0.
By (1), if α = a + bρ, we have a2 + b2 ≤ 2N (α). So
lg a, lg b = O(lg N (α)).
4
Therefore, we can compute α ± β using O(lg N (α) + lg N (β)) bops. We can also compute
α· β using O(M (lg N (α), lg N (β))) bops. As for long division, we can take q = ⌊α ¯β/N (β)⌉
(The brackets indicate that
and then γ = α − qβ; this will use M (lg N (αβ)) bops.
coordinates are rounded to nearest elements of Z.) If we only want the remainder, there
is a formula of Jacobi [14, pp. 249-250]:
γ =
[α ¯β mod N (β)]β
N (β)
(absolutely least remainders for the expression in brackets), with the same complexity. We
will say more about long division in Section 5.
3. Cubic Jacobi Symbols via Iterated Division.
Just as in Z, one can use continued fractions, or a variation thereof, to compute cubic
Jacobi symbols. This was already known to Jacobi. In a paper from the autumn of 1837
[15] that introduced the Jacobi symbol for Z, he noted that Gauss's method for computing
the Legendre symbol (cid:16) p
q(cid:17) from the continued fraction of p/q (presumably from [11]) would
still work with p, q just relatively prime, and then wrote (translation ours):
Exactly the same can be applied to biquadratic and cubic residues, for which I have
introduced similar symbols. The application of the symbols thus generalized affords,
with some practice, agreeable relief.
Although it is reassuring to read that even the great masters needed to hone their compu-
tational skills, we know of no place where Jacobi explicitly wrote down such an algorithm
for the cubic symbol. (In his number theory lectures [14], delivered the preceding winter,
he sketched an algorithm for the quartic symbol employing factorization, but said nothing
directly about computing cubic symbols.)
Eisenstein mentioned a cubic symbol procedure in 1844 [8, p. 32], but gave no details.
He did present a full algorithm for the quartic symbol a short time later [9, §10]. In the
cubic case, full disclosure apparently had to wait until 1977, when Williams and Holte
[26] wrote up the details for an algorithm that we now recognize as the cubic analog of
Eisenstein's quartic method. We now give their algorithm, with an extension that allows
for inputs that are not relatively prime.
1. [Inputs.] Let α = a + bρ and β = c + dρ, with c 6≡ 0, d ≡ 0 mod 3. (That is, β is
primary.)
2. [Base cases.] If β = ±1 or α = ±1, return 1 and stop.
3. [Find remainder.] Let q = ⌊α/β⌉. Compute
γ = e + f ρ = α − qβ.
If γ = 0, return 0 and stop.
4. [Remove ramified factors.] Find the least m ≥ 0 for which
γ′ = e′ + f ′ρ =
e + f ρ
(1 − ρ)m
5
is not divisible by 1 − ρ. (That is, has e′ + f ′ 6≡ 0 mod 3.)
5. [Make primary.] Find n with 0 ≤ n < 3 so that
e′ + f ′ρ
γ′′ =
ρn
is primary.
6. [Recursion.] Return the right side of
(cid:18) α
β(cid:19) = ρ−m(c2−1)/3+n(c2−cd−1)/3(cid:18) β
γ′′(cid:19)
The algorithm must terminate because N (β) strictly decreases. We briefly indicate
why the output is correct. When β is a unit, the value of the symbol is 1, and if not, but
with α = ±1, the same is true. So let there be at least one division step. Using indexing
to denote recursion depth, we have
α1 = q1β1 + γ1
α2 = q2β2 + γ2
...
= q1β1 + (1 − ρ)m1ρn1 γ′′
1 ,
= q2β2 + (1 − ρ)m2ρn2 γ′′
2 ,
(8)
αt−1 = qt−1βt−1 + γt−1 = qt−1βt−1 + (1 − ρ)mt−1ρnt−1 γ′′
t−1,
αt = qtβt,
where α2 = β1, β2 = γ′′
1 , and so on. (We imagine the last division is done, whether needed
or not.) Since β is primary, gcd(α, β) is coprime to 1 − ρ, so (using induction), βt is a
gcd of α, β. If N (βt) > 1, the last call returns 0, which becomes, correctly, the output.
If N (βt) = 1 (that is, γ′′
t−1 is ±1), the t-th call returns 1, and correctness of the output
follows from (3)–(5).
Note that the algorithm computes a gcd, free of charge, in cases where the symbol is
0. This is similar to the situation in Z [21, p. 608].
Let us now discuss some implementation details. Williams and Holte computed the
ratio in Step 3 as
α ¯β
N (β)
In Step 4, we can divide by 1 − ρ using the formula
α
β
=
.
e + f ρ
1 − ρ
=
2e − f
3
+
e + f
3
ρ;
the result is integral iff 3 divides e + f . Regarding Step 5, division by ρ (that is, multipli-
cation by ρ2) permutes
e + f ρ,
(f − e) − eρ, −f + (e − f )ρ
6
in cyclic order. Examination of cases shows that exactly one of these is primary. Finally,
the numerators in the exponent in Step 6 need only be computed mod 9. Williams and
Holte gave this exponent in the equivalent form (2m + n)(c2 − 1)/3 − ncd/3.
4. A "Bad" Sequence of Input Pairs.
Our goal is to construct a sequence of inputs for which the Williams-Holte algorithm
needs nM (n) bit operations.
Consider the recurrence relation
ξ0 = −1,
ξ1 = 2,
ξn = 3ρξn−1 + ξn−2,
n ≥ 2.
This defines a sequence of 2-primary integers in Z[ρ], because all ξn are 2 mod 3, as can
be proved by induction. The first few sequence elements are
n
ξn
N (ξn)
−1
0
1
2
1
4
−1 + 6ρ
2
43
−16 − 21ρ
3
361
62 + 21ρ
4
2983
−79 + 102ρ
5
24703
−244 − 522ρ
6
204652
1487 + 936ρ
7
1695433
−3052 + 1131ρ
8
14045677
9 −1906 − 11613ρ 116360227
31787 + 30252ρ 963976549
10
We now show that these grow rapidly. The characteristic polynomial for the recurrence
relation is
with zeroes
X 2 − 3ρX − 1,
3ρ ±p9ρ2 + 4
2
λ, µ =
(9)
.
(Incidentally, λ 6∈ Q(ρ) since (9) is irreducible modulo any prime divisor of 19.) Numeri-
cally,
.
= −1.7059 + 2.3182i,
.
= 0.2059 + 0.2798i,
λ
µ
.
= 2.8783,
.
= 0.3474.
λ
µ
By the theory of linear recurrence relations, there are numbers A and B for which
ξn = Aλn + Bµn.
7
(Numerically, A
.
= −0.5331 + 0.6442i.)
For large n the second term is insignificant, and we have
.
= −0.4669 − 0.6442i, B
log(N (ξn)) ∼ n log(N (λ))
.
= 2.1144n.
So, in norm, the length of ξn grows linearly, and the number of recursive calls for the
algorithm on the inputs ξn, ξn−1 is n.
We now show that applying the algorithm with α = ξn and β = ξn−1 is tantamount
to running the recurrence backwards. Observe first that if a + bρ = x + yi, we have
a = x +
b =
y,
1
√3
2
√3
y.
So for large n,
ξn/ξn−1 ∼ λ
.
= −0.3675 + 2.6769ρ,
which rounds to 3ρ. (Apparently n ≥ 3 is large enough.) Therefore, the algorithm will
produce
for the next stage.
If ξn = an + bnρ, it can be shown that the vector
γ = ξn − 3ρξn−1 = ξn−2
vn = (cid:18) an
bn(cid:19)
satisfies the recurrence vn = M vn−1 + vn−2, where M is a 2 × 2 matrix. The question of
the lengths of an and bn is more delicate, as they don't always increase. This is already
apparent from the indicated values.
To get around this problem, we note that when N (a + bρ) is large, at least one of a
or b is large. Indeed, by (1), a2 + b2 ≥ 2
3 N (a + bρ), so
max{a, b} ≥ (N (a + bρ)/3)1/2 .
(10)
Suppose we implement the Williams-Holte algorithm by first computing an exact
fractional quotient α ¯β/N (β), using the usual norm formula.
If the inputs are ξn, ξn−1,
just computing the denominators alone will use Θ(n) multiplications in Z[ρ], of numbers
whose norms decrease geometrically over time. At the halfway point, we must still square
a integer whose length is Ω(n), by (10). Since (7) implies M (n/c) ≥ M (n)/c2, we have
already used Ω(nM (n)) bops by this point.
In particular, with standard arithmetic, the worst-case bit complexity of the algorithm
is cubic: on the inputs α, β it is Θ(lg N (αβ))3.
8
Instead of the standard norm formula, we could use others, such as (a − b)2 + ab
or ((2a − b)2 + 3b2)/4. However, any reasonable alternative would lead us to the same
conclusion. Let us restrict attention to algorithms that compute the norm as
1
s
k
Xi=1
(tia + uib)(via + wib)
where s, t1, . . . , wk are fixed integers. (Algebraic complexity theory supports this choice
[27, §IIIc].) If ℓ = max{lg a, lg b}, then
lg (tia + uib), lg (tia + uib) ≤ ℓ + O(1).
On the other hand, lg (N (a + bρ)) ≥ 2ℓ − O(1), by (1). We must multiply at least one
pair of integers whose lengths are ≥ ℓ/2, for if not, the sum on the right hand side would
have length ≤ 3ℓ/2 + O(1), and this would still hold after division by s. (Recall that k is
fixed.) Since M (ℓ/2) ≥ M (ℓ)/4, this is Ω(M (ℓ)) bops.
Incidentally, the minimum value of k is 2. To see this, observe that any one-
multiplication algorithm that yields a2 − ab + b2 could be normalized into (a + ub)(a + wb),
with u, w ∈ Q. This forces uw = 1, u + w = −1, which is a system with no real solutions.
5. A Quadratic Bit Complexity Bound.
This section provides a complete analysis for the the bit complexity of an approximate-
division version of the Williams-Holte algorithm. Many of the techniques we use reflect
previous analyses of Euclidean gcd algorithms, so the quadratic bit complexity bound is
probably a "folk theorem." Nevertheless, we thought it worthwhile to record a proof,
because no reference known to us does this.
Let us first review the standard Newton iteration for computing β−1, where β is a
nonzero complex number. This iteration is
Suppose that ξ = (1 − ǫ)/β with ǫ < 1. Then
ξ′ = ξ(2 − βξ).
ξ′ =
1 − ǫ
β
(1 + ǫ) =
1 − ǫ2
β
.
By induction, the i-th iterate will be
ξi =
1 − ǫ2i
β
.
(The starting value is denoted ξ0.) To obtain relative error ≤ 2−m, we need ǫ2i
which we can solve to obtain
≤ 2−m,
i ≥ log m − log log ǫ−1.
9
(Here, the logs are binary.) With the usual expedient of doubling the precision at each
step, this will require O(M (m)) bops. Here we are assuming that ǫ belongs to a compact
subset of the unit interval (0, 1); the implied constant will depend on this subset.
We now consider a particular starting point. Let β = c + dρ be a nonzero Eisenstein
integer. Then, there are r, s making 2r−1 ≤ c < 2r and 2s−1 ≤ d < 2s. (If c = 0 we take
r = −∞, and similarly for d.) In Q(ρ), the exact inverse is
β−1 =
c − d
N (β) −
d
N (β)
ρ.
Then, with e = 2 max{r, s} + 2, the starting point
ξ0 =
c − d
2e −
d
2e ρ
satisfies ξ0 = ηβ−1 with 1/32 ≤ η ≤ 3/4. (To prove this, observe that (c2 + d2)/2 ≤
22 max{r,s} ≤ 4(c2 + d2), and then use (1) to get (4/3)N (β) ≤ 2e ≤ 32N (β).) Since
ǫ = 1 − η, this implies 1/4 ≤ ǫ ≤ 31/32.
We now consider the bit complexity of long division. Let us argue informally. The
coefficients of any number in Q(ρ) can be written in base 2, and if these parts are aligned,
each bit pair can be chosen in four possible ways. Thus, 35/2 + 3ρ, equal to 1001.1 +
11ρ in base 2, becomes ( 1
0 ) . . ., or, using a straightforward encoding,
2013.200 · · · .
of β, have the representations:
Let α, β ∈ Z[ρ] be the inputs for long division. Let these, along with the exact inverse
1 ).( 1
0 )( 0
0 )( 0
1 )( 1
0 )( 0
0 )( 0
.
{
m digits
z
∗ ∗ · · · ∗ ∗∗ ∗ ∗ · · ·
}
k digits
α =
β =
β−1 =
ℓ digits
{
{
z
∗ ∗ ∗ ∗ ∗ ∗ · · · ∗ ∗∗ .000 · · ·
}
z
∗ ∗ ∗ · · · ∗ ∗∗ .000 · · ·
}
ℓ digits
{
z
000 · · · 000
}
.
Since the quotient is 0 if k < ℓ, we can assume ℓ ≤ k. To obtain the rounded quotient,
numerical inversion of β must deliver m = k − ℓ + O(1) digits past the leading ℓ 0's.
For Newton's method with precision doubling, as indicated above, this uses O(M (m))
bops. Multiplying this approximate inverse by α and then rounding the result to q uses
O(M (m, k)) bops. (One can show that if the error in each computed coefficient of αβ−1 is
≤ δ, the constant in (2) is increased only slightly, to 3/4 + 3δ + O(δ2).) Finally, we obtain
qβ with O(M (m, ℓ)) bops. This totals to
O(M (k, m) + M (m) + M (ℓ, m))
(11)
There is one further subtraction to produce γ, which we will ignore since it will use no
more work than any bound we prove.
10
For standard arithmetic, M (m, n) = O(mn) so (11) becomes
O(km + (m + ℓ)m) = O(km) = O(lg N (α)lg N (q)).
(12)
As a worst-case bound, this is not impressive, since the quotient could be almost as long
as the dividend. However, for a Jacobi symbol calculation, not all quotients can be large.
This makes it useful to have the quotient length as a factor.
Our final goal in this section is a quadratic bit complexity bound for the Williams-Holte
algorithm, with long division implemented in the above manner. This will be achieved using
only standard arithmetic. As we will see this bound cannot be improved.
Since the norms can be compared within our time bound, we will assume that N (α) ≥
N (β). By (2) the maximum recursion depth t (from (8)) is O(lg N (β)). Now we consider
each part of the algorithm separately.
First, we consider the work for long division (Step 3). We first note that
log(x + 1) ≤ log(x) + 1
(13)
(binary logs), provided that x ≥ 1. This can be proved by comparing derivatives. Using
(12), the number of bops in the long divisions is at most a constant times
t
Xi=1
lg N (αi)lg N (qi) ≤ lg N (α)
t
(2 log αi/βi + 3)
≤ lg N (α)
≤ lg N (α)(2 logα1/βt + 3t)
= O(lg N (α))2,
(2 log qi + 1)
Xi=1
Xi=1
t
using the triangle inequality and (13).
Next we consider Step 4. Let us call a division by 1 − ρ successful if the quotient
is integral, and unsuccessful if it is not. There are ≤ t unsuccessful divisions, and since
N (1 − ρ) > 1, O(lg N (β)) successful ones. The work for each such division is linear, so
Step 4 accounts for O(lg N (β))2 bops.
The same is true of Step 5, since each invocation of it uses O(lg N (β)) bops.
Finally, for Step 6, we assume c and d are reduced mod 9 before computing the
exponent. This uses O(lg N (β)) bops. The multiplier n is bounded, and it is simplest
conceptually to imagine multiplication by m as repeated addition, since the sum of all such
multipliers is O(lg N (β)). Then, Step 6 will account for O(lg N (β))2 bops.
Adding the last four bounds and using N (β) ≤ N (α) gives us O(lg N (α))2, as claimed.
To end this section, we point out that the quadratic bit complexity estimate is best
possible for this algorithm. Indeed, it holds for any varition of it that computes sufficiently
accurate quotients. To see this, suppose the inputs are α = ξn and β = ξn−1, as defined
in Section 4. The ratio of two successive ξi's is about −0.37 + 2.7ρ. If their quotient is
computed accurately enough, there will be no ambiguity about the rounded coefficients,
11
since the rounding can only have more than one possible result if a coefficient is an integer
multiple of 1/2. Therefore, the approximate-quotient algorithm will "track" the exact
algorithm for all but the last few steps (and produce primary remainders). Just to write
down the successive remainders will require Ω(lg N (α))2 bops.
Assuming Step 4 searches by incrementing m, we can can also force a linear number
of divisions here, by choosing α = 3m + (1 − ρ)m + 1 and β = 3m + 1, which is primary.
Then, α/β = 1 + (1 − ρ)m/(3m + 1) which rounds to 1 since (1 − ρ)m/(3m + 1) < 3−m/2.
At the top level, the number of divisions by 1 − ρ in Step 4 will be m + 1, with the last
(unsuccessful) one leaving γ′′ = 1. With these inputs (at least with binary representation),
Step 4 alone would consume Ω(lg N (α))2 bops.
6. Cubic Residue Tests.
The reciprocity law suggests a procedure for determining if a ∈ Z is a cubic residue
of the prime p ∈ Z [19]. We need only consider p ≡ 1 mod 3, and assume a is prime to p.
The first step is to solve the norm equation
for integers x, y. It is sufficient to solve the related equation
p = x2 − xy + y2
p = s2 + 3t2
(15)
(16)
and then set x = s + t, y = 2t. To solve (16), we can first solve s2 + 3 ≡ 0 mod p, and then
determine t from the ordinary continued fraction expansion for s/p. See [25, p. 366] for
details of this. Using standard algorithms for the square root mod p, the bit complexity
of this is O(nM (n)); the expensive part is finding s. By (15), one of the prime ideals P
lying above p is generated by π = x + ρy, which we can make primary if it is not already.
The cubic Jacobi symbol (cid:0) a
π(cid:1) is 1 iff a is a cubic residue mod p.
Whether this beats Euler's criterion depends on several factors. If we test many a
against one p, then the effort of computing π is, in effect, amortized over the sequence of
a's. So the procedure is definitely worth using, once we have about log p values of a. For
a "one-off" cubic residue test, Euler's criterion should be used, since all the most efficient
algorithms to find square roots mod p use exponentiation anyway. Before digital computers,
however, quadratic partitions such (16) were systematically tabulated. (Jacobi seems to
have been the first to do this; see the 1846 version of [15].) With this representation of p
in hand, the advantage then shifts to the cubic Jacobi algorithm, assuming long division
is implemented efficiently.
As an aside, we note that for p ≡ 1 mod 3, finding √−3 mod p reduces to solving the
norm equation for p. For, suppose we know integers x, y making x2 − xy + y2 = p. Then
we have
2 (cid:19)2
(cid:18) x + y
+ 3(cid:18) x − y
2 (cid:19)2
= p,
so that
z =
x + y
x − y
12
is a square root of −3 mod p. The denominator cannot vanish mod p; if it did, then from
(x − y)x + y2 = p we could conclude p y2 and then py, which contradicts x2 + y2 ≤
2N (x + ρy) = 2p (a consequence of (1)).
We end this section with a few remarks on the tabulation and computation of solutions
to (16). By 1925, Cunningham had tabulated the (essentially unique) solutions for all
eligible primes up to p = 125683. Lehmer's 1941 survey of number-theoretic tables [17]
mentioned nothing better, so this record likely still stood at the dawn of the computer age.
It is not clear to us if anyone went further, but it is worth noting that a list for all such
p up to N could be made, in amortized polynomial time, by looping through all positive
x ≤ √N and y ≤ pN/3, combining their squares, sorting to remove duplicates, and then
comparing to a list of primes, obtained from the sieve of Eratosthenes.
7. The Quartic Jacobi Symbol.
The results we have proved generalize to testing for fourth powers in finite fields. Since
the arguments are so similar, we will be brief. The ring of Gaussian integers is Z[i]; this
is a principal ideal domain with units ±1, ±i. Here, the norm coincides with the square of
Euclidean length: N (a + bi) = a2 + b2. Again, we have division with remainder: (2) still
holds, but with 3/4 replaced by 1/2 [13, p. 12]. Only p = 2 is ramified; primes p ≡ 1 mod
4 are split and primes p ≡ 1 mod 4 are inert.
When P 6= (1 + i) is a prime ideal, the quartic character is defined by the condition
P(cid:17) ≡ α(NP −1)/4 mod P ;
(cid:16) α
we define a Jacobi symbol for Z[i] just as before.
mod 4 [13, p. 121].
Finally, primary means ≡ 1 mod (1 + i)3. Then, a + bi is primary iff 2 b and a + b ≡ 1
The replacements for (3)–(6) are as follows. Let α = a + bi and β = c + di be primary.
Then
and
(17)
(cid:18) α
β(cid:19) = (−1)(a−1)(c−1)/4(cid:18) β
α(cid:19),
β (cid:19) = i(c−d−d2−1)/4,
(cid:18) 1 + i
(cid:18) i
β(cid:19) = i−(c−1)/2,
(cid:18)−1
β (cid:19) = (−1)(c−1)/2.
See Halter-Koch [12, §7.3-7.4] for proofs.
A quartic Jacobi symbol algorithm, entirely analogous to the cubic one in Section 3
(including the possibility that gcd(α, β) 6= 1) was given by Eisenstein in 1844 [9, §10]. For
compatibility with Section 3, we express the division step as
(20)
(19)
(18)
α = qβ + γ = qβ + (1 + i)minγ′′,
13
where N (γ) ≤ N (β)/2 and γ′′ = e′′ + f ′′i is primary. The algorithm of Section 3 can then
be modified as follows. Step 1 should return 1 on the condition that β = 1 or α = 1. Steps
4 and 5 are the same, but with 1 − ρ and ρ replaced by 1 + i and i, where 0 ≤ n < 4.
Finally, the recursion becomes
β(cid:19) = im(c−d−d2−1)/4−n(c−1)/2(−1)(e′′−1)(c−1)/4(cid:18) β
(cid:18) α
γ′′(cid:19).
As before, suppose that we compute q by computing α ¯β/N (β) exactly, and then
rounding coefficients. Consider the sequence defined by the recurrence relation ξn = 2(i +
1)ξn−1+ξn−2, with initial conditions ξ0 = 1, ξ1 = 5. The dominant root of its characteristic
.
= 2.2720 + 1.7862i. This makes the
sequence grow rapidly, and for large n, the coordinatewise nearest Gaussian integer to
ξn/ξn−1 is 2 + 2i. Therefore, arguing as we did in Section 4, Eisenstein's algorithm will
use Ω(nM (n)) bops in the worst case.
polynomial X 2 − 2(i + 1)X − 1 is (i + 1) + √2i + 1
As we did for the cubic algorithm, we can obtain a quadratic bit-complexity bound
with standard arithmetic by using Newton iteration to determine quotients for long divi-
sion. Our starting point for inverting β = c + di is
ξ0 =
c
2e −
d
2e i,
where e is defined from c and d just as in Section 5. As before, 2N (β) ≤ 2e ≤ 16N (β),
which gives 1/2 ≤ ǫ ≤ 15/16. Again, a quadratic bound on bops is best possible for
Eisenstein's quartic algorithm.
Finally, all of the conclusions in Section 6 hold true for testing quartic residues modulo
primes p ≡ 1 mod 4. In particular, we can reduce computing √−1 mod p to solving the
Diophantine equation p = x2 + y2 by observing that in such a representation, p divides
neither x nor y. So (y/x)2 ≡ −1
8. Generalizing the Even-Quotient Algorithm.
The method most commonly presented in textbooks for computing the Jacobi symbol
in Z could be called an even-to-odd remainder algorithm, because a maximal power of
2 is removed from the remainder before the recursive call. There is another algorithm,
published by Eisenstein in [7], in which the quotients are even. This algorithm uses division
steps of the form
α = qβ + γ,
in which q is even, γ < β, and γ is odd. Eisenstein called it a "simpler" algorithm, but it
is actually more complex, as far as worst-case behavior is concerned. In particular, when
applied to consecutive odd numbers, the number of division steps is exponential in bit
length [21, p. 608].
The even-quotient idea generalizes to both cubic and quartic symbols. The cubic
version seems not to have been discussed before, so we treat that first. Let N (β) > 1.
Then, if α is any other element of Z[ρ], we can find a q, divisible by 1 − ρ, for which
α = qβ + γ,
N (γ) < N (β).
14
To prove this, consider a nonzero ζ ∈ Q(ρ). We want to write ζ = q + ω with q ≡ 0
mod 1 − ρ and ω < 1. Form q in the usual way, by rounding coefficients of ζ to nearest
integers. This makes ω < 1. If q is not divisible by 1− ρ, consider choosing a sixth root of
unity ǫ and replacing q by q + ǫ. As we go around the unit circle, the sixth roots of unity,
reduced mod 1 − ρ, alternate between 1 and 2. Therefore, three of the choices for ǫ will
make q divisible by 1−ρ. Let the new remainder be ω′ = ω−ǫ. We now show that ω′ < 1
is still possible. If the angle θ = arg ω lies between 0 and π/3, we have ω − 1 < 1. (To see
this, observe that there are two congruent equilateral triangles, with vertices 0, 1, −ρ2 and
−1, 0, ρ.) By symmetry, we also have ω − (−ρ2) < 1. The same argument can be made
for each of the other five sectors. The desired result then follows upon taking ζ = α/β.
A cubic analog to Eisenstein's even-quotient algorithm can then be defined so as to
use (8), but with all qi ≡ 0 mod 1 − ρ, and all mi = 0. (That is, the division in Step 3 has
an additional constraint, and Step 4 is deleted from the algorithm.)
Just as for the algorithm in Z, the worst-case bit complexity is exponential. As far
as we know, this is not easy to discover by manipulating formulas. (Some fruitless effort
went into this quest.) Instead, we turned to the "tools of ignorance." After implementing
the algorithm in Maple, we ran it on all possible input coefficients in [0, 10]. Keeping
track of "record values" for the number of iterations revealed that for α = (3k + 2)ρ, and
β = 1 + (3k + 3)ρ, the number of iterations is 4k + 3. For these inputs, the bit length n is
∼ 2 log2 k, making the number of iterations about 2n/2.
iterations, the algorithm settled into a repeating cycle of four quotients:
To prove that our observation would always hold, we noted that after two "warmup"
−2 − ρ, −1 − 2ρ, 2 + ρ, −2 − ρ.
We then executed the algorithm symbolically, using Laurent series around k = 0 (that
is, expansion in powers of 1/k) to guarantee that rounding would behave correctly for all
sufficiently large k. Doing this verified that the observed number of iterations holds holds
for all but a finite number of k, from which the exponential bit complexity follows.
In 1859, Smith [22, p. 87] presented the analogous algorithm for computing quartic
Jacobi symbols. He asserted without proof that the corresponding long division (i.e., with
all quotients divisible by 1 − i) is possible. That can be proved with a similar argument;
it is only necessary to observe that all powers of i are 1 mod i + 1, and then consider
quarter-circular sectors with angles centered on multiples of π/2.
It seems to have escaped notice that the worst-case bit complexity of this algorithm is
also exponential. If we start with inputs 4m + 1, 4m − 3, there could be about m division
steps, since
4m + 1 = 2(4m − 3) − (4m − 7).
(Incidentally, this is also a bad example for Eisenstein's algorithm.) To be fair to Smith, he
did point out that Eisenstein's version of the quartic algorithm "terminates more rapidly"
than his did, but without making this quantitative.
Acknowledgments. The authors were supported by the National Science Foundation
(CCF-1420750).
15
References.
1. A. V. Aho, J. E. Hopcroft, and J. D. Ullman, The Design and Analysis of Computer
Algorithms, Addison-Wesley, 1974.
2. W. Bosma, Cubic reciprocity and explicit primality tests for h · 3k ± 1, in A. van
der Poorten, A. Stein, eds., High Primes and Misdemeanours: Lectures in Honour of
the 60th Birthday of Hugh Cowie Williams (Fields Institute Communications v. 41),
AMS, 2004.
3. G. E. Collins, A fast Euclidean algorithm for Gaussian integers, J. Symbolic Comput.
33, 385-392, 2002.
4. M. J. Collison, The origins of the cubic and biquadratic reciprocity laws, Arch. Hist.
Exact Sci., v. 17, pp. 63-69, 1977.
5. I. B. Damgard and G. S. Frandsen, Efficient algorithms for the gcd and cubic residu-
osity in the ring of Eisenstein integers, J. Symbol. Comput. 39, 643-652, 2005.
6. G. Eisenstein, Beweis des Reciprocitatssatzes fur die cubischen Reste in der Theorie
der aus dritten Wurzeln der Einheit zusammengesetzen complexen Zahlen, J. reine
angew. Math. 27, 289-310, 1844.
7. G. Eisenstein, Einfacher Algorithmus zur Bestimmung des Werthes von (cid:0) a
angew. Math. 27, 317-318, 1844.
b(cid:1), J. reine
8. G. Eisenstein, Nachtrag zum cubischen Reciprocitatssatze fur die aus dritten Wurzeln
der Einheit zusammengesetzen complexen Zahlen. Criterien des cubischen Characters
der Zahl 3 und ihrer Theiler, J. reine angew. Math. 28, 28-35, 1844.
9. G. Eisenstein, Einfacher Beweis un Verallgemeinerung des Fundamentaltheorems fur
die biquadratischen Reste, J. reine angew. Math. 28, 223-245, 1844.
10. G. Frei, The reciprocity law from Euler to Eisenstein, in S. Chikara, S. Mitsuo, and
J. W. Dauben, eds., The Intersection of History and Mathematics, Birkhauser, 1994.
11. C.F. Gauss, Theorematis fundamentalis in doctrina de residuis quadraticis demonstra-
tiones et ampliationes novae, Comment. Soc. Reg. Sci. Gottingensis 4, 1816-1818.
Presented to the society Feb. 10, 1817. [German translation: Neue Beweise und Er-
weiterungen des Fundamentalsatzes in der Lehre von den quadratischen Resten, in
C.F. Gauss, Untersuchen uber Hohere Arithmetik, ed. H. Maser, New York: Chelsea,
1965, pp. 496-510.]
12. F. Halter-Koch, Quadratic Irrationals: An Introduction to Classical Number Theory,
CRC Press, 2013.
13. K. Ireland and M. Rosen, A Classical Introduction to Modern Number Theory,
Springer-Verlag, 1982.
14. C. G. J. Jacobi, Vorlesungen uber Zahlentheorie (Wintersemester 1836/37,
Konigsberg), ed. F. Lemmermeyer and H. Pieper, Erwin Rauner Verlag, Augsburg,
2007.
15. C. G. J. Jacobi, Uber die Kriestheilung und ihre Anwendung auf die Zahlentheorie,
Bericht uber die zur Bekanntmachung geeigneten Verhandlungen der Konigl. Preuss.
Akademie der Wissenschaften zu Berlin, pp. 127-136, 1837. Reprinted with comments
and numerical tables in J. Reine angew. Math., v. 30, 166-182, 1846.
16. E. Kaltofen and H. Rolletschek, Computing greatest common divisors and factoriza-
tions in quadratic number fields, Math. Comp. 53, 697-720, 1989.
16
17. D. H. Lehmer, Guide to Tables in the Theory of Numbers, National Academy of
Sciences, Washington, 1941.
18. F. Lemmermeyer, Reciprocity Laws: From Euler to Eisenstein, Springer-Verlag, 2000,
19. S. Muller, On the computation of cube roots mod p, in A. van der Poorten, A. Stein,
eds., High Primes and Misdemeanours: Lectures in Honour of the 60th Birthday of
Hugh Cowie Williams (Fields Institute Communications v. 41), AMS, 2004.
20. R. Scheidler and H. C. Williams, A public-key cryptosystem utilizing cyclotomic fields,
Designs, Codes, and Cryptography 6, 117-131 (1995).
21. J. O. Shallit, On the worst case of three algorithms for computing the Jacobi symbol,
J. Symbol. Comput. 10, 593-610 (1990).
22. H. J. S. Smith, Report on the Theory of Numbers, Part I, Report of the British
Association for 1859, pp. 228-267. Parts I-VI reprinted as H. J. S. Smith, Report on
the Theory of Numbers, Chelsea, New York, 1965.
23. A. Weilert, Fast computation of the biquadratic residue symbol, J. Number Theory
96, 133-151, 1992.
24. D. Wikstrom, On the ℓ-ary GCD algorithm and computing residue symbols, KTH
(Stockholm) report TRITA NA 04-39, May 2004.
25. H. C. Williams, An M 3 public-key encryption scheme, Proc. CRYPTO 85, LNCS
218, pp. 358-368 (1986).
26. H. C. Williams and R. C. Holte, Computation of the solution of x3 + Dy3 = 1, Math.
Comp. 31, 778-785 (1977).
27. S. Winograd, Arithmetic Complexity of Computations, SIAM, 1980.
17
|
1409.2116 | 2 | 1409 | 2015-02-04T17:34:20 | Smart Sampling for Lightweight Verification of Markov Decision Processes | [
"cs.DS"
] | Markov decision processes (MDP) are useful to model optimisation problems in concurrent systems. To verify MDPs with efficient Monte Carlo techniques requires that their nondeterminism be resolved by a scheduler. Recent work has introduced the elements of lightweight techniques to sample directly from scheduler space, but finding optimal schedulers by simple sampling may be inefficient. Here we describe "smart" sampling algorithms that can make substantial improvements in performance. | cs.DS | cs | Smart Sampling for Lightweight Verification
of Markov Decision Processes
Pedro D'Argenio∗, Axel Legay†, Sean Sedwards† and Louis-Marie Traonouez†
∗Universidad Nacional de C´ordoba, Argentina, and †Inria Rennes -- Bretagne Atlantique, France
5
1
0
2
b
e
F
4
]
S
D
.
s
c
[
2
v
6
1
1
2
.
9
0
4
1
:
v
i
X
r
a
Abstract -- Markov decision processes (MDP) are useful to
model optimisation problems in concurrent systems. To verify
MDPs with efficient Monte Carlo techniques requires that their
nondeterminism be resolved by a scheduler. Recent work has
introduced the elements of lightweight techniques to sample
directly from scheduler space, but finding optimal schedulers by
simple sampling may be inefficient. Here we describe "smart"
sampling algorithms that can make substantial improvements in
performance.
I. INTRODUCTION
Markov decision processes describe systems that interleave
nondeterministic actions and probabilistic transitions. This
model has proved useful in many real optimisation problems
[27], [28], [29] and may be used to represent concurrent
probabilistic programs (see, e.g., [3], [1]). Such models com-
prise probabilistic subsystems whose transitions depend on
the states of the other subsystems, while the order in which
concurrently enabled transitions execute is nondeterministic.
This order may radically affect the behaviour of a system and
it is thus useful to calculate the upper and lower bounds of
quantitative aspects of performance.
As an example, consider the network of computational
nodes depicted in Fig. 1 (relating to the case study in Section
VI-D). Given that one of the nodes is infected by a virus,
we would like to calculate the probability that a target node
becomes infected. If we know the probability that the virus will
pass from one node to the next, we could model the system
as a discrete time Markov chain and analyse it to find the
probability that any particular node will become infected. Such
a model ignores the possibility that the virus might actually
choose which node to infect, e.g., to maximise its probability
of passing through the barrier layer. Under such circumstances
some nodes might be infected with near certainty or with only
very low probability, but this would not be adequately captured
by the Markov chain. By modelling the virus's choice of node
as a nondeterministic transition in an MDP, the maximum and
minimum probabilities of infection can be considered.
Fig. 2 shows a typical fragment of an MDP. Its execution
semantics are as follows. In a given state (s0), an action
(a1, a2, . . . ) is chosen nondeterministically to select a distri-
bution of probabilistic transitions (p1, p2, . . . or p3, p4, etc.).
A probabilistic choice is then made to select the next state
(s1, s2, s3, s4, . . . ). In this work we use the term scheduler to
refer to a particular way the nondeterminism in an MDP is
resolved.
Classic analysis of MDPs is concerned with finding the
target
barrier
layer
infected
s0
a1 a2
p1
s1
p2
s2
p3
s3
p4
s4
Figure 1: Model of network
virus infection.
Figure 2: Fragment of a
Markov decision process.
expected maximum or minimum reward for an execution of
the system, given individual rewards assigned to each of the
actions [2], [25]. Rewards may also be assigned to states or
transitions between states [16]. In this work we focus on MDPs
in the context of model checking concurrent probabilistic
systems, to find schedulers that maximise or minimise the
probability of a property. Model checking is an automatic
technique to verify that a system satisfies a property specified
in temporal logic [7]. Probabilistic model checking quantifies
the probability that a probabilistic system will satisfy a prop-
erty [9]. Numerical model checking algorithms to solve purely
probabilistic systems are costly in time and space. Finding
extremal probabilities in MDPs is generally more so, but is
nevertheless a polynomial function of the explicit description
of the MDP [3].
Statistical model checking (SMC) describes a collection
of Monte Carlo sampling techniques that make probabilistic
model checking more tractable by returning approximative
results with statistical confidence [31]. Recent approaches to
apply SMC to MDPs [4], [21], [11], [10], [22] are memory-
intensive and do not fully address the nondeterministic model
checking problem. Classic sampling approaches for MDPs,
such as the Kearns algorithm [14], address a related but
different problem.
In [22]
This work extends [22].
the authors provide
sampling techniques that can form the basis of memory-
efficient ("lightweight") verification of MDPs. The principal
contributions of [22] are (i) specifying the infinite behaviour
of history-dependent or memoryless schedulers using O(1)
memory, (ii) sampling directly and uniformly from scheduler
space, and (iii) quantifying the statistical confidence of mul-
tiple estimates or multiple hypothesis tests. As in the case
of standard SMC, sampling makes the verification problem
independent of the size of the space of samples, with a
convergence to the correct result almost surely guaranteed
with an infinite number of samples. The use of lightweight
techniques opens up the possibility to efficiently distribute the
problem on high performance massively parallel architectures,
such as general purpose computing on graphics processing
units (GPGPU).
Sampling schedulers makes a significant advance over mere
enumeration. For example, suppose half of all schedulers for
a given MDP and property are 'near optimal', i.e., have a
probability of satisfying the property that is deemed adequately
close to the true optimum. If all such near optimal schedulers
lie in the second half of the enumeration, it will be necessary
to enumerate half of all schedulers before finding one that
is near optimal. In contrast, one would expect to see a near
optimal scheduler after just two random selections, i.e., the
expectation with two samples is one. This phenomenon is
not limited to the case when schedulers are pathologically
distributed with respect to the enumeration. Since the total
number of schedulers increases exponentially with path length
(the number of history-dependent schedulers increases doubly
exponentially with states and path length), the total number
of schedulers is usually very large. Hence, even when near
optimal schedulers are more uniformly distributed with re-
spect to the enumeration, it is typically not tractable to use
enumeration to find one. Note that sampling also works with
non-denumerable spaces. The cost of finding a near optimal
scheduler with sampling is simply proportional to the relative
mass of near optimal schedulers in scheduler space. Our
experiments with standard case studies suggest that this cost
is typically reasonable.
It was demonstrated in [22] that simple undirected sampling
may be adequate for some case studies. In this work we present
"smart sampling" algorithms that make significantly better
use of a simulation budget. For a given number of candidate
schedulers, smart sampling can reduce the simulation cost of
extremal probability estimation by more than N/⌈2+log2 N⌉,
where N is the minimum number of simulations necessary to
achieve the required statistical confidence, as given by (1).
The basic notions of smart sampling were hinted at in [22].
Simply put, a small part of the budget is used to perform an
initial assessment of the problem and to generate an optimal
candidate set of schedulers. The remaining budget is used to
test and refine the candidate set: sub-optimal schedulers are
removed and their budget is re-allocated to good ones. Here
we give a full exposition of smart sampling and explain its
limitations. We have implemented the algorithms in our stat-
istical model checking platform, PLASMA1, and demonstrate
their successful application to a number of case studies from
the literature. We include some examples that are intractable
to numerical techniques and compare the performance of our
techniques with an alternative sampling approach [11]. We also
give an example where smart sampling is less effective, but
show that the results may nevertheless be useful in bounding
the range of extremal probabilities.
1project.inria.fr/plasma-lab/
Structure of the Paper
In Section II we introduce some basic concepts and notation
necessary for the sequel. In Section III we briefly survey
closely related work. In Sections IV we recall the basis of our
lightweight verification techniques. In Section V we describe
the notion of smart sampling and present our smart estimation
and smart hypothesis testing algorithms. In Section VI we
give the results of experiments with a number of case studies
from the literature. In Section VII we discuss the limitations
of smart sampling and in Section VIII we summarise the
challenges and prospects for our approach.
II. PRELIMINARIES
Given an MDP with set of actions A, having a set of states
S that induces a set of sequences of states Ω = S+, a history-
dependent (general) scheduler is a function S : Ω → A. A
memoryless scheduler is a function M : S → A. Intuitively,
at each state in the course of an execution, a general scheduler
(S) chooses an action based on the sequence of previous states
and a memoryless scheduler (M) chooses an action based only
on the current state. History-dependent schedulers therefore
include memoryless schedulers.
a1
p1
a2
p2
1
a0
1 − p1
= ¬ψ
s0
Fig. 3 illustrates a simple
MDP for which memoryless
and history-dependent sched-
ulers give different optima
for the bounded temporal lo-
gic property X(ψ ∧ XGt¬ψ)
when p1 6= p2 and t > 0.
The property makes use of the
temporal operators next (X)
and globally (G). Intuitively,
the property states that on the
next step ψ will be true and,
on the step after that, ¬ψ will
be remain true for t + 1 time
steps. The property is satisfied by the sequence of states
s0s1s0s0 ··· . If p1 > p2, the maximum probability for s0s1 is
achieved with action a2, while the maximum probability for
s0s0 is achieved with action a1. Given that both transitions
start in the same state, a memoryless scheduler will not achieve
the maximum probability achievable with a history-dependent
scheduler.
Figure 3: MDP with differ-
ent optima for general and
memoryless schedulers.
s1
= ψ
1 − p2
Statistical Model Checking with PLASMA
The algorithms we present here are implemented in our
SMC platform PLASMA (Platform for Learning and Advanced
Statistical Model checking Algorithms [5]). PLASMA is mod-
ular, allowing new modelling languages, logics and algorithms
to be plugged-in and take advantage of its graphical user
interface, its integrated development environment and its abil-
ity to correctly divide simulations on parallel computational
architectures. We introduce here the basic features of SMC
with PLASMA that are relevant to what follows.
SMC algorithms work by constructing an automaton to
accept only traces that satisfy a specified property. The state
space of the system is not constructed explicitly -- states are
generated on the fly during simulation -- hence SMC is efficient
for large, possibly infinite state, systems. Moreover, since the
simulations are required to be statistically independent, SMC
may be efficiently divided on parallel computing architectures.
The automaton may then be used to estimate the probability of
the property or to decide an hypothesis about the probability.
Typically, the probability p of property ϕ is estimated by
the proportion of traces that individually satisfy it, i.e., p ≈
N PN
1(ωi = ϕ), where ω1, . . . , ωN are N independently
generated simulation traces and 1(·) is an indicator function
that corresponds to the output of the automaton: it returns 1 if
the trace is accepted and 0 if it is not. As a statistical process,
the results of SMC are given with probabilistic confidence.
i=1
1
In the case of estimation, PLASMA calculates a priori the
required number of simulations according to a Chernoff bound
[24] that allows the user to specify an error ε and a probability
δ that the estimate p will not lie outside the true value ±ε.
Given that a system has true probability p of satisfying a
property, the Chernoff bound ensures P( p − p ≥ ε) ≤ δ.
Parameter δ is related to the number of simulations N by
δ = 2e−2N ε2 [24], giving
N = (cid:6)(ln 2 − ln δ)/(2ε2)(cid:7) .
(1)
In the case of hypothesis testing, PLASMA adopts the
sequential probability ratio test (SPRT) of Wald [26] to test
hypotheses of the form P(ω = ϕ) ⊲⊳ θ, where ⊲⊳∈ {≤,≥}
and θ is a user-specified probability threshold. The number
of simulations required to decide the test is typically fewer
than (1) but is dependent on how close θ is to the true
probability. The number is therefore not known in advance.
To evaluate P(ω = ϕ) ⊲⊳ θ, the SPRT constructs hypotheses
H 0 : P(ω = ϕ) ≥ p0 and H 1 : P(ω = ϕ) ≤ p1,
where p0 = θ + ε and p1 = θ − ε for some user-defined
interval specified by ε [26]. The SPRT also requires parameters
α and β to specify, respectively, the maximum acceptable
probabilities of incorrectly rejecting a true H0 and incorrectly
accepting a false H0. To choose between H0 and H1, the
SPRT defines the probability ratio
ratio =
n
Y
i=1
(p1)1(ωi=ϕ)(1 − p1)1(ωi6=ϕ)
(p0)1(ωi=ϕ)(1 − p0)1(ωi6=ϕ) ,
where n is the number of simulation traces ωi, i ∈ {1, . . . , n},
generated so far. The test proceeds by performing a simulation
and calculating ratio until one of two conditions is satisfied:
H1 is accepted if ratio ≥ (1 − β)/α and H0 is accepted if
ratio ≤ β/(1 − α).
Parallelisation of SMC is conceptually simple with light-
weight algorithms, but balancing the simulation load on unre-
liable or heterogeneous computing devices must be achieved
without introducing a "selection bias". The problem arises
because simulation traces that satisfy a property will,
in
general, take a different time to generate than those which
do not. If the SMC task is divided among a number of
clients of different speed or reliability, a naive balancing
approach will be biased in favour of results that are generated
quickly. To overcome this phenomenon, PLASMA adopts the
load balancing algorithm proposed in [30]. PLASMA's GUI
facilitates easy parallelisation on ad hoc networked computers
or on dedicated grids and clusters. The server application
(an instance of PLASMA) starts the job and waits to be
contacted by available clients (instances of PLASMA Service).
Our estimation experiments in Section VI were distributed on
the IGRIDA computing grid2.
III. RELATED WORK
The classic algorithms to solve MDPs are policy iteration
and value iteration [25]. Model checking algorithms for MDPs
may use value iteration applied to probabilities [1, Ch. 10]
or solve the same problem using linear programming [3]. All
consider history-dependent schedulers. The principal challenge
of finding optimal schedulers is what has been described as the
'curse of dimensionality' [2] and the 'state explosion problem'
[7]. In essence, these two terms refer to fact that the number
of states of a system increases exponentially with respect to
the number of interacting components and state variables.
This phenomenon has motivated the design of lightweight
sampling algorithms that find 'near optimal' schedulers to
optimise rewards in discounted MDPs, but the standard model
checking problem of finding extremal probabilities in non-
discounted MDPs is significantly more challenging. Since
nondeterministic and probabilistic choices are interleaved in an
MDP, schedulers are typically of the same order of complexity
as the system as a whole and may be infinite. As a result,
previous SMC algorithms for MDPs have considered only
memoryless schedulers or have other limitations.
The Kearns algorithm [14] is the classic 'sparse sampling
algorithm' for large, infinite horizon, discounted MDPs. It
constructs a 'near optimal' scheduler by approximating the
best action from a current state, using a stochastic depth-first
search. Importantly, optimality is with respect to discounted
rewards, not probability. The algorithm can work with large,
potentially infinite state MDPs because it explores a probabil-
istically bounded search space. This, however, is exponential
in the discount. To find the action with the greatest expected
reward in the current state of a trace, the algorithm recursively
estimates the rewards of successor states, up to some max-
imum depth defined by the discount and desired error. Actions
are enumerated while probabilistic choices are explored by
sampling, with the number of samples set as a parameter. The
discount guarantees that the algorithm eventually converges.
The stopping criterion is when successive estimates differ by
less than some error threshold. Since the actions of a state are
re-evaluated every time the state is visited (because actions are
history-dependent), the performance of the Kearns algorithm
is critically dependent on its parameters.
There have been several recent attempts to apply SMC
to nondeterministic models [4], [21], [11], [10], [22]. In
[4], [10] the authors present on-the-fly algorithms to remove
2igrida.gforge.inria.fr
'spurious' nondeterminism, so that standard SMC may be
used. This approach is limited to the class of models whose
nondeterminism does not affect the resulting probability of a
property. The algorithms therefore do not attempt to address
the standard MDP model checking problems related to finding
optimal schedulers.
In [21] the authors first find a memoryless scheduler that is
near optimal with respect to a reward scheme and discount,
using an adaptation of the Kearns algorithm. This induces a
Markov chain whose properties may be verified with standard
SMC. By storing and re-using the choices in visited states,
the algorithm improves on the performance of the Kearns
algorithm, but is thus limited to tractable memoryless sched-
ulers. The near optimality of the induced Markov chain is with
respect to rewards, not probability, hence [21] does not address
the standard model checking problems of MDPs.
In [11] the authors present an SMC algorithm to decide
whether there exists a memoryless scheduler for a given MDP,
such that
the probability of a property is above a given
threshold. The algorithm has an inner loop that generates
candidate schedulers by iteratively improving a probabilistic
scheduler, according to sample traces that satisfy the property.
The algorithm is limited to memoryless schedulers because the
improvement process learns by counting state-action pairs. The
outer loop tests the candidate scheduler against the hypothesis
using SMC and is iterated until an example is found or
sufficient attempts have been made. The approach has several
problems. The inner loop does not in general converge to the
true optimum (the number of state-actions does not actually
indicate scheduler probability), but is sometimes successful
because the outer loop randomly explores local maxima. This
makes the number of samples used by the inner loop critical:
too many may reduce the randomness of the outer loop's
exploration and thus significantly reduce the probability of
finding examples. A further problem is that
the repeated
hypothesis tests of the outer loop will eventually produce
erroneous results.
The present work builds on the elements of lightweight
verification for MDPs introduced in [22]. In [22] the authors
use an incremental hash function and a pseudo-random number
generator to define history-dependent schedulers using only
O(1) memory. This allows the schedulers to be selected at
random and tested individually, thus facilitating Monte Carlo
algorithms that are indifferent to the size of the sample space.
The full details of these techniques are described in Section
IV.
IV. LIGHTWEIGHT VERIFICATION OF MDPS
In this section we recall the elemental sampling techniques
of [22].
Storing schedulers as explicit mappings does not scale, so
we represent schedulers using uniform pseudo-random number
generators (PRNG) that are initialised by a seed and iterated
to generate the next pseudo-random value. In general, such
PRNGs aim to ensure that arbitrary subsets of sequences of
iterates are uniformly distributed and that consecutive iterates
are statistically independent. PRNGs are commonly used to
implement the uniform probabilistic scheduler, which chooses
actions uniformly at random and thus explores all possible
combinations of nondeterministic choices. Executing such
an implementation twice with the same seed will produce
identical traces. Executing the implementation with a different
seed will produce an unrelated set of choices:
individual
schedulers cannot be identified, so it is not possible to estimate
the probability of a property under a specific scheduler. We use
a PRNG to resolve nondeterministic choices, but not to make
those choices probabilistically. We use the PRNG to range over
the possible choices, such that repeated scheduler samplings
will eventually consider all possible sequences of actions. We
also rely on the fact that the seed of a PRNG uniquely defines
the sequence of pseudo-random values. Hence, in contrast
to the uniform probabilistic scheduler, actions are consistent
between simulations.
An apparently plausible solution is to use independent
PRNGs to resolve nondeterministic and probabilistic choices.
It is then possible to generate multiple probabilistic simulation
traces per scheduler by keeping the seed of the PRNG for
nondeterministic choices fixed while choosing random seeds
for a separate PRNG for probabilistic choices. Unfortunately,
the schedulers generated by this approach do not span the full
range of general or even memoryless schedulers. Since the
sequence of iterates from the PRNG used for nondeterministic
choices will be the same for all instantiations of the PRNG
used for probabilistic choices, the ith iterate of the PRNG for
nondeterministic choices will always be the same, regardless
of the state arrived at by the previous probabilistic choices.
The ith chosen action can be neither state nor trace dependent,
as required by memoryless (M) and history-dependent (S)
schedulers, respectively.
A. General Schedulers Using Hash Functions
We therefore construct a per-step PRNG seed that is a hash
of the integer identifying a specific scheduler concatenated
with an integer representing the sequence of states up to the
present.
is represented by a number of bits bi,
We assume that a state of an MDP is an assignment of
values to a vector of system variables vi, i ∈ {1, . . . , n}.
Each vi
typically
corresponding to a primitive data type (int, float, double, etc.).
The state can thus be represented by the concatenation of the
bits of the system variables, such that a sequence of states
may be represented by the concatenation of the bits of all the
states. Without loss of generality, we interpret such a sequence
of states as an integer of Pn
i=1 bi bits, denoted s, and refer
to this in general as the trace vector. A scheduler is denoted
by an integer σ, which is concatenated to s (denoted σ : s) to
uniquely identify a trace and a scheduler. Our approach is to
generate a hash code h = H(σ : s) and to use h as the seed
of a PRNG that resolves the next nondeterministic choice.
The hash function H thus maps σ : s to a seed that is
deterministically dependent on the trace and the scheduler. The
PRNG maps the seed to a value that is uniformly distributed
but nevertheless deterministically dependent on the trace and
the scheduler. In this way we approximate the schedulers
functions S and M described in Section II. Importantly,
the technique only relies on the standard properties of hash
functions and PRNGs. Algorithm 1 is the basic simulation
function used by our algorithms.
Algorithm 1: Simulate
Input:
M: an MDP with initial state s0
ϕ: a property
σ: an integer identifying a scheduler
Output:
ω: a simulation trace
1 Let Uprob,Unondet be uniform PRNGs with respective
samples rpr, rnd
2 Let H be a hash function
3 Let s denote a state, initialised s ← s0
4 Let ω denote a trace, initialised ω ← s
5 Let s be the trace vector, initially empty
6 Set seed of Uprob randomly
7 while ω = ϕ is not decided do
s ← s : s
Set seed of Unondet to H(σ : s)
Iterate Unondet to generate rnd and use to resolve
nondeterministic choice
Iterate Uprob to generate rpr and use to resolve
probabilistic choice
Set s to the next state
ω ← ω : s
8
9
10
11
12
13
B. An Efficient Iterative Hash Function
To implement our approach, we use an efficient hash func-
tion that constructs seeds incrementally. The function is based
on modular division [15, Ch. 6], such that h = (σ : s) mod m,
where m is a suitably large prime.
Since s is a concatenation of states,
is usually very
much larger than the maximum size of integers supported as
primitive data types. Hence, to generate h we use Horner's
method [12][15, Ch. 4]: we set h0 = σ and find h ≡ hn (n
as in Section IV-A) by iterating the recurrence relation
it
hi = (hi−12bi + vi) mod m.
(2)
The size of m defines the maximum number of different
hash codes. The precise value of m controls how the hash
codes are distributed. To avoid collisions, a simple heuristic
is that m should be a large prime not close to a power of
2 [8, Ch. 11]. The number of schedulers is typically much
larger than the number of possible hash codes, hence collisions
are theoretically inevitable. This means that not all possible
schedulers are realisable with a given hash function and
PRNG. Since our chosen hash function and PRNG are drawn
from respective families of hash functions and PRNGs that
potentially span all schedulers, the problem of collisions can
conceivably be addressed by also choosing the hash function
and PRNG at random. A scheduler would then be defined by
its label, its hash function and its PRNG. We do not implement
this idea here to avoid unnecessary complication and because
collisions are not the principal limitation. There are typically
many orders of magnitude more seeds than we can test,
hence the problem of finding the best available scheduler
supersedes the problem that the best available scheduler may
not be optimal. We anticipate that our proposed solutions to
accelerate convergence (property-focused scheduler space and
piecewise construction of schedulers) will effectively bypass
the collision problem.
In practical implementations it is an advantage to perform
calculations using primitive data types that are native to the
computational platform, so the sum in (2) should always
be less than or equal to the maximum permissible value.
To achieve this, given x, y, m ∈ N, we note the following
congruences:
(x + y) mod m ≡ (x mod m + y mod m) mod m (3)
(xy) mod m ≡ ((x mod m)(y mod m)) mod m(4)
The addition in (2) can thus be re-written in the form of (3),
such that each term has a maximum value of m − 1:
hi = ((hi−12bi ) mod m + (vi) mod m) mod m (5)
To prevent overflow, m must be no greater than half the
maximum possible integer. Re-writing the first term of (5) in
the form of (4), we see that before taking the modulus it will
have a maximum value of (m − 1)2, which will exceed the
maximum possible integer. To avoid this, we take advantage
of the fact that hi−1 is multiplied by a power of 2 and that m
has been chosen to prevent overflow with addition. We thus
apply the following recurrence relation:
(hi−12j) mod m = (hi−12j−1) mod m
+ (hi−12j−1) mod m
(6)
Equation (6) allows our hash function to be implemented
using efficient native arithmetic. Moreover, we infer from (2)
that to find the hash code corresponding to the current state
in a trace, we need only know the current state and the hash
code from the previous step. When considering memoryless
schedulers we need only know the current state.
C. Hypothesis Testing Multiple Schedulers
To decide whether there exists a scheduler such that P(ω =
ϕ) ⊲⊳ p, we apply the SPRT to multiple (randomly chosen)
schedulers. Since the probability of error with the SPRT
applied to multiple hypotheses is cumulative, we consider
the probability of no errors in any of M tests. Hence, in
order to ensure overall error probabilities α and β, we adopt
αM = 1 − M√1 − α and βM = 1 − M√1 − β in our stopping
conditions. H1 is accepted if ratio ≥ (1−βM )/αM and H0 is
accepted if ratio ≤ βM /(1 − αM ). Algorithm 2 demonstrates
the sequential hypothesis test for multiple schedulers. If the
algorithm finds an example, the hypothesis is true with at least
the specified confidence.
Algorithm 2: SPRT for multiple schedulers
Input:
M, ϕ: the MDP and property of interest
H ∈ {H0, H1}: the hypothesis with interval θ ± ε
α, β: the desired error probabilities of H
M : the maximum number of schedulers to test
Output: The result of the hypothesis test
1 Let p0 = θ + ε and p1 = θ − ε be the bounds of H
2 Let αM = 1 − M√1 − α and βM = 1 − M√1 − β
3 Let A = (1 − βM )/αM and B = βM /(1 − αM )
4 Let Useed be a uniform PRNG and σ be its sample
5 for i ∈ {1, . . . , M} while H is not accepted do
Iterate Useed to generate σi
Let ratio = 1
while ratio > A ∧ ratio < B do
ω ← Simulate(M, ϕ, σi)
ratio ← (p1)1(ω=ϕ)(1−p1)1(ω6=ϕ)
(p0)1(ω=ϕ)(1−p0)1(ω6=ϕ) ratio
if ratio ≤ A ∧ H = H0 ∨ ratio ≥ B ∧ H = H1 then
accept H
6
7
8
9
10
11
12
Algorithm 3: Estimation with multiple schedulers
Input:
M, ϕ: the MDP and property of interest
ε, δ: the required Chernoff bound
M : the number of schedulers to test
Output: Extremal estimates pmin and pmax
1 Let N = (cid:6)ln(2/(1 − M√1 − δ ))/(2ε2)(cid:7) be the no. of
simulations per scheduler
2 Let Useed be a uniform PRNG and σ its sample
3 Initialise pmin ← 1 and pmax ← 0
4 Set seed of Useed randomly
5 for i ∈ {1, . . . , M} do
Iterate Useed to generate σi
Let truecount = 0 be the initial number of traces
that satisfy ϕ
for j ∈ {1, . . . , N} do
6
7
ωj ← Simulate(M, ϕ, σi)
truecount ← truecount + 1(ωj = ϕ)
8
9
10
11
12
13
14
15
Let pi = truecount/N
if pmax < pi then
pmax = pi
if pi > 0 ∧ pmin > pi then
pmin = pi
D. Estimating Multiple Schedulers
16 if pmax = 0 then
17
No schedulers were found to satisfy ϕ
We consider the strategy of sampling M schedulers to es-
timate the optimum probability. We thus generate M estimates
{p1, . . . , pM}, corresponding to true values {p1, . . . , pM},
and take either the maximum (pmax) or minimum (pmin), as
required. To overcome the cumulative probability of error with
the standard Chernoff bound, we specify that all estimates pi
must be within ε of their respective true values pi, ensuring
that any pmin, pmax ∈ {p1, . . . , pM} are within ε of their true
value. Given that all estimates pi are statistically independent,
the probability that all estimates are less than their upper bound
is expressed by P(VM
)M .
Hence, P(WM
)M ,
giving N = (cid:6)− ln(cid:0)1 − M√1 − δ(cid:1) /(2ε2)(cid:7) for user-specified
parameters M , ε and δ. This ensures that P(pmin − pmin ≥
ε) ≤ δ and P(pmax − pmax ≥ ε) ≤ δ. To ensure the usual
stronger conditions that P( pmax − pmax ≥ ε) ≤ δ and
P( pmin − pmin ≥ ε) ≤ δ, we have
i=1 pi − pi ≤ ε) ≥ (1 − e−2N ε2
i=1 pi − pi ≥ ε) ≤ 1 − (1 − e−2N ε2
N = l(cid:16)ln 2 − ln(cid:16)1 − M√1 − δ(cid:17)(cid:17) /(2ε2)m .
(7)
N scales logarithmically with M , making it tractable to
consider many schedulers. In the case of M = 1, (7) degen-
erates to (1). Note, however, that the confidence expressed by
(7) is with respect to the sampled set, not with respect to the
true extrema. Algorithm 3 is the resulting extremal probability
estimation algorithm for multiple schedulers.
Figure 4 shows the empirical cumulative distribution of
schedulers generated by Algorithm 3 applied to the MDP of
y
t
i
l
i
b
a
b
o
r
P
e
v
i
t
l
a
u
m
u
C
1
8
.
0
6
.
0
4
.
0
2
.
0
0
0
memoryless
0.1
0.2
0.3
Probability
Figure 4: Empirical cumulative distribution of estimates from
Algorithm 3 applied to MDP of Fig. 3.
Fig. 3, using p1 = 0.9, p2 = 0.5, ϕ = X(ψ ∧ XG4¬ψ),
ε = 0.01, δ = 0.01 and M = 300. The vertical red and blue
lines mark the true probabilities of ϕ under each of the history-
dependent and memoryless schedulers, respectively. The grey
rectangles show the ±ε error bounds, relative to the true
probabilities. There are multiple estimates per scheduler, but
all estimates are within their respective confidence bounds.
V. SMART SAMPLING
The simple sampling strategies used by Algorithms 2 and
3 have the disadvantage that they allocate equal simulation
budget to all schedulers, regardless of their merit. In general,
›
›
the problem we address has two independent components:
the rarity of near optimal schedulers and the rarity of the
property under a near optimal scheduler. We should allocate
our simulation budget accordingly and not waste budget on
schedulers that are clearly not optimal.
Motivated by the above, our smart estimation algorithm
comprises three stages: (i) an initial undirected sampling
experiment to discover the nature of the problem, (ii) a targeted
sampling experiment to generate a candidate set of schedulers
with high probability of containing an optimal scheduler and
(iii) iterative refinement of the candidates to estimate the
probability of the best scheduler with specified confidence. By
excluding the schedulers with the worst estimated probabilities
and re-allocating their simulation budget to the schedulers
that remain, at each iterative step of stage (iii) the number
of schedulers reduces while the confidence of their estimates
increases. With a suitable choice of per-iteration budget, the
algorithm is guaranteed to terminate.
In the following subsection we develop the theoretical basis
of stage (ii).
A. Maximising the Probability of Seeing a Good Scheduler
In what follows we assume the existence of an MDP
and a property ϕ whose probability we wish to maxim-
ise by choosing a suitable scheduler from the set S. Let
P : S → [0, 1] be a function mapping schedulers to their
probability of satisfying ϕ and let pmax = maxσ∈S(P(s)).
For the sake of exposition, we consider the problem of finding
a scheduler that maximises the probability of satisfying ϕ
and define a "good" (near optimal) scheduler to be one in
the set Sg = {σ ∈ S P(σ) ≥ pmax − ε} for some
ε ∈ (0, pmax]. Intuitively, a good scheduler is one whose
probability of satisfying ϕ is within ε of pmax, noting that
we may similarly define a good scheduler to be one within ε
of pmin = minσ∈S(P(σ)), or to be in any other subset of S.
In particular, to address reward-based MDP optimisations, a
good scheduler could be defined to be the subset of S that is
near optimal with respect to a reward scheme. The notion of
a "best" scheduler follows intuitively from the definition of a
good scheduler.
If we sample uniformly from S, the probability of finding
a good scheduler is pg = Sg/S. The average probability
of a good scheduler is pg = Pσ∈Sg P(σ)/Sg. If we select
M schedulers uniformly at random and verify each with N
simulations, the expected number of traces that satisfy ϕ using
a good scheduler is thus M pgN pg. The probability of seeing a
trace that satisfies ϕ using a good scheduler is the cumulative
probability
(1 − (1 − pg)M )(1 − (1 − pg)N ).
(8)
Hence, for a given simulation budget Nmax = N M ,
to
implement stage (ii) the idea is to choose N and M to
maximise (8) and keep any scheduler that produces at least
one trace that satisfies ϕ. Since, a priori, we are generally
unaware of even the magnitudes of pg and pg, stage (i) is
necessarily uninformed and we set N = M = ⌈√Nmax⌉. The
results of stage (i) allow us to estimate pg and pg (see Fig.
9a) and thus maximise (8). This may be done numerically, but
we have found the heuristic N = ⌈1/pg⌉ to be near optimal
in all but extreme cases.
B. Smart Estimation
Algorithm 4 is our smart estimation algorithm to find
schedulers that maximise the probability of a property. The
algorithm to find minimising schedulers is similar. Lines 1
to 5 implement stage (i),
lines 6 to 10 implement stage
(ii) and lines 11 to 23 implement stage (iii). Note that the
algorithm distinguishes pmax (the notional
true maximum
probability), pmax (the true probability of the best candidate
scheduler found) and pmax (the estimated probability of the
best candidate scheduler).
The per-iteration simulation budget Nmax must be greater
than the number needed by the standard Chernoff bound (1),
to ensure that there will be sufficient simulations to guarantee
the specified confidence if the algorithm refines the candidate
set to a single scheduler. Typically, the per-iteration budget
will be greater than this, such that the required confidence
is reached before refining the set of schedulers to a single
element. Under these circumstances the confidence is judged
according to the Chernoff bound for multiple estimates (7). In
addition, lines 16 to 19 allow the algorithm to quit as soon as
the minimum number of simulations is reached.
Algorithm 4 may be further optimised by re-using the
simulation results from previous iterations of stage (iii). The
contribution is small, however, because confidence decreases
exponentially with the age (in terms of iterations) of the
results.
C. Smart Hypothesis Testing
We wish to test the hypothesis that there exists a scheduler
such that property ϕ has probability ⊲⊳ θ, where ⊲⊳∈ {≥,≤}.
Two advantages of sequential hypothesis testing are that it is
not necessary to estimate the actual probability to know if
an hypothesis is satisfied, and the easier the hypothesis is to
satisfy, the quicker it is to get a result. Algorithm 5 maintains
these advantages and uses smart sampling to improve on the
performance of Algorithm 2. For the purposes of exposition,
Algorithm 5 tests H0, as described in Section II. The algorithm
to test H1 is similar.
A sub-optimal approach would be to use Algorithm 4 to
refine a set of schedulers until one is found whose estimate sat-
isfies the hypothesis with confidence according to a Chernoff
bound. Our approach is to exploit the fact that the average
estimate at each iteration of Algorithm 4 is known with
high confidence, i.e., confidence given by the total simulation
budget. This follows directly from the result of [6], where the
bound is specified for a sum of arbitrary random variables, not
necessarily with identical expectations. By similar arguments
based on [26], it follows that the sequential probability ratio
test may also be applied to the sum of results produced during
the course of an iteration of Algorithm 4. Moreover, it is
possible to test each scheduler with respect to its individual
Algorithm 4: Smart Estimating
Input:
M: an MDP
ϕ: a property
ǫ, δ: the required Chernoff bound
Nmax > ln(2/δ)/(2ǫ2): the per-iteration budget
Output: pmax ≈ pmax, where pmax ≈ pmax and
P(pmax − pmax ≥ ǫ) ≤ δ
1 N ← ⌈√Nmax⌉; M ← ⌈√Nmax⌉
2 S ← {M seeds chosen uniformly at random}
3 ∀σ ∈ S,∀i ∈ {1, . . . , N} : ωσ
i ← Simulate(M, ϕ, σ)
4 R : S → N maps scheduler seeds to number of traces
satisfying ϕ:
R ← {(σ, n) σ ∈ S ∧ N ∋ n = PN
i = ϕ)}
1(ωσ
i=1
5 pmax ← maxσ∈S(R(σ)/N )
6 N ← ⌈1/pmax⌉, M ← ⌈Nmax pmax⌉
7 S ← {M seeds chosen uniformly at random}
8 ∀σ ∈ S,∀i ∈ {1, . . . , N} : ωσ
9 R ← {(σ, n) σ ∈ S ∧ N ∋ n = PN
1(ωσ
10 S ← {σ ∈ S R(σ) > 0}
11 ∀σ ∈ S, R(σ) ← 0; i ← 0; conf ← 1
12 while conf > δ ∧ S 6= ∅ do
i=1
i ← Simulate(M, ϕ, σ)
i = ϕ)}
13
14
15
16
17
18
19
20
21
22
23
i ← i + 1
Mi ← S
Ni ← 0
while conf > δ ∧ Ni < ⌈Nmax/Mi⌉ do
Ni ← Ni + 1
conf ← 1 − (1 − e−2ǫ2Ni)Mi
∀σ ∈ S : ωσ
Ni ← Simulate(M, ϕ, σ)
1(ωσ
R ← {(σ, n) σ ∈ S ∧ N ∋ n = PNi
pmax ← maxσ∈S(R(σ)/Ni)
R′ : {1, . . . ,S} → S is an injective function s.t.
∀(n, σ), (n′, σ′) ∈ R′, n > n′ =⇒ R(σ) ≥ R(σ′)
S ← {σ ∈ S σ = R′(n) ∧ n ∈ {⌊S/2⌋, . . . ,S}}
j = ϕ)}
j=1
results and the current number of schedulers, according to the
bound given in Section IV-C.
Hence, if the "average scheduler" or an individual scheduler
ever satisfies the hypothesis (lines 23 and 24), the algorithm
immediately terminates and reports that
the hypothesis is
satisfied with the specified confidence. If the "best" scheduler
ever individually falsifies the hypothesis (lines 25 and 26), the
algorithm also terminates and reports the result. Note that this
outcome does not imply that there is no scheduler that will
satisfy the hypothesis, only that no scheduler was found with
the given budget. Finally, if the algorithm refines the initial
set of schedulers to a single instance and the hypothesis was
neither satisfied nor falsified, an inconclusive result is reported
(line 29).
We implement one further important optimisation. We use
the threshold probability θ to directly define the simulation
budget to generate the candidate set of schedulers, i.e. N =
⌈1/θ⌉, M = ⌈θNmax⌉ (line 3). This is justified because we
need only find schedulers whose probability of satisfying ϕ
is greater than θ. By setting N = ⌈1/θ⌉, (8) ensures that
such schedulers, if they exist, have high probability of being
observed. The initial coarse exploration used in Algorithm 4
is thus not necessary.
Algorithm 5 is our smart hypothesis testing algorithm. Note
that we do not set a precise minimum per-iteration simulation
budget because we expect the hypothesis to be decided with
many fewer simulations than would be required to estimate
the probability. In practice it is expedient to initially set a low
per-iteration budget (e.g., 1000) and repeat the algorithm with
an increased budget (e.g., increased by an order of magnitude)
if the previous test was inconclusive.
VI. CASE STUDIES
To demonstrate the performance of smart sampling, we
have implemented Algorithms 4 and 5 in our statistical model
checking platform PLASMA [5]. We performed a number
of experiments on standard models taken from the numer-
ical model checking literature, most of which can be found
illustrated on the PRISM website3. We found that all of
our estimation experiments achieved their specified Chernoff
bounds (ε = δ = 0.01 in all cases) with a relatively modest
per-iteration simulation budget of 105 simulations. The actual
number of simulation cores used for the estimation results
was subject to availability and varied between experiments.
To facilitate comparisons, in what follows we normalise all
timings to be with with respect to 64 cores. Typically, each
data point was produced in a few tens of seconds. Our
hypothesis tests were performed on a single machine, without
distribution. Despite this, most experiments completed in just
a few seconds (some in fractions of a second), demonstrating
that our smart hypothesis testing algorithm is able to take
advantage of easy hypotheses.
A. IEEE 802.11 Wireless LAN Protocol
We consider a reachability property of the IEEE 802.11
Wireless LAN (WLAN) protocol model of [20]. The pro-
tocol aims to avoid "collisions" between devices sharing a
communication channel, by means of an exponential back-
off procedure when a collision is detected. We therefore
estimate the probability of the second collision at various
time steps, using Algorithm 4 with per-iteration budget of
105 simulations. Fig. 5 illustrates the estimated maximum
probabilities (pmax) and minimum probabilities (pmin) for time
steps k ∈ {0, 10, . . . , 100}. The property is expressed as
Fkcol = 2. The shaded areas indicate the true probabilities
±0.01,
the specified absolute error bound using Chernoff
bound ε = δ = 0.01. Our results are clearly very close to
the true values. Table I gives the results of hypothesis tests
based on the same model using property F100col = 2. See
Section VI-B for a description.
The results illustrated in Fig. 5 refer to the same property
and confidence as the those shown in Fig. 4 of [22]. The total
3www.prismmodelchecker.org/casestudies/
Algorithm 5: Smart Hypothesis Testing
Input:
M: an MDP
ϕ: a property
H0 : P(ω = ϕ) ≥ θ ± ε is the hypothesis
α, β: the desired error probabilities of H0
Nmax: the per-iteration simulation budget
Output: The result of the hypothesis test
1 Let p0 = θ + ε, p1 = θ − ε
2 Let A = (1 − β)/α, B = β/(1 − α)
3 N ← ⌈1/θ⌉; M ← ⌈θNmax⌉
4 S ← {M seeds chosen uniformly at random}
5 ∀σ ∈ S,∀i ∈ {1, . . . , N} : ωσ
6 R ← {(σ, n) σ ∈ S ∧ N ∋ n = PN
1(ωσ
7 if (p1)P R(σ)(1−p1)Nmax−P R(σ)
8
(p0)P R(σ)(1−p0)Nmax−P R(σ) ≤ A then
Accept H0 and quit
i ← Simulate(M, ϕ, σ)
i = ϕ)}
i=1
9 S ← {σ ∈ S R(σ) > 0}, M ← S + 1
10 while M > 1 do
11
M ← S
Let αM = 1 − M√1 − α, βM = 1 − M√1 − β
Let AM = (1 − βM )/αM , BM = βM /(1 − αM )
Let ratio = 1
for σi ∈ S, i ∈ {1, . . . , M} do
Let ratio i = 1
for j ∈ {1, . . . , N} do
ω ← Simulate(M, ϕ, σi)
if ω = ϕ then
ratio ← p1
p0
ratio; ratio i ← p1
p0
ratio i
else
ratio i
ratio ← 1−p1
1−p0
ratio; ratio i ← 1−p1
1−p0
Accept H0 and quit
if ratio ≤ A ∨ ratio i ≤ AM then
if ratio i ≥ BM then
Reject H0 (given budget) and quit
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
R′ : {1, . . . ,S} → S is an injective function s.t.
∀(n, σ), (n′, σ′) ∈ R′, n > n′ =⇒ R(σ) ≥ R(σ′)
S ← {σ ∈ S σ = R′(n) ∧ n ∈ {⌊S/2⌋, . . . ,S}}
28
29 Inconclusive result (given budget)
simulation cost to generate a point in Fig. 5 is 1.2 × 106
(12 iterations of 105 simulations using smart sampling),
compared to a cost of 2.7 × 108 per point in Fig. 4 of
[22] (4000 schedulers tested with 67937 simulations using
simple sampling). This demonstrates a more than 200-fold
improvement in performance.
B. IEEE 802.3 CSMA/CD Protocol
The IEEE 802.3 CSMA/CD protocol is a wired network
protocol that is similar in operation to that of IEEE 802.11,
but using collision detection instead of collision avoidance. In
n
o
s
i
i
l
l
o
c
f
o
y
t
i
l
i
b
a
b
o
r
P
2
.
0
5
1
.
0
1
.
0
5
0
.
0
0
p^
p^
max
min
0
20
40
60
80
100
Time steps (k)
Figure 5: Estimated maximum and minimum probabilities of
second collision in WLAN protocol. Shaded regions denote
true values ±0.01.
CSMA 3 4
CSMA 3 6
CSMA 4 4
WLAN 5
WLAN 6
θ
time
θ
time
θ
time
θ
time
time
0.5
0.5
0.3
1.3
0.5
0.2
0.1
0.8
1.3
0.8
3.5
0.4
5.2
0.7
0.3
0.15
2.6
2.2
0.85
737
0.45
79
0.8
4.0
0.18
*
*
0.86
*
0.48
*
0.9
8.6
0.2
2.9
6.5
0.9
2.9
0.5
39
0.93
*
0.25
2.9
1.3
0.95
2.5
0.8
2.6
0.95
3.8
0.5
1.3
1.3
Table I: Hypothesis test results for CSMA/CD and WLAN
protocols. θ is the threshold probability or the true probability
(marked by asterisk). time is simulation time in seconds to
achieve the correct result on a single machine.
Table I we give the results of applying Algorithm 5 to the
IEEE 802.3 CSMA/CD protocol model of [18]. The models
and parameters are chosen to compare with results given in
Table III in [11], hence we also give results for hypothesis
tests performed on the WLAN model used in Section VI-A.
In contrast to the results of [11], our results are produced on a
single machine, with no parallelisation. There are insufficient
details given about the experimental conditions in [11] to make
a formal comparison (e.g., error probabilities of the hypothesis
tests and number of simulation cores), but it seems that the
performance of our algorithm is generally much better. We set
α = β = δ = 0.01, which constitute a fairly tight bound, and
note that, as expected, the simulation times tend to increase
as the threshold θ approaches the true probability.
C. Choice Coordination
To demonstrate the scalability of our approach, we consider
the choice coordination model of [23] and estimate the min-
imum probability that a group of six tourists will meet within
T steps. The model has a parameter (BOUND) that limits the
state space. We set BOUND = 100, making the state space
of ≈ 5 × 1016 states intractable to numerical model checking.
Fortunately, it is possible to infer the correct probabilities from
tractable parametrisations. For T = 20 and T = 25 the true
minimum probabilities are respectively 0.5 and 0.75. Using
smart sampling and a Chernoff bound of ε = δ = 0.01, we
correctly estimate the probabilities to be 0.496 and 0.745 in a
few tens of seconds on 64 simulation cores.
Infection probability
1
0.8
0.6
0.4
0.2
Error
-0.008
-0.006
-0.004
-0.002
0
0
0.9
0.92
0.94
0.96
0.98
Detection probability
200
150
100
Time steps
50
1
0
0.002
0.9
0.92
0.94
0.96
0.98
Detection probability
(a) Estimates.
n
o
i
t
c
e
n
n
o
c
f
o
y
t
i
l
i
b
a
b
o
r
P
0
.
1
8
.
0
6
.
0
4
.
0
2
.
0
0
.
0
p^
p^
max
min
200
150
100
Time steps
50
1
0
(b) Errors.
Figure 6: Minimum probability of network infection.
0
50
100
150
200
Time steps (T)
Infection probability
Error
1
0.8
0.6
0.4
0.2
0
0.9
0.92
0.94
0.96
0.98
Detection probability
0.01
0.008
0.006
0.004
0.002
0
-0.002
0.9
0.92
0.94
0.96
0.98
Detection probability
200
150
100
Time steps
50
1
0
(a) Estimates.
200
150
100
Time steps
50
1
0
(b) Errors.
Figure 7: Maximum probability of network infection.
D. Network Virus Infection
Network virus infection is a subject of increasing relevance.
Hence, using a per-iteration budget of 105 simulations, we
demonstrate the performance of Algorithm 4 on the PRISM
virus infection case study based on [19]. The network is
illustrated in Fig. 1 and comprises three sets of linked nodes:
a set of nodes containing one infected by a virus, a set of
nodes with no infected nodes and a set of barrier nodes which
divides the first two sets. A virus chooses which node to infect
nondeterministically. A node detects a virus probabilistically
and we vary this probability as a parameter for barrier nodes.
We consider time as a second parameter. Figs. 6 and 7
illustrate the estimated probabilities that the target node in the
uninfected set will be infected. We observe in Figs. 6b and 7b
that the estimated minimums are within [−0.0070, +0.00012]
and the estimated maximums are within [−0.00012, +0.0083]
of their true values. The respective negative and positive
biases to these error ranges reflects the fact that Algorithm
4 converges from respectively below and above (as illustrated
in Fig. 9b). The average time to generate a point in Fig. 6 was
approximately 100 seconds, using 64 simulation cores. Points
in Fig. 7 took on average approximately 70 seconds.
E. Gossip Protocol
Gossip protocols are an important class of network al-
gorithms that rely on local connectivity to propagate informa-
tion globally. Using the gossip protocol model of [17], we used
Algorithm 4 with per-simulation budget of 105 simulations to
estimate the maximum (pmax) and minimum (pmin) probabil-
ities that the maximum path length between any two nodes is
less than 4 after T time steps. This is expressed by property
FT max path len < 4. The results are illustrated in Fig. 8.
Figure 8: Estimated probabilities that maximum path length is
< 4 in gossip protocol model. Shaded regions denote ±0.01
of true values.
Estimates of maximum probabilities are within [−0, +0.0095]
of the true values. Estimates of minimum probabilities are
within [−0.007, +0] of the true values. Each point in the figure
took on average approximately 60 seconds to generate using
64 simulation cores.
VII. CONVERGENCE AND COUNTEREXAMPLES
The techniques described in the preceding sections open up
the possibility of efficient lightweight verification of MDPs,
with the consequent possibility to take full advantage of
parallel computational architectures, such as multi-core pro-
cessors, clusters, grids, clouds and general purpose computing
on graphics processors (GPGPU). These architectures may
potentially divide the problem by the number of available
computational devices (i.e., linearly), however this must be
considered in the context of scheduler space increasing expo-
nentially with path length. Although Monte Carlo techniques
are essentially impervious to the size of the state space (they
also work with non-denumerable space), it is easy to construct
verification problems for which there is a unique optimal
scheduler. Such examples do not necessarily invalidate the
approach, however, because it may not be necessary to find
the possibly unique optimal scheduler to return a result with
a level of statistical confidence. The nature of the distribution
of schedulers nevertheless affects efficiency, so in this section
we explore the convergence properties of smart sampling and
give an example from the literature that does not converge as
well as the case studies in Section VI.
Essentially, the problem is that of exponentially distributed
schedulers, i.e., having a very low mass of schedulers close
to the optimum. Fig. 10 illustrates the difference between
exponentially decreasing and linearly decreasing distributions
with the same overall mass. In both cases pmax ≈ 0.2 (the
density at 0.2 is zero), but the figure shows that there is more
probability mass near 0.2 in the case of the linear distribution.
Figure 9 illustrates the convergence of Algorithm 4, using a
per-iteration budget of 106 applied to schedulers whose prob-
ability of success (i.e., of satisfying a hypothetical property) is
distributed according to the exponential distribution of Fig. 10.
Fig. 9a shows how the initial undirected sampling (black dots)
crudely approximates pmax. This approximation is then used
y
t
i
l
i
b
a
b
o
r
p
e
v
i
t
a
u
m
u
C
l
1
8
.
0
6
.
0
4
.
0
2
.
0
0
y
t
i
l
i
b
a
b
o
r
P
3
.
0
2
.
0
1
.
0
0
p^
max
s max
p^
mean
0.1
0.2
0
1
2
3
4
5
6
Scheduler probability
Iteration
(a) Scheduler distributions. Dots de-
note the results of initial sampling.
The red line is the set of schedulers
to refine. Black lines show the result
of subsequent refinements.
(b) Estimates and schedulers. At each
iterative step: pmax is the maximum
estimate, pmean is the mean estimate
and σmax is the true maximum prob-
ability of the available schedulers.
Figure 9: Convergence of Algorithm 4 with exponentially
distributed scheduler probabilities (Fig. 10) and per-iteration
budget of 106 simulations.
y
t
i
s
n
e
D
1
8
.
0
6
0
.
4
0
.
2
0
.
0
total mass » 0.0144
y
t
i
l
i
b
a
b
o
r
p
m
u
m
n
M
i
i
15 processes
20 processes
1
8
0
.
6
.
0
4
.
0
2
.
0
0
0.1
0.2
0
100
200
300
400
500
Scheduler probability
Path length (k)
Figure 10: Theoretical
lin-
ear
(blue) and exponential
(red) scheduler densities with
probability mass ≈ 0.0144
and zero density at probab-
ility 0.2.
to
Figure 11: Performance of
smart sampling (black dots)
applied
self-stabilising
models of [13]. Red shaded
areas
values
±0.01.
denote
true
to generate the candidate set of schedulers (red distribution).
The black lines illustrate five iterations of refinement, resulting
in a shift of the distribution towards pmax. Fig. 9b illustrates
the same shift in terms of the convergence of probabilities.
Iteration 0 corresponds to the undirected sampling. Iteration 1
corresponds to the generation of the candidate set of sched-
ulers. Note that for these first two iterations, pmean includes
schedulers that have zero probability of success. The expected
value of pmean in these two iterations is equal to the expected
probability obtained by the uniform probabilistic scheduler.
This fact can be used to verify that the hash function and
PRNG described in Section IV sample uniformly. In sub-
sequent iterations the candidates all have non-zero probability
of success. Importantly, the figure demonstrates that there is a
significant increase in the maximum probability of scheduler
success (σmax) between iteration 0 and iteration 1, and that
this maximum is maintained throughout the subsequent refine-
ments. Despite the apparently very low density of schedulers
near pmax, Algorithm 4 is able to make a good approximation.
The theoretical performance demonstrated in Fig. 9 explains
why we are able to achieve good results in Section VI. It
is nevertheless possible to find examples for which accurate
results are difficult to achieve. Fig. 11 illustrates the results
of applying Algorithm 4 to instances of the self-stabilising al-
gorithm of [13], using a per-iteration budget of 105. Although
the estimates (black dots) do not lie within our statistical
confidence bounds of the true values (red shaded areas), we
nevertheless claim that the results are useful. The problem of
quantifying the confidence of estimates with respect to optimal
values remains open, however.
To improve the performance of smart sampling, it is possible
to make an even better allocation of simulation budget. For
example, if good schedulers are very rare it may be beneficial
to increase the per-iteration budget (thus increasing the pos-
sibility of seeing a good scheduler in the initial candidate set)
but increase the proportion of schedulers rejected after each
iteration (thus reducing the overall number of iterations and
maintaining a fixed total number of simulations). To avoid
rejecting good schedulers under such a regime, it may be
necessary to reject fewer schedulers in the early iterations
when confidence is low.
VIII. PROSPECTS AND CHALLENGES
The use of sampling facilitates algorithms that scale inde-
pendently of the sample space, hence we anticipate that it
will be possible to apply our techniques to nondeterministic
models with non-denumerable schedulers. We believe it is
immediately possible to apply smart sampling to reward-based
MDP optimisation problems.
The success of sampling depends on the relative abundance
of near optimal schedulers in scheduler space and our experi-
ments suggest that these are not rare in standard case studies.
While it is possible to construct pathological examples, where
near optimal schedulers cannot easily be found by sampling,
it is perhaps even simpler to confound numerical techniques
with state explosion (three independent counters ranging over
0 to 1000 is typically sufficient with current hardware). Hence,
as with numerical model checking, our ongoing challenge is
essentially to increase performance and increase the number of
models and problems that may be efficiently addressed. Smart
sampling has made significant
improvements over simple
sampling, but we recognise that it will be necessary to develop
other techniques to accelerate convergence. We anticipate that
the most fruitful approaches will be (i) to reduce the sampled
scheduler space to only those that satisfy the property and (ii)
to construct schedulers piecewise. Such techniques will also
reduce the potential of hash function collisions.
An important remaining challenge is to quantify the con-
fidence of our estimates and hypothesis tests with respect to
optimality. In the case of hypothesis tests that satisfy the hypo-
thesis, the statistical confidence of the result is sufficient. If an
hypothesis is not satisfied, however, the statistical confidence
does not relate to whether there exists a scheduler to satisfy
it. Likewise, the statistical confidence bounds of probability
estimates imply nothing about how close they are to the
true optima. We nevertheless know that our estimates of the
extrema must lie within the true extrema or exceed them with
the specified statistical confidence. This is already useful and
[19] Marta Kwiatkowska, Gethin Norman, David Parker, and Maria Grazia
Vigliotti. Probabilistic mobile ambients. Theoretical Computer Science,
410(12-13):1272 -- 1303, 2009.
[20] Marta Z. Kwiatkowska, Gethin Norman, and Jeremy Sproston. Probab-
ilistic model checking of the IEEE 802.11 wireless local area network
protocol. In Proc. 2nd Joint International Workshop on Process Algebra
and Probabilistic Methods, Performance Modeling and Verification,
pages 169 -- 187. Springer, 2002.
[21] Richard Lassaigne and Sylvain Peyronnet. Approximate planning and
verification for large Markov decision processes. In Proc. 27th Annual
ACM Symposium on Applied Computing, pages 1314 -- 1319. ACM, 2012.
[22] A. Legay, S. Sedwards, and L.-M. Traonouez. Scalable verification of
Markov decision processes. In 4th Workshop on Formal Methods in the
Development of Software (FMDS 2014), LNCS. Springer, 2014.
[23] U. Ndukwu and A. McIver. An expectation transformer approach to pre-
dicate abstraction and data independence for probabilistic programs. In
Proc. 8th Workshop on Quantitative Aspects of Programming Languages
(QAPL'10), 2010.
[24] Masashi Okamoto. Some inequalities relating to the partial sum of
binomial probabilities. Annals of the Institute of Statistical Mathematics,
10(1):29 -- 35, 1958.
[25] Martin L. Puterman. Markov Decision Processes: Discrete Stochastic
Dynamic Programming. Wiley-Interscience, 1994.
[26] Abraham Wald. Sequential tests of statistical hypotheses. The Annals
of Mathematical Statistics, 16(2):117 -- 186, 1945.
[27] Douglas J. White. Real applications of Markov decision processes.
Interfaces, 15(6):73 -- 83, 1985.
[28] Douglas J. White. Further real applications of Markov decision pro-
cesses. Interfaces, 18(5):55 -- 61, 1988.
[29] Douglas J. White. A survey of applications of Markov decision
processes. Journal of the Operational Research Society, 44(11):1073 --
1096, Nov. 1993.
[30] H. L. S. Younes. Verification and Planning for Stochastic Processes with
Asynchronous Events. PhD thesis, Carnegie Mellon University, 2005.
[31] Hakan L. S. Younes and Reid G. Simmons. Probabilistic verification of
discrete event systems using acceptance sampling. In Computer Aided
Verification, pages 223 -- 235. Springer, 2002.
a significant improvement over the results produced using the
uniform probabilistic scheduler. In addition, given the number
of simulations performed, we may at least quantify confidence
with respect to the product pgpg (the rarity of near optimal
schedulers times the average probability of the property with
near optimal schedulers).
ACKNOWLEDGEMENTS
We are grateful to Benoıt Delahaye for useful prior dis-
cussions. This work was partially supported by the European
Union Seventh Framework Programme under grant agreement
no. 295261 (MEALS).
REFERENCES
[1] Christel Baier and Joost-Pieter Katoen. Principles of model checking.
MIT Press, 2008.
[2] Richard Bellman. Dynamic Programming. Princeton University Press,
1957.
[3] Andrea Bianco and Luca De Alfaro. Model checking of probabilistic
and nondeterministic systems. In Foundations of Software Technology
and Theoretical Computer Science, pages 499 -- 513. Springer, 1995.
[4] Jonathan Bogdoll, Luis Mar´ıa Ferrer Fioriti, Arnd Hartmanns, and
Holger Hermanns. Partial order methods for statistical model checking
and simulation.
In Formal Techniques for Distributed Systems, pages
59 -- 74. Springer, 2011.
[5] Benoıt Boyer, Kevin Corre, Axel Legay, and Sean Sedwards. PLASMA-
lab: A flexible, distributable statistical model checking library. In Kaus-
tubh Joshi, Markus Siegle, Marielle Stoelinga, and PedroR. D'Argenio,
editors, Quantitative Evaluation of Systems, volume 8054 of LNCS,
pages 160 -- 164. Springer, 2013.
[6] Herman Chernoff. A measure of asymptotic efficiency for tests of
a hypothesis based on the sum of observations. Ann. Math. Statist.,
23(4):493 -- 507, 1952.
[7] E.M. Clarke, E. A. Emerson, and J. Sifakis. Model checking: algorithmic
verification and debugging. Commun. ACM, 52(11):74 -- 84, November
2009.
[8] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford
Stein. Introduction to Algorithms. MIT Press, 3rd edition, 2009.
[9] Hans Hansson and Bengt Jonsson. A logic for reasoning about time and
reliability. Formal aspects of computing, 6(5):512 -- 535, 1994.
[10] Arnd Hartmanns and Mark Timmer. On-the-fly confluence detection for
statistical model checking. In NASA Formal Methods, pages 337 -- 351.
Springer, 2013.
[11] David Henriques, Joao G. Martins, Paolo Zuliani, Andr´e Platzer, and
Edmund M. Clarke. Statistical model checking for Markov decision
processes. In Quantitative Evaluation of Systems, 2012 Ninth Interna-
tional Conference on, pages 84 -- 93. IEEE, 2012.
[12] William George Horner. A new method of solving numerical equations
of all orders, by continuous approximation. Philosophical Transactions
of the Royal Society of London, 109:308 -- 335, 1819.
[13] A. Israeli and M. Jalfon. Token management schemes and random
walks yield self-stabilizating mutual exclusion.
In Proc. 9th Annual
ACM Symposium on Principles of Distributed Computing (PODC '90),
pages 119 -- 131. ACM New York, 1990.
[14] Michael Kearns, Yishay Mansour, and Andrew Y. Ng. A sparse sampling
algorithm for near-optimal planning in large Markov decision processes.
Machine Learning, 49(2-3):193 -- 208, 2002.
[15] Donald E. Knuth. The Art of Computer Programming. Addison-Wesley,
3rd edition, 1998.
[16] M. Kwiatkowska, G. Norman, and D. Parker. Stochastic model checking.
In M. Bernardo and J. Hillston, editors, Formal Methods for the Design
of Computer, Communication and Software Systems: Performance Eval-
uation (SFM'07), volume 4486 of LNCS (Tutorial Volume), pages 220 --
270. Springer, 2007.
[17] Marta Kwiatkowska, Gethin Norman, and David Parker. Analysis of a
gossip protocol in PRISM. SIGMETRICS Perform. Eval. Rev., 36(3):17 --
22, November 2008.
[18] Marta Kwiatkowska, Gethin Norman, David Parker, and Jeremy Spro-
ston. Performance analysis of probabilistic timed automata using digital
clocks. Formal Methods in System Design, 29:33 -- 78, August 2006.
|
1209.4971 | 1 | 1209 | 2012-09-22T08:37:17 | Streaming Complexity of Checking Priority Queues | [
"cs.DS",
"cs.CC"
] | This work is in the line of designing efficient checkers for testing the reliability of some massive data structures. Given a sequential access to the insert/extract operations on such a structure, one would like to decide, a posteriori only, if it corresponds to the evolution of a reliable structure. In a context of massive data, one would like to minimize both the amount of reliable memory of the checker and the number of passes on the sequence of operations. Chu, Kannan and McGregor initiated the study of checking priority queues in this setting. They showed that use of timestamps allows to check a priority queue with a single pass and memory space O(N^(1/2)), up to a polylogarithmic factor. Later, Chakrabarti, Cormode, Kondapally and McGregor removed the use of timestamps, and proved that more passes do not help. We show that, even in the presence of timestamps, more passes do not help, solving a previously open problem. On the other hand, we show that a second pass, but in reverse direction, shrinks the memory space to O((log N)^2), extending a phenomenon the first time observed by Magniez, Mathieu and Nayak for checking well-parenthesized expressions. | cs.DS | cs |
Streaming Complexity of Checking Priority Queues∗
Nathanael Fran¸cois1 and Fr´ed´eric Magniez2
1Univ Paris Diderot, Sorbonne Paris-Cit´e, LIAFA, CNRS, 75205 Paris, France,
[email protected]
2CNRS, LIAFA, Univ Paris Diderot, Sorbonne Paris-Cit´e, 75205 Paris, France,
[email protected]
Abstract
This work is in the line of designing efficient checkers for testing the reliability of some
massive data structures. Given a sequential access to the insert/extract operations on such a
structure, one would like to decide, a posteriori only, if it corresponds to the evolution of a
reliable structure. In a context of massive data, one would like to minimize both the amount of
reliable memory of the checker and the number of passes on the sequence of operations.
Chu, Kannan and McGregor [9] initiated the study of checking priority queues in this setting.
They showed that the use of timestamps allows to check a priority queue with a single pass and
memory space O(√N ). Later, Chakrabarti, Cormode, Kondapally and McGregor [7] removed
the use of timestamps, and proved that more passes do not help.
We show that, even in the presence of timestamps, more passes do not help, solving an open
problem of [9, 7]. On the other hand, we show that a second pass, but in reverse direction,
shrinks the memory space to O((log N )2), extending a phenomenon the first time observed by
Magniez, Mathieu and Nayak [15] for checking well-parenthesized expressions.
1
Introduction
The reliability of memory is central and becomes challenging when it is massive. In the context
of program checking [4] this problem has been addressed by Blum, Evans, Gemmell, Kannan and
Naor [3]. They designed on-line checkers that use a small amount of reliable memory to test the
behavior of some data structures. Checkers are allowed to be randomized and to err with small
error probability. In that case the error probability is not over the inputs but over the random
coins of the algorithm.
Chu, Kannan and McGregor [9] revisited this problem for priority queue data structures, where
the checker only has to detect an error after processing an entire sequence of data accesses. This can
be rephrased as a one-pass streaming recognition problem. Streaming algorithms sequentially scan
the whole input piece by piece in one sequential pass, or in a small number of passes, while using
sublinear memory space. In our context, the stream is defined by the sequence of insertions and
extractions on the priority queue. Using a streaming algorithm, the objective is then to decide if
the stream corresponds to a correct implementation of a priority queue. We also consider collection
data structures that implement multisets.
∗Supported by the French ANR Defis program under contract ANR-08-EMER-012 (QRAC project)
1
Definition 1 (Collection,PQ). Let Σ0 be some alphabet. Let Σ = {ins(a), ext(a) : a ∈ Σ0}.
For w ∈ ΣN , define inductively multisets Mi by M0 = ∅, Mi = Mi−1 \ {a} if w[i] = ext(a), and
Mi = Mi−1 ∪ {a} if w[i] = ins(a).
Then w ∈ Collection(Σ0) if and only if Mn = ∅ and a ∈ Mi−1 when w[i] = ext(a), for
i = 1, . . . , N . Moreover, w ∈ PQ(U ), for U ∈ N, if and only if w ∈ Collection({0, 1, . . . , U})
and a = max(Mi−1) when w[i] = ext(a), for i = 1, . . . , N .
Streaming algorithms were initially designed with a single pass: when a piece of the stream has
been read, it is gone for ever. This makes those algorithms of practical interest for online context,
such as network monitoring, for which first streaming algorithms were developed [1]. Motivated
by the explosion in the size of the data that algorithms are called upon to process in everyday
real-time applications, the area of streaming algorithms has experienced tremendous growth over
the last decade in many applications. In particular, a streaming algorithm can model an external
read-only memory. Examples of such applications occur in bioinformatics for genome decoding, or
in Web databases for the search of documents. In that context, considering multi-pass streaming
algorithm is relevant.
Using standard arguments one can establish that every p-pass randomized streaming algorithm
needs memory space Ω(N/p) for recognizing Collection. Nonetheless, Chakrabarti, Cormode,
Kondapally and McGregor [7] gave a one-pass randomized for PQ using memory space O(√N ).
They also showed that several passes do not help, since any p-pass randomized algorithm would
require memory space Ω(√N /p). A similar lower bound was showed independently, but using
different tools, by Jain and Nayak [10]. The case of a single pass was established previously by
Magniez, Mathieu and Nayak [15] for checking the well-formedness of parenthesis expressions, or
equivalently the behavior of a stack.
A simpler variant of PQ with timestamps was in fact first studied by Chu, Kannan and Mc-
Gregor [9], where now each item is inserted to the queue with its index.
Definition 2 (PQ-TS). Let Σ = {ins(a), ext(a) : a ∈ {0, 1, . . . , U}} × N. Let w ∈ ΣN . Then
w ∈ PQ-TS(U ) if and only if w ∈ Collection(Σ), w[1, . . . , N ][1] ∈ PQ(U ), and w[i][2] = i when
w[i][1] = ins(a).
Nonetheless the two works [9, 7] let open two problems. The lower bound of [7] was only proved
for PQ, and no significant lower bounds for PQ-TS was established. Moreover, the streaming
complexity of PQ for algorithms that can process the stream in any direction has not been studied.
Even though recognizing PQ-TS is obviously easier than recognizing PQ, our first contribution
(Section 3) consists in showing that they both obey the same limitation, even with multiple passes
in the same direction.
Theorem 3. Every p-pass randomized streaming algorithm recognizing PQ-TS(3N/2) with bounded
error 1/3 requires memory space Ω(√N /p) for inputs of length N .
As a consequence, since this lower bound uses very restricted hard instances, it models most of
possible variations. For instance, assuming that the input is in Collection and has no duplicates,
is not sufficient to guarantee a faster algorithm. The proof of Theorem 3 consists in introducing
a related communication problem with Θ(√N ) players. Then we reduce the number of players
to 3, and prove a lower bound on the information carried by players, leading to the desired lower
bound. We are following the information cost approach taken in [8, 17, 2, 12, 11], among other
works. Recently, the information cost appeared as one of the most central notion in communication
2
complexity [6, 5, 13]. The information cost of a protocol is the amount of information that messages
carry about players' inputs. We adapt this notion to suit both the nature of streaming algorithms
and of our problem.
Even if our result suggests that allowing multiple passes does not help, one could also consider
the case of bidirectional passes. We believe that it is a natural relaxation of multi-pass streaming
algorithms where the stream models some external read-only memory.
In that case, we show
that a second pass, but in reverse order, makes the problem of checking PQ easy, even with no
timestamps (Section 4). A similar phenomenon has been established previously in [15] for checking
the well-formedness of parenthesis expressions. Their problem is simpler than ours, and therefore
our algorithm is more general.
Theorem 4. There is a bidirectional 2-pass randomized streaming algorithm recognizing PQ(U )
with memory space O((log N )(log U + log N )), time per processing item polylog(N, U ), and one-
sided bounded error N −c, for inputs of length N and any constant c > 0.
Our algorithm uses a hierarchical data structure similar to the one introduced in [15] for checking
well-parenthesized expressions. At high level, it also behaves similarly. It performs one pass in each
direction and makes an on-line compression of past information in at most log N hashcodes. While
this compression can loose information, the compression technique ensures that a mistake is always
detected in one of the two directions. Nonetheless our algorithm differs on two main points. First,
unlike parenthesized expressions, PQ is not symmetric. Therefore one has to design an algorithm
for each pass. Second, the one-pass algorithm for PQ [7] is technically more advanced than the one
of [15]. Thus designing a bidirectional 2-pass algorithm for PQ is more challenging.
Theorems 3 and 4 point out a strange situation but not isolated at all. Languages studied
in [9, 15, 7, 14] and in this paper have space complexity Θ(√N polylog(N )) for a single pass,
Ω(√N /p) for p passes in the same direction, and polylog(N ) for 2 passes but one in each direction.
We hope this paper makes progress in the study that phenomenon.
2 Preliminaries
In streaming algorithms (see [16] for an introduction), a pass on an input w ∈ ΣN , for some alphabet
Σ, means that w is given as an input stream w[1], w[2], . . . , w[N ], which arrives sequentially, i.e.,
letter by letter in this order. For simplicity, we assume throughout this article that the input length
N is always given to the algorithm in advance. Nonetheless, all our algorithms can be adapted to
the case in which N is unknown until the end of a pass.
Definition 5 (Streaming algorithm). A p-pass randomized streaming algorithm with space s(N )
and time t(N ) is a randomized algorithm that, given w ∈ ΣN as an input stream,
• performs k sequential passes on w;
• maintains a memory space of size at most s(N ) bits while reading w;
• has running time at most t(N ) per processed letter w[i];
• has preprocessing and postprocessing time at most t(N ).
The algorithm is bidirectional if it is allowed to access to the input in the reverse order, after
reaching the end of the input. Then p is the total number of passes in either direction.
3
The proof of our lower bound uses the language of communication complexity with multi-players,
and is based on information theory arguments. We consider number-in-hand and message-passing
communication protocols. Each player is given some input, and can communicate with another
player according to the rules of the protocol. Our players are embedded into a directed circle,
so that each player can receive (resp.
transmit) a message from its unique predecessor (resp.
successor). Each player send a message after receiving one, until the end of the protocol is reached.
Players have no space and time restriction. Only the number of rounds and the size of messages
are constrained.
Consider a randomized multi-player communication protocol P . We consider only two types of
random source, that we call coins. Each player has access to its own independent source of private
coins. In addition, all players share another common source of public coins. The output of P is
announced by the last player. This is therefore the last message of the last player. We say that P
is with bounded error ǫ when P errs with probability at most ε over the private and public coins.
The transcript Π of P is the concatenation of all messages sent by all players, including all public
coins. In particular, it contains the output of P , since it is given by the last player. Given a subset
S of players, we let ΠS be the concatenation of all messages sent by players in S, including again
all public coins.
We now remind the usual notions of entropy H and mutual information I. Let X, Y, Z be random
variables. Then H(X) = − Ex←X log Pr(X = x), H(XY = y) = − Ey←Y log Pr(X = xY = y),
H(XY ) = Ey←Y H(XY = y), and I(X : Y Z) = H(XZ) − H(XY, Z). The entropy and the
mutual information are non negative and satisfy I(X : Y Z) = I(Y : XZ).
The mutual information between two random variables is connected to the Hellinger distance h
between their respective distribution probabilities. Given a random variable X we also denote by
X its underlying distribution.
Proposition 6 (Average encoding). Let X, Y be random variables. Then Ey←Y h2(XY =y, X) ≤
κI(X : Y ), where κ = ln 2
2 .
The Hellinger distance also generalizes the cut-and-paste property of deterministic protocols to
randomized ones.
Proposition 7 (Cut and paste). Let P be a 2-player randomized protocol. Let Π(x, y) denote the
random variable representing the transcript in P when Players A, B have resp. inputs x, y. Then
h(Π(x, y), Π(u, v)) = h(Π(x, v), Π(u, y)), for all pairs (x, y) and (u, v).
Last we use that the square of the Hellinger distance is convex, and the following connexion to
2kX − Y k1 ≤ √2h(X, Y ). For a reference on these
the more convention ℓ1-distance: h(X, Y )2 ≤ 1
results, see [10].
3 Lower bound for PQ-TS
The proof of our lower bound consists in first translating it into a 3m-player communication prob-
lem, for some large m; then reducing the number of players to 3 using the information cost approach;
and last studying the base case of 3 players using information theory arguments.
4
i = 3
i = 2
k = 3
i = 1
k = 3
23
23
23
23
i = 3
18
17
18
16
16
17
i = 2
14
14
18
17
18
16
16
14
17
14
A2
B2
C2
9
8
9
8
7
7
5
i = 1
5
2
2
A1
B1
5
2
C1
9
8
9
8
7
7
5
2
Figure 1: Left: Instance of Raindrops(m, 4) with one error: 17 is extracted after 16. Insertions ai
are circled. Right: Cutting Raindrops(m, 4) into 3m pieces to make it a communication problem.
Players' input are within each corresponding region.
3.1 From streaming algorithms to communication protocols
In this section, we write a instead of ins(a) and ¯a instead of ext(a). Consider the following set of
hard instances of size N = (2n + 2)m:
Raindrops(m, n) (see LHS of Figure 1)
• For i = 1, 2, . . . , m, repeat the following motif:
-- For j = 1, 2, . . . , n, insert either vi,j = 3(ni − j) or vi,j = 3(ni − j) + 2
-- Insert either ai = 3(ni − (ki − 1)) + 1 or ai = 3(ni − ki) + 1, for some ki ∈
{2, . . . , n}
-- Extract vi,1, vi,2, . . . , vi,ki−1, ai in decreasing order
• Extract everything left in decreasing order
Observe that such an instance is in Collection. One can compute the timestamps for each
value by maintaining only O(log N ) additionnal bits. Last, there is only one potential error in each
motif that can make it outside of PQ-TS. Indeed, vi,1, vi,2, . . . , vi,ki−1, ai are in decreasing order
up to a switch between ai and vi,ki−1.
Given such an instance as a stream, an algorithm for PQ-TS must decide if an error occurs
between ai and vi,ki, for some i. Intuitively, if the memory space is less than εn, for a small enough
constant ε > 0, then the algorithm cannot remember all the values (vi,j)j when ai is extracted, and
therefore cannot check a potential error with ai. The next opportunity is during the last sequence
of extractions. But then, the algorithm has to remember all values (ai)i, which is again impossible
if the memory space is less than εm.
In order to formalize this intuition, Lemma 8 (proof in Appendix A) first translates our problem
into a communication one between 3m players as shown on the RHS of Figure 1. Then we analyze
its complexity using information theory arguments in Section 3.2.
5
Any insertion and extraction of an instance in Raindrops(m, n) can be described by its index
and a single bit. Let xi[j] ∈ {0, 1} such that vi,j = 3(ni − j) + 2xi[j]. Similarly, let di ∈ {0, 1} such
that ai = 3(ni − ki) + 1 + 3di. For simplicity, we write x instead of (xi)1≤i≤m. Similarly, we use
the notations k and d. Then our related communication problem is:
WeakIndex(m, n)
• Input for players (Ai, Bi, Ci)1≤i≤m:
-- Player Ai has a sequence xi ∈ {0, 1}n
-- Player Bi has xi[1, ki − 1], with ki ∈ {2, . . . , n} and di ∈ {0, 1}
-- Player Ci has xi[ki, n]
• Output: fm(x, k, d) = Wm
• Communication settings:
i=1 f (xi, ki, di), where f (x, k, d) = [(d = 0) ∧ (x[k] = 1)]
-- One round: each player sends a message to the next player according to the
diagram A1 → B1 → A2 → ··· → Bm → Cm → Cm−1 → ··· → C1.
-- Multiple rounds: If there is at least one round left, C1 sends a message to A1,
and then players continue with the next round.
Lemma 8. Assume there is a p-pass randomized streaming algorithm for deciding if an instance
of Raindrops(n, m) is in PQ-TS(3mn) with memory space s(m, n) and bounded error ε. Then
there is a p-round randomized protocol for WeakIndex(n, m) with bounded error ε such that each
message has size at most s(m, n).
We are now ready to give the structure of the proof of Theorem 3, which has techniques based on
information theory. Define the following collapsing distribution µ0 of hard inputs (x, k, d), encoding
instances of Raindrops(1, n), where f always takes value 0. Distribution µ0 is such that (x, k) is
uniform on {0, 1}n × {2, . . . , n} and, given x, k, the bit d ∈ {0, 1} is uniform if x[k] = 0, and d = 1
if x[k] = 1. From now on, (X, K, D) are random variables distributed according to µ0, and (x, k, d)
denote any of their values.
Then the proof of Theorem 3 consists in studying the information cost of any communica-
tion protocol for WeakIndex(n, m), which is a lower bound on its communication complexity.
Using that µ0 is collapsing for f , Lemma 9 establishes a direct sum on the information cost of
WeakIndex(n, m). Then, even if f is constant on µ0, Lemma 12 lower bounds the information
cost of a single instance of WeakIndex(n, 1).
Proof of Theorem 3. Let n, N be positive integers such that N = (2n + 2)n. Assume that there
exists a p-pass randomized algorithm that recognizes PQ-TS(3N/2), with memory space αn and
bounded error ε, for inputs of size N . Then, by Lemma 8, there a p-round randomized protocol P for
WeakIndex(n, n) such that each message has size at most αn. By Lemma 9, one can derive from
P another (p + 1)-round randomized protocol P ′ for WeakIndex(n, 1) with bounded error ε, and
transcript Π′ satisfying Π′ ≤ 3(t + 1)αn and max {I(D : Π′
CX)} ≤ (p + 1)α.
Then by Lemma 12, 3(p + 1)α ≥ (1 − 2ε)/10, that is α = O(1/p), concluding the proof.
BX, K), I(K, D : Π′
6
3.2 Communication complexity lower bound
We first reduce the general problem WeakIndex(n, m) with 3m players to a single instance of
WeakIndex(n, 1) with 3 players.
In order to do so we exploit the direct sum property of the
information cost. The use of a collapsing distribution where f is always 0 is crucial.
Lemma 9. If there is a p-round randomized protocol P for WeakIndex(n, m) with bounded error
ε and messages of size at most s(m, n), then there is a (p + 1)-round randomized protocol P ′ for
WeakIndex(n, 1) with bounded error ǫ, and transcript P ′ satisfying Π′ ≤ 3(p + 1)s(m, n) and
max{I(D : Π′
Sketch of proof. Given a protocol P , we show how to construct another protocol P ′ for any instance
(x, k, d) of WeakIndex(n, 1). In order to avoid any confusion, we denote by A, B and C the three
players of P ′, and by (Ai, Bi, Ci)i the ones of P .
BX, K), I(K, D : Π′
CX)} ≤ p+1
m s(m, n).
Protocol P ′
• Using public coins, all players generate uniformly at random j ∈ {1, . . . , m}, and
xi ∈ {0, 1}n for i 6= j
• Players A, B and C set respectively their inputs to the ones of Aj, Bj, Cj
• For all i > j, Player B generates, using its private coins, uniformly at random ki ∈
{2, . . . , n}, and then it generates uniformly at random di such that f (xi, ki, di) = 0
• For all i < j, Player C generates, using its private coins, uniformly at random ki ∈
{2, . . . , n}, and then it generates uniformly at random di such that f (xi, ki, di) = 0
• Players A, B and C run P as follows. A simulates Aj only, B simulates Bj and
(Ai, Bi, Ci)i>j, and C simulates Cj and (Ai, Bi, Ci)i<j.
Observe that A starts the protocol if j = 1, and C starts otherwise. Moreover C stops the simulation
after p rounds if j = 1, and after p+1 rounds otherwise. For all i 6= j, entries are generated such that
f (xi, ki, ai) = 0, therefore fm(X, k, d) = f (xj, kj, aj) = f (x, k, a), and P ′ has the same bounded
error than P .
Then we show in Appendix A that P ′ satisfies the required conditions of the lemma.
We now prove a trade-off between the bounded error of a protocol for a single instance of
WeakIndex(n, 1) and its information cost. The proof involves some of the tools of [10] but with
some additional obstacles to apply them. The inherent difficulty is due to that we have 3 players
whereas the cute-and-paste property applies to 2-player protocols. Therefore we have to group 2
players together.
Given some parameters (x, k, a) for an input of WeakIndex(n, 1), we denote by Π(x, k, a) the
random variable describing the transcript Π of our protocol. We start by two lemmas exploiting
the average encoding theorem (proofs in Appendix A).
Lemma 10. Let P be a randomized protocol for WeakIndex(n, 1) with transcript Π satisfying
Π ≤ αn and I(K, D : ΠCX) ≤ α. Then
E
x[1,l−1],l
h2(Π(x[1, l − 1]0X[l + 1, n], l, 1), Π(x[1, l − 1]1X[l + 1, n], l, 1)) ≤ 28α,
where l ∈ [ n
2 + 1, n] and x[1, l − 1] are uniformly distributed.
7
Lemma 11. Let P be a randomized protocol for WeakIndex(n, 1) with transcript Π satisfying
I(D : ΠBX, K) ≤ α. Then
E
x[1,l−1],l
h2(Π(x[1, l − 1]0X[l + 1, n], l, 0), Π(x[1, l − 1]0X[l + 1, n], l, 1)) ≤ 12α,
where l ∈ [ n
2 + 1, n] and x[1, l − 1] are uniformly distributed.
We now end with the main lemma which combines both previous ones and applies the cut-and-
paste property, where Players A, C are grouped.
Lemma 12. Let P be a randomized protocol for WeakIndex(n, 1) with bounded error ǫ, and
transcript Π satisfying Π ≤ αn and max {I(D : ΠBX, K), I(K, D : ΠCX)} ≤ α. Then α ≥
(1 − 2ε)/10.
Proof. Let L be a uniform integer random variable in [ n
2 + 1, n]. Remind that we enforce the output
of P to be part of Π. Therefore, any player, and in particular B, can compute f with bounded
error ε given Π. Since f (x[1, l − 1]0X[l + 1, n], l, 0) = 0 and f (x[1, l − 1]1X[l + 1, n], l, 1) = 1, the
error parameter ε must satisfies
E
x[1,l−1],lkΠ(x[1, l − 1]0X[l + 1, n], l, 0) − Π(x[1, l − 1]1X[l + 1, n], l, 0)k1 ≥ 2(1 − 2ε).
The rest of the proof consists in upper bounding the LHS by 19α.
Applying the triangle inequality and that (u+ v)2 ≤ 2(u2 + v2) on the inequalities of Lemmas 10
and 11 gives
E
x[1,l−1],l
h2(Π(x[1, l − 1]0X[l + 1, n], l, 0), Π(x[1, l − 1]1X[l + 1, n], l, 1)) ≤ 30α.
We then apply the cut-and-paste property by considering (A, C) as a single player with transcript
ΠA,C. Therefore
E
x[1,l−1],l
h2(Π(x[1, l − 1]0X[l + 1, n], l, 1), Π(x[1, l − 1]1X[l + 1, n], l, 0)) ≤ 30α.
Combining again with the inequality from Lemma 11 gives
E
x[1,l−1],l
h2(Π(x[1, l − 1]0X[l + 1, n], l, 0), Π(x[1, l − 1]1X[l + 1, n], l, 0)) ≤ 42α.
Last, we get the requested upper bound by using the connexion between the Hellinger distance and
the ℓ1-distance, and the convexity of the square function.
4 Bidirectional streaming algorithm for PQ
Remember that in this section our stream is given without any timestamps. Therefore we consider
in this section only streams w of ins(a), ext(a), where a ∈ [0, U ]. For the sake of clarity, we assume
for now that the stream has no duplicate. Our algorithms can be extended to the general case, but
the technical difficulties shadow the main ideas.
Up to padding we can assume that N is a power of 2: we append a sequence of
ins(a)ext(a)ins(a + 1)ext(a + 1) . . . of suitable length, where a is large enough so that there
8
is no duplicate (assuming that w is of even size, otherwise w 6∈ PQ(U )). We use O(log N ) bits of
memory to store, after the first pass, the number of letters padded.
We use a hash function based on the one used by the Karp-Rabin algorithm for pattern match-
ing. For all this section, let p be a prime number in {max(2U + 1, N c+1), . . . , 2 max(2U + 1, N c+1)},
for some fixed constant c ≥ 1. Since our hash function is linear we only define it for single inser-
tion/extraction as
hash(ins(a)) = αa mod p,
and
hash(ext(a)) = −αa mod p,
where α is a randomly chosen integer in [0, p − 1]. This is the unique source of randomness of
our algorithm. A hashcode h encodes a sequence w if h = hash(w) as a formal polynomial in α.
In that case we say that h includes w[i], for all i. Moreover w is balanced if the same integers
have been inserted and extracted.
In that case it must be that h = 0. We also say that h is
balanced it it encodes a balanced sequence w. The converse is also true with high probability by
the Schwartz-Zippel lemma.
Fact 13. Let w be some unbalanced sequence. Then Pr(hash(w) = 0) ≤ N
p ≤ 1
N c .
The forward-pass algorithm was introduced in [7], but the reverse-pass one is even simpler. As a
warming up, we start by introducing the later algorithm. In order to keep it simple to understand,
we do not optimize it fully. Last define the instruction Update(h, v) that returns (h + hash(v)
mod p) and updates h to that value.
4.1 One-reverse-pass algorithm for PQ
Our algorithm decomposes the stream w into blocks. We call a valley an extraction w[t] = ext(a)
with w[t + 1] = ins(b). A new block starts at every valley. To the i-th block we associate a
hashcode hi and an integer mi. Hashcode hi encodes all the extractions within the block and the
matching insertions. Integer mi is the minimum of extractions in the block. With the values (mi)i,
one can encode insertions in the correct hi if w ∈ PQ. Observe that we use index notations for
block indices and bracket notations for stream positions.
Algorithm 1 uses memory space O(r), where r is the number of valleys in w. We could make it
run with memory space O(√N log N ) by reducing the number of valleys as in [7]. We do not need
to as we use another compression in the two-pass algorithm.
We first state a crucial property of Algorithm 1, and then show that it satisfies Theorem 15,
when there is no duplicate. We remind that we process the stream from right to left.
Lemma 14. Consider Algorithm 1 right after processing ins(a). Assume that ext(a) has been
already processed. Let hk, hk′ be the respective hashcodes including ext(a), ins(a). Then k = k′ if
and only if all ext(b) occurring between ext(a) and ins(a) satisfy b > a.
Theorem 15. There is a 1-reverse-pass randomized streaming algorithm for PQ(U ) with memory
space O(r(log N + log U )) and one-sided bounded error N −c, for inputs of length N with r valleys,
and any constant c > 0.
Proof. We show that Algorithm 1 suits the conditions, assuming there is no duplicate. Let w ∈
PQ(U ). Then w always passes the test at line 10. Moreover, by Lemma 14, each insertion ins(a)
is necessarily in the same hashcode than its matching extraction ext(a). Therefore, all hashcodes
9
Algorithm 1: One-reverse-pass algorithm for PQ
1 m0 ← −∞; h0 ← 0; t ← N ; i ← 0 // i is called the block index
2 While t > 0
If w[t] = ins(a)
k ← max{j ≤ i : mj ≤ a}; // C o m p u t e the h a s h c o d e index of a
Update(hk, w[t])
Else w[t] = ext(a)
If w[t + 1] = ins(b) // This is a valley . We start a new block
Else w[t + 1] = ext(b)
i ← i + 1; mi ← a; hi ← 0 // Create a new h a s h c o d e
Check (a ≥ b)
// Check that e x t r a c t i o n s are well - o r d e r e d
Update(hi, w[t])
t ← t − 1
3
4
5
6
7
8
9
10
11
12
13 For j = 0 to i: Check (hj = 0) // Check that h a s h c o d e s are b a l a n c e d w . h . p .
14 Accept
// w s u c c e e d e d to all checks
equal 0 at line 13 since they are balanced. In conclusion, the algorithm accepts w with probability
1.
Assume now that w 6∈ PQ. First we show that unbalanced w are rejected with high probability,
that is at least 1 − N −c, at line 13, if they are not rejected before.
Indeed, since each w[t] is
encoded in some hj, at least one hj must be unbalanced. Then by Fact 13, the algorithm rejects
w.h.p. We end the proof assuming w balanced. We remind that we process the stream from right
to left. The two remaining possible errors are: (1) ins(a) is processed before ext(a), for some
a; and (2) ext(a), ext(b), ins(a) are processed in this order with b < a and possibly intermediate
insertions/extractions. In both cases, we show that some hashcodes are unbalanced at line 13, and
therefore fail the test w.h.p by Fact 13, except if the algorithm rejects before.
Consider case (1). Since ins(a) is processed before ext(a), there is at least one valley between
ins(a) and ext(a). Therefore ins(a) and ext(a) are encoded into two different hashcodes, that
are unbalanced at line 13.
Consider now case (2). Lemma 14 gives that ext(a) and ins(a) are encoded in two different
hashcodes, that are again unbalanced at line 13.
4.2 Bidirectional two-pass algorithm
Algorithm 2 performs one pass in each direction using Algorithm 3. We use the hierarchical
data structure of [15] in order to reduce the number of blocks. A block of size 2i is of the form
[(q − 1)2i + 1, q2i], for 1 ≤ q ≤ N/2i. Observe that, given two such blocks, either they are disjoint
or one is included in the other. We decompose dynamically the letters of w, that have been already
processed, into nested blocks of 2i letters as follows. Each new processed letter of w defines a new
block. When two blocks have same size, they merge. All processed blocks are pushed on a stack.
Therefore, only the two topmost blocks of the stack may potentially merge. Because the size of
each block is a power of 2 and at most two blocks have the same size (before merging), there are
at most log N + 1 blocks at any time.
Moreover, since our stream size is a power of 2, all blocks eventually appear in the hierarchical
decomposition, whether we read the stream from left to right or from right to left. In fact, if two
same-sized blocks appear simultaneously in one decomposition before merging, the same is true in
10
Algorithm 2: Bidirectional 2-pass algorithm for PQ
1 O n e P a s s A l g o r i t h m (w)
2 O n e P a s s A l g o r i t h m (w)
3 Accept
// w s u c c e e d e d to all checks
r e a d i n g stream from left to right
r e a d i n g stream from right to left
Algorithm 3: OnePassAlgorithm
1 S ← [];
2 If left - to - right - pass Then Push (S ,(0,−∞, 0)) // I n i t i a l i z a t i o n of S
3 While stream is not empty
4
5
6
7
Read ( next letter v on stream ) // See below
While the 2 t o p m o s t e l e m e n t s of S have same block size ℓ
(h1, m1, ℓ) ←Pop (S ); (h2, m2, ℓ) ←Pop (S )
Push (S ,(h1 + h2 mod p, min(m1, m2), 2ℓ)) // Merge of 2 blocks
8 If left - to - right - pass Then Check (S = [(0,−∞, 0), (0, 0, N )])
9 Else Check (S = [(0, 0, N )])}
10 Return
11
12 F u n c t i o n Read ( v ):
13 Case v = ins(a) // When r e a d i n g an i n s e r t i o n
17 Case v = ext(a) and left - to - right - pass // When r e a d i n g an e x t r a c t i o n
Let (h, m, ℓ) be the first item of S from top such that a ≥ m
R e p l a c e (h, m, ℓ) by (Update(h, v), m, ℓ)
Push (S, (0, +∞, 1))
For all items (h, m, ℓ) on S such that m > a: Check (h = 0)
Let (h, m, ℓ) be the first item of S from top such that a > m
R e p l a c e (h, m, ℓ) by (Update(h, v), m, ℓ)
Push (S ,(0, a, 1))
14
15
16
18
19
20
21
22 Case v = ext(a) and right - to - left - pass // When r e a d i n g an e x t r a c t i o n
23
24
For all items (h, m, ℓ) on S such that m > a: Check (h = 0)
Push (S ,(hash(v), a, 1))
the other decomposition. This point is crucial for our analysis.
Algorithm 3 uses the following description of a block B: its hashcode hB , the minimum mB of
its extractions, and its size ℓB. For the analysis, we also note tB the index such that w[tB] = mB.
Among those parameters, only hB can change without B being merged with another block. On the
pass from right to left, all extractions from the block and the matching insertions are included in
hB. On the pass from left to right, insertions are included in the hashcode of the earliest possible
block where they could have been, and the extractions are included with their matching insertions.
The minimums (mB)B are used to decide where to include values (except extractions on the pass
from right to left). Observe that it is important to check that hB = 0 whenever possible and not
at the end of the execution of the algorithm, since only one block is left at the end.
When there is some ambiguity, we denote by h→
B and h←
B the hashcodes for the left-to-right and
right-to-left passes. Observe that mB, tB, ℓB are identical in both directions.
Proof of Theorem 4. We show that Algorithm 2 suits the conditions, assuming there is no duplicate.
The space constraints are satisfied because each element of S takes space O(log N + log U ) and S
has size at most log N + 1. The processing time is from inspection.
11
ins(a): case 1
ext(mB)
ins(a): case 2
ext(b)
ext(a)
ins(a): case 3
ext(mc)
ins(a): case 1
ρ′
tB
ρ′
B
τ
ρ
ρ′
tC
ρ′
C
Figure 2: Relative positions of insertions and extractions used in the proof of Theorem 4
As with Theorem 15, inputs in PQ(U ) are accepted with probability 1, and unbalanced inputs
are rejected with high probability (at least 1−N −c). Let w 6∈ PQ be balanced. For ease of notations,
let w[−1] = ins(−∞) and w[0] = ext(−∞). Then, there are τ < ρ such that w[τ ] = ext(b),
w[ρ] = ext(a), a > b, and w[t] 6= ins(a) for all τ < t < ρ.
Among those pairs (τ, ρ), consider the ones with the smallest ρ. From those, select the one with
the smallest b, with w[τ ] = ext(b). Let B, C be the largest possible disjoint blocks such that τ is
in B and ρ in C. Then B and C have same size, are contiguous, and appear simultaneously in each
direction before they merge. Let ρ′ and τ ′ be such that w[ρ′] = ins(a) and w[τ ′] = ins(b). The
minimality of ρ and the minimality of b guarantee that w[t] is an insertion for all τ < t < ρ. Indeed
if w[t] = ext(c) either b > c, which contradicts the minimality of b, or c > b and (τ, t) contradicts
the minimality of ρ. In particular, tC ≥ ρ and tB ≤ τ . Similarly τ < τ ′, otherwise τ would be a
better candidate than ρ.
6∈ [tB, tC],
tB < ρ′ < τ , and ρ < ρ′ < tC. These cases determine in which hashcode ins(a) is included. We
analyze Algorithm 3 when some letter is processed before blocks potentially merge.
We distinguish three cases based on the position ρ′ of ins(a) (see Figure 2): ρ′
Case 1: ρ′ 6∈ [tB, tC]. One can prove that h→
h←
C is unbalanced when w[tB] is processed; therefore Algorithm 3 detects w.h.p. h→
depending on whether mB > mC (see Lemma 19 in Appendix B).
B is unbalanced when w[tC ] is processed and that
C 6= 0
Case 2: tB < ρ′ < τ . We show that when Algorithm 3 processes w[tB] = ext(mB), it checks
B 6= 0 or h←
h←
D = 0 at line 23 for some h←
D including ins(a) but not ext(a). Thus it rejects w.h.p.
When w[ρ′] = ins(a) is processed on the right-to-left pass, τ ∈ B1 with B1 a block in the stack.
τ ∈ B, therefore B1 intersects B. Because B1 6⊆ B, we have B1 ⊆ B. Because w[τ ] = ext(b), we
have a > b ≥ mB1, and block B1 is eligible at line 14 of Algorithm 3, meaning that w[ρ′] = ins(a)
is included in either h←
B2. Since ρ′ ∈ B, again B2 ⊆ B. Last, when
Algorithm 3 processes w[tB] = ext(mB), since we are still within B, some hashcode hB3, with
B3 ⊆ B, includes w[ρ′]. Moreover, h←
B3 does not include w[ρ] = ext(a) since ρ ∈ C and C comes
before B. Last, mB3 > mB, by definition of mB. Hence, Algorithm 3 checks h←
B3 = 0 at line 23
when processing w[tB]. B3 satisfies the conditions for D when w[tB] is processed, and Algorithm 3
rejects w.h.p.
B1 or a more recent hashcode h←
Case 3: ρ < ρ′ < tC. The proof is the same as case 2, replacing τ , B, B1, B2, B3, h←
B1, h←
B2 ,
C3, tC, B and line 23 with line 18. Note that we only
h←
B3, tB, C with ρ, C, C1, C2, C3, h→
have a ≥ mC1 this time, so it is important that the inequality at line 14 is large and not strict.
C2, h→
C1 , h→
4.3 Generalization when duplicates occur
We maintain two additional parameters δB and CB for each block B. The difference between
the number of insertions and extractions included in hB is stored in δB. Whenever δB = 0, we
12
check hB = 0. The number of unmatched occurrences of ins(mB) for the left-to-right pass (resp.
ext(mB) for the right-to-left pass) is stored in CB. We can then appropriately determine whether
each ext(mB) (resp. ins(mB)) should be included in hB.
The change on the criterion of line 14 of Algorithm 3 makes the proof of case 3 of the theorem
longer and breaks the symmetry.
Acknowledgements
Authors would like to thank Rahul Jain and Ashwin Nayak for sharing their intuition and possible
extensions of [10]. In particular, N.F. thanks Ashwin Nayak for having hosted him at IQC, Univer-
sity of Waterloo. They also thanks Christian Konrad, Wei Yu, Qin Zhang for related discussions,
and Andrew McGregor for motivating us to study restricted instances of PQ.
References
[1] N. Alon, Y. Matias, and M. Szegedy. The space complexity of approximating the frequency
moments. Journal of Computer and System Sciences, 58(1):137 -- 147, 1999.
[2] Z. Bar-Yossef, T. S. Jayram, R. Kumar, and D. Sivakumar. An information statistics approach
to data stream and communication complexity. Journal of Computer and System Sciences,
68(4):702 -- 732, 2004.
[3] M. Blum, W. S. Evans, P. Gemmell, S. Kannan, and M. Naor. Checking the correctness of
memories. Algorithmica, 12(2):225 -- 244, 1994.
[4] M. Blum and S. Kannan. Designing programs that check their work. Journal of the ACM,
42(1):269 -- 291, 1995.
[5] M. Braverman.
Interactive information complexity.
In Proc. of ACM Symp. on Theory of
Computing, pages 505 -- 524, 2012.
[6] M. Braverman and A. Rao. Information equals amortized communication. In 748-757, editor,
Proc. of IEEE Symp. on Foundations of Computer Science, 2011.
[7] A. Chakrabarti, G. Cormode, R. Kondapally, and A. McGregor. Information cost tradeoffs for
augmented index and streaming language recognition. In Proc. of IEEE Symp. on Foundations
of Computer Science, pages 387 -- 396, 2010.
[8] A. Chakrabarti, Y. Shi, A. Wirth, and A. C.-C. Yao. Informational complexity and the direct
sum problem for simultaneous message complexity. In Proc. of IEEE Symp. on Foundations
of Computer Science, pages 270 -- 278, 2001.
[9] M. Chu, S. Kannan, and A. McGregor. Checking and spot-checking the correctness of priority
queues. In Proc. of Int. Colloquium on Automata, Languages and Programming, pages 728 -- 739,
2007.
[10] R. Jain and A. Nayak. The space complexity of recognizing well-parenthesized expressions in
the streaming model: the index function revisited, 2010. ECCC Tech. Rep. TR10-071.
13
[11] R. Jain, J. Radhakrishnan, and P. Sen. A lower bound for the bounded round quantum
communication complexity of Set Disjointness. In Proc. of IEEE Symp. on Foundations of
Computer Science, pages 220 -- 229, 2003.
[12] T. S. Jayram, Ravi Kumar, and D.Sivakumar. Two applications of information complexity. In
Proc. of ACM Symp. on Theory of Computing, pages 673 -- 682, 2003.
[13] I. Kerenidis, S. Laplante, V. Lerays, J. Roland, and D. Xiao. Lower bounds on information
complexity via zero-communication protocols and applications. In Proc. of IEEE Symp. on
Foundations of Computer Science, 2012. To appear.
[14] C. Konrad and F. Magniez. Validating XML documents in the streaming model with external
memory. In Proc. of Int. Conf. on Database Theory, pages 34 -- 45, 2012.
[15] F. Magniez, C. Mathieu, and A. Nayak. Recognizing well-parenthesized expressions in the
streaming model. In Proc. of ACM Symp. on Theory of Computing, pages 261 -- 270, 2010.
[16] S. Muthukrishnan. Data Streams: Algorithms and Applications. Now Publishers Inc., 2005.
[17] M. Saks and X. Sun. Space lower bounds for distance approximation in the data stream model.
In Proc. of ACM Symp. on Theory of Computing, pages 360 -- 369, 2002.
A Missing proofs for the lower bound
We start by proving the lemma relating the streaming complexity of deciding if an in-
stance of Raindrops(m, n) belongs to PQ-TS(3mn) to the communication complexity of
WeakIndex(n, m).
Proof of Lemma 8. Assume that there exists a p-pass randomized streaming algorithm with mem-
ory space s(m, n), that decides if an instance of Raindrops(m, n) belongs or not to PQ-TS(3nm).
Each instance of Raindrops(m, n) can be encoded by an input of WeakIndex(n, m), where each
of the 3m players has one part of it. Then, the rest of the proof consists in showing how the players
can use the algorithm in order to construct a protocol that satisfies the required properties of the
lemma.
Each player simulates alternatively the algorithm. A player performs the simulation until the
algorithm reaches the part of the input of the next player. Then the player sends the current state
of the algorithm, so that the next player can continue the simulation. Since the algorithm uses
at most memory space s(m, n), the current state can be encoded using s(m, n) bits. Each pass
corresponds to one round of communication, implying the result.
Before giving the next missing proofs of Section 3, we state some useful properties of entropy
and mutual information that we need. See [10] for more information.
Fact 16. Let X, Y, Z, R be random variables such X and Z are independent when conditioning on
R, namely when conditioning on R = r, for each possible values of r. Then I(X : Y Z, R) ≥ I(X :
Y R).
14
Proof. From the definition of mutual information and the independence of X, Z when conditioning
on R, we get that
I(X : Y Z, R) = H(XZ, R) − H(XY, Z, R) = H(XR) − H(XY, Z, R).
Using that entropy can only decrease under conditioning, and using again the definition of mutual
information, we conclude by bounding the last term as
H(XR) − H(XY, Z, R) ≥ H(XR) − H(XY, R) = I(X : Y R).
Proposition 17 (Chain rule). Let X, Y, Z, R be random variables. Then I(X, Y : ZR) = I(X :
ZR) + I(Y : ZX, R).
Proposition 18 (Data processing inequality). Let X, Y, Z, R be random variables such that R is
independent from X, Y, Z. Then I(X : Y Z) ≥ I(f (X, R) : Y Z), for every function f .
Note that the previous property is usually stated with no variable T . Nonetheless, since T is
independent from the other variables, we have I(X : Y Z) = I(X, R : Y Z), and then we can apply
the usual data processing inequality.
We can now prove our three lemmas.
End of proof of Lemma 9. Let Π, Π′ be the respective transcripts of P, P ′. For convenience, note
ΠCm+1 = ΠBm, ΠB0 = ΠCm and ΠCm+1 = ΠA1. Remind that the public coins of a protocol are
included in its transcript.
First, each player of P ′ sends 3 messages by round, and there are (p + 1) rounds. Since each
message has size at most s(m, n), we derive that the length of Π′ is at most 3(p + 1)s(m, n).
Then, in order to prove that there is only a small amount of information in the transcripts of
Bob and Charlie, we show a direct sum of some appropriated notion of information cost. Consider
first the transcript of Player C1. Because of the restriction on the size of his messages, we know
that ΠC1 ≤ (p + 1)s(m, n). From this we derive a first inequality on the amount of information
this transcript can carry, using that the entropy of a variable is at most its bit-size:
I(K, D : ΠC1X) ≤ ΠC1 ≤ (p + 1)s(m, n).
We now use the chain rule in order to get a bound about the information carried by P ′ on a single
instance.
I(K, D : ΠC1X) =
≥
m
X
j=1
m
X
j=1
m
I((Ki, Di)j≥i : ΠC1X, (Ki, Di)i<j)
(by chain rule)
I(Kj, Dj : ΠBj−1X, (Ki, Di)i<j)
(by data processing inequality)
X
j=1
I(Kj, Dj : ΠBj−1X)
(by Fact 16)
≥
= m × I(KJ , DJ : ΠBJ −1X, J)
= m × I(KJ , DJ : ΠBJ −1, J, (Xi)i6=JXJ ) (independence of J, (Xi)i6=J )
= m × I(K, D : Π′
(since J, (Xi)i6=J are public coins of P ′).
(by conditioning on J)
CX)
15
We then do similarly for Player Bm and therefore conclude the proof. First the size bound on
messages of Bm gives I(ΠBm : DX, K) ≤ (p + 1)s(m, n). Then as before we get:
I(D : ΠBmX, K) =
m
X
j=1
I(Dj : ΠBmX, K, (Di)i>j) ≥
m
X
j=1
I(Dj : ΠCj+1X, K, (Di)i>j)
≥ m × I(DJ : ΠCJ +1, J, (Xi)i6=JXJ , KJ ) = m × I(D : Π′
BX, K).
Proof of Lemma 10. From the second hypothesis and the data processing inequality we
get that I(K, D : ΠA,CX) ≤ α, which after applying the average encoding leads to
Ex,k,d h2(ΠA,C(x, k, d), ΠA,C (x, K, D)) ≤ κα. We now restrict µ0 by conditioning on D = 1.
Then (X, K) is uniformly distributed. Moreover, since D = 1 with probability 3/4 on µ0, we
get Ex,k h2(ΠA,C(x, k, 1), ΠA,C (x, K, 1)) ≤ 4
3 κα. Let J, L be uniform integer random variables re-
spectively in [2, n
2 + 1, n]. Then the above implies Ex,j h2(ΠA,C(x, j, 1), ΠA,C (x, K, 1)) ≤
8
3 κα and Ex,l h2(ΠA,C(x, l, 1), ΠA,C (x, K, 1)) ≤ 8
3 κα. Applying the triangle inequality and that
(u + v)2 ≤ 2(u2 + v2), we get
2 ] and [ n
E
x,j,l
h2(ΠA,C(x, j, 1), ΠA,C (x, l, 1)) ≤ 32
3 κα.
Using the convexity of h2, we finally obtain for b = 0, 1:
E
x[1,l−1],j,l
h2(ΠA,C(x[1, l − 1]bX[l + 1, n], j, 1), ΠA,C (x[1, l − 1]bX[l + 1, n], l, 1)) ≤ 64
3 κα.
Now the chain rule allow us to measure the information about a single bit in ΠA,C as
I(X[L] : ΠA,C(X, J, 1)X[1, L − 1]) = E
l←L
2
n × I(X[ n
=
I(X[l] : ΠA,C(X, J, 1)X[1, l − 1])
2 + 1, n] : ΠA,C(X, J, 1)X[1, n
2 ]).
Since the entropy of a variable is at most its bit-size, we get that the last term is upper bounded
by ΠA,C, which is at most αn by the first hypothesis. Then as before, the average encoding and
the triangle inequality lead to
E
x[1,l−1],j,l
h2(ΠA,C(x[1, l − 1]0X[l + 1, n], j, 1), ΠA,C (x[1, l − 1]1X[l + 1, n], j, 1)) ≤ 16κα.
Combining gives
E
x[1,l−1],l
h2(ΠA,C (x[1, l − 1]0X[l + 1, n], l, 1), ΠA,C (x[1, l − 1]1X[l + 1, n], l, 1)) ≤ 28α.
Let RB be the random coins of B. Since they are independent from all variables, including the
messages, the previous inequality is still true when we concatenate RB to ΠA,C. Then ΠB is
uniquely determined from RB once K, D, X[1, K − 1] are fixed, which is the case in that inequality.
Therefore replacing RB by ΠB can only decrease the distance, concluding the proof.
16
Proof of Lemma 11. Using the data processing inequality and the hypothesis we get that I(D :
ΠX, K)) ≤ α. Therefore by average encoding, Ex,k,d h2(Π(x, k, d), Π(x, k, D)) ≤ κα.
2κα. Using the convexity of h2 and the fact that X[l] is a uniform random bit, we derive
2 + 1, n]. Then Ex,l,d h2(Π(x, l, d), Π(x, l, D)) ≤
Let L be a uniform integer random variable in [ n
E
x[1,l−1],l,d
h2(Π(x[1, l − 1]0X[l + 1, n], l, d), Π(x[1, l − 1]0X[l + 1, n], l, D)) ≤ 4κα.
Since D = 0 with probability 1/2 when X[l] = 0 and K = l, we finally get the two inequalities
E
x[1,l−1],l
h2(Π(x[1, l − 1]0X[l + 1, n], l, 0), Π(x[1, l − 1]0X[l + 1, n], l, D)) ≤ 8κα,
E
h2(Π(x[1, l − 1]0X[l + 1, n], l, 1), Π(x[1, l − 1]0X[l + 1, n], l, D)) ≤ 8κα,
leading to the conclusion using the triangle inequality and that (u + v)2 ≤ 2(u2 + v2).
x[1,l−1],l
B Missing proofs for the algorithm
We start by proving the property of Algorithm 1 we use in the proof of Theorem 15.
Proof of Lemma 14. Remind again, that we process the stream from right to left in this proof, and
that hk, hk′ are the respective hashcodes including ext(a), ins(a). First assume that all ext(b)
between ext(a) and ins(a) satisfy b > a. Let i be the current block index while processing ins(a).
Observe that k is the current block index right after processing ext(a). Since ext(a) is processed
before ins(a) and since there is a valley between ext(a) and ins(a), we have k < i.
We prove that k′ = max{j ≤ imj ≤ a} = k. The first equality is from line 4 of Algorithm 1.
We now prove the second equality. For each j ∈ {k +1, . . . , i}, value mj is extracted between ext(a)
and ins(a). Then, our assumption leads to mj > a. Moreover, because the algorithm checks at
line 10 that extraction sequences included in the same hashcode are decreasing, we have mk ≤ a,
leading to the second equality.
We now prove the converse by contrapositive. Assume that some ext(b) between ext(a) and
ins(a) satisfies b ≤ a. Since we forbid duplicates, in fact b < a. Let j be the current block index
right after processing ext(b). Then line 10 ensures that mj ≤ b. Again, k is the current block index
right after processing ext(a), and therefore k ≤ j. If k = j, then the extraction sequence is not
decreasing and line 10 rejects, contradicting the hypotheses that the algorithm has not rejected yet
after processing ins(a). Therefore k < j. But, line 4 and the fact that mj ≤ b imply that k′ ≥ j,
and therefore k < k′.
We now give the missing part of the proof of Theorem 4.
Lemma 19. If ρ′ 6∈ [tB, tC ], then Algorithm 2 rejects w with probability at least 1 − N −c.
Proof. We prove that h→
C is unbalanced when
w[tB] is processed. From that, we deduce that the algorithm rejects with high probability unless
mB ≤ mC and mC ≤ mB, i.e. mB = mC, which is impossible because w has no duplicates and B
and C are disjoint.
B is unbalanced when w[tC ] is processed and that h←
17
Indeed if mC < mB then Algorithm 3 checks that h→
and rejects with high probability because h→
high probability at line 23 when processing w[tB] on the right-to-left pass.
B = 0 at line 18 when processing w[tC],
B is unbalanced. Similarly, if mC < mB, it rejects with
B (resp. h←
Now we only have to prove that h→
C ) is unbalanced when w[tC] (resp. w[tB]) is
processed. Let us assume there exists B1 ( B such that ins(a) is included in h→
B1 when w[tB]
is processed. Then, by definition of mB, mB1 > mB. Moreover, ρ ∈ C, so w[ρ] = ext(a) is not
processed yet and not included in B1. Therefore, Algorithm 3 checks h→
B1 = 0 at line 18, and rejects
w.h.p. We can now assume that there is no such B1 ( B, and therefore that hB, does not include
ins(a) when w[tC ] is processed. Since h→
B is unbalanced when tC is processed.
B1 , B, B1, tB and tC with h←
C1 , C,
B includes ext(a), h→
B , h→
The proof for h←
C is the same as above, replacing h→
C , h←
C1, tC and tB, and line 18 with line 23.
18
|
1906.07887 | 1 | 1906 | 2019-06-19T02:56:11 | Tutorial on algebraic deletion correction codes | [
"cs.DS",
"cs.IT",
"eess.SP",
"cs.IT"
] | The deletion channel is known to be a notoriously diffcult channel to design error-correction codes for. In spite of this difficulty, there are some beautiful code constructions which give some intuition about the channel and about what good deletion codes look like. In this tutorial we will take a look at some of them. This document is a transcript of my talk at the coding theory reading group on some interesting works on deletion channel. It is not intended to be an exhaustive survey of works on deletion channel, but more as a tutorial to some of the important and cute ideas in this area. For a comprehensive survey, we refer the reader to the cited sources and surveys.
We also provide an implementation of VT codes that correct single insertion/deletion errors for general alphabets at https://github.com/shubhamchandak94/VT_codes/. | cs.DS | cs |
Tutorial on algebraic deletion correction codes
Kedar Tatwawadi∗, Shubham Chandak
Stanford University
April 2019
Abstract
The deletion channel is known to be a notoriously difficult channel to design error-correction codes
for. In spite of this difficulty, there are some beautiful code constructions which give some intuition about
the channel and about what good deletion codes look like. In this tutorial we will take a look at some of
them. This document is a transcript of my talk at the coding theory reading group on some interesting
works on deletion channel. It is not intended to be an exhaustive survey of works on deletion channel,
but more as a tutorial to some of the important and cute ideas in this area. For a comprehensive survey,
we refer the reader to the cited sources and the following papers and surveys [8, 10, 12].
We also provide implementation of VT codes that correct single insertion/deletion errors for general
alphabets at https://github.com/shubhamchandak94/VT_codes/.
1
Introduction
Deletion channel is one of the most fundamental of the channels, and is still not well understood. Most of
you would know what a deletion channel/error looks like. But, to give an example, this is what a single
deletion looks like:
Here the decoder receives the message 1010 and needs to recover the original message. This corresponds
10110 → 1010
to a single deletion, either in position 3 or 4. We cannot say what position the deletion occurred.
In relation to our favourite erasure channel, where such a single error might look like:
Here, we know what positions the erasure happened. The deletion channel is in fact a strictly worse channel,
as we can convert the erasure channel output into a deletion channel output, by simply removing all the 'e'
symbols from the output.
10110 → 10e10
We will also see that it has connections to our second favourite channel the Binary Symmetric Channel
(BSC) channel.
1.1 Capacity of Deletion Channel
We define a binary deletion channel, BDC with deletion rate α, where each symbol in input xn can get
deleted with probability α. This definition has similarities with the BEC and BSC, but that's where the
similarity ends. Surprisingly unlike BEC and BSC, we know quite less about BDC.
1. We still do not know the capacity of the BDC channel. One reason is that BDC is not really a discrete
memoryless channel (DMC). For a DMC, you should be able to write:
∗Corresponding author: [email protected]
p(ynxn) =
n
Yi=1
p(yixi)
1
We cannot write the same for deletion channels, one reason: as the output does not have "length" n,
and is in fact a random variable. Shannon theory gives us a nice characterization for the capacity of
DMCs:
C = max
p(x)
I(X; Y )
(1)
But, without this nice expression, finding the capacity is a much more difficult task.
2. We can of course get some bounds on the deletion channel capacity. For example: we know that binary
erasure channel is strictly better than the BDC. Thus:
CBDC (α) ≤ 1 − α
(2)
This has been an open problem since quite some time now, but there has been some recent progress
on this, which I will briefly talk about.
Most of the results are a part of the survey by Mitzmacher [10]. Kalai, Mitzenmacher, and Sudan [7]
showed that when α → 0 the capacity is almost equal to the BSC. Best known capacity lower bounds
I am aware of are [5]: 0.11(1 − α) . Recently, improved capacity upper bounds were obtained by
Cheraghchi [4] which state that:
CBDC ≤ (1 − α) log Φ
(3)
for α ≥ 0.5, where Φ is the golden ratio. For those of you who are interested, they essentially find
capacity bounds for a channel known as Poisson repeat channel (the number of times each symbol is
repeated is a Poisson random variable), which are then ported over for deletion channel as a special
case.
This is mainly to give a sense of how difficult the deletion channel analysis is, and how little we really
know about it. Typically, not knowing the capacity directly translates to not knowing good codes.
But, surprisingly we know some nice code constructions in specific cases, which we will discuss next.
Deletion channel is also related to the insertion channel, or the indel channel, where both insertions
and deletions happen, or general edit distance channels:
indels + substitutions/reversals. A better
understanding of the deletion channel is useful not just for communication, but for the problem of
denoising, which is quite common (say when we type things and miss a character).
2 Adversarial deletion error
We spoke about "Shannon-type" random deletion error capacity. But, for majority of the talk, we will
mainly talk about the adversarial error, where when we say e errors, we mean at max e symbols are deleted.
In this context let us define an e-deletion error correction code.
Definition 1 For any vector u ∈ Fn
obtained after e deletions.
2 we define De(u) as the descendants; i.e. all the vectors in Fn−e
2
For example:
For u = 01101, D1(u) = {0110, 0111, 0101, 1101}
For u = 00000, D1(u) = {0000}
Thus the descendant balls need not be of the same sizes unlike the Hamming balls we are familiar with.
And this will lead to some interesting scenarios as we will see. In fact we can analyze the size of 1-deletion
balls:
Lemma 1 The size of 1-deletion Descendant ball D1(un) of a vector un is equal to the number of "runs" in
un.
2
For example:
u7 = 0010111 has 4 runs in total 00, 1, 0, 111. Thus, D1(u7) = 4. It is easy to see why this is true: Any
deletion in a run essentially leads to the same n − 1 length sequence in the descendant ball.
Definition 2 We call a subset C ⊂ Fn
De(v) = φ
2 to be a e-error correction code, if for any u, v ∈ C, u 6= v, De(u) ∩
We will mainly deal with e = 1 during the talk.
2.1 Repetition coding
Lets start with the simplest of the 1-deletion correction codes you can think of: "repetition codes". We
repeat every symbol 2 times. Is that sufficient? Let us say, we observe some symbol an odd number of times,
then we know that a symbol was deleted in that runs of 0's or 1's. Note that, we still cannot figure out the
location of the deletion, but can figure out what was deleted.
This idea can in fact be extended to correct e deletions by repeating every symbol e + 1 times. But, this
is quite bad: For correcting 1 error, our communication rate is 0.5. Still better that the BSC case, where we
had to repeat things 3 times.
Cool! Can we do better? We will next look at a cool class of codes known as the VT-codes; or the
Varshamov-Tenengolts code. But before doing let's look at a puzzle.
2.2 A puzzle
I believe the puzzle has some connection with VT-codes, might help with the understanding; but if not, it
is a cool simple puzzle! So lets say, Mary is the Queen of the seven kingdoms, and she had ordered 100
big barrels filled with gold coins, where each coin has a weight of 10gm. But, she knows from her secret
agency that one of the barrels contains counterfeit coins weighing only 9gm. She has an electric weighing
scale which she can use; so the question is:
How can she determine which barrel contains the counterfeit coins with a single measurement
The solution is simple: she takes j coins from the jth barrel and places them on the electronic weighing
scale. Now if the weight is less that expected by r grams, then it is the rth barrel which is counterfeit! We
will come back to this puzzle :)
3 VT-codes
Allright! We are all set to define VT-codes.
Definition 3 Varshamov-Tenengolts code V Ta(n) ⊂ Fn
2 is defined as:
V Ta(n) =(xn
n
Xi=1
ixi ≡ a mod (n + 1))
(4)
Some historical context on these codes: These codes were first proposed as error correction codes for 1-bit
Z-channel error! Which means essentially 1 → 0, but not the other way round. Z-channel errors are known
as asymmetric bit-flips. Varshamov and Tenengolts proposed these codes in 1965 [14], and then Levenshtein
discovered that these codes in fact work well also for the deletion channel as well!
Z-channel correction
So before we look into 1 deletion correction, let us see how they can correct one Z-channel error, i.e. one of
the 1's can flip to a 0.
Let xn be the received symbol: then we can still compute:
S = a −
ixi
n
Xi=1
3
In case, there is a 1 → 0 flip at position j, then S = j. Thus, we can correct the Z-channel error! Here is
where the similarity with the puzzle can be seen.
3.1 VT-codes decoding
We are all set to discuss the decoding for deletion channel:
1. First of all, note that if it is only a deletion channel, then "error detection" comes for free from the
length of the code, unlike the bit-flipping error
2. Let yn−1 be the received erroneous codeword after 1 deletion. Define:
n−1
w =
Xi=1
S = a −
yi
n−1
Xi=1
iyi
(5)
(6)
3. Let p be the position at which deletion occurred, as in xp was deleted. Let L0, L1 be the counts of 1,0
to the left of position p. and R0, R1 be the counts to the right. In that case: S = R1 if xp = 0 and
S = p + R1 if xp = 1.
S = p + R1
= L0 + L1 + 1 + R1
= w + L0 + 1
Thus, as R1 < w + 1, if S ≤ w, 0 is deleted at a position with R1 1's to its right, otherwise, 1 is deleted
such that there are L0 = S − w − 1 zeros to its left.
4. Note that we can thus uniquely determine the xn sequence, but we cannot determine the exact location
of the deletion here, as it can be any 0 or 1 in the run we identified.
This beautiful decoding algorithm was given by Levenshtein in his 1965 work [8]. Also most of the things
which I will talk are from the survey by Sloane [12].
3.2 VT-codes rate
As the VT-codes are combinatorial, we can get exact formulae for their sizes. Exact combinatorial formulate
can be found in [12]. Here are few interesting things:
1. Lemma 2 For some a ∈ [0, n]:
As every xn ∈ Fn
V Ta(n) ≥
1 lies in exactly one of V Ta(n) sets:
2n
n + 1
This leads to the lemma.
n
Xa=0
V Ta(n) = 2n
2. It can be in fact shown that a = 0 leads to the largest code size, and a = 1 the smallest size.
3. For n = 2q − 1, all the VT-code sizes are in fact equal to 2n
n+1 .
V T0(n) ≤ V Ta(n) ≤ V T1(a)
4
3.3 Optimality of VT-codes
We say a e-error correction code is "optimal" if it has the smallest size amongst all e-error correction codes.
Let us analyze the "optimality" of VT-codes.
1. Levenshtein [8] showed that, optimal 1-deletion correction codes have asymptotic sizes ≈ 2n
n . This
makes VT-codes asymptotically optimal.
2. People have not been able to prove that VT-codes are optimal non-asymptotically. Finding the "op-
timal" 1-deletion code is a NP-hard problem, as it involves finding the independent set on the graph
where vertices are connected if they lead to the same deletion descendants [12]. But for n ≤ 8 it is
known that they are optimal, using computer programs. Sizes of these codes are:
1, 2, 2, 4, 6, 10, 16, 30
For higher n due to the exponential nature of the algorithms, we cannot say anything yet.
3. VT-codes also have the property that they are "perfect codes" [9] which implies that their descendent
sets which are disjoint cover the entire 2n−1 sized. For example:
V T0(3) = {000, 101}
Descendants = {00, 10, 01, 11}
Levenshtein showed that surprisingly, this is true for all a, which is quite cool in itself! The perfect
codes analogy comes from Hamming codes being Perfect. But, unlike the Hamming distortion case,
here perfect codes does not imply optimal codes? Why? For example, code {000, 111} is not perfect
but is perhaps a better code that V T0(3), potentially there might be a larger code for larger n. Why
does this happen? Because, the number of descendants are not fixed, some have 1 and some have more.
4. Linearity: VT codes are linear until n = 4 but never after! [12]. Variants of VT-code (restrictions on
VT-codes) can be made linear by considering redundancy to be √n as against log n. Althought I am
not sure how is linearity of codes useful, if the decoding is still non-linear (linear time complexity, but
non-linear in nature).
3.4 Systematic Encoding
Now that we have taken a look at the linear-time decoding of VT-codes, it is a natural question to ask if
there exist a nice way to encode data. This problem surprisingly remained open for more than 30 years until
1998 when Abdel-Gaffar et al. [1] provided a very convenient way of in fact "systematic encoding" of data.
1. For n = 2q − 1, let m = 2q − q − 1 be the number of data bits, and q are the "parity" bits. Let the
data bits be b1, b2, . . . , bm, and the code-word to be formed xn in V Ta(n).
2. Fill in the data except in positions 1, 2, 22, 23, . . . , 2q−1. Thus xn codeword looks like:
x1, x2, b1, x4, b2, b3, b4, x8, . . . , bm
We can compute: S = a−P2q
are decided, to obtain S = 0, we need:
−1
i=1 ixi mod (2q). As most of the positions of xn (except x2j , j ∈ [0, q−1])
2jx2j mod (2q) = a
q−1
Xj=0
We can now conveniently choose x2j , j ∈ [0, q − 1] as the q-bit binary expansion of a.
3. Note that m = 2q − q − 1, n = 2q − 1 for 1-deletion correction is exactly same as the rate of Hamming
code. Not sure if this is a coincidence or something more!
5
4
Insertion + Deletion + Substitution codes
We looked at 1-deletion correction channels in depth. In the next part of the tutorial, we will extend this
understanding to more general scenarios. The first scenario is 1 insertion instead of 1 deletion error.
4.1 General indel error codes
Levenshtein [8] showed this general lemma:
Lemma 3 Any s-deletion correction code can also correct s1 deletions and s2 insertion errors where s1+s2 ≤
s.
Note that here we do not need to know s1, s2 beforehand. The general proof is quite simple. Here, we will
prove the simpler version of 1-insertion error, as that is sufficient to get an intuitive understanding.
1. Let us assume that C is a 1-deletion correction code. We want to show that C can correct 1-insertion
errors as well.
2. Let us assume that on the contrary, there exist codewords xn, yn ∈ C such that after one insertion
error in them the resulting noisy codewords xn+1, yn+1 are equal. Let the insertion occured at position
i in xn+1 and at position j in yn+1.
3. As xn+1 = yn+1, even after deleting symbols in position i, j in both x, y, results in n− 1 vectors which
are equal. However, the n − 1 length codewords are in fact 1-deletion descendants of the codewords
xn, yn ∈ C. This is contradictory to the definition of deletion correction codes, as no descendants can
be equal. Thus, C has to be 1-insertion correction as well.
Note that, this is more of an existential result, and efficient deletion error decoding algorithms might not
translate into efficient insertion detection algorithms.
4.2 VT-codes for 1-insertion, deletion, substitution correction
Levenshtein showed the surprising fact that with a simple modification standard VT-codes can be converted
into 1-insertion, deletion and substitution correction codes. The modification is as follows:
V T a(n) ⊂ Fn
2 is defined as:
V T a(n) =(xn
n
Xi=1
ixi ≡ a mod (2n + 1))
(7)
Let us try to understand why V T a(n) work:
1. First of all, from the length of the code, we know whether there is an insertion, deletion or a substitution.
2. Recollect that deletion error correction in V Ta(n) codes only depends on distinct remainders modulus
(n + 1). This should still hold true if the modulus is taken m ≥ n + 1. Thus, with m = 2n + 1,
1-deletion correction still holds.
3. 1-insertion code ability was already shown from the general lemma earlier. However, using a similar
remainder trick mod 2n + 1, insertions can in fact be corrected efficiently using the V T a(n) codes.
4. The only case remaining to analyze is the 1-substitution or 1-bitflip case. Let 0 → 1 at position
p ∈ [1, n] in the codeword xn resulting in the noisy codework xn. Clearly:
ixi mod (2n + 1)
n
S = a −
Xi=1
= 2n − p
6
If 1 → 0, then:
n
Xi=1
S = a −
= p
ixi mod (2n + 1)
As all these S values are distinct, we can correct for 1-bitflip. Note that, this construction is not
optimal just for 1-bitflip, as it essentially encodes 1 bit less than Hamming codes.
5 VT-codes for larger alphabet
Creating deletion codes for larger alphabets becomes a bit tricky. Of course, repetition coding still works on
non-binary alphabets.
Code which does not work
When I started thinking about this problem, I came up with this code, which has a bug! Let us still take a
look at it, as it gives some understanding as to the intricacies of code-design:
Define βi for i ∈ [2, n] as:
Then we consider code as:
βi = 1, xi 6= 0
= 0, otherwise
C =(cid:26)xn
n
n
Xi=1
Xj=1
iβi = a mod (n + 1),
xi = b mod (q)(cid:27)
Let us look at the argument as to why the code works, and try to find the bug!
Argument: The first equation, similar to the binary VT code will tell us the position of the deletion, and
the second equation tells us the value.
Why does the above argument not work. The reason is that binary VT-codes not not actually tell us the
position of the deletion correctly. They can tell us in which "run" the deletion happened, and hence obtain
the codeword correctly, but not the exact position.
How do we solve for that?
Code which works
This code appears in the work of Tenengolts in 1984. [13]
Define αi for i ∈ [2, n] as:
αi = 1, xi ≥ xi−1
= 0, xi < xi−1
Then we consider the code:
C =(cid:26)xn
n
n
Xi=2
(i − 1)αi = a mod (n),
xi = b mod (q)(cid:27)
Xj=1
7
Let us try to analyze the decoding for this code:
1. First of all, as in the previous code, from the second constraint Pn
what is the value of the deleted symbol. What remains to determine is its position.
j=1 xi = b mod (q), we figure out
2. The α sequence is essentially capturing the monotonous regions of the xn sequence. αi = 1, when the
sequence is increasing (non-decreasing), and 0 when it is decreasing. Thus, a deletion in xn sequence
will in fact lead to exactly 1 deletion in the α sequence (the position of deletion might be shifted by 1,
but it does not matter as VT-codes do not correct for position).
3. As the first equation is a VT-code, it can determine the run in which the deletion occurred. As every
run in α sequence corresponds to monotonic increasing/decreasing subsequence in xn, from the value
of the deleted symbol, we can correctly place it and complete the decoding!
Efficient systematic encoding for this code was discovered recently in a work by Abroshan et al. [2] and
is included in the implementation at https://github.com/shubhamchandak94/VT_codes/.
6 Bursty deletion codes
In this section we will look at Bursty deletions. By a single bursty deletion of size s, we mean that some
consecutive s symbols were deleted. Note that, s-bursty deletion correction codes can correct for exactly 1
burst of size s, but surprisingly they need not correct for a burst of size s− 1. For example: C = {0101, 1010}
can correct 2-bursty errors, but not single deletion errors!
6.1 VT-code based construction
We consider a construction based on single-deletion correction VT-codes to correct a single burst of s
deletions. How should one do that? One simple trick is to distribute these s deletions across the n length
sequence, so that each n/s length subsequence has exactly 1 deletion. For simplicity, let n/s = k.
Then the codeword xn has the property that each of the s rows below, belong to a V Ta(n) code.
x1, xs+1, x2s+1, . . . , xn−s+1
x2, xs+2, x2s+2, . . . , xn−s+2
...
xs, x2s, x2s+2, . . . , xn
1. Let us analyze the scenario of 1 bursty error of size s from position xj, . . . , xj+s−1 are deleted. This
corresponds to exactly 1 deletion in each of the rows.
2. We still need to figure out which symbols of the deleted codeword belong to which rows, as the alignment
might no longer be true. The cool thing is that, if there are exactly s consecutive deletions, then every
sth symbol will still be correctly aligned to the rows.
3. Thus, our code can in fact correct a bursty error of s deletions, but need not correct lower number of
bursts, which is quite unusual!
One important caveat to observe here is that the position of deletion in each of the rows is the same, or
shifted by 1. Thus, if one of the rows is a VT-code, and the other rows just tell whether the error position is
odd or even, that is enough to resolve the bursty error. This observation is the basis of further improvements
to bursty error correction. For more details take a look at this paper: [11].
8
7 Multiple Deletions
One would imagine that it should be possible to extend the elegant construction of VT-codes from single
deletion correction to multiple deletions. However, this problem has proved to be much more difficult.
There have been some recent works which extend the single deletion errors to multiple errors. Gabrys
et al [6] provide an extension of the VT-codes idea to correct two deletions. There have been other recent
works which provide multiple deletion correcting codes using different (non VT-code based) ideas [3].
Acknowledgement
I would like to thank Jay Mardia and Mary Wootters for interesting discussions on deletion codes.
References
[1] Khaled AS Abdel-Ghaffar and Hendrik C Ferreira. Systematic encoding of the Varshamov-Tenengol'ts
codes and the Constantin-Rao codes. IEEE Transactions on Information Theory, 44(1):340 -- 345, 1998.
[2] Mahed Abroshan, Ramji Venkataramanan, and Albert Guillen I Fabregas. Efficient systematic encoding
of non-binary vt codes. In 2018 IEEE International Symposium on Information Theory (ISIT), pages
91 -- 95. IEEE, 2018.
[3] Joshua Brakensiek, Venkatesan Guruswami, and Samuel Zbarsky. Efficient low-redundancy codes for
correcting multiple deletions. IEEE Transactions on Information Theory, 64(5):3403 -- 3410, 2017.
[4] Mahdi Cheraghchi. Capacity upper bounds for deletion-type channels. Journal of the ACM (JACM),
66(2):9, 2019.
[5] Eleni Drinea and Michael Mitzenmacher. Improved lower bounds for the capacity of iid deletion and
duplication channels. IEEE Transactions on Information Theory, 53(8):2693 -- 2714, 2007.
[6] Ryan Gabrys and Frederic Sala. Codes correcting two deletions. IEEE Transactions on Information
Theory, 65(2):965 -- 974, 2018.
[7] Adam Kalai, Michael Mitzenmacher, and Madhu Sudan. Tight asymptotic bounds for the deletion
In 2010 IEEE International Symposium on Information
channel with small deletion probabilities.
Theory, pages 997 -- 1001. IEEE, 2010.
[8] Vladimir I Levenshtein. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet
physics doklady, pages 707 -- 710, 1966.
[9] Vladimir I Levenshtein. On perfect codes in deletion and insertion metric. Discrete Mathematics and
Applications, 2(3):241 -- 258, 1992.
[10] Michael Mitzenmacher et al. A survey of results for deletion channels and related synchronization
channels. Probability Surveys, 6:1 -- 33, 2009.
[11] Clayton Schoeny, Antonia Wachter-Zeh, Ryan Gabrys, and Eitan Yaakobi. Codes correcting a burst of
deletions or insertions. IEEE Transactions on Information Theory, 63(4):1971 -- 1985, 2017.
[12] Neil JA Sloane. On single-deletion-correcting codes. Codes and designs, 10:273 -- 291, 2000.
[13] Grigory Tenengolts. Nonbinary codes, correcting single deletion or insertion (Corresp.). IEEE Trans-
actions on Information Theory, 30(5):766 -- 769, 1984.
[14] RR Varshamov and GM Tenengolts. Codes which correct single asymmetric errors (in Russian). Au-
tomatika i Telemkhanika, 161(3):288 -- 292, 1965.
9
|
1506.08518 | 2 | 1506 | 2015-12-22T08:31:43 | Fast Computation of Abelian Runs | [
"cs.DS"
] | Given a word $w$ and a Parikh vector $\mathcal{P}$, an abelian run of period $\mathcal{P}$ in $w$ is a maximal occurrence of a substring of $w$ having abelian period $\mathcal{P}$. Our main result is an online algorithm that, given a word $w$ of length $n$ over an alphabet of cardinality $\sigma$ and a Parikh vector $\mathcal{P}$, returns all the abelian runs of period $\mathcal{P}$ in $w$ in time $O(n)$ and space $O(\sigma+p)$, where $p$ is the norm of $\mathcal{P}$, i.e., the sum of its components. We also present an online algorithm that computes all the abelian runs with periods of norm $p$ in $w$ in time $O(np)$, for any given norm $p$. Finally, we give an $O(n^2)$-time offline randomized algorithm for computing all the abelian runs of $w$. Its deterministic counterpart runs in $O(n^2\log\sigma)$ time. | cs.DS | cs | Fast Computation of Abelian Runs
Gabriele Ficia,1, Tomasz Kociumakab,2, Thierry Lecroqc, Arnaud Lefebvrec, ´Elise Prieur-Gastonc
aDipartimento di Matematica e Informatica, Universit`a di Palermo, Italy
bFaculty of Mathematics, Informatics and Mechanics, University of Warsaw, Poland
cNormandie Universit´e, LITIS EA4108, NormaStic CNRS FR 3638, IRIB, Universit´e de Rouen, 76821 Mont-Saint-Aignan
Cedex, France
5
1
0
2
c
e
D
2
2
]
S
D
.
s
c
[
2
v
8
1
5
8
0
.
6
0
5
1
:
v
i
X
r
a
Abstract
Given a word w and a Parikh vector P, an abelian run of period P in w is a maximal occurrence of a
substring of w having abelian period P. Our main result is an online algorithm that, given a word w of
length n over an alphabet of cardinality σ and a Parikh vector P, returns all the abelian runs of period P
in w in time O(n) and space O(σ + p), where p is the norm of P, i.e., the sum of its components. We also
present an online algorithm that computes all the abelian runs with periods of norm p in w in time O(np),
for any given norm p. Finally, we give an O(n2)-time offline randomized algorithm for computing all the
abelian runs of w. Its deterministic counterpart runs in O(n2 log σ) time.
Keywords: Combinatorics on Words, Text Algorithms, Abelian Period, Abelian Run
1. Introduction
Computing maximal (non-extendable) repetitions in a word is a classical topic in the area of string
algorithms (see for example [1] and references therein). Maximal repetitions of substrings, also called runs,
give information on the repetitive regions of a word, and are used in many applications, for example in the
analysis of genomic sequences.
Kolpakov and Kucherov [2] gave the first linear-time algorithm for computing all the runs in a word and
conjectured that any word of length n contains less than n runs. Recently, Bannai et al. [3, 4], using the
notion of Lyndon roots of a run, proved this conjecture and designed a much simpler algorithm computing
the runs.
Here we deal with a generalization of this problem to the commutative setting. Recall that an abelian
power is a concatenation of two or more words that have the same Parikh vector, i.e., that have the same
number of occurrences of each letter of the alphabet. For example, aababa is an abelian square, since aab
and aba both have two a's and one b, i.e., the same Parikh vector P = (2, 1). When an abelian power
occurs within a word, one can search for its "maximal" occurrence by extending it to the left and to the
right character by character without violating the condition on the number of occurrences of each letter.
Following the approach of Constantinescu and Ilie [5], we say that a Parikh vector P is an abelian period of
a word w if w can be written as w = u0u1 · · · uk−1uk for some k ≥ 1 where for 0 < i < k all the ui's have
the same Parikh vector P and the Parikh vectors of u0 and uk are contained in P. If k > 2, we say that the
Email addresses: [email protected] (Gabriele Fici), [email protected] (Tomasz Kociumaka),
[email protected] (Thierry Lecroq), [email protected] (Arnaud Lefebvre),
[email protected] (´Elise Prieur-Gaston)
1Supported by the Italian Ministry of Education (MIUR) project PRIN 2010/2011 "Automi e Linguaggi Formali: Aspetti
Matematici e Applicativi".
2Supported by the Polish Ministry of Science and Higher Education under the 'Iuventus Plus' program in 2015-2016 grant
no 0392/IP3/2015/73. The author is also supported by Polish budget funds for science in 2013-2017 as a research project under
the 'Diamond Grant' program.
To appear in Theoretical Computer Sicence
August 16, 2018
word w is periodic with period P. Note that the factorization above is not necessarily unique. For example,
a · bba · bba · ε and ε · abb · abb · a (ε denotes the empty word) are two factorizations of the word abbabba both
corresponding to the abelian period (1, 2). Moreover, the same word can have different abelian periods.
In this paper we define an abelian run of period P in a word w as an occurrence of a substring v of w
such that v is periodic with abelian period P and this occurrence cannot be extended to the left nor to the
right by one letter into a substring periodic with period P.
For example, let w = ababaaa. Then the prefix ab · ab · a = w[0. . 4] has abelian period (1, 1) but it is not
an abelian run since the prefix a · ba · ba · a = w[0. . 5] also has abelian period (1, 1). The latter, on the other
hand, is an abelian run of period (1, 1) in w.
Looking for abelian runs in a word can be useful to detect regions in the word where there is some kind
of non-exact repetitiveness, for example regions with several consecutive occurrences of a substring or its
reversal.
Matsuda et al. [6] recently presented an offline algorithm for computing all abelian runs of a word of
length n in O(n2) time. Notice that, however, the definition of abelian run in [6] is slightly different from
the one we consider here. We compare both versions in Section 2. Basically, our notion of abelian run is
more restrictive than the one of [6], for which we use the term "anchored run".
We first present an online algorithm that, given a word w of length n over an alphabet of cardinality
σ and a Parikh vector P, returns all the abelian runs of period P in w in time O(n) and space O(σ + p),
where p is the norm of P, that is, the sum of its components. This algorithm improves upon the one given
in [7] which runs in time O(np). Next, we give an O(np)-time online algorithm for computing all the abelian
runs with periods of norm p of a word of length n, for any given p. Finally, we present an O(n2) (resp.
O(n2 log n)) -time offline randomized (resp. deterministic) algorithm for computing all the abelian runs of
a word of length n.
The rest of this article is organized as follows. Sect. 2 introduces central concepts and fixes the notation.
In Sect. 3 we review the results on abelian runs given in [6]. Sect. 4 is devoted to the presentation of our
main result: a new solution for computing the abelian runs for a given Parikh vector. In Sect. 5 we apply
this algorithm in a procedure for computing the abelian runs with periods of a given norm. Next, in Sect. 6,
we design a solution for computing all the abelian runs, which builds upon the result recalled in Sect. 3.
Finally, we conclude in Sect. 7.
2. Definitions and Notation
Let Σ = {a1, a2, . . . , aσ} be a finite ordered alphabet of cardinality σ, and let Σ∗ be the set of finite words
over Σ. We assume that the mapping between ai and i can be evaluated in constant time for 1 ≤ i ≤ σ. We
let w denote the length of the word w. Given a word w = w[0. . n − 1] of length n > 0, we write w[i] for the
(i + 1)-th symbol of w and, for 0 ≤ i ≤ j < n, we write w[i. . j] to denote a fragment of w from the (i + 1)-th
symbol to the (j + 1)-th symbol, both included. This fragment is an occurrence of a substring w[i] · · · w[j].
For 0 ≤ i ≤ n, w[i. . i − 1] denotes the empty fragment. We let wa denote the number of occurrences of
the symbol a ∈ Σ in the word w. The Parikh vector of w, denoted by Pw, counts the occurrences of each
letter of Σ in w, that is, Pw = (wa1 , . . . , waσ ). Notice that two words have the same Parikh vector if and
only if one word is a permutation (i.e., an anagram) of the other. Given the Parikh vector Pw of a word
w, we let Pw[i] denote its i-th component and Pw its norm, defined as the sum of its components. Thus,
for w ∈ Σ∗ and 1 ≤ i ≤ σ, we have Pw[i] = wai and Pw = Pσ
i=1 Pw[i] = w. Finally, given two Parikh
vectors P, Q, we write P ⊆ Q if P[i] ≤ Q[i] for every 1 ≤ i ≤ σ. If additionally P 6= Q, we write P ⊂ Q
and say that P is contained in Q.
Definition 1 (Abelian period [5]). A factorization w = u0u1 · · · uk−1uk satisfying k ≥ 1, Pu1 = · · · =
Puk−1 = P, and Pu0 ⊂ P ⊃ Puk is called a periodic factorization of w with respect to P. If a word w
admits such a factorization, we say that P is an (abelian) period of w.
We call fragments u0 and uk respectively the head and the tail of the factorization, while the remaining
factors are called cores. Note that the head and the tail are of length strictly smaller than P; in particular
they can be empty.
2
i
j
j − i − h − t + 1
h
P
P
P
t
Figure 1: The tuple (i, h, t, j) denotes an occurrence of a substring starting at position i, ending at position j, and having
abelian period P with head length h and tail length t.
Observe that a periodic factorization with respect to a fixed period is not unique. However, it suffices
to specify u0 to indicate a particular factorization; see [5]. Dealing with factorizations of fragments of a
fixed text, it is more convenient to use a different quantity for this aim. Suppose w[i. . j] = u0 · · · uk is a
factorization with respect to abelian period P with p = P. Observe that consecutive starting positions
i1, . . . , ik of factors u1, . . . , uk differ by exactly p. Hence, they share a common remainder modulo p, which
we call the anchor of the factorization. Note that the anchor does not change if we trim a factorization of
w[i. . j] to a factorization of a shorter fragment, or if we extend it to a factorization of a longer fragment.
Definition 2 (Anchored period). A fragment w[i. . j] has an abelian period P anchored at k if it has a
periodic factorization with respect to P whose anchor is k mod p.
If w has a factorization with at least two cores, we say that w is periodic with period P (anchored at k
if k mod p is the anchor of the factorization).
Definition 3 (Abelian run). A fragment w[i. . j] is called an abelian run with period P if it is periodic
with period P and maximal with respect to this property (i.e., each of w[i − 1. . j] and w[i. . j + 1] either
does not exist or it is not periodic with period P).
We shall often represent an abelian run w[i. . j] as a tuple (i, h, t, j) where h and t are respectively the lengths
of the head and the of the tail of a periodic factorization of w[i. . j] with period P and at least two cores
(see Figure 1). Note that (i + h) mod p is the anchor of the factorization, and that 1
p (j − i − h − t − 1) is
the number of cores, in particular it is an integer.
Observe that an abelian run with period P may have several valid factorizations. For example, a·ba·ba·ε
and ε · ab · ab · a are factorization of a run w[0..4] with period P = (1, 1) in w = ababa. Therefore the run
can be represented as (0, 1, 0, 4) and as (0, 0, 1, 4). However, in v = abab only (0, 0, 0, 3) is a representation
of v[0..3] as an abelian run with period P = (1, 1). This is because (0, 1, 1, 3) corresponds to a factorization
v[0..3] = a · ba · b with one core only, and such a factorization does not indicate that v[0..3] is an abelian run.
Matsuda et al. [6] gave a different definition of abelian runs, where maximality is with respect to extending
a fixed factorization. In this paper, we call such fragments anchored (abelian) runs.
Definition 4 (Anchored run [6]). A fragment w[i. . j] is a k-anchored abelian run with period P if w[i. . j]
is periodic with period P anchored at k and maximal with respect to this property (i.e., each of w[i − 1. . j]
and w[i. . j + 1] either does not exist or it is not periodic with period P anchored at k).
Note that every abelian run is an anchored run with the same period (for some anchor). The converse
is not true, since it might be possible to extend an anchored run preserving the period but not the anchor.
For example, in the word w = ababaaa considered in the introduction, the fragment w[0. . 4] = ε · ab · ab · a is
a 0-anchored run but not an abelian run, since w[0. . 5] = a · ba · ba · a is periodic with abelian period (1, 1).
Since a factorization is uniquely determined by the anchor, standard inclusion-maximality is equivalent
to the condition in Definition 4.
Observation 5. Let w[i. . j] and w[i′. . j ′] be fragments of w with abelian period P anchored at k. If w[i. . j]
is properly contained in w[i′. . j ′] (i.e, i′ < i and j ≤ j ′, or i′ ≤ i and j < j ′), then w[i. . j] is not a k-anchored
abelian run with period P.
3
Abelian runs enjoy the same property, but its proof is no longer trivial.
Lemma 6. Let w[i. . j] and w[i′. . j ′] be fragments of w with abelian period P. If w[i. . j] is properly contained
in w[i′. . j ′], then w[i. . j] is not an abelian run with period P.
Proof. We assume that i′ < i. The case of j < j ′ is symmetric. For a proof by contradiction suppose that
w[i. . j] is an abelian run and let w[i. . j] = u0 · · · uk be a periodic factorization with period P and at least
two cores (i.e., satisfying k ≥ 3). A periodic factorization of w[i′. . j ′] can be trimmed to a factorization
w[i − 1. . j] = v0 · · · vℓ. However, since w[i. . j] is an abelian run, this factorization must have at most one
core (i.e., ℓ ≤ 2). Moreover, u0 · · · uk cannot be extended to a factorization of w[i − 1. . j] = u′
0u1 · · · uk. In
other words u′
0, the extension of u0 by one letter to the left, must satisfy Pu′
0 6⊆ P.
Let p = P. The conditions on the number of cores imply u0+2p ≤ w[i. . j] and w[i−1]. . j] < v0+2p.
0 ⊂ Pv0 ⊂ P, which is in
(cid:3)
0 is a proper prefix of v0. This yields Pu′
0 = u0 + 1 < v0, i.e., u′
Consequently, u′
contradiction with Pu′
0 6⊆ P.
Corollary 7. Let w be a word. For a fixed Parikh vector P, there is at most one abelian run with abelian
period P starting at each position of w.
3. Previous Work
Matsuda et al. [6] presented an algorithm that computes all the anchored runs of a word w of length n in
O(n2) time and space complexity. The initial step of the algorithm is to compute maximal abelian powers
in w. Recall that an abelian power is a concatenation of several abelian-equivalent words. In other words,
an abelian power of period P is a word admitting a periodic factorization with respect to P with an empty
head, an empty tail and at least two cores. A fragment w[i. . j] is a maximal abelian power if it cannot be
extended to a longer power of period P (preserving the anchor). Formally, the maximality conditions are
1. Pw[i−p..i−1] 6= Pw[i..i+p−1] or i − p < 0, and
2. Pw[j−p+1..j] 6= Pw[j+1..j+p] or j + p ≥ n,
where p = P.
The approach of [6] is to first compute all the abelian squares using the algorithm by Cummings &
Smyth [8]. The next step is to group squares into maximal abelian powers. For this, it suffices to merge
pairs of overlapping abelian squares of the form w[i. . i + 2p − 1] and w[i + p. . i + 3p − 1]. This way maximal
abelian powers are computed in O(n2) time.
Observe that there is a natural one-to-one correspondence between maximal abelian powers and anchored
runs: it suffices to trim the head and the tail of the factorization of an anchored run to obtain a maximal
abelian power. Hence, the last step of the algorithm is to compute the maximal head and tail by which each
abelian power can be extended. This could be done naively in O(n3) time overall, but a clever computation
enables to find all the abelian runs in time and space O(n2) (see [6] for further details).
In Section 6, we extend this result to compute the abelian runs only rather than all the anchored runs.
Both these algorithms work offline: they need to know the whole word before reporting any abelian run.
In the following two sections we give several online algorithms, which are able to report a run ending at
position i − 1 of a word w before reading w[i + 1] and the following letters. Clearly, not knowing w[i] one
cannot decide whether the run could be extended to the right, so this is the optimal delay. However, these
methods are restricted to finding runs of a given period or a given norm of the periods, respectively.
4. Computing Abelian Runs with Fixed Parikh Vector
In this section we present our online solution for computing all the abelian runs of a given Parikh vector
P of norm p in a given word w. The algorithm works in O(n) time and O(σ + p) space where n = w.
First, in Sect. 4.1, we show how to compute all anchored runs of period P. Later, in Sect. 4.2, we modify
the algorithm to return abelian runs only. We conclude in Sect. 4.3 with an example course of actions in
our solution.
4
w
i − p
bi
i
impossible tail
6⊆ P
Figure 2: Bi[k] = ∞ for i − p + 1 < k < bi.
4.1. Algorithm for Anchored Runs
We begin with a description of data maintained while scanning the string w. For an integer k, let Bi[k]
be the starting position of the longest suffix of w[0. . i] which has period P anchored at k. If there is no such
a suffix, we set Bi[k] = ∞. Since this notion depends on k mod p only, we store Bi[k] for 0 ≤ k < p only.
Let bi be the starting position of the longest suffix of w[0. . i] whose Parikh vector is contained in or equal
to P. In other words, we have Pw[bi..i] ⊆ P and Pw[bi−1..i] 6⊆ P (or bi = 0). Note that bi = i + 1 if w[i] does
not occur in P.
Observe that the tail of any periodic factorization of a suffix of w[0. . i] must be contained in w[bi. . i].
This leads to the following characterization:
Lemma 8. Let 0 ≤ i < w. We have Bi[k] ≤ k for bi ≤ k ≤ i + 1 and Bi[k] = ∞ for i − p + 1 < k < bi.
Proof. For bi ≤ k ≤ i + 1, the fragment w[k. . i] has abelian period P anchored at k. (The underlying
factorization has empty head, no cores and tail w[k. . i], unless k = bi = i − p + 1, when the factorization
has one core, empty head and empty tail). Hence, we have Bi[k] ≤ k directly from the definition.
For i − p + 1 < k < bi, the tail of the factorization with anchor k mod p would need to start at position
(cid:3)
k, which is impossible (see Figure 2).
The values bi−1 and bi are actually sufficient to describe Bi based on Bi−1.
Lemma 9. For 0 ≤ i < w the following equalities hold:
1. Bi[k] = ∞ 6= Bi−1[k] for max(i − p + 1, bi−1) ≤ k < bi,
2. Bi[k] = Bi−1[k] for bi ≤ k ≤ i and for i − p + 1 < k < bi−1,
3. Bi[i + 1] = bi if bi > i − p + 1 and Bi[i + 1] = Bi−1[i − p + 1] otherwise.
Proof. Lemma 8 implies that Bi[k] = ∞ for i − p + 1 < k < bi and Bi−1[k] = ∞ for i − p + 1 < k < bi−1
(hence Bi−1[k] = Bi[k] in this latter case). For bi ≤ k < i, we have Pw[k..i] ⊆ P, so we can extend the
factorization of a suffix of w[0. . i − 1] whose tail starts at position k (see Figure 3).
Finally, note that Bi[i + 1] is the starting position of the maximal suffix of w[0. . i] with an empty-tail
periodic factorization. If Pw[i−p+1..i] 6= P (i.e., if bi > i − p + 1), this is just w[bi. . i]. Otherwise, we can
extend the factorization of a suffix of w[0. . i − 1] whose tail starts at position i − p + 1.
(cid:3)
Having read letter w[i], we need to report anchored runs which end at position i − 1. For this, we use
the following characterization.
Lemma 10. Let i − p < k ≤ i. A fragment w[b. . i − 1] is a k-anchored run with period P if and only if
Bi−1[k] = b ≤ k − 2p and Bi[k] > b.
Proof. Clearly an anchored run ending at position i − 1 must be a left-maximal suffix of w[0. . i − 1] with
a given anchor. Moreover, we must have b ≤ k − 2p so that the factorization has at least two cores and
Bi[k] > b due to right-maximality. It is easy to see that these conditions are sufficient.
(cid:3)
5
By Lemma 9, most entries of Bi are inherited from Bi−1, so we use a single array B and having read
w[i], we update its entries. As evident from Lemma 10, each anchored run to be reported corresponds to a
modified entry.
The algorithm AnchoredRun(P, p, w, n) in Figure 4 implements our approach. The while loop incre-
ments k from bi−1 to bi. For k > i − p, we set B[k] to ∞ and possibly report a run. Note that k = i − p + 1
is within the scope of Case 3 rather than Case 1 in Lemma 9. However, later we set B[i + 1] to bi if
bi > i − p + 1 (as described in Case 3). Nevertheless, if an (i + 1)-anchored run needs to be reported, we
have Bi−1[i − p + 1] < ∞ = Bi[i − p + 1], so bi−1 ≤ i − p + 1 and thus k = i − p + 1 is considered in the loop.
Theorem 11. The algorithm AnchoredRun(P, p, w, n) computes all the anchored runs with period P of
norm p in a word w of length n in time O(n) and additional space O(σ + p).
Proof. The correctness of the algorithm comes from Lemmas 9-10 and the discussion above. The external
for loop in lines 3 -- 14 runs n + 1 times. The internal while loop in lines 4 -- 12 cannot iterate more than n + 1
times since it starts with k equal to 0 and ends when k is equal to n and k can only be incremented by 1
(in line 12). The test Pw[k..i] 6⊆ P in line 4 can be realized in constant time once we store Pw[k..i] and a
counter of its components for which the value is greater than in P. This data needs to be updated once we
increment i in the for loop and k in line 12. We then need to increment the component w[i] or decrement
the component w[k] of Pw[k..i], respectively. The global counter needs to be updated accordingly. All the
w
Bi−1[k]
bi−1
kbi
i
⊆ P
fragment
ending at i − 1
fragment
ending at i
· · ·
· · ·
Figure 3: Bi[k] = Bi−1[k] for bi ≤ k ≤ i.
AnchoredRun(P, p, w, n)
1 k ← 0
2 B[0] ← k
3 for i ← 0 to n do
4
while k ≤ n and (i = n or Pw[k..i] 6⊆ P) do
5
6
7
8
9
10
11
12
13
14
if k > i − p then
b ← B[k mod p]
B[k mod p] ← ∞
if b ≤ k − 2p then
h ← (k − b) mod p
t ← i − k
Output(b, h, t, i − 1)
¯ ¯k ← k + 1
¯if k > i − p + 1 then
B[(i + 1) mod p] ← k
¯ ¯
Figure 4: Algorithm computing all the anchored runs of period P of norm p in a word w of length n.
6
other operations run in constant time. Thus the total time complexity of the algorithm is O(n). The space
complexity comes from the number of counters (σ) and the size of the array B (p).
(cid:3)
Note that the space consumption can be reduced to O(p) at the price of introducing (Monte Carlo)
randomization. Instead of storing the Parikh vectors in a plain form, we can use dynamic hash tables [9] so
that the size is proportional to the number of non-zero entries.
4.2. Algorithm for Abelian Runs
In this section we extend our algorithm so that it reports abelian runs only. For an offline solution, we
could simply determine the anchored runs (using the procedure developed above) and filter out those which
are not maximal. However, in order to obtain an online algorithm, we need a more subtle approach, which
is based on the following characterization.
Lemma 12. A fragment w[b. . i − 1] is an abelian run with period P if and only if it is an anchored run
(with period P) and for each k′ the inequalities Bi−1[k′] ≥ b and Bi[k′] > b hold.
Proof. By Lemma 6, an abelian run of period P cannot be properly contained in a fragment with pe-
riod P (anchored at some k′). Conditions involving Bi−1[k′] and Bi[k′] enforce left-maximality and right-
maximality, respectively. Since each abelian run is an anchored run (with the same period) and since all
anchored runs are periodic, the claim follows.
(cid:3)
To apply Lemma 12, it suffices to find an anchor k such that b = Bi−1[k] = mink′ Bi−1[k′] < mink′ Bi[k′].
There can be several such anchors and in case of ties we are going to detect the one for which the factorization
of w[b. . i − 1] has shortest tail. This factorization maximizes the number of cores, so if w[b. . i − 1] is an
anchored run with any anchor, it is with that one in particular. Note that the while loop in lines 4 -- 12
of Algorithm AnchoredRun processes anchors bi−1 ≤ k < bi in the order of decreasing tail lengths and
Run(P, p, w, n)
1 k ← 0
2 L ← ∅
3 B[0] ← k
4 P tr[0] ← insertAtTheEnd(L, 0)
5 for i ← 0 to n do
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
bmin ← B[getFirst(L)]
while k ≤ n and (i = n or Pw[k..i] 6⊆ P) do
if k > i − p then
b ← B[k mod p]
B[k mod p] ← ∞
Delete(L, P tr[k mod p])
P tr[k mod p] ← Nil
if b = bmin and B[getFirst(L)] > b and b ≤ k − 2p then
h ← (k − b) mod p
t ← i − k
Output(b, h, t, i)
¯ ¯k ← k + 1
if k > i − p + 1 then
B[(i + 1) mod p] ← k
P tr[(i + 1) mod p] ← insertAtTheEnd(L, (i + 1) mod p)
¯ ¯ ¯
Figure 5: Algorithm computing all the abelian runs of period P of norm p in a word w of length n.
7
updates the underlying values B[k] from Bi−1[k] to Bi[k]. For the sought anchor k this update strictly
increases the value mink′ B[k′], and moreover this is the first increase of the minimum within a given
iteration of the outer for loop. Hence, we record the original minimum and check for an abelian run only if
mink′ B[k′] increases from that value.
To implement the procedure described above, we need to efficiently compute the smallest element in the
array B. For this, recall that B[j] can only be modified from ∞ to k (and the value k does not decrease
throughout the algorithm) or from some value back to ∞. We maintain a doubly-linked list L of all indices
j with finite B[j] such that the order of indices j in the list is consistent with the order of values B[j]. To
update the list, it suffices to insert the index j to the end of list while setting B[j] to k, and remove it
from the list setting B[j] to ∞. Then the smallest value in B is attained at an argument stored as the first
element of the list L (or ∞, if the list is empty).
The algorithm Run(P, p, w, n), depicted in Figure 5, implements the approach described above. It uses
the following constant-time functions to operate on lists:
• insertAtTheEnd(L, e) that inserts e at the end of the doubly-linked list L and returns a pointer to
the location of e in the list;
• Delete(L, ptr) that deletes the element pointed by ptr from the doubly-linked list L;
• getFirst(L) that returns the first element of the list L (0 if the list is empty).
The algorithm also uses an array P tr which maps any anchor j to a pointer of the corresponding location
in the list L (or Nil if B[j] = ∞).
The discussion above proves that Run(P, p, w, n) correctly computes abelian runs with period P in w.
Its running time is the same as that of AnchoredRun(P, p, w, n) since the structure of the computations
remains the same while additional instructions run in constant time. Memory consumption is still O(p + σ)
because both L and P tr take O(p) space.
Theorem 13. The algorithm Run(P, p, w, n) computes all the abelian runs with period P of norm p in
a word w of length n in time O(n) and additional space O(σ + p), which can be reduced to O(p) using
randomization.
4.3. Example
Let us see the behaviour of the algorithm on Σ = {a, b}, w = abaababaabbb and P = (2, 2):
k = 0 B = [0, ∞, ∞, ∞] L = (0)
i = 0 Pw[0] ⊆ P B = [0, 0, ∞, ∞] L = (0, 1)
i = 1 Pw[0..1] ⊆ P B = [0, 0, 0, ∞] L = (0, 1, 2)
i = 2 Pw[0..2] ⊆ P B = [0, 0, 0, 0] L = (0, 1, 2, 3)
i = 3 Pw[0..3] 6⊆ P b = 0 B = [∞, 0, 0, 0] L = (1, 2, 3) k = 1
Pw[1..3] ⊆ P B = [1, 0, 0, 0] L = (1, 2, 3, 0)
i = 4 Pw[1..4] ⊆ P
i = 5 Pw[1..5] 6⊆ P k = 2
Pw[2..5] 6⊆ P b = 0 B = [1, 0, ∞, 0] L = (1, 3, 0) k = 3
Pw[3..5] ⊆ P B = [1, 0, 3, 0] L = (1, 3, 0, 2)
i = 6 Pw[3..6] ⊆ P
i = 7 Pw[3..7] 6⊆ P k = 4
Pw[4..7] ⊆ P
i = 8 Pw[4..8] 6⊆ P k = 5
Pw[5..8] 6⊆ P b = 0 B = [1, ∞, 3, 0] L = (3, 0, 2) k = 6
Pw[6..8] ⊆ P B = [1, 6, 3, 0] L = (3, 0, 2, 1)
i = 9 Pw[6..9] ⊆ P
i = 10 Pw[6..10] 6⊆ P k = 7
8
Pw[7..10] ⊆ P
i = 11 Pw[7..11] 6⊆ P k = 8
Pw[8..11] 6⊆ P b = 1 B = [∞, 6, 3, 0] L = (3, 2, 1) k = 9
Pw[9..11] 6⊆ P b = 6 B = [∞, ∞, 3, 0] L = (3, 2) k = 10
Pw[10..11] ⊆ P B = [10, ∞, 3, 0] L = (3, 2, 0)
i = 12
b = 3 B = [10, ∞, ∞, 0] L = (3, 0) k = 11
b = 0 B = [10, ∞, ∞, ∞] L = (0)
h = 3 t = 1 Output(0, 3, 1, 11) k = 12
b = 10 B = [∞, ∞, ∞, ∞] L = () k = 13
5. Computing Abelian Runs with Fixed Parikh Vector Norm
In this section we develop an O(np)-time algorithm to compute all abelian runs with periods of norm p.
First, we describe the algorithm for anchored runs and later generalize it to abelian runs.
5.1. Anchored Runs
Let us start with a simple offline algorithm which works in O(n) time to compute k-anchored runs with
period of norm p for fixed values p and k. This method is similar to the algorithm of Matsuda et al. [6]
briefly described in Section 3. Namely, it suffices to compute maximal abelian powers with periods of norm
p anchored at k, and then extend them by a head and a tail.
Define a block as any fragment of the form w[i. . i + p − 1] such that i ≡ k (mod p). Note that the
cores in decompositions with anchor k mod p are blocks. Finding k-anchored powers with periods of a given
norm p is very easy if the anchor is fixed. We consider consecutive blocks, naively check if they are abelian-
equivalent and merge any maximal chains of abelian-equivalent blocks. Determining the head and the tail
of the k-anchored runs is also simple. For each i ≡ k (mod p) we compute the longest suffix of w[0. . i − 1]
and the longest prefix of w[i + p. . n − 1] whose Parikh vectors are contained in Pw[i..i+p−1].
This approach can be implemented online in O(σ + p) space as follows: we scan consecutive blocks and
(naively) check their abelian equivalence. Whenever we read a full block (say, starting at position i), we
compute the longest suffix w[bi−1. . i − 1] of w[0. . i − 1] whose Parikh vector is contained in Pw[i..i+p−1].
This gives a periodic factorization of w[bi−1. . i + p − 1] anchored at k mod p. We then try to extend it
to the right while reading further characters. Once it is impossible to extend the factorization, say by
letter w[j + 1], we declare w[bi−1. . j] as a maximal fragment with period Pw[i..i+p−1] anchored at k. If the
decomposition has at least two cores, we report an anchored run. If we succeed to extend by a full block
(i.e., if Pw[i..i+p−1] = Pw[i+p..i+2p−1]), we do not restart the algorithm but instead we continue to extend
the factorization. This way, we guarantee that bi−1 > i − p whenever we start building a new factorization.
Clearly, the procedure described above computes all k-anchored runs with period of norm p. To compute
all anchored runs, we simply run it in parallel for all p possible anchors.
Theorem 14. There is an algorithm which computes online all the anchored runs with periods of norm p
in a word w of length n over an alphabet of size σ in time O(np) and additional space O(p(σ + p)), which
can be reduced to O(p2) using randomization.
5.2. Abelian Runs
Let us first slightly modify the algorithm presented in the previous section. Observe that whenever we
start a new phase having just read a block w[i. . i + p − 1], instead of performing the computations using a
simple procedure described above, we could launch the algorithm of Section 4.1 for w[max(i − p, 0). . ] and
P = Pw[i..i+p−1], simulate it until it needs to read w[i + p] and then feed it with newly read letters until the
maximal extension of w[i. . i + p − 1] anchored at k is found (i.e., until the respective entry of the B array
is set to ∞). Other anchored runs output by the algorithm should be ignored, of course. As before, if such
a process is running while we have completed reading a subsequent block, we do not start a new phase.
9
It is easy to see that such an algorithm is equivalent to the previous one. However, if we use the
algorithm of Section 4.2 instead, we automatically get a possibility to check whether the maximal extension
of w[i. . i + p − 1] anchored at k is a maximal fragment with period Pw[i..i+p−1]. Note that we start the
simulation at a position max(i − p, 0) which is smaller than bi−1, unless the latter is 0. This guarantees that
left-maximality is correctly verified despite the fact the fragment prior to position i − p is ignored in the
simulation. As before, we disregard any other abelian run that the algorithm of Section 4.2 may return. We
run this process in parallel for all possible anchors to guarantee that each abelian run with period of norm p
is reported exactly once. More precisely, in ambiguous cases a run is reported for the anchor corresponding
to the factorization with shortest tail, just as in Section 4.2.
Theorem 15. There is an algorithm which computes online all the abelian runs with periods of norm p in
a word w of length n over an alphabet of size σ in time O(np) and additional space O(p(σ + p)), which can
be reduced to O(p2) using randomization.
6. Offline Algorithm for Computing All Abelian Runs
In this section we present an O(n2)-time offline algorithm which computes all the abelian runs. As
a starting point, we use the set of all anchored runs computed by the algorithm by Matsuda et al. (see
Section 3). Recall that all abelian runs are anchored runs with the same period. Hence, it suffices to filter
out those anchored runs which are properly contained in another anchored run with the same period. We
also need to make sure that every abelian run is reported once only (despite possibly being k-anchored for
different anchors k).
Note that this filtering can be performed independently for distinct periods. If we have a list of anchored
runs with a fixed period, sorted by the starting position, it is easy to retrieve the abelian runs of that period
with a single scan of the list. Ordering by the starting position can be performed together for all periods
so that it takes O(r + n) time where r is the number of all anchored runs. Hence, the main difficulty is
grouping according to the period. For this, we shall assign to each fragment of w an identifier, so that
two fragments are abelian-equivalent if and only if their identifiers are equal. The identifiers of periods can
be easily retrieved since given a k-anchored run, we can easily locate one of the cores of the underlying
factorization.
Thus, in the remaining part of this section we design a naming algorithm which assigns the identifiers.
A naive solution would be to generate the Parikh vectors of all substrings of w, sort these vectors removing
duplicates, and give each fragment a rank of its Parikh vector in that order. However, already storing the
Parikh vectors can take prohibitive Θ(n2σ) space.
To overcome this issue, we use the concept of diff-representation, originally introduced in the context of
abelian periods [10]. Observe that in a sense the Parikh vectors of fragments can be generated efficiently: for
a fixed p, we can first generate Pw[0..p−1], then update it to Pw[1..p], and so on until we reach Pw[n−p..n−1].
In other words, the Parikh vectors of all fragments of length p can be represented in a sequence so that the
total Hamming distance of the adjacent vectors is O(n). The diff-representation, designed to manipulate
sequences satisfying such a property, is formally defined as a sequence of single-entry changes such that the
original sequence of vectors is a subsequence of intermediate results when applying this operations starting
from the null vector (of the fixed dimension r). Note that the diff-representation of a sequence of Parikh
vectors of all fragments of w can be computed in time O(n2) proportional to its size. The following result
lets us efficiently assign identifiers to its elements.
Lemma 16 ([10]). Given a sequence of vectors of dimension r represented using a diff-representation of size
m, consider the problem of assigning integer identifiers of size nO(1) so that equality of vectors is equivalent
to equality of their identifiers. It can be solved in O(r + m log r) time using a deterministic algorithm and
in O(r + m) time using a Monte Carlo algorithm which is correct with high probability (1 − 1
(r+m)c where c
can be chosen arbitrarily large).
In our setting this yields the following result.
10
Theorem 17. There exists an O(n2)-time randomized algorithm (Monte Carlo, correct with high-probability)
which computes all abelian runs in a given word of length n. Additionally, there exists an O(n2 log σ)-time
deterministic algorithm solving the same problem.
7. Conclusions
We gave algorithms that, given a word w of length n over an alphabet of cardinality σ, return all the
abelian runs of a given period P in w in time O(n) and space O(σ + p), or all the abelian runs with periods
of a given norm p in time O(np) and space O(p(σ + p)). These algorithms work in an online manner.
We also presented an O(n2) (resp. O(n2 log n))-time offline randomized (resp. deterministic) algorithm for
computing all the abelian runs in a word of length n. One may wonder if it is possible to reduce further
the complexities of these latter algorithms. We believe that further combinatorial results on the structure
of the abelian runs in a word could lead to novel solutions.
References
[1] W. F. Smyth, Computing regularities in strings: a survey, European Journal of Combinatorics 34 (1) (2013) 3 -- 14.
doi:10.1016/j.ejc.2012.07.010.
[2] R. Kolpakov, G. Kucherov, Finding maximal repetitions in a word in linear time,
40th Annual Sympo-
sium on Foundations of Computer Science, FOCS 1999, IEEE Computer Society, New-York, 1999, pp. 596 -- 604.
doi:10.1109/SFFCS.1999.814634.
in:
[3] H. Bannai, T. I, S. Inenaga, Y. Nakashima, M. Takeda, K. Tsuruta, A new characterization of maximal repetitions by
Lyndon trees, in: P. Indyk (Ed.), Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA 2015, San Diego, CA, USA, January 4-6, 2015, SIAM, 2015, pp. 562 -- 571. doi:10.1137/1.9781611973730.38 .
[4] H. Bannai, T. I, S. Inenaga, Y. Nakashima, M. Takeda, K. Tsuruta, The "runs" theorem, CoRR abs/1406.0263v7.
URL http://arxiv.org/abs/1406.0263v7
[5] S. Constantinescu, L. Ilie, Fine and Wilf's theorem for abelian periods, Bulletin of the European Association for Theoretical
Computer Science 89 (2006) 167 -- 170.
[6] S. Matsuda, S. Inenaga, H. Bannai, M. Takeda, Computing abelian covers and abelian runs, in: J. Holub, J. Zd´arek (Eds.),
Prague Stringology Conference, PSC 2014, Czech Technical University in Prague, 2014, pp. 43 -- 51.
[7] G. Fici, T. Lecroq, A. Lefebvre, ´E. Prieur-Gaston, Online computation of abelian runs, in: A. H. Dediu, E. Formenti,
C. Mart´ın-Vide, B. Truthe (Eds.), Language and Automata Theory and Applications, LATA 2015, Vol. 8977 of LNCS,
Springer International Publishing, 2015, pp. 391 -- 401. doi:10.1007/978-3-319-15579-1_30.
[8] L. J. Cummings, W. F. Smyth, Weak repetitions in strings, Journal of Combinatorial Mathematics and Combinatorial
Computing 24 (1997) 33 -- 48.
[9] M. Dietzfelbinger, A. R. Karlin, K. Mehlhorn, F. Meyer auf der Heide, H. Rohnert, R. E. Tarjan, Dynamic perfect hashing:
Upper and lower bounds, SIAM Journal on Computing 23 (4) (1994) 738 -- 761. doi:10.1137/S0097539791194094 .
[10] T. Kociumaka, J. Radoszewski, W. Rytter, Fast algorithms for abelian periods in words and greatest common di-
visor queries,
in: N. Portier, T. Wilke (Eds.), 30th International Symposium on Theoretical Aspects of Computer
Science, STACS 2013, Vol. 20 of LIPIcs, Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2013, pp. 245 -- 256.
doi:10.4230/LIPIcs.STACS.2013.245.
11
|
1108.3683 | 1 | 1108 | 2011-08-18T08:31:11 | Substring Range Reporting | [
"cs.DS"
] | We revisit various string indexing problems with range reporting features, namely, position-restricted substring searching, indexing substrings with gaps, and indexing substrings with intervals. We obtain the following main results. {itemize} We give efficient reductions for each of the above problems to a new problem, which we call \emph{substring range reporting}. Hence, we unify the previous work by showing that we may restrict our attention to a single problem rather than studying each of the above problems individually. We show how to solve substring range reporting with optimal query time and little space. Combined with our reductions this leads to significantly improved time-space trade-offs for the above problems. In particular, for each problem we obtain the first solutions with optimal time query and $O(n\log^{O(1)} n)$ space, where $n$ is the length of the indexed string. We show that our techniques for substring range reporting generalize to \emph{substring range counting} and \emph{substring range emptiness} variants. We also obtain non-trivial time-space trade-offs for these problems. {itemize} Our bounds for substring range reporting are based on a novel combination of suffix trees and range reporting data structures. The reductions are simple and general and may apply to other combinations of string indexing with range reporting. | cs.DS | cs |
Substring Range Reporting∗
Philip Bille
[email protected]
Inge Li Gørtz
[email protected]
September 4, 2018
Abstract
We revisit various string indexing problems with range reporting features, namely, position-
restricted substring searching, indexing substrings with gaps, and indexing substrings with in-
tervals. We obtain the following main results.
• We give efficient reductions for each of the above problems to a new problem, which we
call substring range reporting. Hence, we unify the previous work by showing that we may
restrict our attention to a single problem rather than studying each of the above problems
individually.
• We show how to solve substring range reporting with optimal query time and little space.
Combined with our reductions this leads to significantly improved time-space trade-offs
for the above problems. In particular, for each problem we obtain the first solutions with
optimal time query and O(n logO(1) n) space, where n is the length of the indexed string.
• We show that our techniques for substring range reporting generalize to substring range
counting and substring range emptiness variants. We also obtain non-trivial time-space
trade-offs for these problems.
Our bounds for substring range reporting are based on a novel combination of suffix trees and
range reporting data structures. The reductions are simple and general and may apply to other
combinations of string indexing with range reporting.
1
Introduction
Given a string S of length n the string indexing problem is to preprocess S into a compact rep-
resentation that efficiently supports substring queries, that is, given another string P of length m
report all occurrences of substrings in S that match P . Combining the classic suffix tree data struc-
ture [14] with perfect hashing [13] leads to an optimal time-space trade-off for string indexing, i.e.,
an O(n) space representation that supports queries in O(m + occ) time, where occ is the number
of occurrences of P in S.
In recent years, several extensions of string indexing problems that add range reporting features
have been proposed. For instance, Makinen and Navarro proposed the position-restricted substring
searching problem [21, 22]. Here, queries take an additional range [a, b] of positions in S and the
goal is to report the occurrences of P within S[a, b]. For such extensions of string indexing no
optimal time-space trade-off is known. For instance, for position-restricted substring searching one
∗An extended abstract of this paper appeared at the 22nd Conference on Combinatorial Pattern Matching.
1
can either get O(n logε n) space (for any constant ε > 0) and O(m + log log n + occ) query time or
O(n1+ε) space with O(m + occ) query time [8, 21, 22]. Hence, removing the log log n term in the
query comes at the cost of significantly increasing the space.
In this paper, we revisit a number string indexing problems with range reporting features,
namely position-restricted substring searching, indexing substrings with gaps, and indexing sub-
strings with intervals. We achieve the following results.
• We give efficient reductions for each of the above problems to a new problem, which we
call substring range reporting. Hence, we unify the previous work by showing that we may
restrict our attention to a single problem rather than studying each of the above problems
individually.
• We show how to solve substring range reporting with optimal query time and little space.
Combined with our reductions this leads to significantly improved time-space trade-offs for
all of the above problems. For instance, we show how to solve position-restricted substring
searching in O(n logε n) space and O(m + occ) query time.
• We show that our techniques for substring range reporting generalize to substring range
counting and substring range emptiness variants. We also obtain non-trivial time-space trade-
offs for these problems.
Our bounds for substring range reporting are based on a novel combination of suffix trees and
range reporting data structures. The reductions are simple and general and may apply to other
combinations of string indexing with range reporting.
1.1 Substring Range Reporting
Let S be a string where each position is associated with a integer value in the range [0, u]. The
integer associated with position i in S is the label of position i, denoted label(i), and we call S
a labeled string. Given a labeled string S, the substring range reporting problem is to compactly
represent S while supporting substring range reporting queries, that is, given a string P and a pair
of integers a and b, 0 ≤ a ≤ b ≤ u, report all starting positions in S that match P and whose labels
are in the range [a, b].
We assume a standard unit-cost RAM model with word size w and a standard instruction set
including arithmetic operations, bitwise boolean operations, and shifts. We assume that a label
can be stored in a constant number of words and therefore w = Θ(log u). The space complexity is
the number of words used by the algorithm. All bounds mentioned in this paper are valid in this
model of computation.
To solve substring range reporting a basic approach is to combine a suffix tree with a 2D range
reporting data structure. A query for a pattern P and range [a, b] consists of a search in the suffix
tree and then a 2D range reporting query with [a, b] and the lexicographic range of suffixes defined
P . This is essentially the overall approach used in the known solutions for position-restricted
substring searching [4, 8, 9, 21, 22, 31], which is a special case of substring range reporting (see the
next section).
Depending on the choice of the 2D range reporting data structure this approach leads to different
trade-offs. In particular, if we plug in the 2D range reporting data structure of Alstrup et al. [2],
we get a solution with O(n logε n) space and O(m + log log u + occ) query time (see Makinen
2
and Navarro [21, 22]). The log log u term in the query time is from the range reporting query.
Alternatively, if we use a fast data structure for the range successor problem [8, 31] to do the range
reporting, we get optimal O(m+occ) query time but increase the space to at least Ω(n1+ε). Indeed,
since any 2D range reporting data structure with O(n logO(1) n) space must use Ω(log log u) query
time [26], we cannot hope to avoid this blowup in space with this approach.
Our first main contribution is a new and simple technique that overcomes the inherent problem
of the previous approach. We show the following result.
Theorem 1 Let S be a labeled string of length n with labels in the range [0, u]. For any constants
ε, δ > 0, we can solve substring range reporting using O(n(logε n + log log u)) space, O(n(log n +
logδ u)) expected preprocessing time, and O(m + occ) query time, for a pattern string of length m.
Compared to the previous results we achieve optimal query time with an additional O(n log log u)
term in the space. For the applications considered here, we have that u = O(n) and therefore the
space bound simplifies to O(n(logε n + log log u)) = O(n logε n). Hence, in this case there is no
asymptotic space overhead.
The key idea to obtain Theorem 1 is a new and simple combination of suffix trees with multiple
range reporting data structures for both 1 and 2 dimensions. Our solution handles queries differently
depending on the length of the input pattern such that the overall query is optimized accordingly.
Interestingly, the idea of using different query algorithms depending on the length of the pattern
is closely related to the concept of filtering search introduced for the standard range reporting
problem by Chazelle as early as 1986 [6]. Our new results show that this idea is also useful in
combinatorial pattern matching.
Finally, we also consider substring range counting and substring range emptiness variants. Here,
the goal is to count the number of occurrences in the range and to determine whether or not the
range is empty, respectively. Similar to substring range reporting, these problems can also be
solved in a straightforward way by combining a suffix with a 2D range counting or emptiness data
structure. We show how to extend our techniques to obtain improved time-space trade-offs for both
of these problems.
1.2 Applications
Our second main contribution is to show that substring range reporting actually captures several
other string indexing problems. In particular, we show how to reduce the following problems to
substring range reporting.
• Position-restricted substring searching: Given a string S of length n, construct a data struc-
ture supporting the following query: Given a string P and query interval [a, b], with 1 ≤ a ≤
b ≤ n, return the positions of substrings in S matching P whose positions are in the interval
[a, b].
• Indexing substrings with intervals: Given a string S of length n, and a set of intervals π =
{[s1, f1], [s2, f2], . . . , [sπ, fπ]} such that si, fi ∈ [1, n] and si ≤ fi, for all 1 ≤ i ≤ π, construct
a data structure supporting the following query: Given a string P and query interval [a, b],
with 1 ≤ a ≤ b ≤ n, return the positions of substrings in S matching P whose positions are
in [a, b] and in one of the intervals in π.
3
• Indexing substrings with gaps: Given a string S of length n and an integer d, the problem is
to construct a data structure supporting the following query: Given two strings P1 and P2
return all positions of substrings in S matching P1 ◦ (cid:63)d ◦ P2. Here ◦ denotes concatenation
and (cid:63) is a wildcard matching all characters in the alphabet.
Previous results Let m be the length of P . Makinen and Navarro [21, 22] introduced the
position-restricted substring searching problem. Their fastest solution uses O(n logε n) space,
O(n log n) expected preprocessing time, and O(m+log log n+occ) query time. Crochemore et al. [8]
proposed another solution using O(n1+ε) space, O(n1+ε) preprocessing time, and O(m + occ) query
time (see also Section 1.1). Using techniques from range non-overlapping indexing [7] it is possible
to improve these bounds for small alphabet sizes [27]. Several succinct versions of the problem have
also been proposed [4, 21, 22, 31]. All of these have significantly worse query time since they require
superconstant time per reported occurrence. Finally, Crochemore et al. [10] studied a restricted
version of the problem with a = 1 or b = n.
For the indexing substrings with intervals problem, Crochemore et al. [8, 9] gave a solution
with O(n log2 n) space, O(π + n log3 n) expected preprocessing time, and O(m + log log n + occ)
query time. They also showed how to achieve O(n1+ε) space, O(n1+ε + π) preprocessing time,
and O(m + occ) query time. Several papers [3,17,20] have studied the property matching problem,
which is similar to the indexing substrings with intervals problem, but where both start and end
point of the match must be in the same interval.
Iliopoulos and Rahman [18] studied the problem of indexing substrings with gaps. They gave a
solution using O(n logε n) space, O(n log n) expected preprocessing time, and O(m + loglog n + occ)
query time, where m is the length of the two input strings. Crochemore and Tischler recently
proposed a variant of the problem [11].
Our results We reduce position-restricted substring searching, indexing substrings with intervals,
and indexing substrings with gaps to substring range reporting. Applying Theorem 1 with our new
reductions, we get the following result.
Theorem 2 Let S be a string of length n and let m be the length of the query. For any constant
ε > 0, we can solve
(i) Position-restricted substring searching using O(n logε n) space, O(n log n) expected preprocess-
ing time, and O(m + occ) query time.
(ii) Indexing substrings with intervals using O(n logε n) space, O(π + n log n) expected prepro-
cessing time, and O(m + occ) query time.
(iii) Indexing substrings with gaps using O(n logε n) space, O(n log n) expected preprocessing time,
and O(m + occ) query time (m is the size of the two input strings).
This improves the best known time-space trade-offs for all three problems, that all suffer from the
trade-off inherent in 2D range reporting.
The reductions are simple and general and may apply to other combinations of string indexing
with range reporting.
4
2 Basic Concepts
2.1 Strings and Suffix Trees
Throughout the section we will let S be a labeled string of length S = n with labels in [0, u]. We
denote the character at position i by S[i] and the substrings from position i to j by S[i, j]. The
substrings S[1, j] and S[i, n] are the prefixes and suffixes of S, respectively. The reverse of S is SR.
We denote the label of position i by labelS(i). The order of suffix S[i, n], denoted orderS(i), is the
lexicographic order of S[i, n] among the suffixes of S.
The suffix tree for S, denoted TS, is the compacted trie storing all suffixes of S [14]. The depth
of a node v in TS is the number of edges on the path from v to the root. Each of the edges in TS
is associated with some substring of S. The children of each node are sorted from left to right in
increasing alphabetic order of the first character of the substring associated with the edge leading
to them. The concatenation of substrings from the root to v is denoted strS(v). The string depth
of v, denoted strdepthS(v), is the length of strS(v). The locus of a string P , denoted locusS(P ), is
the minimum depth node v such that P is a prefix of strS(v). If P is not a prefix of a substring in
S we define locusS(P ) to be ⊥.
Each leaf (cid:96) in TS uniquely corresponds to a suffix in S, namely, the suffix strS((cid:96)). Hence, we
will use labelS((cid:96)) and orderS((cid:96)) to refer to the label and order of the corresponding suffix. For an
internal node v we extend the notation such that
labelS(v) = {labelS((cid:96)) (cid:96) is a descendant leaf of v}
orderS(v) = {orderS((cid:96)) (cid:96) is a descendant leaf of v}.
Since children of a node are sorted, the left to right order of the leaves in TS corresponds to the
lexicographic order of the suffixes of S. Hence, for any node v, orderS(v) is an interval. We denote
the left and right endpoints of this interval by lv and rv. When the underlying string S is clear
from the context we will often drop the subscript S for brevity.
The suffix tree for S uses O(n) space and can be constructed in O(sort(n)) time, where sort(n)
is the time for sorting n values in the model of computation [12]. We only need a standard
comparison-based O(n log n) suffix tree construction in our results. Let P be a string of length m.
If locusS(P ) = ⊥ then P does not occur as a substring in S. Otherwise, the substrings in S that
match P are the suffixes in orderS(locusS(P )). Hence, we can compute all occurrences of P in S
by traversing the suffix tree from the root to locusS(P ) and then report all suffixes stored in the
subtree. Using perfect hashing [13] to represent the outgoing edges of each node in TS we achieve
an O(n) solution to string indexing that supports queries in O(m + occ) time (here occ is the total
number of occurrences of P in S).
2.2 Range Reporting
Let X ⊆ {0, . . . , u}d be a set of points in a d-dimensional grid. The range reporting problem in
d-dimensions is to compactly represent X while supporting range reporting queries, that is, given a
rectangle R = [a1, b1]×···× [ad, bd] report all points in the set R∩ X. We use the following results
for range reporting in 1 and 2 dimensions.
Lemma 1 (Alstrup et al. [1], Mortensen et al. [24]) For a set of n points in [0, u] and any
constant γ > 0, we can solve 1D range reporting using O(n) space, O(n logγ u) expected preprocess-
ing time and O(1 + occ) query time.
5
Lemma 2 (Alstrup et al. [2]) For a set of n points in [0, u] × [0, u] and any constant ε > 0, we
can solve 2D range reporting using O(n logε n) space, O(n log n) expected preprocessing time, and
O(log log u + occ) query time.
3 Substring Range Reporting
We now show Theorem 1. Recall that S is a labeled string of length n with labels from [0, u].
3.1 The Data Structure
Our substring range reporting data structure consists of the following components.
• The suffix tree TS for S. For each node v in TS we also store lv and rv. We partition TS into
a top tree and a number of bottom trees. The top tree consists of all nodes in TS whose string
depth is at most log log u and all their children. The trees induced by the remaining nodes
are the forest of bottom trees.
• A 2D range reporting data structure on the set of points {(orderS(i), labelS(i)) i ∈ {1, . . . , n}}.
• For each node v in the top tree, a 1D range reporting data structure on the set {labelS(i)
i ∈ orderS(v)}.
We analyze the space and preprocessing time for the data structure. We use the range reporting
data structures from Lemmas 1 and 2. The space for the suffix tree is O(n) and the space for the
2D range reporting data structure is O(n logε n), for any constant ε > 0. We bound the space for
the (potentially Ω(n)) 1D range reporting data structures stored for the top tree. Let Vd be the
set of nodes in the top tree with depth d. Since the sets orderS(v), v ∈ Vd, partition the set of
descendant leaves of nodes in Vd, the total size of these sets is as most n. Hence, the total size
of the 1D range reporting data structures for the nodes in Vd is therefore O(n). Since there are
at most log log u + 1 levels in the top tree, the space for all 1D range reporting data structures is
O(n log log u). Hence, the total space for the data structure is O(n(logε n + log log u)).
We can construct the suffix tree in O(sort(n)) time and the 2D range reporting data structure
in O(n log n) expected time. For any constant γ > 0, the expected preprocessing time for all 1D
range reporting data structures is
(cid:88)
O
= O(n log log u logγ u) = O(n log2γ u).
orderS(v) logγ u
v in top tree
Setting δ = 2γ we use O(n(log n + logδ u)) expected preprocessing time in total.
3.2 Substring Range Queries
Let P be a string of length m, and let a and b be a pair of integers, 0 ≤ a ≤ b ≤ u. To answer a
substring range query we want to compute the set of starting positions for P whose labels are in
[a, b]. First, we compute the node v = locusS(P ). If v = ⊥ then P is not a substring of S, and we
return the empty set. Otherwise, we compute the set of descendant leaves of v with labels in [a, b].
There are two cases to consider.
6
(i) If v is in the top tree we query the 1D range reporting data structure for v with the interval
[a, b].
(ii) If v is in a bottom tree, we query the 2D range reporting data with the rectangle [lv, rv]×[a, b].
Given the points returned by the range reporting data structures, we output the corresponding
starting positions of the corresponding suffixes. From the definition of the data structure it follows
that these are precisely the occurrences of P within the range [a, b]. Next consider the time
complexity. We find locusS(P ) in O(m) time (see Section 2). In case (i) we use O(1 + occ) time to
compute the result by Lemma 1. Hence, the total time for a substring range query for case (i) is
O(m + occ). In case (ii) we use O(log log u + occ) time to compute the result by Lemma 2. We have
that v = locusS(P ) is in a bottom tree and therefore m ≥ strdepth(parent(locusS(v))) > log log u.
Hence, the total time to answer a substring range query in case (ii) is O(m + log log u + occ) =
O(m + occ). Thus, in both cases we use O(m + occ) time.
Summing up, our solution uses O(n(logε n + log log u) space, O(n(log n + logδ u)) expected
preprocessing time, and O(m + occ) query time. This completes the proof of Theorem 1.
4 Applications
In this section we show how to improve the results for the three problems position-restricted sub-
string searching, indexing substrings with intervals, and indexing gapped substrings, using our data
structure for substring range reporting. Let reportS(P, a, b) denote a substring range reporting
query on string S with parameters P , a, and b.
4.1 Position-Restricted Substring Searching
We can reduce position-restricted substring searching to substring range reporting by setting
label(i) = i for all i = 1, . . . , n. To answer a query we return the result of the substring range
query reportS(P, a, b). Since each label is equal to the position, it follows that the solution to
the substring range reporting instance immediately gives a solution to position-restricted substring
searching. Applying Theorem 1 with u = n, this proves Theorem 2(i).
4.2 Indexing Substrings with Intervals
We can reduce indexing substrings with intervals to substring range reporting by setting
(cid:40)
i
0
label(i) =
if i ∈ ϕ for some ϕ ∈ π,
otherwise.
To answer a query we return the result of the substring range reporting query reportS(P, a, b).
Let I be the solution to the indexing substrings with intervals instance and let I(cid:48) be the solution
to the substring range reporting instance derived by the above reduction. Then i ∈ I ⇔ i ∈ I(cid:48).
To prove this assume i ∈ I. Then i ∈ ϕ for some ϕ ∈ π and i ∈ [a, b]. From i ∈ ϕ and the
definition of label(i) it follows that label(i) = i. Thus, label(i) = i ∈ [a, b] and thus i ∈ I(cid:48). Assume
i ∈ I(cid:48). Then label(i) ∈ [a, b]. Since a > 0 also label(i) > 0, and it follows that label(i) = i. By the
reduction this means that i ∈ ϕ for some ϕ ∈ π. Since i = label(i), we have i ∈ [a, b] and therefore
i ∈ I.
7
Figure 1: A string S, the labeling for d = 2 (below the string), and the suffix tree of TSR. Given
a query P1 = ab and P2 = bac we find v = locusSR(ba) (marked in the suffix tree). We have
lv = 6 and rv = 7 from the left-to-right-order in the TSR. The substring range reporting query
reports(P2, 6, 7) returns 7. Hence, we report the occurrence at position 7 − 2 − 2 = 3.
We can construct the labeling in O(n +π) if the intervals are sorted by startpoint or endpoint.
Otherwise additional time for sorting is needed. A similar approach is used in the solution by
Crochemore et al. [8].
Applying Theorem 1 with u = n, this proves Theorem 2(ii).
4.3 Indexing Substrings with Gaps
We can reduce the indexing substrings with gaps problem to substring range reporting as follows.
Construct the suffix tree of the reverse of S, i.e., the suffix tree TSR for SR. For each node v in
TSR also store lv and rv. Set
(cid:40)
orderSR(n − i + d + 2)
0
for i ≥ d + 2,
otherwise.
labelS(i) =
To answer a query find the locus node v of P R
in TSR. Then use the substring range reporting
1
data structure to return all positions of substrings in S matching P2 whose labels are in the range
[lv, rv]. For each position i returned by reportS(P2, lv, rv), return i − P1 − d. See Fig. 1 for an
example.
Correctness of the reduction We will now show that the reduction is correct. Let I be the
solution to the indexing substrings with gaps instance and let I(cid:48) be the solution to the substring
range reporting instance derived by the above reduction. We will show i ∈ I ⇔ i ∈ I(cid:48). Let mi = Pi
for i = 1, 2.
If i ∈ I then there is an occurrence of P1 at position i in S and an occurrence of P2 at
position i(cid:48) = i + m1 + d in S. It follows directly, that there is an occurrence of P R
1 at position
i(cid:48)(cid:48) = n − (i + m1) + 2 in SR. By definition,
(cid:48)
labelS(i
) = labelS(i + m1 + d) = orderSR(n − (i + m1 + d) + d + 2) = orderSR(i
(cid:48)(cid:48)
),
where the second equality follows from the fact that i + m1 + d ≥ d + 2. Since there is an
occurrence of P R
1 )).
1 at position i(cid:48)(cid:48) in SR, we have labelS(i(cid:48)) = orderSR(i(cid:48)(cid:48)) ∈ orderSR(locusSR(P R
8
cababa$cababa$cabcababa$aba$acabbaab$cababa$ba$$ba$$68311079425vlvrv7aaabbcabca123456891079861000S=25 Thus, labelS(i(cid:48)) ∈ [lv, rv], and since there is an occurrence of P2 at position i(cid:48) in S, we have
i(cid:48) − m1 − d = i ∈ I(cid:48).
If i ∈ I(cid:48) then there is an occurrence of P2 at position i(cid:48) = i + m1 + d with label(i(cid:48)) in the range
1 ). We need to show that this implies that there is an occurrence of
[lv, rv], where v = locusSR(P R
P1 at position i in S. By definition,
(cid:48)
labelS(i
) = orderSR(n − i
(cid:48)
+ d + 2) = orderSR(n − i − m1 + 2).
Let i(cid:48)(cid:48) = n − i − m1 + 2. Since orderSR(i(cid:48)(cid:48)) = labelS(i(cid:48)) ∈ [lv, rv], there is an occurrence of P R
1 at
position i(cid:48)(cid:48) in SR. It follows directly, that there is an occurrence of P1 at position n− i(cid:48)(cid:48) − m1 + 2 =
n − (n − i − m1 + 2) − m1 + 2 = i in S. Therefore, i ∈ I.
Complexity Construction of the suffix tree TSR takes time O(n log n) and the labeling can be
constructed in time O(n). Both use space O(n). It takes O(m1) time to find the locus nodes of
P R
in TSR. The substring range reporting query takes time O(m2 + occ). Thus the total query
1
time is O(m + occ).
Applying Theorem 1 with u = n, this completes the proof of Theorem 2(iii).
5 Substring Range Counting and Emptiness
We now show how to apply our techniques to substring range counting and substring range empti-
ness. Analogous to substring range reporting, the goal is here to count the number of occurrences
in the range and to determine whether or not the range is empty, respectively. A straightforward
way to solve these problems is to combine a suffix tree with a 2D range counting data structure
and a 2D range emptiness data structure, respectively. Using the techniques from Section 3 we
show how to significantly improve the bounds of this approach in both cases. We note that by
the reductions in Section 4 the bounds for substring range counting and substring range emptiness
also immediately imply results for counting and emptiness versions of position-restricted substring
searching, indexing substrings with intervals, and indexing substrings with gaps.
5.1 Preliminaries
Let X ⊆ {0, . . . , u} be a set of points in a d-dimensional grid. Given a query rectangle R =
[a1, b1] × ··· × [ad, bd], a range counting query computes R ∩ X, and a range emptiness query
computes if R ∩ X = ∅. Given X the range counting problem and the range emptiness problem is
to compactly represent X, while supporting range counting queries and range emptiness queries,
respectively. Note that any solution for range reporting or range counting implies a solution for
range emptiness with the same complexity (ignoring the occ term for range reporting queries). We
will need the following additional geometric data structures.
Lemma 3 (J´aJ´a et al. [19]) For a set of n points in [0, u]×[0, u] we can solve 2D range counting
in O(n) space, O(n log n) preprocessing time, and O(log n/ log log n + log log u) query time.
Lemma 4 (van Emde Boas et al. [29, 30], Mehlhorn and Naher [23]) For a set of n points
in [0, u] we can solve 1D range counting in O(n) space, O(n log log n) preprocessing time, and
O(log log u) query time.
9
To achieve the result of Lemma 4 we use a van Emde Boas data structure [29, 30] implemented
in linear space [23] using perfect hashing. This data structure supports predecessor queries in
O(log log u) time. By also storing for each point it's rank in the sorted order of the points, we can
compute a range counting query by two predecessor queries. To build the data structure efficiently
we need to sort the points and build suitable perfect hash tables. We can sort deterministically in
O(n log log n) time [16], and we can build the needed hash tables in O(n) time using deterministic
hashing [15] combined with a standard two-level approach (see e.g., Thorup [28]).
Lemma 5 (Chan et al. [5]) For a set of n points in [0, u]×[0, u] we can solve 2D range emptiness
in O(n log log n) space, O(n log n) preprocessing time, and O(log log u) query time.
5.2 The Data Structures
We now show how to efficiently solve substring range counting and substring range emptiness.
Recall that S is a labeled string of length n with labels from [0, u].
We can directly solve substring range counting by combining a suffix tree with the 2D range
counting result from Lemma 3. This leads to a solution using O(n) space and O(m+log n/ log log n+
log log u) query time. We show how to improve the query time to O(m + log log u) at the cost of
increasing the space to O(n log n/ log log n). Hence, we remove the log n/ log log n term from the
query time at the cost of increasing the space by a log n/ log log n factor. We cannot hope to achieve
such a bound using a suffix tree combined with a 2D range counting data structure since any 2D
range counting data structure using O(n logO(1) n) space requires Ω(log n/ log log n) query time [25].
We can also directly solve substring range emptiness by combining a suffix tree with the 2D range
emptiness result from Lemma 5. This solution uses O(n log log n) space and O(m + log log u) query
time. We show how to achieve optimal O(m) query time with space O(n log log u).
Our data structure for substring range counting and existence follows the construction in Sec-
tion 3. We partition the suffix tree into a top and a number of bottom trees and store a 1D data
structure for each node in the top tree and a single 2D data structure. To answer a query for a
pattern string P of length m, we search the suffix tree with P and use the 1D data structure if the
search ends in the top tree and otherwise use the 2D data structure.
We describe the specific details for each problem. First we consider substring range counting.
In this case the top tree consists of all nodes of string depth at most log n/ log log n. The 1D and 2D
data structures used are the ones from Lemma 4 and 3. By the same arguments as in Section 3 the
total space used for the 1D data structures for all nodes in the top tree at depth d is at most O(n)
and hence the total space for all 1D data structures is O(n(log n/ log log n)). Since the 2D data
structure uses O(n) space, the total space is O(n log n/ log log n). The time to build all 1D data
structures is O(n(log n/ log log n) · log log n)) = O(n log n). Since the suffix tree and the 2D data
structure can be built within the same bound, the total preprocessing time is O(n log n). Given
a pattern of length m, a query uses O(m + log log u) time if the search ends in the top tree, and
O(m + log n/ log log n + log log u) time if the search ends in a bottom tree. Since bottom trees
consists of nodes of string depth more than log n/ log log n the time to answer a query in both cases
is O(m + log log u). In summary, we have the following result.
Theorem 3 Let S be a labeled string of length n with labels in the range [0, u]. We can solve
substring range counting using O(n log n/ log log n) space, O(n log n) preprocessing time, and O(m+
log log u) query time, for a pattern string of length m.
10
Next we consider substring range emptiness.
In this case the top tree consists of all nodes
of string depth at most log log u. We use the 1D and 2D data structures from Lemma 1 and
Lemma 5. The total space for all 1D data structures is O(n log log u). Since the 2D data structure
uses O(n log log n) space the total space is O(n log log u). For any constant γ > 0, the expected time
to build all 1D data structures is O(n log log u logγ u) = O(n logδ u) for suitable constant δ > 0. The
suffix tree and the 2D data structure can be built in O(n log n) time and hence the total expected
preprocessing time is O(n(log n + logδ u)). If the search for a pattern string ends in the top tree
the query time is O(m) and if the search ends in a bottom tree the query time is O(m + log log u).
As above, the partition in top and bottom trees ensures that the query time in both cases is O(m).
In summary, we have the following result.
Theorem 4 Let S be a labeled string of length n with labels in the range [0, u]. For any constant
δ > 0 we can solve substring range existence using O(n log log u) space, O(n(log n+logδ u)) expected
preprocessing time, and O(m) query time, for a pattern string of length m.
6 Acknowledgments
We thank Christian Worm Mortensen and Kasper Green Larsen for clarifications on the prepro-
cessing times for the results in Lemma 3 and Lemma 5.
References
[1] S. Alstrup, G. Brodal, and T. Rauhe. Optimal static range reporting in one dimension. In
Proc. 33rd STOC, pages 476 -- 482, 2001.
[2] S. Alstrup, G. Stølting Brodal, and T. Rauhe. New data structures for orthogonal range
searching. In Proc. 41st FOCS, pages 198 -- 207, 2000.
[3] A. Amir, E. Chencinski, C. S. Iliopoulos, T. Kopelowitz, and H. Zhang. Property matching
and weighted matching. Theoret. Comput. Sci., 395(2-3):298 -- 310, 2008.
[4] P. Bose, M. He, A. Maheshwari, and P. Morin. Succinct orthogonal range search structures
on a grid with applications to text indexing. In Proc. 11th WADS, pages 98 -- 109, 2009.
[5] T. M. Chan, K. Larsen, and M. Patra¸scu. Orthogonal range searching on the ram, revisited.
In Proc. 27th SoCG, pages 354 -- 363, 2011.
[6] B. Chazelle. Filtering search: A new approach to query-answering. SIAM J. Comput.,
15(3):703 -- 724, 1986.
[7] H. Cohen and E. Porat. Range non-overlapping indexing. In Proc. 20th ISAAC, pages 1044 --
1053, 2009.
[8] M. Crochemore, C. S. Iliopoulos, M. Kubica, M. S. Rahman, and T. Walen.
gorithms for the range next value problem and applications.
205 -- 216, 2008.
Improved al-
In Proc. 25th STACS, pages
[9] M. Crochemore, C. S. Iliopoulos, M. Kubica, M. S. Rahman, and T. Walen. Finding patterns
in given intervals. Fundam. Inform., 101(3):173 -- 186, 2010.
11
[10] M. Crochemore, C. S. Iliopoulos, and M. S. Rahman. Optimal prefix and suffix queries on
texts. Inf. Process. Lett., 108(5):320 -- 325, 2008.
[11] M. Crochemore and G. Tischler. The gapped suffix array: A new index structure. In Proc.
17th SPIRE, pages 359 -- 364, 2010.
[12] M. Farach-Colton, P. Ferragina, and S. Muthukrishnan. On the sorting-complexity of suffix
tree construction. J. ACM, 47(6):987 -- 1011, 2000.
[13] M. L. Fredman, J. Koml´os, and E. Szemer´edi. Storing a sparse table with O(1) worst case
access time. J. ACM, 31:538 -- 544, 1984.
[14] D. Gusfield. Algorithms on strings, trees, and sequences: computer science and computational
biology. Cambridge, 1997.
[15] T. Hagerup, P. B. Miltersen, and R. Pagh. Deterministic dictionaries. J. Algorithms, 41(1):69 --
85, 2001.
[16] Y. Han. Deterministic sorting in O(n log log n) time and linear space. J. Algorithms, 50(1):96 --
105, 2004.
[17] C. S. Iliopoulos and M. S. Rahman. Faster index for property matching. Inf. Process. Lett.,
105(6):218 -- 223, 2008.
[18] C. S. Iliopoulos and M. S. Rahman. Indexing factors with gaps. Algorithmica, 55(1):60 -- 70,
2009.
[19] J. J´aJ´a, C. W. Mortensen, and Q. Shi. Space-efficient and fast algorithms for multidimensional
dominance reporting and counting. In Proc. 15th ISAAC, pages 558 -- 568, 2004.
[20] M. Juan, J. Liu, and Y. Wang. Errata for "Faster index for property matching". Inf. Process.
Lett., 109(18):1027 -- 1029, 2009.
[21] V. Makinen and G. Navarro. Position-restricted substring searching.
In Proc. 7th LATIN
2006, pages 703 -- 714, 2006.
[22] V. Makinen and G. Navarro. Rank and select revisited and extended. Theoret. Comput. Sci.,
387(3):332 -- 347, 2007.
[23] K. Mehlhorn and S. Nahler. Bounded ordered dictionaries in O(log log N ) time and O(n)
space. Inform. Process. Lett., 35(4):183 -- 189, 1990.
[24] C. W. Mortensen, R. Pagh, and M. Patra¸ccu. On dynamic range reporting in one dimension.
In Proc. 37th STOC, pages 104 -- 111, 2005.
[25] M. Patra¸scu. Lower bounds for 2-dimensional range counting. In Proc. 39th STOC, pages
40 -- 46, 2007.
[26] M. Patra¸scu and M. Thorup. Time-space trade-offs for predecessor search.
In Proc. 38th
STOC, pages 232 -- 240, 2006.
[27] E. Porat, 2011. Personal communication.
12
[28] M. Thorup. Space efficient dynamic stabbing with fast queries. In Proceedings of the 33rd
Annual ACM Symposium on Theory of Computing, pages 649 -- 658, 2003.
[29] P. van Emde Boas. Preserving order in a forest in less than logarithmic time and linear space.
Inform. Process. Lett., 6(3):80 -- 82, 1977.
[30] P. van Emde Boas, R. Kaas, and E. Zijlstra. Design and implementation of an efficient priority
queue. Mathematical Systems Theory, 10:99 -- 127, 1977. Announced at FOCS 1975.
[31] C.-C. Yu, W.-K. Hon, and B.-F. Wang. Improved data structures for the orthogonal range
successor problem. Comput. Geometry, 44(3):148 -- 159, 2011.
13
|
1811.11856 | 1 | 1811 | 2018-11-27T16:42:21 | Efficient Measuring of Congruence on High Dimensional Time Series | [
"cs.DS",
"cs.DM"
] | A time series is a sequence of data items; typical examples are streams of temperature measurements, stock ticker data, or gestures recorded with modern virtual reality motion controllers. Quite some research has been devoted to comparing and indexing time series. Especially, when the comparison should not be affected by time warping, the ubiquitous Dynamic Time Warping distance function ($\texttt{DTW}$) is one of the most analyzed time series distance functions. The Dog-Keeper distance ($\texttt{DK}$) is another example for a distance function on time series which is truely invariant under time warping.
For many application scenarios (e.$\,$g. motion gesture recognition in virtual reality), the invariance under isometric spatial transformations (i.$\,$e. rotation, translation, and mirroring) is as important as the invariance under time warping. Distance functions on time series which are invariant under isometric transformations can be seen as measurements for the congruency of two time series. The congruence distance ($\texttt{CD}$) is an example for such a distance function. However, it is very hard to compute and it is not invariant under time warpings.
In this work, we are taking one step towards developing a feasable distance function which is invariant under isometric spatial transformations and time warping: We develop four approximations for $\texttt{CD}$. Two of these even satisfy the triangle inequality and can thus be used with metric indexing structures. We show that all approximations serve as a lower bound to $\texttt{CD}$. Our evaluation shows that they achieve remarkable tightness while providing a speedup of more than two orders of magnitude to the congruence distance. | cs.DS | cs |
Efficient Measuring of Congruence on High Dimensional Time
Series
Jorg P. Bachmann1 and Johann-Christoph Freytag2
1 [email protected]
2 [email protected]
1,2Humboldt-Universitat zu Berlin, Germany
November 30, 2018
Abstract
A time series is a sequence of data items; typical
examples are streams of temperature measure-
ments, stock ticker data, or gestures recorded
with modern virtual reality motion controllers.
Quite some research has been devoted to com-
paring and indexing time series. Especially,
when the comparison should not be affected by
time warping, the ubiquitous Dynamic Time
Warping distance function (DTW) is one of the
most analyzed time series distance functions.
The Dog-Keeper distance (DK) is another exam-
ple for a distance function on time series which
is truely invariant under time warping.
For many application scenarios (e. g. motion
gesture recognition in virtual reality), the in-
variance under isometric spatial transformations
(i. e. rotation, translation, and mirroring) is as
important as the invariance under time warping.
Distance functions on time series which are in-
variant under isometric transformations can be
seen as measurements for the congruency of two
time series. The congruence distance (CD) is an
example for such a distance function. However,
it is very hard to compute and it is not invariant
under time warpings.
In this work, we are taking one step towards
developing a feasable distance function which
is invariant under isometric spatial transforma-
tions and time warping: We develop four ap-
proximations for CD. Two of these even satisfy
the triangle inequality and can thus be used with
metric indexing structures. We show that all ap-
proximations serve as a lower bound to CD. Our
evaluation shows that they achieve remarkable
tightness while providing a speedup of more than
two orders of magnitude to the congruence dis-
tance.
1
Introduction
Multimedia retrieval is a common application
which requires finding similar objects to a query
object. We consider examples such as gesture
recognition with modern virtual reality motion
controllers and classification of handwritten let-
1
ters where the objects are multi-dimensional
time series.
In many cases, similarity search is performed
using a distance function on the time series,
where small distances imply similar time series.
A nearest neigbor query to the query time series
can be a k-nearest neighbor (k-NN) query or an
ε-nearest neighbor (ε-NN) query: A k-NN query
retrieves the k most similar time series; an ε-NN
query retrieves all time series with a distance of
at most ε.
In our examples, the time series of the same
classes (e. g., same written characters or same
gestures) differ by temporal as well as spatial
displacements. Time warping distance functions
such as dynamic time warping (DTW) [14] and
the Dog-Keeper distance (DK) [6, 10] are robust
against temporal displacements. They map pairs
of time series representing the same trajectory to
small distances. Still, they fail when the time se-
ries are rotated or translated in space.
The distance functions defined and analyzed
in this paper measure the (approximate) congru-
ence of two time series. Thereby, the distance be-
tween two time series S and T shall be 0 iff S can
be transformed into T by rotation, translation,
and mirroring; in this case, S and T are said to
be congruent. A value greater than 0 shall corre-
late to the amount of transformation needed to
turn the time series into congruent ones.
The classical Congruence problem basically
determines whether two point sets A, B ⊆ Rk
are congruent considering isometric transforma-
tions (i. e., rotation, translation, and mirroring)
[3, 11]. For 2- and 3-dimensional spaces, there
are results providing algorithms with runtime
O(n · log n) when n is the size of the sets [3].
For larger dimensionalities, they provide an al-
gorithm with runtime O(nk−2·log n). For various
reasons (e. g. bounded floating point precision,
physical measurement errors), the approximated
Congruence problem is of much more interest
in practical applications. Different variations of
the approximated Congruence problem have
been studied (e. g. what types of transforma-
tions are used, is the assignment of points from
A to B known, what metric is used) [2, 3, 11, 12].
The Congruence problem is related to our
work, since the problem is concerned with the ex-
istence of isometric functions such that a point
set maps to another point set. The main differ-
ence is, that we consider ordered lists of points
(i. e. time series) rather than pure sets. It turned
out, that solving the approximated Congru-
ence problem is NP-hard regarding length and
dimensionality [5].
With this work, we contribute by evaluating
the congruence distance with an implementation
based on a nonlinear optimizer. We propose
two approximations to the congruence distance
which have linear runtime regarding the dimen-
sionality and (quasi-) quadratic runtime regard-
ing the length of the time series. We improve
the complexity of both approximations at cost
of approximation quality, such that their com-
plexity is (quasi-) linear regarding the length of
the time series. We evaluate the approximations
experimentally.
1.1 Basic Notation
We denote the natural numbers including zero
with N and the real numbers with R. For a, b ∈ N
we denote the modulo operator by a%b. The
:=
set of all powers of two is denoted via 2
{1, 2, 4, 8,···}.
Elements of a k-dimensional vector v ∈ Rk are
accessed using subindices, i. e. v3 is the third el-
ement of the vector. Sequences (here also called
time series) are usually written using capital let-
N
2
ters, e. g. S = (s0,··· , sn−1) is a sequence of
length n. Suppose si ∈ Rk, then si,j denotes the
j-th element of the i-th vector in the sequence
S. The projection to the j-th dimension is de-
noted via Sj, i. e. Sj = (s0,j,··· , sn−1,j). The
Euclidean norm of a vector v is denoted via (cid:107)v(cid:107)2,
thus d(v, w) := (cid:107)v − w(cid:107)2 denotes the Euclidean
distance between v and w ∈ Rk.
We denote the set of k-dimensional orthog-
onal matrices with MO(k), the identity ma-
trix with I and the transposed of a matrix M
with M T . For a matrix M in Rk, we denote
the matrix holding the absoloute values with
M := (Mi,j)0(cid:54)i,j<k.
1.2 Congruence Distance
While DTW compares two time series S and T ,
it is (nearly) invariant under time warpings.
In detail, consider σ(S) and τ (T ) as warp-
ings by duplicating elements (e. g.
σ(S) =
(s0, s1, s1, s1, s2, s3, s3, s4,··· )), then DTW mini-
mizes the L1 distance under all time warps:
σ(S)−1(cid:88)
i=0
d (σ(S)i, τ (T )i)
DTW(S, T ) := min
σ,τ
with σ(S) = τ (T ).
On the other hand, the congruence distance
is invariant under all isometric transformations.
The difference to DTW is, that it minimizes the
L1 distance by multiplying an orthogonal matrix
and adding a vector:
f (M, v)
dC(S, T ) := min
M,v
:= min
M,v
n−1(cid:88)
i=0
d (si, M · ti + v)
(1)
to the objective function. For time series in Rk,
the orthogonality of M yields a set of k2 equality
based constraints.
2 Approximating the Congru-
ency
Consider two time series S, T , an arbitrary or-
thogonal matrix M ∈ MO(k), and a vector
v ∈ Rk. Using the triangle inequality, we ob-
tain
d(si, M · sj + v) (cid:54)
d(si, M · ti + v) + d(ti, tj) + d(M · tj + v, sj)
⇒d(si, sj) − d(ti, tj) (cid:54)
d(si, M · ti + v) + d(sj, M · tj + v)
(cid:80)n−1
i. e. we can estimate the congruence distance
i=0 d(si, M · ti + v) without actually solving
the optimization problem. We unroll this idea
to propose two approximating algorithms in Sec-
tion 2.1 and 2.2.
(2)
Considering the well-known self-similarity
matrix of a time series, the left hand side of
Equation (2) matches the entry of the differ-
ence between two self-similarity matrices. Usu-
ally, the self-similarity matrix is used to analyze
a time series for patterns (e. g. using Recurrence
Plots [9]). The important property that makes
the self-similarity matrix useful for approximat-
ing the congruence distance, is its invariance un-
der transformations considered for the congru-
ence distance,
rotation, translation, and
mirroring.
i. e.
The self-similarity matrix of an arbitrary time
series T = (t0, . . . , tn−1) is defined as follows:
∆T := (cid:0) d (ti, tj) (cid:1)
0(cid:54)i,j<n
The computation of dC(S, T ) is an optimization
problem where f (cf. Equation (1)) corresponds
Note, that ∆Ti,j = ∆Tj,i and ∆Ti,i = 0.
In
fact, the self-similarity matrix ∆T completely
3
describes the sequence T up to congruence, i.e.,
up to rotation, translation, and mirroring of the
whole sequence in Rk [5]: Two time series S
and T are congruent iff they have the same self-
similarity matrix, i. e.
∃M ∈ MO(k), v ∈ Rk :
S = M · T + v ⇐⇒ ∆S = ∆T.
(3)
2.1 Metric Approximation
Equation (2) and (3) yield the approach for ap-
proximating the congruency: We measure the
congruency of two time series S and T via a met-
ric on their self-similarity matrices.
Definition 2.1 (Delta Distance). Let S, T be
two time series of length n. The delta distance
d∆(S, T ) is defined as follows:
Since d is symmetric, d∆ inherits its symme-
try. Hence, d∆ is a pseudo metric on the set of
time series of length n where all time series of an
equivalence class are congruent to each other.
We omit providing pseudo code since the com-
putation matches the formula in Definition 2.1.
The complexity of computing the delta distance
d∆ grows quadratically with the length of the
time series.
Our next aim is to show that the the delta dis-
tance d∆ provides a lower bound on the congru-
ence distance dC, as formulated in the following
theorem.
Theorem 2.3. For all time series S and T , the
following holds:
d∆(S, T ) (cid:54) dC(S, T ).
(cid:12)(cid:12)d(cid:0)si, s(i+δ)%n
(cid:1) − d(cid:0)ti, t(i+δ)%n
(cid:1)(cid:12)(cid:12)
Proof. Fixiate a δ∗ which maximizes d∆(S, T ) in
Definition 2.1. Using the triangle inequality as
in Equation (2) yields
d∆(S, T ) :=
1
2
max
0<δ<n
n−1(cid:88)
i=0
(cid:1)(cid:12)(cid:12)
d∆(S, T ) =
1
2
i=0
i=0
1
2
n−1(cid:88)
(cid:12)(cid:12)d(cid:0)si, s(i+δ∗)%n
(cid:1) − d(cid:0)ti, t(i+δ∗)%n
n−1(cid:88)
(cid:12)(cid:12)d(cid:0)si, s(i+δ∗)%n
(cid:1)−
d(cid:0)M · ti + v, M · t(i+δ∗)%n + v(cid:1)(cid:12)(cid:12)
n−1(cid:88)
d(cid:0)s(i+δ∗)%n, M · t(i+δ∗)%n + v(cid:1)(cid:1)
n−1(cid:88)
(d (si, M · ti + v) +
d (si, M · ti + v)
(cid:54) 1
2
i=0
=
i=0
for arbitrary M ∈ MO(k) and v ∈ Rk. Hence,
d∆(S, T ) (cid:54) dC(S, T ).
4
Proposition 2.2. The delta distance satisfies
the triangle inequality.
Proof. Consider three time series R, S, and T
and fixiate a δ∗ which maximizes d∆(R, T ) in
Definition 2.1. Then
=
dδ(R, T ) =
i=0
n−1(cid:88)
(cid:12)(cid:12)d(cid:0)ri, r(i+δ∗)%n
n−1(cid:88)
(cid:12)(cid:12)d(cid:0)ri, r(i+δ∗)%n
(cid:12)(cid:12)d(cid:0)si, s(i+δ∗)%n
(cid:1)(cid:12)(cid:12)
(cid:1) − d(cid:0)ti, t(i+δ∗)%n
(cid:1)(cid:12)(cid:12) +
(cid:1) + d(cid:0)si, s(i+δ∗)%n
(cid:1)(cid:12)(cid:12)
(cid:1) − d(cid:0)ti, t(i+δ∗)%n
i=0
(cid:54)
(cid:54) d∆(R, S) + d∆(S, T )
prooves the triangle inequality.
In this section, we provided the delta distance,
which is a metric lower bound to the congruence
distance.
2.2 Greedy Approximation
The approach of the delta distance is simple: For
time series S and T , it only sums up values along
a (wrapped) diagonal in ∆S − ∆T and chooses
the largest value. However, another combination
of elements within ∆S − ∆T as addends might
provide a better approximation of the congru-
ence distance. Since it is a computational expen-
sive task, to try all combinations, we try to find
a good combination using a greedy algorithm for
selecting the entries of ∆S − ∆T.
The greedy algorithm first sorts the ele-
ments di,j = d(si, sj) − d(ti, tj)
in descend-
ing order and stores them in a sequence Q =
(di1,j1, di2,j2,··· ). While iterating over the se-
quence Q,
it adds dir,jr to a global sum and
masks the indices ir and jr as already seen. El-
ements in the queue which access already seen
indices are skipped, thus each index is used at
most once. Basically, this is the reason, why the
greedy delta distance (denoted as dG(S, T )) is a
lower bound to the congruence distance. The-
orem 2.4 proves the last statement and Algo-
rithm 1 provides the pseudo code for the com-
putation.
ements, which takes n2 · log(n2) steps.
The complexity is dominated by sorting n2 el-
Theorem 2.4. For all time series S and T , the
following holds:
dG(S, T ) (cid:54) dC(S, T ).
Algorithm 1 Greedy Delta Distance
Input: time series S, T of length n
1 Algorithm: greedydelta
2
3 Output: distance d
4
5
6
7
8
9
10
11
12
13
14
15
16
let Q = () // empty sequence
for i = 0, . . . , n − 2
for j = i + 1, . . . , n − 1
append di,j := d (si, sj) − d (ti, tj) to
Q
sort Q // (descending)
let S = ∅
let d = 0
for each dia,ja in Q
if ia ∈ S or ja ∈ S continue
let d = d + dia,ja
let S = S ∪ {ia, ja}
return d
have not been skipped. Since each index is ap-
pears at most once in this list, the following in-
equality holds for arbitrary orthogonal matrices
M and vectors vRk:
r(cid:88)
dG(S, T ) =
dia,ja
a=1
r(cid:88)
r(cid:88)
n−1(cid:88)
a=1
(cid:54)
(cid:54)
(cid:54)
i=0
a=1
d (sia, sja) − d (M · tia + v, M · tja + v)
d (sia, tia) + d (M · tia + v, M · tja + v)
d (si, M · ti + v)
Proof. Let Q∗ = (di1,j1,··· , dir,jr ) be the list of
elements from the queue in Algorithm 1 which
Hence, dG(S, T ) (cid:54) dC(S, T ).
5
2.3 Runtime improvement
The complexity of the delta distance and greedy
delta distance is linear regarding the dimension-
ality but quadratic in length. In this section, we
motivate an optimization for both algorithms.
Time series usually do not contain random
points, but they come from continuous processes
in the real world, i. e. the distance between two
successive elements is rather small. Hence, the
distances d (ti, tj) and d (ti, tj+1) are probably
close to each other if i (cid:28) j, i. e.
if j is much
larger than i. This insight leads to the idea, to
only consider elements d (ti, tj) where i − j is a
power of two, i. e. we consider less elements for
larger temporal distances.
The Fast Delta Distance: Adapting the idea
to the delta distance d∆ yields the following def-
inition.
Definition 2.5 (Fast Delta Distance). Let S, T
be two time series of length n. The fast delta
distance d∆(S, T ) is defined as follows:
d∆(S, T ) :=
1
2
max
0<δ<n,δ∈2N
n−1(cid:88)
i=0
(cid:12)(cid:12)d(cid:0)si, s(i+δ)%n
(cid:1) − d(cid:0)ti, t(i+δ)%n
(cid:1)(cid:12)(cid:12)
Since we omit some values δ in Definition 2.1,
the fast version d∆ is a lower bound to d∆, i. e.
the following theorem holds:
Theorem 2.6. For all time series S and T , the
following holds:
d∆(S, T ) (cid:54) d∆(S, T )
Especially, the fast delta distance is also a
lower bound to the congruence distance. For
time series of length n the complexity of the
fast delta distance d∆ improves to n log n. On
the other hand, equivalence classes regarding
the fast delta distance might include time series
which are not congruent.
The Fast Greedy Delta Distance:
Incor-
porating the idea for improving the runtime into
the greedy delta distance simply changes Line 7
of Algorithm 1: We only consider values for the
variable j, which add a power of 2 to the vari-
able i. Algorithm 2 provides the line to change
in Algorithm 1 in order to achieve the fast greedy
delta distance.
Algorithm 2 Distinction between Greedy Delta
Distance (cf. Algorithm 1) and Fast Greedy
Delta Distance
for j ∈ 2
N
with j (cid:54) n − 1
7
The fast greedy delta distance is again dom-
inated by the sorting of elements. This time,
n log n elements have to be sorted, thus its
complexity is n log(n) log(n log n) = n log(n)2.
Hence, the fast versions both have quasi linear
runtime regarding length and linear runtime re-
garding dimensionality.
An inequality such as in Theorem 2.6 does not
exist for the fast greedy delta distance. Also,
there is no correlation between the (fast) delta
distance and the (fast) greedy distance. Though,
the evaluation shows that the greedy delta dis-
tance provides a much better approximation in
most cases. Call for Section 3 for an evaluation
of their tightness to the congruence distance.
3 Evaluation
Since the exact computation of the congruence
distance is a computational hard problem (an
6
thus not feasable in practical applications), we
are mainly interested in the evaluation of the ap-
proximations. Unfortunately, there is no direct
algorithm for the computation of the congruence
distance and we have to consider the computa-
tion of the congruence distance as a nonlinear
optimization problem. For two time series S and
T , we will denote the distance value computed by
an optimizer with dO(S, T ). Since an optimizer
might not find the global optimum, all values for
the congruence distance (computed by an opti-
mizer) in this section, are in fact upper bounds to
the correct but unknown value of the congruence
distance, i. e. dC(S, T ) (cid:54) dO(S, T ). This given
circumstance complicates the evaluation of our
approximations to the congruence distance.
To estimate the tightness of the approxima-
tions, we first evaluate our optimizer on prob-
lems for which we know the correct results
(cf. Section 3.1). In those cases, where the er-
ror of the optimizer is small, the estimation of
the tightness of our approximations is accurate.
On the other hand, when the error of the opti-
mizer increases, our estimation of the tightness
of our approximations are loose and the approx-
imation might be tighter than the experiments
claim.
For a detailed explanation, consider a lower
bound (cid:96)(S, T ) for the congruence distance (e. g.
(cid:96) might be one of dG, dG, d∆, or d∆) and suppose
dC(S, T ) = dO(S, T ) − ε, i. e. ε (cid:62) 0 is the error
of the optimizer. Then, we have the following
correlation between the estimated tightness and
the real tightness:
(cid:96)(S, T )
dC(S, T )
=
(cid:96)(S, T )
dO(S, T ) − ε
(cid:62) (cid:96)(S, T )
dO(S, T )
Hence, for small errors ε, the estimated tightness
is accurate and for large errors ε we underesti-
mate the tightness. In Section 3.2 and 3.3, we
Figure 1: Boxplot: distance values (left) and
runtimes (right) from our Optimizer on congru-
ent time series.
evaluate the tightness and the speedup of ap-
proximations to the (optimizer based) congru-
ence distance, respectively.
3.1 Congruence Distance: An Opti-
mization Problem
Consider fixed time series S and T in Rk with
length n. The congruence distance is a nonlinear
optimization problem with equality based con-
straints. The function to minimize is
n−1(cid:88)
f (M, v) =
d (si, M · ti + v)
i=0
while the k2 equality based constraints corre-
spond the the constraints for orthogonal matri-
ces:
M · M T = I.
As a initial "solution" for the optimizer, we sim-
ply choose M = I and v = 0.
We manually transformed time series T with a
random orthogonal matrix M∗ and a small ran-
dom vector v∗ and solved the optimization prob-
lem dC(T, M∗ · T + v∗) to examine whether our
optimizer is working properly. Clearly, we ex-
pect the optimizer to find a solution with value
7
0 1 2 3 4 5 6 7 81234Congruence DistanceDimensionality 0 5 10 15 20 251234Computation TimeDimensionality 0. Whenever the optimizer claimed large dis-
tance values, we concluded that the optimizer is
not working. We tried different optimizer strate-
gies and chose an augmented lagrangian algo-
rithm [7] with the BOBYQA algorithm [13] as
local optimizer for further experiments because
it promised the best performance with these ex-
periments. We used the implementations pro-
vided by the NLopt library [1].
We used the RAM dataset generator [4] to eval-
uate the optimizer on time series with vary-
ing dimensionality. We solved 400 optimiza-
tion problems for varying dimensionality and
removed all of those runs where the optimizer
did not find a reasonable solution (i. e.
runs
where the optimizer yielded solutions larger than
100). Figure 1 shows the distance values pro-
posed by the optimizer (and therefore the error
it makes) per dimensionality up to dimensional-
ity 4. For higher dimensionalities, the optimizer
completely failed to find any reasonable value
near 0 although we gave it enough resources of
any kind (e. g. number of iterations, computa-
tion time, etc.). Figure 1 also shows that the
computation times rapidly increase with increas-
ing dimensionality. Because of the raising error
and runtime with increasing dimensionality, an
evaluation of the congruence distance on higher
dimensionality is not feasable. Hence, we can
only consider up to 4-dimensional time series in
all further experiments.
3.2 Tightness of Approximations
In order to evaluate the tightness of the (fast)
delta distance and (fast) greedy delta distance
as lower bounds to the congruence distance, we
used the RAM dataset generator [4] as well as a
real world dataset with 2-dimensional time se-
ries (character trajectories [8], contains over 2800
Figure 2: Average tightness of the delta distance
(top left), the fast delta distance (top right), the
greedy delta distance (bottom left), and the fast
greedy delta distance (bottom right) to the con-
gruence distance, respectively.
time series). Other real world datasets with
higher dimensionality have not been suitable be-
cause the optimizer failed to compute the con-
gruence distance.
Since making the (greedy) delta distance time
warping aware is future work, we have to deal
with time warping another way. We simply pre-
process our datasets, such that each time series,
seen as a trajectory, moves with constant speed,
i. e. for each dewarped time series, the following
holds:
d (ti, ti+1) ≈ d (ti+1, ti+2) .
We achieve this property by simply reinterpolat-
ing the time series regarding the arc length.
Figure 2 shows the tightness of the approxi-
mations on RAM datasets. As we expected, the
greedy delta distance provides the tightest ap-
8
0 0.2 0.4 0.6 0.8 11234TightnessDimensionalityDelta Distance 0 0.2 0.4 0.6 0.8 11234TightnessDimensionalityFast Delta Distance 0 0.2 0.4 0.6 0.8 11234TightnessDimensionalityGreedy Delta Distance 0 0.2 0.4 0.6 0.8 11234TightnessDimensionalityFast Greedy Delta Distance Figure 3: Average speedup of the approxima-
tions to our optimizer
proximation to the congruence distance (pro-
vided by our optimizer).
As we observed in Section 3.1, the error of
our optimizer increases with increasing dimen-
sionality. Hence, the tightness of the optimizer
to the real congruence distance is decreasing.
Since we can observe a similar behaviour here
(the tightness of the approximation is decreas-
ing with increasing dimensionality), the reason
might be the inaccuracy of the optimizer. Ei-
ther way, we can see that the tightness is above
50% in most cases. Especially when using the
greedy delta distance, the tightness is above 75%
in most cases.
On the character trajectories dataset, the
delta distance and the greedy delta distance
achieved a tightness of 63% and 83%, respec-
tively.
3.3 Speedup of Approximations
Figure 3 shows the speedup of the approxima-
tions to the optimizer. As expected, the speedup
increases exponentially with increasing dimen-
sionality. While the fast delta distance is the
fastest algorithm, it also provides the worst ap-
proximation (compare with Figure 2). On the
other hand, the greedy delta distance provides
the best approximation while being the slowest
algorithm. Still, the greedy delta distance is mul-
tiple orders of magnitudes faster than our opti-
mizer.
The following speedups have been achieved on
the character trajectory dataset: 1642 with the
delta distance; 8040 with the fast delta distance;
321 with the greedy delta distance; 2287 with
the fast greedy delta distance. The results are
similar to those on the RAM generated datasets.
4 Conclusion and Future Work
In this paper, we analyzed the problem of mea-
suring the congruence between two time series.
We provided four measures for approximating
the congruence distance which are at least 2 or-
ders of magnitude faster than the congruence
distance. The first (namely, the delta distance)
provides the additional ability to be used in
metrix indexing structures. The second (greedy
delta distance) loses this benefit, but seems to
achieve a better approximation. Both approx-
imations have linear complexity regarding the
dimensionality but at least quadratic complex-
ity regarding the length of the time series. The
other two approximations address this problem
at a cost of approximation quality. They have
quasi-linear runtime regarding the length.
In practical applications, time series distance
functions need to be robust against time warp-
ing. The approximations provided in this work
are based on comparing self-similarity matrices
of time series. Based on this idea, our next step
is to develop a time warping distance function
measuring the congruency.
9
10 100 1000 10000 100000 1x106 1 1.5 2 2.5 3 3.5 4Average Computation TimeDimensionalityDelta DistanceFast Delta DistanceGreedy Delta DistanceFast Greedy Delta Distance [10] T. Eiter and H. Mannila. Computing dis-
crete fr´echet distance. Technical report,
Technische Universitat Wien, 1994.
[11] P. J. Heffernan and S. Schirra. Approximate
decision algorithms for point set congru-
ence. In Proc. of the Symposium on Com-
put. Geometry, SCG, pages 93 -- 101. ACM,
1992.
[12] P. Indyk and S. Venkatasubramanian. Ap-
proximate congruence in nearly linear time.
Comput. Geom. Theory Appl., 24(2):115 --
128, Feb. 2003.
[13] M. J. D. Powell. The bobyqa algorithm
for bound constrained optimization without
derivatives. 01 2009.
[14] H. Sakoe and S. Chiba. Readings in speech
recognition. In A. Waibel and K.-F. Lee, ed-
itors, Readings in Speech Recognition, chap-
ter Dynamic Programming Algorithm Op-
timization for Spoken Word Recognition,
pages 159 -- 165. Morgan Kaufmann Publish-
ers Inc., San Francisco, CA, USA, 1990.
References
[1] Nlopt.
https://nlopt.readthedocs.io.
Accessed: 2018-05-15.
[2] H. Alt and L. J. Guibas. Discrete geomet-
ric shapes: Matching, interpolation, and ap-
proximation: A survey. Technical report,
Handbook of Comput. Geometry, 1996.
[3] H. Alt, K. Mehlhorn, H. Wagener, and
E. Welzl. Congruence, similarity and sym-
metries of geometric objects. Discrete Com-
put. Geom., 3(3):237 -- 256, Jan. 1988.
[4] J. P. Bachmann and J. Freytag. High di-
mensional time series generators. CoRR,
abs/1804.06352, 2018.
[5] J. P. Bachmann, J. Freytag, B. Hauskeller,
and N. Schweikardt. Measuring congruence
on high dimensional time series. CoRR,
abs/1805.10697, 2018.
[6] J. P. Bachmann and J.-C. Freytag. Dynamic
Time Warping and the (Windowed) Dog-
Keeper Distance, pages 127 -- 140. Springer
International Publishing, Cham, 2017.
[7] E. G. Birgin and J. M. Mart´ınez. Improv-
ing ultimate convergence of an augmented
lagrangian method. Optimization Methods
Software, 23(2):177 -- 195, Apr. 2008.
[8] D. Dheeru and E. Karra Taniskidou. UCI
machine learning repository, 2017.
[9] J.-P. Eckmann, S. O. Kamphorst, and
D. Ruelle. Recurrence plots of dynami-
cal systems. EPL (Europhysics Letters),
4(9):973, 1987.
10
|
1304.7067 | 1 | 1304 | 2013-04-26T04:55:07 | Detecting regularities on grammar-compressed strings | [
"cs.DS"
] | We solve the problems of detecting and counting various forms of regularities in a string represented as a Straight Line Program (SLP). Given an SLP of size $n$ that represents a string $s$ of length $N$, our algorithm compute all runs and squares in $s$ in $O(n^3h)$ time and $O(n^2)$ space, where $h$ is the height of the derivation tree of the SLP. We also show an algorithm to compute all gapped-palindromes in $O(n^3h + gnh\log N)$ time and $O(n^2)$ space, where $g$ is the length of the gap. The key technique of the above solution also allows us to compute the periods and covers of the string in $O(n^2 h)$ time and $O(nh(n+\log^2 N))$ time, respectively. | cs.DS | cs |
Detecting regularities on grammar-compressed
strings
Tomohiro I1,2, Wataru Matsubara3, Kouji Shimohira1,
Shunsuke Inenaga1, Hideo Bannai1, Masayuki Takeda1,
Kazuyuki Narisawa3, and Ayumi Shinohara3
1 Department of Informatics, Kyushu University, Japan
{tomohiro.i, inenaga, bannai, takeda}@inf.kyushu-u.ac.jp
2 Japan Society for the Promotion of Science (JSPS)
3 Graduate School of Information Sciences, Tohoku University, Japan
{narisawa,ayumi}@ecei.tohoku.ac.jp
Abstract. We solve the problems of detecting and counting various
forms of regularities in a string represented as a Straight Line Program
(SLP). Given an SLP of size n that represents a string s of length N ,
our algorithm compute all runs and squares in s in O(n3h) time and
O(n2) space, where h is the height of the derivation tree of the SLP. We
also show an algorithm to compute all gapped-palindromes in O(n3h +
gnh log N ) time and O(n2) space, where g is the length of the gap. The
key technique of the above solution also allows us to compute the periods
and covers of the string in O(n2h) time and O(nh(n + log2 N )) time,
respectively.
1
Introduction
Finding regularities such as squares, runs, and palindromes in strings, is a fun-
damental and important problem in stringology with various applications, and
many efficient algorithms have been proposed (e.g., [12,6,1,7,13,2,9]). See also [5]
for a survey.
In this paper, we consider the problem of detecting regularities in a string
s of length N that is given in a compressed form, namely, as a straight line
program (SLP), which is essentially a context free grammar in the Chomsky
normal form that derives only s. Our model of computation is the word RAM:
We shall assume that the computer word size is at least ⌈log2 N ⌉, and hence,
standard operations on values representing lengths and positions of string s can
be manipulated in constant time. Space complexities will be determined by the
number of computer words (not bits).
Given an SLP whose size is n and the height of its derivation tree is h,
Bannai et al. [3] showed how to test whether the string s is square-free or
not, in O(n3h log N ) time and O(n2) space. Independently, Khvorost [8] pre-
sented an algorithm for computing a compact representation of all squares in
s in O(n3h log2 N ) time and O(n2) space. Matsubara et al. [14] showed that a
compact representation of all maximal palindromes occurring in the string s can
be computed in O(n3h) time and O(n2) space. Note that the length N of the
decompressed string s can be as large as O(2n) in the worst case. Therefore, in
such cases these algorithms are more efficient than any algorithm that work on
uncompressed strings.
In this paper we present the following extension and improvements to the
above work, namely,
1. an O(n3h)-time O(n2)-space algorithm for computing a compact represen-
tation of squares and runs;
2. an O(n3h+gnh log N )-time O(n2)-space algorithm for computing a compact
representation of palindromes with a gap (spacer) of length g.
We remark that our algorithms can easily be extended to count the number of
squares, runs, and gapped palindromes in the same time and space complexities.
Note that Result 1 improves on the work by Khvorost [8] which requires
O(n3h log2 N ) time and O(n2) space. The key to the improvement is our new
technique of Section 3.3 called approximate doubling, which we believe is of
independent interest. In fact, using the approximate doubling technique, one
can improve the time complexity of the algorithms of Lifshits [10] to compute
the periods and covers of a string given as an SLP, in O(n2h) time and O(nh(n+
log2 N )) time, respectively.
If we allow no gaps in palindromes (i.e., if we set g = 0), then Result 2 implies
that we can compute a compact representation of all maximal palindromes in
O(n3h) time and O(n2) space. Hence, Result 2 can be seen as a generalization
of the work by Matsubara et al. [14] with the same efficiency.
2 Preliminaries
2.1 Strings
Let Σ be the alphabet, so an element of Σ∗ is called a string. For string s = xyz,
x is called a prefix, y is called a substring, and z is called a suffix of s, respectively.
The length of string s is denoted by s. The empty string ε is a string of length
0, that is, ε = 0. For 1 ≤ i ≤ s, s[i] denotes the i-th character of s. For
1 ≤ i ≤ j ≤ s, s[i..j] denotes the substring of s that begins at position i and
ends at position j. For any string s, let sR denote the reversed string of s, that
is, sR = s[s] · · · s[2]s[1]. For any strings s and u, let lcp(s, u) (resp. lcs(s, u))
denote the length of the longest common prefix (resp. suffix) of s and u.
We say that string s has a period c (0 < c ≤ s) if s[i] = s[i + c] for any
1 ≤ i ≤ s − c. For a period c of s, we denote s = uq, where u is the prefix of
s of length c and q = s
c . For convenience, let u0 = ε. If q ≥ 2, s = uq is called
a repetition with root u and period u. Also, we say that s is primitive if there
is no string u and integer k > 1 such that s = uk. If s is primitive, then s2 is
called a square.
We denote a repetition in a string s by a triple hb, e, ci such that s[b..e] is a
repetition with period c. A repetition hb, e, ci in s is called a run (or maximal
2
periodicity in [11]) if c is the smallest period of s[b..e] and the substring cannot
be extended to the left nor to the right with the same period, namely neither
s[b − 1..e] nor s[b..e + 1] has period c. Note that for any run hb, e, ci in s, every
substring of length 2c in s[b..e] is a square. Let Run(s) denote the set of all runs
in s.
A string s is said to be a palindrome if s = sR. A string s said to be a gapped
palindrome if s = xuxR for some string u ∈ Σ∗. Note that u may or may not be
a palindrome. The prefix x (resp. suffix xR) of xuxR is called the left arm (resp.
right arm) of gapped palindrome xuuR. If u = g, then xuxR is said to be a
g-gapped palindrome. We denote a maximal g-gapped palindrome in a string s
by a pair hb, eig such that s[b..e] is a g-gapped palindrome and s[b − 1..e + 1] is
not. Let gPals(s) denote the set of all maximal g-gapped palindromes in s.
Given a text string s ∈ Σ+ and a pattern string p ∈ Σ+, we say that p
occurs at position i (1 ≤ i ≤ s − p + 1) iff s[i..i + p − 1] = p. Let Occ(s, p)
denote the set of positions where p occurs in s. For a pair of integers 1 ≤ b ≤ e,
[b, e] = {b, b + 1, . . . , e} is called an interval.
Lemma 1 ([15]). For any strings s, p ∈ Σ+ and any interval [b, e] with 1 ≤ b ≤
e ≤ b + p, Occ(s, p) ∩ [b, e] forms a single arithmetic progression if Occ(s, p) ∩
[b, e] 6= ∅.
2.2 Straight-line programs
A straight-line program (SLP ) S of size n is a set of productions S = {Xi →
expr i}n
i=1, where each Xi is a distinct variable and each expr i is either expri =
XℓXr (1 ≤ ℓ, r < i), or expri = a for some a ∈ Σ. Note that Xn derives only a
single string and, therefore, we view the SLP as a compressed representation of
the string s that is derived from the variable Xn. Recall that the length N of the
string s can be as large as O(2n). However, it is always the case that n ≥ log N .
For any variable Xi, let val (Xi) denote the string that is derived from variable
Xi. Therefore, val (Xn) = s. When it is not confusing, we identify Xi with the
string represented by Xi.
Let Ti denote the derivation tree of a variable Xi of an SLP S. The derivation
tree of S is Tn (see also Fig. 5 in Appendix C). Let height (Xi) denote the height
of the derivation tree Ti of Xi and height (S) = height (Xn). We associate each
leaf of Ti with the corresponding position of the string val (Xi). For any node z
of the derivation tree Ti, let ℓz be the number of leaves to the left of z in Ti.
The position of z in Ti is ℓz + 1.
Let [u, v] be any integer interval with 1 ≤ u ≤ v ≤ val (Xi). We say that
the interval [u, v] crosses the boundary of node z in Ti, if the lowest common
ancestor of the leaves u and v in Ti is z. We also say that the interval [u, v]
touches the boundary of node z in Ti, if either [u − 1, v] or [u, v + 1] crosses the
boundary of z in Ti. Assume p = w[u..u + p − 1] and interval [u, u + p − 1]
crosses or touches the boundary of node z in Ti. When z is labeled by Xj, then
we also say that the occurrence of p starting at position u in val (Xi) crosses or
touches the boundary of Xj.
3
Lemma 2 ([4]). Given an SLP S of size n describing string w of length N ,
we can pre-process S in O(n) time and space to answer the following queries in
O(log N ) time:
-- Given a position u with 1 ≤ u ≤ N , answer the character w[u].
-- Given an interval [u, v] with 1 ≤ u ≤ v ≤ N , answer the node z the interval
[u, v] crosses, the label Xi of z, and the position of z in TS = Tn.
For any production Xi → XℓXr and a string p, let Occξ(Xi, p) be the set of
occurrences of p which begin in Xℓ and end in Xr. Let S and T be SLPs of sizes
n and m, respectively. Let the AP-table for S and T be an n × m table such
that for any pair of variables X ∈ S and Y ∈ T the table stores Occξ(X, Y ).
It follows from Lemma 1 that Occξ(X, Y ) forms a single arithmetic progression
which requires O(1) space, and hence the AP-table can be represented in O(nm)
space.
Lemma 3 ([10]). Given two SLPs S and T of sizes n and m, respectively, the
AP-table for S and T can be computed in O(nmh) time and O(nm) space, where
h = height (S).
Lemma 4 ([10], local search (LS)). Using AP-table for S and T that de-
scribe strings p in s, we can compute, given any position b and constant α > 0,
Occ(s, p) ∩ [b, b + αp] as a form of at most ⌈α⌉ arithmetic progressions in O(h)
time, where h = height (S).
Note that, given any 1 ≤ i ≤ j ≤ s, we are able to build an SLP of size
O(n) that generates substring s[i..j] in O(n) time. Hence, by computing the
AP-table for S and the new SLP, we can conduct the local search LS operation
on substring s[i..j] in O(n2h) time.
For any variable Xi of S and positions 1 ≤ k1, k2 ≤ Xi, we define the
"right-right" longest common extension query by
LCE(Xi, k1, k2) = lcp(Xi[k1..Xi], Xi[k2..Xi]).
Using a technique of [15] in conjunction with Lemma 3, it is possible to answer
the query in O(n2h) time for each pair of positions, with no pre-processing.
We will later show our new algorithm which, after O(n2h)-time pre-processing,
answers to the LCE query for any pair of positions in O(h log N ) time.
3 Finding runs
In this section we propose an O(n3h)-time and O(n2)-space algorithm to com-
pute O(n log N )-size representation of all runs in a text s of length N represented
by SLP S = {Xi → expri }n
i=1 of height h.
For each production Xi → Xℓ(i)Xr(i) with i ≤ n, we consider the set
Run ξ(Xi) of runs which touch or cross the boundary of Xi and are completed
in Xi, i.e., those that are not prefixes nor suffixes of Xi. Formally,
Run ξ(Xi) = {hb, e, ci ∈ Run(Xi) 1 ≤ b − 1 ≤ Xℓ(i) < e + 1 ≤ Xi}.
4
It is known that for any interval [b, e] with 1 ≤ b ≤ e ≤ s, there exists a unique
occurrence of a variable Xi in the derivation tree of SLP, such that the interval
[b, e] crosses the boundary of Xi. Also, wherever Xi appears in the derivation
tree, the runs in Run ξ(Xi) occur in s with some appropriate offset, and these
occurrences of the runs are never contained in Run ξ(Xj) with any other variable
Xj with j 6= i. Hence, by computing Run ξ(Xi) for all variables Xi with i ≤ n,
we can essentially compute all runs of s that are not prefixes nor suffixes of s.
In order to detect prefix/suffix runs of s, it is sufficient to consider two auxiliary
variables Xn+1 → X$Xn and Xn+2 → Xn+1X$′, where X$ and X$′ respectively
derive special characters $ and $′ that are not in s and $ 6= $′. Hence, the
problem of computing the runs from an SLP S reduces to computing Run ξ(Xi)
for all variables Xi with i ≤ n + 2.
Our algorithm is based on the divide-and-conquer method used in [3] and
also [8], which detect squares crossing the boundary of each variable Xi. Roughly
speaking, in order to detect such squares we take some substrings of val (Xi) as
seeds each of which is in charge of distinct squares, and for each seed we detect
squares by using LS and LCE constant times. There is a difference between [3]
and [8] in how the seeds are taken, and ours is rather based on that in [3].
In the next subsection, we briefly describe our basic algorithm which runs in
O(n3h log N ) time.
3.1 Basic algorithm
Consider runs in Run ξ(Xi) with Xi → XℓXr. Since a run in Run ξ(Xi) contains
a square which touches or crosses the boundary of Xi, our algorithm finds a run
by first finding such a square, and then computing the maximal extension of its
period to the left and right of its occurrence.
We divide each square ww by its length and how it relates to the boundary
of Xi. When w > 1, there exists 1 ≤ t < log val (Xi) such that 2t ≤ w < 2t+1
and there are four cases (see also Fig. 1); (1) wℓ ≥ 3
2 w > wℓ ≥ w,
(3) w > wℓ ≥ 1
2 w > wℓ, where wℓ is a prefix of ww which is also
a suffix of val (Xℓ).
2 w, (4) 1
2 w, (2) 3
The point is that in any case we can take a substring p of length 2t−1 of s
which touches the boundary of Xi, and is completely contained in w. By using
p as a seed we can detect runs by the following steps:
Step 1: Conduct local search of p in an "appropriate range" of Xi, and find a
copy p′ (= p) of p.
Step 2: Compute the length plen of the longest common prefix to the right of
p and p′, and the length slen of the longest common suffix to the left of p
and p′, then check that plen + slen ≥ d − p, where d is the distance between
the beginning positions of p and p′.
Notice that Step 2 actually computes maximal extension of the repetition.
Since d = w, it is sufficient to conduct local search in the range satisfying
2t ≤ d < 2t+1, namely, the width of the interval for local search is smaller than
5
Xi
Xℓ
Xr
case (1)
case (2)
case (3)
case (4)
w
p'
w
p'
t-1
2
t-1
2
w
p
w
p
w
p
w
p'
w
p
w
p'
Fig. 1. The left arrows represent the longest common suffix between the left substrings
immediately to the left of p and p′. The right arrows represent the longest common
prefix between the substrings immediately to the right of p and p′.
2p, and all occurrences of p′ are represented by at most two arithmetic progres-
sions. Although exponentially many runs can be represented by an arithmetic
progression, its periodicity enables us to efficiently detect all of them, by using
LCE only constant times, and they are encoded in O(1) space. We put the details
in Appendix A since the employed techniques are essentially the same as in [8].
By varying t from 1 to log N , we can obtain an O(log N )-size compact rep-
resentation of Run ξ(Xi) in O(n2h log N ) time. More precisely, we get a list of
j=0 hδ1−cj, δ2+
cj, δ3 + cji for all elements of the list equals to Run ξ(Xi) without duplicates. By
applying the above procedure to all the n variables, we can obtain an O(n log N )-
size compact representation of all runs in s in O(n3h log N ) time. The total space
requirement is O(n2), since we need O(n2) space at each step of the algorithm.
In order to improve the running time of the algorithm to O(n3h), we will use
O(log N ) quintuplets hδ1, δ2, δ3, c, ki such that the union of setsSk−1
new techniques of the two following subsections.
3.2 Longest common extension
In this subsection we propose a more efficient algorithm for LCE queries.
Lemma 5. We can pre-process an SLP S of size n and height h in O(n2h) time
and O(n2) space, so that given any variable Xi and positions 1 ≤ k1, k2 ≤ Xi,
LCE(Xi, k1, k2) is answered in O(h log N ) time.
6
To compute LCE(Xi, k1, k2) we will use the following function: For an SLP
S = {Xi → expri}n
i=1, let Match be a function such that
Match(Xi, Xj, k) =(true
false
if k ∈ Occ(Xi, Xj),
if k /∈ Occ(Xi, Xj).
Lemma 6. We can pre-process a given SLP S of size n and height h in O(n2h)
time and O(n2) space so that the query Match(Xi, Xj, k) is answered in O(log N )
time.
Proof. We apply Lemma 2 to every variable Xi of S, so that the queries of
Lemma 2 is answered in O(log N ) time on the derivation tree Ti of each variable
Xi of S. Since there are n variables in S, this takes a total of O(n2) time and
space. We also apply Lemma 3 to S, which takes O(n2h) time and O(n2) space.
Hence the pre-processing takes a total of O(n2h) time and O(n2) space.
To answer the query Match(Xi, Xj, k), we first find the node of Ti the interval
[k, k + Xj − 1] crosses, its label Xq, and its position r in Ti. This takes O(log N )
time using Lemma 2. Then we check in O(1) time if (k − r) ∈ Occξ(Xq, Xj) or
not, using the arithmetic progression stored in the AP-table. Thus the query is
⊓⊔
answered in O(log N ) time.
The following function will also be used in our algorithm: Let FirstMismatch
be a function such that
FirstMismatch(Xi, Xj, k) =(lcp(Xi[k..Xi], Xj)
undefined
if Xi − k + 1 ≤ Xj,
otherwise.
Using Lemma 6 we can establish the following lemma. See Appendix B for a
full proof.
Lemma 7. We can pre-process a given SLP S of size n and height h in O(n2h)
time and O(n2) space so that the query FirstMismatch(Xi, Xj, k) is answered in
O(h log N ) time.
We are ready to prove Lemma 5:
Proof. Consider to compute LCE(Xi, k1, k2). Without loss of generality, assume
k1 ≤ k2. Let z be the lca of the k1-th and (k2−k1+Xi)-th leaves of the derivation
tree Ti. Let Pℓ be the path from z to the k1-th leaf of the derivation tree Ti, and
let L be the list of the right child of the nodes in Pℓ sorted in increasing order of
their position in Ti. The number of nodes in L is at most height (Xi) ≤ h, and L
can be computed in O(height (Xi)) = O(h) time. Let Pr be the path from z to
the (k2 −k1+Xi)-th leaf of the derivation tree Ti, and let R be the list of the left
child of the nodes in Pr sorted in increasing order of their position in Ti. R can
be computed in O(h) time as well. Let U = L∪R = {Xu(1), Xu(2), . . . , Xu(m)} be
the list obtained by concatenating L and R. For each Xu(p) in increasing order
q=1 Xu(q)) until
of p = 1, 2, . . . , m, we perform query Match(Xi, Xu(p), k1 +Pp−1
7
either finding the first variable Xu(p′) for which the query returns false (see also
Fig. 6 in Appendix C), or all the queries for p = 1, . . . , m have returned true.
In the latter case, clearly LCE(Xi, k1, k2) = Xi − k1 + 1. In the former case,
the first mismatch occurs between Xi and Xu(p′), and hence LCE(Xi, k1, k2) =
Pp′−1
q′=1 Xu(q′) + FirstMismatch(Xi, Xu(p′), k1 +Pp′−1
Since U contains at most 2 · height (Xi) variables, we perform O(h) Match
queries. We perform at most one FirstMismatch query. Thus, using Lemmas 6
and 7, we can compute LCE(Xi, k1, k2) in O(h log N ) time after O(n2h)-time
O(n2)-space pre-processing.
⊓⊔
q′=1 Xu(q′)).
We can use Lemma 5 to also compute "left-left", "left-right", and "right-
left" longest common extensions on the uncompressed string s = val (S): We
can compute in O(n) time an SLP S R of size n which represents the reversed
string sR [14]. We then construct a new SLP S′ of size 2n and height h + 1 by
concatenating the last variables of S and S R, and apply Lemma 5 to S′.
3.3 Approximate doubling
Here we show how to reduce the number of AP-table computation required in
Step 1 of the basic algorithm, from O(log N ) to O(1) times per variable.
Consider any production Xi → XℓXr. If we build a new SLP which contains
variables that derive the prefixes of length 2t of Xr for each 0 ≤ t < log Xr,
we can obtain the AP-tables for Xi and all prefix seeds of Xr by computing the
AP-table for Xi and the new SLP. Unfortunately, however, the size of such a
new SLP can be as large as O(n log N ). Here we notice that the lengths of the
seeds do not have to be exactly doublings, i.e., the basic algorithm of Section 3.1
works fine as long as the following properties are fulfilled: (a) the ratio of the
lengths for each pair of consecutive seeds is constant; (b) the whole string is
covered by the O(log N ) seeds 4. We show in the next lemma that we can build
an approximate doubling SLP of size O(n).
Lemma 8. Let S = {Xi → expr i}n
i=1 be an SLP that derives a string s. We
can build in O(n) time a new SLP S′ = {Yi → expr ′
i=1 with n′ = O(n)
and height (S′) = O(height (S)), which derives s and contains O(log N ) variables
Ya1 , Ya2, . . . , Yak satisfying the following conditions:
i}n′
-- For any 1 ≤ j ≤ k, Yaj derives a prefix of s, Ya1 = 1 and Yak = s.
-- For any 1 ≤ j < k, Yaj < Yaj+1 ≤ 2Yaj .
Proof. First, we copy the productions of S into S′. Next we add productions
needed for creating prefix variables Ya1 , Ya2, . . . , Yak in increasing order. We con-
sider separating the derivation tree Tn of Xn into segments by a sequence of nodes
v1, v2, . . . , vk such that the i-th segment enclosed by the path from vi to vi+1
represents the suffix of Yai+1 of length Yai+1 − Yai , namely, Yai+1 → Yai Ybi
4 A minor modification is that we conduct local search for a seed p at Step 1 with the
range satisfying 2p ≤ d < 2q, where q is the next longer seed of p.
8
where Ybi is a variable for the i-th segment. Each node vi is called an l-node
(resp. r-node) if the node belongs to the left (resp. right) segment of the node.
We start from v1 which is the leftmost node that derives s[1]. Suppose we
have built prefix variables up to Yai and now creating Yai+1. At this moment we
are at vi. We move up to the node ui such that ui is the deepest node on the
path from the root to vi which contains position 2Yai , and move down from
ui towards position 2Yai. The traversal ends when we meet a node vi+1 which
satisfies one of the following conditions; (1) the rightmost position of vi+1 is
2Yai, (2) vi+1 is labeled with Xj, and we have traversed another node labeled
with Xj before.
-- If Condition (1) holds, vi+1 is set to be an l-node. It is clear that the length
of the i-th segment is exactly Yai and Yai+1 = 2Yai.
-- If Condition (1) does not hold but Condition (2) holds, vi+1 is set to be an
r-node. Since vi+1 contains position 2Yai, the length of the i-th segment
is less than Yai and Yai+1 < 2Yai. We remark that since Xj appears in
Yai+1 , then Yai+1+ Xj ≤ 2Yai+1 , and therefore, we never move down vi+1
for the segments to follow.
We iterate the above procedures until we obtain a prefix variable Yak−1 that
satisfies Xn ≤ 2Yak−1. We let uk be the deepest node on the path from the
root to vk−1 which contains position s, and let vk be the right child of uk. Since
Yai < 2Yai+2 for any 1 ≤ i < k, k = O(log N ) holds.
We note that the i-th segment can be represented by the concatenation of
"inner" nodes attached to the path from vi to vi+1, and hence, the number of
new variables needed for representing the segment is bounded by the number of
such nodes. Consider all the edges we have traversed in the derivation tree Tn
of Xn. Each edge contributes to at most one new variable for some segment (see
also Fig. 7 in Appendix C). Since each variable Xj is used constant times for
moving down due to Condition (2), the number of the traversed edges as well
as n′ is O(n). Also, it is easy to make the height of Ybi be O(height (S)) for any
1 ≤ i < k. Thus O(height (S′)) = O(log N + height (S)) = O(height (S)).
⊓⊔
3.4 Improved algorithm
Using Lemmas 5 and 8, we get the following theorem.
Theorem 1. Given an SLP S of size n and height h that describes string s
of length N , an O(n log N )-size compact representation of all runs in s can be
computed in O(n3h) time and O(n2) working space.
Proof. Using Lemma 5, we first pre-process S in O(n2h) time so that any "right-
right" or "left-left" LCE query can be answered in O(h log N ) time. For each
variable Xi → XℓXr, using Lemma 8, we build temporal SLPs T and T ′ which
have respectively approximately doubling suffix variables of Xℓ and prefix vari-
ables of Xr, and compute two AP-tables for S and each of them in O(n2h) time.
For each of the O(log N ) prefix/suffix variables, we use it as a seed and find
9
Xi
wR
p
w
p
g
Xℓ
w
p'
g
w
Xr
wR
p'
g
wR
case (1)
case (2)
case (3)
Fig. 2. Three groups of g-gapped palindromes to be found in Xi.
all corresponding runs by using LS and LCE queries constant times. Hence the
time complexity is O(n2h + n(n2h + (h + h log N ) log N )) = O(n3h). The space
requirement is O(n2), the same as the basic algorithm.
⊓⊔
4 Finding g-gapped palindromes
A similar strategy to finding runs on SLPs can be used for computing a compact
representation of the set gPals(s) of g-gapped palindromes from an SLP S that
describes string s. As in the case of runs, we add two auxiliary variables Xn+1 →
X$Xn and Xn+2 → Xn+1X$′ . For each production Xi → XℓXr with i ≤ n + 2,
we consider the set gPals ξ(Xi) of g-gapped palindromes which touch or cross
the boundary of Xi and are completed in Xi, i.e., those that are not prefixes nor
suffixes of Xi. Formally,
gPals ξ(Xi) = {hb, eig ∈ gPals(Xi) 1 ≤ b − 1 ≤ Xℓ < e + 1 ≤ Xi}.
Each g-gapped palindrome in Xi can be divided into three groups (see also
Fig. 2); (1) its right arm crosses or touches with its right end the boundary of
Xi, (2) its left arm crosses or touches with its left end the boundary of Xi, (3)
the others.
For Case (3), for every Xℓ−g +1 ≤ j < Xℓ we check if lcp(Xi[1..j]R, Xi[j +
g + 1..Xi]) > 0 or not. From Lemma 5, it can be done in O(gh log N ) time for
any variable by using "left-right" LCE (excluding pre-processing time for LCE).
Hence we can compute all such g-gapped palindromes for all productions in
O(n2h + gnh log N ) time, and clearly they can be stored in O(ng) space.
For Case (1), let wℓ be the prefix of the right arm which is also a suffix of
val (Xℓ). We take approximately doubling suffixes of Xℓ as seeds. Let p be the
longest seed that is contained in wℓ. We can find g-gapped palindromes by the
following steps:
10
Step 1: Conduct local search of p′ = pR in an "appropriate range" of Xi and
find it in the left arm of palindrome.
Step 2: Compute "right-left" LCE of p′ and p, then check that the gap can be g.
The outward maximal extension can be obtained by computing "left-right"
LCE queries on the occurrences of p′ and p.
As in the case of runs, for each seed, the length of the range where the local
search is performed in Step 1 is only O(p). Hence, the occurrences of p′ can
be represented by a constant number of arithmetic progressions. Also, we can
obtain O(1)-space representation of g-gapped palindromes for each arithmetic
progression representing overlapping occurrences of p′, by using a constant num-
ber of LCE queries. Therefore, by processing O(log N ) seeds for every variable
Xi, we can compute in O(n2h + n(n2h + (h + h log N ) log N )) = O(n3h) time an
O(n log N )-size representation of all g-gapped palindromes for Case (1) in s.
In a symmetric way of Case (1), we can find all g-gapped palindromes for
Case (2). Putting all together, we get the following theorem.
Theorem 2. Given an SLP of size n and height h that describes string s of
length N , and non-negative integer g, an O(n log N + ng)-size compact represen-
tation of all g-gapped palindromes in s can be computed in O(n3h + gnh log N )
time and O(n2) working space.
5 Discussions
Let R and G denote the output compact representations of the runs and g-gapped
palindromes of a given SLP S, respectively, and let R and G denote their size.
Here we show an application of R and G; given any interval [b, e] in s, we can
count the number of runs and gapped palindromes in s[b..e] in O(n + R) and
O(n+G) time, respectively. We will describe only the case of runs, but a similar
technique can be applied to gapped palindromes. As is described in Section 3.2,
s[b..e] can be represented by a sequence U = (Xu(1), Xu(2), . . . , Xu(m)) of O(h)
variables of S. Let T be the SLP obtained by concatenating the variables of U .
There are three different types of runs in R: (1) runs that are completely within
the subtree rooted at one of the nodes of U ; (2) runs that begin and end inside
[b, e] and cross or touch any border between consecutive nodes of U ; (3) runs
that begin and/or end outside [b, e]. Observe that the runs of types (2) and (3)
cross or touch the boundary of one of the nodes in the path from the root to
the b-th leaf of the derivation tree TS, or in the path from the root to the e-th
leaf of TS. A run that begins outside [b, e] is counted only if the suffix of the
run that intersects [b, e] has an exponent of at least 2. The symmetric variant
applies to a run that ends outside [b, e]. Thus, the number of runs of types (2)
and (3) can be counted in O(n + 2R) time. Since we can compute in a total of
O(n) time the number of nodes in the derivation tree of T that are labeled by
Xi for all variables Xi, the number of runs of type (1) for all variables Xu(j) can
be counted in O(n + R) time. Noticing that runs are compact representation
11
of squares, we can also count the number of occurrences of all squares in s[b..e]
in O(n + R) time by simple arithmetic operations.
The approximate doubling and LCE algorithms of Section 3 can be used
as basis of other efficient algorithms on SLPs. For example, using approximate
doubling, we can reduce the number of pairs of variables for which the AP-table
has to be computed in the algorithms of Lifshits [10], which compute compact
representations of all periods and covers of a string given as an SLP. As a result,
we improve the time complexities from O(n2h log N ) to O(n2h) for periods, and
from O(n2h log2 N ) to O(nh(n + log2 N )) for covers.
References
1. Apostolico, A., Breslauer, D.: An optimal O(log log N )-time parallel algorithm for
detecting all squares in a string. SIAM Journal on Computing 25(6), 1318 -- 1331
(1996)
2. Apostolico, A., Breslauer, D., Galil, Z.: Parallel detection of all palindromes in a
string. Theor. Comput. Sci. 141(1&2), 163 -- 173 (1995)
3. Bannai, H., Gagie, T., I, T., Inenaga, S., Landau, G.M., Lewenstein, M.: An efficient
algorithm to test square-freeness of strings compressed by straight-line programs.
Inf. Process. Lett. 112(19), 711 -- 714 (2012)
4. Bille, P., Landau, G.M., Raman, R., Sadakane, K., Satti, S.R., Weimann, O.: Ran-
dom access to grammar-compressed strings. In: Proc. SODA 2011. pp. 373 -- 389
(2011)
5. Crochemore, M., Ilie, L., Rytter, W.: Repetitions in strings: Algorithms and com-
binatorics. Theor. Comput. Sci. 410(50), 5227 -- 5235 (2009)
6. Crochemore, M., Rytter, W.: Efficient parallel algorithms to test square-freeness
and factorize strings. Information Processing Letters 38(2), 57 -- 60 (1991)
7. Jansson, J., Peng, Z.: Online and dynamic recognition of squarefree strings. Inter-
national Journal of Foundations of Computer Science 18(2), 401 -- 414 (2007)
8. Khvorost, L.: Computing all squares in compressed texts. In: Proceedings of the
2nd Russian Finnish Symposium on Discrete Mathemtics. vol. 17, pp. 116 -- 122
(2012)
9. Kolpakov, R.M., Kucherov, G.: Finding maximal repetitions in a word in linear
time. In: FOCS. pp. 596 -- 604 (1999)
10. Lifshits, Y.: Processing compressed texts: A tractability border. In: Proc. CPM
2007. LNCS, vol. 4580, pp. 228 -- 240 (2007)
11. Main, M.G.: Detecting leftmost maximal periodicities. Discrete Applied Mathe-
matics 25(1-2), 145 -- 153 (1989)
12. Main, M.G., Lorentz, R.J.: An O(n log n) algorithm for finding all repetitions in a
string. Journal of Algorithms 5(3), 422 -- 432 (1984)
13. Manacher, G.K.: A new linear-time "on-line" algorithm for finding the smallest
initial palindrome of a string. J. ACM 22(3), 346 -- 351 (1975)
14. Matsubara, W., Inenaga, S., Ishino, A., Shinohara, A., Nakamura, T., Hashimoto,
K.: Efficient algorithms to compute compressed longest common substrings and
compressed palindromes. Theoretical Computer Science 410(8 -- 10), 900 -- 913 (2009)
15. Miyazaki, M., Shinohara, A., Takeda, M.: An improved pattern matching algorithm
for strings in terms of straight-line programs. In: Proceedings of the 8th Annual
Symposium on Combinatorial Pattern Matching. pp. 1 -- 11 (1997)
12
Appendix A: Details of the algorithm to find runs
In this section, we describe how we process occurrences of p′ at Step 2 of the basic
algorithm. To handle occurrences of p′ that are represented by an arithmetic
progression, we make use of its periodicity.
For any string s and positive integer c ≤ s, let −→rep c(s) (resp. ←−rep c(s)) denote
the length of the longest prefix (resp. suffix) of s having period c.
Lemma 9. Let s, p ∈ Σ+ and {a0, a1, . . . , ak} be consecutive occurrences of p in
s that form a single arithmetic progression with common difference c ≤ p. Let
zj = s[aj + p..s] and z′
j = s[1..aj − 1] for any 0 ≤ j ≤ k. For any non-empty
strings x, x′ ∈ Σ+, it holds that
−→
β }
−→
β + lcp(z0[
lcp(zj, x) =(min{−→α − cj,
j, x′) =(min{←−α + cj,
lcs(z′
−→
if −→α − cj 6=
β ,
otherwise, and
if ←−α + cj 6=
otherwise,
←−
β ,
−→
β + 1..x])
−→
β + 1..z0], x[
←−
β }
0[1..z′
−→
β = −→repc(px) − p, ←−α = ←−repc(z′
←−
β ], x′[1..x′ −
←−
β ])
0 −
←−
β + lcs(z′
where −→α = −→rep c(pz0) − p,
←−rep c(x′p).
Proof. Since −→rep c(pzj) = −→α − cj + p, both pzj and px have a prefix of length
−→
min{−→α − cj,
β , either pzj or
−→
px has a prefix of length min{−→α − cj,
β } + p + 1 with period c while the other
−→
does not, and hence lcp(zj, x) = lcp(pzj, px) − p = min{−→α − cj,
β }. Only when
−→
the period breaks the periodicity, i.e., −→α −cj =
β , lcp(zj, x) could expand. Note
that such expansion occurs at most once. Similarly, since ←−rep c(z′
jp) = ←−α + cj we
get the statement for lcs(z′
⊓⊔
−→
β } + p with period c (see also Fig. 3). If −→α − cj 6=
0p) − p and
←−
β =
j, x′).
a0
a1
a2
a3
a4
s a b b c a b c a b c a b c a b c a b c a b c a b c a c b a a b
x
c a b c a b c a c b a a c
p0
c
p1
α
β
β
α
c
p2
p3
p4
x'
a c c a b c a b c a b c
s a b b c a b c a b c a b c a b c a b c a b c a b c a c b a a b
Fig. 3. Illustration for Lemma 9.
In the next lemma, we show how to handle one of the arithmetic progressions
computed in Step 2 of Case (3).
13
Lemma 10. Let Xi → XℓXr be a production of an SLP of size n and p be the
suffix of val (Xℓ) of length 2t−1. Let {a0, a1, . . . , ak} be consecutive occurrences
of p′ in val (Xi) which form a single arithmetic progression, which are computed
in Step 2 of Case (3). We can detect all runs corresponding to the occurrences
of p′ by using LCE constant times. Also, such runs are represented in constant
space.
Proof. We apply Lemma 9 by letting s = val (Xi), x = val (Xr) and x′ =
val (Xℓ)[1..val (Xℓ) − p]. First we compute −→α = lcp(pz0, p[c + 1..p]z0) + c − p,
−→
β = lcp(px, p[c + 1..p]x) + c − p, ←−α = lcs(z′
0p[1..p − c]) + c − p and
←−
β = lcs(x′p, x′p[1..p − c]) + c − p by using lcp and lcs four times.
0p, z′
−→
β + ←−α ≥ a0 − 1 + c, the root of any repetition detected from aj is not
Claim. If
primitive.
−→
β + ←−α ≥ a0 − 1 + c, pyp must have period c, where y is the
Proof of Claim. If
prefix of length a1 − 1 of x. Since pyp[c + 1..c + p] = p, yp − c is a period of
yp. It follows from the periodicity lemma that py, as well as every aj + p − 1,
is divisible by greatest common divisor of c and yp − c, and hence the root of
⊓⊔
any repetition detected from aj is not primitive.
−→
β + ←−α < a0 − 1 + c.
Let dj = aj − 1 + p = a0 − 1 + p + cj, and then we want to check if lcp(zj, x) +
j, x′) − cj ≥
lcs(z′
a0 − 1.
j, x′) ≥ dj − p = a0 − 1 + cj, or equivalently, lcp(zj, x) + lcs(z′
From the above claim, in what follows we assume that
Let j′ = min{j ≥ 0 −→α − cj ≤
−→
β } and j′′ = min{j ≥ 0 ←−α + cj ≥
any 0 ≤ j < min{j′, j′′}, it follows from lcp(zj, x) =
j, x′) − cj =
that lcp(zj, x) + lcs(z′
−→
β +←−α ≥ a0−1, where δ1 = x′+1−←−α , δ2 = a0+p+
cj, δ3+cji appears iff
and δ3 = a0 + p − 1 are constants.
←−
β }. For
j, x′) = ←−α + cj
−→
β + ←−α , and hence a repetition hδ1 − cj, δ2 +
−→
β −1
−→
β and lcs(z′
We show that the root of such repetition hδ1 − cj, δ2 + cj, δ3 + cji is primitive.
Assume on the contrary that it is not primitive, namely, s′ = s[δ1 − cj..δ2 + cj] =
uq with u ≤ (δ3 + cj)/2 and q ≥ 4. Evidently, −→rep c(s′) =
j, x′) + p =
−→
−→
β + ←−α + p + cj. It follows from a0 − 1 ≤
β + ←−α < a0 − 1 + c that δ3 + cj ≤
−→rep c(s′) < δ3 +cj +c < s′. Since 2u ≤ −→repc(s′) and c ≤ p ≤ (δ3 +cj)/2 ≤ δ3 +
cj − u ≤ −→rep c(s′)− u, −→repc(s′[1..s′ − u) = −→repc(s′[u + 1..s′])+ u, however
both s′[1..s′−u] and s′[u+1..s′] are uq−1, a contradiction. Therefore, for all
0 ≤ j < min{j′, j′′}, hδ1 − cj, δ2 + cj, δ3 + cji are runs, and they can be encoded
by a quintuplet hδ1, δ2, δ3, c, min{j′, j′′}i.
−→
β + lcs(z′
For any min{j′, j′′} ≤ j ≤ k except for j = j′ or j′′, lcp(zj, x)+lcs(z′
j, x′)−cj
is monotonically decreasing by at least c and satisfies lcp(zj, x)+lcs(z′
j, x′)−cj <
−→
β + ←−α − c < a0 − 1, and hence, no repetition appears. For j′ and j′′, we can
check whether these two occurrences become runs or not by using LCE constant
⊓⊔
times.
14
Xi
Xℓ
Xr
t-1
2
p
p
p
p
p
β
p0
c
p1
α
p4
p2
p3
β
α
c
Illustration for Lemma 10. Four runs are found. Here j ′ = 3 and j ′′ = 2.
Fig. 4.
The runs from p0 and p1 are encoded by a quintuplet. For each j ′ and j ′′, the run is
separately encoded by a quintuplet that shows a single run.
The other cases can be processed in a similar way.
A minor technicality is that we may redundantly find the same run in different
cases. However, we can avoid duplicates by simply looking into the currently
computed runs when we add new runs, spending O(log N ) time. Also, we can
remove repetitions whose root are not primitive by just choosing the smallest
period among the repetitions with the same interval.
15
Appendix B: Proof of Lemma 7
Proof. The outline of our algorithm to compute FirstMismatch follows [15] which
used a slower algorithm for Match. Assume Xi − k + 1 ≤ Xj holds.
If Xj → a with a ∈ Σ, then
FirstMismatch(Xi, Xj, k) =(1
0
if Match(Xi, Xj, k) = true,
if Match(Xi, Xj, k) = false.
If Xj → Xℓ(j)Xr(j), then we can recursively compute FirstMismatch(Xi, Xj, k)
as follows:
FirstMismatch(Xi, Xj, k)
=(FirstMismatch(Xi, Xr(j), k + Xℓ)
FirstMismatch(Xi, Xℓ(j), k)
if Match(Xi, Xℓ(j), k) = true,
if Match(Xi, Xℓ(j), k) = false.
(1)
We apply Lemma 6 to S, pre-processing SLP S in O(n2h) time and O(n2)
space, so that query Match(Xi, Xj ′ , k′) is answered in O(log N ) time for any vari-
able Xj ′ and integer k′. Note that in either case of Equation 1, the height of the
second variable decreases by 1. Hence we can compute FirstMismatch(Xi, Xj, k)
in O(h log N ) time, after the O(n2h)-time O(n2)-space pre-processing.
⊓⊔
16
Appendix C: Figures
X8
X7
X7
X5
X6
X5
X6
X1
X3
X4
X3
X1
X3
X4
X3
X2 X2
X1 X2
X2 X2
X2 X2
X1 X2
X2 X2
a
1
b b
2 3
a
4
bb
65
b a b b a b b b
7 8 9 10 11 12 13 14
Fig. 5. The derivation tree of SLP S = {X1 → a, X2 → b, X3 → X2X2, X4 →
X1X2, X5 → X1X3, X6 → X4X3, X7 → X5X6, X8 → X7X7 }, representing string
s = abbabbbabbabbb.
k1
k2
Xi
Pl
Z
Xi
a
b
Pr
Fig. 6. Lemma 5: Illustration for computing LCE(Xi, k1, k2). The roots of the gray
subtrees are labeled by the variables in U . We find the first variable Xu(p′) in the list
U with which the Match query returns false. We then perform the FirstMismatch query
for Xi and Xu(p′) using the appropriate offset.
17
v2
v1
v3
v4
v5
2Ya1 2Ya2
2Ya3
2Ya4
2Ya5
Ya1
Ya2
Ya3
Ya4
Ya5
Fig. 7. Lemma 8: Illustration for approximate doubling. The prefix variables up to
Ya5 have been created. The traversals for v2, v3, v4 end due to Condition 1 and that
for v5 ends due to Condition 2. Each traversed edge (depicted in bold) contributes to
at most one new variable for some segment. Next, we will resume the traversal from v5
targeting position 2Ya5 , and iterate the procedure until we get the last variable Yak .
The total number of bold edges can be bounded by O(n) thanks to Condition 2.
18
|
1908.06320 | 2 | 1908 | 2019-10-28T16:23:28 | Revisiting the Graph Isomorphism Problem with Semidefinite Programming | [
"cs.DS",
"cs.CC"
] | It is well-known that the graph isomorphism problem can be posed as an equivalent problem of determining whether an auxiliary graph structure contains a clique of specific order. However, the algorithms that have been developed so far for this problem are either not efficient or not exact. In this paper, we present a new algorithm which solves this equivalent formulation via semidefinite programming. Specifically, we show that the problem of determining whether the auxiliary graph contains a clique of specific order can be formulated as a semidefinite programming problem, and can thus be (almost exactly) solved in polynomial time. Furthermore, we show that we can determine if the graph contains such a clique by rounding the optimal solution to the nearest integer. Our algorithm provides a significant complexity result in graph isomorphism testing, and also represents the first use of semidefinite programming for solving this problem. | cs.DS | cs |
Revisiting the Graph Isomorphism Problem with
Semidefinite Programming
Giannis Nikolentzos1 and Michalis Vazirgiannis1,2
2Athens University of Economics and Business, Greece
1 ´Ecole Polytechnique, France
{nikolentzos,mvazirg}@aueb.gr
preliminary version∗
October 28, 2019
Abstract
It is well-known that the graph isomorphism problem can be posed as an equivalent
problem of determining whether an auxiliary graph structure contains a clique of specific
order. However, the algorithms that have been developed so far for this problem are either
not efficient or not exact. In this paper, we present a new algorithm which solves this equiv-
alent formulation via semidefinite programming. Specifically, we show that the problem of
determining whether the auxiliary graph contains a clique of specific order can be formu-
lated as a semidefinite programming problem, and can thus be (almost exactly) solved in
polynomial time. Furthermore, we show that we can determine if the graph contains such
a clique by rounding the optimal solution to the nearest integer. Our algorithm provides a
significant complexity result in graph isomorphism testing, and also represents the first use
of semidefinite programming for solving this problem.
1 Introduction
Graph isomorphism is one of those few fundamental problems in NP whose computational status
still remains unknown [1]. Roughly speaking, the graph isomorphism problem asks whether two
graphs are structurally identical or not. The problem is clearly in NP. However, it has been
neither proven NP-complete nor found to be solved by a polynomial time algorithm. In fact,
there is strong evidence that graph isomorphism is not NP-complete since it has been shown
that the problem is located in the low hierarchy of NP [2]. This implies that if the problem was
NP-complete, then the polynomial time hierarchy would collapse to its second level.
Over the years, algorithms of different nature have been developed to attack the problem.
Traditionally, those that draw ideas from group theory turn out to be the most promising.
One of these group-theoretic algorithms was proposed by Babai and Luks in 1983 [3]. The
algorithm combines a preprocessing procedure proposed by Zemlyachenko et al.
[4] with an
efficient algorithm for solving graph isomorphism on graphs of bounded degree [5]. Its com-
√
putational complexity is 2O(
nlogn) where n denotes the number of vertices. Despite decades
of active research, no progress had been achieved, and this was the best known algorithm un-
til recently when Babai presented an algorithm that solves the graph isomorphism problem in
∗Note: this article has not been peer reviewed yet.
1
quasi-polynomial time [6]. It should be mentioned that while the complexity status of the graph
isomorphism for general graphs remains a mystery, for many restricted graph classes, polyno-
mial time algorithms are known. This is, for example, the case for planar graphs [7], graphs
of bounded degree [5], or graphs with bounded eigenvalue multiplicity [8]. It should also be
noted that there exist several algorithms which have proven very efficient for graphs of practical
interest [9, 10, 11]. Interestingly, these algorithms are very different from the ones that offer
the lowest worst case complexities. This indicates that there is a wide gap between theory and
practice.
Besides the above algorithms, there are also several scalable heuristics for graph isomorphism
which are based on continuous optimization. In these heuristics, the discrete search problem
in the space of permutation matrices is replaced by an optimization problem with continuous
variables, enabling the use of efficient continuous optimization algorithms. Formally, for any
two graphs on n vertices with respective n× n adjacency matrices A1 and A2, the optimization
problem consists in minimizing the function A1 − P A2P (cid:62)F over all P ∈ Π, where Π denotes
the set of n× n permutation matrices, and ·F is the Froebenius matrix norm [12]. Therefore,
the problem of graph isomorphism can be reformulated as the problem of minimizing the above
function over the set of permutation matrices. The two graphs are isomorphic to each other if
there exists a permutation matrix P for which the above function is equal to 0. Note also that
other objectives have also been proposed in the literature, this being perhaps the most common.
This problem has a combinatorial nature and there is no known polynomial algorithm to solve
it. Numerous approximate methods have been developed. Most of these methods replace the
space of permutations by the space of doubly-stochastic matrices. Let D denote the set of n× n
doubly stochastic matrices, i. e., nonnegative matrices with row and column sums each equal to
F over all D ∈ D. There
1. The convex relaxed problem minimizes the function A − DAD(cid:62)2
is a polynomial-time algorithm for exactly solving the convex relaxed graph matching problem
[13]. However, due to relaxation, even if there exists a doubly stochastic matrix D for which
the objective function is equal to 0, there is no guarantee that the two graphs are isomorphic
to each other.
The main contribution of this work is a novel algorithm which attacks efficiently the problem
of graph isomorphism. The main tools employed are a compatibility graph, i. e., an auxiliary
graph structure that is useful for solving general graph and subgraph isomorphism problems,
and a semidefinite programming formulation. Given two graphs of order n, we build their com-
patibility graph of order n2. We show that testing the two graphs for isomorphism is equivalent
to determining whether the compatibility graph contains a clique of order n. We show that
this problem can be formulated as a semidefinite programming optimization problem, and can
thus be (almost exactly) solved in polynomial time with readily available solvers. We show that
the two graphs are isomorphic to each other if the optimal value of the semidefinite program
is arbitrarily close to n(n − 1). Our algorithm demonstrates the usefulness of semidefinite pro-
gramming in combinatorial optimization, and provides a significant complexity result in graph
isomorphism testing. It should be mentioned that our work is not the first to apply continuous
optimization approaches to the problem of determining whether the compatibility graph con-
tains a clique of order n. In a previous study, Pelillo developed a heuristic for computing the
clique number of the compatibility graph [14]. However, in contrast to the proposed algorithm,
this method provides no guarantees, and may get stuck on some local optimum of the objective
function.
2
2 Preliminaries
In this Section, we first define our notation, and we then introduce the concept of a compatibility
graph. We show that the graph isomorphism problem is equivalent to finding if the compatibility
graph contains a clique of specific size. We also present the basic concepts of semidefinite
programming and an algorithm which is based on Lov´asz ϑ number and for almost all classes
of graphs can decide if two instances are isomorphic to each other.
2.1 Graph Theory Notation and Terminology
Let G = (V, E) be an undirected and unweighted graph consisting of a set V of vertices and
a set E of edges between them. We will denote by n the number of vertices. The adjacency
matrix of G is a symmetric matrix A ∈ Rn×n defined as follows: Ai,j = 1 if (i, j) ∈ E, and 0
otherwise. Note that since the graph is undirected, (i, j) ∈ E if and only if (j, i) ∈ E, for all
∼= G2), if
i, j ∈ V . Two graphs G1 = (V1, E1) and G2 = (V2, E2) are isomorphic (denoted by G1
there is a bijective mapping φ : V1 → V2 such that (vi, vj) ∈ E1 if and only if (φ(vi), φ(vj)) ∈ E2.
We next present the notion of a compatibility graph, i. e., a structure that is very useful
for solving graph/subgraph isomorphism and maximum common subgraph problems. These
auxiliary graph structures have been proposed independently by several authors [15, 16, 17],
while they also lie at the core of several algorithms [14, 18, 19]. Furthermore, different authors
have used different names to describe them. For instance, compatibility graphs, association
graphs, derived graphs, M-graphs, and product graphs are all names that have been coined to
describe these structures. In what follows, we will use the name compatibility graph to refer
to them. Formally, given two graphs G1 = (V1, E1) and G2 = (V2, E2), their compatibility
graph Gc = (Vc, Ec) is a graph with vertex set Vc = V1 × V2. An edge is drawn between two
vertices (v1, u1), (v2, u2) ∈ Vc if and only if v1 (cid:54)= v2, u1 (cid:54)= u2 and either e1 = (v1, v2) ∈ E1
and e2 = (u1, u2) ∈ E2 or e1 (cid:54)∈ E1 and e2 (cid:54)∈ E2. Clearly, there are two types of edges in a
compatibility graph: (1) edges that represent common adjacency, and (2) edges that represent
common non-adjacency. An example of the compatibility graph that emerges from two P3
graphs is illustrated in Figure 1.
Levi established a relation between isomorphic subgraphs of two graphs and cliques in their
compatibility graph [15]. Specifically, if some vertices (v1, u1), (v2, u2), . . . , (vk, uk) ∈ Vc form
a clique, and are thus pairwise adjacent, then the subgraph in G1 induced by the vertices
v1, v2, . . . , vk is isomorphic to the subgraph in G2 induced by the vertices u1, u2, . . . , uk. The
isomorphism is given by the vertices (v1, u1), (v2, u2), . . . , (vk, uk) ∈ Vc of the compatibility
graph that form the clique, i. e., φ(u1) = v1, φ(u2) = v2, . . . , φ(uk) = vk.
The next Theorem establishes an equivalence between the graph isomorphism problem and
the maximum clique problem on the compatibility graph, and corresponds to a sub-instance of
Levi's result. For completeness, we also provide the proof. Note that if G1 and G2 are two
graphs of order n, the maximum clique of their compatibility graph will consist of at most n
vertices.
Theorem 1. Let G1 and G2 be two graphs of order n, and let Gc be their compatibility graph.
Then, G1 and G2 are isomorphic if and only if Gc contains a clique of order n, i. e., ω(Gc) = n.
Proof. If G1 = (V1, E1) and G2 = (V2, E2) are isomorphic, then there exists a bijective mapping
φ : V1 → V2 such that (vi, vj) ∈ E1 if and only if (φ(vi), φ(vj)) ∈ E2. Then, by construction,
there are n vertices (v1, φ(v1)), . . . , (vn, φ(vn)) ∈ Vc which are connected to each other by an
edge, i. e.,(cid:0)(v1, φ(v1)), (v2, φ(v2))(cid:1), . . .,(cid:0)(vn−1, φ(vn−1)), (vn, φ(vn))(cid:1) ∈ Ec. These vertices form
a clique of order n, and therefore, ω(Gc) = n. For the second part, given a compatibility graph
Gc that contains a clique of order n, let (v1, u1), . . . , (vn, un) ∈ Vc denote the n vertices that
3
Figure 1: Two graphs (top left and right) and their compatibility graph (bottom).
a bijective mapping φ : V1 → V2 as follows: φ(vi) = ui. Since(cid:0)(vi, φ(vi)), (vj, φ(vj))(cid:1) ∈ Ec for
form the clique. By definition, V1 = {v1, . . . , vn} and V2 = {u1, . . . , un}. Then, we can construct
i ∈ {0, . . . , n}, we have that (vi, vj) ∈ E1 if and only if (φ(vi), φ(vj)) ∈ E2. Therefore, G1 and
G2 are isomorphic to each other.
Note that the number of isomorphisms between two graphs can be exponential to the number
of vertices of the graphs n. Specifically, if G1 and G2 are isomorphic, then the number of
automorphisms of G1 (or of G2) is equal to the number of isomorphisms from G1 to G2. Hence,
if, for instance, G1 and G2 are complete graphs on n vertices, i. e., both correspond to the
complete graph Kn, then the number of isomorphisms between the two graphs is equal to n!
since Aut(G1) = Aut(G2) = Aut(Kn) = n!. Each isomorphism φ : V → V (cid:48) corresponds to
an n-clique in the compatibility graph. Therefore, the compatibility graph can contain up to
n! cliques. As an example, consider the graphs shown in Figure 1. There are two isomorphisms
between G1 and G2. Hence, their compatibility graph contains exactly two cliques of order 3.
We now introduce the following Lemma which we will use in the next Section.
Lemma 1. Let G1 and G2 be two graphs of order n, and let Gc be their compatibility graph.
Then, the vertices of Gc can be grouped into n partitions such that there are no edges between
vertices that belong to the same partition.
compatibility graph Gc, and (a, b), (c, d) ∈ Vc. Then, by definition, if a = c,(cid:0)(a, b), (c, d)(cid:1) (cid:54)∈ Ec.
Proof. Let V1 and V2 denote the sets of vertices of G1 and G2, respectively. Then, V1 =
{v1, . . . , vn} and V2 = {u1, . . . , un}. Let also Vc and Ec denote the set of vertices and edges of the
The set of nodes Vc can be decomposed into the following n disjoint sets: P1 = {(v1, u1), (v1, u2),
. . . , (v1, un)}, P2 = {(v2, u1), (v2, u2), . . . , (v2, un)}, . . ., Pn = {(vn, u1), (vn, u2), . . . , (vn, un)}.
4
123(1,a)(1,b)(1,c)GcG1G2abc(2,a)(2,b)(2,c)(3,a)(3,b)(3,c) Then, Vc = P1 ∪ P2 ∪ . . . ∪ Pn. Clearly, there is no edge between each pair of vertices of each
partition. This concludes the proof.
2.2 Semidefinite Programming
A semidefinite program (SDP) is the problem of optimizing a linear function over the intersection
of the cone of positive semidefinite matrices with an affine space. Semidefinite programming
has attracted a lot of attention in recent years since many practical problems in operations
research and combinatorial optimization can be modeled or approximated as semidefinite pro-
gramming problems. For instance, in control theory, SDPs are used in the context of linear
matrix inequalities. In graph theory, the problem of computing the Lov´asz number of a graph
can be formulated as a semidefinite program. Given any > 0, semidefinite programs can be
solved within an additive error of in polynomial time. There are several different algorithms
for solving SDPs. For instance, this can be done through the ellipsoid algorithm [20] or through
interior-point methods [21].
2.3 An Algorithm for Almost all Classes of Graphs
It follows from Lemma 1 that the set of vertices Vc of a compatibility graph can be decomposed
into n disjoint sets P1, P2, . . . , Pn such that there is no edge that connects vertices of the same
set. Since a compatibility graph can be decomposed into n disjoint sets such that every edge
e ∈ Ec connects a vertex in Pi to one in Pj with i (cid:54)= j, the minimum number of colors required
for a proper coloring of Gc is no more than n. Therefore, χ(Gc) ≤ n.
For an arbitrary graph G, it is well-known that ω(G) ≤ χ(G). Therefore, a compatibility
graph can contain a clique of order n only if χ(Gc) = n. Computing the chromatic number of
a graph is in general an NP-complete problem. However, we can compute in polynomial time a
real number ϑ( ¯G) that is "sandwiched" between the clique number and the chromatic number
of a graph G, that is ω(G) ≤ ϑ( ¯G) ≤ χ(G) [22]. This number is known as the Lov´asz number
of G. Again, a compatibility graph can contain a clique of order n only if ϑ( ¯Gc) = χ(Gc) = n.
However, the fact that ϑ( ¯Gc) = χ(Gc) = n does not imply that ω(G) = n. Instead, it may hold
that ω(Gc) < ϑ( ¯Gc) = χ(Gc) = n. One class of graphs for which the above holds is the family of
latin square graphs. Although for the class of compatibility graphs, it may hold that ω(G) = n
whenever ϑ( ¯G) = χ(G) = n, we do not study this any further, but we leave it as future work.
3 The Main Result
Next, we give a SDP based solution for the graph isomorphism problem. We show that the
problem of identifying whether a compatibility graph contains a clique of order n can be ex-
pressed as a SDP. For the ease of presentation, we start with a simple formulation where we
assume that the rank of the matrix involved in the objective function of the SDP is 1. Unfor-
tunately, this constraint is not convex, and renders the problem NP-hard. We then drop the
rank constraint, and show that the emerging SDP can successfully deal with the general case.
3.1 Rank-1 Case
Define a variable xi for every vertex i ∈ Vc, and let X = xx(cid:62). Let also m = n2. Then, we have
that:
X = xx(cid:62) ⇔ X ∈ Sm, X (cid:23) 0, rank(X) = 1
5
where Sm is the set of all m × m real symmetric matrices and X (cid:23) 0 means that the matrix
variable X is positive semidefinite. Let also J denote the m × m matrix of ones. Consider now
the following optimization problem:
maximize
X
subject to
trace(JX)
trace(X) = n,
Xi,j = 0, (i, j) (cid:54)∈ Ec,
Xi,j ≤ 1, (i, j) ∈ Ec,
X (cid:23) 0,
rank(X) = 1.
(1a)
(1b)
(1c)
(1d)
(1e)
(1f)
We can determine if there is a clique of order n in Gc by solving the above optimization problem.
Specifically, if there exists such a clique in Gc, the value of the optimal solution to the problem
is equal to n2.
Lemma 2. The value of problem (1) is no greater than n2.
Proof. The value of the objective function of the optimization problem is:
trace(JX) =
Ji,jXi,j = J1,1 X1,1 + . . . + Jm,m Xm,m
i=1
j=1
= x1x1 + . . . + xmxm
From constraint (1b), it follows that the sum of the diagonal terms of matrix X is equal to n,
i. e., x1x1 + x2x2 + . . . + xmxm = n. We next replace these terms with their sum. Therefore,
the objective function of the optimization problem becomes:
m(cid:88)
m(cid:88)
m(cid:88)
m(cid:88)
i=1
j=1
trace(JX) =
Ji,jXi,j = n + x1x2 + . . . + xmxm−1
We will next show that in our setting, we can have at most n vertices for which xi (cid:54)= 0. It
follows from the constraint (1c) that ∀(i, j) (cid:54)∈ Ec, Xi,j = xixj = 0. Therefore, ∀(i, j) (cid:54)∈ Ec, one
of the following three conditions holds: (1) xi = 0, xj (cid:54)= 0, (2) xi (cid:54)= 0, xj = 0, (3) xi = 0,
xj = 0. It follows also from Lemma 1 that the vertices of Gc can be grouped into n partitions
such that the n vertices that belong to each partition are not connected by an edge. Therefore,
there can be at most one vertex i from each partition for which xi (cid:54)= 0. It thus turns out that
we can have at most n vertices for which xi (cid:54)= 0. From these n vertices, we can obtain n(n − 1)
(ordered) pairs of vertices. For each pair, Xi,j = xixj ≤ 1 holds. Therefore, it follows that:
m(cid:88)
m(cid:88)
trace(JX) =
Ji,jXi,j ≤ n + n(n − 1) = n2
This proves the Lemma.
i=1
j=1
Theorem 2. If Gc contains a clique of order n, problem (1) can attain its largest possible value.
Proof. Let us assume that Gc contains a clique of order n, and let S be the set of vertices that
form the n-clique. We set xi = 1 for i ∈ S, and xj = 0 for j ∈ Vc \ S. Clearly, since Ji,j = 1 for
6
each i, j ∈ S, there are n(n − 1) (ordered) pairs of vertices, and for each of those pairs, we have
that xixj = 1. Therefore, the objective value of problem (1) is equal to:
trace(JX) = n + n(n − 1) = n2
By Lemma 2, the objective value is equal to the largest possible value of problem (1). Further-
more, all the constraints hold. Specifically, trace(X) = S = n, and Xi,j (cid:54)= 0 only if i, j ∈ S
(with i (cid:54)= j). Moreover, since matrix X corresponds to the outer product of a vector and itself,
it is symmetric, positive semidefinite and of rank 1.
Theorem 3. If Gc contains no clique of order n, the optimal solution of problem (1) takes
some value no greater than n + (n − 1)(n − 2).
Proof. As described above, Lemma 1 implies that we can have at most n vertices for which
x (cid:54)= 0. Furthermore, since Gc does not contain any n-clique, for each subset of vertices S ⊂ Vc
with S = n, there is at least one pair of vertices i, j for which (i, j) (cid:54)∈ Ec. Then, to satisfy the
condition that Xi,j = 0 if (i, j) (cid:54)∈ Ec, either xi = 0 or xj = 0. Hence, there can be at most n− 1
vertices for which x (cid:54)= 0. Then, the objective value of the problem is equal to:
trace(JX) = x1x1 + . . . + xmxm
= n + x1x2 + . . . + xmxm−1
≤ n + (n − 1)(n − 2)
< n2
The first inequality follows from the fact that xixj ≤ 1 for any i, j with i (cid:54)= j (from con-
straint (1d)), and as mentioned above there are at most n − 1 vertices for which x (cid:54)= 0.
3.2 General Case
Rank constraints in semidefinite programs are usually hard. Specifically, it turns out that
problem (1) is a nonconvex problem in X ∈ Sm. If we simply drop the rank constraint, we
obtain the following relaxation:
trace(JX)
maximize
X
subject to
trace(X) = n,
Xi,j = 0, (i, j) (cid:54)∈ Ec,
Xi,j ≤ 1, (i, j) ∈ Ec,
X (cid:23) 0.
(2a)
(2b)
(2c)
(2d)
(2e)
which is a semidefinite program in X ∈ Sm. Clearly, the optimal value of problem (2) is an
upper bound of the optimal value of problem (1).
We will next show that in case Gc contains one or more cliques of order n, then the optimal
value of problem (2) is equal to the optimal value of problem (1) (i. e., equal to n2). Furthermore,
we will show that if Gc contains no cliques, then the optimal value of the problem is not greater
than n(n − 1).
Let vi ∈ Rd be the vector representation of vertex i ∈ Vc, and U = [v1, . . . , vm](cid:62) ∈ Rm×d a
matrix whose ith row contains the vector representation of vertex i ∈ Vc. Then, X = U U(cid:62).
Lemma 3. The value of problem (2) is no greater than n2.
7
m(cid:88)
m(cid:88)
m(cid:88)
m(cid:88)
trace(JX) =
Ji,jXi,j = J1,1 X1,1 + . . . + Jm,m Xm,m
i=1
j=1
= (cid:104)v1, v1(cid:105) + . . . + (cid:104)vm, vm(cid:105)
From constraint (2b), it follows that the sum of the diagonal terms of matrix X is equal to
n, i. e., (cid:104)v1, v1(cid:105) + (cid:104)v2, v2(cid:105) + . . . + (cid:104)vm, vm(cid:105) = n. We next replace these terms with their sum.
Therefore, the objective function of the optimization problem becomes:
trace(JX) =
Ji,jXi,j = n + (cid:104)v1, v2(cid:105) + . . . + (cid:104)vm, vm−1(cid:105)
(3)
i=1
j=1
From Lemma 1, it follows that the set of vertices Vc can be decomposed into n disjoint sets
P1, P2, . . . , Pn such that there is no edge that connects vertices of the same set. Then, from
constraint (2c), it follows that the vector representations of vertices that belong to the same
partition are pairwise orthogonal. Let v1, v2, . . . , vn denote the vector representations of the
vertices that belong to partition P1, vn+1, vn+2, . . . , v2n denote the representations of the vertices
that belong to partition P2, and so on.
Let w1, w2, . . . , wn be the sum of the representations of the vertices that belong to partitions
P1, P2, . . . , Pn, respectively. Then, w1 = v1 + v2 + . . . + vn, w2 = vn+1 + vn+2 + . . . + v2n, and
so on. Given two partitions (e. g., partitions P1 and P2), the sum of the inner products of all
pairs of vectors is equal to:
(cid:104)w1, w2(cid:105) = (cid:104)v1 + v2 + . . . + vn, vn+1 + vn+2 + . . . + v2n(cid:105)
From the Cauchy-Schwarz inequality, it is known that (cid:104)v, u(cid:105) ≤(cid:112)(cid:104)v, v(cid:105)(cid:104)u, u(cid:105). Hence, fow two
= (cid:104)v1, vn+1(cid:105) + (cid:104)v1, vn+2(cid:105) + . . . + (cid:104)vn, v2n(cid:105)
(4)
Proof. The objective value of problem (2) is equal to:
partitions (e. g., partitions P1 and P2), we have that:
(cid:104)w1, w2(cid:105) ≤(cid:112)(cid:104)w1, w1(cid:105)(cid:104)w2, w2(cid:105)
(cid:112)w12w22
=
(5)
Now, since the inner product of each pair of vectors of a partition is equal to 0, for each partition
(e. g., partition P1), we have that:
w12 = (cid:104)w1, w1(cid:105) = (cid:104)v1 + . . . + vn, v1 + . . . + vn(cid:105) = (cid:104)v1, v1(cid:105) + . . . + (cid:104)vn, vn(cid:105)
Therefore, from constraint (2b), it follows that:
w12 + . . . + wn2 = n
(6)
8
Now, from Equation (3), we have:
trace(JX) = n + (cid:104)v1, v2(cid:105) + . . . + (cid:104)vm, vm−1(cid:105)
= n + (cid:104)w1, w2(cid:105) + . . . + (cid:104)wn, wn−1(cid:105)
≤ n +
≤ n +
(cid:112)w12w22 + . . . +
+ . . . +
(cid:112)wn2wn−12
2
2
2
wn2wn−12
w12 + w22
(n − 1)w12 + (w22 + . . . + wn2)
(n − 1)w12 + (n − w12)
(n − 2)w12 + n
2
(n − 2)(w12 + . . . + wn2) + n2
(n − 2)n + n2
2
2n2 − 2n
+ . . . +
+ . . . +
2
2
2
2
= n +
= n +
= n +
= n +
= n +
= n +
= n2
(n − 1)wn2 + (w12 + . . . + wn−12)
+ . . . +
(n − 1)wn2 + (n − wn2)
2
(n − 2)wn2 + n
2
The second equality follows from Equation (4), and the first inequality from Equation (5). The
last inequality follows from the well-know inequality of arithmetic and geometric means, while
the fourth equality follows from Equation (6).
Clearly, a solution to problem (1) is also a solution to problem (2). Hence, if the compatibility
graph Gc contains a clique of order n, the rank-1 solution that was presented above is also an
optimal solution in this case. However, if Gc contains multiple cliques of order n, there is an
optimal solution of higher rank as shown below.
Theorem 4. If Gc contains d cliques of order n, we can construct an optimal solution to
problem (2) as follows: We assign a vector vi ∈ Rd to each vertex i ∈ Vc. We consider a feaure
space with a feature corresponding to each one of the d cliques. If vertex i participates in the
jth clique, then the jth component of vi is set equal to 1√
Proof. If vertices i, j ∈ Vc participate in all d cliques of Gc, then (cid:104)vi, vj(cid:105) = 1. Otherwise,
(cid:104)vi, vj(cid:105) < 1. From the above, it is clear that Xi,j ≤ 1, ∀(i, j) ∈ Ec. If a vertex i participates in
no clique, then vi is a zero vector, and (cid:104)vi, vj(cid:105) = 0, ∀j ∈ Vc. Furthermore, if two vertices i and j
participate in one or more cliques, but (i, j) (cid:54)∈ Ec, then the two vertices participate in different
cliques and they have no common components taking nonzero values, and thus (cid:104)vi, vj(cid:105) = 0.
Therefore, Xi,j = 0, ∀(i, j) (cid:54)∈ Ec. Furthermore, we have that:
.
d
trace(X) = (cid:104)v1, v1(cid:105) + . . . + (cid:104)vm, vm(cid:105) =
1vi
1 + . . . + vi
mvi
m
(cid:0)vi
d(cid:88)
i=1
(cid:1)
where vi denotes the ith component of vector v. Since each clique contains n vertices, there are
n vectors whose ith component is nonzero and equal to 1√
d
. Hence, it follows that:
trace(X) =
n
1
d
= n
d(cid:88)
i=1
9
1
1
1
1
1
1
1
1
1
JX =
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
0
0
1/2
0
1/2
0
0
0
0
1/2
0
0
1/2
1/2
0
0
0
1/2
0
0
1/2
1/2
1
0
0
0
1/2
1/2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1/2
1/2
0
0
0
1/2
0
trace(JX) = 9
0
1/2
0
1/2
0
0
0
0
1/2
=
0
0
0
0
0
0
0
0
0
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3
3
3
3
3
3
3
3
3
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
3/2
Figure 2: An optimal solution for the compatibility graph of Figure 1 (top). There are two
cliques of order 3 in this graph. Therefore, we assign a 2-dimensional vector to each vertex.
Objective value of the constructed solution (bottom). The i, jth component of matrix X corre-
sponds to the inner product of the representations of vertices i and j.
Hence, all the constraints are satisfied. The objective value of the solution is equal to:
trace(JX) =
Ji,jXi,j = J1,1 (cid:104)v1, v1(cid:105) + . . . + Jm,m (cid:104)vm, vm(cid:105)
m(cid:88)
m(cid:88)
i=1
j=1
= n +
(cid:0)vi
d(cid:88)
i=1
(cid:1)
1vi
2 + . . . + vi
mvi
m−1
Since each clique contains n vertices, there are n(n − 1) pairs of vertices i, j ∈ Vc such that
(i, j) ∈ Ec. Furthermore, since these vertices participate in a clique, their component that
corresponds to that clique is equal to 1√
, and therefore, the product of these components is
d
equal to 1
d . Since there are n(n − 1) such pairs for each clique, it follows that:
d(cid:88)
trace(JX) = n +
n(n − 1)
1
d
= n + n(n − 1) = n2
This concludes the proof.
i=1
Fugure 2 illustrates how an optimal solution is constructed for the compatibility graph of
Figure 1. There are two cliques of order 3 in this graph. Therefore, we assign a 2-dimensional
vector to each of its vertices.
10
(1,a)(1,b)(1,c)Gc(2,a)(2,b)(2,c)(3,a)(3,b)(3,c)[1√2,1√2][0,0][0,0][1√2,0][0,1√2][0,0][0,0][0,1√2][1√2,0] Theorem 5. If Gc contains no clique of order n, the optimal solution of problem (2) takes
some value no greater than n(n − 1).
Proof. From the Cauchy-Schwarz inequality, it is known that (cid:104)v, u(cid:105) ≤(cid:112)(cid:104)v, v(cid:105)(cid:104)u, u(cid:105). Further-
more, for nonzero vectors, equality holds if and only if v and u are linearly dependent (i. e.,
either v is a multiple of u or the opposite). As shown above, by applying the Cauchy-Schwarz
inequality, we obtain:
trace(JX) = n + (cid:104)v1, v2(cid:105) + . . . + (cid:104)vm, vm−1(cid:105)
= n + (cid:104)w1, w2(cid:105) + . . . + (cid:104)wn, wn−1(cid:105)
≤ n +
(cid:112)w12w22 + . . . +
(cid:112)wn2wn−12
Equality (between the last two quantities) holds if and only if the n vectors w1, w2, . . . , wn are
pairwise linearly dependent. Let us assume that these vectors are pairwise linearly dependent.
Then, without loss of generality, let us also assume that the ith component of the first vector
w1 is nonzero. Since the vectors are linearly dependent, then the ith component of all the other
vectors w2, . . . , wn is also nonzero. Since each vector is equal to the sum of orthogonal vectors
(sum of representations of vertices of each partition), for all n partitions there exists some vector
(corresponding to some vertex of that partition) whose ith component is nonzero. Then, there
exist n vectors whose ith component is nonzero, and hence the inner product of every pair of
these vectors is nonzero. That means that every pair of these vertices is connected by an edge
(from constraints (2c) and (2d)). Hence, these vertices form a clique of order n. We have
reached a contradiction since Gc does not contain a clique of order n. Therefore, the objective
function is:
(cid:112)w12w22 + . . . +
trace(JX) = n + (cid:104)v1, v2(cid:105) + . . . + (cid:104)vm, vm−1(cid:105)
= n + (cid:104)w1, w2(cid:105) + . . . + (cid:104)wn, wn−1(cid:105)
< n +
≤ n +
= n2
w12 + w22
+ . . . +
2
(cid:112)wn2wn−12
wn2wn−12
2
We will now establish an upper bound on the value of the objective function. We have that:
trace(JX) = n + (cid:104)v1, v2(cid:105) + . . . + (cid:104)vm, vm−1(cid:105)
= n + (cid:104)w1, w2(cid:105) + . . . + (cid:104)wn, wn−1(cid:105)
= n +
1wi
2 + . . . + wi
nwi
n−1
(cid:0)wi
d(cid:88)
i=1
where wi denotes the ith component of vector w. We also have that:
trace(X) = (cid:104)v1, v1(cid:105) + (cid:104)v2, v2(cid:105) + . . . + (cid:104)vm, vm(cid:105)
= (cid:104)w1, w1(cid:105) + (cid:104)w2, w2(cid:105) + . . . + (cid:104)wn, wn(cid:105)
d(cid:88)
(cid:0)wi
=
1wi
1 + wi
2wi
2 + . . . + wi
nwi
n
(cid:1)
(cid:1) = n
Since Gc contains no cliques of order n, for each component i of w1, w2, . . . , wn, at most n − 1
out of these n vectors can have a nonzero value. Hence, for each i ∈ {1, 2, . . . , d}, at least one
i=1
11
(7)
of the following n components wi
(n − 1)(n − 2) of the terms of the following summation are nonzero:
2, . . . , wi
1, wi
n is equal to zero. Therefore, for each i, at most
(cid:1)
1wi
2 + . . . + wi
nwi
n−1
(cid:0)wi
d(cid:88)
d(cid:88)
i=1
zi
(8)
trace(JX) = n +
= n +
where zi denotes the contribution of the ith component to the above summation, i. e., zi =
wi
j is equal to zero. Then, we have that:
n−1. Let us assume that for a given i, wi
2 + . . . + wi
nwi
1wi
i=1
1wi
1wi
1wi
zi = wi
= wi
≤ wi
wi
jwi
1wi
2 + . . . + wi
2 + . . . + wi
1 + wi
2wi
2
2
jwi
n + . . . + wi
1 + . . . + wi
1wi
j−1 + wi
j+1 + . . . + wi
1 + wi
wi
1wi
2
nwi
wi
j−1wi
j−1
+
nwi
j−1wi
wi
j+1wi
j+1wi
n−1
n + wi
1 + wi
1wi
2
nwi
wi
j−1wi
j−1
+ . . . +
j+1 + wi
1wi
1
+ . . . +
n + wi
2
+
j+1wi
j+1
n + wi
2
+ . . . +
1 + . . . + wi
j+1
+ . . . +
j−1 + wi
nwi
wi
j−1wi
nwi
j−1 + wi
nwi
n
j+1 + . . . + wi
nwi
n−1
+
j+1wi
= (n − 2)(wi
= (n − 2)(wi
2
1wi
1wi
1 + . . . + wi
1 + . . . + wi
j−1wi
jwi
j−1 + wi
j + . . . + wi
j+1wi
j+1 + . . . + wi
nwi
n)
nwi
n)
2
nwi
wi
n−1wi
n−1
n + wi
2
(9)
The inequality follows from the fact that for two scalars a and b, a2+b2
Furthermore, the last equality holds since we have assumed that wi
we have:
2 ≥ ab always holds.
j = 0. From Equation (8),
(cid:1)
1wi
2 + . . . + wi
nwi
n−1
trace(JX) = n +
= n +
≤ n +
(cid:0)wi
i=1
d(cid:88)
d(cid:88)
d(cid:88)
i=1
i=1
zi
(n − 2)(cid:0)wi
d(cid:88)
(cid:0)wi
i=1
= n + (n − 2)
= n(n − 1)
(cid:1)
(cid:1)
1wi
1 + . . . + wi
nwi
n
1wi
1 + . . . + wi
nwi
n
The inequality follows from Equation (9), while the last equality follows from Equation (7).
This concludes the proof.
Note that problem (2) has a strictly feasible solution, i. e., a solution that satisfies the
n for i ∈ 1, . . . , m
positive semidefiniteness requirement strictly. For instance, by setting Xi,i = 1
and Xi,j = 0 for i, j ∈ 1, . . . , m with i (cid:54)= j, we obtain a positive definite feasible solution,
i. e., X (cid:31) 0 and all constraints are satisfied. Therefore, strong duality holds [23]. There is an
algorithm that for any > 0, returns a rational closer than to the solution of problem (2) in
time bounded by a polynomial in n and log(1/) [21]. We can thus find a number closer than 1
2
to the optimal value of problem (2) in time polynomial in n. By comparing this number with
n2, we can answer if the compatibility graph contains a clique of order n or not.
12
4 Conclusion
In this paper, we have presented an algorithm for the graph isomorphism problem. The algo-
rithm capitalizes on previous results that transform the problem into an equivalent problem
of determining whether there exists a clique of specific order in an auxiliary graph structure.
We have shown that the answer to the above question can be given by solving a semidefi-
nite program. Given the frequent use of semidefinite programming in the design of algorithms
[24, 25, 26], it seemed worthwhile to investigate its effectiveness in addressing the graph isomor-
phism problem. The results of this paper constitute a first step in this direction. This paper
still leaves some open questions. Studying the dual of the proposed semidefinite program is
perhaps the most interesting of them.
Acknowledgments
The authors would like to thank Prof. Leo Liberti for discussions and comments on early
versions of this paper. Giannis Nikolentzos is supported by the project "ESIGMA" (ANR-17-
CE40-0028).
References
[1] M. R. Garey and D. S. Johnson, Computers and Intractability: A Guide to the Theory of
NP-Completeness. W. H. Freeman & Co., 1979.
[2] U. Schoning, "Graph Isomorphism Is in the Low Hierarchy," Journal of Computer and
System Sciences, vol. 37, no. 3, pp. 312 -- 323, 1988.
[3] L. Babai and E. M. Luks, "Canonical Labeling of Graphs," in Proceedings of the 15th
Annual Symposium on Theory of Computing, pp. 171 -- 183, 1983.
[4] V. N. Zemlyachenko, N. M. Korneenko, and R. I. Tyshkevich, "Graph Isomorphism Prob-
lem," Journal of Soviet Mathematics, vol. 29, no. 4, pp. 1426 -- 1481, 1985.
[5] E. M. Luks, "Isomorphism of Graphs of Bounded Valence Can Be Tested in Polynomial
Time," Journal of Computer and System Sciences, vol. 25, no. 1, pp. 42 -- 65, 1982.
[6] L. Babai, "Graph Isomorphism in Quasipolynomial Time," in Proceedings of the 48th An-
nual ACM Symposium on Theory of Computing, pp. 684 -- 697, 2016.
[7] J. E. Hopcroft and J.-K. Wong, "Linear Time Algorithm for Isomorphism of Planar
Graphs," in Proceedings of the 6th Annual Symposium on Theory of Computing, pp. 172 --
184, 1974.
[8] L. Babai, D. Y. Grigoryev, and D. M. Mount, "Isomorphism of Graphs with Bounded Eigen-
value Multiplicity," in Proceedings of the 14th Annual Symposium on Theory of Computing,
pp. 310 -- 324, 1982.
[9] B. D. McKay and A. Piperno, "Practical graph isomorphism, ii," Journal of Symbolic
Computation, vol. 60, pp. 94 -- 112, 2014.
[10] T. Junttila and P. Kaski, "Engineering an Efficient Canonical Labeling Tool for Large
and Sparse Graphs," in Proceedings of the 9th Workshop on Algorithm Engineering and
Experiments, pp. 135 -- 149, 2007.
13
[11] P. T. Darga, K. A. Sakallah, and I. L. Markov, "Faster Symmetry Discovery using Sparsity
of Symmetries," in Proceedings of the 45th Design Automation Conference, pp. 149 -- 154,
2008.
[12] Y. Aflalo, A. Bronstein, and R. Kimmel, "On convex relaxation of graph isomorphism,"
Proceedings of the National Academy of Sciences, vol. 112, no. 10, pp. 2942 -- 2947, 2015.
[13] D. Goldfarb and S. Liu, "An O(n3l) primal interior point algorithm for convex quadratic
programming," Mathematical Programming, vol. 49, no. 1-3, pp. 325 -- 340, 1990.
[14] M. Pelillo, "Replicator Equations, Maximal Cliques, and Graph Isomorphism," in Advances
in Neural Information Processing Systems, pp. 550 -- 556, 1999.
[15] G. Levi, "A note on the derivation of maximal common subgraphs of two directed or
undirected graphs," Calcolo, vol. 9, no. 4, p. 341, 1973.
[16] H. G. Barrow and R. M. Burstall, "Subgraph isomorphism, matching relational structures
and maximal cliques," Information Processing Letters, vol. 4, no. 4, pp. 83 -- 84, 1976.
[17] D. Kozen, "A clique problem equivalent to graph isomorphism," ACM SIGACT News,
vol. 10, no. 2, pp. 50 -- 52, 1978.
[18] I. Koch, "Enumerating all connected maximal common subgraphs in two graphs," Theo-
retical Computer Science, vol. 250, no. 1-2, pp. 1 -- 30, 2001.
[19] N. Kriege and P. Mutzel, "Subgraph Matching Kernels for Attributed Graphs," in Proceed-
ings of the 29th International Coference on International Conference on Machine Learning,
pp. 291 -- 298, 2012.
[20] M. Grotschel, L. Lov´asz, and A. Schrijver, Geometric algorithms and combinatorial opti-
mization, vol. 2. Springer Science & Business Media, 2012.
[21] Y. Nesterov and A. Nemirovskii, Interior-point polynomial algorithms in convex program-
ming, vol. 13. SIAM, 1994.
[22] M. Grotschel, L. Lov´asz, and A. Schrijver, "The ellipsoid method and its consequences in
combinatorial optimization," Combinatorica, vol. 1, no. 2, pp. 169 -- 197, 1981.
[23] L. Vandenberghe and S. Boyd, "Semidefinite programming," SIAM review, vol. 38, no. 1,
pp. 49 -- 95, 1996.
[24] M. X. Goemans and D. P. Williamson, "Improved Approximation Algorithms for Maximum
Cut and Satisfiability Problems Using Semidefinite Programming," Journal of the ACM,
vol. 42, no. 6, pp. 1115 -- 1145, 1995.
[25] A. Srivastav and K. Wolf, "Finding dense subgraphs with semidefinite programming," in
Proceedings of the International Workshop on Approximation Algorithms for Combinatorial
Optimization, pp. 181 -- 191, 1998.
[26] U. Feige and R. Krauthgamer, "Finding and certifying a large hidden clique in a semiran-
dom graph," Random Structures & Algorithms, vol. 16, no. 2, pp. 195 -- 208, 2000.
14
|
1712.09473 | 1 | 1712 | 2017-12-27T01:26:52 | Sketching for Kronecker Product Regression and P-splines | [
"cs.DS",
"cs.LG",
"stat.ML"
] | TensorSketch is an oblivious linear sketch introduced in Pagh'13 and later used in Pham, Pagh'13 in the context of SVMs for polynomial kernels. It was shown in Avron, Nguyen, Woodruff'14 that TensorSketch provides a subspace embedding, and therefore can be used for canonical correlation analysis, low rank approximation, and principal component regression for the polynomial kernel. We take TensorSketch outside of the context of polynomials kernels, and show its utility in applications in which the underlying design matrix is a Kronecker product of smaller matrices. This allows us to solve Kronecker product regression and non-negative Kronecker product regression, as well as regularized spline regression. Our main technical result is then in extending TensorSketch to other norms. That is, TensorSketch only provides input sparsity time for Kronecker product regression with respect to the $2$-norm. We show how to solve Kronecker product regression with respect to the $1$-norm in time sublinear in the time required for computing the Kronecker product, as well as for more general $p$-norms. | cs.DS | cs | Sketching for Kronecker Product Regression and P-splines
7
1
0
2
c
e
D
7
2
]
S
D
.
s
c
[
1
v
3
7
4
9
0
.
2
1
7
1
:
v
i
X
r
a
Huaian Diao
Zhao Song
Wen Sun
[email protected]
[email protected]
[email protected]
David P. Woodruff
[email protected]
Northeast Normal University Harvard U & UT-Austin Carnegie Mellon University Carnegie Mellon University
Abstract
TensorSketch is an oblivious linear sketch
introduced in (Pagh, 2013) and later used
in (Pham and Pagh, 2013) in the context of
SVMs for polynomial kernels. It was shown
in (Avron et al., 2014) that TensorSketch
provides a subspace embedding, and there-
fore can be used for canonical correlation
analysis, low rank approximation, and princi-
pal component regression for the polynomial
kernel. We take TensorSketch outside of
the context of polynomials kernels, and show
its utility in applications in which the under-
lying design matrix is a Kronecker product of
smaller matrices. This allows us to solve Kro-
necker product regression and non-negative
Kronecker product regression, as well as reg-
ularized spline regression. Our main tech-
nical result is then in extending TensorS-
ketch to other norms. That is, TensorS-
ketch only provides input sparsity time for
Kronecker product regression with respect to
the 2-norm. We show how to solve Kronecker
product regression with respect to the 1-norm
in time sublinear in the time required for com-
puting the Kronecker product, as well as for
more general p-norms.
1
INTRODUCTION
In the overconstrained least squares regression prob-
lem, we are given an n× d matrix A called the "design
matrix", n ≫ d, and an n × 1 vector b, and the goal
is to find an x which minimizes kAx − bk2. There are
many variants to this problem, such as regularized ver-
sions of the problem in which one seeks an x so as to
Proceedings of the 21st International Conference on Artifi-
cial Intelligence and Statistics (AISTATS) 2018, Lanzarote,
Spain. JMLR: W&CP volume 7X. Copyright 2018 by the
author(s).
2 + kLxk2
minimize kAx − bk2
2 for a matrix L, or regres-
sion problems which seek to minimize more robust loss
functions, such as ℓ1-regression kAx − bk1.
In the era of big data, large scale matrix computa-
tions have attracted considerable interest. To obtain
reasonable computational and time complexities for
large scale matrix computations, a number of random-
ized approximation algorithms have been developed.
For example, in (Woodruff, 2014), it was shown how
to output a vector x′ ∈ Rd for which kAx′ − bk2 ≤
(1 + ǫ) minx∈Rd kAx − bk2 in nnz(A) + poly(d/ǫ) time,
where nnz(A) denotes the number of non-zero en-
tries of matrix A. We refer the reader to the recent
surveys (Kannan and Vempala, 2009; Mahoney, 2011;
Woodruff, 2014) for a detailed treatment of this topic.
In this work we focus on regression problems for which
the design matrix is a Kronecker product matrix, that
is, it has the form A ⊗ B for A ∈ Rn1×d1 and B ∈
Rn2×d2. Also, b ∈ Rn1n2 , and one seeks to solve the
problem minx∈Rd1d2 k(A ⊗ B)x − bk2 in the standard
setting, which can also be generalized to regularized
and robust variants. One can also ask the question
when the design matrix is a Kronecker product of more
than two matrices.
Kronecker product matrices have many applications
in statistics, linear system theory, signal processing,
photogrammetry, multivariate data fitting, etc.; see
(Golub and Van Loan, 2013; Van Loan and Pitsianis,
1993; Van Loan, 1992). The linear least squares prob-
lem involving the Kronecker product arises in many
applications, such as structured linear total least norm
on blind deconvolution problems (Oh and Yun, 2005),
constrained linear least squares problems with a Kro-
necker product structure, the bivariate problem of sur-
face fitting and multidimensional density smoothing
(Eilers and Marx, 2006).
One way to solve Kronecker product regression is
to form the matrix C = A ⊗ B explicitly, where
A ∈ Rn1×d1, B ∈ Rn2×d2, and then apply a ran-
domized algorithm to C. However, this takes at
least nnz(A) · nnz(B) time and space.
It is natural
Sketching for Kronecker Product Regression and P-splines
to ask if it is possible to solve the Kronecker prod-
uct regression problem in time faster than computing
A ⊗ B. This is in fact the case, as Fausett and Fulton
(Fausett and Fulton, 1994) show that one can solve
Kronecker product regression in O(n1d2
2) time;
indeed, the solution vector x = vec((B⊥)⊤D−1A⊥),
where D = vec(b) and vec(E) for a matrix E denotes
the operation of stacking the columns of E into a sin-
gle long vector. While such a computation does not
involve computing A ⊗ B, it is more expensive than
what one would like.
1 + n2d2
A natural question is if one can approximately solve
Kronecker product regression in nnz(A) + nnz(B) +
poly(d/ǫ) time, which, up to the poly(d/ǫ) term, would
match the description size of the input. Another natu-
ral question is Kronecker product regression with reg-
ularization. Such regression problems arise frequently
in the context of splines (Eilers and Marx, 2006). Fi-
nally, what about Kronecker product regression with
other, more robust, norms such as the ℓ1-norm?
1.1 Our Contributions
is exactly suited for this task.
We first observe that the random linear map TensorS-
ketch, introduced in the context of problems for the
polynomial kernel by Pagh (2013), Pham and Pagh
(2013),
Namely,
in (Avron et al., 2014) it was shown that for a d-
dimensional subspace C of Rn, represented as an n× d
matrix, there is a distribution on linear maps S with
O(d2/ǫ2) rows such that with constant probability, si-
multaneously for all x ∈ Rd, kSCxk2 = (1 ± ǫ)kCxk2.
That is, S provides an Oblivious Subspace Embed-
ding (OSE) for the column span of C. Further, it
is known that if b is an n-dimensional vector, then
one has that kS[C, b]xk2 = (1 ± ǫ)k[C, b]xk2 for all
x ∈ Rd+1. Consequently, to solve the regression prob-
lem minx∈Rd kCx−bk2, one can instead solve the much
smaller problem minx∈Rd kSCx − Sbk2, and the solu-
tion x′ to the latter problem will satisfy kCx′ − bk2 ≤
(1 + ǫ) minx∈Rd kCx − bk2.
Importantly, if n = n1 · n1 and there is a basis for
the column span of C of the form A1 ⊗ A1, A2 ⊗
A2, . . . , Ad ⊗ Ad, then given A1, . . . , Ad, it holds that
SC can be computed in nnz(A) time, where A is
the n1 × d matrix whose i-th column is Ai. Fur-
ther, given a vector b with nnz(b) non-zero entries,
one can compute Sb in nnz(b) time. Thus, one ob-
tains a (1 + ǫ)-approximate solution to the regression
minx∈Rd kCx−bk2 in nnz(A)+nnz(b)+poly(d/ǫ) time.
While not immediately useful for our problem, we
show that via simple modifications, the claim about
TensorSketch above can be generalized to the case
when there is a basis for the column span of C of
the form Ai ⊗ Bj for arbitrary vectors A1, . . . , Ad1 ∈
Rn1 and vectors B1, . . . , Bd2 ∈ Rn2 . That is, we
observe that in this case SC can be computed in
nnz(A) + nnz(B) time, where A is the n1 × d1 matrix
whose i-th column is Ai, and B is the n2 × d2 matrix
whose i-th column is Bi.
In this case C = A ⊗ B,
which is exactly the case of Kronecker product regres-
sion. Using the above connection to regression, we
obtain an algorithm for Kronecker product regression
in nnz(A) + nnz(B) + nnz(b) + poly(d1d2/ǫ) time. Us-
ing the fact that kSCx − Sbk2 = (1 ± ǫ)kCx − bk2 for
all x, we have in particular that this holds for all non-
negative x, and so also obtain the same reduction in
problem size for non-negative Kronecker product re-
gression, which occurs often in image and text data;
see e.g., (Chen and Plemmons, 2010).
2 + λkxk2
The above observation allows us to extend many ex-
isting variants of least squares regression to the case
when C is a Kronecker product of two matrices. For ex-
ample, the results in (Avron et al., 2016) for the ridge
regression problem minx∈Rd kAx − bk2
2 imme-
diately hold when C is a Kronecker product matrix,
since the conditions needed for the oblivious embed-
ding in (Avron et al., 2016) come down to a subspace
embedding and an approximate matrix product condi-
tion, both of which are known to hold for TensorS-
ketch (Avron et al., 2014). More interestingly we are
able to extend the results in (Avron et al., 2016) for
Kronecker ridge regression to the case when the reg-
ularizer is a general matrix, namely, to the problem
minx∈Rd kAx − bk2
2, where L is an arbitrary
matrix. Such problems occur in the context of spline
regression (Eilers and Marx, 1996, 2006; Eilers et al.,
2015). The number of rows of TensorSketch de-
pends on a generalized notion of statistical dimen-
sion depending on the generalized singular values γi
of [A; L], and may be much smaller than d1 or d2. In
(Avron et al., 2016), only results for L equal to the
identity were obtained.
2 + λkLxk2
Finally, our main technical result is to extend our re-
sults to least absolute deviation Kronecker product re-
gression minx∈Rd kCx − bk1, which, since it involves
the 1-norm, is often considered more robust than least
squares regression (Rousseeuw and Leroy, 2005) and
has been widely used in applications such as com-
puter vision (Zheng et al., 2012). We in fact extend
this to general p-norms but focus the discussion on
p = 1. We give the first algorithms for this problem
that are faster than computing C = A ⊗ B explicitly,
which would take at least nnz(A)·nnz(B) ≥ n1n2 time.
Namely, and for simplicitly focusing on the case when
n1 = n2, for which the goal is to do better than n2
1
time, we show how to output an x ∈ Rd1×d2 for which
kCx′ − bk1 ≤ (1 + ǫ) minx∈Rd1d2 kCx − bk1 in time
n3/2 poly(d1d2/ǫ). While this is more expensive than
Diao, Song, Sun, Woodruff
solving Kronecker product least squares, the 1-norm
may lead to significantly more robust solutions. From
a technical perspective, TensorSketch when applied
to a vector actually destroys the 1-norm of a vector,
preserving only its 2-norm, so new ideas are needed.
We show how to use multiple TensorSketch matri-
ces to implicitly obtain very crude estimates to the so-
called ℓ1-leverage scores of C, which can be interpreted
as probabilities of sampling rows of C in order to ob-
tain an ℓ1-subspace embedding of the column span of
C (see, e.g., (Woodruff, 2014)). However, since C has
n1n2 rows, we cannot even afford to write down the
ℓ1-leverage scores of C. We show how to implicitly rep-
resent such leverage scores and to sample from them
without ever explicitly writing them down. Balancing
the phases of our algorithm leads to our overall time.
A1 ⊗ A2 ⊗ ...⊗ Aq, n =Qq
i=1 ni, and d =Qq
a vector x ∈ Rd is defined as kxkp = (Pd
1.2 Notation
We consider Kronecker Product of q 2-d matrices A1⊗
A2⊗ ...⊗ Aq, where each Ai ∈ Rni×di. We denote A =
i=1 di. We
denote [n] as the set {1, 2, 3,··· , n}. The ℓp norm for
i=1 xip)1/p,
where xi stands for the i'th entry of the vector x. For
any matrix M , we use Mi,∗ to represent the i'th row
of M and M∗,j as the j'th column of M .
We define a Well-Conditioned Basis and Statistical
Dimension as
follow (similar definitions can be
found in Clarkson (2005); Dasgupta et al. (2009);
Sohler and Woodruff
(2011); Meng and Mahoney
(2013); Song et al. (2017a,b) and Avron et al. (2016)):
Definition 1.1 (Well-Conditioned Basis). Let A be
an n × m matrix with rank d, let p ∈ [1,∞), and let
k · kq be the dual norm of k · kp, i.e., 1/p + 1/q = 1.
Then an n×d matrix U is an (α, β, p)-well-conditioned
basis for the column space of A, if the columns of U
span the column space of A and (1) kUkp ≤ α, and (2)
for all z ∈ Rd, kzkq ≤ βkU zkp.
Consider the classic Ridge Regression: minx kAx −
bk2
2. The Statistical Dimension is defined as:
Definition 1.2 (Statistical Dimension). Let A be an
n×m matrix with rank d and singular values σi, i ∈ [d].
For λ ≥ 0, the statistical dimension sdλ(A) is defined
as the quantity sdλ(A) =Pi∈[d] 1/(1 + λ/σ2
2 BACKGROUND:TensorSketch
2 + λkxk2
i ).
We briefly introduce TensorSketch (Pagh, 2013;
Avron et al., 2014) and how to apply TensorSketch
to the Kronecker product of multiple matrices effi-
ciently without explicitly computing the tensor prod-
uct.1
1We refer readers to (Avron et al., 2014) for more de-
tails about TensorSketch.
, ..., Aq∗,iq
··· × [nq] → [m] as H(i1, i2, ..., iq) = ((Pq
{−1, 1} as S(i1, i2, ..., iq) = Qq
We want to find a oblivious subspace embedding S
such that for any x ∈ Rn, we have kSAxk2 = (1 ±
ǫ)kAxk2, where the notation a = (1 ± ǫb) stands for
(1 − ǫ)b ≤ a ≤ (1 + ǫ)b, for any a, b ∈ R. Con-
sider the (i1, i2, ..., iq)'th column of A (ij ∈ [dj]):
A1∗,i1 ⊗ A2∗,i2 ⊗···⊗ Aq∗,iq . Assume the sketching tar-
get dimension is m. TensorSketch is defined using q
3-wise indepedent hash functions hi : [ni] → [m], and
q 4-wise independent sign functions si : [ni] → {−1, 1},
∀i ∈ [q]. Define hash function H :
[n1] × [n2] ×
k=1 hk(ik))
mod m), and sign function S : [n1]× [n2]×···× [nq] →
k=1 sk(ik). Applying
Tensorsketch to the Kronecker product of vectors
is equivalent to applying CountS-
A1∗,i1
ketch (Charikar et al., 2004) defined with H and S
to the vector (A1∗,i1 ⊗ ...⊗ Aq∗,iq ). To apply Tensors-
ketch to A, we just need to apply CountSketch
defined with H and S to the columns of A one by one.
Applying tensorsketch to the Kronecker product of
, ..., Aq∗,iq ) naively would require at least O(n)
(A1∗,i1
time. Pagh (2013) shows that one can apply tensors-
, ..., Aq∗,iq )
ketch to the Kronecker product of (A1∗,i1
without explicitly computing the Kronecker product
of these vectors using the Fast Fourier Transforma-
tion. Particularly, Pham and Pagh (2013) show that
) + qm log(m)) time
to compute S(A1∗,i1 ⊗...⊗Aq∗,iq ) where Ai∗,j stands for
the j'th column of Ai. As A has d columns, comput-
i=1 nnz(Ai)) + dqm log(m)) time,
i=1 ni), which is the
least amount of time one needs for explicitly comput-
In the rest of the paper, we assume that we
ing A.
compute SA using the above efficient procedure with-
out explicitly computing A1 ⊗ . . . ⊗ Aq.
one only needs O(Pq
ing SA takes O(d(Pq
which is much smaller than O(Qq
j=1 nnz(Aj∗,ij
3 TENSOR PRODUCT LEAST
SQUARES REGRESSION
Consider the tensor product least squares regression
problem minx kAx − bk2 where A ∈ Rn×d and b ∈ Rn.
Let S ∈ Rm×n be the matrix form of TensorSketch
of Section 2. We propose a TensorSketch -type
Algorithm 1 for the tensor product regression prob-
lem. The following theorem shows that the solution
obtained from Alg. 1 is a good approximation of the
optimal solution of the original tensor product regres-
sion.
Let us define OPT to be the optimal cost of the opti-
mization problem, e.g., OPT = minx kAx − bk2. The
following theorem shows that Alg. 1 computes a good
approximate solution.
Theorem 3.1. (Tensor regression) Suppose ex is the
output of Algorithm 1 with tensorsketch S ∈
Sketching for Kronecker Product Regression and P-splines
Algorithm 1 Tensor product regression
1: procedure TRegression(A, b, ǫ, δ)
2:
3:
m ← (d1d2 ··· dq + 1)2(2 + 3q)/(ǫ2δ)
Choose S to be an m× (n1n2 ··· nq) TensorS-
Compute S(A1 ⊗ A2 ⊗ ··· ⊗ Aq) and Sb
ketch matrix
4:
5:
6:
7: end procedure
ex ← minx kS(A1 ⊗ A2 ⊗ ··· ⊗ Aq)x − Sbk2
return ex
Aq)ex − bk2 ≤ (1 + ǫ) OPT, holds with probability at
Rm×n, where m = 8(d1d2 ··· dq + 1)2(2 + 3q)/(ǫ2δ).
Then the following approximation k(A1 ⊗ A2 ⊗ ··· ⊗
least 1 − δ.
The proof of Theorem 3.1 can be found in Appendix
C.1. Theorem 3.1 shows that we can achieve an ǫ-close
solution by solving a much smaller regression problem
with a number of samples of order O(poly(d/ǫ)), which
is independent of the large dimension n. Using the
technique we introduced in Sec. 2, we can also compute
SA without explicitly computing the tensor product.
We can extend Theorem 3.1 to the nonnegative ten-
sor product regression problem minx≥0 k(A1 ⊗ A2 ⊗
··· ⊗ Aq)x − bk2, where Ai ∈ Rni×di, i = 1, . . . , q and
b ∈ Rn1n2···nq . Suppose x is the optimal solution. Sim-
ilarly, let S ∈ Rm×(n1n2···nq) be the matrix form of
lution to minx≥0 kS(A1 ⊗ A2 ⊗ ··· ⊗ Aq)x − Sbk2., we
have the following:
TensorSketch of Section 2. If ex is the optimal so-
Corollary 3.2. (Sketch for tensor nonnegative regres-
sion) Suppose x = minx≥0 kSAx − Sbk2 with ten-
sorsketch S ∈ Rm×n, where m = 8(d1d2 ··· dq +
1)2(2 + 3q)/(ǫ2δ). Then the following approxima-
tion k(A1 ⊗ A2 ⊗ ··· ⊗ Aq)ex − bk2 ≤ (1 + ǫ) OPT
holds with probability at least 1 − δ, where OPT =
minx≥0 k(A1 ⊗ A2 ⊗ ··· ⊗ Aq)x − bk2.
The proof of Corollary 3.2 can be found in Ap-
pendix C.1.
4 P-SPLINES
B-splines are local basis functions, consisting of low de-
gree (e.g., quadratic, cubic) polynomial segments. The
positions where the segments join are called the knots.
B-splines have local support and are thus suitable for
smoothing and interpolating data with complex pat-
terns. Unfortunately, control over smoothness is lim-
ited: one can only change the number and positions
of the knots. If there are no reasons to assume that
smoothness is non-uniform, the knots will be equally
spaced and the only tuning parameter is their (dis-
crete) number. In contrast P-spline (Eilers and Marx,
1996) equally spaces B-splines, discards the deriva-
tive completely, and controls smoothness by regular-
izing the sum of squares of differences of coefficients.
Specifically Eilers and Marx Eilers and Marx (1996)
proposed the P-spline recipe: (1) use a (quadratic or
cubic) B-spline basis with a large number of knots,
say 10-50; (2) introduce a penalty on (second or third
order) differences of the B-spline coefficients; (3) min-
imize the resulting penalized likelihood function; (4)
tune smoothness with the weight of the penalty, us-
ing cross-validation or AIC to determine the optimal
weight.
We give a brief overview of B-Splines and P-Splines
below. Let b and u, each vectors of length n, represent
the observed and explanatory variables, respectively.
Once a set of knots is chosen, the B-spline basis A fol-
lows from u. If there are d basis functions then A is
n× d. In the case of normally distributed observations
the model is b = Ax + e, with independent errors e. In
the case of B-spline regression the sum of squares of
residuals kb − Axk2 is minimized and the normal equa-
tions A⊤Ax = A⊤b are obtained; the explicit solution
x = (A⊤A)−1A⊤b results. The P-spline approach min-
imizes the penalized least-squares function
kb − Axk2
2 + λkLxk2
2,
(1)
where L ∈ Rp×n is a matrix that forms differences of
order ℓ, i.e., Lℓx = ∆ℓx. Examples of this matrix, for
ℓ = 1 and ℓ = 2 are :
L1 =
−1
1
0 −1
0
0
0
0
1
0 −1 1 , L2 =
1 −2
1
0
1 −2
0
0
0
0
0
1
1 −2 1 .
The parameter λ determines the influence of the
penalty. If λ is zero, we are back to B-spline regression;
increasing λ makes x, and hence b = Ax, smoother.
Let x∗ denote argminx∈RdkAx − bk2
2 + λkLxk2
2, and
2. In general x∗ =
OPT denote kAx∗ − bk2
2 + λkLx∗k2
(A⊤A + λL⊤L)−1A⊤b = A⊤(AA⊤ + λLL⊤)−1b, so x∗
can be found in O(nnz(A) min(n, d)) time using an it-
erative method (e.g., LSQR). Our first goal in this
section is to design faster algorithms that find an ap-
proximateex in the following sense:
kAex − bk2
2 + λkLexk2
4.1 Sketching for P-Spline
2 ≤ (1 + ǫ) OPT .
(2)
We first introduce a new definition of Statistical Di-
mension that extends the statistical dimension defined
for Ridge Regression (i.e., L is an identity matrix in
Eq. 1) (Avron et al., 2016) to P-Spline regression.
The problem (1) can also be analyzed by gen-
eralized
(GSVD)
decomposition
singular
value
Diao, Song, Sun, Woodruff
Σ
0p×(n−p)
0(n−p)×p
d, the GSVD of (A, L) is given by the pair of factoriza-
Golub and Van Loan (2013). For matrices A ∈ Rn×d
and L ∈ Rp×d with rank(L) = p and rank(cid:18)(cid:20) A
L (cid:21)(cid:19) =
Id−p (cid:21) RQ⊤ and L =
tions A = U(cid:20)
V (cid:2)Ω 0p×(n−p)(cid:3) RQ⊤, where U ∈ Rm×n has or-
thonormal columns, V ∈ Rp×p, Q ∈ Rd×d are
orthogonal, R ∈ Rd×d is upper triangular and non-
singular, and Σ and Ω are p × p diagonal matrices:
Σ = diag(σ1, σ2, . . . , σp) and Ω = diag(µ1, µ2, . . . , µp)
with 0 ≤ σ1 ≤ σ2 ≤ . . . ≤ σp < 1 and 1 ≥ µ1 ≥
µ2 ≥ . . . ≥ µp > 0, satisfying Σ⊤Σ + Ω⊤Ω = Ip. The
generalized singular values γi of (A, L) are defined by
the ratios γi = σi/µi (i = [p]).
In this section we design an algorithm that is aimed
at the case when n ≫ d. The general strategy is to
design a distribution on matrices of size m-by-n (m is
a parameter), sample an S from that distribution, and
2 + λkLxk2
2 .
solveex ≡ argminx∈RdkS(Ax − b)k2
The following lemma defines conditions on the distribu-
tion of S that guarantees Eq. (2) holds with constant
probability (which can be boosted to high probability
by repetition and taking the minimum objective).
Lemma 4.1. Let x∗ ∈ Rd, A ∈ Rn×d and b ∈ Rn as
above. Let U1 ∈ Rn×d denote the first n rows of an
orthogonal basis for h A√λLi ∈ R(n+p)×d. Let sketching
matrix S ∈ Rm×n have a distribution such that with
constant probability
(I) kU⊤1 S⊤SU1 − U⊤1 U1k2 ≤ 1/4,
and
(II) kU⊤1 (S⊤S − I)(b − Ax∗)k2 ≤pǫ OPT /2.
Letex denote argminx∈RdkS(Ax− b)k2
with probability at least 9/10,
2 + λkLxk2
2. Then
kAex − bk2
2 + λkLexk2
2 ≤ (1 + ǫ) OPT .
Define the statistical dimension for P-Splines as fol-
lows:
i ) + d − p.
Definition 4.2 (Statistical Dimension for P-Splines).
For S-Spline in Eq. (1), the statistical dimension is
defined as sdλ(A, L) =Pi 1/(1 + λ/γ2
The following theorem shows that there is a sparse
subspace embedding matrix S ∈ Rm×n (e.g., CountS-
ketch), with m ≥ K(sdλ(A, L)/ǫ + sdλ(A, L)2), that
satisfies Property (I) and (II) of Lemma 4.1, and hence
achieves an ǫ−approximation solution to problem 1:
Theorem 4.3. (P-Spline regression) There is a con-
stant K > 0 such that for m ≥ K(ǫ−1 sdλ(A, L) +
Algorithm 2 P-Spline Tensor product regression
1: procedure PTRegression(A, b, K, L, ǫ, δ)
2:
3:
4:
5:
6:
7: end procedure
m ← K(ǫ−1 sdλ(A, L) + sdλ(A, L)2)
Choose S to be a m×n TensorSketch matrix
Compute S(A1 ⊗ A2 ⊗ ··· ⊗ Aq) and Sb
2 + λkLxk2
2.
ex = argminx∈RdkS(Ax − b)k2
return ex
ing ex = argminx∈RdkS(Ax − b)k2 + λkLxk2
sdλ(A, L)2) and S ∈ Rm×n a sparse embedding ma-
trix (e.g., Countsketch) with SA computable in
O(nnz(A)) time, Property (I) and (II) of Lemma 4.1
apply, and with constant probability the correspond-
2 is an ǫ-
2 + λkLxk2
approximate solution to minx∈Rdkb − Axk2
2.
Note sdλ(A, L) is upper bounded by d. The above
theorem shows that the statistical dimension al-
lows us to design smaller sketch matrices whose
size only depends on O(poly(sdλ(A, L)/ǫ)) instead of
O(poly(d/ǫ)), without sacrificing the approximation
accuracy.
4.2 Tensor Sketching for Multi-Dimensional
P-Spline
Tensor products allow a natural extension of one-
dimensional P-spline smoothing to multi-dimensional
P-Spline. We focus on 2-dimensional P-Spline but our
results can be generalized to the multi-dimensional set-
ting. Assume that in addition to u we have a sec-
ond explanatory variable v. We have data triples
(ui, vj, b(i−1)·n2+j) for i = 1, . . . , n1 and j = 1, . . . , n2.
We seek a smooth surface f (u, v) which gives a good
approximation to the response b. Let A1, n × d1, be a
B-spline basis along u, and A2, n × d2, be a B-spline
basis along v. We form the tensor product basis as
A1 ⊗ A2. When n1 and n2 are large, we do not com-
pute A1⊗ A2. We apply tensorsketch here to avoid
explicitly forming A1 ⊗ A2 to speed up computation.
Let X = [xkl] be a d1 × d2 matrix of co-
the value fit
efficients.
at
(u, v)
and so X may be chosen using least
is f (u, v) = PkPl A2,k(v)A1,l(u)xkl
by minimizing Pi,j[b(i−1)·n2+j − f (ui, vi)]2 =
Pi(cid:2)b(i−1)·n2+j −PkPl A2,k(vi)A1,l(ui)xkl(cid:3)2
. Using
Kronecker product, the above minimization can be
written in the form min kb − Axk2 , where A = A1 ⊗
A2 ∈ Rn1n2×d1d2 and x = vec(X). Again the P-
spline approach minimizes the penalized least-squares
function kb − (A1 ⊗ A2)xk2
2 + λkLxk2
2. Consider the
tensor p-spline regression problem minx k(A1 ⊗ A2 ⊗
··· ⊗ Aq)x − bk2
2 + λkLxk2
2, where L ∈ Rp×n, Ai ∈
Rni×di, i = 1, . . . , q and b ∈ Rn. Let S ∈ Rm×n
be the matrix form of TensorSketch of Section 2.
for given X,
squares
Then,
Sketching for Kronecker Product Regression and P-splines
Algorithm 2 summarizes the procedure for efficiently
solving multi-dimensional P-Spline.
Let A = A1⊗ A2⊗···⊗ Aq. Replacing the matrix A in
Theorem 4.3 with A, we have the following corollary
for multi-dimensional P-Spline:
Corollary 4.4 (P-Spline tensor regression). Suppose
λ ≤ σ2
1/ǫ. There is a constant K > 0 such that
for m ≥ K(ǫ−1 sdλ(A, L) + sdλ(A, L)2) and S ∈
Rm×n a TensorSketch matrix with SA computable
in O(nnz(A)) time, Property (I) and Property (II) of
Lemma 4.1 apply, and with constant probability the
corresponding ex = argminx∈RdkS(Ax − b)k2 + λkLxk2
is an ǫ-approximate solution to minx∈Rdkb − Axk2
λkLxk2
2.
5 TENSOR PRODUCT ABSOLUTE
2
2 +
DEVIATION REGRESSION
i=1 d2
running a cross validation procedure similar to that in
Algorithm 4 of Liang et al. (2014), to find one which
succeeds with probability at least 1 − δ:
Lemma 5.1 ((Liang et al., 2014)). For δ, ǫ ∈ (0, 1),
i (2 + 3q)/ǫ2.
Running algorithm Algorithm 4 in (Liang et al., 2014)
with parameters t, m, we can obtain a tensorsketch
i=1 wi such that with probability at least 1−δ,
i2 ⊗ ··· ⊗ A(q)
i2 ⊗
let t = O(log 1/δ) and m ≥ 100Qq
S ∈ Rm×Qq
i1 ⊗ A(2)
kSA(1)
··· ⊗ A(q)
iq xk2 for all x ∈ Rd.
computing a tensorsketch Si1i2...iq
i=1 wi
∈
After
Rm×Qq
each
row-block A(1)
compose
=
a
S
diag(S1,··· ,1,··· , Si1,··· ,iq ,··· , S(n1/w1),··· ,(nq/wq))
∈
RmQq
i1 ⊗ ... ⊗ A(q)
iq ,
we
for A as
i=1 ni , which is defined as:
iq xk2 = (1 ± ǫ)kA(1)
i=1 ni/wi×Qq
i1 ⊗ A(2)
tensorsketch
lemma
single
using
5.1
for
We extend our previous results for the ℓ2 norm (i.e.,
least squares regression) to general ℓp norms, with a
focus on p = 1 (i.e., absolute deviation regression).
Specifically we consider minx kAx − bkp, where A =
A1 ⊗ A2 ⊗ ··· ⊗ Aq. We will show in this section that
with probability at least 2/3, we can quickly find an x
for which kAx − bk1 ≤ (1 + ǫ) minx kAx − bk1.
As in Clarkson and Woodruff (2013), for each i,
in
O(nnz(Ai) log(di)) + O(r3
i ) time we can replace the
input matrix Ai ∈ Rni×di with a new matrix with the
same column space of Ai and full column rank. We
therefore assume A has full rank in what follows.
Suppose S is the TensorSketch matrix defined in
Section 2. Let wi ∈ N and assume wi ni. Split Ai
into ni/wi matrices A(i)
, each wi × di, so
that A(i)
is the submatrix of Ai indexed by the j-th
j
block of wi rows. Note A can be written as:
⊗ A(q)
⊗ A(q)
1 ⊗ ··· ⊗ A(q−1)
1 ⊗ ··· ⊗ A(q−1)
A(1)
1 ⊗ A(2)
1 ⊗ A(2)
A(1)
1 , . . . , A(i)
ni/wi
1
1
1
2
...
n1/w1 ⊗ A(2)
A(1)
n2/w2 ⊗ ··· ⊗ A(q−1)
nq−1/wq−1 ⊗ A(q)
nq/wq
.
iq
i1 ⊗ A(2)
i=1 d2
i2 ⊗ ··· ⊗ A(q)
For each A(1)
, we can use the Ten-
sorSketch matrix Si1i2...iq ∈ Rm×Qq
i=1 wi , where we
set m ≥ 100Qq
i (2 + 3q)/ǫ2, such that with proba-
i1 ⊗A(2)
bility at least .99, kSi1i2...iq A(1)
iq xk2 =
i2 ⊗ ··· ⊗ A(q)
i1 ⊗ A(2)
(1 ± ǫ)kA(1)
xk2 simultaneously for
all x ∈ Rd as Si1i2...iq is an oblivious subspace embed-
ding (Lemma. B.3) . Now we use Algorithm 4 from
(Liang et al., 2014) to boost the success probability
by computing t = O(log(1/δ)) independent TensorS-
ketch products S(j)
, j = [t],
each with only constant success probability, and then
i2 ⊗···⊗A(q)
i2 ⊗···⊗A(q)
i1 ⊗A(2)
A(1)
i1i2...iq
iq
iq
.
S1,··· ,1
. . .
Si1,··· ,iq
. . .
S(n1/w1),··· ,(nq/wq)
,
i=1 ni/wi×Qq
let S ∈ RmQq
Note that A has in total Qq
where each block on the diagonal is from Lemma 5.1.
i=1(ni/wi) many blocks.
Using Lemma 5.1 with a union bound over all blocks
of A, we have the following theorem which shows S is
an oblivious subspace embedding for A in ℓ2 norm:
Theorem 5.2 (ℓ2 OSE for tensor matrices). Given
i=1 ni denote
δ, ǫ ∈ (0, 1),
the matrix that has Qq
i=1 ni/wi diagonal block matri-
ces where each diagonal block Si1i2...iq ∈ Rm×Qq
i=1 wi
is from Lemma 5.1. With probability at least 1 −
Qq
i=1(ni/wi)δ, kSAxk2 = (1 ± ǫ)kAxk2,∀x ∈ Rd.
It is known that for any matrix A ∈ Rn×r, we can
compute a change of basis U ∈ Rr×r such that AU
is an (α, β, p) well-conditioned basis of A (see Defi-
nition 1.1), in time polynomial with respect to n, r
(Dasgupta et al., 2009). Specifically, Theorem 5 in
(Dasgupta et al., 2009) shows that we can compute a
change of basis U for which AU is a well-conditioned
basis of A in time O(nr5 log(n)) time. However
we cannot afford to directly use the results from
(Dasgupta et al., 2009) to compute a well-conditioned
basis for A, which requires time at least Ω(n). Instead
we compute a well-conditioned basis for A through a
sketch SA where S is the tensorsketch from Theo-
rem 5.2.
the
Specifically we define
following procedure
Condition(A) for computing a well-conditioned basis
for A. Given A = A1 ⊗ ··· ⊗ Aq ∈ Rn×r, 1) Compute
SA; 2) Compute a d × d change of basis matrix U
Diao, Song, Sun, Woodruff
i=1(ni/wi)δ.
i=1 mi and w = Qq
so that SAU is an (α, β, p)-well-conditioned basis of
the column space of SA; 3) Output AU/(dγp), where
γp ≡ √2t1/p−1/2 for p ≤ 2, and γp ≡ √2w1/2−1/p for
p ≥ 2, where t = Qq
i=1 wi. The
following Lemma 5.3 (the proof can be found in the
Appendix) is the analogue of that in Clarkson et al.
(2013) proved for the Fast Johnson Lindenstauss
Transform.
Lemma 5.3. For any p ≥ 1. Condition(A) computes
AU/(dγp) which is an (α, β√3d(tw)1/p−1/2, p)-well-
conditioned basis of A, with probability at least 1 −
Qq
Lemma 5.3 indicates that Condition(A) computes a
(α, β · poly(max(d, log n)), p)-well-conditioned basis. A
well-conditioned basis can be used to solve ℓp regres-
sion problems, via sampling a subset of rows of the
well-conditioned basis AU with probabilities propor-
tional to the p-th power of the ℓp norm of the rows
(Woodruff, 2014). However the first issue for sampling
is that we cannot afford to compute AU as this re-
quires O(nnz(A)d) time. To fix this, we apply a Gaus-
sian sketch matrix G ∈ Rd×log(n) with i.i.d normal ran-
dom variables (Drineas et al., 2011) on the right hand
side of AU . Note that AU G can be computed effi-
ciently by first compuing U G and then A(U G) in time
O(d2 log(n) + nnz(A) log(n)). The second issue is that
even with AU G, computing the ℓp norm of each row
of AU G takes O(n) time, but we want sublinear time.
This leads us to the following sampling technique.
M = A1 ⊗ ... ⊗ Aq1 E(Aq1+1 ⊗ ... ⊗ Aq)⊤,
The high level
idea is that since AU G only has
O(log(n)) columns, we can afford to sample columns of
AU G with probability proportional to the p-th power
of the ℓp norms of the columns, if we can efficiently es-
timate the ℓp norms of the columns (note that naıvely
computing the ℓp norm of a column also takes O(n)
time). Let us denote the first of column of U G as
e ∈ Rlog(n). We focus on how to efficiently estimate
the ℓp norm of the first column of AU G, which is Ae,
and all the left columns can be estimated in the same
way. We first reshape the vector Ae into a 2-d matrix
(3)
where M ∈ R(n1n2...nq1 )×(nq1+1...nq ), E is obtained
from reshaping e into a (d1d2...dq1 ) × (dq1+1..dq) ma-
trix and q1 ∈ [1, q] is chosen such that (n1n2...nq1 ) ≈
(nq1+1...nq). Namely we reshape the column Ae into a
(nearly) square matrix. Focusing now on p = 1, note
that kAek1 =Pnq1+1...nq
kM∗,ik1. Hence to estimate
kAek1, we only need to estimate the ℓ1 norm of the
columns of M .
Let us apply a sketch matrix R ∈ RO(log(n))×Qq1
i=1 ni,
whose entries are sampled i.i.d. from the Cauchy dis-
tribution, to the left hand side of M . For the i'th col-
umn of M , let us define random variables zi
Mi,
i=1
l = R⊤l,∗
Algorithm 3 ℓ1 tensor product regression
1: procedure L1TRregression(A, b, ǫ, δ)
2:
a
R
i=1
ni
wi
)×n.
tensorsketch
Construct
∈
(mQq
Run Condition(A) using S to compute U/(dγp).
Generate a Gaussian matrix G ∈ Rd×O(log(n))
for for each column e in U G do
and a Cauchy sketch matrix R ∈ Rlog(n)×n.
S
end for
Reshape Ae to M (Eq. 3).
Compute λi and λe =Pi λi (Eq 4 and 5).
for i ∈ [pQq
Sample a column (AU G)∗,e with probabil-
⊲ e ∈ [O(log(n))]
Reshape (AU G)∗,e to M , sample a column
i=1 wi poly(d)] do
ity proportional to λe. .
Sample an entry Mk,j with probability pro-
M∗,j with probability proportional to λj.
portional to Mk,j.
row index in A, denoted as ri.
end for
Convert (k, j) back to the corresponding
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
where D is a diagonal matrix to select the
r1, r2, ..., rN -th rows of A and b with N =
ex ← minx kD(A1 ⊗ A2 ⊗ ··· ⊗ Aq)x − Dbk1,
pQq
⊲ex ∈ Rd1d2···dq
return ex
i=1 wi poly(d).
16:
17: end procedure
l=1
l}O(log n)
for l ∈ [O(log(n))], where Rl,∗ is the l'th row of R.
Due to the 1-stability property of the Cauchy distri-
bution, we have that {zi
are O(log n) inde-
pendent Cauchys scaled by kMik1. Applying a Cher-
noff bound to independent half-Cauchys (see Claims
1, 2 and Lemmas 1, 2 in Indyk (2006)), we have
0.5kMik1 ≤ medianl∈[O(log n)]{zi
j} ≤ 1.5kMik1 with
probability at least 1 − 2e−cO(log(n)), with constant
l}O(log(n))
c ≥ 0.07. Denote the median of {zi
as λi.
By a union bound over all columns of M , we have that
with probability at least 1 − n[q]\[q1]2e−cO(log(n)):
l=1
(4)
(5)
λi = (1 ± 0.5)kMik1,∀i ∈ [n[q]\[q1]],
λi = (1 ± 0.5)kAek1
λe =
n[q]\[q1]Xi=1
where n[q]\[q1] =Qq
i=q1+1 ni.
Note that since AU G only has O(log n) columns, we
can afford to compute the ℓ1 norm of all the columns
using the above procedure. Let us denote the ℓ1 norms
of the columns of AU G by λ1, λ2, ..., λO(log n). We can
sample a column (AU G)∗,i with probability propor-
tional to λi (Line 10 in Alg. 3). Once we sample a
column AU G∗,i, we need to sample an entry j ∈ [n]
Sketching for Kronecker Product Regression and P-splines
with probability proportional to the absolute value of
the entry (AU G)j,i. As we cannot afford to compute
(AU G)j,i for all j ∈ [n], we use the reshaped 2-d ma-
trix M of (AU G)∗,i. Note that sampling an entry in M
with probability proportional to the absolute values of
entries of M is equivalent to sampling an entry j ∈ [n]
from (AU G)∗,i with probability proportional to the ab-
solute value of the entries of (AU G)∗,i. We first sample
a column M∗,j from all the columns of M with proba-
i=q1+1 ni]. We then
sample an entry Mk,j with probability proportional to
i=1 ni]
i=q1+1 ni], the pair (k, j) uniquely deter-
mines a corresponding row index r in A, for some
i=1 ni]. Hence we successfully sample a row
from AU G without ever computing the ℓp norm of the
rows. The above sampling procedure is summarized
in Line 10 to Line 13 in Alg. 3. We use the above
i=1 wi poly(d) rows of A. Let
D be a diagonal matrix that selects the corresponding
sampled rows from A. We can now solve a smaller
ADL problem as minx kDA − Dbk1. Note that our
analysis focuses on the ℓ1 norm. We can extend the
analysis to general ℓp norms by using a sketching ma-
i=1 ni with entries sampled i.i.d.
bility proportional to λj for j ∈ [Qq
Mk,j for k ∈ [Qq1
and j ∈ [Qq
r ∈ [Qq
procedure to samplepQq
i=1 ni]. Noting that k ∈ [Qq1
trix R ∈ RO(log n)×Qq1
from a p-stable distribution for p ∈ [1, 2].
We now present our main theorem and defer the proofs
to the appendix:
Theorem 5.4. (Main result) Given ǫ ∈ (0, 1), A ∈
with probability at least 1/2, kAx − bk1 ≤ (1 +
ǫ) minx∈Rd kAx − bk1. For the special case when
q = 2, n1 = n2, the algorithm's running time is
O(n1
Rn×d and b ∈ Rn, Alg. 3 computes bx such that
i=1 di/ǫ)).
3/2 poly(Q2
For the special case where q = 2 and n1 = n2, we
3/2 poly( d)), which is faster than O(n1
can see from theorem 5.4 our algorithm computesbx in
2)-
time O(n1
the time needed for forming A1 ⊗ A2. Note that we
can run Alg. 3 O(log(1/δ)) times independently and
pick the best solution among these independent runs to
boost the success probability to be 1 − δ, for δ ∈ [0, 1).
6 NUMERICAL EXPERIMENTS
We generate matrices A1, A2 and b with all entries
sampled i.i.d from a normal distribution. The base-
line we compared to is directly solving regression with-
out sketching. We let T1 be the time for directly
solving the regression problem, and T2 be the time
of our algorithm. The time ratio is rt = T2/T1.
The relative residual percentage is defined by re =
100 k(A1⊗A2)ex−bk2−k(A1⊗A2)x∗−bkp
put of our algorithms and x∗ is the optimal solution.
k(A1⊗A2)x∗−bkp
, whereex is the out-
Table 1: Examples 6.1 and 6.2: the values of re and rt
with respect to different sampling parameters m.
m
8000
12000
16000
re
rt
1.79% 0.11
1.24% 0.18
1.01% 0.25
m
8000
12000
16000
re
rt
1.89% 0.06
1.33% 0.11
0.992% 0.18
ℓ1
ℓ2
Throughout the simulations, we use a moderate input
matrix size in order to accommodate the brute force
algorithm and to compare to the exact solution.
Example 6.1 (ℓ2 Regression). We create a design ma-
trix with moderate size by fixing n1 = n2 = 300 and
d1 = d2 = 15. Thus A ∈ R90000×225. We do 10 rounds
to compute the mean values of re and rt, which is re-
ported in Table 1 (Left).
From Table 1 (left), we can see that when the number
m of sampled rows in Algorithm 1 increases from 8000
to 12000, the mean values of re decrease while the
mean values of rt increase. In general we can see that
we can achieve around 1% relative error while being 5
times faster than the direct method.
Example 6.2 (ℓ1 Regression). We set n1 = n2 =
300, d1 = d2 = 15.
For minx kAx − bk1, we
solve it by a Linear Programming solver in Gurobi
(Gurobi Optimization, 2016). We tested different num-
bers of sampled rows m (the number of rows in D in
Alg. 3). The results are summarized in Table 1 (Right).
As we can see from Table 1 (right), our method is
around 10 times faster than directly solving the prob-
lem, with relative error only around 1%.
Example 6.3 (P-Spline Regression). For P-spline re-
gression, L3 is fixed. We use 30 knots and cubic B-
splines. The data u = (ui) ∈ Rn1, v = (vj ) ∈ Rn2
and b ∈ Rn1n2 are generated i.i.d from the normal
distribution. We compute the B-spline basis matrices
A1 ∈ Rn1×d1 and A2 ∈ Rn2×d2 separately, if there are
d1 and d2 basis functions for the B-spline. We set
n1 = n2, d1 = d2, with n1n2 = 104 and d1d2 = 529.
The original and sketched P-spline regression problem
are both solved by Regularization Tools (Hansen,
1994) via computing their GSVDs. We test different
choices of λ. The results are shown in Table 2.
From Table 2, sampling only 20% of the rows can give
around 0.05% relative error, while the computation
time is half of the time of directly solving the p-spline.
7 CONCLUSION
We propose algorithms for efficiently solving tensor
product least squares regression, regularized P-splines,
as well as tensor product least absolute deviation re-
gression, using sketching techniques. Our main con-
tributions are: (1) we apply tensorsketch to least
Diao, Song, Sun, Woodruff
Table 2: Example 6.3: the mean values of re and rt
with respect to different sampling parameters m and
regularization parameters λ.
λ
1
0.1
0.01
m
2000
4000
6000
2000
4000
6000
2000
4000
6000
re
rt
4.43e-2% 0.52
2.99e-2% 0.70
7.92e-2% 0.91
6.98e-2% 0.47
4.07e-2% 0.72
5.41e-2% 0.94
3.78e-2% 0.46
1.07e-1% 0.71
2.97e-2% 0.95
square regression problems, (2) we propose new statis-
tical dimension measures for P-splines, extending the
previous statistical dimension defined only for classic
Ridge regression, and (3) we extend tensorsketch
to ℓp norms and propose an algorithm that can solve
tensor product ℓ1 regression in time sublinear in the
time for explicitly computing the tensor product. Sim-
ulation results support our theorems and demonstrate
that our algorithms are much faster than brute-force
algorithms and can achieve approximate solutions that
are close to optimal.
ACKNOWLEDGEMENT
Huaian Diao is supported in part by the Fundamental
Research Funds for the Central Universities under the
grant 2412017FZ007. Wen Sun is supported in part by
Office of Naval Research contract N000141512365.
References
H. Avron, H. Nguyen, and D. Woodruff. Subspace em-
In Advances
beddings for the polynomial kernel.
in Neural Information Processing Systems(NIPS),
pages 2258–2266, 2014.
H. Avron, K. L. Clarkson, and D. P. Woodruff.
Sharper bounds for regression and low-rank approxi-
mation with regularization. CoRR, abs/1611.03225,
2016.
L. Carter and M. N. Wegman. Universal classes of
hash functions. J. Comput. Syst. Sci., 18(2):143–
154, 1979.
M. Charikar, K. Chen, and M. Farach-Colton. Finding
frequent items in data streams. Theor. Comput. Sci.,
312(1):3–15, 2004.
D. Chen and R. J. Plemmons. Nonnegativity con-
straints in numerical analysis. In The birth of nu-
merical analysis, pages 109–139. World Sci. Publ.,
Hackensack, NJ, 2010.
K. Clarkson, P. Drineas, M. Magdon-Ismail, M. Ma-
honey, X. Meng, and D. P. Woodruff. The fast
Cauchy transform and faster robust linear regres-
sion. In SODA, 2013.
K. L. Clarkson. Subgradient and sampling algorithms
for ℓ1 regression. In Proceedings of the sixteenth an-
nual ACM-SIAM symposium on Discrete algorithms
(SODA), pages 257–266, 2005.
K. L. Clarkson and D. P. Woodruff. Low rank ap-
proximation and regression in input sparsity time.
In Symposium on Theory of Computing Conference,
STOC'13, Palo Alto, CA, USA, June 1-4, 2013,
pages 81–90. https://arxiv.org/pdf/1207.6365,
2013.
A. Dasgupta, P. Drineas, B. Harb, R. Kumar, and
M. W. Mahoney. Sampling algorithms and coresets
for ℓp regression. SIAM J. Comput., 38(5):2060–
2078, 2009.
P. Drineas, M. Magdon-Ismail, M. W. Mahoney, and
D. P. Woodruff. Fast approximation of matrix coher-
ence and statistical leverage. CoRR, abs/1109.3843,
2011.
P. H. Eilers and B. D. Marx. Multidimensional den-
sity smoothing with p-splines. In Proceedings of the
21st international workshop on statistical modelling,
2006.
P. H. C. Eilers and B. D. Marx. Flexible smoothing
with B-splines and penalties. Statist. Sci., 11(2):
89–121, 1996.
P. H. C. Eilers, B. D. Marx, and M. Durb´an. Twenty
years of P-splines. SORT, 39(2):149–186, 2015. ISSN
1696-2281.
D. W. Fausett and C. T. Fulton. Large least squares
problems involving Kronecker products. SIAM J.
Matrix Anal. Appl., 15(1):219–227, 1994.
G. H. Golub and C. F. Van Loan. Matrix compu-
tations. Johns Hopkins Studies in the Mathemati-
cal Sciences. Johns Hopkins University Press, Balti-
more, MD, 2013.
I. Gurobi Optimization. Gurobi optimizer reference
manual, 2016. URL http://www.gurobi.com.
P. C. Hansen. Regularization tools: A matlab package
for analysis and solution of discrete ill-posed prob-
lems. Numerical Algorithms, 6(1):1–35, 1994.
P. Indyk. Stable distributions, pseudorandom genera-
tors, embeddings, and data stream computation. J.
ACM, 53(3):307–323, 2006.
R. Kannan and S. Vempala.
Spectral algorithms.
Foundations and Trends in Theoretical Computer
Science, 4(3-4):157–288, 2009.
Sketching for Kronecker Product Regression and P-splines
C. F. Van Loan and N. Pitsianis. Approximation with
Kronecker products. In Linear algebra for large scale
and real-time applications (Leuven, 1992), volume
232 of NATO Adv. Sci. Inst. Ser. E Appl. Sci., pages
293–314. Kluwer Acad. Publ., Dordrecht, 1993.
D. P. Woodruff. Sketching as a tool for numerical lin-
ear algebra. Foundations and Trends in Theoretical
Computer Science, 10(1-2):1–157, 2014.
Y. Zheng, G. Liu, S. Sugimoto, S. Yan, and M. Oku-
tomi. Practical low-rank matrix approximation un-
der robust l 1-norm. In Computer Vision and Pat-
tern Recognition (CVPR), 2012 IEEE Conference
on, pages 1410–1417. IEEE, 2012.
Y. Liang, M.-F. F. Balcan, V. Kanchanapally, and
D. Woodruff. Improved distributed principal com-
ponent analysis. In Advances in Neural Information
Processing Systems, pages 3113–3121, 2014.
M. W. Mahoney. Randomized algorithms for matri-
ces and data. Foundations and Trends in Machine
Learning, 3(2):123–224, 2011.
X. Meng and M. W. Mahoney. Low-distortion sub-
space embeddings in input-sparsity time and ap-
plications to robust linear regression.
In Pro-
ceedings of the forty-fifth annual ACM symposium
on Theory of computing, pages 91–100. ACM,
https://arxiv.org/pdf/1210.3135, 2013.
J. Nelson and H. L. Nguyen.
Osnap: Faster
linear algebra algorithms via sparser
numerical
subspace
embeddings.
In 2013 IEEE 54th
Annual Symposium on Foundations of Com-
puter Science (FOCS), pages 117–126.
IEEE,
https://arxiv.org/pdf/1211.1002, 2013.
S. Oh, S. Kwon and J. Yun. A method for structured
linear total least norm on blind deconvolution prob-
lem. Applied Mathematics and Computing, 19:151–
164, 2005.
R. Pagh. Compressed matrix multiplication. ACM
Trans. Comput. Theory, 5(3):9:1–9:17, 2013.
M. Patrascu and M. Thorup. The power of simple
tabulation hashing. J. ACM, 59(3):14, 2012.
N. Pham and R. Pagh. Fast and scalable polynomial
kernels via explicit feature maps. In Proceedings of
the 19th ACM SIGKDD international conference on
Knowledge discovery and data mining(KDD), pages
239–247. ACM, 2013.
P. J. Rousseeuw and A. M. Leroy. Robust regression
and outlier detection, volume 589. John wiley & sons,
2005.
C. Sohler and D. P. Woodruff. Subspace embeddings
for the ℓ1-norm with applications. In Proceedings of
the forty-third annual ACM symposium on Theory
of computing (STOC), pages 755–764. ACM, 2011.
Z. Song, D. P. Woodruff, and P. Zhong.
Low
rank approximation with entrywise ℓ1-norm er-
ror.
In Proceedings of the 49th Annual Sympo-
sium on the Theory of Computing (STOC). ACM,
https://arxiv.org/pdf/1611.00898, 2017a.
Z. Song, D. P. Woodruff, and P. Zhong. Relative er-
ror tensor low rank approximation. arXiv preprint
arXiv:1704.08246, 2017b.
C. Van Loan. Computational frameworks for the fast
Fourier transform, volume 10 of Frontiers in Ap-
plied Mathematics. Society for Industrial and Ap-
plied Mathematics (SIAM), Philadelphia, PA, 1992.
A Background: CountSketch and TensorSketch
Diao, Song, Sun, Woodruff
i of the output, i = 1, 2, . . . , m is Pjh(j)=i s(j)vj. Note that CountSketch can be represented as an m × n
We start by describing the CountSketch transform Charikar et al. (2004). Let m be the target dimension.
When applied to n-dimensional vectors, the transform is specified by a 2-wise independent hash function h :
[n] → [m] and a 2-wise independent sign function s : [n] → {−1, +1}. When applied to v, the value at coordinate
matrix in which the j-th column contains a single non-zero entry s(j) in the h(j)-th row.
We now describe the TensorSketch transform Pagh (2013). Suppose we are given points vi ∈ Rni , where
i = 1, . . . , q and so φ(v1, . . . , vq) = v1 ⊗ v2 ⊗ ··· ⊗ vq ∈ Rn1n2···nq , and the target dimension is again m. The
transform is specified using q 3-wise independent hash functions hi : [ni] → [m], and q 4-wise independent sign
functions si : [ni] → {+1,−1}, where i = 1, . . . , q. TensorSketch applied to v1, . . .⊗ vq is then CountSketch
applied to φ(v1, . . . , vq) with hash function H : [n1n2 ··· nq] → [m] and sign function S : [n1n2 ··· nq] → {+1,−1}
defined as follows:
and
H(i1, . . . , iq) = h1(i1) + h2(i2) + ··· + hq(iq) mod m,
S(i1, . . . , iq) = s1(i1) · s2(i2)··· sq(iq),
is 3-wise in-
where ij ∈ [nj].
dependent Carter and Wegman (1979); Patrascu and Thorup (2012).
Unlike the work of Pham and
Pagh Pham and Pagh (2013), which only used that H was 2-wise independent, our analysis needs this stronger
property of H.
if H is constructed this way,
is well-known that
then it
It
The TensorSketch transform can be applied to v1, . . . , vq without computing φ(v1, . . . , vq) as follows. Let
vj = (vjℓ ) ∈ Rnj . First, compute the polynomials
pℓ(x) =
B−1Xi=0
xi Xjℓhℓ(jℓ)=i
vjℓ · sℓ(jℓ),
for ℓ = 1, 2, . . . , q. A calculation Pagh (2013) shows
qYℓ=1
pℓ(x) mod (xB − 1) =
xi
B−1Xi=0
X
vj1 ··· vjq S(j1, . . . , jq),
(j1,...,jq)H(j1,...,jq)=i
the q polynomials mod (xm − 1)
is,
the coefficients of
that
form the value of
TensorSketch(v1, . . . , vq). Pagh observed that this product of polynomials can be computed in O(qm log m)
time using the Fast Fourier Transform. As it takes O(q max(nnz(vi))) time to form the q polynomials, the overall
time to compute TensorSketch(v) is O(q(max(nnz(vi)) + m log m)).
the product of
B TensorSketch is an Oblivious Subspace Embedding (OSE)
Let S be the m × (n1n2 ··· nq) matrix such that TensorSketch (v1, . . . , vq) is S · φ(v1, . . . , vq) for a randomly
selected TensorSketch. Notice that S is a random matrix. In the rest of the paper, we refer to such a matrix
as a TensorSketch matrix with an appropriate number of rows, i.e., the number of hash buckets. We will
show that S is an oblivious subspace embedding for subspaces in Rn1n2···nq for appropriate values of m. Notice
that S has exactly one non-zero entry per column. The index of the non-zero in the column (i1, . . . , iq) is
j=1 hj(ij) mod m. Let δa,b be the indicator random variable of whether Sa,b is non-zero. The
j=1 sj(ij). We show that the embedding
matrix S of TensorSketch can be used to approximate matrix product and is an oblivious subspace embedding
(OSE).
Theorem B.1. Let S be the m × (n1n2 ··· nq) matrix such that
H(i1, . . . , iq) =Pq
sign of the non-zero entry in column (i1, . . . , iq) is S(i1, . . . , iq) = Qq
TensorSketch(v1, . . . , vq)
is S · φ(v1, . . . , vq) for a randomly selected TensorSketch. The matrix S satisfies the following two properties.
Sketching for Kronecker Product Regression and P-splines
1. (Approximate Matrix Product :) Let A and B be matrices with n1n2 ··· nq rows. For m ≥ (2 + 3q)/(ǫ2δ),
we have
Pr
S (cid:2)kA⊤S⊤SB − A⊤Bk2
F ≤ ǫ2kAk2
FkBk2
F(cid:3) ≥ 1 − δ.
2. (Subspace Embedding :) Consider a fixed k-dimensional subspace V . If m ≥ k2(2 + 3q)/(ǫ2δ), then with
probability at least 1 − δ, kSxk = (1 ± ǫ)kxk simultaneously for all x ∈ V .
We establish the theorem via two lemmas as in Avron et al. (2016). The first lemma proves the approximate
matrix product property via a careful second moment analysis.
Lemma B.2 (Approximate matrix product). Let A and B be matrices with n1n2 ··· nq rows. For m ≥ (2 +
3q)/(ǫ2δ), we have
Pr
S (cid:2)kA⊤S⊤SB − A⊤Bk2
F ≤ ǫ2kAk2
FkBk2
F(cid:3) ≥ 1 − δ.
Proof. The proof follows that in Avron et al. (2016). Let C = A⊤S⊤SB. We have
Cu,u′ =
mXt=1 Xi,j∈[n1n2···nq]
S(i)S(j)δt,iδt,jAi,uBj,u′ =
mXt=1 Xi6=j∈[n1n2···nq]
S(i)S(j)δt,iδt,iAi,uBj,u′ + (A⊤B)u,u′
Thus, E[Cu,u′ ] = (A⊤B)u,u′ .
Next, we analyze E[((C − A⊤B)u,u′ )2]. We have
((C − A⊤B)u,u′ )2 =
mXt1,t2=1
X
i16=j1,i26=j2∈[n1n2···nq]
S(i1)S(i2)S(j1)S(j2) · δt1,i1 δt1,j1 δt2,i2δt2,j2 · Ai1,uAi2,uBj1,u′Bj2,u′
For a term in the summation on the right hand side to have a non-zero expectation, it must be the case
that E[S(i1)S(i2)S(j1)S(j2)] 6= 0. Note that S(i1)S(i2)S(j1)S(j2) is a product of random signs (possibly with
multiplicities) where the random signs in different coordinates in {1, . . . , q} are independent and they are 4-wise
independent within each coordinate. Thus, E[S(i1)S(i2)S(j1)S(j2)] is either 1 or 0. For the expectation to be
1, all random signs must appear with even multiplicities. In other words, in each of the q coordinates, the 4
coordinates of i1, i2, j1, j2 must be the same number appearing 4 times or 2 distinct numbers, each appearing
twice. All the subsequent claims in the proof regarding i1, i2, j1, j2 agreeing on some coordinates follow from this
property.
Let S1 be the set of coordinates where i1 and i2 agree. Note that j1 and j2 must also agree in all coordinates
in S1 by the above argument. Let S2 ⊂ [q] \ S1 be the coordinates among the remaining where i1 and j1
agree. Finally, let S3 = [q] \ (S1 ∪ S2). All coordinates in S3 of i1 and j2 must agree. Similarly as before,
note that i2 and j2 agree on all coordinates in S2 and i2 and j1 agree on all coordinates in S3. We can rewrite
i1 = (a, b, c), i2 = (a, e, f ), j1 = (g, b, f ), j2 = (g, e, c) where a = (aℓ), g = (gℓ) with ℓ ∈ S1, b = (bℓ), e = (eℓ) with
ℓ ∈ S2 and c = (cℓ), f = (fℓ) with ℓ ∈ S3.
First we show that the contribution of the terms where i1 = i2 or i1 = j2 is bounded by 2kAuk2
, where Au
is the uth column of A and Bu′ is the u′th column of B. Indeed, consider the case i1 = i2. As observed before,
we must have j1 = j2 to get a non-zero contribution. Note that if t1 6= t2, we always have δt1,i1δt2,i2 = 0 as H(i1)
cannot be equal to both t1 and t2. Thus, for fixed i1 = i2, j1 = j2,
2kBu′k2
m
2
S(i1)S(i2)S(j1)S(j2) · δt1,i1 δt1,j1 δt2,i2 δt2,j2 · Ai1,uAi2,uBj1,u′ Bj2,u′#
E" mXt1,t2=1
= E" mXt1=1
=
A2
i1,uB2
m
j1,u′
i1,t1δ2
δ2
j1,t1A2
i1,uB2
j1,u′#
Diao, Song, Sun, Woodruff
Summing over all possible values of i1, j1, we get the desired bound of kAuk2
. The case i1 = j2 is analogous.
Next we compute the contribution of the terms where i1 6= i2, j1, j2 i.e., there are at least 3 distinct numbers
among i1, i2, j1, j2. Notice that E[δt1,i1δt1,j1δt2,i2 δt2,j2 ] ≤ 1
m3 because the δt,i's are 3-wise independent. For fixed
i1, j1, i2, j2, there are m2 choices of t1, t2 so the total contribution to the expectation from terms with the same
i1, j1, i2, j2 is bounded by m2 · 1
Therefore,
m3 · Ai1,uAi2,uBj1,u′ Bj2,u′ = 1
mAi1,uAi2,uBj1,u′Bj2,u′.
2kBu′k2
m
2
A(a,b,c),uB(g,b,f ),u′A(a,e,f ),uB(g,e,c),u′
2
E[((C − A⊤B)u,u′ )2]
2kBu′k2
2kAuk2
1
m
2kBu′k2
2kAuk2
m
3q
+
+
2
2kBu′k2
2kAuk2
m
2
2kBu′k2
2kAuk2
m
2
2kBu′k2
2kAuk2
m
2
2kBu′k2
2kAuk2
m
2
2kBu′k2
2kAuk2
m
2
+
+
+
+
+
(2 + 3q)kAuk2
2kBu′k2
2
m
,
≤
≤
≤
=
≤
=
≤
=
3q
m Xpartition S1,S2,S3 Xa,g,b,e,c,f
m Xa,b,c,g,e,f
m Xg,e,f(cid:18)Xa,b,c
m Xg,e,f(cid:18)Xb
m Xe (cid:18)Xb,g,f
3qkAuk
3qkAuk
B2
A2
(a,b,c),u(cid:19)1/2(cid:18)Xa,b,c
3qkAuk · kBu′k
m
3qkAuk · kBu′k
m
A(a,b,c),uB(g,b,f ),u′A(a,e,f ),uB(g,e,c),u′
B2
(g,b,f ),u′ A2
(a,e,f ),uB2
(g,e,c),u′(cid:19)1/2
(g,e,c),u′(cid:19)1/2
(g,e,c),u′(cid:19)1/2
(a,e,f ),uB2
(a,e,f ),uB2
A2
A2
B2
(g,b,f ),u′(cid:19)1/2(cid:18)Xa,c
(g,b,f ),u′(cid:19)1/2(cid:18) Xa,c,g,f
Xe (cid:18)Xa,f
(a,e,f ),u(cid:19)1/2(cid:18)Xg,c
(cid:18)Xa,e,f
(a,e,f ),u(cid:19)1/2(cid:18)Xg,e,c
A2
A2
B2
B2
(g,e,c),u′(cid:19)1/2
(g,e,c),u′(cid:19)1/2
where the second inequality follows from the fact that there are at most 3q partitions of [q] into 3 sets. The other
inequalities are from Cauchy-Schwarz.
Combining the above bounds, we have E[((C − A⊤B)u,u′ )2] ≤ (2+3q)kAuk2
Markov inequality, kA⊤S⊤SB − A⊤Bk2
2kBu′k2
F with probability 1 − δ.
F ≤ ǫ2kAk2
FkBk2
m
2
. For m ≥ (2 + 3q)/(ǫ2δ), by the
The second lemma proves that the subspace embedding property follows from the approximate matrix product
property.
Lemma B.3 (Oblivious subspace embeddings). Consider a fixed k-dimensional subspace V ⊂ Rn1n2···nq . If
m ≥ k2(2 + 3q)/(ǫ2δ), then with probability at least 1 − δ, kSxk2 = (1 ± ǫ)kxk2 simultaneously for all x ∈ V .
Proof. Let B be a (n1n2 ··· nq) × k matrix whose columns form an orthonormal basis of V . Thus, we have
B⊤B = Ik and kBk2
F = k. The condition that kSxk2 = (1 ± ǫ)kxk2 simultaneously for all x ∈ V is equivalent to
the condition that the singular values of SB are bounded by 1 ± ǫ. By Lemma B.2, for m ≥ (2 + 3q)/((ǫ/k)2δ),
with probability at least 1 − δ, we have
kB⊤S⊤SB − B⊤Bk2
F ≤ (ǫ/k)2kBk4
F = ǫ2
Thus, we have kB⊤S⊤SB − Ikk2 ≤ kB⊤S⊤SB − IkkF ≤ ǫ. In other words, the squared singular values of SB
are bounded by 1 ± ǫ, implying that the singular values of SB are also bounded by 1 ± ǫ. Note that kAk2 for a
matrix A denotes its operator norm.
Sketching for Kronecker Product Regression and P-splines
C Missing Proofs
C.1 Proofs for Tensor Product Least Square Regression
Theorem 3.1. (Tensor regression) Suppose ex is the output of Algorithm 1 with tensorsketch S ∈ Rm×n,
where m = 8(d1d2 ··· dq + 1)2(2 + 3q)/(ǫ2δ). Then the following approximation k(A1 ⊗ A2 ⊗ ··· ⊗ Aq)ex − bk2 ≤
(1 + ǫ) OPT, holds with probability at least 1 − δ.
Proof. It is easy to see that
k(A1 ⊗ A2 ⊗ ··· ⊗ Aq)x − bk2 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:2)(A1 ⊗ A2 ⊗ ··· ⊗ Aq)
b(cid:3)(cid:20) x
−1(cid:21) ∈ Rn1n2···nq
y =(cid:2)(A1 ⊗ A2 ⊗ ··· ⊗ Aq)
,
−1(cid:21)(cid:13)(cid:13)(cid:13)(cid:13)2
b(cid:3)(cid:20) x
and identifying
and y is a vector of a subspace V ⊂ Rn1n2···nq with dimension at most d1d2 ··· dq + 1, we can use Lemma B.3 to
conclude that
when m = (d1d2 ··· dq + 1)2(2 + 3q)/(ǫ2δ).
Thus we have
Pr [kSyk2 − kyk2 ≤ ǫkyk2] ≥ 1 − δ
k(A1 ⊗ A2 ⊗ ··· ⊗ Aq)ex − bk2 ≤
1
1 − ǫkS(A1 ⊗ A2 ⊗ ··· ⊗ Aq)ex − Sbk2
and
kS(A1 ⊗ A2 ⊗ ··· ⊗ Aq)x − Sbk2 ≤ (1 + ǫ)k(A1 ⊗ A2 ⊗ ··· ⊗ Aq)x − bk2
hold with probability at least 1 − δ. Then using a union bound, we have
1
1
k(A1 ⊗ A2 ⊗ ··· ⊗ Aq)ex − bk2
1 − ǫkS(A1 ⊗ A2 ⊗ ··· ⊗ Aq)ex − Sbk2
1 − ǫkS(A1 ⊗ A2 ⊗ ··· ⊗ Aq)x − Sbk2
1 − ǫk(A1 ⊗ A2 ⊗ ··· ⊗ Aq)x − bk2
1 + ǫ
≤
≤
≤
holds with probability at least 1 − 2δ.
Corollary 3.2. (Sketch for tensor nonnegative regression) Suppose x = minx≥0 kSAx − Sbk2 with tensors-
ketch S ∈ Rm×n, where m = 8(d1d2 ··· dq +1)2(2+3q)/(ǫ2δ). Then the following approximation k(A1⊗A2⊗···⊗
Aq)ex−bk2 ≤ (1+ǫ) OPT holds with probability at least 1−δ, where OPT = minx≥0 k(A1 ⊗ A2 ⊗ ··· ⊗ Aq)x − bk2.
Proof. The proof of Theorem. 3.2 is similar to the proof of theorem 3.1. Denote x = minx≥0 kSAx − Sbk2 and
x∗ = minx≥0 kAx − bk2. Using Lemma. B.3, we have:
with probability at least 1 − δ, and
kAx − bk2 ≤
1
1 − ǫkSAx − Sbk2,
kSAx∗ − Sbk2 ≤ (1 + ǫ)kAx∗ − bk2,
(6)
(7)
Diao, Song, Sun, Woodruff
with probability at least 1 − δ. Hence applying a union bound we have:
≤
≤
≤
with probability at least 1 − 2δ.
C.2 Proofs for P-Splines
1
1
kAx − bk2
1 − ǫkSAx − Sbk2
1 − ǫkSAx∗ − Sbk2
1 − ǫkAx∗ − bk2,
1 + ǫ
(8)
(9)
Lemma 4.1. Let x∗ ∈ Rd, A ∈ Rn×d and b ∈ Rn as above. Let U1 ∈ Rn×d denote the first n rows of an
orthogonal basis for h A√λLi ∈ R(n+p)×d. Let sketching matrix S ∈ Rm×n have a distribution such that with
constant probability
and
(I) kU⊤1 S⊤SU1 − U⊤1 U1k2 ≤ 1/4,
(II) kU⊤1 (S⊤S − I)(b − Ax∗)k2 ≤pǫ OPT /2.
2 + λkLxk2
2. Then with probability at least 9/10,
Let ex denote argminx∈RdkS(Ax − b)k2
Proof. Let A ∈ R(n+d)×d have orthonormal columns with range( A) = range(h A√λLi). (An explicit expression for
one such A is given below.) Let b ≡(cid:2) b
kAex − bk2
0d(cid:3). We have
2 + λkLexk2
2 ≤ (1 + ǫ) OPT .
(10)
min
y∈Rdk Ay − bk2
2, , in the sense that for any Ay ∈ range( A), there is x ∈ Rd with Ay =h A√λLi x,
√λLx∗i.
2. Let y∗ = argminy∈Rdk Ay−bk2, so that Ay∗ =h Ax∗
2 = kb−Axk2
2+λkLxk2
2 + λkLxk2
2 = kh A√λLi x−bk2
equivalent to kb − Axk2
so that k Ay−bk2
Let A =(cid:2) U1
We define S to beh S 0m×d
U2(cid:3), where U1 ∈ Rn×d and U2 ∈ Rd×d, so that U1 is as in the lemma statement.
Id i and S satisfies Property (I) and (II) of Lemma 4.1.
0d×n
Using kU⊤1 S⊤SU1 − U⊤1 U1k2 ≤ 1/4, with constant probability
k A⊤ S⊤ S A − Idk2 = kU⊤1 S⊤SU1 + U⊤2 U2 − Idk2 = kU⊤1 S⊤SU1 − U⊤1 U1k2 ≤ 1/4.
(11)
Using the normal equations for Eq. (10), we have
0 = A⊤(b − Ay∗) = U⊤1 (b − Ax∗) −
√λU⊤2 x∗,
and so
A⊤ S⊤ S(b − Ay∗) = U⊤1 S⊤S(b − Ax∗) −
√λU⊤2 x∗ = U⊤1 S⊤S(b − Ax∗) − U⊤1 (b − Ax∗).
Using Property (II) of Lemma 4.1, with constant probability
k A⊤ S⊤ S(b − Ay∗)k2
= kU⊤1 S⊤S(b − Ax∗) − U⊤1 (b − Ax∗)k2
≤pǫ OPT /2
=pǫ/2kb − Ay∗k2.
(12)
Sketching for Kronecker Product Regression and P-splines
It follows by a standard result from (11) and (12) that the solution y ≡ argminy∈Rdk S( Ay−b)k2 has k Ay−bk2 ≤
(1 + ǫ) miny∈Rdk Ay − bk2, and therefore that x satisfies the claim of the theorem.
For convenience we give the proof of the standard result: (11) implies that A⊤ S⊤ S A has smallest singular value
at least 3/4. The normal equations for the unsketched and sketched problems are
A⊤(b − Ay∗) = 0 = A⊤ S⊤ S(b − Ay).
The normal equations for the unsketched case imply k Ay − bk2
show that k A(y − y∗)k2
2 ≤ ǫ OPT. We have
2 = ky − y∗k2
2 = k A(y − y∗)k2
2 + kb − Ay∗k2
2, so it is enough to
(3/4)ky − y∗k2 ≤ k A⊤ S⊤ S A(y − y∗)k2
= k A⊤ S⊤ S A(y − y∗) − A⊤ S⊤ S(b − Ay)k2
= k A⊤ S⊤ S(b − Ay∗)k2
≤pǫ OPT /2
so that ky − y∗k2
2 ≤ (4/3)2ǫ OPT /2 ≤ ǫ OPT. The lemma follows.
by Eq. (11)
by Normal Equation
by Eq. (12),
The following lemma computes the statistical dimension sdλ(A, L) that will be used for computing the number
of rows of sketching matrix S.
Lemma C.1. For U1 as in Lemma 4.1, kU1k2
values σi. Also kU1k2 = max{1/p1 + λ/γ2
1 , 1}.
F = sdλ(A, L) =Pi 1/(1 + λ/γ2
i ) + d − p, where A has singular
Proof. Suppose we have the GSVD of (A, L). Let
Then
.
0(n−p)×p
D ≡(cid:20)Σ⊤Σ + λΩ⊤Ω 0p×(n−p)
A =
U
0p×(n−p)
D
0(n−p)×p
Id−p
√λVhΩ 0p×(n−p)iD
Id−p (cid:21)−1/2
Σ
2
F
Σ
Σ
0p×(n−p)
0p×(n−p)
0(n−p)×p
0(n−p)×p
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:20)
Id−p (cid:21) D(cid:13)(cid:13)(cid:13)(cid:13)
has A⊤ A = Id, and for given x, there is y = D−1RQ⊤x with Ay = h A√λLi x. We have kU1k2
(cid:13)(cid:13)(cid:13)(cid:13)U(cid:20)
Theorem 4.3. (P-Spline regression) There is a constant K > 0 such that for m ≥ K(ǫ−1 sdλ(A, L)+sdλ(A, L)2)
and S ∈ Rm×n a sparse embedding matrix (e.g., Countsketch) with SA computable in O(nnz(A)) time, Prop-
erty (I) and (II) of Lemma 4.1 apply, and with constant probability the correspondingex = argminx∈RdkS(Ax −
b)k2 + λkLxk2
Proof. Recall that sdλ(A, L) = kU1k2
multiplication
2 is an ǫ-approximate solution to minx∈Rdkb − Axk2
F . Sparse embedding distributions satisfy the bound for approximate matrix
Id−p (cid:21) D(cid:13)(cid:13)(cid:13)(cid:13)
i ) + d − p as claimed.
2 + λkLxk2
2.
=Pp
i=1 1/(1 + λ/γ2
F =
2
F
kW ⊤S⊤SH − W ⊤HkF ≤ CkWkFkHkF /√m,
for a constant C (Clarkson and Woodruff, 2013; Meng and Mahoney, 2013; Nelson and Nguyen, 2013); this is
also true of OSE matrices. We set W = H = U1 and use kXk2 ≤ kXkF for all X and m ≥ KkU1k4
F to obtain
Property (I) of Lemma 4.1, and set W = U1, H = b − Ax∗ and use m ≥ KkU1k2
F /ǫ to obtain Property (II) of
Lemma 4.1. (Here the bound is slightly stronger than Property (II), holding for λ = 0.) With Property (I) and
Property (II), the claim for x from a sparse embedding follows using Lemma 4.1.
C.3 Proofs for Tensor Product ℓ1 Regression
Diao, Song, Sun, Woodruff
Lemma 5.3. For any p ≥ 1. Condition(A) computes AU/(dγp) which is an (α, β√3d(tw)1/p−1/2, p)-well-
conditioned basis of A, with probability at least 1 −Qq
for completeness. Applying Theorem 5.2, we have that with probability at least 1−Qq
i=1 ni/wi ]⊤, then for all i ∈ [Qq
Proof. This lemma is similar to arguments in Clarkson et al. (2013), we simply adjust notation and parameters
i=1(ni/wi)δ, for all x ∈ Rr,
i=1 ni/wi],
if we consider y = Ax and write y⊤ = [z⊤1 , z⊤2 , . . . , zQq
i=1(ni/wi)δ.
where Si ∈ Rmi×Qq
we have
i=1 wi . In the following, suppose mi = t. By relating the 2-norm and the p-norm, for 1 ≤ p ≤ 2,
2kzik2,
2kzik2 ≤ kSizik2 ≤q 3
q 1
kSizikp ≤ t1/p−1/2kSzik2 ≤ t1/p−1/2q 3
2kzik2 ≥q 1
kSizikp ≥ kSizik2 ≥q 1
and similarly,
If p > 2, then
and similarly,
kSizikp ≤ kSizik2 ≤q 3
2kzik2 ≤q 3
kSizikp ≥ t1/p−1/2kSizik2 ≥ t1/p−1/2q 1
p = Pi kSizikp
p = Pi kzikp
2 w1/2−1/pkAxkp ≤ kSAxkp ≤q 3
q 1
q 1
2 t1/p−1/2kAxkp ≤ kSAxkp ≤q 3
p and kSAxkp
i=1(ni/wi)δ
Since kAxkp
p = kykp
i=1(ni/wi)δ
1 −Qq
and for p ∈ [2,∞) with probability 1 −Qq
In either case,
2kzik2 ≤ t1/p−1/2q 3
qYj=1
2 w1/2−1/pkzikp, w =
2kzikp,
wj.
2 w1/2−1/pkzikp,
2kzik2 ≥ t1/p−1/2q 1
2kzikp.
2 t1/p−1/2kAxkp,
2 w1/2−1/pkAxkp.
p,
for p ∈ [1, 2] we have with probability
kAxkp ≤ γpkSAxkp ≤
√3(tw)1/p−1/2kAxkp.
We have, from the definition of an (α, β, p)-well-conditioned basis, that
and for all x ∈ Rd,
kSAUkp ≤ α
kxkq ≤ βkSAU xkp.
kAU/(rγp)kp ≤Xi
Combining (13) and (14), we have that with probability at least 1 −Qq
Combining (13) and (15), we have that with probability at least 1 −Qq
kxkq ≤ βkSAU xkp ≤ β√3r(tw)1/p−1/2kAU
Hence AU/(rγp) is an (α, β√3r(tw)1/p−1/2, p)-well-conditioned basis.
kAUi/rγpkp ≤Xi
i=1(ni/wi)δ,
kSAUi/rkp ≤ α.
i=1(ni/wi)δ, for all x ∈ Rr,
1
rγp
xkp.
(13)
(14)
(15)
Sketching for Kronecker Product Regression and P-splines
i=1 di/ǫ)).
k=1 nnz(A(k)
ik
3/2 poly(Q2
i=q1+1 di. For any row-block A1
probability at least 1/2, kAx − bk1 ≤ (1 + ǫ) minx∈Rd kAx − bk1. For the special case when q = 2, n1 = n2, the
algorithm's running time is O(n1
Theorem 5.4. (Main result) Given ǫ ∈ (0, 1), A ∈ Rn×d and b ∈ Rn, Alg. 3 computes bx such that with
Proof. For notational simplicity, let us denote n[q1] = Qq1
d[q]\[q1] = Qq
i1 ⊗ . . . ⊗ A(q)
O(d(Pq
i=q1+1 n1, d[q1] = Qq1
i1 ⊗ . . . ⊗ A(q−1)
)) + dqm log(m)) (see Sec 2). Hence for SA, it takes:
i=1 ni, n[q]\[q1] = Qq
qYi=1
where S ∈ R(mQq
i (2 + 3q)/ǫ2 = O(poly(d/ǫ)). We need to compute an
orthogonal factorization SA = QRA in O(qmd2) and then compute U = R−1
in O(d3) time. Hence the total
A
running time of Algorithm Condition(A) is O(qmd2 + d3). Thus the total running time of computing SA and
Condition(A) is
qYi∈[q]\{k}
qXk=1
i=1 wi and m ≥ 100Qq
ni/wi + dqm log(m)
iq , computing Si1i2...iq (A1
ni/wi!
i=1(ni/wi))×Qq
i=1 di, and
) takes
iq
d
nnz(Ak)
i=1 d2
nnz(Ak)
qYi∈[q]\{k}
O
qXk=1
ni/wi + qYi=1
ni/wi! poly(d/ǫ) + qmd2 + d3 ,
We will compute U G in O(d2 log n) time. We compute eE = E(Aq1+1 ⊗ . . . ⊗ Aq)T in O(dn[q]\[q1]) time.
Then we can compute R(A1 ⊗ ··· ⊗ Aq1 )eEj in O(n[q1]d[q1] log n + d[q1]n[q]\[q1] log n) time.
Since computation of the median λi takes O(log n) time, computing all λi and then λe takes O(n[q]\[q1] log n)
time.
As AU G has O(log n) columns, we need to compute λe for each AU G using the above procedure and hence it
takes in total O(d(n[q1] + n[q]\[q1]) log2 n) time.
Sampling a column of AU G using λe takes O(log n) time, sampling an entry in M takes in total O(n[q1] + n[q]\[q1])
time.
k=1 wk poly(r) samples to select rows, the running time is d(n[q1] + n[q]\[q1]) log2 n ·
Now for simplicity, we set q = 2, ni = n0 for i ∈ [2]. Note that it is optimal to choose wi = w for i ∈ [2].
Substituting q = 2, ni = n0 and wi = w, we that the total running time of Alg. 3:
O(cid:0)dw−1n0(nnz(A1) + nnz(A2)) + w−2n2
0 poly(d/ǫ) + wn0 poly(d) log(n)(cid:1) .
For dense A1 and A2, nnz(A1) + nnz(A2) = O(n0) time, and so ignoring poly and log terms that do not depend
on n0, the total running time can be simplified to:
Setting w = √n0, we can minimize the above running time to O(n3/2
solving the problem by forming A1 ⊗ A2.
0
), which is faster than the n2
0 time for
O(w−1n2
0 + wn0).
Since we need pQq
pQq
k=1 wk poly(r).
|
1804.02269 | 2 | 1804 | 2019-05-06T14:15:28 | A Subquadratic Approximation Scheme for Partition | [
"cs.DS"
] | The subject of this paper is the time complexity of approximating Knapsack, Subset Sum, Partition, and some other related problems. The main result is an $\widetilde{O}(n+1/\varepsilon^{5/3})$ time randomized FPTAS for Partition, which is derived from a certain relaxed form of a randomized FPTAS for Subset Sum. To the best of our knowledge, this is the first NP-hard problem that has been shown to admit a subquadratic time approximation scheme, i.e., one with time complexity of $O((n+1/\varepsilon)^{2-\delta})$ for some $\delta>0$. To put these developments in context, note that a quadratic FPTAS for \partition has been known for 40 years.
Our main contribution lies in designing a mechanism that reduces an instance of Subset Sum to several simpler instances, each with some special structure, and keeps track of interactions between them. This allows us to combine techniques from approximation algorithms, pseudo-polynomial algorithms, and additive combinatorics.
We also prove several related results. Notably, we improve approximation schemes for 3SUM, (min,+)-convolution, and Tree Sparsity. Finally, we argue why breaking the quadratic barrier for approximate Knapsack is unlikely by giving an $\Omega((n+1/\varepsilon)^{2-o(1)})$ conditional lower bound. | cs.DS | cs |
A Subquadratic Approximation Scheme for Partition
Marcin Mucha∗
Karol Węgrzycki∗
Michał Włodarczyk∗
The subject of this paper is the time complexity of approximating Knapsack, Subset
Abstract
time randomized FPTAS for Partition, which is derived from a certain relaxed form of a
randomized FPTAS for Subset Sum. To the best of our knowledge, this is the first NP-hard
problem that has been shown to admit a subquadratic time approximation scheme, i.e., one
Sum, Partition, and some other related problems. The main result is an eO(n + 1/ε5/3)
with time complexity of O((n + 1/ε)2−δ) for some δ > 0. To put these developments in context,
note that a quadratic FPTAS for Partition has been known for 40 years.
Our main contribution lies in designing a mechanism that reduces an instance of Subset Sum
to several simpler instances, each with some special structure, and keeps track of interactions
between them. This allows us to combine techniques from approximation algorithms, pseudo-
polynomial algorithms, and additive combinatorics.
We also prove several related results. Notably, we improve approximation schemes for 3SUM,
(min, +)-convolution, and TreeSparsity. Finally, we argue why breaking the quadratic barrier
for approximate Knapsack is unlikely by giving an Ω((n + 1/ε)2−o(1)) conditional lower bound.
∗Institute of Informatics, University of Warsaw, Poland, {mucha, k.wegrzycki, m.wlodarczyk}@mimuw.edu.pl
1 Introduction
The Knapsack-type problems are among the most fundamental optimization challenges. These
problems have been studied for more than a century already, as their origins can be traced back to
the 1897's paper by Mathews [52].
The Knapsack problem is defined as follows:
Definition 1.1 (Knapsack). Given a set of n items En = {1, . . . , n}, with item j having a positive
integer weight wj and value vj, together with knapsack capacity t. Select a subset of items E ⊆ En,
such that the corresponding total weight w(E) = Pi∈E wi does not exceed the capacity t and the
total value v(E) =Pi∈E vi is maximized.
Knapsack is one of the 21 problems featured in Karp's list of NP-complete problems [43]. We
also study the case where we are allowed to take each element multiple times, called Unbounded
Knapsack. Let Σ(S) denote the sum of elements S. Subset Sum is defined as follows:
Definition 1.2 (Subset Sum). Given a set S ⊂ N of n numbers (sometimes referred to as items)
and an integer t, find a subset S′ ⊆ S with maximal Σ(S′) that does not exceed t.
Subset Sum is a special case of Knapsack, where item weights are equal to item values. This
problem is NP-hard as well. In fact, it remains NP-hard even if we fix t to be Σ(S)/2. This problem
is called the Number Partitioning Problem (or Partition, as we will refer to it):
Definition 1.3 (Partition). Given a set S ⊂ N of n numbers, find a subset S′ ⊆ S with maximal
Σ(S′) not exceeding Σ(S)/2.
The practical applications of Partition problem range from scheduling [42] to minimization of
circuits sizes, cryptography [53], or even game theory [35, 54]. The decision version of this problem
is sometimes humorously referred to as "the easiest NP-complete problem" [35]. In this paper we
will demonstrate that there is a grain of truth in this claim.
All the aforementioned problems are weakly NP-hard and admit pseudo-polynomial time algo-
rithms. The first such an algorithm for the Knapsack was proposed by Bellman [15] and runs in
time O(nt). This bound was improved for the Subset Sum [48] and the current best (randomized)
results see Section 2). The strong dependence on t in all of these algorithms makes them impractical
for a large t (note that t can be exponentially larger than the size of the input). This dependence
time complexity for this problem is eO(n + t), due to Bringmann [17] (for more on these and related
has been shown necessary as an O(cid:0)poly(n)t0.99(cid:1) algorithm for the Subset Sum would contradict
both the SETH [4] and the SetCover conjecture [23].
One possible approach to avoid the dependence on t is to settle for approximate solutions. The
notion of approximate solution we focus on in this paper is that of a Polynomial Time Approximation
Scheme (PTAS). A PTAS for a maximization problem is an algorithm that, given an instance of
size n and a parameter ε > 0, returns a solution with value S, such that OPT(1 − ε) ≤ S ≤ OPT.
It also needs to run in time polynomial in n, but not necessarily in 1/ε (so, e.g., we allow time
complexities like O(n1/ε)). A PTAS is a Fully Polynomial Time Approximation Scheme (FPTAS)
if it runs in time polynomial in both n and 1/ε. Equivalently, one can require the running time to
be polynomial in (n + 1/ε). For example, O(n2/ε4) = O((n + 1/ε)6). For definitions of problems
in both exact and approximate sense see Appendix C.
The first approximation scheme for Knapsack (as well as Subset Sum and Partition as spe-
cial cases) dates back to 1975 and is due to Ibarra and Kim [39]. Its running time is O(n/ε2). After
1
a long line of improvements [29, 30, 45, 44, 50, 47], the current best algorithms for each problem are:
the O(min{n/ε, n+1/ε2}) algorithm for Partition due to [31], the O(min{n/ε, n+1/ε2 log (1/ε)})
algorithm for Subset Sum due to [46] and, a very recent eO(n + 1/ε12/5) for Knapsack, due to [19].
Observe that all of these algorithms work in Ω((n + 1/ε)2) time. In fact, we are not aware of
the existence of any FPTAS for an NP-hard problem working in time O((n + 1/ε)2−δ).
Open Question 1. Can we get an O((n + 1/ε)2−δ) FPTAS for any Knapsack-type problem (or
any other NP-hard problem) for some constant δ > 0 or justify that it is unlikely?
In this paper we resolve this question positively, by presenting the first such algorithm for the
Partition problem. This improves upon almost 40 years old algorithm by Gens and Levner [31].
On the other hand, we also provide a conditional lower bound suggesting that similar improvement
for the more general Knapsack problem is unlikely.
After this paper was announced, Bringmann [18] showed that for any δ > 0, an O((n + 1/ε)2−δ)
algorithm for Subset Sum would contradict the (min, +)-convolution-conjecture. This not only
shows a somewhat surprising separation between the approximate versions of Partition and Sub-
set Sum, but also explains why our techniques do not seem to transfer to approximating Subset
Sum.
1.1 Related Work
In this paper we avoid the dependence on t by settling on approximate instead of exact solutions.
Another approach is to allow running times exponential in n. This line of research has been very
active with many interesting results. The naive algorithm for Knapsack works in O∗(2n) time
by simply enumerating all possible subsets. Horowitz and Sahni [36] introduced the meet-in-the-
middle approach and gave an exact O∗(2n/2) time and space algorithm. Schroeppel and Shamir
[56] improved the space complexity of that algorithm to O∗(2n/4). Very recently Bansal et al. [11]
showed an O∗(20.86n)-algorithm working in polynomial space.
An interesting question (and very relevant for applications in cryptography) is how hard Knap-
sack type problems are for random instances. For results in this line of research see [6, 7, 8, 37].
1.2 History of Approximation Schemes for Knapsack-type problems
To the best of our knowledge, the fastest approximation for Partition dates back to 1980 [31]
with eO(min{n/ε, n + 1/ε2}) running time1. The majority of later research focused on matching this
running time for the Knapsack and Subset Sum. In this section we will present an overview of
the history of the FPTAS for these problems.
The first published FPTAS for the Knapsack is due to Ibarra and Kim [39]. This naturally
In their approach,
gives approximations for the Subset Sum and Partition as special cases.
the items were partitioned into large and small classes. The profits are scaled down and then the
problem is solved optimally with dynamic programming. Finally, the remaining empty space is filled
up greedily with the small items. This algorithm has a complexity O(n/ε2) and requires O(n+1/ε3)
space.2 Lawler [50] proposed a different method of scaling and obtained O(n + 1/ε4) running time.
1 As is common for Knapsack-type problems, the eO notation hides terms poly-logarithmic in n and 1/ε, but not
in t.
2In [47, Section 4.6] there are claims, that 1975 Karp [44] also gives O(n/ε2) approximation for Subset Sum.
2
Knapsack
Knapsack
Subset Sum
Knapsack
Partition
Subset Sum
Subset Sum
Knapsack
Partition
Reference
[15] [47]
[39, 44]
* [32, 29]
[50]
* [31]
[30]
* [46]
* [19]
This Paper
O(n2/ε)
O(n/ε2)
O(n/ε)
O(n + 1/ε4)
O(n + 1/ε2)
O(n + 1/ε3)
eO(n + 1/ε2)
eO(n + 1/ε12/5)
eO(n + 1/ε5/3)
Table 1: Brief history of FPTAS for Knapsack-type problems. Since Partition is a special case
of Subset Sum, and Subset Sum is a special case of Knapsack, an algorithm for Knapsack
also works for Subset Sum and Partition. We omit redundant running time factors for clarity,
For a more robust history see [47, Section 4.6]. A star (*) marks the papers that match the previous
e.g., [46] actually runs in eO(min{n/ε, n + 1/ε2}) time but [32, 29] gave O(n/ε) algorithm earlier.
best eO(cid:0)(n + 1/ε)2(cid:1) complexity for Partition problem.
Running Time Problem
Later, Gens and Levner [32, 29] obtained an O(n/ε) algorithm for the Subset Sum based on a
different technique. Then, in 1980 they proposed an even faster O(min{n/ε, n+1/ε2}) algorithm [31]
for the Partition. To the best of our knowledge this algorithm remained the best (until this paper).
Subsequently, Gens and Levner [30] managed to generalize their result to Subset Sum with an
increase of running time and obtained O(min{n/ε, n + 1/ε3}) time and O(min{n/ε, n + 1/ε2})
space algorithm [30]. Finally, Kellerer et al. [46] improved this algorithm for Subset Sum by giving
O(min{n/ε, n + 1/ε2 log (1/ε)}) time and O(n + 1/ε) space algorithm. This result matched (up to
the polylogarithmic factors) the running time for Partition.
For the Knapsack problem Kellerer and Pferschy [45] gave an O(n min{log n, log (1/ε)} +1/ε2 log (1/ε)
min{n, 1/ε log (1/ε)}) time algorithm (note that the exponent in the parameter (n + 1/ε) is 3 here)
and for Unbounded Knapsack Jansen and Kraft [40] gave an O(n + 1/ε2 log3 (1/ε)) time algo-
rithm (the exponent in (n+1/ε) is 2, see Appendix C for the definition of Unbounded Knapsack).
Very recently Chan [19] presented the currently best eO(n + 1/ε12/5) algorithm for the Knapsack.
1.3 Our Contribution
Our main result is the design of the mechanism that allows us to merge the pseudo-polynomial time
algorithms for Knapsack-type problems with algorithms on dense Subset Sum instances. The
most noteworthy application of these reductions is the following.
5
3 ) randomized time FPTAS for Partition.
and Levner [31]. Our algorithm also generalizes to a weak (1− ε)-approximation for Subset Sum.3
Theorem 1.5. There is a randomized weak (1 − ε)-approximation algorithm for Subset Sum
Theorem 1.4. There is an eO(n + 1/ε
This improves upon the previous, 40 year old bound of eO(n+1/ε2) for this problem, due to Gens
running in eO(cid:0)n + 1/ε
3Weak approximation can break the capacity constraint by a small factor. Definition 2.1 specifies formally what
3(cid:1) time.
5
weak (1 − ε)-approximation for Subset Sum is.
3
For a complete proof of these theorems see Section 5. We also present a conditional lower bound
for Knapsack and Unbounded Knapsack.
Theorem 1.6. For any constant δ > 0, an FPTAS for Knapsack or Unbounded Knapsack
with O((n + 1/ε)2−δ) running time would refute the (min, +)-convolution conjecture.
This means that a similar improvement is unlikely for Knapsack. Also, this shows that the
algorithm of [40] for Unbounded Knapsack is optimal (up to polylogarithmic factors). This lower
bound is relatively straightforward and follows from previous works [24, 49] and was also observed
in [19]. The lower bound also applies to the relaxed, weak (1 − ε)-approximation for Knapsack
and Unbounded Knapsack, which separates these problems from weak (1 − ε)-approximation
approximation for Subset Sum. This result was recently extended by Bringmann [18] who showed
a conditional hardness for obtaining a strong subquadratic approximation for Subset Sum, which
explains why we need to settle for a weak approximation.
Lately it has been shown that the exact pseudo-polynomial algorithms for Knapsack and Un-
bounded Knapsack are subquadratically equivalent to the (min, +)-convolution [24, 49]. There-
fore, as a possible first step towards obtaining an improved FPTAS for Knapsack, we focus our
attention on (min, +)-convolution.
This also entails an improvement for the related TreeSparsity problem (see Section 7). The
Theorem 1.7. (1 + ε)-approximate (min, +)-convolution can be computed in eO((n/ε) log W ) time.
best previously known algorithms for both problems worked in time eO((n/ε2)polylog(W )) (see Back-
The techniques used to improve the approximation algorithm for (min, +)-convolution also apply
to approximation algorithms for 3SUM. For this problem we are able to show an algorithm that
matches its asymptotic lower bounds.
urs et al. [10]).
Theorem 1.8. There is a deterministic algorithm for (1 + ε)-approximate 3SUM running in time
eO((n + 1/ε)polylog(W )).
Theorem 1.9. Assuming the Strong-3SUM conjecture, there is no eO((n + 1/ε1−δ)polylog(W ))
algorithm for (1 + ε)-approximate 3SUM, for any constant δ > 0.
For proofs of these theorems and detailed running times see Sections 8
1.4 Organization of the Paper
In the Section 2 we present the building blocks of our framework and a sketch of the approximation
scheme for Partition. Section 3 contains the notation and preliminaries, and the main proof is
divided into Sections 4 and 5. In Sections 6 and 7 we present the algorithms for (min, +)-convolution
and TreeSparsity. In the Section 8 we present the algorithms for 3SUM. The proofs of technical
lemmas can be found in Appendix A and Appendix B. In Appendix C we give formal definitions of
all problems.
2 Connecting Dense, Pseudo-polynomial and Approximation Algo-
rithms for Knapsack-type problems: An Overview
In this section we describe main building blocks of our framework. We also briefly discuss the recent
advances in the pseudo-polynomial algorithms for Subset Sum and discuss how to use them. Then,
4
we explain the intuition behind the trade-off we exploit and give a sketch of the main algorithm.
The formal arguments are located in Section 5.
Difficulties with Rounding for Subset Sum There is a strong connection between approxi-
mation schemes and pseudo-polynomial algorithms [59]. For example, a common theme in approx-
imating knapsack is to reduce the range of the values (while keeping the weights intact) and then
apply a pseudo-polynomial algorithm. Rounding the weights would be tricky because of the hard
knapsack constraint. In particular, if one rounds the weights down, some feasible solutions to the
rounded instance might correspond to infeasible solutions in the original instance. On the other
hand, when rounding up, some feasible solutions might become infeasible in the rounded instance.
Recently, new pseudo-polynomial algorithms have been proposed for Subset Sum (see Koiliaris
and Xu [48] and Bringmann [17]). A natural idea is to use these to design an improved approximation
scheme for Subset Sum. However, this seems to be difficult due to rounding issues discussed above.
After this paper was announced, Bringmann [18] explained this difficulty by giving a conditional
lower bound on a quadratic approximation of Subset Sum.
2.1 Weak Approximation for Subset Sum and Application to Partition
Because of these rounding issues, it seems hard to design a general rounding scheme that, given a
pseudo-polynomial algorithm for Subset Sum, produces an FPTAS for Subset Sum. What we
can do, however, is to settle for a weaker notion of approximation.
Definition 2.1 (Weak apx for Subset Sum). Let Z ∗ be the optimal value for an instance (Z, t) of
Subset Sum. Given (Z, t), a weak (1 − ε)-approximation algorithm for Subset Sum returns Z H
such that (1 − ε)Z ∗ ≤ Z H < (1 + ε)t.
Compared to the traditional notion of approximation, here we allow a small violation of the
packing constraint. This notion of approximation is interesting in itself. Indeed, it has been already
considered in the stochastic regime for Knapsack [16].
Before going into details of constructing the weak (1 − ε)-approximation algorithms for the
Subset Sum, let us establish a relationship with the approximation for the Partition.
Corollary 2.2. If we can weakly (1 − ε)-approximate Subset Sum in time eO(T (n, ε)), then we
can (1 − ε)-approximate Partition in the same eO(T (n, ε)) time.
This is because of the symmetric structure of Partition problem: If a subset Z ′ violates the
hard constraint (t ≤ Σ(Z ′) ≤ (1 + ε)t), then the set Z − Z ′ is a good approximation and does not
violate it (recall that in Partition problem we always have t = Σ(Z)/2). For a formal proof see
Section A.
2.2 Constructing Weak Approximation Algorithms for Subset Sum: A Sketch
Fact 2.3. Given an eO(T (n, t)) exact algorithm for Subset Sum, we can construct a weak (1 − ε)-
approximation algorithm for Subset Sum working in time eO(T (n, n
Proof. We assume that the exact algorithm for the Subset Sum works also for multisets. We will
address this issue in more detail in Section 4.1.
2ε )).
5
Let Z = {v1, . . . , vn} and t constitute a Subset Sum instance. Let I be the set of indices
of elements of some optimal solution, and let OPT be their sum. Let us also introduce a scaled
approximation parameter ε′ = ε
4 .
Let k = 2ε′t
n . Define a rounded instance as follows: the (multi)-set of Z contains a copy of
k(cid:5) for each i ∈ {1, . . . , n}, and t =(cid:4) t
k(cid:5).
vi =(cid:4) vi
Apply the exact algorithm A to the rounded instance ( Z, t). Let I ′ be the set of indices of
We claim that {vi : i ∈ I ′} is a weak (1 − ε) approximation for Z and t. First let us show that
elements of the solution found.
this solution is not much worse than OPT:
Xi∈I ′
vi ≥ kXi∈I ′
vi ≥ kXi∈I
vi = kXi∈Ij vi
kk ≥Xi∈I
(vi − k) ≥ OPT − nk = OPT − 2ε′t ≥ OPT(1 − ε).
The last inequality holds because we can assume OPT ≥ t/2 (see Section 4.3 for details).
Similarly, we can show that this solution does not violate the hard constraint by too much:
Xi∈I ′
vi ≤Xi∈I ′
(kvi + k) ≤ nk + kXi∈I ′
vi ≤ nk + tk ≤ nk + k + t ≤ 3ε′t + t ≤ t(1 + ε).
Finally, since the exact algorithm is applied to a (multi)-set of n items with t =(cid:4) t
the resulting algorithm runs in the claimed time.
k(cid:5) =(cid:4) n
2ε′(cid:5),
We state the above proof only to give the flavour of the basic form of reductions in this paper.
Usually reductions that we will consider are more complex for technical reasons. One thing to note
in particular is that the relation between k and ε is dictated by the fact, that there may be as many
as n items in the optimal solution. Given some control over the solution size, one can improve this
reasoning (see Lemma 4.7).
2.3 Approximation via Pseudo-polynomial time Subset Sum algorithm
Currently, the fastest pseudo-polynomial algorithm for Subset Sum runs in time eO(n + t), ran-
domized. S(Z, t) denotes the set of all possible subsums of set Z up to integer t (see Section 3).
Theorem 2.4 (Bringmann [17]). There is a randomized, one-sided error algorithm with running
time O(n + t log t log3 n
δ log n), that returns a set Z ′ ⊆ S(Z, t), containing each element from S(Z, t)
with probability at least 1 − δ.
This suffices to solve Subset Sum exactly with high probability. Here S(Z, t) is represented
by a binary array which for a given index i tells whether there is a subset that sums up to i (see
Section 3 for a formal definition). For our trade-off, we actually need a probabilistic guarantee on
all elements of S(Z, t) simultaneously. Fortunately, this kind of bound holds for this algorithm as
well (see [24, Appendix B.3.2] for detailed analysis).
Corollary 2.5. There is a randomized eO(n + t) algorithm that computes S(Z, t) with a constant
probability of success.
6
The first case where this routine comes in useful occurs when all items are in the range [γt, t]
(think of γ as a trade-off parameter set to ε−2/3). Note, that any solution summing to at most t
can consist of at most 1/γ such elements. This observation allows us to round the elements with
lower precision and still maintain a good approximation ratio, as follows:
v′
γεt(cid:23) ,
i =(cid:22) 2vi
t′ =(cid:22) 2t
γεt(cid:23) =(cid:22) 2
γε(cid:23) .
γε ) and
returns an array of solutions with an additive error ±εt with high probability (see Lemma 5.1).
Similar reasoning about sparseness also applies if the number of items is bounded (i.e., when n =
ε2 ) and provides the same guarantees
Bringmann's [17] algorithm on the rounded instance runs in time eO(n + t′) = eO(n + 1
eO( γ
ε )). In that case Bringmann's [17] algorithm runs in time eO( γ
(see Lemma 5.2 and also the next section).
2.4 Approximation via Dense Subset Sum
Now we need a tool to efficiently solve the instances where all items are in range [0, γt), so-called
dense instances. More formally, an instance consisting of m items is dense if all items are in the
range [1, mO(1)].
Intuitively, rounding does not work well for these instances since it introduces
large rounding errors. On the other hand, if an instance contains many distinct numbers on a small
interval, one can exploit its additive structure.
Theorem 2.6 (Galil and Margalit [27]). Let Z be a set of m distinct numbers in the interval (0, ℓ]
such that
and let L := 100·Σ(Z)ℓ log ℓ
m2
m > 1000 ·
√ℓ log ℓ,
.
Then in O(m + ((ℓ/m) log ℓ)2) preprocessing time we can build a structure that can answer the
following queries in constant time. In a query the structure receives a target number t ∈ (L, Σ(Z)−L)
and decides whether there is a Z ′ ⊆ Z such that Σ(Z ′) = t. The structure is deterministic.
In fact we will use a more involved theorem that can also construct a solution in O(log(l))
time but we omit it here to keep this section relatively free of technicalities (see Section 5.2 for a
discussion regarding these issues).
Observe that L = eO(ℓ1.5) (because Σ(Z) < mℓ) and the running time is bounded by eO(m + ℓ)
(because ℓ/m = O(√ℓ)). We will apply this result for the case ℓ = γt (see Lemma 5.5). Recall,
that Bringmann's [17] algorithm runs in time eO(m + t), which would be slower by the factor γ (the
trade-off parameter). For simplicity, within this overview we will assume, that Theorem 2.6 provides
a data structure that can answer queries with the target numbers in [0, Σ(Z)]. In the actual proof,
we need to overcome this obstacle, by merging this data structure with other structures, responsible
for targets near the boundary, which we call marginal targets (see Lemma 5.3).
Suppose our instance consists of m elements in the range [0, γt]. We use the straightforward
rounding scheme, as in the proof of Fact 2.3.
v′
εt (cid:23) ,
i =(cid:22) 2mvi
t′ =(cid:22) 2mt
εt (cid:23) =(cid:22) 2m
ε (cid:23) .
7
We chose γt as the upper bound on item size, so that ℓ′ = mγ/ε is an upper bound on v′
i.
Now, if the number of items satisfies the inequality ℓ′ < m2, then we can use the Theorem 2.6 with
cannot use the approach from [27].
from the range that is of our interest (for a careful proof see Section 5).
running time eO(m + ℓ′) = eO(m + mγ/ε). This provides a data structure that can answer queries
Still, it can happen that most of the items are in the sparse instance (i.e., ℓ′ ≥ m2) and we
In that case we use Theorem 2.4 again, with running time
eO(m + γ
eO(m + γ
In the end, we are able to compute an array of solutions, for items in range [0, γt] in time
ε2 ) with additive error ±εt and high probability (see Lemma 5.6). The last term in
time complexity comes from handling the marginal queries.
ε2 ) (see Lemma 5.2).
ε2 + mγ2
2.5 A Framework for Efficient Approximation
In this section we will sketch the components of our mechanism (see Algorithm 1). The mechanism
combines pseudo-polynomial Bringmann's [17] algorithm with Galil and Margalit [27] algorithm for
dense instances of Subset Sum.
Algorithm 1 Roadmap for the weak (1−ε)-approximation for Subset Sum. Input: item set Z, t, ε
1: ensure OP T ≥ t/2
3: repeat
4:
2: reduce Z to eO(1/ε)
partition items into Zlarge and Zsmall
divide [0, γt] into ℓ = O(γ log(n)/ε) · Zsmall segments
round down small items
remove item repetitions in Zsmall
5:
6:
7:
8: until ℓ = O(γ log(n)/ε) · Zsmall
9: build a data structure for large items
10: if Zsmall = eO(√ℓ) then
build a data structure for small items
11:
12: else
13:
14:
15: end if
16: merge the data structures for large and small items
build data structures for marginals
exploit the density of the instance to cover the remaining case
We begin by reducing the number of items in the instance Z to roughly eO(1/ε) items to get a
near linear running time (see Lemma 4.2). After that our goal is to divide items into small and
large and process each part separately, as described earlier.
However, Theorem 2.6 requires a lower bound on the number of distinct items. To control
this parameter, we merge identical items into larger ones, until each item appears at most twice.
However, this changes the number of items, and so the procedure might have to be restarted.
Lemma 4.4 guarantees that we require at most log n such refinement steps.
In the next phase we decide which method to use to solve the instance depending on its density
(line 10). We encapsulate these methods into data structures (lines 11-14). Finally we will need
to merge the solutions. For this task we introduce the concept of membership oracles (see Defi-
8
nition 4.5) that are based on FFT and backtracking to retrieve solutions (see Lemma 4.6). The
simplified trade-off schema is presented on the Figure 1.
γt
t
Dense Instance
ǫt
m
ǫγt
Sparse Instance
m2 ≤ ℓ
eO(m + γ
ε2 )
γǫ )
eO(n + 1
If m2 > ℓ
ε2 + mγ2
ε2 )
eO( γ
Figure 1: Overall schema of trade-off and usage of building blocks. The parameter m denotes
number of items in the dense instance, n is the number of all elements, γ is the trade-off parameter,
ℓ is the upper bound on the item size after rounding, t is the target sum. The buckets in the
sparse/dense instance depict the rounding scheme for small and large items.
The final running time of our framework is eO(n + 1
γ(n, ε) > 0 (see Lemma 5.9). For γ = ε−2/3, this gives us an eO(n + ε−5/3) time weak (1 − ε)-
approximation approximation for Subset Sum.
ε3 ) with high probability for any
γε + γ
ε2 + γ2
3 Preliminaries
For a finite multiset Z ⊂ N we denote its size as Z, the number of distinct elements as Z, and
the sum of its elements as Σ(Z).
For a number x we define pow(x) as the largest power of 2 not exceeding x. If x < 2 we set
pow(x) = 1.
For sets A, B ⊂ N their bounded algebraic sum A ⊕t B is a set {a + b : a ∈ {0} ∪ A, b ∈
{0} ∪ B} ∩ [0, t].
Definition 3.1 (Subsums). For a finite multiset Z ⊂ N we define S(Z)k as a set of all possible
subset sums of Z of size at most k, i.e., x ∈ S(Z)k iff there exists S′ ⊆ Z, such that Σ(S′) = x and
S′ ≤ k. S(Z) is the set without the constraint on the size of the subsets, i.e., S(Z) := S(Z)∞.
The capped version is defined as S(Z, t)k := S(Z)k ∩ [0, t] and S(Z, t) := S(Z) ∩ [0, t].
We call two multisets Z1, Z2 ⊂ N equivalent if S(Z1) = S(Z2).
Note that 0 ∈ S(Z, t)k for all sets Z and t, k > 0.
Definition 3.2 ((ε, t)-closeness). We say that set B is (ε, t)-close to A if there is a surjection
φ : A → B such that x − εt ≤ φ(x) ≤ x + εt. A Subset Sum instance (Z2, t) is ε-close to (Z1, t) if
S(Z2, t) is (ε, t)-close to S(Z1, t).
Sometimes, when there is no other notation on t, we will use the notion of ε-closeness as a
(ε, t)-close.
Usually the surjection from the definitions will come by rounding down the item sizes and each
item set will get a moderately smaller total size. We will also apply the notion of (ε, t)-closeness to
binary arrays having in mind the sets they represent.
9
Fact 3.3. If A is (ε, t)-close to S(Z1, t) and B is (ε, t)-close to S(Z2, t) then A ⊕t B is (2ε, t)-close
to S(Z1 ∪ Z2, t)
We will also need to say, that there are no close elements in a set. It will come in useful to show,
that after rounding down all the elements are distinct.
Definition 3.4 ((x)-distinctness). The set S is said to be (x)-distinct if every interval of length x
contains at most one item from S. The set S is said to be (x, 2)-distinct if every interval of length
x contains at most two items from S.
4 Preprocessing
This section is devoted to simplify the instance of Subset Sum in order to produce a more readable
proof of the main algorithm. In here we will deal with:
• multiplicities of the items,
• division of the instance into large and small items,
• proving that rounding preserves ε-closeness,
• reducing a number of items from n to eO(1/ε) items.
The solutions to these problems are rather technical and well known in the community [46, 47,
48, 17]. We include it in here because these properties are used in approximation algorithms [47, 46]
and exact pseudo-polynomial algorithms [48, 17] communities separately. We expect that reader
may not be familiar with both of these technical toolboxes simultaneously and accompany this
section with short historical references and pointers to the original versions of proofs.
4.1 From Multisets to Sets
The general instance of Subset Sum may consists of plenty of items with equal size. Intuitively,
these instances seem to be much simpler than instances where almost all items are different. The
next lemma will allow us to formally capture this intuition with the appropriate reduction. This
lemma was proposed in [48, Lemma 2.2] but was also used in [17].
Lemma 4.1 (cf. Lemma 2.2 from [48]). Given a multiset S of integers from {1, . . . , t}, such that
S = n and the number of distinct items S is n′, one can compute, in O(n log n) time, a multiset
T , such that:
• S(S, t) = S(T, t)
• T ≤ S
• T = O(n′ log n)
• no element in T has multiplicity exceeding two.
10
Proof. We follow the proof from [48, Lemma 2.2], however the claimed bound on T is only
O(n′ log t) therein. Consider an element x with the multiplicity 2k + 1. We can replace it with
a single copy of x and k copies of 2x while keeping the multiset equivalent. If the multiplicity is
2k + 2 we need 2 copies of x and k copies of 2x. We iterate over items from the smallest one and
for each with at least 3 copies we perform the replacement as described above. Observe that this
procedure generates only elements of form 2ix where i ≤ log n and x is an element from S. This
yields the bound on T. The routine can be implemented to take O(log n) time for creating each
new item using tree data structures.
later found applications in Knapsack-type problems [47].
4.2 From n Items to eO(1/ε) Items
To reduce number of items n to eO(1/ε) Kellerer et al. [46] gave a very intuitive construction that
we want to get rid of factor O(n). We divide an instance to k = (cid:6) 1
ε(cid:7) intervals of length εt, i.e.,
jm) items,
Ij := (jt, (j + 1)t]. Next notice that for interval Ij we do not need to store more than O(l k
Intuitively, rounding scheme described in Section 2 could divide the items into O(n/ε) intervals
and this would result with an ε-close instance to the original one. In here we start similarly but
because their sum would exceed t (this is the step where ε factor will come in). Finally, the number
of items is upper bounded (up to the constant factors):
kXj=1(cid:24) k
j(cid:25) ≤ k
kXj=1
1
j
< k log k = O(1/ε log (1/ε))
εt
ki ·
εt
2
+
log( 2n
ε )Xi=1
ε ) log(n)(cid:1).
11
t
ε log( n
The last inequality is just an upper bound on harmonic numbers. This was a very informal
sketch of the proof of [46] construction to give some intuition. The next technical lemma is based
on their trick.
Lemma 4.2. Given a Subset Sum instance (Z, t), Z = n, one can find an ε-close instance (Z2, t)
2i−1 ) to the closest multiplicity of(cid:4) εt
such that Z2 = O(cid:0) 1
Proof. We begin with constructing Z1 as follows. For i = 1, . . . , log( 2n
in Z ∩ [ t
2i ,
that Z1 = O(cid:0) 1
most t and its counterpart Y1 ⊆ Z1. We lose at most n· εt
Let ki = I ∩ [ t
2i ,
[ t
2i ,
ε ) log(n)(cid:1). The running time of this procedure is O(Z + Z2).
2i+1(cid:5). We neglect elements smaller than εt
ε )(cid:1).
2i−1 ) and ti denote the sum of elements in I ∩ [ t
2i ,
2i+1 and ki · t
2i ≤ ti, we have
log( 2n
εt
2i+1 ≤
We argue that (Z1, t) is ε-close to (Z, t). To see this, consider any subset I ⊆ Z summing to at
2 by omitting items smaller than εt
2n .
t
2i−1 ). Since each element in
ε ) we round down each element
2n . Observe
t
2i−1 ) has been decreased by at most
Σ(I) − Σ(Y1) ≤
εti
2 ≤ εt.
2n = εt
εt
2
+
ε log( n
t
ε )Xi=1
In the end we take advantage of Lemma 4.1 to transform Z1 into an equivalent multiset Z2 such
that Z2 ≤ Z1 log(Z1) = O(cid:0) 1
ε log( n
Note, that we discarded items smaller than εt
2n . We do this because sum of these elements is
just too small to influence the worst case approximation factor. We do not consider them just for
the simplicity of analysis. To make this algorithm competitive in practice, one should probably just
greedily add these small items to get a little better solution.
4.3 From One Instance to Small and Large Instances
First we need a standard technical assumption, that says that we can cheaply transform an instance
to one with a lower bounded solution. We will need it just to simplify the proofs (e.g., it will allow
us to use Lemma B.2 multiple times).
Lemma 4.3. One may assume w.l.o.g. that for any Subset Sum instance OPT ≥ t
2 .
Proof. Let us remove from the item set Z all elements exceeding t since they cannot belong to any
solution. If Σ(Z) ≤ t then the optimal solution consists of all items. Otherwise consider a process
in which Y1 = Z and in each turn we obtain Yk+1 by dividing Yk into two arbitrary non-empty parts
and taking the one with a larger sum. We terminate the process when Ylast contains only one item.
Since Σ(Y1) > t, Σ(Ylast) ≤ t, and in each step the sum decreases by at most factor two, for some
k it must be Σ(Yk) ∈ [ t
2 , OPT cannot be
lower.
2 , t]. Because there is a feasible solution of value at least t
One of the standard ways of solving Subset Sum is to separate the large and small items [47].
Usually these approximations consider items greater and smaller than some trade-off parameter.
Our techniques require a bound on the multiplicities of small items, which is provided by the next
lemma.
Lemma 4.4 (Partition into Small / Large Items). Given an instance (Z, t) of Subset Sum,
an approximation factor ε, and a trade-off parameter γ, one can deterministically transform the
instance (Z, t), in time O(n log2 n), to an ε-close instance (Zsmall ∪ Zlarge, t) such that:
• ∀zs ∈ Zsmall,
• ∀zl ∈ Zlarge,
• The set Zsmall is (
it holds that zs < γt,
it holds that zl ≥ γt,
at most 2 occurrences of each item.
εt
m·log n , 2)-distinct where m = O(Zsmall), i.e., after rounding there can be
εt
Proof. We call an item x large if x ≥ γt and small otherwise. Let Y0 be the initial set of small
m0 log n ). We round down the size of each small item to the closest
items and m0 = Y0, q0 = pow(
multiplicity of q0. Then we apply Lemma 4.1 to the set of small items to get rid of items with 3
or more copies. Note that this operation might introduce new items that are large. We obtain a
mi log n ). It
new set of small items Y1 and repeat this procedure with notation mi = Yi, qi = pow(
holds that mi+1 ≤ mi and qi qi+1. We stop the process when mi+1 ≥ mi
2 , which means there can
be at most log n iterations. Let m denote the final number of small items and q ≥ εt
4m log n -- the last
power of 2 used for rounding. All small items now occur with multiplicities at most 2.
εt
Let us fix Zsmall as the set of small items after the modification above and likewise Zlarge. In the
mi log n , so each new instance
ε
log n -close to the previous one. There are at most log n steps and the removal of copies keeps the
i-th rounding step values of mi items are being decreased by at most
is
instance equivalent, therefore (Zsmall ∪ Zlarge, t) is ε-close to (Z, t).
εt
Our algorithm works independently on these two instances and produces two arrays ε-close to
them. The construction below allows us to join these solutions efficiently. We want to use them
even if we have only access to them by queries. We formalize this as an (ε, t)-membership-oracle.
The asymmetry of the definition below will become clear in Lemma 4.7.
12
Definition 4.5 ((ε, t)-membership-oracle). The (ε, t)-membership-oracle of a set X is a data struc-
ture, that given an integer q answers yes/no obeying following conditions:
1. if X contains an element in [q − εt, q + εt], then the answer is yes,
2. if the answer was yes, then X contains an element in [q − 2εt, q + 2εt].
A query to the oracle takes eO(1) time. Moreover, if the oracle answers yes, then it can return a
witness x in eO(1) time.
Below we present an algorithm that can efficiently join the solutions. We assume, that we have
only query-access to them and want to produce an (ε, t)-membership-oracle of the merged solution.
Lemma 4.6 (Merging solutions). Given S(Z1, t) and S(Z2, t) as (ε, t)-membership-oracles
• S1 that is (ε, t)-close instance to S(Z1, t),
• S2 that is (ε, t)-close instance to S(Z2, t),
ε ) time, construct a (2ε, t)-membership-oracle for S(Z1 ∪ Z2, t).
Proof. For an ease of presentation, only in this proof we will use interval notation of inclusion, i.e.,
we will say that (a, b] ⊓ A iff ∃xx ∈ (a, b] ∧ x ∈ A. Let p = O(εt). For each interval (ip, (i + 1)p]
we can, deterministically in eO( 1
where i ∈n0, . . .j t
interval, having in mind that the answer is approximate. The number of intervals is O( 1
ε ).
for interval (ip, (i + 1)p].
pko we query oracles whether S(Z1, t) and S(Z2, t) contain some element in the
We store the answers in arrays S1 and S2, namely Sj[i] = 1 if the oracle for S(Zj, t) answers yes
1[i] =(1 if the oracle for
0 otherwise
S′
(ip, (i + 1)p] ⊓ S1 or i = 0
Then we perform a fast convolution on S1, S2 with FFT.
x = x1 + x2. We have (S1 ⊕FFT S2)[k] =Pk
If x ∈ S(Z1 ∪ Z2, t)∩ (kp, (k + 1)p], then there is some x1 ∈ S(Z1, t) and x2 ∈ S(Z2, t) such that
i=0 S1[i] · S2[k − i] and thus (S1 ⊕FFT S2)[k′] is nonzero
for k′ = k or k′ = k + 1. This defines the rule for the new oracle. The additive error of the oracle
gets doubled with the summation. On the other hand, if one of these fields is nonzero, then there
are corresponding indices i1, i2 summing to k or k + 1. The second condition from Definition 4.5
allows the corresponding value x1 to lie within one of the intervals with indices i1 − 1, i1, or i1 + 1
and likewise for x2. Therefore, the additive error is O(p) = O(εt).
iff there is i such that (ip, (i + 1)p] ⊓ S1 ∪ {0} and ((k − i)p, (k − i + 1)p] ⊓ S2 ∪ {0}.
Now, we promised only oracle output to our array. When a query comes, we scale down the
query interval, then we check whether any of adjacent interval in our structure is nonzero (we lose
a constant factor of O(ε) accuracy here) and output yes if we found it and no otherwise.
Moreover, with additional polylogarithmic factors we can also retrieve the solution. The idea is
similar to backtracking from [46]. Namely, the fast convolution algorithm can compute the table of
witnessing indexes (of only one). We store a witnessing index if there is solution and −1 otherwise.
Then we ask the oracles of S(Z1, t) and S(Z2, t) for a solution with a proper indexes and return the
combination of those.
13
4.4 From Exact Solution to ε-close Instance
In Section 2 we presented an overall approach of rounding elements and explained why it gives us
the weak approximation of Subset Sum. Here we will focus on formally proving these claims.
In our subroutines, we round down the items, execute the exact algorithm on the rounded
instance, and retrieve the solution. We want to argue, that in the end we lose only an additive
factor of ±εt. We presented a sketch of this reasoning in Fact 2.3. For our purposes we will describe
the procedure in the case, when the number of items in any solution is bounded by k (i.e., we are
interested only in S(Z, t)k). We can always assume k ≤ n.
Lemma 4.7. Given an exact algorithm that outputs the set S(Z, t)k and works in time T (n, t), where
n = Z, we can construct an (ε, t)-membership-oracle of set S(Z, t)k in time eO(n + T (n, k/ε)).
If the exact algorithm retrieves solution in eO(1) time, then so does the oracle.
Proof. For sake of legibility, we assume that we are interested in S(Z, t)k−1 - this only allows us to
write simpler formulas. Let (zi) denote the items. We perform rounding in the following way:
z′
εt(cid:23) ,
i =(cid:22) kzi
t′ =(cid:22) kt
εt(cid:23) =(cid:22) k
ε(cid:23) .
We run the exact algorithm on the rounded instance (Z ′, t′). It takes time T (n, t′) = O(T (n, k/ε)).
This algorithm returns S(Z ′, t′)k−1, which we store in array Q[1, t′]. We construct (ε, t)-membership-
oracle in array Q′[1, t′] as follows: we set Q′[i] = 1 iff Q contains 1 in range (i− 2k, i + k]. If we want
to be to able retrieve a solution, we need to also remember a particular index j(i) ∈ (i − 2k, i + k]
such that Q[j(i)] = 1. Such a data structure can be constructed in a linear time with a help of
a queue. Given a query q, the oracle returns Q′[q′], where q′ = j kq
Definition 4.5 is satisfied.
εtk. It remains to prove that
Let I ⊆ Z be a set of at most k − 1 items and I ′ be the set of their counterparts after rounding.
Since for all zi ∈ Z it holds
we obtain
kzi
εt − 1 ≤ z′
i ≤
kzi
εt
,
k · Σ(I)
εt
− k + 1 ≤ Σ(I ′) ≤
k · Σ(I ′)
εt
.
(1)
Therefore, if Σ(I) ∈ [q − εt, q + εt], then
kq
εt − 2k + 1 =
Σ(I ′) ≤
εt
k · (q + εt)
εt
k · (q − εt)
kq
εt
=
− k + 1 ≤ Σ(I ′),
+ k,
and Σ(I ′) ∈ (q′ − 2k, q′ + k], because Σ(I ′) is integer. On the other hand, we can invert relation (1)
to obtain
εt
k · Σ(I ′) ≤ Σ(I) ≤
εt
k ·(cid:0)Σ(I ′) + k − 1(cid:1) .
14
To satisfy the second condition we assume Σ(I ′) ∈ (q′ − 2k, q′ + k] and check that
q − 2εt =
εt
k ·(cid:18) kq
εt − 2k(cid:19) ≤
εt
k · (q′ + 2k − 1) ≤ q + 2εt,
εt
k · (q′ − 2k + 1) ≤ Σ(I),
what finishes the proof.
Σ(I) ≤
We apply Lemma B.2 with {z1, . . . , zk} = Y , q = t/2 and k and ε as in the statement.
guarantees that:
It
(1 − ε)Σ(Y ) ≤
εt
2k
Σ(Y ′).
And finally, (1 − ε)Σ(Y ) ≥ Σ(Y ) − εt (because we are only interested in solutions smaller than
t). So if Y is a optimal solution, then an exact algorithm after rounding would return something
greater or equal Σ(Y ) − εt.
Conversely, it can turn out that an exact algorithm would find something with a sum greater
than q (this is where we can violate the hard constraint). We need to bound it as well (because
the definition of (ε, t)-membership-oracle requires that). Note, that analogous argument proves it.
Namely, the solution can consist of at most k items and each of them lose only O(εt/k). Moreover,
exact oracle gave us only the solution that its rounded version sums up to exactly t′. Formally,
we prove it again with Lemma B.2 with the same parameters as before. By dividing both sides by
(1 − ε) we know that:
kXi=1(cid:22) 2kzi
tε (cid:23) = ⌊k⌋ · ε.
Once again, we can use Lemma B.2 with the same parameters (we divided both sides by (1−ε) >
0):
1 − ε(cid:19) εt
2k
xi ≤(cid:18) 1
kXi=1
tε (cid:23) =(cid:18) 1
kXi=1(cid:22) 2kzi
1 − ε(cid:19) εt
tε (cid:23) .
kXi=1(cid:22) 2kzi
2k(cid:22) 2k
ε (cid:23) ≤
1
1 − ε
t < (1 + 2ε)t.
The right side satisfies:
1 − ε(cid:19) εt
(cid:18) 1
2k
(we can always rescale the approximation factor by setting ε′ = ε/2 at the beginning).
The constant before ε does not change much since we only need (O(ε), t)-membership-oracle
The main obstacle with returning a solution that obeys the capacity constraint comes from
the above lemma. If we could provide a reduction from an exact algorithm without widening the
interval [q − εt, q + εt], this would automatically entail a strong approximation for Subset Sum.
This seems unlikely due to conditional hardness result for a strong subquadratic approximation for
Subset Sum [18].
At the end, we need to prove, that an (ε, t)-membership-oracle gives us the correct solution for
weak (1 − ε)-approximation Subset Sum.
15
Lemma 4.8. Given an ( ε
6 , t)-membership-oracle of S(Z, t), we can read the answer to the weak
(1 − ε)-approximation Subset Sum in time eO( 1
Proof. We query the oracle for q = i · εt
the interval [q − εt
[x − εt
Otherwise the witness might belong to (t, (1 + ε
2 ]. If OPT < (1 − ε
6 , q + εt
2 , x + εt
ε ).
2 )t].
6 for i = 0, . . . , 6
6 ] contains an x ∈ S(Z, t), then the oracle returns an element within
2 , OPT].
2 )t, then the oracle will return a witness within (OPT − εt
ε . Each query takes time eO(1) and if
By taking advantage of Lemma 4.3, we can assume that OPT ≥ t/2, therefore the relative error
gets bounded with respect to OPT.
5 The weak (1 − ε)-approximation algorithm for Subset Sum
5.1 Large Items
We will use Theorem 2.4 to compute S(Zlarge, t) on a large instance. On that instance, this algorithm
is more efficient than Kellerer et al. [46] algorithm because one can round items less aggressively.
Lemma 5.1 (Algorithm for Large Items). Given a large instance (Zlarge, t) of Subset Sum (i.e., all
items are greater than γt), we can construct an (ε, t)-membership-oracle of S(Zlarge, t) in randomized
γε ) time with a constant probability of success.
eO(n + 1
Proof. We use Bringmann's [17] algorithm, namely Corollary 2.5, that solves the Subset Sum
problem exactly. Since all elements are greater than γt, any subset that sums up to at most t
must contain at most 1
γ items. The parameter k in Lemma 4.7 is an upper bound on number of
elements in the solution, hence we set k = 1
γ . The Bringmann's [17] algorithm runs in time eO(n + t)
and Lemma 4.7 guarantees that we can build an (ε, t)-membership-oracle in time eO(n + k/ε) =
eO(n + 1/(γε)), which is what we needed.
5.2 Small Items
Now we need an algorithm that solves the problem for small items. As mentioned in Section 2 we will
consider two cases depending on the density of instance. The initial Subset Sum instance consists
of n elements. The m is the number of elements in the small instance and let m′ = O(m log n) be
as in Lemma 4.4. For now we will assume, that the set of elements is (εt/m′)-distinct (we will deal
with multiplicities 2 in Lemma 5.7).
ε
Let q = εt/m′ be the rounding parameter (the value by which we divide) and ℓ = γm′/ε =
) be the upper bound on items' sizes in the small instance after rounding. Parameter
m2 ) describes the boundaries of Theorem 2.6. We deliberately use O notation to hide
O( γm log n
L = O(Σ(S)·
constant factors (note that Galil and Margalit [27] algorithm requires that m > 1000 · √l log l).
l
Lemma 5.2 (Small items and m2 < ℓ log2 ℓ). Suppose we are given an instance (Zsmall, t) of
Subset Sum (i.e., all items are smaller than γt) with size satisfying m2 < ℓ log2 ℓ. Then we can
compute (ε, t)-membership-oracle of S(Zsmall, t) in randomized eO(m + γ
Proof. In here we need to deal with the case, where small instance is sparse. So just as in the proof
of Lemma 5.1, we can use Bringmann's [17] algorithm.
ε2 ) time.
16
S(Zsmall, L) in time eO(m + mγ2
ε2 ).
Proof. We round down items with rounding parameter q = εt/m′ = Ω(
small) · cℓ
rounded items as Z ′
Σ(Zsmall) with Σ(Z ′
small. After scaling down we have L′ = Σ(Z ′
small) and ℓ remains the same). Recall that ℓ = O( mγ log n
ε
εt
The total sum of items in Z ′
m log n ) and denote the set of
m2 (note that we only replace
).
ε ).
Combining Corollary 2.5 and Lemma 4.7 allows us to construct an (ε, t)-membership-oracle in
We will use the reduction from exact to weak (1− ε)-approximation algorithm for Subset Sum
from Lemma 4.7. We set m as the maximal number of items in the solution, as there are at most
ε ). After dividing both
m small items. Recall that ℓ is eO(mγ/ε). This gives us m2 = eO(ℓ) = eO( mγ
sides by m we obtain m = eO( γ
eO(m + T (m, m/ε)) = eO(m + γ
Now we have to handle the harder m2 ≥ ℓ log2 ℓ case. In this situation we again consider two
cases. The Galil and Margalit [27] algorithm allows only to ask queries in the range (L, Σ(S) − L)
where L = O(Σ(S)·
m2 ). In the next lemma we take care of ranges [0, L] and [Σ(S)− L, Σ(S)]. We
focus on the range [0, L], because the sums within [Σ(S) − L, Σ(S)] are symmetric to [0, L].
Lemma 5.3 (Small items, range (0, L)). Given an instance (Zsmall, t) of Subset Sum, such that
Zsmall = m and the items' sizes are at most γt, we can compute an (ε, t)-membership-oracle for
ε2 ) randomized time.
l
size at most ℓ). Hence ,L′ = O(ℓ2/m) = O( γ2m log2 n
in time eO(m + L′) = eO(m + mγ2
(ε, t)-membership-oracle for S(Zsmall, L).
small is smaller or equal to ℓm (because there are m elements of
). Therefore Bringmann's [17] algorithm runs
ε2 ). Combining it with the analysis of the Lemma 4.7 gives us an
ε2
5.3 Applying Additive Combinatorics
Before we proceed forward, we need to present the full theorem of Galil and Margalit [26, Theorem
6.1] (in Section 2.4 we presented only a short version to keep it free from technicalities). We need
a full running time complexity (with dependence on ℓ, m, Σ(S)). We copied it in here with a slight
change of notation (e.g., [26] use SA but we use notation from [48] paper of Σ(A)).
Theorem 5.4 (Theorem 6.1 from [26]). Let A be a set of m different numbers in interval (0, ℓ] such
that
m > 1000 · ℓ0.5 log2 ℓ;
then we can build in O(cid:16)m + ((ℓ/m) log l)2 + Σ(A)
m2 ℓ0.5 log2 ℓ(cid:17) preprocessing time a structure which
allows us to solve the Subset Sum problem for any given integer N in the interval (L, Σ(A) −
L). Solving means finding a subset B ⊆ A, such that Σ(B) ≤ N and there is no subset C ⊆ A
such that Σ(B) < Σ(C) ≤ N . An optimal subset B is build in O(log ℓ) time per target number
and is listed in time O(B). For finding the optimal sum Σ(B) only, the preprocessing time is
O(cid:16)m + ((ℓ/m) log ℓ)2(cid:17) and only constant time is needed per target number.
, however in the next [27] the authors improved it
to L := O(Σ(A) ℓ
m2 ) without any damage on running time [28]. For both of these possible choices
of L we obtain a subquadratic algorithm. We will use the improved version [27] because it provides
a better running time.
In [26] authors defined L := 100·Σ(A)ℓ0.5 log2 ℓ
m
17
Lemma 5.5 (Small items, range (L, Σ(S)− L)). Given a small instance (Zsmall, t) of Subset Sum
(i.e., all items are smaller than γt) such that Zsmall is (εt/m′)-distinct (where m′ = O(m log n)), we
ε(cid:1)2 +
can compute an (ε, t)-membership-oracle of S(Zsmall, t) ∩ (L, Σ(Zsmall) − L) in time eO(n +(cid:0) γ
ε ·(cid:0) γn
Proof. We round items to multiplicities of q = εt/m′. Precisely:
ε (cid:1)0.5).
γ
z′
i =(cid:22) zi
q(cid:23) ,
t′ =(cid:22) t
q(cid:23) =(cid:22) m′
ε (cid:23) .
We know that zi < γt. Therefore
z′
i ≤
zi
q
<
γt
q
=
γm′
ε
= ℓ.
By the same inequalities as in the proof of Lemma 4.7 we know that if we compute S(Z ′
and multiply all results by q, we obtain an (ε, t)-membership-oracle for S(Zsmall, t).
small, t′)
Checking conditions of the algorithm Now we will check that we satisfy all assumptions of
small. First note that m2 < ℓ log2 ℓ, ℓ
Galil and Margalit [27] algorithm on the rounded instance Z ′
is the upper bound on the items' sizes in Z ′
small are distinct
because we assumed that Zsmall is (εt/m′)-distinct.
small, and we know that all items in Z ′
Preprocessing Next Galil and Margalit [27] algorithm constructs a data structure on the set of
rounded items Z ′
small. The preprocessing of Galil and Margalit [27] algorithm requires
O(cid:18)m + (ℓ/m log ℓ)2 +
Σ(Z ′
small)
m2
ℓ0.5 log2 ℓ(cid:19)
time. If we put it in terms of m, ε, t and hide polylogarithmic factors we see that preprocessing runs
in:
+
γ
ε (cid:17)0.5(cid:19)
ε(cid:16) γm
because Σ(Z ′
small) ≤ ℓm.
eO(cid:18)m +(cid:16) γ
ε(cid:17)2
L). After scaling down we have L′ = eO(cid:0)Σ(Z ′
Naively, one could run queries for all elements in range (L′, Σ(Z ′
Queries With this data structure we need to compute a set ε-close to S(Zsmall, t)∩(L, Σ(Z ′
small)−
small) − L′) and check if there is
a subset of Z ′
small that sums up to the query value. However this is too expensive. In order to deal
with this issue, we take advantage of the fact that each query returns the closest set whose sum is
smaller or equal to the query value.
mε2 ) = eO( mγ2
m ) = eO( γ2m2
m2(cid:1) = eO( ℓ2
small) ·
ε2 ).
ℓ
Since we have rounded down items with q = εt
m′ , we only need to ask εt
to learn sufficient information. The queries will reveal if Zsmall contains at least one element in each
range [iεt, (i + 1)εt), what matches the definition of the (ε, t)-membership-oracle.
q = eO(m) queries in order
18
Retrieving the solution Galil and Margalit [27] algorithm can retrieve the solution in time
O(log ℓ).
by the preprocessing time.
This finalizes the construction of the (ε, t)-membership-oracle. The running time is dominated
5.4 Combining the Algorithms
Now we will combine the algorithms for small items.
Lemma 5.6 (Small Items). Given a (Zsmall, t) instance of Subset Sum (i.e., all elements in
Zsmall are smaller than γt), such that the set Zsmall is (εt/m)-distinct, we can compute an (ε, t)-
membership-oracle of S(Zsmall, t) in time eO(m + γ
Proof. We will combine two cases:
ε2 + mγ2
ε2 ) with high probability.
Case When m2 < ℓ log2 ℓ:
In such case we use Lemma 5.2 that works in eO(m + γ
ε2 ) time.
ε ·(cid:0) γm
ε(cid:1)2 + γ
ε (cid:1)0.5) time.
Case When m2 ≥ ℓ log2 ℓ: First we take advantage of Lemma 5.5. This gives us an (ε, t)-
It requires
membership-oracle that answers queries within set S(Zsmall, t) ∩ (L, Σ(Zsmall) − L).
oracle for interval set [Σ(Zsmall) − L, Σ(Zsmall)] is obtained by symmetry.
We combine it (using Lemma 4.6) with the (ε, t)-membership-oracle that gives us answers to a
ε2 ). The
eO(nm +(cid:0) γ
set S(Zsmall, t) ∩ [0, L] from Lemma 5.3. This oracle can be constructed in time eO(m + mγ2
Running Time: The running time of merging the solutions from Lemma 4.6 is eO(1/ε) which is
ε(cid:1)2) is suppressed by
suppressed by the running time of Lemma 5.3 and Lemma 5.5. Factor eO((cid:0) γ
eO(cid:16)(cid:16) mγ2
ε2 (cid:17)(cid:17).
ε ·(cid:0) γm
ε (cid:1)0.5 is also suppressed by eO( mγ2
ε2 ). The algorithm is randomized because Lemma 5.3
is randomized.
Term γ
The Lemma 4.4 allowed us to partition our instance into small and large items. We additionally
know that each interval of length εt/m′ contains at most 2 items. However in the previous proofs
we assumed there can be only one such item, i.e., the set should be (εt/m′)-distinct.
Lemma 5.7 (From multiple to distinct items). Given an instance (Zsmall, t) of Subset Sum,
where Zsmall = m and zzsmall is (εt/m′, 2)-distinct for m′ = O(m log n), we can compute an
Proof. We divide the set Zsmall into two sets Z 1
sets Z 1
(ε, t)-membership-oracle for instance (Zsmall, t) in eO(n + γ
Zsmall and dividing items into odd-indexed and even-indexed . It takes eO(m) time.
ε2 + nγ2
small such that Zsmall = Z 1
Next we use Lemma 5.6 to compute an (ε, t)-membership-oracle for (Z 1
small, the
small are disjoint, (εt/m′)-distinct, and have size Ω(m). This can be done by sorting
ε2 ) time with high probability.
and merge them using Lemma 4.6.
small, t) and (Z 2
small ∪ Z 2
small and Z 2
small,Z 2
small, t),
19
Now we will combine the solutions for small and large items.
γ
εt
ε2 ) time with high probability.
Theorem 5.8. Let 0 < γ be a trade-off parameter (that depends on n, ε). Given an (Z, t) instance
γε +
of Subset Sum, we can construct the (ε, t)-membership-oracle of instance S(Z, t) in eO(n + 1
ε2 + nγ2
Proof. We start with Lemma 4.4, that in O(n log2 n) time partitions the set into Zlarge and Zsmall,
m log n , 2)-distinct, where m = Zsmall. To deal with small items, we use
such that Zsmall is (
Lemma 5.7. The algorithm for small items returns an (ε, t)-membership-oracle of S(Zsmall, t). For
large items we can use Lemma 5.1. It also returns an (ε, t)-membership-oracle of S(Zlarge, t).
Finally, we use Lemma 4.6 to merge these oracles in time eO(1/ε). All the subroutines run with
a constant probability of success.
Finally, we have combined all the pieces and we can get a faster algorithm for weak (1 − ε)-
approximation for Subset Sum.
Corollary 5.9 (Subset Sum with tradeoff). There is a randomized weak (1 − ε)-approximation
ε2 ) time with high probability for any
ε2 + nγ2
γε + γ
γ(n, ε) > 0.
algorithm for Subset Sum running in eO(n + 1
Proof. It follows from Lemma 4.8 and Theorem 5.8.
γε + γ
ε2 + nγ2
The weak (1 − ε)-approximation Subset Sum gives us the approximation for Partition via
Corollary 2.2.
Corollary 5.10 (Partition with trade-off). There is a randomized (1−ε)-approximation algorithm
ε2 ) time with high probability for any γ(n, ε) > 0.
for Partition running in eO(n + 1
To get running time of form eO(n+1/εc) and prove our main result we need to reduce the number
of items from n to eO(1/ε) and choose the optimal γ.
algorithm for Subset Sum running in eO(cid:0)n + ε− 5
Proof. We apply Lemma 4.2 to ensure that the number of items is eO(cid:0) 1
Theorem 5.11 (Weak apx for Subset Sum). There is a randomized weak (1 − ε)-approximation
ε(cid:1) and work with an O(ε)-close
instance. Then we take advantage of Corollary 5.9 with γ = ε
3(cid:1) time.
3 .
2
Analogously for Partition we get that:
Theorem 5.12 (Apx for Partition). There is a randomized (1 − ε)-approximation algorithm for
Partition running in eO(cid:0)n + ε− 5
3(cid:1) time.
6 Approximate (min, +)-convolution
Approximate (min, +)-convolution
Input: Sequences A[0, . . . , n − 1], B[0, . . . , n − 1] of positive integers and approximation pa-
rameter 0 < ε < 1
Task: Let OPT[k] = min0≤i≤k(A[i] + B[k − i]) be the (min, +)-convolution of A and B. Find
a sequence C[0, . . . , n − 1] such that ∀i OPT[i] ≤ C[i] ≤ (1 + ε)OPT[i]
20
Backurs et al. [10] described a (1 + ε)-approximation algorithm for (min, +)-convolution, that
ǫ2 log n log2 W ). In their paper [10] it is used as a building block
runs deterministically in time O( n
to show a near-linear time approximation algorithm for TreeSparsity. With the approximation
algorithm for (min, +)-convolution, they managed to solve TreeSparsity approximately in eO( n
ε2 )
time, which in practical applications may be faster than solving this problem exactly in time eO(n2).
We begin with explaining its connection with the Subset Sum problem. A natural generaliza-
tion of Subset Sum is Knapsack. In this scheme each item has value and weight and our task
is to pack items into the knapsack of capacity C, so that their cumulative weight does not exceed
capacity and in the same time we want to maximize their total value. In the special case when all
weights and values are equal we obtain the Subset Sum problem.
To certify the existence of a subset with a given sum, we have used the fast convolution using
FFT as subroutine.
If we want to generalize it and capture maximal value subset of items of a
given weight, we require (max, +)-convolution, which is computationally equivalent to (min, +)-
convolution.
Cygan et al. [24] exploited this idea to show subquadratic equivalence between exact (min, +)-
convolution, Knapsack, and other problems. Here we focus on the approximate setting. From [24]
ε1.99 ) approximation algorithm for Unbounded Knapsack is unlikely.
ε2 ) FPTAS for Unbounded
it follows that the eO(n + 1
This lower bounds proves the optimality of Jansen and Kraft [40] eO(n+ 1
Knapsack. The current best FPTAS for Knapsack is burdened with time complexity of eO(n +
In this section we improve upon the eO(n/ε2) approximation algorithm. for (min, +)-convolution.
Similar techniques have been exploited to obtain the eO(nω/ε)-time approximation for APSP [60]
1/ε12/5) Chan [19]. We hope, that our approximation schemes are a step towards a faster FPTAS
for Knapsack.
and they have found use in the approximate pattern matching over l∞ [51]. The basic idea is to
propose a fast exact algorithm depending on W (upper bound on the weights) and apply it after
rounding weights into smaller space. Our result also applies to (max, +)-convolution.
6.1 Exact eO(nW ) algorithm
The (min, +)-convolution admits a brute force O(n2)-algorithm. From the other hand, when all
values in sequences are binary, then applying FFT and performing convolution yields an O(n log n)-
algorithm. Our exact eO(nW ) algorithm is an attempt to capture this trade-off. Note, that this
algorithm is worse than a brute force whenever W > n which is often the case. However, this
algorithm turns out useful for approximation.
Lemma 6.1. The (min, +)-convolution [(max, +)-convolution] problem can be solved deterministi-
cally in O(nW log (nW )) time and O(nW ) space.
Proof. Given sequences A[0, . . . , n − 1] and B[0, . . . , n − 1] with values at most W , we transform
them into binary sequences of length 2nW . We encode every number in the natural unary manner.
For 0 ≤ i < n, 1 ≤ k ≤ W we define:
a[2W i + k] =(0 if A[i] 6= k
1 if A[i] = k
and similarly we define sequence B. For example, sequence (2, 3, 1) with W = 3 gets encoded as
010'000'001'000'100'000 (the separators ' are used to visually separate sections of length W ).
21
We compute convolution C = A⊕ B using FFT in time O(nW log n log W ). Since C[2W i + k] =
P i1+i2=i
A[2W i1 + k1] · B[2W i2 + k2], the first nonzero occurrence in the i-th block of length 2W
encodes the value of the i-th element of the requested (min, +)-convolution. If we are interested in
computing (max, +)-convolution, we should similarly seek for last nonzero value in each block.
k1+k2=k
The time complexity is dominated by performing convolution with FFT. As the additional space
we need O(nW ) bits for the transformed sequences.
6.2 Approximating Algorithm
We start with a lemma inspired by [60, Lemma 5.1] and [51, Lemma 1].
Lemma 6.2. For natural numbers x, y and positive q, ε satisfying q ≤ x + y and 0 < ε < 1 it holds:
x + y ≤(cid:16)l 2x
(x + y)(1 − ǫ) <(cid:16)j 2x
qǫm +l 2y
qǫk +j 2y
qǫm(cid:17) qǫ
qǫk(cid:17) qǫ
2 ≤ x + y.
Proof. The proof is a special case of Lemmas B.1 and B.2 for k = 2.
2 < (x + y)(1 + ǫ),
Lemma 6.3. Assume the (min, +)-convolution [(max, +)-convolution] can be solved exactly in time
T (n, W ). Then we can approximate (min, +)-convolution [(max, +)-convolution] in time O((T (n, 4
ε )+
n) log W ).
Algorithm 2 ApproximateMinConv(A, B). We use a simplified notation to transform all ele-
ments in the sequences A[i] and B[i].
1: Output[i] = ∞
2: for l = 2⌈log W⌉, . . . , 0 do
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
q := 2l
A′[i] = ⌈ 2A[i]
qε ⌉
if A′[i] > ⌈4/ε⌉ then
A′[i] = ∞
B′[i] = ⌈ 2B[i]
qε ⌉
if B′[i] > ⌈4/ε⌉ then
B′[i] = ∞
end if
end if
V = runExact(A′, B′)
if V [i] < ∞ then
end if
Output[i] = V [i] · qε
2
15:
16: end for
17: return Output[0, . . . , n − 1]
Proof. The idea is based on [51, Section 6.2]. We focus on the variant with (min, +)-convolution,
however the proofs works alike for (max, +)-convolution.
22
array by qǫ
We iterate the precision parameter q through 2W, W, . . . , 4, 2, 1. In each iteration we apply the
2 and store them in the output array C, possibly overwriting some elements.
qǫ(cid:7)) to all elements in A, B, we set ∞ for each value exceeding
transform from Lemma 6.2 (x →(cid:6) 2x
(cid:6) 4
ε(cid:7), and launch the exact algorithm on such input. We multiply all finite elements in the returned
m are at most (cid:6) 4
parameter q such that q ≤ C[k] < 2q. The rounded numbers l 2A[i]
ε(cid:7),
Assume the correct value of C[k] equals A[i] + B[k − i]. For some iteration we get the precision
so we will update the k-th index in the output array. On the other hand, the assumption of
Lemma 6.2 is satisfied, therefore the generated value lies between C[k] and C[k](1 + ε).
In the
following iterations, we will still have q ≤ C[k], therefore any further updates to the k-th index will
remain valid.
The algorithm performs O(log W ) iterations and in each step we run the exact algorithm in
ǫ ), thanks to the pruning procedure. Transforming the sequences takes O(n) time in
qǫ m,l 2B[k−i]
qǫ
time T (n, 4
each step.
Theorem 6.4 (Apx for (min / max, +)-conv). There is a deterministic algorithm for (1 + ε)-
approximate (min, +)-convolution [(max, +)-convolution] running in O(cid:0) n
Proof. From Lemma 6.1 the running time of exact algorithm is T (n, W ) = O(nW log n log W ). This
quantity dominates the additive term O(n log W ). Hence by replacing each W with 1/ε we get the
claimed running time.
ε log ( n
ε ) log W(cid:1) time.
7 Tree Sparsity
running time for (1 + ε)-tail approximation.
The TreeSparsity problem has been stated as follows: given a node-weighted binary tree and
an integer k, find a rooted subtree of size k with the maximal weight. Its approximation version
comes with two flavors: as a head approximation where we are supposed to maximize the weight
of the solution, and as a tail approximation where we minimize the total weight of nodes that
do not belong to the solution. Note that a constant approximation for one of the variants does
not necessarily yield a constant approximation for the other one. Backurs et al. [10] proposed an
for approximating (min, +) and (max, +) convolutions. Our construction is based on the approach
by Cygan et al. [24] which also results in a simpler analysis than for the previously known approxi-
mation schema [10]. In particular, a single proof suffices to cover both head and tail variants.
ε2 · log12 n · log2 W(cid:1) running time for (1 − ε)-head approximation, and an O(cid:0) n
O(cid:0) n
In this section we improve the running times for both variants relying on the eO(cid:0) n
ε · log(n/ε)· log3 n· log W(cid:1)-time algorithm that computes the maximal weights of rooted subtrees
O(cid:0) n
(1 + ε)-approximate TreeSparsity can be solved in time O(cid:0)(cid:0)n + T (n, W, ε/ log2 n)(cid:1) log n(cid:1).
ε3 · log9 n · log3 W(cid:1)
ε(cid:1) algorithm
Theorem 7.1. If (1 + ε)-approximate (min, +)-convolution can be solved in time T (n, W, ε), then
The following theorem, combined with our approximation for (min, +)-convolution yields an
for each size k = 1, . . . , n with a relative error at most ε in both head and tail variant.
Proof. We exploit the heavy-light decomposition introduced by Sleator and Tarjan [57]. This tech-
nique has been utilized by Backurs et al. [10] in their work on TreeSparsity approximation and
later by Cygan et al. [24] in order to show a subquadratic equivalence between TreeSparsity and
(min, +)-convolution.
23
We construct a spine with a head s1 at the root of the tree. We define si+1 to be the child of
si with the larger subtree (in case of draw we choose any child) and the last node in the spine is a
leaf. The remaining children of nodes si become heads for analogous spines so the whole tree gets
covered. Observe that every path from a leaf to the root intersects at most log n spines because
each spine transition doubles the subtree size.
At first we express the head variant in the convolutional paradigm. For a node v with a subtree
of size m we define the sparsity vector (xv[0], xv[1], . . . , xv[m]) of weights of the heaviest subtrees
rooted at v with fixed sizes. This vector equals the (max, +)-convolution of the sparsity vectors
for the children of v. We are going to compute sparsity vectors for all heads of spines in the tree
recursively. Having this performed we can read the solution from a sparsity vector of the root. Let
(si)ℓ
i=1 be a spine with a head v and let ui indicate the sparsity vector for the child of si being
a head (i.e., the child with the smaller subtree). If si has less than two children we treat ui as a
vector (0).
For an interval [a, b] ⊆ [1, ℓ] let ua,b = ua⊕max ua+1⊕max···⊕max ub and ya,b[k] be the maximum
2 (cid:5). The ⊕max
weight of a subtree of size k rooted at sa and not containing sb+1. Let c = (cid:4) a+b
operator is associative so ua,b = ua,c ⊕max uc+1,b. To compute the second vector we consider two
cases: whether the optimal subtree contains sc+1 or not.
ya,b[k] = max(cid:20)ya,c[k],
= max(cid:20)ya,c[k],
cXi=a
cXi=a
max
x(si) +
k1+k2=k−(c−a+1)(cid:16)ua,c[k1] + yc+1,b[k2](cid:17)(cid:21)
x(si) +(cid:16)ua,c ⊕max yc+1,b(cid:17)(cid:2)k − (c − a + 1)(cid:3)(cid:21)
(2)
2 ] and [ ℓ
Using the presented formulas we reduce the problem of computing xv = y1,ℓ to subproblems for
intervals [1, ℓ
2 + 1, ℓ] and results are merged with two (max, +)-convolutions. Proceeding
further we obtain log ℓ levels of recursion. Since there are O(log n) spines on a path from a leaf
to the root, the whole computation tree has O(log2 n) layers, each node being expressed as a pair
of convolutions on vectors from its children. Each vertex of the graph occurs in at most log n
convolutions so the sum of convolution sizes is O(n log n).
In order to deal with the tail variant we consider a dual sparsity vector (xv[0], xv[1], . . . , xv[m]),
where xv[i] stands for the total weight of the subtree rooted at v minus xv[i]. The dual sparsity
vector of v equals the (min, +)-convolution of the vectors for the children of v. We can use an analog
of equation (2) and also express the problem as a computation tree based on convolutions.
We take advantage of Theorem 6.4 to perform each convolution with a relative error δ. The
formula (2) contains an additive termPc
i=a x(si) but this can only decrease the relative error. The
cumulative relative error is bounded by (1− δ)log2 n for head approximation and (1 + δ)log2 n for tail
approximation, therefore setting δ = Θ(ε/ log2 n) guarantees that the sparsity vector for the root is
burdened with relative error at most ε.
The sum of running times for all convolutions is O(cid:0)T (n, W, δ) log n(cid:1), what gives the postulated
running time for the whole algorithm. In order to retrieve the solution for a given k, we need to
find the pair of indices that produced the value of the k-th index of the last convolution. Then we
proceed recursively and traverse back the computation tree. Since finding arg max and arg min can
be performed in linear time, the total time of analyzing all convolutions is O(n log n).
24
8
In the abstract we have claimed that our result for Partition constitutes the first approximation
algorithm for NP-hard problem that breaks the quadratic barrier. However this is not necessary the
eO(n + 1/ε) approximation algorithm for 3SUM
case for the problems in P. In this section we will show an eO(n + 1/ε) approximation algorithm for
3SUM and prove accompanying lower bound under a reasonable assumption. To the best of our
knowledge, this is also the first nontrivial linear approximation algorithm for a natural problem.
k-SUM
Input: Sets A1, A2, . . . , Ak−1, S, each with cardinality at most n.
Task: Decide if there is a tuple (a1, . . . , ak−1, s) ∈ A1×. . .×Ak−1×S such that a1+. . .+ak−1 =
s.
The 3SUM problem is a special case of k-SUM for k = 3. The 3SUM is one of the most notori-
ous problems with a quadratic running time and has been widely accepted as a hardness assumption
(see [58] for overview). The fastest known algorithm for 3SUM is slightly subquadratic: Jørgensen
and Pettie [41] gave an O(n2(log log n/ log n)2/3)-time deterministic algorithm and then indepen-
dently Freund [25] and Gold and Sharir [34] improved this result by presenting an O(n2 log log n/ log n)-
time algorithm.
eO( n
algorithm for 3SUM. In this section we show how to solve 3SUM approximately in time eO(n + 1/ε)
The approximation variant for 3SUM was considered by Gfeller [33] who showed a deterministic
If we are not
ε ) algorithm as a byproduct of finding longest approximate periodic patterns.
interested in exact solution, the Gfeller [33] algorithm is polynomially faster than the best exact
time and prove this tight up to the polylogarithmic factors.
Approximate 3SUM ([33])
Input: Three sets A, B, C of positive integers, each with cardinality at most n.
Task: The algorithm:
• concludes that no triple (a, b, c) ∈ A × B × C with a + b = c exists, or
• it outputs a triple (a, b, c) ∈ A × B × C with a + b ∈ [c/(1 + ε), c(1 + ε)].
This definition generalizes to k-SUM, however we are unaware about any previous works on
approximate k-SUM.
8.1 Faster approximation algorithm for 3SUM
approximation algorithm. In the Section 9 we will show a conditional optimality of this result.
technique from Section 6, where we gave the fast approximation algorithm for (min, +)-convolution.
In this section we present an eO(n + 1/ε)-time approximation scheme for 3SUM problem. We use a
As previously, we start with a fast eO(n + W ) exact algorithm and then utilize rounding to get an
8.1.1 Exact eO(n + W ) algorithm for 3SUM
Let W denote the upper bound on the integers in the sets A,B and C. The exact eO(n + W )-time
algorithm for 3SUM is already well known [22, 20]. In here we will place the proof for completeness.
For formal reasons we need to take care of the special symbol ∞. What is more, we will generalize
this result to k-SUM.
25
Theorem 8.1 (Based on [22, 20]). The k-SUM can be solved deterministically in eO(kn+kW log W )
time and eO(kn + W ) space.
Proof. We will encode the numbers in the sets as binary arrays of size O(W ) and iteratively perform
fast convolution using FFT. Because we will use only O(1) tables at once, the space complexity will
not depend on k. At the end we will need to check if any entry in the final array is in S.
Encoding: We iterate for every set A1, . . . , Ak−1 and for l-th iteration encode it as a binary vector
V of length W + 1, such that:
Vl[i] =(1 iff t ∈ Al
0 otherwise
to save space we will use only one Vl vector at the time. The encoding can be done in O(n + W )
time. If the special symbol ∞ ∈ Al appears then we simply discard it.
FFT: We want to perform a convolution with FFT on all vectors Vl. We do it one at a time
and discard all elements larger than W . Let Ul be the result of up to l-th iteration. We know
that the proper polynomial is Ul(x) =P(a1,...,al)∈(A1×...×Al) xa1+...al. And if we multiply it by the
polynomial Vl+1 =Pal+1∈Al+1
xal+1, we get Ul+1(x) =P(a1,...,al+1)∈(A1×...×Al+1) xa1+...al+1.
Hence at the end we obtain the vector Vk−1 that encodes all the sums of elements in subsets
truncated up to W place.
Comparing At the end we need to get the binary vector for S and compare it with the resulting
vector Vk−1.
Time and Space We did k iterations. In each of them we transformed a set into a vector in
time O(n). The fast convolution works in O(T log T ) by using FFT. Hence, the running time is
O(kn + kW log W ). Algorithm needs O(nk) space to encode input and O(W ) space to store binary
vectors.
8.1.2 Approximation algorithm
Next we will use an exact algorithm to propose the fast approximation. We will use the same
reasoning as in Section 6.4.
Lemma 8.2. Assume the k-SUM can be solved exactly in T (n, k, W ) time. Then approximate
k-SUM can be solved in O((T (n, k, k/ε) + nk) log W ) time.
Because the proof is just a small modification of the Lemma 6.3 we have included it in Sec-
tion 8.1.3. At the end we need to connect the exact algorithm from Lemma 8.1 and the reduction
from Lemma 8.2.
ε log k
ε log W ) time.
Theorem 8.3. There is a deterministic algorithm for (1 + ε)-approximate k-SUM running in
O(nk log W + k2
Proof. From Lemma 8.1 the running time of k-SUM is T (n, k, W ) = O(nk + kW log W ). Applying
this running time to the reduction in Lemma 8.2 results in the claimed running time, because the
O(nk) term is dominated by O(nk log W ) term in the reduction.
26
To get an approximate algorithm for 3SUM we set k = 3.
Corollary 8.4. The approximate 3SUM can be solved deterministically in O((n + 1
time.
ε log 1
ε ) log W ))
8.1.3 Proof of Lemma 8.2
Algorithm 3 ApproximateKSum(a1, a2, . . . , ak−1, s, ε). We use a shorten notation to transform
all elements in the sequences al[i] and s[i].
1: Output[i] = ∞
2: for l = 2⌈log W⌉, . . . , 0 do
q := 2l
for l = 1 . . . k − 1 do
qε m
l[i] =l kal[i]
a′
if a′
l[i] > ⌈4k/ε⌉ then
a′
l[i] = ∞
end if
end for
qε m
s′[i] =l ks[i]
if s′[i] > ⌈4k/ε⌉ then
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
s′[i] = ∞
end if
if runExactKsum(a′
1, . . . , a′
k−1, s′) then
return True
end if
15:
16:
17: end for
18: return False
Proof. The proof basically follows the approach approximating (min, +)-convolution in Lemma 6.3.
Assume, that there is some number s, for each there exists a tuple (a1, a2, . . . , ak−1) ∈ A1×. . . Ak−1,
i=1 ai < s(1 + ε). Then look at Algorithm 3 in which we iterate precision parameter q.
Hence there is some q, such that q ≤ s < 2q. From Lemma B.1 we know, that then, we can round
the numbers a′
that s <Pk
qεm and then their sum should be approximately:
i =l kai
kXi=1
ai ≤X(cid:24) kai
qε(cid:25) < (1 + ε)
ai
kXi=1
So if there is some number s ∈ S, then ApproximateKSum algorithm would find a tuple, that
sum up to s′ ∈ [s, (1 + ε)s].
From the other hand, if for all s ∈ S no tuple sums up to [s, (1 + ε)s] then our Approxi-
mateKSum can also return YES. It is because before rounding items could sum up to something in
[(1−ε)s, s] (see Section 2). However, if there for such a parameter there always exists a precision pa-
rameter q, that q ≤ s < 2q. Then rounding the numbers according to Lemma B.1 gives only (1± ε)
27
error and they cannot sum tol ks
qεm. Hence if for all s ∈ S no tuple sums up to [(1 − ε)s, (1 + ε)s]
then our ApproximateKSum will return NO.
A technicality is hidden in the Definition 8 we need to return approximation of the form
1+ε ≈ 1 − ε so we can take care of if by
OPT/(1 + ε) ≤ OUR ≤ OPT(1 + ε) but note that
adjusting ε.
1
9 Conditional Lower Bounds
Proving conditional lower bounds in P under a plausible assumption is a very active line of re-
search [58, 23, 2, 3, 4, 1]. One of the first problems with a truly subquadratic running time ruled
out was EditDistance [9]. It admits a linear time approximation algorithm for ε = 1/√n, that
follows from the exact O(n + d2) algorithm. Subsequently, this (linear-time) approximation factor
was improved by Bar-Yossef et al. [12] to n3/7+o(1), then by Batu et al. [14] to n1/3+o(1), and most
recently Andoni et al. [5] proposed an O(n1+ε)-time algorithm with factor (log n)O(1/ε) for every
fixed ε > 0. From the other hand Abboud and Backurs [1] ruled out a truly subquadratic PTAS for
Longest Common Subsequence using circuit lower bounds. Our results are somehow of similar
flavor to this line of research.
9.1 Conditional Lower Bound for Approximate 3SUM
we rule out such a possibility and prove the optimality of Theorem 8.4.
To show the conditional lower bound we will assume the hardness of the exact 3SUM. The
We have shown an approximate algorithm for 3SUM running in eO(n + 1/ε) time. Is this the best
we can hope for? Perhaps one could imagine an eO(n + 1/√ε) time algorithm. In this subsection
3SUM conjecture says, that the eO(n2) algorithm is essentially the best we can hope for up to
Conjecture 9.1 (3SUM conjecture [58]). In the Word RAM model with O(log n) bit words, any al-
gorithm requires Ω(n2−o(1)) time in expectation to determine whether given set S ⊂ {−n3+o(1), . . . , n3+o(1)}
of size n contains three distinct elements a, b, c such that a + b = c.
subpolynomial factors.
This definition of 3SUM in [58] is equivalent to the one in Section C (see discussion in [13]). What
is more, solving 3SUM with only polynomially bounded numbers can be reduced to solving it with
the upper bound W = O(n3) [13]. 3SUM can be solved in subquadratic time when W = O(n2−δ)
via FFT, but doing so assuming only W = O(n2) constitutes a major open problem. Hsu and
Umans [38] have considered it as a yet another hardness assumption.
Conjecture 9.2 (Strong-3SUM conjecture [38]). 3SUM on a set of n integers in the domain of
{−n2, . . . , n2} requires time Ω(n2−o(1)).
Theorem 9.3. Assuming the Strong-3SUM conjecture, there is no eO(n + 1/ε1−δ) algorithm for
(1 + ε)-approximate 3SUM, for any constant δ > 0.
Proof. Consider the exact variant of 3SUM within the domain {−n2, . . . , n2}. We can assume that
the numbers are divided into sets A, B, C and we can restrict ourselves to triples a ∈ A, b ∈ B,
c ∈ C [13]. We add n2 + 1 to all numbers in A ∪ B and likewise 2n2 + 2 to numbers in C to obtain
an equivalent instance with all input numbers greater than 0 and W = O(n2).
28
We can use it to solve the problem above exactly by setting ε = 1
Suppose, that for some small δ > 0 the approximate 3SUM admits an eO(n + 1/ε1−δ)-algorithm.
the exact algorithm is strongly subquadratic, namely eO(n + 1/ε1−δ) = eO(n2−2δ). This contradicts
the Strong-3SUM conjecture.
2 ). The running time of
2W = Ω(n
1
9.2 Conditional Lower Bounds for Knapsack-type Problems
The conditional lower bounds for Knapsack and Unbounded Knapsack are corollaries from [24].
We commence by introducing the main theorem from that work, truncated to problems that are of
interest to us.
Theorem 9.4 (Theorem 2 from [24]). The following statements are equivalent:
1. There exists an O(n2−ε) algorithm for (min, +)-convolution for some ε > 0.
2. There exists an O(cid:0)(n + t)2−ε(cid:1) algorithm for Unbounded Knapsack for some ε > 0.
3. There exists an O(cid:0)(n + t)2−ε(cid:1) algorithm for Knapsack for some ε > 0.
We allow randomized algorithms.
Conjecture 9.5 ((min, +)-convolution conjecture [24]). Any algorithm computing (min, +)-convolution
requires Ω(n2−o(1)) running time.
Basically [24, Theorem 2] says that assuming the (min, +)-convolution conjecture both Un-
bounded Knapsack and Knapsack require Ω((n + t)2−o(1)) time. The pseudo-polynomial algo-
rithm for Knapsack running in time O(nt) can be modified to work in time O(nv), where v is an
upper bound on value of the solution. In similar spirit, the reductions from [24] can use a hypothet-
algorithm for (min, +)-convolution (modify Theorem 4 from [24]).
ical O(cid:16)(n + v)2−δ(cid:17) algorithm for Knapsack or Unbounded Knapsack to get a subquadratic
Knapsack with O(cid:16)(n + v)2−δ(cid:17) running time would refute the (min, +)-convolution conjecture.
Corollary 9.6 ([24]). For any constant δ > 0, an exact algorithm for Knapsack or Unbounded
We need this modification because in the definition of FPTAS for Knapsack we consider relative
error with respect to the optimal value (not weight). We can use a hypothetical faster approximation
algorithm to get a faster pseudo-polynomial exact algorithm, what would contradict the (min, +)-
convolution conjecture. More formally:
Theorem 9.7 (restated Theorem 1.6). For any constant δ > 0, obtaining a weak (1−ε)-approximation
for Knapsack or Unbounded Knapsack with O((n + 1/ε)2−δ) running time would refute the
(min, +)-convolution conjecture.
Unbounded Knapsack) with running time O(cid:0)(n + 1/ε)2−δ(cid:1). If we set ε = 2/v, then the ap-
Proof. Suppose, that for some δ > 0 we have a weak (1 − ε)-approximation for Knapsack (or
proximation algorithm would solve the exact problem because the absolute error gets bounded by
1/2. By [24, Theorem 2] we know that such an algorithm contradicts the (min, +)-convolution
conjecture. The claim follows.
29
A similar argument works for the Subset Sum problem. Abboud et al. [4] showed that assuming
SETH there can be no O(t1−δpoly(n)) algorithm for Subset Sum (Cygan et al. [23] obtained the
same lower bound before but assuming the SetCover conjecture).
Theorem 9.8 (Conditional Lower Bound for approximate Subset Sum). For any constant δ > 0,
a weak (1− ε)-approximation for Subset Sum with running time O(cid:16)poly(n)(cid:0) 1
SETH and SetCover conjecture.
ε(cid:1)1−δ(cid:17) would refute
Proof. We set ε = 2/t and obtain an algorithm solving the exact Subset Sum, because all numbers
are integers and the absolute error is at most 1/2. The running time is O(t1−δpoly(n)), what refutes
SETH due to [4] and the SetCover conjecture due to [23].
10 Conclusion and Open Problems
In this paper we study the complexity of the Knapsack, Subset Sum and Partition. In the
exact setting, if we are only concerned about the dependence on n, Knapsack and Subset Sum
were already known to be equivalent up to the polynomial factors. Nederlof et al. [55, Theorem
2] showed, that if there exists an exact algorithm for Subset Sum working in O∗(T (n)) time and
O∗(S(n)) space, then we can construct an algorithm for Knapsack working in the same O∗(T (n))
time and O∗(S(n)) space. In contrast, in the realm of pseudo-polynomial time complexity, Subset
Sum seems to be simpler than Knapsack (see Bringmann [17], Cygan et al. [24]). In this paper,
we show similar separation for Knapsack and Partition in the approximation setting.
rithm for Subset Sum is optimal assuming (min, +)-convolution conjecture. Can we improve the
After this paper was announced, Bringmann [18] showed that the current eO(n + 1/ε2) algo-
approximation algorithm for Knapsack to an eO(n + 1/ε2) and match the quadratic lower bound?
It also remains open whether 3SUM and (min, +)-convolution admit FPTAS algorithms with
no dependence on W . To add weight to this open problem, note that it is this issue that makes the
FPTAS algorithms for TreeSparsity inefficient in practice.
Closing the time complexity gap for Partition is another open problem, either by improving
the eO((n + 1/ε)5/3)) FPTAS or the Ω((n + 1/ε)1−o(1)) conditional lower bound. It is worth noting,
that if the Freiman's Conjecture [26] is true, then our techniques would automatically lead to even
faster FPTAS for Partition.
Finally, one can also ask whether randomization is necessary to obtain subquadratic FPTAS
for Partition. We believe that the randomized building blocks can be replaced with deterministic
algorithms by Kellerer et al. [46] and Koiliaris and Xu [48].
11 Acknowledgements
This work is part of the project TOTAL that has received funding from the European Research
Council (ERC) under the European Union's Horizon 2020 research and innovation programme
(grant agreement No 677651). Karol Węgrzycki is supported by the grants 2016/21/N/ST6/01468
and 2018/28/T/ST6/00084 of the Polish National Science Center. We would like to thank Marek
Cygan, Artur Czumaj, Zvi Galil, Oded Margalit and Piotr Sankowski for helpful discussions. Also we
are grateful to the organizers and participants of the Bridging Continuous and Discrete Optimization
program at the Simons Institute for the Theory of Computing, especially Aleksander Mądry.
30
References
[1] Amir Abboud and Arturs Backurs. Towards hardness of approximation for polynomial time
problems.
In Christos H. Papadimitriou, editor, 8th Innovations in Theoretical Computer
Science Conference, ITCS 2017, January 9-11, 2017, Berkeley, CA, USA, volume 67 of LIPIcs,
pages 11:1 -- 11:26. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2017. ISBN 978-3-95977-
029-3. URL http://www.dagstuhl.de/dagpub/978-3-95977-029-3.
[2] Amir Abboud, Ryan Williams, and Huacheng Yu. More applications of the polynomial method
to algorithm design. In Proceedings of the Twenty-sixth Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA '15, pages 218 -- 230, Philadelphia, PA, USA, 2015. Society for
Industrial and Applied Mathematics.
[3] Amir Abboud, Virginia Vassilevska Williams, and Joshua R. Wang. Approximation and
fixed parameter subquadratic algorithms for radius and diameter in sparse graphs. In Robert
Krauthgamer, editor, Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016, pages 377 -- 391.
SIAM, 2016.
[4] Amir Abboud, Karl Bringmann, Danny Hermelin, and Dvir Shabtay. Seth-based lower bounds
for subset sum and bicriteria path. arXiv preprint arXiv:1704.04546, to appear at SODA 2019,
2019.
[5] Alexandr Andoni, Robert Krauthgamer, and Krzysztof Onak. Polylogarithmic approximation
for edit distance and the asymmetric query complexity. In 51th Annual IEEE Symposium on
Foundations of Computer Science, FOCS 2010, October 23-26, 2010, Las Vegas, Nevada, USA,
pages 377 -- 386. IEEE Computer Society, 2010.
[6] Per Austrin, Petteri Kaski, Mikko Koivisto, and Jussi Määttä. Space-time tradeoffs for sub-
set sum: An improved worst case algorithm. In Fedor V. Fomin, Rusins Freivalds, Marta Z.
Kwiatkowska, and David Peleg, editors, Automata, Languages, and Programming - 40th Inter-
national Colloquium, ICALP 2013, Riga, Latvia, July 8-12, 2013, Proceedings, Part I, volume
7965 of Lecture Notes in Computer Science, pages 45 -- 56. Springer, 2013.
[7] Per Austrin, Petteri Kaski, Mikko Koivisto, and Jesper Nederlof. Subset sum in the absence of
concentration. In Ernst W. Mayr and Nicolas Ollinger, editors, 32nd International Symposium
on Theoretical Aspects of Computer Science, STACS 2015, March 4-7, 2015, Garching, Ger-
many, volume 30 of LIPIcs, pages 48 -- 61. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik,
2015.
[8] Per Austrin, Petteri Kaski, Mikko Koivisto, and Jesper Nederlof. Dense subset sum may be
the hardest. In Nicolas Ollinger and Heribert Vollmer, editors, 33rd Symposium on Theoretical
Aspects of Computer Science, STACS 2016, February 17-20, 2016, Orléans, France, volume 47
of LIPIcs, pages 13:1 -- 13:14. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2016.
[9] Arturs Backurs and Piotr Indyk. Edit distance cannot be computed in strongly subquadratic
time (unless SETH is false). In Rocco A. Servedio and Ronitt Rubinfeld, editors, Proceedings of
the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC 2015, Portland,
OR, USA, June 14-17, 2015, pages 51 -- 58. ACM, 2015.
31
[10] Arturs Backurs, Piotr Indyk, and Ludwig Schmidt. Better approximations for tree sparsity in
nearly-linear time. In Philip N. Klein, editor, Proceedings of the Twenty-Eighth Annual ACM-
SIAM Symposium on Discrete Algorithms, SODA 2017, Barcelona, Spain, Hotel Porta Fira,
January 16-19, pages 2215 -- 2229. SIAM, 2017.
[11] Nikhil Bansal, Shashwat Garg, Jesper Nederlof, and Nikhil Vyas. Faster space-efficient algo-
rithms for subset sum and k-sum.
In Hamed Hatami, Pierre McKenzie, and Valerie King,
editors, Proceedings of the 49th Annual ACM SIGACT Symposium on Theory of Computing,
STOC 2017, Montreal, QC, Canada, June 19-23, 2017, pages 198 -- 209. ACM, 2017.
[12] Ziv Bar-Yossef, T. S. Jayram, Robert Krauthgamer, and Ravi Kumar. Approximating edit
distance efficiently. In 45th Symposium on Foundations of Computer Science (FOCS 2004),
17-19 October 2004, Rome, Italy, Proceedings, pages 550 -- 559. IEEE Computer Society, 2004.
[13] Ilya Baran, Erik D. Demaine, and Mihai Patrascu. Subquadratic algorithms for 3sum. Algo-
rithmica, 50(4):584 -- 596, 2008.
[14] Tugkan Batu, Funda Ergün, and Süleyman Cenk Sahinalp. Oblivious string embeddings and
edit distance approximations. In Proceedings of the Seventeenth Annual ACM-SIAM Sympo-
sium on Discrete Algorithms, SODA 2006, Miami, Florida, USA, January 22-26, 2006, pages
792 -- 801. ACM Press, 2006.
[15] Richard Bellman. Dynamic Programming. Princeton University Press, Princeton, NJ, USA,
1957.
[16] Anand Bhalgat, Ashish Goel, and Sanjeev Khanna. Improved approximation results for stochas-
tic knapsack problems.
In Dana Randall, editor, Proceedings of the Twenty-Second Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 2011, San Francisco, California, USA,
January 23-25, 2011, pages 1647 -- 1665. SIAM, 2011.
[17] Karl Bringmann. A near-linear pseudopolynomial time algorithm for subset sum. In Proceedings
of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA '17, pages
1073 -- 1084, Philadelphia, PA, USA, 2017. Society for Industrial and Applied Mathematics.
[18] Karl Bringmann. personal communication, April 2018.
[19] Timothy M. Chan. Approximation schemes for 0-1 knapsack.
In Raimund Seidel, editor,
1st Symposium on Simplicity in Algorithms, SOSA 2018, January 7-10, 2018, New Orleans,
LA, USA, volume 61 of OASICS, pages 5:1 -- 5:12. Schloss Dagstuhl - Leibniz-Zentrum fuer
Informatik, 2018.
[20] Timothy M. Chan and Moshe Lewenstein. Clustered integer 3sum via additive combinatorics.
In Proceedings of the Forty-seventh Annual ACM Symposium on Theory of Computing, STOC
'15, pages 31 -- 40, New York, NY, USA, 2015. ACM.
[21] Ioannis Chatzigiannakis, Piotr Indyk, Fabian Kuhn, and Anca Muscholl, editors. 44th Inter-
national Colloquium on Automata, Languages, and Programming, ICALP 2017, July 10-14,
2017, Warsaw, Poland, volume 80 of LIPIcs, 2017. Schloss Dagstuhl - Leibniz-Zentrum fuer
Informatik.
32
[22] Thomas H Cormen. Introduction to algorithms. MIT press, 2009.
[23] Marek Cygan, Holger Dell, Daniel Lokshtanov, Dańiel Marx, Jesper Nederlof, Yoshio Okamoto,
Ramamohan Paturi, Saket Saurabh, and Magnus Wahlstrom. On problems as hard as cnf-sat.
In Proceedings of the 2012 IEEE Conference on Computational Complexity (CCC), CCC '12,
pages 74 -- 84, Washington, DC, USA, 2012. IEEE Computer Society.
[24] Marek Cygan, Marcin Mucha, Karol Wegrzycki, and Michal Wlodarczyk. On problems equiv-
alent to (min, +)-convolution. In Chatzigiannakis et al. [21], pages 22:1 -- 22:15.
[25] Ari Freund. Improved subquadratic 3sum. Algorithmica, 77(2):440 -- 458, 2017.
[26] Zvi Galil and Oded Margalit. An almost linear-time algorithm for the dense subset-sum prob-
lem. SIAM J. Comput., 20(6):1157 -- 1189, 1991.
[27] Zvi Galil and Oded Margalit. An almost linear-time algorithm for the dense subset-sum
problem. In Javier Leach Albert, Burkhard Monien, and Mario Rodríguez-Artalejo, editors,
Automata, Languages and Programming, 18th International Colloquium, ICALP91, Madrid,
Spain, July 8-12, 1991, Proceedings, volume 510 of Lecture Notes in Computer Science, pages
719 -- 727. Springer, 1991. ISBN 3-540-54233-7.
[28] Zvi Galil and Oded Margalit. personal communication, 2017.
[29] George Gens and Eugene Levner. Computational complexity of approximation algorithms for
combinatorial problems. In Jirí Becvár, editor, Mathematical Foundations of Computer Science
1979, Proceedings, 8th Symposium, Olomouc, Czechoslovakia, September 3-7, 1979, volume 74
of Lecture Notes in Computer Science, pages 292 -- 300. Springer, 1979.
[30] George Gens and Eugene Levner. A fast approximation algorithm for the subset-sum problem.
INFOR: Information Systems and Operational Research, 32(3):143 -- 148, 1994.
[31] Georgii V Gens and Eugenii V Levner. Fast approximation algorithms for knapsack type
problems. In Optimization Techniques, pages 185 -- 194. Springer, 1980.
[32] GV Gens and EV Levner. Approximation algorithm for some scheduling problems. Engrg.
Cybernetics, 6:38 -- 46, 1978.
[33] Beat Gfeller. Finding longest approximate periodic patterns. In Frank Dehne, John Iacono, and
Jörg-Rüdiger Sack, editors, Algorithms and Data Structures - 12th International Symposium,
WADS 2011, New York, NY, USA, August 15-17, 2011. Proceedings, volume 6844 of Lecture
Notes in Computer Science, pages 463 -- 474. Springer, 2011.
[34] Omer Gold and Micha Sharir. Improved bounds for 3sum, k-sum, and linear degeneracy. In
Kirk Pruhs and Christian Sohler, editors, 25th Annual European Symposium on Algorithms,
ESA 2017, September 4-6, 2017, Vienna, Austria, volume 87 of LIPIcs, pages 42:1 -- 42:13.
Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2017.
[35] Brian Hayes. Computing science: The easiest hard problem. American Scientist, 90(2):113 -- 117,
2002.
33
[36] Ellis Horowitz and Sartaj Sahni. Computing partitions with applications to the knapsack
problem. J. ACM, 21(2):277 -- 292, 1974.
[37] Nick Howgrave-Graham and Antoine Joux. New generic algorithms for hard knapsacks. In
Henri Gilbert, editor, Advances in Cryptology - EUROCRYPT 2010, 29th Annual International
Conference on the Theory and Applications of Cryptographic Techniques, French Riviera, May
30 - June 3, 2010. Proceedings, volume 6110 of Lecture Notes in Computer Science, pages
235 -- 256. Springer, 2010.
[38] Chloe Ching-Yun Hsu and Chris Umans. On multidimensional and monotone k-sum. To appear
at MFCS 2017, 2017.
[39] Oscar H. Ibarra and Chul E. Kim. Fast approximation algorithms for the knapsack and sum
of subset problems. J. ACM, 22(4):463 -- 468, 1975.
[40] Klaus Jansen and Stefan Erich Julius Kraft. A faster FPTAS for the unbounded knapsack
In Zsuzsanna Lipták and William F. Smyth, editors, Combinatorial Algorithms -
problem.
26th International Workshop, IWOCA 2015, Verona, Italy, October 5-7, 2015, Revised Selected
Papers, volume 9538 of Lecture Notes in Computer Science, pages 274 -- 286. Springer, 2015.
[41] Allan Grønlund Jørgensen and Seth Pettie. Threesomes, degenerates, and love triangles. In
55th IEEE Annual Symposium on Foundations of Computer Science, FOCS 2014, Philadelphia,
PA, USA, October 18-21, 2014, pages 621 -- 630. IEEE Computer Society, 2014.
[42] Edward G. Coffman Jr. and George S. Lueker. Probabilistic analysis of packing and partitioning
algorithms. Wiley-Interscience series in discrete mathematics and optimization. Wiley, 1991.
[43] Richard M. Karp. Reducibility among combinatorial problems.
In Raymond E. Miller and
James W. Thatcher, editors, Proceedings of a symposium on the Complexity of Computer Com-
putations, held March 20-22, 1972, at the IBM Thomas J. Watson Research Center, Yorktown
Heights, New York., The IBM Research Symposia Series, pages 85 -- 103. Plenum Press, New
York, 1972.
[44] R.M. Karp. The fast approximate solution to hard combinatorial problems. Proceedings of the
6th Southeastern Conference on Combinatorics, Graph Theory and Computing, pages 15 -- 31,
1975.
[45] Hans Kellerer and Ulrich Pferschy.
Improved dynamic programming in connection with an
FPTAS for the knapsack problem. J. Comb. Optim., 8(1):5 -- 11, 2004.
[46] Hans Kellerer, Ulrich Pferschy, and Maria Grazia Speranza. An efficient approximation scheme
for the subset-sum problem. In Hon Wai Leong, Hiroshi Imai, and Sanjay Jain, editors, Al-
gorithms and Computation, 8th International Symposium, ISAAC '97, Singapore, December
17-19, 1997, Proceedings, volume 1350 of Lecture Notes in Computer Science, pages 394 -- 403.
Springer, 1997.
[47] Hans Kellerer, Ulrich Pferschy, and David Pisinger. Knapsack problems. Springer, 2004.
34
[48] Konstantinos Koiliaris and Chao Xu. A faster pseudopolynomial time algorithm for subset sum.
In Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA '17, pages 1062 -- 1072, Philadelphia, PA, USA, 2017. Society for Industrial and Applied
Mathematics.
[49] Marvin Künnemann, Ramamohan Paturi, and Stefan Schneider. On the fine-grained complexity
of one-dimensional dynamic programming. In Chatzigiannakis et al. [21], pages 21:1 -- 21:15.
[50] Eugene L Lawler. Fast approximation algorithms for knapsack problems. Mathematics of
Operations Research, 4(4):339 -- 356, 1979.
[51] Ohad Lipsky and Ely Porat. Approximate pattern matching with the L1, L2 and L∞ metrics.
Algorithmica, 60(2):335 -- 348, 2011.
[52] George B Mathews. On the partition of numbers. Proceedings of the London Mathematical
Society, 1(1):486 -- 490, 1896.
[53] Ralph C. Merkle and Martin E. Hellman. Hiding information and signatures in trapdoor
knapsacks. IEEE Trans. Information Theory, 24(5):525 -- 530, 1978.
[54] Stephan Mertens. The easiest hard problem: Number partitioning. Computational Complexity
and Statistical Physics, 125(2):125 -- 139, 2006.
[55] Jesper Nederlof, Erik Jan van Leeuwen, and Ruben van der Zwaan. Reducing a target interval
to a few exact queries. In Branislav Rovan, Vladimiro Sassone, and Peter Widmayer, editors,
Mathematical Foundations of Computer Science 2012 - 37th International Symposium, MFCS
2012, Bratislava, Slovakia, August 27-31, 2012. Proceedings, volume 7464 of Lecture Notes in
Computer Science, pages 718 -- 727. Springer, 2012.
[56] Richard Schroeppel and Adi Shamir. A t=o(2n/2), s=o(2n/4) algorithm for certain np-
complete problems. SIAM J. Comput., 10(3):456 -- 464, 1981.
[57] Daniel D. Sleator and Robert Endre Tarjan. A data structure for dynamic trees. J. Comput.
Syst. Sci., 26(3):362 -- 391, June 1983. ISSN 0022-0000.
[58] Virginia Vassilevska Williams. Hardness of easy problems: Basing hardness on popular con-
jectures such as the strong exponential time hypothesis (invited talk). In Thore Husfeldt and
Iyad A. Kanj, editors, 10th International Symposium on Parameterized and Exact Computa-
tion, IPEC 2015, September 16-18, 2015, Patras, Greece, volume 43 of LIPIcs, pages 17 -- 29.
Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2015.
[59] Gerhard J. Woeginger. When does a dynamic programming formulation guarantee the existence
of a fully polynomial time approximation scheme (fptas)? INFORMS Journal on Computing,
12(1):57 -- 74, 2000.
[60] Uri Zwick. All pairs shortest paths in weighted directed graphs -- exact and almost exact
algorithms.
In 39th Annual Symposium on Foundations of Computer Science, FOCS '98,
November 8-11, 1998, Palo Alto, California, USA, pages 310 -- 319. IEEE Computer Society,
1998.
35
A Proof of Theorem 2.2
Corollary A.1 (restated Observation 2.2). If we can weakly (1 − ε)-approximate Subset Sum in
time eO(T (n, ε)), then we can (1 − ε)-approximate Partition in the same eO(T (n, ε)) time.
Proof. Let Z = n be the initial set of items. We run a weak (1 − ε)-approximation algorithm for
Subset Sum with target b = Σ(Z)/2. Let Z ∗ denote the optimal partition of set Z:
Z ∗ = arg max
Σ(Z ′).
Z ′⊆Z, Σ(Z ′)≤b
By the definition of weak (1 − ε)-approximation for Subset Sum we get a solution ZW such
that:
(1 − ε)Σ(Z ∗) ≤ Σ(ZW ) and Σ(ZW ) < (1 + ε)b
If Σ(ZW ) ≤ b then it is a correct solution for Partition. Otherwise we take a set Z ′
W = Z\ZW .
W ) < b. Additionally we know, that Σ(ZW ) < (1 + ε)b, so
Because Σ(Z)/2 = b we know that Σ(Z ′
(1 − ε)b < Σ(Z ′
W ). Similarly, because Z ∗ ≤ b, we have:
(1 − ε)Σ(Z ∗) ≤ (1 − ε)b < Σ(Z ′
W ) ≤ Σ(Z ∗) ≤ b.
So Σ(Z ′
W ) follows the definition of approximation for Partition. The running time follows
because T (n, 1/ε) must be superlinear (algorithm needs to read input at least) and we executed the
weak (1 − ε)-approximation Subset Sum algorithm only constant number of times.
B Proofs of the Rounding Lemmas
Lemma B.1. For k natural numbers x1, x2, . . . , xk and positive q, ε such that q ≤ Pk
0 < ε < 1, it holds:
i=1 xi and
xi ≤
qε
k
kXi=1
kXi=1(cid:24) kxi
qε(cid:25) < (1 + ε)
kXi=1
xi
Proof. Let xi = qε
k ci + di where 0 < di ≤ qε
know there is at least one positive di (we will use this fact later). We have l kxi
note that:
qεm = ci + 1. First,
k . If some xi = 0 then we set ci = di = 0, however we
xi =
qε
k
kXi=1
kXi=1
ci +
kXi=1
di ≤
qε
k
ci + qε =
qε
k
(ci + 1),
kXi=1
what proves the left inequality. To handle the right inequality we take advantage of the assumption
i=1 xi ≥ q and get:
Pk
(1 + ǫ)
kXi=1
xi = ε
xi +
kXi=1
kXi=1
xi ≥ qǫ +
xi = qǫ +
ci +
qε
k
kXi=1
di >
qε
k
kXi=1
di =
kXi=1
(ci + 1).
kXi=1
kXi=1
kXi=1
(ci + 1) +
qε
k
kXi=1
36
Lemma B.2. For k natural numbers x1, x2, . . . , xk and positive q, ε such that q ≤ Pk
0 < ε < 1, it holds:
i=1 xi and
(1 − ε)
xi <
qε
k
kXi=1
qε(cid:23) ≤
kXi=1(cid:22) kxi
kXi=1
xi.
Proof. The proof is very similar to the proof of Lemma B.1, however now we represent xi as qε
where 0 ≤ di < qε
qεk = ci. The right inequality holds because:
k . We havej kxi
kXi=1
di ≥
qε
k
qε
k
xi =
k ci +di
and the left one can be proven as follows:
ci +
kXi=1
kXi=1
xi − qε = kXi=1
ci
kXi=1
ci + di! − qε <
qε
k
qε
k
kXi=1
ci.
(1 − ε)
xi =
kXi=1
kXi=1
xi − ε
kXi=1
xi ≤
kXi=1
C Problems Definitions
C.1 Exact problems
Knapsack
Input: A set of n items {(v1, w1), . . . , (vn, wn)}
Task: Find x1, . . . , xn such that:
maximize
subject to
nXj=1
nXj=1
vjxj
wjxj ≤ t,
xj ∈ {0, 1}n,
j = 1, . . . n.
Sometimes, instead of exact solution x1, . . . , xn in Knapsack-type problems one needs to return
the value of such solution. In decision version of such problems we are given capacity t and value v
and ask if there is a subset of items with the total capacity not exceeding t and total value exactly
v (e.g., see discussion in [24, 47]).
37
Unbounded Knapsack
Input: A set of n items {(v1, w1), . . . , (vn, wn)}
Task: Find x1, . . . , xn such that:
maximize
subject to
nXj=1
nXj=1
vjxj
wjxj ≤ t,
xj ∈ N ∪ {0},
j = 1, . . . n.
Subset Sum
Input: A set of n integers {w1, . . . , wn}
Task: Find x1, . . . , xn such that:
maximize
subject to
nXj=1
nXj=1
wjxj
wjxj ≤ t,
xj ∈ {0, 1}n,
j = 1, . . . n.
Partition
2Pn
Input: A set of n integers {w1, . . . , wn} and b = 1
Task: Find x1, . . . , xn such that:
i=1 wi
maximize
subject to
nXj=1
nXj=1
wjxj
wjxj ≤ b,
xj ∈ {0, 1}n,
j = 1, . . . n.
(min, +)-convolution
Input: Sequences (a[i])n−1
Task: Output sequence (c[i])n−1
i=0 , (b[i])n−1
i=0
i=0 , such that c[k] = mini+j=k(a[i] + b[j])
k-SUM
Input: k − 1 sets A1, A2, . . . , Ak−1 and the set S of integers, each with cardinality at most n.
Task: Is there a (a1, . . . , ak−1, s) ∈ A1 × . . . × Ak−1 × S such that a1 + . . . + ak−1 = s
3SUM
Input: 3 sets A, B, C of integers, each with cardinality at most n.
Task: Is there a triple (a, b, c) ∈ A × B × C such that a + b = c
38
TreeSparsity
Input: A rooted tree T with a weight function x : V (T ) → N, parameter k
Task: Find the maximal total weight of a rooted subtree of size k
C.2 Approximate problems definition
Let Σ(S) denote the sum of elements in S. The V (I) denotes the total value of items I and W (I)
denotes the total weight of items.
(1 − ε)-approximation of Knapsack
Input: A set S = {(v1, w1), . . . , (vn, wn)} items and a target number t
Task: Let Z ∗ be the optimal solution of exact Knapsack with target t. The (1 − ε)-
approximate algorithm for Knapsack returns ZH such that (1 − ε)V (Z ∗) ≤ V (ZH ) ≤ V (Z ∗)
and W (ZH) ≤ t.
Analogous definition is for Unbounded Knapsack.
(1 − ε)-approximation of Subset Sum
Input: A set S = {a1, . . . , an} of positive integers and a target number t
Task: Let Z ∗ be the optimal solution of exact Subset Sum with target t. The (1 − ε)-
approximate algorithm returns ZH such that (1 − ε)Σ(Z ∗) ≤ Σ(ZH) ≤ Σ(Z ∗)
(1 − ε)-approximation of Partition
Input: A set S = {a1, . . . , an} of positive integers
Task: Let Z ∗ be the optimal solution of exact Partition. The (1− ε)-approximate algorithm
returns ZH such that (1 − ε)Σ(Z ∗) ≤ Σ(ZH) ≤ Σ(Z ∗)
Weak (1 − ε)-approximation of Subset Sum
Input: A set S = {a1, . . . , an} of positive integers and a target number t
Task: Let Z ∗ be the optimal solution of exact Subset Sum with target t. The (1 − ε)-
approximate algorithm returns ZH such that (1 − ε)Σ(Z ∗) ≤ Σ(ZH) ≤ Σ(Z ∗) or t ≤ Σ(ZH) ≤
(1 + ε)t
Approximate 3SUM ([33])
Input: Three sets A, B, C of positive integers, each with cardinality at most n.
Task: The algorithm:
• concludes that no triple (a, b, c) ∈ A × B × C with a + b = c exists, or
• it outputs a triple (a, b, c) ∈ A × B × C with a + b ∈ [c/(1 + ε), c(1 + ε)]
39
Approximate (min, +)-convolution
Input: Sequences A[0, . . . , n − 1], B[0, . . . , n − 1] of positive integers and approximation pa-
rameter 0 < ε ≤ 1
Task: Assume that OPT[k] = min0≤i≤k(A[i] + B[k − i]) is the exact (min, +)-convolution
of A and B. The task is to output a sequence C[0, . . . , n − 1] such that ∀i OPT[i] ≤ C[i] ≤
(1 + ε)OPT[i]
40
|
1611.03385 | 1 | 1611 | 2016-11-10T16:27:09 | Approximately Sampling Elements with Fixed Rank in Graded Posets | [
"cs.DS",
"math.CO"
] | Graded posets frequently arise throughout combinatorics, where it is natural to try to count the number of elements of a fixed rank. These counting problems are often $\#\textbf{P}$-complete, so we consider approximation algorithms for counting and uniform sampling. We show that for certain classes of posets, biased Markov chains that walk along edges of their Hasse diagrams allow us to approximately generate samples with any fixed rank in expected polynomial time. Our arguments do not rely on the typical proofs of log-concavity, which are used to construct a stationary distribution with a specific mode in order to give a lower bound on the probability of outputting an element of the desired rank. Instead, we infer this directly from bounds on the mixing time of the chains through a method we call $\textit{balanced bias}$.
A noteworthy application of our method is sampling restricted classes of integer partitions of $n$. We give the first provably efficient Markov chain algorithm to uniformly sample integer partitions of $n$ from general restricted classes. Several observations allow us to improve the efficiency of this chain to require $O(n^{1/2}\log(n))$ space, and for unrestricted integer partitions, expected $O(n^{9/4})$ time. Related applications include sampling permutations with a fixed number of inversions and lozenge tilings on the triangular lattice with a fixed average height. | cs.DS | cs | Approximately Sampling Elements with Fixed Rank in
Prateek Bhakta ∗
Graded Posets
Ben Cousins †
Dana Randall §
August 13, 2018
Abstract
Matthew Fahrbach ‡
Graded posets frequently arise throughout combinatorics, where it is natural to try
to count the number of elements of a fixed rank. These counting problems are often #P-
complete, so we consider approximation algorithms for counting and uniform sampling.
We show that for certain classes of posets, biased Markov chains that walk along edges
of their Hasse diagrams allow us to approximately generate samples with any fixed
rank in expected polynomial time. Our arguments do not rely on the typical proofs
of log-concavity, which are used to construct a stationary distribution with a specific
mode in order to give a lower bound on the probability of outputting an element of
the desired rank. Instead, we infer this directly from bounds on the mixing time of the
chains through a method we call balanced bias.
A noteworthy application of our method is sampling restricted classes of integer
partitions of n. We give the first provably efficient Markov chain algorithm to uniformly
sample integer partitions of n from general restricted classes. Several observations
allow us to improve the efficiency of this chain to require O(n1/2 log(n)) space, and for
unrestricted integer partitions, expected O(n9/4) time. Related applications include
sampling permutations with a fixed number of inversions and lozenge tilings on the
triangular lattice with a fixed average height.
6
1
0
2
v
o
N
0
1
]
S
D
.
s
c
[
1
v
5
8
3
3
0
.
1
1
6
1
:
v
i
X
r
a
[email protected]. Supported in part by NSF grant CCF-1526900.
∗Department of Math and Computer Science, University of Richmond, Richmond, VA 23173. Email:
†School of Computer Science, Georgia Institute of Technology, Atlanta, GA 30332.
[email protected]. Supported in part by NSF grants CCF-1217793 and EAGER-1415498.
‡School of Computer Science, Georgia Institute of Technology, Atlanta, GA 30332.
Email:
[email protected]. Supported in part by NSF grant DGE-1650044 and a Tau Beta Pi Fellow-
ship.
Email:
§School of Computer Science, Georgia Institute of Technology, Atlanta, GA 30332.
Email:
[email protected]. Supported in part by NSF grants CCF-1526900 and CNS-1544090.
1
1
Introduction
Graded posets are partially ordered sets equipped with a unique rank function that both
respects the partial order and such that neighboring elements in the Hasse diagram of the
poset have ranks that differ by ±1. Graded posets arise throughout combinatorics, including
permutations ordered by numbers of inversions, geometric lattices ordered by volume, and
independent sets and matchings ordered by cardinality. Sometimes we find rich underlying
structures that allow us to directly count, and therefore sample, fixed rank elements of a
graded poset. In other cases, efficient methods are unlikely to exist, so Markov chains offer
the best approach to sampling and approximate counting.
Jerrum and Sinclair [18] observed that we could sample matchings of any fixed size with
the addition of a bias parameter λ that gives weight proportional to λm to each matching m.
For any graph G, they showed that the sequence ai, the number of matchings of G of size i,
is log-concave, from which it follows that f (i) = aiλi is also. In particular, f (i) must be
unimodal for all λ. Setting λ = ak/ak+1 makes k the mode of distribution f (i), and therefore
samples with this weighting will be of the appropriate size with probability at least 1/(n+1).
Jerrum and Sinclair showed that the matching Markov chain is rapidly mixing for all λ, so
it can find matchings of fixed size k efficiently whenever 1/poly(n) < λ < poly(n). (This
condition is not always satisfied, but the more involved algorithm of Jerrum, Sinclair, and
Vigoda circumvents this issue [19].) Log-concavity is critical to this argument in order to
conclude that there is a value of λ for which samples of the desired size occur with high
enough probability.
This follows a common approach used in physics for which we would like to sample from
a microcanonical ensemble, i.e., the states with a fixed energy, from a much larger canonical
(or grand canonical) ensemble, where the energies are allowed to vary due to interactions
with the external environment.
In particular, given input parameter λ, often related to
temperature, a configuration σ has Gibbs (or Boltzmann) weight π(σ) = λr(σ)/Z, where r(σ)
is the rank of σ and Z is the normalizing constant. Elements σ sampled from this distribution
are uniformly distributed, conditioned on their rank. The choice of λ controls the expected
rank of the distribution, so simulations of the Markov chain at various λ can be useful for
understanding properties of configurations with a fixed energy. Typically, however, there is
no a priori guarantee that this approach will enable us to sample configurations of a given
size efficiently.
Our main example throughout will be sampling and counting (possibly restricted) integer
partitions. An integer partition of nonnegative integer n is a decomposition of n into a
nonincreasing sequence of positive integers that sum to n. The seven partitions of 5 are: (5),
(4, 1), (3, 2), (3, 1, 1), (2, 2, 1), (2, 1, 1, 1), and (1, 1, 1, 1, 1). Integer partitions are commonly
represented by staircase walks in Z2 known as Young (or Ferrers) diagrams, where the heights
of the columns represent distinct pieces of the partition. Partitions of n have exactly n
squares, i.e., the area of the diagram, and their column heights are nonincreasing. Partitions
arise in many contexts, include exclusion processes [8], random matrices [29], representation
theory [15], juggling patterns [4], and growth processes [13] (see, e.g., [2]).
2
1.1 Sampling Elements from Graded Posets
Several general approaches have been developed to sample elements of fixed rank from a
graded poset, with varying success. The three main approaches for sampling are dynamic
programming algorithms using self-reducibility, Boltzmann samplers using geometric random
variables, and Markov chains. The first two approaches require methods to estimate the
number of configurations of each size, so Markov chains offer the most promising approach
for sampling when these are unavailable.
Each of these approaches has been studied extensively in the context of sampling integer
partitions. The first class of approaches uses dynamic programming and generating functions
to iteratively count the number of partitions of a given type. Nijinhuis and Wilf [28] give
a recursive algorithm using dynamic programming that computes tables of exact values.
This algorithm takes O(n5/2) time and space for preprocessing and O(n3/2) time per sample.
Squire [34] improved this to O(n2) time and space for preprocessing and O(n3/2 log(n))
time per sample using Euler's pentagonal recurrence and a more efficient search method.
The time and space complexity bounds of these algorithms account for the fact that each
value of p(n), as well as the intermediate summands, requires O(n1/2) space by the Hardy-
Ramanujan formula. Therefore, even when available, dynamic programming approaches
for exact sampling break down in practice on single machines when n ≥ 106 due to space
constraints.
Boltzmann samplers offer a more direct method for sampling that avoids the compu-
tationally expensive task of counting partitions. A Boltzmann sampler generates samples
from a larger combinatorial class with probability proportional to the Boltzmann weight λσ,
where σ is the size of the partition. Samples of the same size are drawn uniformly at ran-
dom, and the algorithm rejects those that fall outside of the target size [10, 11]. The value λ
is chosen to maximize the yield of samples of our target size n. Fristedt [12] suggested an ap-
proach that quickly generates a random partition using appropriate independent geometric
random variables. His approach exploits the factorization of the generating function for p(n)
and can be interpreted as sampling Young diagrams σ in the n × ∞ grid with probability
proportional to the Boltzmann weight λσ. Recently Arratia and DeSalvo [3] gave a proba-
bilistic approach that is substantially more efficient than previous algorithms, thus allowing
for fast generation of random partitions for significantly larger numbers, e.g., n ≥ 106. Build-
ing on the work of Fristedt [12], they introduce the probabilistic divide-and-conquer (PDC)
method to generate random partitions of n in optimal (cid:101)O(n1/2) expected time and space
(where (cid:101)O suppresses log factors). Their PDC algorithm also uses independent geometric
random variables to generate a partition, but does so recursively in phases. PDC achieves
superior performance relative to conventional Boltzmann Sampling by rejecting impossible
configurations in early phases.
Stochastic approaches using Markov chains have produced a similarly rich corpus of work,
but until now have not provided rigorous polynomial bounds. One popular direction uses
Markov chains based on coagulation and fragmentation processes that allow pieces of the
partition to be merged and split [1, 6]. Ayyer et al. [4] recently proposed several natural
Markov chains on integer partitions in order to study juggling patterns. In all of these works,
3
most of the effort has been to show that the Markov chains converge to the uniform distribu-
tion over partitions and often use stopping rules in order to generate samples. Experimental
evidence suggests that these chains may converge quickly to the correct equilibrium, but
they lack explicit bounds.
1.2 Results
For any graded poset, let Ωk be the elements of rank k and let Ω =(cid:83)n
i=0 Ωi be the entire
poset. We show that provably efficient Boltzmann samplers on Ωk can be easily constructed
from certain rapidly mixing Markov chains on the Hasse diagram of the entire poset Ω,
under very mild conditions. We apply this technique to design the first provably efficient
Markov chain based algorithms for sampling integer partitions of an integer n, permutations
with a fixed number of inversions, and lozenge tilings with fixed average height. Unlike all
other methods for sampling that depend on efficient counting techniques, our results extend
to interesting subspaces of these posets, such as partitions with at least k pieces with size
greater than (cid:96), or partitions into pieces with distinct sizes, or many other such restricted
classes. For these subspaces, our results provide the first sampling algorithms that do not
require the space-expensive task of counting.
We focus on the example of integer partitions of n and prove that there is a Markov chain
Monte Carlo algorithm for uniformly sampling partitions of n from a large family of region-
restricted partitions, i.e., Young diagrams restricted to any simply-connected bounding re-
gion. The Markov chain on the Hasse diagram for partitions is the natural "mountain-valley"
chain studied for staircase walks, tilings, and permutations. The transition probabilities are
designed to generate a diagram σ with weight proportional to λσ. Previous work on biased
card shuffling [5] and growth processes [5, 13, 24] shows that this chain is rapidly mixing for
any constant λ on well-behaved regions.
In the general setting of sampling from a graded poset, our algorithm is similar to current
Boltzmann samplers that heuristically sample elements of a given size, but often without
rigorous analysis. However, we establish conditions under which these algorithms can be
shown to be efficient, including restricted settings for which no other methods provide guar-
antees on both efficiency and accuracy. For example, we show that our method can produce
random partitions of n in O(n9/4) expected time with only O(n1/2 log(n)) space. Using cou-
pling from the past, we can in fact generate samples of the desired size exactly uniformly, if
this is desirable.
Although our algorithm is slower than recent results for sampling unrestricted partitions
using independent geometric random variables [3, 12] (in the settings where those methods
apply), our method is significantly more versatile. The Markov chain algorithm readily
adapts to various restricted state spaces, such as sampling partitions with bounded size and
numbers of parts, partitions with bounded Durfee square, and partitions with prescribed
gaps between successive pieces including partitions into pieces with distinct sizes. For general
bounding regions, our algorithm still uses O(n1/2 log(n)) space, and hence is usually much
more suitable than other approaches with substantially larger space requirements.
4
Finally, we achieve similar results for sampling from fixed a rank in other graded posets.
These include permutations with a fixed number of inversions and lozenge tilings with a
given average height, referring to the height function representation of the tilings (see, e.g.,
[25]). Kenyon and Okounkov [20] explored limit shapes of tilings with fixed volume, and
showed such constraints simplified some arguments, but there has not been work addressing
sampling.
1.3 Techniques
i≤k Pr[Ωi] ≥ 1/c and(cid:80)
distribution satisfying(cid:80)
First, we present a new argument that shows how to build Boltzmann samplers with per-
formance guarantees, even in cases where the underlying distributions are not known (or
necessarily even believed) to be unimodal, provided the Markov chain is rapidly mixing on
the whole Hasse diagram. We prove that there must be a balanced bias parameter λ that
we can find efficiently allowing us to generate configurations of the target size with prob-
ability at least 1/poly(n). The desired set is no longer guaranteed to be the mode of the
distribution, as generally required, but we still show that rejection probabilities will not
be too high. We carefully define a polynomial sized set from which the bias parameter λ
will be chosen. Then we show that at least one bias parameter in this set will define a
i>k Pr[Ωi] ≥ 1/c, for some constant c. Be-
cause the Markov chain M changes the rank by at most 1 in each step, we must generate
samples of size exactly k with probability at least 1/τ (M), where τ (M) is the mixing time
of M, which we prove using conductance. Thus, when the chain is rapidly mixing, samples of
size k must occur with non-negligible probability. This new method based on balanced biases
is quite general and circumvents the need to make any assumptions about the underlying
distributions.
We use biased Markov chains and Boltzmann sampling to generate samples of the de-
sired size k. We assign probability λr(σ)/Z to every element σ ∈ Ω, where r(σ) is its rank
and Z is the normalizing constant. When the underlying distributions on f (i) = Ωiλi are
known to be log-concave in i, such as unrestricted integer partitions or permutations with a
fixed number of inversions, we can provide better guarantees than the general balanced bias
algorithm.
Several observations allow us to improve the running time of our algorithm, especially
in the case of unrestricted integer partitions. First, instead of sampling Young diagrams
in an n × n lattice region, we restrict to diagrams lying in the first quadrant of Z2 below
the curve y = 2n/x, since this region contains all the Young diagrams of interest and has
area Θ(n log(n)), allowing the Markov chain to converge faster. Next, we improve the bounds
on the mixing time for our particular choice of λ given in [13] using a careful analysis of
a recent result in [24]. Last, we show how to salvage many of the samples rejected by
Boltzmann sampling to increase the success probability to at least Ω(1/n1/4). With all of
these improvements we conclude that the chain will converge in O(n2) time and O(n1/4) trials
are needed in expectation before generating a sample corresponding to a partition of n. We
also optimize the space required to implement the Markov chain. All Young diagrams in the
region R have at most O(n1/2) corners, so each diagram in stored in O(n1/2 log(n)) space.
5
2 Bounding Rejection with Balanced Bias
Let Ω be the elements of any graded poset with rank function r : Ω → Z≥0. The rank of the
poset Ω is R = max({r(σ) : σ ∈ Ω}) and the rank generating function of Ω is
FΩ(x) =
xr(σ).
(cid:88)
σ∈Ω
R(cid:88)
Let Ωk be the set of elements of Ω with rank k and let aΩ,k = Ωk. For any λ > 0, the Gibbs
measure of each σ ∈ Ω is π(σ) = λr(σ)/Z, where
Z = FΩ(λ) =
aΩ,iλi
is the normalizing constant.
We define the natural Markov chain M that traverses the Hasse diagram of Ω as follows.
Let ∆ be the maximum number of neighbors of any element σ ∈ Ω in the Hasse diagram.
For any pair of neighboring elements σ, ρ ∈ Ω, we define the transition probabilities
i=0
P (σ, ρ) =
min(1, π(ρ)/π(σ))
2∆
,
and with all remaining probability we stay at σ. This Markov chain is known as the
lazy, Metropolis-Hastings algorithm [26] with respect to the Boltzmann distribution π(σ) =
λr(σ)/Z. If M connects the state space of the poset, the process σt is guaranteed to converge
to the stationary distribution π starting from any initial σ0 [24].
The number of steps needed for the Markov chain M with state space Ω to get arbitrarily
close to this stationary distribution is known as its mixing time τ (ε), defined as
τ (ε) = min({t : (cid:107)P t(cid:48)
, π(cid:107)tv ≤ ε for all t(cid:48) ≥ t}),
for all ε > 0, where (cid:107)·,·(cid:107)tv is the total variation distance (see, e.g., [32]). We say that a
Markov chain is rapidly mixing if the mixing time is bounded above by a polynomial in n
and log(ε−1).
We wish to uniformly sample a random element σ ∈ Ωk, for a fixed k ∈ [R]. To achieve
this, we repeatedly sample from a favorable Boltzmann distribution over all of Ω until we
have an element of rank k. We show that under very mild conditions on the coefficients of the
rank generating function, it is sufficient that the Markov chain M over Ω be rapidly mixing
in order for the Boltzmann sampling procedure to be efficient. Specifically, we require only
that R = O(poly(n)) and 1 ≤ aΩ,i ≤ c(n)i for some polynomial c(n).
We formalize our claim by assuming the polynomial c = c(n) ≥ 2. For t ≥ 0, let
βt = ln(1/c) + t ln(c)/R
and
Then let Prt[σ] = λr(σ)
a way that at most R2 values need to be considered.
/Zt, where Zt = FΩ(λt). The sequence {λt}∞
t
t=0 is constructed in such
λt = eβt = ct/R−1.
6
Lemma 2.1. For all σ ∈ Ω and t ≥ 0, we have
Prt+1[σ]
Prt[σ]
≥ 1
c
.
Proof. By the definition of βt+1, we have
1 ≥ eβtr(σ)
eβt+1r(σ) = e− ln(c)r(σ)/R ≥ 1
c
.
It follows that
Prt+1[σ]
Prt[σ]
=
eβt+1r(σ)
eβtr(σ)
· Zt
Zt+1
≥ Zt
Zt+1
=
(cid:80)
(cid:80)
σ∈Ω eβt+1r(σ) ≥
σ∈Ω eβtr(σ)
(cid:80)
(cid:80)
σ∈Ω eβt(σ)
σ∈Ω ceβtr(σ) =
1
c
.
The following lemma is critical to our argument and states that there exists a balanced
bias parameter λ relative to our target set Ωk that assigns nontrivial probability mass to
elements with rank at most k and elements with rank greater than k.
Lemma 2.2. Let Ω be the elements of a graded poset with rank R ≥ 1 such that 1 ≤ aΩ,i ≤ ci
for all i ∈ {0, 1, . . . , R} and some c ≥ 2. If k ∈ [R − 1], there exists a t ∈ [R2] for which
[r(σ) ≤ k] ≥ 1
c + 1
Pr
t
and
To prove the existence of t∗, recall that Prt[r(σ) > k] > 1/(c + 1) if and only if
To prove the second inequality, it suffices to show (c + 1)λR
λt = cR−1 ≥ 1 because R ≥ 1. Therefore (c + 1)λR
minimum t ∈ [R2] satisfying
t > 1. Letting t = R2, we have
t > 1 as desired. Finally, let t∗ be the
1
.
c + 1
[r(σ) > k] >
Pr
t
7
Proof. Suppose there exists a minimum t∗ ∈ Z≥1 such that
Then
so by Lemma 2.1 we have
[r(σ) > k] ≥ 1
c + 1
.
Pr
t
Pr
t∗ [r(σ) > k] >
1
c + 1
.
[r(σ) ≤ k] ≥ c
c + 1
Pr
t∗−1
,
t∗ [r(σ) ≤ k] ≥ 1
Pr
c + 1
.
R(cid:88)
i=k+1
(c + 1)
aΩ,iλi
t > 1.
We now prove our main theorem, which depends on the mixing time τ (ε) of the Markov
chain M for the balanced bias λt, given by Lemma 2.2. The proof uses a characterization
of the mixing time of a Markov chain in terms of its conductance [17, 33]. For an ergodic
Markov chain M with stationary distribution π, the conductance of a subset S ⊆ Ω is
defined as
(cid:88)
σ∈S,ρ∈S
Φ(S) =
π(σ)P (σ, ρ)
.
π(S)
The conductance of the chain M is the minimum conductance over all subsets
ΦM = min
S⊆Ω
({Φ(S) : π(S) ≤ 1/2}) ,
and is related to the mixing time τ (ε) of M as follows.
Theorem 2.1 ([17]). The mixing time of a Markov chain M with conductance Φ satisfies
(cid:18)1 − 2Φ
(cid:19)
2Φ
ln(cid:0)ε−1(cid:1) .
τ (ε) ≥
Theorem 2.2 (Balanced Bias). Let M be a rapidly mixing Markov chain with state space Ω
and mixing time τ = τ (e−1) such that the transitions of M induce a graded partial order
on Ω with rank function r : Ω → Z≥0 and rank R. If there exists a polynomial c ≥ 2 such
that 1 ≤ aΩ,i ≤ ci for all i ∈ {0, 1, . . . , R}, then
π(Ωk) ≥
1
2(c + 1)(τ + 1)
for any fixed k ∈ [R] with the balanced bias. If M can be used to generate exact samples
from π in expected O(τ ) time, then we can uniformly sample from Ωk in expected O(cτ 2) time.
Proof. Let M have conductance Φ and assume k < R. Considering the cut S = Ω≤k and
using Lemma 2.2, we have min(π(S), π(S)) ≥ 1/(c + 1) for the balanced bias λt. It follows
that
Φ(S) ≤
σ∈S,ρ∈S π(σ)P (σ, ρ)
min(π(S), π(S))
≤ (c + 1)
π(σ)P (σ, ρ) ≤ (c + 1) π(Ωk).
(cid:88)
σ∈Ωk,ρ∈Ωk+1
(cid:80)
By Theorem 2.1, we have
Φ ≥
1
2(τ + 1)
,
so
π(Ωk) ≥
1
2(c + 1)(τ + 1)
.
It follows that O(cτ ) samples from π are needed in expectation to generate a uniform σ ∈ Ωk
for any fixed k ∈ [R − 1] with the given balanced bias. Moreover, if each sample is exactly
generated in O(τ ) expected time, then the total running time of this sampling algorithm
is O(cτ 2). The argument when k = R − 1 extends to k = R by the detailed balance
equation.
8
For simplicity, this theorem assumes we have a method for generating samples exactly
from π. In many graded posets, including all considered here, we can use the coupling from
the past algorithm to generate perfect samples in expected O(τ ) steps per sample [30]. In
cases when we cannot sample exactly, we have the following corollary of Theorem 2.2 that
only requires samples be chosen close to π.
Corollary 2.1. We can use M to approximately generate samples from Ωk to within ε of
the total variation distance of π in expected O(cτ 2 max(log(ε−1), log(cτ ))) time.
(cid:18)
ε,
(cid:19)
1
8(c + 1)(τ + 1)
Proof. Let
ε∗ = min
be the desired bound on the total variation distance between the t-step distribution P t(σ,·)
starting from any initial σ ∈ Ω and the stationary distribution π. Then
P t(σ, Ωk) − π(Ωk).
(cid:88)
P t(σ, ρ) − π(ρ) ≥ 1
2
ε∗ ≥ 1
2
ρ∈Ω
Theorem 2.2 and our choice of ε∗ imply that
P t(σ, Ωk) ≥ π(Ωk) − 2ε∗ ≥
1
4(c + 1)(τ + 1)
.
Each sample can be generated in O(τ (ε)) = O(τ log(ε−1)) steps, so the expected runtime is
O(cτ 2 max(log(ε−1), log(cτ ))).
3 Sampling Integer Partitions
We demonstrate how to use the balanced bias technique to sample from general classes
of restricted integer partitions. Integer partitions have a natural representation as Young
diagrams, which formally are finite subsets σ ⊆ Z2≥0 with the property that if (a, b) ∈ σ, then
{(x, y) ∈ Z2≥0 : 0 ≤ x ≤ a and 0 ≤ y ≤ b} ⊆ σ.
Young diagrams can be visualized as a connected set of unit squares on the integer lattice
with a corner at (0, 0) and a nonincreasing upper boundary from left to right. Each square in
the Young diagram must be supported below by the x-axis or another square and supported
to the left by the y-axis or another square. We are interested in region-restricted Young
diagrams, a variant of Young diagrams whose squares are restricted to lie in a connected
region R ⊆ Z2≥0 such that each square is supported below and to the left by the boundary
of R or another square. Note that we use R in this section to denote a region instead of the
rank of a poset. We will see that the rank of the poset induced by the natural partial order
on R-restricted Young diagrams is R.
9
We call Young diagrams σ ⊆ Z2≥0 such that σ = n unrestricted integer partitions of n
and use this term interchangeably with integer partitions. Many well-studied classes of
restricted integer partitions have natural interpretations as region-restricted Young diagrams.
For example, the set of integer partitions of n with at most k parts and with each part at
most size (cid:96) give rise to the Gaussian binomial coefficients and can be thought of as the set
of Young diagrams of size n contained in a k × (cid:96) box.
Figure 1: Unrestricted and restricted integer partitions.
3.1 The Biased Markov Chain
Let the state space Ω be the set of all Young diagrams restricted to lie in a region R. Young
diagrams have a natural graded partial order via inclusion, where σ ≤ ρ if and only if σ ⊆ ρ,
so the rank of a diagram σ is r(σ) = σ. The following Markov chain M on the Hasse
diagram of this partial order makes transitions that add or remove a square on the boundary
of the diagram in each step according to the Metropolis-Hastings algorithm. Therefore the
stationary distribution is a Boltzmann distribution parameterized by a bias value λ. Let R
be a region such that every partition restricted to this region has at most ∆ neighboring
configurations.
Biased Markov Chain on Integer Partitions M
Starting at any Young diagram σ0 ⊆ R, repeat:
• Choose a neighbor ρ of σt uniformly at random with probability 1/2∆.
• Set σt+1 = ρ with probability min(1, λρ−σt).
• With all remaining probability, set σt+1 = σt.
10
The state space Ω is connected, because any configuration can eventually reach the
minimum configuration σ = ∅ with positive probability. By construction, M is lazy (i.e.,
it is always possible that σt = σt+1), so it follows that M is an ergodic Markov chain, and
hence has a unique stationary distribution π. Using the detailed balance equation for Markov
chains [31], we see that π(σ) = λσ, for all σ ∈ Ω.
This Markov chain can be used to efficiently approximate the number of partitions of n
restricted to R within arbitrarily small specified relative error, because this problem is self-
reducible [16]. Observe that we can run M restricted to R polynomially many times and
compute the mean height m in the first column of the sampled Young diagrams. Then we
use M to recursively approximate the number of partitions of n− m restricted to the region
R(cid:48) = {(x, y) ∈ R : 1 ≤ x and y ≤ m},
and return the product of m and this approximation.
3.2 Sampling Using Balanced Bias
In the following general sampling theorem for restricted integer partitions, the mixing time
of M must hold for all bias parameters λt.
Theorem 3.1. Let τ = τ (e−1) be the mixing time of M on the region R. We can uniformly
sample partitions of k restricted to a region R in expected O(∆τ 2) time.
Proof. There is only one such partition when k = 0 or k = R, so assume k ∈ [R − 1]. By
construction Ωk+1/Ωk ≤ ∆ for all fixed k, so 1 ≤ Ωi ≤ ∆i for all i ∈ {0, 1, . . . ,R}. By
Lemma 2.2 there exists a balanced bias λ, which we can identify adaptively in O(log(R))
time with a binary search as we are sampling, since Boltzmann distributions increase mono-
tonically with increasing λt. Therefore, we can generate Young diagrams restricted to R
with any fixed rank in expected O(∆τ 2) steps of M by Theorem 2.2.
If more is known about the number of elements at each rank or the geometry of R, then
we can give better bounds on the runtime of this algorithm. For example, if R is the region
of a skew Young diagram (see Figure 1), a region contained between two Young diagrams,
then we can adapt Levin and Peres' mixing results about biased exclusion processes to this
setting.
Theorem 3.2 ([23]). Consider the biased exclusion process with bias β = βn = 2pn − 1 > 0
on the segment of length 2n and with n particles. Set α =(cid:112)pn/(1 − pn). For ε > 0, if n is
large enough, then
τ (ε) ≤ 4n
β2
(cid:34)
log(cid:0)ε−1(cid:1) + log
(cid:34)
α
(cid:18) αn − 1
α − 1
(cid:19)2(cid:35)(cid:35)
.
Corollary 3.1. If the region R is a skew Young diagram contained in an n × n box, we can
uniformly sample partitions of k restricted to R in expected O(n16) time.
11
Proof. The biased exclusion process on a segment of length 2n with n particles is in bijection
with M when the restricting region is an n × n box. The proof of Theorem 3.2 in [23] uses
a path coupling argument that directly extends to and gives an upper bound for the mixing
time of M when the region R is a skew Young diagram, since the expected change in distance
of two adjacent states in the more restricted setting can only decrease. Let λn,t denote λt in
an instance of size n. We analyze the three cases λn,t < 1, λn,t = 1, and λn,t > 1, and then
bound the mixing time of M for all λn,t.
To prove the existence of a balanced bias using Lemma 2.2, observe that aΩ,i ≤ p(i) ≤ 2i
for all i ∈ {0, 1, . . . , n2}. In the first case, assume λt,n < 1. Then we have t ∈ [n2 − 1] since
R = n2. Translating M to the biased exclusion process,
pn,t =
βn,t =
and
,
,
1
1 + λn,t
1 − λn,t
1 + λn,t
(cid:115)
1
λn,t
To use Theorem 3.2, we first prove 1/β2
minimizes βn,t, hence maximizes the desired quantity. Then
.
αn =
n,t ≤ 10n4. To see this, observe that t = n2 − 1
(cid:19)2
lim
n→∞
βn,n2−1
(cid:18) 1
(cid:19)2 ≤ αn,t
=
(cid:0)nαn
n,t
4n2
log(2)2 ≤ 10n4.
(cid:1)2 ≤ n2
λn+1/2
n,t
≤ n22n+1/2,
Next, since an,t > 1, we have
n,t − 1
αn,t − 1
αn,t
(cid:18) αn
because λn,t ≥ 1/2. Thus, τ (ε) = O(n5(log(ε−1) + n)) for all λn,t by Theorem 3.2.
is Θ(n3 log(n/ε)). In the third case, λt,n > 1 so t ∈ {n2 + 1, n2 + 2, . . . , n4},
In the unbiased case when λn,t = 1, Wilson [35] proved that the mixing time of M
and
pn,t =
λn,t
,
,
1 + λn,t
λn,t − 1
1 + λn,t
βn,t =
αn,t =(cid:112)λn,t.
12
By similar analysis, 1/β2
n,t ≤ 10n4 and
(cid:18) αn
n,t − 1
αn,t − 1
(cid:19)2 ≤ n2λn+1/2
n,t
≤ n2(cid:16)
2n2−1(cid:17)n+1/2
,
αn,t
since λn,t ≤ 2n2−1. Thus τ (ε) = O(n5(log(ε−1) + n3)), so by Theorem 2.2 we can uniformly
sample partitions of k restricted to R in expected O(n16) time.
3.3 Sampling Using Log-concavity
When more is known about the stationary distribution π, specifically the sequence {Ωi}∞
i=0,
we can typically improve the bounds on the running time of our algorithm.
In particu-
lar, we show that we can sample unrestricted integer partitions in expected O(n9/4) time.
Our primary techniques involve using a compressed representation of partitions and using
log-concavity to show strong probability concentration around partitions of the desired size.
These techniques extend to a variety of settings where log-concavity or probability concen-
tration can be shown.
To sample integer partitions of n, we set the bias parameter λn = p(n−1)/p(n) to force the
stationary distribution to concentrate at n. The sequence {p(k)}∞
k=26 is log-concave [9, 27], so
it follows that the sequence {p(k)λk
k=26 is, too. Log-concave sequences of positive terms are
unimodal, which implies that the mode of our stationary distribution is at k = n. Moreover,
we show how log-concavity gives exponential decay on both sides of the mode, and hence
strong concentration.
n}∞
We now argue that we need only consider Young diagrams that lie under the curve
y = 2n/x to sample partitions of n, as all Young diagrams with squares above that curve
must have more than 2n squares total.
Proposition 3.1. A Young diagram that lies under the curve y = 2n/x can be stored in
O(n1/2 log(n)) space.
√
2n,
Proof. For any square in the Young diagram, both of its coordinates are not greater than
for then it would lie above y = 2n/x. We may record the height of each column and the
2n(cid:99) − 1} to capture the position of every square
2n(cid:99) heights
width of each row in the range {0, 1, . . . ,(cid:98)√
in the diagram. Therefore, we can represent the diagram using exactly these 2(cid:98)√
and widths.
Using the compressed representation in the previous proposition, we see that there will
not be more than O(n1/2) possible transitions at any possible state, since our algorithms
adds or removes at most one square on the upper boundary in each step. Note that we can
adapt this technique in the general case for any region R that lies under the curve y = 2n/x.
√
Proposition 3.2. There are at most 4
2n potential transitions for any Young diagram that
lies under the curve y = 2n/x.
13
Proof. Observe that since the squares in any row or column must be connected, there are at
most two valid moves in any particular row or column. Therefore, by Proposition 3.1, there
2n(cid:99) possible transitions from any such Young diagram.
are at most 4(cid:98)√
We now shift our attention to bounding λn and the consequences it has on both the
mixing time of M and the concentration of π. Hardy and Ramanujan [14] gave the classical
asymptotic formula for the partition numbers
√
p(n) ∼ 1
4
2n/3,
√
3n
eπ
and we use related bounds given in [9] for the following lemma. The proof is deferred to the
next subsection.
Lemma 3.1. For all n ≥ 30, we have
1 − 2√
n
< λn < 1 − 1√
n
.
Theorem 3.3. The Markov chain M with bias λn restricted to the region R bounded by the
curve y = 2n/x mixes in O(n3/2(log(ε−1) + n1/2)).
β = (1− λ)/(1 + λ), and α =(cid:112)1/λ. By Proposition 3.2, there are at most 4
Proof. We modify Theorem 3.2 and its proof in [23]. In this biased exclusion process, λ = λn,
2n transitions
√
from any state, so for n large enough
√
τ (ε) ≤ 8
2n
β2
(cid:2)log(cid:0)ε−1(cid:1) + log (diam (Ω))(cid:3) ,
where diam(Ω) is the maximum length path between any two states, as defined in [24].
Therefore, we have diam(Ω) ≤ Rα2n and
R ≤ 2nH2n ≤ 2n(log(2n) + 1),
so
diam(Ω) ≤ 2n(log(2n) + 1)α2n = 2n(log(2n) + 1)λ−n.
By Lemma 3.1 and the bound 1 + x ≤ ex, for all x ∈ R,
log(cid:0)λ−n(cid:1) ≤ log
(cid:18)(cid:18)
(cid:19)n(cid:19)
√
≤ 3
n,
1 +
2√
n − 2
for n sufficiently large. We have
√
≤ 2
n − 1
1
β
by Lemma 3.1. Therefore, τ (ε) = O(n3/2(log(ε−1) + n1/2)).
14
Another key observation we make to generate partitions of n more efficiently is to sal-
vage samples larger than n instead of rejecting them, while preserving uniformity on the
distribution Ωn. For any k ≥ 0, consider the function fk : Ωn → Ωn+k that maps a parti-
tion σ = (σ1, σ2, . . . , σm) to fk(σ) = (σ1 + k, σ2, . . . , σm). Note that σ1 ≥ σ2 ≥ ··· ≥ σm
since σ is a Young diagram. Clearly fk is injective, so we can consider the inverse map
k ((ρ1, ρ2, . . . , ρ(cid:96))) that subtracts k from ρ1 if ρ1 − k ≥ ρ2, and is invalid otherwise. Then,
f−1
define g : Ω≥n → Ωn ∪ {0} as
g((ρ1, ρ2, . . . , ρ(cid:96))) =
(ρ1 − k, ρ2, . . . , ρ(cid:96))
0
if ρ1 + ρ2 + ··· + ρ(cid:96) = n + k and ρ1 − k ≥ ρ2
otherwise.
(cid:40)
The following lemma, whose proof is deferred to the next subsection, uses the log-concavity
of the partition numbers to give a strong lower bound on the success of the map g.
Lemma 3.2. Let σ be a random Young diagram from the stationary distribution of M, and
let g be the function defined above. Then for all n sufficiently large,
Pr[g(σ) generates a partition of n] ≥
1
160n1/4 .
Assembling the ideas in this section, we now formally present our Markov chain Monte
Carlo algorithm for generating partitions of n uniformly at random.
Algorithm for Sampling Integer Partitions
Repeat until success:
• Sample σ ∈ Ω using M.
• If n ≤ σ ≤ 2n and g(σ) (cid:54)= 0, return g(σ).
Note that we restrict σ ≤ 2n instead of σ ≤ 2n log(n) so that g maps to Ωn uniformly.
All partitions of 2n are elements of Ω2n, but the same is not true for larger partitions since
the bounding region R is the curve y = 2n/x. Lastly, recall that coupling from the past can
be used efficiently in this setting to generate perfectly uniform samples, because the natural
coupling is monotone and there is a single minimum and maximum configuration [13].
Theorem 3.4. Our Markov chain Monte Carlo algorithm for generating a uniformly random
partition of n runs in expected O(n9/4) time and O(n1/2 log(n)) space.
Proof. The proof directly follows from Proposition 3.1, Lemma 3.2, and Theorem 3.3.
3.4 Proofs of Lemma 3.1 and Lemma 3.2
We prove Lemma 3.1 using bounds for p(n) given in [9]. Let
√
24n − 1
6
,
µ(n) = µn =
π
15
and
T (n) = νn
(cid:19)
√
12
24n − 1
,
eµn +
(−1)n√
2
eµn/2
(cid:21)
.
ν(n) = νn =
(cid:20)(cid:18)
1 − 1
µn
The function T (n) is the sum of the three largest terms in the Hardy-Ramanujan formula,
and the explicit error bounds in [9] that we use were first proved by Lehmer [22]. We only
prove upper bounds in the following two proofs, as the lower bounds are proved similarly.
Lemma 3.3. For all n ≥ 2, we have
(cid:18)
(cid:12)(cid:12)(cid:12)(cid:12)p(n) − νn
1 − 1
µn
(cid:19)
Proof. By Lemma 2.3 and Proposition 2.4 in [9],
p(n) < T (n) + 1 +
eµn/2 < νn
16
µ3
n
Proof of Lemma 3.1. By Lemma 3.3,
(cid:16)
(cid:17)
(cid:17)
(cid:16)
1 − 1
1 − 1
µn−1
λn <
1 + eµn−1/2 + νn−1
− (1 + eµn/2) + νn
− (e−µn + e−µn/2) + νn
for all n ≥ 14, because the lower bound for p(n) is initially negative. We have
eµn
µn
eµn−1
=
eµn−1
eµn
(cid:16)
(cid:16)
(cid:17)
µn−1
1 − 1
1 − 1
µn
(cid:17)
,
eµn
1 − 1
µn
eµ
n + 1 + eµn/2.
(cid:12)(cid:12)(cid:12)(cid:12) < 1 + eµn/2.
(cid:19)
(cid:18)
e−µn−1 + e−µn−1/2 + νn−1
(cid:18)
=
(cid:19)
(cid:18) n
(cid:19)2(cid:18) µ3
1 − 1
µn
νn
n
,
eµn−1
eµn
n − 1
for all n ≥ 65, so it follows that
e−µn + e−µn/2 <
(cid:0)1 + 1
(cid:0)1 − 1
n−1
(cid:16)
(cid:1) νn−1
(cid:16)
(cid:1) νn
n
(cid:17)
1 − 1
1 − 1
µn
µn−1
(cid:17)
λn <
eµn−1
eµn
√
for all n ≥ 2. Observe that µn−1 − µn < −π/
6n and
n (µn−1 − 1)
µ3
n−1 (µn − 1)
µ3
<
for all n ≥ 2. Using ex ≤ 1 + x + x2/2, for all x ≤ 0,
n
n − 1
,
(cid:18) n
(cid:19)3
(cid:18) n
(cid:18)
(cid:19)3 ≤
− π√
(cid:19)
,
n (µn−1 − 1)
n−1 (µn − 1)
µ3
(cid:19)(cid:18) n
(cid:19)3
eµn−1
eµn
< e
n − 1
λn <
,
where the final inequality is true for all n ≥ 160. When 30 ≤ n < 160, we verify the claim
numerically.
n − 1
n − 1
+
6n
1 − π√
6n
π2
12n
< 1 − 1√
n
16
Now we prove Lemma 3.2 by showing that the truncation scheme g(σ) succeeds with
sufficient probability. By Hardy-Ramanujan formula, we have that for any constant c > 0
and n sufficiently large,
1 − c
√
3n
4
eπ
√
√
2n/3 ≤ p(n) ≤ 1 + c
3n
4
√
2n/3.
eπ
Letting λ = λn, the Hardy-Ramanujan formula implies that for all n ≥ 20,
√
e−πk/
√
6n ≤ λk ≤ ek/n−πk/
6n.
Lemma 3.4. Let Zn be the normalizing constant of the desired distribution. Then we have
for all n sufficiently large.
Proof. Clearly
Zn < 40n3/4λnp(n),
∞(cid:88)
k=0
Zn ≤
p(k)λk.
We further know that f (k) = p(k)λk is unimodal with a maximum at k = n. By the
log-concavity of {f (k)}∞
k=26, we have
f (n + k)
f (n)
≥ f (n + 2k)
f (n + k)
and
f (n − k)
for all k ≥ 1. Therefore, we can bound Zn as
≥ f (n − 2k)
f (n − k)
,
f (n)
(cid:32)
Zn ≤ kf (n)
1
1 − f (n+k)
f (n)
+
1
1 − f (n−k)
f (n)
(cid:33)
,
for any k ≥ 1. Specifically, if both f (n + k)/f (n) and f (n − k)/f (n) are at most some fixed
constant less than 1, then Zn = O(kf (n)). Using the bounds above,
√
ek/n−π(
√
ek/n−π(k−2
6n =
k−√
n/6.
√
/
kn)/
√
n)2
6n+π
√
f (k) = p(k)λk ≤ 1 + c
3k
√
4
√
1 + c
4
3k
√
Letting n + k = (
n + n1/4)2, for n large enough,
(cid:16)(cid:0)√
n + n1/4(cid:1)2(cid:17) ≤ 1 + c
f
√
4
3n
√
e1.1−π/
6+π
√
n/6.
17
We can then bound the density value at (
f (n)
Taking c ≤ 0.01, we have
(cid:16)(cid:0)√
f
√
√
≤ 1+c
4
3n
1−c
√
3n
4
n + n1/4(cid:1)2(cid:17)
(cid:16)(cid:0)√
n + n1/4(cid:1)2(cid:17)
(cid:16)(cid:0)√
f (n)
f
f
≤ 1.01
0.99
n − n1/4(cid:1)2(cid:17)
Similarly, for n sufficiently large,
n + n1/4)2 relative to the maximum by
√
e1.1−π/
√
eπ
√
e1.1−π/
1 + c
1 − c
√
√
2n/3−π
6+π
6.
=
n/6
n/6
e1.1− π√
6 < 0.85.
f (n)
< 0.85.
Therefore, we have Zn < 40n3/4λnp(n) using the fact that k ≤ 3n3/4.
Proof of Lemma 3.2. We use Lemma 3.4 and Lemma 3.1 to bound the probability that g(σ)
generates a partition of n successfully. Therefore, we have
Pr[g(σ) generates a partition of n] =
n(cid:88)
λn+kp(n)
Zn
n(cid:88)
n(cid:88)
k=0
k=0
√
n
2
k=0
≥ 1
40n3/4
≥ 1
40n3/4
=
≥
1
40n3/4 ·
1
160n1/4 .
(cid:19)k
λk
(cid:18)
(cid:32)
1 − 2√
n
(cid:18)
1 −
1 − 2√
n
(cid:19)n+1(cid:33)
4 Sampling in Other Graded Posets
We demonstrate the versatility of using a Markov chain on the Hasse diagram of a graded
poset to sample elements of fixed rank. When this chain is rapidly mixing for all λt
with t ∈ [R2], we can apply Boltzmann sampling and Theorem 2.2 to generate approxi-
mately uniform samples in polynomial time. Similar to region-restricted integer partitions,
analogous notions of self-reducibility apply to restricted families of permutations and lozenge
tilings, so there exist fully polynomial-time approximation schemes for these enumerations
problems because we can efficiently sample elements of a given rank from their respective
posets [16].
18
4.1 Permutations with Fixed Rank
In the first case, we consider permutations of n elements with a fixed number of inversions.
The Hasse diagram in this setting connects permutations that differ by one adjacent trans-
position. This partial order is in bijection with the weak Bruhat order on the symmetric
group.
In the unbiased case (λ = 1), the nearest neighbor Markov chain mixes in time
Θ(n3 log(n)) [35]. With constant bias the chain is known to converge in time Θ(n2) [5, 13].
The number of permutations of n with k inversions is known to be log-concave in k, so
standard Boltzmann sampling techniques can be used. However, using our balanced bias
method, we avoid the need for bounds on the growth of inversion numbers in restricted
settings, such as permutations where at least i of the first j elements are in the first half of
the permutation. Figure 2 illustrates the distribution of inversions of random permutations
in S100 sampled from various ranks of the inversion poset as Rothe diagrams [21].
(a)
(b)
(c)
(d)
(e)
(f)
(cid:0)100
(cid:1) inversions.
2
Figure 2: Random permutations with (a) 5, (b) 20, (c) 40, (d) 60, (e) 80, (f) 95 percent of
4.2 Lozenge Tilings with Fixed Average Height
Lozenge tilings are tilings of a triangular lattice region with pairs of equilateral triangles
that share an edge. There is a well-studied height function that maps hexagonal lozenge
tilings bijectively to plane partitions lying in an n× n× n box (see, e.g., [25]), and it follows
that lozenge tilings with a fixed average height of k are precisely the plane partitions with
19
volume k. The Markov chain that adds or removes single cubes on the surface of the plane
partition (corresponding to rotating three nested lozenges 180 degrees) is known to mix
rapidly in the unbiased case. Caputo et al. [7] studied the biased version of this chain with
a preference toward removing cubes, and showed that this chain converges in O(n3) time.
Applying the balanced bias method, we can use Boltzmann sampling to generate random
lozenge tilings with any target average height in polynomial time, as shown in Figure 3.
(a)
(b)
(c)
(d)
Figure 3: Random lozenge tilings with average height (a) 5, (b) 15, (c) 35, (d) 50 percent
of 753.
20
References
[1] D. Aldous. Deterministic and stochastic models for coalescence (aggregation and co-
agulation): a review of the mean-field theory for probabilists. Bernoulli, 5(1):3 -- 48,
1999.
[2] G. E. Andrews. The Theory of Partitions. Cambridge mathematical library. Cambridge
University Press, 1998.
[3] R. Arratia and S. DeSalvo. Probabilistic divide-and-conquer: a new exact simulation
method, with integer partitions as an example. Combinatorics, Probability and Com-
puting, 25(3):324 -- 351, 2016.
[4] A. Ayyer, J. Bouttier, S. Corteel, and F. Nunzi. Multivariate juggling probabilities.
Electronic Journal of Probability, 20(5):1 -- 29, 2014.
[5] I. Benjamini, N. Berger, C. Hoffman, and E. Mossel. Mixing times of the biased card
shuffling and the asymmetric exclusion process. Transactions of the American Mathe-
matical Society, 357(8):3013 -- 3029, 2005.
[6] N. Berestycki and J. Pitman. Gibbs distributions for random partitions generated by a
fragmentation process. Journal of Statistical Physics, 127:381 -- 418, 2007.
[7] P. Caputo, F. Martinelli, and F. L. Toninelli. Convergence to equilibrium of biased
plane partitions. Random Structures & Algorithms, 39(1):83 -- 114, 2011.
[8] A. Comtet, S. N. Majumdar, and S. Ouvry. Integer partitions and exclusion statistics.
Journal of Physics A: Mathematical and Theoretical, 40(37):11255 -- 11269, 2007.
[9] S. DeSalvo and I. Pak. Log-concavity of the partition function. The Ramanujan Journal,
38(1):61 -- 73, 2014.
[10] P. Duchon, P. Flajolet, G. Louchard, and G. Shaeffer. Boltzmann samplers for the ran-
dom generation of combinatorial structures. Combinatorics, Probability and Computing,
13(4 -- 5):577 -- 625, 2004.
[11] P. Flajolet, ´E. Fusy, and C. Pivoteau. Boltzmann sampling of unlabelled structures.
In Proceedings of the Fourth Workshop on Analytic Algorithmics and Combinatorics
(ANALCO), pages 201 -- 211, 2007.
[12] B. Fristedt. The structure of random partitions of large integers. Transactions of the
American Mathematical Society, 337(2):703 -- 735, 1993.
[13] S. Greenberg, A. Pascoe, and D. Randall. Sampling biased lattice configurations using
exponential metrics. In Proceedings of the Twentieth Annual ACM-SIAM Symposium
on Discrete Algorithms (SODA), pages 76 -- 85, 2009.
21
[14] G. H. Hardy and S. Ramanujan. Asymptotic formulae in combinatory analysis. Pro-
ceedings of the London Mathematical Society, 17:75 -- 115, 1918.
[15] G. D. James and A. Kerber. The representation theory of the symmetric group. Ency-
clopedia of Mathematics and its Applications. Cambridge University Press, 1984.
[16] M. Jerrum. Counting, Sampling and Integrating: Algorithms and Complexity. Lectures
in Mathematics. ETH Zurich. Birkhauser Basel, 2003.
[17] M. Jerrum and A. Sinclair. Approximate counting, uniform generation and rapidly
mixing Markov chains. Information and Computation, 82:93 -- 133, 1989.
[18] M. Jerrum and A. Sinclair. The Markov chain Monte Carlo method: an approach
to approximate counting and integration. In D. S. Hochbaum, editor, Approximation
Algorithms for NP-hard Problems, pages 482 -- 520. PWS Publishing, 1997.
[19] M. R. Jerrum, A. J. Sinclair, and E. Vigoda. A polynomial-time approximation algo-
rithm for the permanent of a matrix with nonnegative entries. Journal of the ACM,
41:671 -- 697, 2006.
[20] R. Kenyon and A. Okounkov. Limit shapes and the complex Burgers equation. Acta
Mathematica, 199:263 -- 302, 2007.
[21] A. Kerber. Applied Finite Group Actions. Algorithms and Combinatorics. Springer
Berlin Heidelberg, 2013.
[22] D. H. Lehmer. On the series for the partition function. Transactions of the American
Mathematical Society, 43(2):271 -- 295, 1938.
[23] D. Levin and Y. Peres. Mixing of the exclusion process with small bias. Preprint available
at https://arxiv.org/abs/1608.03633, 2016.
[24] D. Levin, Y. Peres, and E. Wilmer. Markov chains and mixing times. American Math-
ematical Society, 1st edition, 2008.
[25] M. Luby, D. Randall, and A. J. Sinclair. Markov chain algorithms for planar lattice
structures. SIAM Journal on Computing, 31:167 -- 192, 2001.
[26] N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller. Equa-
tion of State Calculations by Fast Computing Machines. The Journal of Chemical
Physics, 21(6):1087 -- 1092, 1953.
[27] J. L. Nicolas. Sur les entiers n pour lesquels il y a beaucoup de groupes ab´eliens d'ordre
n. Annales de l'institut Fourier, 28(4):1 -- 16, 1978.
[28] A. Nijenhuis and H. S. Wilf. Combinatorial algorithms. Academic Press, 1978.
22
[29] A. Okounkov. Symmetric Functions and Random Partitions, pages 223 -- 252. Springer
Netherlands, Dordrecht, 2002.
[30] J. G. Propp and D. B. Wilson. Exact sampling with coupled markov chains and ap-
plications to statistical mechanics. Random Structures & Algorithms, 9(1 -- 2):223 -- 252,
1996.
[31] D. Randall. Mixing. In 44th Symposium on Foundations of Computer Science (FOCS),
pages 4 -- 15, 2003.
[32] A. Sinclair. Improved bounds for mixing rates of Markov chains and multicommodity
flow. Combinatorics, Probability and Computing, 1:351 -- 370, 1992.
[33] A. Sinclair. Algorithms for random generation and counting. Progress in theoretical
computer science. Birkhauser, 1993.
[34] M. Squire. Efficient generation of integer partitions. Unpublished manuscript, 1993.
[35] D. B. Wilson. Mixing times of lozenge tiling and card shuffling Markov chains. The
Annals of Applied Probability, 14(1):274 -- 325, 2004.
23
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.